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)
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. */
24 #include "coretypes.h"
31 #include "basic-block.h"
36 #include "diagnostic.h"
38 #include "tree-dump.h"
39 #include "tree-flow.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. */
65 /* If this edge creates a simple equivalence, the LHS and RHS of
66 the equivalence will be stored here. */
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
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
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
107 A NULL node is used to mark the last node associated with the
109 static VEC(tree_on_heap) *block_defs_stack;
111 /* Stack of statements we need to rescan during finalization for newly
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
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
133 /* The value (lhs) of this expression. */
136 /* The expression (rhs) we want to record. */
139 /* The annotation if this element corresponds to a statement. */
142 /* The hash value for RHS/ann. */
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. */
174 long num_exprs_considered;
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. */
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. */
222 /* The actual conditional we recorded. This is needed since we compute
226 /* The basic block where this record was created. We use this to determine
227 when to remove records. */
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. */
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;
262 /* Local functions. */
263 static void optimize_stmt (struct dom_walk_data *,
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 *,
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 *,
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. */
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);
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));
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
337 free_all_edge_infos (void)
345 FOR_EACH_EDGE (e, ei, bb->preds)
347 struct edge_info *edge_info = e->aux;
351 e->aux = edge_info->redirection_target;
352 if (edge_info->cond_equivalences)
353 free (edge_info->cond_equivalences);
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
367 tree_ssa_dominator_optimize (void)
369 struct dom_walk_data walk_data;
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
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
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
422 /* Optimize the dominator tree. */
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;
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
492 for (i = 0; i < num_ssa_names; i++)
494 tree name = ssa_name (i);
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);
514 gate_dominator (void)
516 return flag_tree_dom != 0;
519 struct tree_opt_pass pass_dominator =
522 gate_dominator, /* gate */
523 tree_ssa_dominator_optimize, /* execute */
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 */
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. */
542 thread_across_edge (struct dom_walk_data *walk_data, edge e)
544 block_stmt_iterator bsi;
548 /* Each PHI creates a temporary equivalence, record them. */
549 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
551 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
552 tree dst = PHI_RESULT (phi);
553 record_const_or_copy (dst, src);
554 register_new_def (dst, &block_defs_stack);
557 for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
559 tree lhs, cached_lhs;
561 stmt = bsi_stmt (bsi);
563 /* Ignore empty statements and labels. */
564 if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
567 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
568 value, then stop our search here. Ideally when we stop a
569 search we stop on a COND_EXPR or SWITCH_EXPR. */
570 if (TREE_CODE (stmt) != MODIFY_EXPR
571 || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
574 /* At this point we have a statement which assigns an RHS to an
575 SSA_VAR on the LHS. We want to prove that the RHS is already
576 available and that its value is held in the current definition
577 of the LHS -- meaning that this assignment is a NOP when
578 reached via edge E. */
579 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
580 cached_lhs = TREE_OPERAND (stmt, 1);
582 cached_lhs = lookup_avail_expr (stmt, false);
584 lhs = TREE_OPERAND (stmt, 0);
586 /* This can happen if we thread around to the start of a loop. */
587 if (lhs == cached_lhs)
590 /* If we did not find RHS in the hash table, then try again after
591 temporarily const/copy propagating the operands. */
594 /* Copy the operands. */
595 stmt_ann_t ann = stmt_ann (stmt);
596 use_optype uses = USE_OPS (ann);
597 vuse_optype vuses = VUSE_OPS (ann);
598 tree *uses_copy = xcalloc (NUM_USES (uses), sizeof (tree));
599 tree *vuses_copy = xcalloc (NUM_VUSES (vuses), sizeof (tree));
602 /* Make a copy of the uses into USES_COPY, then cprop into
604 for (i = 0; i < NUM_USES (uses); i++)
608 uses_copy[i] = USE_OP (uses, i);
609 if (TREE_CODE (USE_OP (uses, i)) == SSA_NAME)
610 tmp = SSA_NAME_VALUE (USE_OP (uses, i));
611 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
612 SET_USE_OP (uses, i, tmp);
615 /* Similarly for virtual uses. */
616 for (i = 0; i < NUM_VUSES (vuses); i++)
620 vuses_copy[i] = VUSE_OP (vuses, i);
621 if (TREE_CODE (VUSE_OP (vuses, i)) == SSA_NAME)
622 tmp = SSA_NAME_VALUE (VUSE_OP (vuses, i));
623 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
624 SET_VUSE_OP (vuses, i, tmp);
627 /* Try to lookup the new expression. */
628 cached_lhs = lookup_avail_expr (stmt, false);
630 /* Restore the statement's original uses/defs. */
631 for (i = 0; i < NUM_USES (uses); i++)
632 SET_USE_OP (uses, i, uses_copy[i]);
634 for (i = 0; i < NUM_VUSES (vuses); i++)
635 SET_VUSE_OP (vuses, i, vuses_copy[i]);
640 /* If we still did not find the expression in the hash table,
641 then we can not ignore this statement. */
646 /* If the expression in the hash table was not assigned to an
647 SSA_NAME, then we can not ignore this statement. */
648 if (TREE_CODE (cached_lhs) != SSA_NAME)
651 /* If we have different underlying variables, then we can not
652 ignore this statement. */
653 if (SSA_NAME_VAR (cached_lhs) != SSA_NAME_VAR (lhs))
656 /* If CACHED_LHS does not represent the current value of the underlying
657 variable in CACHED_LHS/LHS, then we can not ignore this statement. */
658 if (var_ann (SSA_NAME_VAR (lhs))->current_def != cached_lhs)
661 /* If we got here, then we can ignore this statement and continue
662 walking through the statements in the block looking for a threadable
665 We want to record an equivalence lhs = cache_lhs so that if
666 the result of this statement is used later we can copy propagate
668 record_const_or_copy (lhs, cached_lhs);
669 register_new_def (lhs, &block_defs_stack);
672 /* If we stopped at a COND_EXPR or SWITCH_EXPR, then see if we know which
673 arm will be taken. */
675 && (TREE_CODE (stmt) == COND_EXPR
676 || TREE_CODE (stmt) == SWITCH_EXPR))
678 tree cond, cached_lhs;
682 /* Do not forward entry edges into the loop. In the case loop
683 has multiple entry edges we may end up in constructing irreducible
685 ??? We may consider forwarding the edges in the case all incoming
686 edges forward to the same destination block. */
687 if (!e->flags & EDGE_DFS_BACK)
689 FOR_EACH_EDGE (e1, ei, e->dest->preds)
690 if (e1->flags & EDGE_DFS_BACK)
696 /* Now temporarily cprop the operands and try to find the resulting
697 expression in the hash tables. */
698 if (TREE_CODE (stmt) == COND_EXPR)
699 cond = COND_EXPR_COND (stmt);
701 cond = SWITCH_COND (stmt);
703 if (COMPARISON_CLASS_P (cond))
705 tree dummy_cond, op0, op1;
706 enum tree_code cond_code;
708 op0 = TREE_OPERAND (cond, 0);
709 op1 = TREE_OPERAND (cond, 1);
710 cond_code = TREE_CODE (cond);
712 /* Get the current value of both operands. */
713 if (TREE_CODE (op0) == SSA_NAME)
715 tree tmp = SSA_NAME_VALUE (op0);
716 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
720 if (TREE_CODE (op1) == SSA_NAME)
722 tree tmp = SSA_NAME_VALUE (op1);
723 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
727 /* Stuff the operator and operands into our dummy conditional
728 expression, creating the dummy conditional if necessary. */
729 dummy_cond = walk_data->global_data;
732 dummy_cond = build (cond_code, boolean_type_node, op0, op1);
733 dummy_cond = build (COND_EXPR, void_type_node,
734 dummy_cond, NULL, NULL);
735 walk_data->global_data = dummy_cond;
739 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
740 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
741 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
744 /* If the conditional folds to an invariant, then we are done,
745 otherwise look it up in the hash tables. */
746 cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
747 if (! is_gimple_min_invariant (cached_lhs))
749 cached_lhs = lookup_avail_expr (dummy_cond, false);
750 if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
751 cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
756 /* We can have conditionals which just test the state of a
757 variable rather than use a relational operator. These are
758 simpler to handle. */
759 else if (TREE_CODE (cond) == SSA_NAME)
762 cached_lhs = SSA_NAME_VALUE (cached_lhs);
763 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
767 cached_lhs = lookup_avail_expr (stmt, false);
771 edge taken_edge = find_taken_edge (e->dest, cached_lhs);
772 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
777 /* If we have a known destination for the conditional, then
778 we can perform this optimization, which saves at least one
779 conditional jump each time it applies since we get to
780 bypass the conditional at our original destination. */
783 struct edge_info *edge_info;
785 update_bb_profile_for_threading (e->dest, EDGE_FREQUENCY (e),
786 e->count, taken_edge);
790 edge_info = allocate_edge_info (e);
791 edge_info->redirection_target = taken_edge;
792 bb_ann (e->dest)->incoming_edge_threaded = true;
799 /* Initialize local stacks for this optimizer and record equivalences
800 upon entry to BB. Equivalences can come from the edge traversed to
801 reach BB or they may come from PHI nodes at the start of BB. */
804 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
807 if (dump_file && (dump_flags & TDF_DETAILS))
808 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
810 /* Push a marker on the stacks of local information so that we know how
811 far to unwind when we finalize this block. */
812 VEC_safe_push (tree_on_heap, avail_exprs_stack, NULL_TREE);
813 VEC_safe_push (tree_on_heap, block_defs_stack, NULL_TREE);
814 VEC_safe_push (tree_on_heap, const_and_copies_stack, NULL_TREE);
815 VEC_safe_push (tree_on_heap, nonzero_vars_stack, NULL_TREE);
816 VEC_safe_push (tree_on_heap, vrp_variables_stack, NULL_TREE);
818 record_equivalences_from_incoming_edge (bb);
820 /* PHI nodes can create equivalences too. */
821 record_equivalences_from_phis (bb);
824 /* Given an expression EXPR (a relational expression or a statement),
825 initialize the hash table element pointed by by ELEMENT. */
828 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
830 /* Hash table elements may be based on conditional expressions or statements.
832 For the former case, we have no annotation and we want to hash the
833 conditional expression. In the latter case we have an annotation and
834 we want to record the expression the statement evaluates. */
835 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
840 else if (TREE_CODE (expr) == COND_EXPR)
842 element->ann = stmt_ann (expr);
843 element->rhs = COND_EXPR_COND (expr);
845 else if (TREE_CODE (expr) == SWITCH_EXPR)
847 element->ann = stmt_ann (expr);
848 element->rhs = SWITCH_COND (expr);
850 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
852 element->ann = stmt_ann (expr);
853 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
857 element->ann = stmt_ann (expr);
858 element->rhs = TREE_OPERAND (expr, 1);
862 element->hash = avail_expr_hash (element);
865 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
866 LIMIT entries left in LOCALs. */
869 remove_local_expressions_from_table (void)
871 /* Remove all the expressions made available in this block. */
872 while (VEC_length (tree_on_heap, avail_exprs_stack) > 0)
874 struct expr_hash_elt element;
875 tree expr = VEC_pop (tree_on_heap, avail_exprs_stack);
877 if (expr == NULL_TREE)
880 initialize_hash_element (expr, NULL, &element);
881 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
885 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
886 state, stopping when there are LIMIT entries left in LOCALs. */
889 restore_nonzero_vars_to_original_value (void)
891 while (VEC_length (tree_on_heap, nonzero_vars_stack) > 0)
893 tree name = VEC_pop (tree_on_heap, nonzero_vars_stack);
898 bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
902 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
903 CONST_AND_COPIES to its original state, stopping when we hit a
907 restore_vars_to_original_value (void)
909 while (VEC_length (tree_on_heap, const_and_copies_stack) > 0)
911 tree prev_value, dest;
913 dest = VEC_pop (tree_on_heap, const_and_copies_stack);
918 prev_value = VEC_pop (tree_on_heap, const_and_copies_stack);
919 SSA_NAME_VALUE (dest) = prev_value;
923 /* Similar to restore_vars_to_original_value, except that it restores
924 CURRDEFS to its original value. */
926 restore_currdefs_to_original_value (void)
928 /* Restore CURRDEFS to its original state. */
929 while (VEC_length (tree_on_heap, block_defs_stack) > 0)
931 tree tmp = VEC_pop (tree_on_heap, block_defs_stack);
934 if (tmp == NULL_TREE)
937 /* If we recorded an SSA_NAME, then make the SSA_NAME the current
938 definition of its underlying variable. If we recorded anything
939 else, it must have been an _DECL node and its current reaching
940 definition must have been NULL. */
941 if (TREE_CODE (tmp) == SSA_NAME)
944 var = SSA_NAME_VAR (saved_def);
952 var_ann (var)->current_def = saved_def;
956 /* We have finished processing the dominator children of BB, perform
957 any finalization actions in preparation for leaving this node in
958 the dominator tree. */
961 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
965 /* If we are at a leaf node in the dominator tree, see if we can thread
966 the edge from BB through its successor.
968 Do this before we remove entries from our equivalence tables. */
969 if (EDGE_COUNT (bb->succs) == 1
970 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
971 && (get_immediate_dominator (CDI_DOMINATORS, EDGE_SUCC (bb, 0)->dest) != bb
972 || phi_nodes (EDGE_SUCC (bb, 0)->dest)))
975 thread_across_edge (walk_data, EDGE_SUCC (bb, 0));
977 else if ((last = last_stmt (bb))
978 && TREE_CODE (last) == COND_EXPR
979 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
980 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
981 && EDGE_COUNT (bb->succs) == 2
982 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
983 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
985 edge true_edge, false_edge;
987 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
989 /* If the THEN arm is the end of a dominator tree or has PHI nodes,
990 then try to thread through its edge. */
991 if (get_immediate_dominator (CDI_DOMINATORS, true_edge->dest) != bb
992 || phi_nodes (true_edge->dest))
994 struct edge_info *edge_info;
997 /* Push a marker onto the available expression stack so that we
998 unwind any expressions related to the TRUE arm before processing
999 the false arm below. */
1000 VEC_safe_push (tree_on_heap, avail_exprs_stack, NULL_TREE);
1001 VEC_safe_push (tree_on_heap, block_defs_stack, NULL_TREE);
1002 VEC_safe_push (tree_on_heap, const_and_copies_stack, NULL_TREE);
1004 edge_info = true_edge->aux;
1006 /* If we have info associated with this edge, record it into
1007 our equivalency tables. */
1010 tree *cond_equivalences = edge_info->cond_equivalences;
1011 tree lhs = edge_info->lhs;
1012 tree rhs = edge_info->rhs;
1014 /* If we have a simple NAME = VALUE equivalency record it.
1015 Until the jump threading selection code improves, only
1016 do this if both the name and value are SSA_NAMEs with
1017 the same underlying variable to avoid missing threading
1020 && TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME
1021 && TREE_CODE (edge_info->rhs) == SSA_NAME
1022 && SSA_NAME_VAR (lhs) == SSA_NAME_VAR (rhs))
1023 record_const_or_copy (lhs, rhs);
1025 /* If we have 0 = COND or 1 = COND equivalences, record them
1026 into our expression hash tables. */
1027 if (cond_equivalences)
1028 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1030 tree expr = cond_equivalences[i];
1031 tree value = cond_equivalences[i + 1];
1033 record_cond (expr, value);
1037 /* Now thread the edge. */
1038 thread_across_edge (walk_data, true_edge);
1040 /* And restore the various tables to their state before
1041 we threaded this edge. */
1042 remove_local_expressions_from_table ();
1043 restore_vars_to_original_value ();
1044 restore_currdefs_to_original_value ();
1047 /* Similarly for the ELSE arm. */
1048 if (get_immediate_dominator (CDI_DOMINATORS, false_edge->dest) != bb
1049 || phi_nodes (false_edge->dest))
1051 struct edge_info *edge_info;
1054 edge_info = false_edge->aux;
1056 /* If we have info associated with this edge, record it into
1057 our equivalency tables. */
1060 tree *cond_equivalences = edge_info->cond_equivalences;
1061 tree lhs = edge_info->lhs;
1062 tree rhs = edge_info->rhs;
1064 /* If we have a simple NAME = VALUE equivalency record it.
1065 Until the jump threading selection code improves, only
1066 do this if both the name and value are SSA_NAMEs with
1067 the same underlying variable to avoid missing threading
1070 && TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
1071 record_const_or_copy (lhs, rhs);
1073 /* If we have 0 = COND or 1 = COND equivalences, record them
1074 into our expression hash tables. */
1075 if (cond_equivalences)
1076 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1078 tree expr = cond_equivalences[i];
1079 tree value = cond_equivalences[i + 1];
1081 record_cond (expr, value);
1085 thread_across_edge (walk_data, false_edge);
1087 /* No need to remove local expressions from our tables
1088 or restore vars to their original value as that will
1089 be done immediately below. */
1093 remove_local_expressions_from_table ();
1094 restore_nonzero_vars_to_original_value ();
1095 restore_vars_to_original_value ();
1096 restore_currdefs_to_original_value ();
1098 /* Remove VRP records associated with this basic block. They are no
1101 To be efficient, we note which variables have had their values
1102 constrained in this block. So walk over each variable in the
1103 VRP_VARIABLEs array. */
1104 while (VEC_length (tree_on_heap, vrp_variables_stack) > 0)
1106 tree var = VEC_pop (tree_on_heap, vrp_variables_stack);
1107 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1110 /* Each variable has a stack of value range records. We want to
1111 invalidate those associated with our basic block. So we walk
1112 the array backwards popping off records associated with our
1113 block. Once we hit a record not associated with our block
1115 varray_type var_vrp_records;
1120 vrp_hash_elt.var = var;
1121 vrp_hash_elt.records = NULL;
1123 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1125 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1126 var_vrp_records = vrp_hash_elt_p->records;
1128 while (VARRAY_ACTIVE_SIZE (var_vrp_records) > 0)
1130 struct vrp_element *element
1131 = (struct vrp_element *)VARRAY_TOP_GENERIC_PTR (var_vrp_records);
1133 if (element->bb != bb)
1136 VARRAY_POP (var_vrp_records);
1140 /* If we queued any statements to rescan in this block, then
1141 go ahead and rescan them now. */
1142 while (VEC_length (tree_on_heap, stmts_to_rescan) > 0)
1144 tree stmt = VEC_last (tree_on_heap, stmts_to_rescan);
1145 basic_block stmt_bb = bb_for_stmt (stmt);
1150 VEC_pop (tree_on_heap, stmts_to_rescan);
1151 mark_new_vars_to_rename (stmt, vars_to_rename);
1155 /* PHI nodes can create equivalences too.
1157 Ignoring any alternatives which are the same as the result, if
1158 all the alternatives are equal, then the PHI node creates an
1161 Additionally, if all the PHI alternatives are known to have a nonzero
1162 value, then the result of this PHI is known to have a nonzero value,
1163 even if we do not know its exact value. */
1166 record_equivalences_from_phis (basic_block bb)
1170 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1172 tree lhs = PHI_RESULT (phi);
1176 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1178 tree t = PHI_ARG_DEF (phi, i);
1180 /* Ignore alternatives which are the same as our LHS. */
1181 if (operand_equal_for_phi_arg_p (lhs, t))
1184 /* If we have not processed an alternative yet, then set
1185 RHS to this alternative. */
1188 /* If we have processed an alternative (stored in RHS), then
1189 see if it is equal to this one. If it isn't, then stop
1191 else if (! operand_equal_for_phi_arg_p (rhs, t))
1195 /* If we had no interesting alternatives, then all the RHS alternatives
1196 must have been the same as LHS. */
1200 /* If we managed to iterate through each PHI alternative without
1201 breaking out of the loop, then we have a PHI which may create
1202 a useful equivalence. We do not need to record unwind data for
1203 this, since this is a true assignment and not an equivalence
1204 inferred from a comparison. All uses of this ssa name are dominated
1205 by this assignment, so unwinding just costs time and space. */
1206 if (i == PHI_NUM_ARGS (phi)
1207 && may_propagate_copy (lhs, rhs))
1208 SSA_NAME_VALUE (lhs) = rhs;
1210 /* Now see if we know anything about the nonzero property for the
1211 result of this PHI. */
1212 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1214 if (!PHI_ARG_NONZERO (phi, i))
1218 if (i == PHI_NUM_ARGS (phi))
1219 bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
1221 register_new_def (lhs, &block_defs_stack);
1225 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1226 return that edge. Otherwise return NULL. */
1228 single_incoming_edge_ignoring_loop_edges (basic_block bb)
1234 FOR_EACH_EDGE (e, ei, bb->preds)
1236 /* A loop back edge can be identified by the destination of
1237 the edge dominating the source of the edge. */
1238 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
1241 /* If we have already seen a non-loop edge, then we must have
1242 multiple incoming non-loop edges and thus we return NULL. */
1246 /* This is the first non-loop incoming edge we have found. Record
1254 /* Record any equivalences created by the incoming edge to BB. If BB
1255 has more than one incoming edge, then no equivalence is created. */
1258 record_equivalences_from_incoming_edge (basic_block bb)
1262 struct edge_info *edge_info;
1264 /* If our parent block ended with a control statment, then we may be
1265 able to record some equivalences based on which outgoing edge from
1266 the parent was followed. */
1267 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
1269 e = single_incoming_edge_ignoring_loop_edges (bb);
1271 /* If we had a single incoming edge from our parent block, then enter
1272 any data associated with the edge into our tables. */
1273 if (e && e->src == parent)
1281 tree lhs = edge_info->lhs;
1282 tree rhs = edge_info->rhs;
1283 tree *cond_equivalences = edge_info->cond_equivalences;
1286 record_equality (lhs, rhs);
1288 if (cond_equivalences)
1290 bool recorded_range = false;
1291 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1293 tree expr = cond_equivalences[i];
1294 tree value = cond_equivalences[i + 1];
1296 record_cond (expr, value);
1298 /* For the first true equivalence, record range
1299 information. We only do this for the first
1300 true equivalence as it should dominate any
1301 later true equivalences. */
1302 if (! recorded_range
1303 && COMPARISON_CLASS_P (expr)
1304 && value == boolean_true_node
1305 && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
1307 record_range (expr, bb);
1308 recorded_range = true;
1316 /* Dump SSA statistics on FILE. */
1319 dump_dominator_optimization_stats (FILE *file)
1323 fprintf (file, "Total number of statements: %6ld\n\n",
1324 opt_stats.num_stmts);
1325 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
1326 opt_stats.num_exprs_considered);
1328 n_exprs = opt_stats.num_exprs_considered;
1332 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1333 opt_stats.num_re, PERCENT (opt_stats.num_re,
1336 fprintf (file, "\nHash table statistics:\n");
1338 fprintf (file, " avail_exprs: ");
1339 htab_statistics (file, avail_exprs);
1343 /* Dump SSA statistics on stderr. */
1346 debug_dominator_optimization_stats (void)
1348 dump_dominator_optimization_stats (stderr);
1352 /* Dump statistics for the hash table HTAB. */
1355 htab_statistics (FILE *file, htab_t htab)
1357 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
1358 (long) htab_size (htab),
1359 (long) htab_elements (htab),
1360 htab_collisions (htab));
1363 /* Record the fact that VAR has a nonzero value, though we may not know
1364 its exact value. Note that if VAR is already known to have a nonzero
1365 value, then we do nothing. */
1368 record_var_is_nonzero (tree var)
1370 int indx = SSA_NAME_VERSION (var);
1372 if (bitmap_bit_p (nonzero_vars, indx))
1375 /* Mark it in the global table. */
1376 bitmap_set_bit (nonzero_vars, indx);
1378 /* Record this SSA_NAME so that we can reset the global table
1379 when we leave this block. */
1380 VEC_safe_push (tree_on_heap, nonzero_vars_stack, var);
1383 /* Enter a statement into the true/false expression hash table indicating
1384 that the condition COND has the value VALUE. */
1387 record_cond (tree cond, tree value)
1389 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
1392 initialize_hash_element (cond, value, element);
1394 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
1395 element->hash, true);
1398 *slot = (void *) element;
1399 VEC_safe_push (tree_on_heap, avail_exprs_stack, cond);
1405 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1406 the new conditional into *p, then store a boolean_true_node
1407 into the the *(p + 1). */
1410 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
1412 *p = build2 (new_code, boolean_type_node, op0, op1);
1414 *p = boolean_true_node;
1417 /* Record that COND is true and INVERTED is false into the edge information
1418 structure. Also record that any conditions dominated by COND are true
1421 For example, if a < b is true, then a <= b must also be true. */
1424 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
1428 if (!COMPARISON_CLASS_P (cond))
1431 op0 = TREE_OPERAND (cond, 0);
1432 op1 = TREE_OPERAND (cond, 1);
1434 switch (TREE_CODE (cond))
1438 edge_info->max_cond_equivalences = 12;
1439 edge_info->cond_equivalences = xmalloc (12 * sizeof (tree));
1440 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
1441 ? LE_EXPR : GE_EXPR),
1442 op0, op1, &edge_info->cond_equivalences[4]);
1443 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1444 &edge_info->cond_equivalences[6]);
1445 build_and_record_new_cond (NE_EXPR, op0, op1,
1446 &edge_info->cond_equivalences[8]);
1447 build_and_record_new_cond (LTGT_EXPR, op0, op1,
1448 &edge_info->cond_equivalences[10]);
1453 edge_info->max_cond_equivalences = 6;
1454 edge_info->cond_equivalences = xmalloc (6 * sizeof (tree));
1455 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1456 &edge_info->cond_equivalences[4]);
1460 edge_info->max_cond_equivalences = 10;
1461 edge_info->cond_equivalences = xmalloc (10 * sizeof (tree));
1462 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1463 &edge_info->cond_equivalences[4]);
1464 build_and_record_new_cond (LE_EXPR, op0, op1,
1465 &edge_info->cond_equivalences[6]);
1466 build_and_record_new_cond (GE_EXPR, op0, op1,
1467 &edge_info->cond_equivalences[8]);
1470 case UNORDERED_EXPR:
1471 edge_info->max_cond_equivalences = 16;
1472 edge_info->cond_equivalences = xmalloc (16 * sizeof (tree));
1473 build_and_record_new_cond (NE_EXPR, op0, op1,
1474 &edge_info->cond_equivalences[4]);
1475 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1476 &edge_info->cond_equivalences[6]);
1477 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1478 &edge_info->cond_equivalences[8]);
1479 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
1480 &edge_info->cond_equivalences[10]);
1481 build_and_record_new_cond (UNLT_EXPR, op0, op1,
1482 &edge_info->cond_equivalences[12]);
1483 build_and_record_new_cond (UNGT_EXPR, op0, op1,
1484 &edge_info->cond_equivalences[14]);
1489 edge_info->max_cond_equivalences = 8;
1490 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1491 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1492 ? UNLE_EXPR : UNGE_EXPR),
1493 op0, op1, &edge_info->cond_equivalences[4]);
1494 build_and_record_new_cond (NE_EXPR, op0, op1,
1495 &edge_info->cond_equivalences[6]);
1499 edge_info->max_cond_equivalences = 8;
1500 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1501 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1502 &edge_info->cond_equivalences[4]);
1503 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1504 &edge_info->cond_equivalences[6]);
1508 edge_info->max_cond_equivalences = 8;
1509 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1510 build_and_record_new_cond (NE_EXPR, op0, op1,
1511 &edge_info->cond_equivalences[4]);
1512 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1513 &edge_info->cond_equivalences[6]);
1517 edge_info->max_cond_equivalences = 4;
1518 edge_info->cond_equivalences = xmalloc (4 * sizeof (tree));
1522 /* Now store the original true and false conditions into the first
1524 edge_info->cond_equivalences[0] = cond;
1525 edge_info->cond_equivalences[1] = boolean_true_node;
1526 edge_info->cond_equivalences[2] = inverted;
1527 edge_info->cond_equivalences[3] = boolean_false_node;
1530 /* A helper function for record_const_or_copy and record_equality.
1531 Do the work of recording the value and undo info. */
1534 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1536 SSA_NAME_VALUE (x) = y;
1538 VEC_safe_push (tree_on_heap, const_and_copies_stack, prev_x);
1539 VEC_safe_push (tree_on_heap, const_and_copies_stack, x);
1543 /* Return the loop depth of the basic block of the defining statement of X.
1544 This number should not be treated as absolutely correct because the loop
1545 information may not be completely up-to-date when dom runs. However, it
1546 will be relatively correct, and as more passes are taught to keep loop info
1547 up to date, the result will become more and more accurate. */
1550 loop_depth_of_name (tree x)
1555 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1556 if (TREE_CODE (x) != SSA_NAME)
1559 /* Otherwise return the loop depth of the defining statement's bb.
1560 Note that there may not actually be a bb for this statement, if the
1561 ssa_name is live on entry. */
1562 defstmt = SSA_NAME_DEF_STMT (x);
1563 defbb = bb_for_stmt (defstmt);
1567 return defbb->loop_depth;
1571 /* Record that X is equal to Y in const_and_copies. Record undo
1572 information in the block-local vector. */
1575 record_const_or_copy (tree x, tree y)
1577 tree prev_x = SSA_NAME_VALUE (x);
1579 if (TREE_CODE (y) == SSA_NAME)
1581 tree tmp = SSA_NAME_VALUE (y);
1586 record_const_or_copy_1 (x, y, prev_x);
1589 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1590 This constrains the cases in which we may treat this as assignment. */
1593 record_equality (tree x, tree y)
1595 tree prev_x = NULL, prev_y = NULL;
1597 if (TREE_CODE (x) == SSA_NAME)
1598 prev_x = SSA_NAME_VALUE (x);
1599 if (TREE_CODE (y) == SSA_NAME)
1600 prev_y = SSA_NAME_VALUE (y);
1602 /* If one of the previous values is invariant, or invariant in more loops
1603 (by depth), then use that.
1604 Otherwise it doesn't matter which value we choose, just so
1605 long as we canonicalize on one value. */
1606 if (TREE_INVARIANT (y))
1608 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1609 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1610 else if (prev_x && TREE_INVARIANT (prev_x))
1611 x = y, y = prev_x, prev_x = prev_y;
1612 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1615 /* After the swapping, we must have one SSA_NAME. */
1616 if (TREE_CODE (x) != SSA_NAME)
1619 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1620 variable compared against zero. If we're honoring signed zeros,
1621 then we cannot record this value unless we know that the value is
1623 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1624 && (TREE_CODE (y) != REAL_CST
1625 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1628 record_const_or_copy_1 (x, y, prev_x);
1631 /* Return true, if it is ok to do folding of an associative expression.
1632 EXP is the tree for the associative expression. */
1635 unsafe_associative_fp_binop (tree exp)
1637 enum tree_code code = TREE_CODE (exp);
1638 return !(!flag_unsafe_math_optimizations
1639 && (code == MULT_EXPR || code == PLUS_EXPR
1640 || code == MINUS_EXPR)
1641 && FLOAT_TYPE_P (TREE_TYPE (exp)));
1644 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1645 hash tables. Try to simplify the RHS using whatever equivalences
1646 we may have recorded.
1648 If we are able to simplify the RHS, then lookup the simplified form in
1649 the hash table and return the result. Otherwise return NULL. */
1652 simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *walk_data,
1653 tree stmt, int insert)
1655 tree rhs = TREE_OPERAND (stmt, 1);
1656 enum tree_code rhs_code = TREE_CODE (rhs);
1659 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1660 In which case we can change this statement to be lhs = y.
1661 Which can then be copy propagated.
1663 Similarly for negation. */
1664 if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR)
1665 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME)
1667 /* Get the definition statement for our RHS. */
1668 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1670 /* See if the RHS_DEF_STMT has the same form as our statement. */
1671 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR
1672 && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code)
1674 tree rhs_def_operand;
1676 rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0);
1678 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1679 if (TREE_CODE (rhs_def_operand) == SSA_NAME
1680 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand))
1681 result = update_rhs_and_lookup_avail_expr (stmt,
1687 /* If we have z = (x OP C1), see if we earlier had x = y OP C2.
1688 If OP is associative, create and fold (y OP C2) OP C1 which
1689 should result in (y OP C3), use that as the RHS for the
1690 assignment. Add minus to this, as we handle it specially below. */
1691 if ((associative_tree_code (rhs_code) || rhs_code == MINUS_EXPR)
1692 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME
1693 && is_gimple_min_invariant (TREE_OPERAND (rhs, 1)))
1695 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1697 /* See if the RHS_DEF_STMT has the same form as our statement. */
1698 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR)
1700 tree rhs_def_rhs = TREE_OPERAND (rhs_def_stmt, 1);
1701 enum tree_code rhs_def_code = TREE_CODE (rhs_def_rhs);
1703 if ((rhs_code == rhs_def_code && unsafe_associative_fp_binop (rhs))
1704 || (rhs_code == PLUS_EXPR && rhs_def_code == MINUS_EXPR)
1705 || (rhs_code == MINUS_EXPR && rhs_def_code == PLUS_EXPR))
1707 tree def_stmt_op0 = TREE_OPERAND (rhs_def_rhs, 0);
1708 tree def_stmt_op1 = TREE_OPERAND (rhs_def_rhs, 1);
1710 if (TREE_CODE (def_stmt_op0) == SSA_NAME
1711 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0)
1712 && is_gimple_min_invariant (def_stmt_op1))
1714 tree outer_const = TREE_OPERAND (rhs, 1);
1715 tree type = TREE_TYPE (TREE_OPERAND (stmt, 0));
1718 /* If we care about correct floating point results, then
1719 don't fold x + c1 - c2. Note that we need to take both
1720 the codes and the signs to figure this out. */
1721 if (FLOAT_TYPE_P (type)
1722 && !flag_unsafe_math_optimizations
1723 && (rhs_def_code == PLUS_EXPR
1724 || rhs_def_code == MINUS_EXPR))
1728 neg ^= (rhs_code == MINUS_EXPR);
1729 neg ^= (rhs_def_code == MINUS_EXPR);
1730 neg ^= real_isneg (TREE_REAL_CST_PTR (outer_const));
1731 neg ^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1));
1734 goto dont_fold_assoc;
1737 /* Ho hum. So fold will only operate on the outermost
1738 thingy that we give it, so we have to build the new
1739 expression in two pieces. This requires that we handle
1740 combinations of plus and minus. */
1741 if (rhs_def_code != rhs_code)
1743 if (rhs_def_code == MINUS_EXPR)
1744 t = build (MINUS_EXPR, type, outer_const, def_stmt_op1);
1746 t = build (MINUS_EXPR, type, def_stmt_op1, outer_const);
1747 rhs_code = PLUS_EXPR;
1749 else if (rhs_def_code == MINUS_EXPR)
1750 t = build (PLUS_EXPR, type, def_stmt_op1, outer_const);
1752 t = build (rhs_def_code, type, def_stmt_op1, outer_const);
1754 t = build (rhs_code, type, def_stmt_op0, t);
1757 /* If the result is a suitable looking gimple expression,
1758 then use it instead of the original for STMT. */
1759 if (TREE_CODE (t) == SSA_NAME
1760 || (UNARY_CLASS_P (t)
1761 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME)
1762 || ((BINARY_CLASS_P (t) || COMPARISON_CLASS_P (t))
1763 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
1764 && is_gimple_val (TREE_OPERAND (t, 1))))
1765 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1772 /* Transform TRUNC_DIV_EXPR and TRUNC_MOD_EXPR into RSHIFT_EXPR
1773 and BIT_AND_EXPR respectively if the first operand is greater
1774 than zero and the second operand is an exact power of two. */
1775 if ((rhs_code == TRUNC_DIV_EXPR || rhs_code == TRUNC_MOD_EXPR)
1776 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0)))
1777 && integer_pow2p (TREE_OPERAND (rhs, 1)))
1780 tree op = TREE_OPERAND (rhs, 0);
1782 if (TYPE_UNSIGNED (TREE_TYPE (op)))
1784 val = integer_one_node;
1788 tree dummy_cond = walk_data->global_data;
1792 dummy_cond = build (GT_EXPR, boolean_type_node,
1793 op, integer_zero_node);
1794 dummy_cond = build (COND_EXPR, void_type_node,
1795 dummy_cond, NULL, NULL);
1796 walk_data->global_data = dummy_cond;
1800 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GT_EXPR);
1801 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1802 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1803 = integer_zero_node;
1805 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1808 if (val && integer_onep (val))
1811 tree op0 = TREE_OPERAND (rhs, 0);
1812 tree op1 = TREE_OPERAND (rhs, 1);
1814 if (rhs_code == TRUNC_DIV_EXPR)
1815 t = build (RSHIFT_EXPR, TREE_TYPE (op0), op0,
1816 build_int_cst (NULL_TREE, tree_log2 (op1)));
1818 t = build (BIT_AND_EXPR, TREE_TYPE (op0), op0,
1819 local_fold (build (MINUS_EXPR, TREE_TYPE (op1),
1820 op1, integer_one_node)));
1822 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1826 /* Transform ABS (X) into X or -X as appropriate. */
1827 if (rhs_code == ABS_EXPR
1828 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0))))
1831 tree op = TREE_OPERAND (rhs, 0);
1832 tree type = TREE_TYPE (op);
1834 if (TYPE_UNSIGNED (type))
1836 val = integer_zero_node;
1840 tree dummy_cond = walk_data->global_data;
1844 dummy_cond = build (LE_EXPR, boolean_type_node,
1845 op, integer_zero_node);
1846 dummy_cond = build (COND_EXPR, void_type_node,
1847 dummy_cond, NULL, NULL);
1848 walk_data->global_data = dummy_cond;
1852 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), LE_EXPR);
1853 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1854 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1855 = build_int_cst (type, 0);
1857 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1861 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GE_EXPR);
1862 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1863 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1864 = build_int_cst (type, 0);
1866 val = simplify_cond_and_lookup_avail_expr (dummy_cond,
1871 if (integer_zerop (val))
1872 val = integer_one_node;
1873 else if (integer_onep (val))
1874 val = integer_zero_node;
1880 && (integer_onep (val) || integer_zerop (val)))
1884 if (integer_onep (val))
1885 t = build1 (NEGATE_EXPR, TREE_TYPE (op), op);
1889 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1893 /* Optimize *"foo" into 'f'. This is done here rather than
1894 in fold to avoid problems with stuff like &*"foo". */
1895 if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF)
1897 tree t = fold_read_from_constant_string (rhs);
1900 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1906 /* COND is a condition of the form:
1908 x == const or x != const
1910 Look back to x's defining statement and see if x is defined as
1914 If const is unchanged if we convert it to type, then we can build
1915 the equivalent expression:
1918 y == const or y != const
1920 Which may allow further optimizations.
1922 Return the equivalent comparison or NULL if no such equivalent comparison
1926 find_equivalent_equality_comparison (tree cond)
1928 tree op0 = TREE_OPERAND (cond, 0);
1929 tree op1 = TREE_OPERAND (cond, 1);
1930 tree def_stmt = SSA_NAME_DEF_STMT (op0);
1932 /* OP0 might have been a parameter, so first make sure it
1933 was defined by a MODIFY_EXPR. */
1934 if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
1936 tree def_rhs = TREE_OPERAND (def_stmt, 1);
1938 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1939 if ((TREE_CODE (def_rhs) == NOP_EXPR
1940 || TREE_CODE (def_rhs) == CONVERT_EXPR)
1941 && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
1943 tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
1944 tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
1947 if (TYPE_PRECISION (def_rhs_inner_type)
1948 > TYPE_PRECISION (TREE_TYPE (def_rhs)))
1951 /* What we want to prove is that if we convert OP1 to
1952 the type of the object inside the NOP_EXPR that the
1953 result is still equivalent to SRC.
1955 If that is true, the build and return new equivalent
1956 condition which uses the source of the typecast and the
1957 new constant (which has only changed its type). */
1958 new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
1959 new = local_fold (new);
1960 if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
1961 return build (TREE_CODE (cond), TREE_TYPE (cond),
1962 def_rhs_inner, new);
1968 /* STMT is a COND_EXPR for which we could not trivially determine its
1969 result. This routine attempts to find equivalent forms of the
1970 condition which we may be able to optimize better. It also
1971 uses simple value range propagation to optimize conditionals. */
1974 simplify_cond_and_lookup_avail_expr (tree stmt,
1978 tree cond = COND_EXPR_COND (stmt);
1980 if (COMPARISON_CLASS_P (cond))
1982 tree op0 = TREE_OPERAND (cond, 0);
1983 tree op1 = TREE_OPERAND (cond, 1);
1985 if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
1988 tree low, high, cond_low, cond_high;
1989 int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
1990 varray_type vrp_records;
1991 struct vrp_element *element;
1992 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1995 /* First see if we have test of an SSA_NAME against a constant
1996 where the SSA_NAME is defined by an earlier typecast which
1997 is irrelevant when performing tests against the given
1999 if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2001 tree new_cond = find_equivalent_equality_comparison (cond);
2005 /* Update the statement to use the new equivalent
2007 COND_EXPR_COND (stmt) = new_cond;
2009 /* If this is not a real stmt, ann will be NULL and we
2010 avoid processing the operands. */
2014 /* Lookup the condition and return its known value if it
2016 new_cond = lookup_avail_expr (stmt, insert);
2020 /* The operands have changed, so update op0 and op1. */
2021 op0 = TREE_OPERAND (cond, 0);
2022 op1 = TREE_OPERAND (cond, 1);
2026 /* Consult the value range records for this variable (if they exist)
2027 to see if we can eliminate or simplify this conditional.
2029 Note two tests are necessary to determine no records exist.
2030 First we have to see if the virtual array exists, if it
2031 exists, then we have to check its active size.
2033 Also note the vast majority of conditionals are not testing
2034 a variable which has had its range constrained by an earlier
2035 conditional. So this filter avoids a lot of unnecessary work. */
2036 vrp_hash_elt.var = op0;
2037 vrp_hash_elt.records = NULL;
2038 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
2042 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
2043 vrp_records = vrp_hash_elt_p->records;
2044 if (vrp_records == NULL)
2047 limit = VARRAY_ACTIVE_SIZE (vrp_records);
2049 /* If we have no value range records for this variable, or we are
2050 unable to extract a range for this condition, then there is
2053 || ! extract_range_from_cond (cond, &cond_high,
2054 &cond_low, &cond_inverted))
2057 /* We really want to avoid unnecessary computations of range
2058 info. So all ranges are computed lazily; this avoids a
2059 lot of unnecessary work. i.e., we record the conditional,
2060 but do not process how it constrains the variable's
2061 potential values until we know that processing the condition
2064 However, we do not want to have to walk a potentially long
2065 list of ranges, nor do we want to compute a variable's
2066 range more than once for a given path.
2068 Luckily, each time we encounter a conditional that can not
2069 be otherwise optimized we will end up here and we will
2070 compute the necessary range information for the variable
2071 used in this condition.
2073 Thus you can conclude that there will never be more than one
2074 conditional associated with a variable which has not been
2075 processed. So we never need to merge more than one new
2076 conditional into the current range.
2078 These properties also help us avoid unnecessary work. */
2080 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records, limit - 1);
2082 if (element->high && element->low)
2084 /* The last element has been processed, so there is no range
2085 merging to do, we can simply use the high/low values
2086 recorded in the last element. */
2088 high = element->high;
2092 tree tmp_high, tmp_low;
2095 /* The last element has not been processed. Process it now. */
2096 extract_range_from_cond (element->cond, &tmp_high,
2099 /* If this is the only element, then no merging is necessary,
2100 the high/low values from extract_range_from_cond are all
2109 /* Get the high/low value from the previous element. */
2110 struct vrp_element *prev
2111 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records,
2116 /* Merge in this element's range with the range from the
2119 The low value for the merged range is the maximum of
2120 the previous low value and the low value of this record.
2122 Similarly the high value for the merged range is the
2123 minimum of the previous high value and the high value of
2125 low = (tree_int_cst_compare (low, tmp_low) == 1
2127 high = (tree_int_cst_compare (high, tmp_high) == -1
2131 /* And record the computed range. */
2133 element->high = high;
2137 /* After we have constrained this variable's potential values,
2138 we try to determine the result of the given conditional.
2140 To simplify later tests, first determine if the current
2141 low value is the same low value as the conditional.
2142 Similarly for the current high value and the high value
2143 for the conditional. */
2144 lowequal = tree_int_cst_equal (low, cond_low);
2145 highequal = tree_int_cst_equal (high, cond_high);
2147 if (lowequal && highequal)
2148 return (cond_inverted ? boolean_false_node : boolean_true_node);
2150 /* To simplify the overlap/subset tests below we may want
2151 to swap the two ranges so that the larger of the two
2152 ranges occurs "first". */
2154 if (tree_int_cst_compare (low, cond_low) == 1
2156 && tree_int_cst_compare (cond_high, high) == 1))
2169 /* Now determine if there is no overlap in the ranges
2170 or if the second range is a subset of the first range. */
2171 no_overlap = tree_int_cst_lt (high, cond_low);
2172 subset = tree_int_cst_compare (cond_high, high) != 1;
2174 /* If there was no overlap in the ranges, then this conditional
2175 always has a false value (unless we had to invert this
2176 conditional, in which case it always has a true value). */
2178 return (cond_inverted ? boolean_true_node : boolean_false_node);
2180 /* If the current range is a subset of the condition's range,
2181 then this conditional always has a true value (unless we
2182 had to invert this conditional, in which case it always
2183 has a true value). */
2184 if (subset && swapped)
2185 return (cond_inverted ? boolean_false_node : boolean_true_node);
2187 /* We were unable to determine the result of the conditional.
2188 However, we may be able to simplify the conditional. First
2189 merge the ranges in the same manner as range merging above. */
2190 low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
2191 high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
2193 /* If the range has converged to a single point, then turn this
2194 into an equality comparison. */
2195 if (TREE_CODE (cond) != EQ_EXPR
2196 && TREE_CODE (cond) != NE_EXPR
2197 && tree_int_cst_equal (low, high))
2199 TREE_SET_CODE (cond, EQ_EXPR);
2200 TREE_OPERAND (cond, 1) = high;
2207 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2208 result. This routine attempts to find equivalent forms of the
2209 condition which we may be able to optimize better. */
2212 simplify_switch_and_lookup_avail_expr (tree stmt, int insert)
2214 tree cond = SWITCH_COND (stmt);
2217 /* The optimization that we really care about is removing unnecessary
2218 casts. That will let us do much better in propagating the inferred
2219 constant at the switch target. */
2220 if (TREE_CODE (cond) == SSA_NAME)
2222 def = SSA_NAME_DEF_STMT (cond);
2223 if (TREE_CODE (def) == MODIFY_EXPR)
2225 def = TREE_OPERAND (def, 1);
2226 if (TREE_CODE (def) == NOP_EXPR)
2231 def = TREE_OPERAND (def, 0);
2233 #ifdef ENABLE_CHECKING
2234 /* ??? Why was Jeff testing this? We are gimple... */
2235 gcc_assert (is_gimple_val (def));
2238 to = TREE_TYPE (cond);
2239 ti = TREE_TYPE (def);
2241 /* If we have an extension that preserves value, then we
2242 can copy the source value into the switch. */
2244 need_precision = TYPE_PRECISION (ti);
2246 if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti))
2248 else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti))
2249 need_precision += 1;
2250 if (TYPE_PRECISION (to) < need_precision)
2255 SWITCH_COND (stmt) = def;
2258 return lookup_avail_expr (stmt, insert);
2268 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2269 known value for that SSA_NAME (or NULL if no value is known).
2271 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2272 even if we don't know their precise value.
2274 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2275 nodes of the successors of BB. */
2278 cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
2283 /* This can get rather expensive if the implementation is naive in
2284 how it finds the phi alternative associated with a particular edge. */
2285 FOR_EACH_EDGE (e, ei, bb->succs)
2290 /* If this is an abnormal edge, then we do not want to copy propagate
2291 into the PHI alternative associated with this edge. */
2292 if (e->flags & EDGE_ABNORMAL)
2295 phi = phi_nodes (e->dest);
2300 for ( ; phi; phi = PHI_CHAIN (phi))
2303 use_operand_p orig_p;
2306 /* The alternative may be associated with a constant, so verify
2307 it is an SSA_NAME before doing anything with it. */
2308 orig_p = PHI_ARG_DEF_PTR (phi, indx);
2309 orig = USE_FROM_PTR (orig_p);
2310 if (TREE_CODE (orig) != SSA_NAME)
2313 /* If the alternative is known to have a nonzero value, record
2314 that fact in the PHI node itself for future use. */
2315 if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
2316 PHI_ARG_NONZERO (phi, indx) = true;
2318 /* If we have *ORIG_P in our constant/copy table, then replace
2319 ORIG_P with its value in our constant/copy table. */
2320 new = SSA_NAME_VALUE (orig);
2322 && (TREE_CODE (new) == SSA_NAME
2323 || is_gimple_min_invariant (new))
2324 && may_propagate_copy (orig, new))
2326 propagate_value (orig_p, new);
2332 /* We have finished optimizing BB, record any information implied by
2333 taking a specific outgoing edge from BB. */
2336 record_edge_info (basic_block bb)
2338 block_stmt_iterator bsi = bsi_last (bb);
2339 struct edge_info *edge_info;
2341 if (! bsi_end_p (bsi))
2343 tree stmt = bsi_stmt (bsi);
2345 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
2347 tree cond = SWITCH_COND (stmt);
2349 if (TREE_CODE (cond) == SSA_NAME)
2351 tree labels = SWITCH_LABELS (stmt);
2352 int i, n_labels = TREE_VEC_LENGTH (labels);
2353 tree *info = xcalloc (n_basic_blocks, sizeof (tree));
2357 for (i = 0; i < n_labels; i++)
2359 tree label = TREE_VEC_ELT (labels, i);
2360 basic_block target_bb = label_to_block (CASE_LABEL (label));
2362 if (CASE_HIGH (label)
2363 || !CASE_LOW (label)
2364 || info[target_bb->index])
2365 info[target_bb->index] = error_mark_node;
2367 info[target_bb->index] = label;
2370 FOR_EACH_EDGE (e, ei, bb->succs)
2372 basic_block target_bb = e->dest;
2373 tree node = info[target_bb->index];
2375 if (node != NULL && node != error_mark_node)
2377 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
2378 edge_info = allocate_edge_info (e);
2379 edge_info->lhs = cond;
2387 /* A COND_EXPR may create equivalences too. */
2388 if (stmt && TREE_CODE (stmt) == COND_EXPR)
2390 tree cond = COND_EXPR_COND (stmt);
2394 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
2396 /* If the conditional is a single variable 'X', record 'X = 1'
2397 for the true edge and 'X = 0' on the false edge. */
2398 if (SSA_VAR_P (cond))
2400 struct edge_info *edge_info;
2402 edge_info = allocate_edge_info (true_edge);
2403 edge_info->lhs = cond;
2404 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
2406 edge_info = allocate_edge_info (false_edge);
2407 edge_info->lhs = cond;
2408 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
2410 /* Equality tests may create one or two equivalences. */
2411 else if (COMPARISON_CLASS_P (cond))
2413 tree op0 = TREE_OPERAND (cond, 0);
2414 tree op1 = TREE_OPERAND (cond, 1);
2416 /* Special case comparing booleans against a constant as we
2417 know the value of OP0 on both arms of the branch. i.e., we
2418 can record an equivalence for OP0 rather than COND. */
2419 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2420 && TREE_CODE (op0) == SSA_NAME
2421 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
2422 && is_gimple_min_invariant (op1))
2424 if (TREE_CODE (cond) == EQ_EXPR)
2426 edge_info = allocate_edge_info (true_edge);
2427 edge_info->lhs = op0;
2428 edge_info->rhs = (integer_zerop (op1)
2429 ? boolean_false_node
2430 : boolean_true_node);
2432 edge_info = allocate_edge_info (false_edge);
2433 edge_info->lhs = op0;
2434 edge_info->rhs = (integer_zerop (op1)
2436 : boolean_false_node);
2440 edge_info = allocate_edge_info (true_edge);
2441 edge_info->lhs = op0;
2442 edge_info->rhs = (integer_zerop (op1)
2444 : boolean_false_node);
2446 edge_info = allocate_edge_info (false_edge);
2447 edge_info->lhs = op0;
2448 edge_info->rhs = (integer_zerop (op1)
2449 ? boolean_false_node
2450 : boolean_true_node);
2454 if (is_gimple_min_invariant (op0)
2455 && (TREE_CODE (op1) == SSA_NAME
2456 || is_gimple_min_invariant (op1)))
2458 tree inverted = invert_truthvalue (cond);
2459 struct edge_info *edge_info;
2461 edge_info = allocate_edge_info (true_edge);
2462 record_conditions (edge_info, cond, inverted);
2464 if (TREE_CODE (cond) == EQ_EXPR)
2466 edge_info->lhs = op1;
2467 edge_info->rhs = op0;
2470 edge_info = allocate_edge_info (false_edge);
2471 record_conditions (edge_info, inverted, cond);
2473 if (TREE_CODE (cond) == NE_EXPR)
2475 edge_info->lhs = op1;
2476 edge_info->rhs = op0;
2480 if (TREE_CODE (op0) == SSA_NAME
2481 && (is_gimple_min_invariant (op1)
2482 || TREE_CODE (op1) == SSA_NAME))
2484 tree inverted = invert_truthvalue (cond);
2485 struct edge_info *edge_info;
2487 edge_info = allocate_edge_info (true_edge);
2488 record_conditions (edge_info, cond, inverted);
2490 if (TREE_CODE (cond) == EQ_EXPR)
2492 edge_info->lhs = op0;
2493 edge_info->rhs = op1;
2496 edge_info = allocate_edge_info (false_edge);
2497 record_conditions (edge_info, inverted, cond);
2499 if (TREE_CODE (cond) == NE_EXPR)
2501 edge_info->lhs = op0;
2502 edge_info->rhs = op1;
2507 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2512 /* Propagate information from BB to its outgoing edges.
2514 This can include equivalency information implied by control statements
2515 at the end of BB and const/copy propagation into PHIs in BB's
2516 successor blocks. */
2519 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2523 record_edge_info (bb);
2524 cprop_into_successor_phis (bb, nonzero_vars);
2527 /* Search for redundant computations in STMT. If any are found, then
2528 replace them with the variable holding the result of the computation.
2530 If safe, record this expression into the available expression hash
2534 eliminate_redundant_computations (struct dom_walk_data *walk_data,
2535 tree stmt, stmt_ann_t ann)
2537 v_may_def_optype v_may_defs = V_MAY_DEF_OPS (ann);
2538 tree *expr_p, def = NULL_TREE;
2541 bool retval = false;
2543 if (TREE_CODE (stmt) == MODIFY_EXPR)
2544 def = TREE_OPERAND (stmt, 0);
2546 /* Certain expressions on the RHS can be optimized away, but can not
2547 themselves be entered into the hash tables. */
2548 if (ann->makes_aliased_stores
2550 || TREE_CODE (def) != SSA_NAME
2551 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
2552 || NUM_V_MAY_DEFS (v_may_defs) != 0)
2555 /* Check if the expression has been computed before. */
2556 cached_lhs = lookup_avail_expr (stmt, insert);
2558 /* If this is an assignment and the RHS was not in the hash table,
2559 then try to simplify the RHS and lookup the new RHS in the
2561 if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR)
2562 cached_lhs = simplify_rhs_and_lookup_avail_expr (walk_data, stmt, insert);
2563 /* Similarly if this is a COND_EXPR and we did not find its
2564 expression in the hash table, simplify the condition and
2566 else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
2567 cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
2568 /* Similarly for a SWITCH_EXPR. */
2569 else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR)
2570 cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert);
2572 opt_stats.num_exprs_considered++;
2574 /* Get a pointer to the expression we are trying to optimize. */
2575 if (TREE_CODE (stmt) == COND_EXPR)
2576 expr_p = &COND_EXPR_COND (stmt);
2577 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2578 expr_p = &SWITCH_COND (stmt);
2579 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
2580 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
2582 expr_p = &TREE_OPERAND (stmt, 1);
2584 /* It is safe to ignore types here since we have already done
2585 type checking in the hashing and equality routines. In fact
2586 type checking here merely gets in the way of constant
2587 propagation. Also, make sure that it is safe to propagate
2588 CACHED_LHS into *EXPR_P. */
2590 && (TREE_CODE (cached_lhs) != SSA_NAME
2591 || may_propagate_copy (*expr_p, cached_lhs)))
2593 if (dump_file && (dump_flags & TDF_DETAILS))
2595 fprintf (dump_file, " Replaced redundant expr '");
2596 print_generic_expr (dump_file, *expr_p, dump_flags);
2597 fprintf (dump_file, "' with '");
2598 print_generic_expr (dump_file, cached_lhs, dump_flags);
2599 fprintf (dump_file, "'\n");
2604 #if defined ENABLE_CHECKING
2605 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
2606 || is_gimple_min_invariant (cached_lhs));
2609 if (TREE_CODE (cached_lhs) == ADDR_EXPR
2610 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
2611 && is_gimple_min_invariant (cached_lhs)))
2614 propagate_tree_value (expr_p, cached_lhs);
2620 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2621 the available expressions table or the const_and_copies table.
2622 Detect and record those equivalences. */
2625 record_equivalences_from_stmt (tree stmt,
2629 tree lhs = TREE_OPERAND (stmt, 0);
2630 enum tree_code lhs_code = TREE_CODE (lhs);
2633 if (lhs_code == SSA_NAME)
2635 tree rhs = TREE_OPERAND (stmt, 1);
2637 /* Strip away any useless type conversions. */
2638 STRIP_USELESS_TYPE_CONVERSION (rhs);
2640 /* If the RHS of the assignment is a constant or another variable that
2641 may be propagated, register it in the CONST_AND_COPIES table. We
2642 do not need to record unwind data for this, since this is a true
2643 assignment and not an equivalence inferred from a comparison. All
2644 uses of this ssa name are dominated by this assignment, so unwinding
2645 just costs time and space. */
2647 && (TREE_CODE (rhs) == SSA_NAME
2648 || is_gimple_min_invariant (rhs)))
2649 SSA_NAME_VALUE (lhs) = rhs;
2651 /* alloca never returns zero and the address of a non-weak symbol
2652 is never zero. NOP_EXPRs and CONVERT_EXPRs can be completely
2653 stripped as they do not affect this equivalence. */
2654 while (TREE_CODE (rhs) == NOP_EXPR
2655 || TREE_CODE (rhs) == CONVERT_EXPR)
2656 rhs = TREE_OPERAND (rhs, 0);
2658 if (alloca_call_p (rhs)
2659 || (TREE_CODE (rhs) == ADDR_EXPR
2660 && DECL_P (TREE_OPERAND (rhs, 0))
2661 && ! DECL_WEAK (TREE_OPERAND (rhs, 0))))
2662 record_var_is_nonzero (lhs);
2664 /* IOR of any value with a nonzero value will result in a nonzero
2665 value. Even if we do not know the exact result recording that
2666 the result is nonzero is worth the effort. */
2667 if (TREE_CODE (rhs) == BIT_IOR_EXPR
2668 && integer_nonzerop (TREE_OPERAND (rhs, 1)))
2669 record_var_is_nonzero (lhs);
2672 /* Look at both sides for pointer dereferences. If we find one, then
2673 the pointer must be nonnull and we can enter that equivalence into
2675 if (flag_delete_null_pointer_checks)
2676 for (i = 0; i < 2; i++)
2678 tree t = TREE_OPERAND (stmt, i);
2680 /* Strip away any COMPONENT_REFs. */
2681 while (TREE_CODE (t) == COMPONENT_REF)
2682 t = TREE_OPERAND (t, 0);
2684 /* Now see if this is a pointer dereference. */
2685 if (INDIRECT_REF_P (t))
2687 tree op = TREE_OPERAND (t, 0);
2689 /* If the pointer is a SSA variable, then enter new
2690 equivalences into the hash table. */
2691 while (TREE_CODE (op) == SSA_NAME)
2693 tree def = SSA_NAME_DEF_STMT (op);
2695 record_var_is_nonzero (op);
2697 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2698 which are known to have a nonzero value. */
2700 && TREE_CODE (def) == MODIFY_EXPR
2701 && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
2702 op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
2709 /* A memory store, even an aliased store, creates a useful
2710 equivalence. By exchanging the LHS and RHS, creating suitable
2711 vops and recording the result in the available expression table,
2712 we may be able to expose more redundant loads. */
2713 if (!ann->has_volatile_ops
2714 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
2715 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
2716 && !is_gimple_reg (lhs))
2718 tree rhs = TREE_OPERAND (stmt, 1);
2721 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2722 is a constant, we need to adjust the constant to fit into the
2723 type of the LHS. If the LHS is a bitfield and the RHS is not
2724 a constant, then we can not record any equivalences for this
2725 statement since we would need to represent the widening or
2726 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2727 and should not be necessary if GCC represented bitfields
2729 if (lhs_code == COMPONENT_REF
2730 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
2732 if (TREE_CONSTANT (rhs))
2733 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
2737 /* If the value overflowed, then we can not use this equivalence. */
2738 if (rhs && ! is_gimple_min_invariant (rhs))
2744 /* Build a new statement with the RHS and LHS exchanged. */
2745 new = build (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
2747 create_ssa_artficial_load_stmt (&(ann->operands), new);
2749 /* Finally enter the statement into the available expression
2751 lookup_avail_expr (new, true);
2756 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2757 CONST_AND_COPIES. */
2760 cprop_operand (tree stmt, use_operand_p op_p)
2762 bool may_have_exposed_new_symbols = false;
2764 tree op = USE_FROM_PTR (op_p);
2766 /* If the operand has a known constant value or it is known to be a
2767 copy of some other variable, use the value or copy stored in
2768 CONST_AND_COPIES. */
2769 val = SSA_NAME_VALUE (op);
2770 if (val && TREE_CODE (val) != VALUE_HANDLE)
2772 tree op_type, val_type;
2774 /* Do not change the base variable in the virtual operand
2775 tables. That would make it impossible to reconstruct
2776 the renamed virtual operand if we later modify this
2777 statement. Also only allow the new value to be an SSA_NAME
2778 for propagation into virtual operands. */
2779 if (!is_gimple_reg (op)
2780 && (get_virtual_var (val) != get_virtual_var (op)
2781 || TREE_CODE (val) != SSA_NAME))
2784 /* Do not replace hard register operands in asm statements. */
2785 if (TREE_CODE (stmt) == ASM_EXPR
2786 && !may_propagate_copy_into_asm (op))
2789 /* Get the toplevel type of each operand. */
2790 op_type = TREE_TYPE (op);
2791 val_type = TREE_TYPE (val);
2793 /* While both types are pointers, get the type of the object
2795 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
2797 op_type = TREE_TYPE (op_type);
2798 val_type = TREE_TYPE (val_type);
2801 /* Make sure underlying types match before propagating a constant by
2802 converting the constant to the proper type. Note that convert may
2803 return a non-gimple expression, in which case we ignore this
2804 propagation opportunity. */
2805 if (TREE_CODE (val) != SSA_NAME)
2807 if (!lang_hooks.types_compatible_p (op_type, val_type))
2809 val = fold_convert (TREE_TYPE (op), val);
2810 if (!is_gimple_min_invariant (val))
2815 /* Certain operands are not allowed to be copy propagated due
2816 to their interaction with exception handling and some GCC
2818 else if (!may_propagate_copy (op, val))
2822 if (dump_file && (dump_flags & TDF_DETAILS))
2824 fprintf (dump_file, " Replaced '");
2825 print_generic_expr (dump_file, op, dump_flags);
2826 fprintf (dump_file, "' with %s '",
2827 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
2828 print_generic_expr (dump_file, val, dump_flags);
2829 fprintf (dump_file, "'\n");
2832 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2833 that we may have exposed a new symbol for SSA renaming. */
2834 if (TREE_CODE (val) == ADDR_EXPR
2835 || (POINTER_TYPE_P (TREE_TYPE (op))
2836 && is_gimple_min_invariant (val)))
2837 may_have_exposed_new_symbols = true;
2839 propagate_value (op_p, val);
2841 /* And note that we modified this statement. This is now
2842 safe, even if we changed virtual operands since we will
2843 rescan the statement and rewrite its operands again. */
2846 return may_have_exposed_new_symbols;
2849 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2850 known value for that SSA_NAME (or NULL if no value is known).
2852 Propagate values from CONST_AND_COPIES into the uses, vuses and
2853 v_may_def_ops of STMT. */
2856 cprop_into_stmt (tree stmt)
2858 bool may_have_exposed_new_symbols = false;
2863 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
2865 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
2866 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
2869 if (may_have_exposed_new_symbols)
2871 rhs = get_rhs (stmt);
2872 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
2873 recompute_tree_invarant_for_addr_expr (rhs);
2876 return may_have_exposed_new_symbols;
2880 /* Optimize the statement pointed by iterator SI.
2882 We try to perform some simplistic global redundancy elimination and
2883 constant propagation:
2885 1- To detect global redundancy, we keep track of expressions that have
2886 been computed in this block and its dominators. If we find that the
2887 same expression is computed more than once, we eliminate repeated
2888 computations by using the target of the first one.
2890 2- Constant values and copy assignments. This is used to do very
2891 simplistic constant and copy propagation. When a constant or copy
2892 assignment is found, we map the value on the RHS of the assignment to
2893 the variable in the LHS in the CONST_AND_COPIES table. */
2896 optimize_stmt (struct dom_walk_data *walk_data, basic_block bb,
2897 block_stmt_iterator si)
2901 bool may_optimize_p;
2902 bool may_have_exposed_new_symbols = false;
2904 stmt = bsi_stmt (si);
2906 get_stmt_operands (stmt);
2907 ann = stmt_ann (stmt);
2908 opt_stats.num_stmts++;
2909 may_have_exposed_new_symbols = false;
2911 if (dump_file && (dump_flags & TDF_DETAILS))
2913 fprintf (dump_file, "Optimizing statement ");
2914 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2917 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2918 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
2920 /* If the statement has been modified with constant replacements,
2921 fold its RHS before checking for redundant computations. */
2924 /* Try to fold the statement making sure that STMT is kept
2926 if (fold_stmt (bsi_stmt_ptr (si)))
2928 stmt = bsi_stmt (si);
2929 ann = stmt_ann (stmt);
2931 if (dump_file && (dump_flags & TDF_DETAILS))
2933 fprintf (dump_file, " Folded to: ");
2934 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2938 /* Constant/copy propagation above may change the set of
2939 virtual operands associated with this statement. Folding
2940 may remove the need for some virtual operands.
2942 Indicate we will need to rescan and rewrite the statement. */
2943 may_have_exposed_new_symbols = true;
2946 /* Check for redundant computations. Do this optimization only
2947 for assignments that have no volatile ops and conditionals. */
2948 may_optimize_p = (!ann->has_volatile_ops
2949 && ((TREE_CODE (stmt) == RETURN_EXPR
2950 && TREE_OPERAND (stmt, 0)
2951 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
2952 && ! (TREE_SIDE_EFFECTS
2953 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
2954 || (TREE_CODE (stmt) == MODIFY_EXPR
2955 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
2956 || TREE_CODE (stmt) == COND_EXPR
2957 || TREE_CODE (stmt) == SWITCH_EXPR));
2960 may_have_exposed_new_symbols
2961 |= eliminate_redundant_computations (walk_data, stmt, ann);
2963 /* Record any additional equivalences created by this statement. */
2964 if (TREE_CODE (stmt) == MODIFY_EXPR)
2965 record_equivalences_from_stmt (stmt,
2969 register_definitions_for_stmt (stmt);
2971 /* If STMT is a COND_EXPR and it was modified, then we may know
2972 where it goes. If that is the case, then mark the CFG as altered.
2974 This will cause us to later call remove_unreachable_blocks and
2975 cleanup_tree_cfg when it is safe to do so. It is not safe to
2976 clean things up here since removal of edges and such can trigger
2977 the removal of PHI nodes, which in turn can release SSA_NAMEs to
2980 That's all fine and good, except that once SSA_NAMEs are released
2981 to the manager, we must not call create_ssa_name until all references
2982 to released SSA_NAMEs have been eliminated.
2984 All references to the deleted SSA_NAMEs can not be eliminated until
2985 we remove unreachable blocks.
2987 We can not remove unreachable blocks until after we have completed
2988 any queued jump threading.
2990 We can not complete any queued jump threads until we have taken
2991 appropriate variables out of SSA form. Taking variables out of
2992 SSA form can call create_ssa_name and thus we lose.
2994 Ultimately I suspect we're going to need to change the interface
2995 into the SSA_NAME manager. */
3001 if (TREE_CODE (stmt) == COND_EXPR)
3002 val = COND_EXPR_COND (stmt);
3003 else if (TREE_CODE (stmt) == SWITCH_EXPR)
3004 val = SWITCH_COND (stmt);
3006 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
3009 /* If we simplified a statement in such a way as to be shown that it
3010 cannot trap, update the eh information and the cfg to match. */
3011 if (maybe_clean_eh_stmt (stmt))
3013 bitmap_set_bit (need_eh_cleanup, bb->index);
3014 if (dump_file && (dump_flags & TDF_DETAILS))
3015 fprintf (dump_file, " Flagged to clear EH edges.\n");
3019 if (may_have_exposed_new_symbols)
3020 VEC_safe_push (tree_on_heap, stmts_to_rescan, bsi_stmt (si));
3023 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
3024 available expression hashtable, then return the LHS from the hash
3027 If INSERT is true, then we also update the available expression
3028 hash table to account for the changes made to STMT. */
3031 update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert)
3033 tree cached_lhs = NULL;
3035 /* Remove the old entry from the hash table. */
3038 struct expr_hash_elt element;
3040 initialize_hash_element (stmt, NULL, &element);
3041 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
3044 /* Now update the RHS of the assignment. */
3045 TREE_OPERAND (stmt, 1) = new_rhs;
3047 /* Now lookup the updated statement in the hash table. */
3048 cached_lhs = lookup_avail_expr (stmt, insert);
3050 /* We have now called lookup_avail_expr twice with two different
3051 versions of this same statement, once in optimize_stmt, once here.
3053 We know the call in optimize_stmt did not find an existing entry
3054 in the hash table, so a new entry was created. At the same time
3055 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
3057 If this call failed to find an existing entry on the hash table,
3058 then the new version of this statement was entered into the
3059 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3060 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3062 If this call succeeded, we still have one copy of this statement
3063 on the BLOCK_AVAIL_EXPRs vector.
3065 For both cases, we need to pop the most recent entry off the
3066 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3067 statement in the hash tables, that will leave precisely one
3068 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3069 we found a copy of this statement in the second hash table lookup
3070 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3072 VEC_pop (tree_on_heap, avail_exprs_stack);
3074 /* And make sure we record the fact that we modified this
3081 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3082 found, return its LHS. Otherwise insert STMT in the table and return
3085 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3086 is also added to the stack pointed by BLOCK_AVAIL_EXPRS_P, so that they
3087 can be removed when we finish processing this block and its children.
3089 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3090 contains no CALL_EXPR on its RHS and makes no volatile nor
3091 aliased references. */
3094 lookup_avail_expr (tree stmt, bool insert)
3099 struct expr_hash_elt *element = xcalloc (sizeof (struct expr_hash_elt), 1);
3101 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
3103 initialize_hash_element (stmt, lhs, element);
3105 /* Don't bother remembering constant assignments and copy operations.
3106 Constants and copy operations are handled by the constant/copy propagator
3107 in optimize_stmt. */
3108 if (TREE_CODE (element->rhs) == SSA_NAME
3109 || is_gimple_min_invariant (element->rhs))
3115 /* If this is an equality test against zero, see if we have recorded a
3116 nonzero value for the variable in question. */
3117 if ((TREE_CODE (element->rhs) == EQ_EXPR
3118 || TREE_CODE (element->rhs) == NE_EXPR)
3119 && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
3120 && integer_zerop (TREE_OPERAND (element->rhs, 1)))
3122 int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
3124 if (bitmap_bit_p (nonzero_vars, indx))
3126 tree t = element->rhs;
3129 if (TREE_CODE (t) == EQ_EXPR)
3130 return boolean_false_node;
3132 return boolean_true_node;
3136 /* Finally try to find the expression in the main expression hash table. */
3137 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
3138 (insert ? INSERT : NO_INSERT));
3147 *slot = (void *) element;
3148 VEC_safe_push (tree_on_heap, avail_exprs_stack,
3149 stmt ? stmt : element->rhs);
3153 /* Extract the LHS of the assignment so that it can be used as the current
3154 definition of another variable. */
3155 lhs = ((struct expr_hash_elt *)*slot)->lhs;
3157 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3158 use the value from the const_and_copies table. */
3159 if (TREE_CODE (lhs) == SSA_NAME)
3161 temp = SSA_NAME_VALUE (lhs);
3162 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
3170 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3171 range of values that result in the conditional having a true value.
3173 Return true if we are successful in extracting a range from COND and
3174 false if we are unsuccessful. */
3177 extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
3179 tree op1 = TREE_OPERAND (cond, 1);
3180 tree high, low, type;
3183 /* Experiments have shown that it's rarely, if ever useful to
3184 record ranges for enumerations. Presumably this is due to
3185 the fact that they're rarely used directly. They are typically
3186 cast into an integer type and used that way. */
3187 if (TREE_CODE (TREE_TYPE (op1)) != INTEGER_TYPE)
3190 type = TREE_TYPE (op1);
3192 switch (TREE_CODE (cond))
3206 high = TYPE_MAX_VALUE (type);
3211 low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
3212 high = TYPE_MAX_VALUE (type);
3218 low = TYPE_MIN_VALUE (type);
3223 high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
3224 low = TYPE_MIN_VALUE (type);
3234 *inverted_p = inverted;
3238 /* Record a range created by COND for basic block BB. */
3241 record_range (tree cond, basic_block bb)
3243 enum tree_code code = TREE_CODE (cond);
3245 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3246 They rarely allow for meaningful range optimizations and significantly
3247 complicate the implementation. */
3248 if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
3249 || code == GE_EXPR || code == EQ_EXPR)
3250 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
3252 struct vrp_hash_elt *vrp_hash_elt;
3253 struct vrp_element *element;
3254 varray_type *vrp_records_p;
3258 vrp_hash_elt = xmalloc (sizeof (struct vrp_hash_elt));
3259 vrp_hash_elt->var = TREE_OPERAND (cond, 0);
3260 vrp_hash_elt->records = NULL;
3261 slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
3264 *slot = (void *) vrp_hash_elt;
3266 free (vrp_hash_elt);
3268 vrp_hash_elt = (struct vrp_hash_elt *) *slot;
3269 vrp_records_p = &vrp_hash_elt->records;
3271 element = ggc_alloc (sizeof (struct vrp_element));
3272 element->low = NULL;
3273 element->high = NULL;
3274 element->cond = cond;
3277 if (*vrp_records_p == NULL)
3278 VARRAY_GENERIC_PTR_INIT (*vrp_records_p, 2, "vrp records");
3280 VARRAY_PUSH_GENERIC_PTR (*vrp_records_p, element);
3281 VEC_safe_push (tree_on_heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
3285 /* Hashing and equality functions for VRP_DATA.
3287 Since this hash table is addressed by SSA_NAMEs, we can hash on
3288 their version number and equality can be determined with a
3289 pointer comparison. */
3292 vrp_hash (const void *p)
3294 tree var = ((struct vrp_hash_elt *)p)->var;
3296 return SSA_NAME_VERSION (var);
3300 vrp_eq (const void *p1, const void *p2)
3302 tree var1 = ((struct vrp_hash_elt *)p1)->var;
3303 tree var2 = ((struct vrp_hash_elt *)p2)->var;
3305 return var1 == var2;
3308 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3309 MODIFY_EXPR statements. We compute a value number for expressions using
3310 the code of the expression and the SSA numbers of its operands. */
3313 avail_expr_hash (const void *p)
3315 stmt_ann_t ann = ((struct expr_hash_elt *)p)->ann;
3316 tree rhs = ((struct expr_hash_elt *)p)->rhs;
3321 /* iterative_hash_expr knows how to deal with any expression and
3322 deals with commutative operators as well, so just use it instead
3323 of duplicating such complexities here. */
3324 val = iterative_hash_expr (rhs, val);
3326 /* If the hash table entry is not associated with a statement, then we
3327 can just hash the expression and not worry about virtual operands
3332 /* Add the SSA version numbers of every vuse operand. This is important
3333 because compound variables like arrays are not renamed in the
3334 operands. Rather, the rename is done on the virtual variable
3335 representing all the elements of the array. */
3336 vuses = VUSE_OPS (ann);
3337 for (i = 0; i < NUM_VUSES (vuses); i++)
3338 val = iterative_hash_expr (VUSE_OP (vuses, i), val);
3344 real_avail_expr_hash (const void *p)
3346 return ((const struct expr_hash_elt *)p)->hash;
3350 avail_expr_eq (const void *p1, const void *p2)
3352 stmt_ann_t ann1 = ((struct expr_hash_elt *)p1)->ann;
3353 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
3354 stmt_ann_t ann2 = ((struct expr_hash_elt *)p2)->ann;
3355 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
3357 /* If they are the same physical expression, return true. */
3358 if (rhs1 == rhs2 && ann1 == ann2)
3361 /* If their codes are not equal, then quit now. */
3362 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
3365 /* In case of a collision, both RHS have to be identical and have the
3366 same VUSE operands. */
3367 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
3368 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
3369 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
3371 vuse_optype ops1 = NULL;
3372 vuse_optype ops2 = NULL;
3373 size_t num_ops1 = 0;
3374 size_t num_ops2 = 0;
3379 ops1 = VUSE_OPS (ann1);
3380 num_ops1 = NUM_VUSES (ops1);
3385 ops2 = VUSE_OPS (ann2);
3386 num_ops2 = NUM_VUSES (ops2);
3389 /* If the number of virtual uses is different, then we consider
3391 if (num_ops1 != num_ops2)
3394 for (i = 0; i < num_ops1; i++)
3395 if (VUSE_OP (ops1, i) != VUSE_OP (ops2, i))
3398 gcc_assert (((struct expr_hash_elt *)p1)->hash
3399 == ((struct expr_hash_elt *)p2)->hash);
3406 /* Given STMT and a pointer to the block local definitions BLOCK_DEFS_P,
3407 register register all objects set by this statement into BLOCK_DEFS_P
3411 register_definitions_for_stmt (tree stmt)
3416 FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_ALL_DEFS)
3419 /* FIXME: We shouldn't be registering new defs if the variable
3420 doesn't need to be renamed. */
3421 register_new_def (def, &block_defs_stack);