1 /* SSA Dominator optimizations for trees
2 Copyright (C) 2001, 2002, 2003, 2004, 2005 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"
37 #include "diagnostic.h"
39 #include "tree-dump.h"
40 #include "tree-flow.h"
43 #include "tree-pass.h"
44 #include "tree-ssa-propagate.h"
45 #include "langhooks.h"
47 /* This file implements optimizations on the dominator tree. */
50 /* Structure for recording edge equivalences as well as any pending
51 edge redirections during the dominator optimizer.
53 Computing and storing the edge equivalences instead of creating
54 them on-demand can save significant amounts of time, particularly
55 for pathological cases involving switch statements.
57 These structures live for a single iteration of the dominator
58 optimizer in the edge's AUX field. At the end of an iteration we
59 free each of these structures and update the AUX field to point
60 to any requested redirection target (the code for updating the
61 CFG and SSA graph for edge redirection expects redirection edge
62 targets to be in the AUX field for each edge. */
66 /* If this edge creates a simple equivalence, the LHS and RHS of
67 the equivalence will be stored here. */
71 /* Traversing an edge may also indicate one or more particular conditions
72 are true or false. The number of recorded conditions can vary, but
73 can be determined by the condition's code. So we have an array
74 and its maximum index rather than use a varray. */
75 tree *cond_equivalences;
76 unsigned int max_cond_equivalences;
78 /* If we can thread this edge this field records the new target. */
79 edge redirection_target;
83 /* Hash table with expressions made available during the renaming process.
84 When an assignment of the form X_i = EXPR is found, the statement is
85 stored in this table. If the same expression EXPR is later found on the
86 RHS of another statement, it is replaced with X_i (thus performing
87 global redundancy elimination). Similarly as we pass through conditionals
88 we record the conditional itself as having either a true or false value
90 static htab_t avail_exprs;
92 /* Stack of available expressions in AVAIL_EXPRs. Each block pushes any
93 expressions it enters into the hash table along with a marker entry
94 (null). When we finish processing the block, we pop off entries and
95 remove the expressions from the global hash table until we hit the
97 static VEC(tree,heap) *avail_exprs_stack;
99 /* Stack of statements we need to rescan during finalization for newly
102 Statement rescanning must occur after the current block's available
103 expressions are removed from AVAIL_EXPRS. Else we may change the
104 hash code for an expression and be unable to find/remove it from
106 static VEC(tree,heap) *stmts_to_rescan;
108 /* Structure for entries in the expression hash table.
110 This requires more memory for the hash table entries, but allows us
111 to avoid creating silly tree nodes and annotations for conditionals,
112 eliminates 2 global hash tables and two block local varrays.
114 It also allows us to reduce the number of hash table lookups we
115 have to perform in lookup_avail_expr and finally it allows us to
116 significantly reduce the number of calls into the hashing routine
121 /* The value (lhs) of this expression. */
124 /* The expression (rhs) we want to record. */
127 /* The stmt pointer if this element corresponds to a statement. */
130 /* The hash value for RHS/ann. */
134 /* Stack of dest,src pairs that need to be restored during finalization.
136 A NULL entry is used to mark the end of pairs which need to be
137 restored during finalization of this block. */
138 static VEC(tree,heap) *const_and_copies_stack;
140 /* Bitmap of SSA_NAMEs known to have a nonzero value, even if we do not
141 know their exact value. */
142 static bitmap nonzero_vars;
144 /* Bitmap of blocks that are scheduled to be threaded through. This
145 is used to communicate with thread_through_blocks. */
146 static bitmap threaded_blocks;
148 /* Stack of SSA_NAMEs which need their NONZERO_VARS property cleared
149 when the current block is finalized.
151 A NULL entry is used to mark the end of names needing their
152 entry in NONZERO_VARS cleared during finalization of this block. */
153 static VEC(tree,heap) *nonzero_vars_stack;
155 /* Track whether or not we have changed the control flow graph. */
156 static bool cfg_altered;
158 /* Bitmap of blocks that have had EH statements cleaned. We should
159 remove their dead edges eventually. */
160 static bitmap need_eh_cleanup;
162 /* Statistics for dominator optimizations. */
166 long num_exprs_considered;
172 static struct opt_stats_d opt_stats;
174 /* Value range propagation record. Each time we encounter a conditional
175 of the form SSA_NAME COND CONST we create a new vrp_element to record
176 how the condition affects the possible values SSA_NAME may have.
178 Each record contains the condition tested (COND), and the range of
179 values the variable may legitimately have if COND is true. Note the
180 range of values may be a smaller range than COND specifies if we have
181 recorded other ranges for this variable. Each record also contains the
182 block in which the range was recorded for invalidation purposes.
184 Note that the current known range is computed lazily. This allows us
185 to avoid the overhead of computing ranges which are never queried.
187 When we encounter a conditional, we look for records which constrain
188 the SSA_NAME used in the condition. In some cases those records allow
189 us to determine the condition's result at compile time. In other cases
190 they may allow us to simplify the condition.
192 We also use value ranges to do things like transform signed div/mod
193 operations into unsigned div/mod or to simplify ABS_EXPRs.
195 Simple experiments have shown these optimizations to not be all that
196 useful on switch statements (much to my surprise). So switch statement
197 optimizations are not performed.
199 Note carefully we do not propagate information through each statement
200 in the block. i.e., if we know variable X has a value defined of
201 [0, 25] and we encounter Y = X + 1, we do not track a value range
202 for Y (which would be [1, 26] if we cared). Similarly we do not
203 constrain values as we encounter narrowing typecasts, etc. */
207 /* The highest and lowest values the variable in COND may contain when
208 COND is true. Note this may not necessarily be the same values
209 tested by COND if the same variable was used in earlier conditionals.
211 Note this is computed lazily and thus can be NULL indicating that
212 the values have not been computed yet. */
216 /* The actual conditional we recorded. This is needed since we compute
220 /* The basic block where this record was created. We use this to determine
221 when to remove records. */
225 /* A hash table holding value range records (VRP_ELEMENTs) for a given
226 SSA_NAME. We used to use a varray indexed by SSA_NAME_VERSION, but
227 that gets awful wasteful, particularly since the density objects
228 with useful information is very low. */
229 static htab_t vrp_data;
231 /* An entry in the VRP_DATA hash table. We record the variable and a
232 varray of VRP_ELEMENT records associated with that variable. */
239 /* Array of variables which have their values constrained by operations
240 in this basic block. We use this during finalization to know
241 which variables need their VRP data updated. */
243 /* Stack of SSA_NAMEs which had their values constrained by operations
244 in this basic block. During finalization of this block we use this
245 list to determine which variables need their VRP data updated.
247 A NULL entry marks the end of the SSA_NAMEs associated with this block. */
248 static VEC(tree,heap) *vrp_variables_stack;
256 /* Local functions. */
257 static void optimize_stmt (struct dom_walk_data *,
259 block_stmt_iterator);
260 static tree lookup_avail_expr (tree, bool);
261 static hashval_t vrp_hash (const void *);
262 static int vrp_eq (const void *, const void *);
263 static hashval_t avail_expr_hash (const void *);
264 static hashval_t real_avail_expr_hash (const void *);
265 static int avail_expr_eq (const void *, const void *);
266 static void htab_statistics (FILE *, htab_t);
267 static void record_cond (tree, tree);
268 static void record_const_or_copy (tree, tree);
269 static void record_equality (tree, tree);
270 static tree update_rhs_and_lookup_avail_expr (tree, tree, bool);
271 static tree simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *,
273 static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int);
274 static tree simplify_switch_and_lookup_avail_expr (tree, int);
275 static tree find_equivalent_equality_comparison (tree);
276 static void record_range (tree, basic_block);
277 static bool extract_range_from_cond (tree, tree *, tree *, int *);
278 static void record_equivalences_from_phis (basic_block);
279 static void record_equivalences_from_incoming_edge (basic_block);
280 static bool eliminate_redundant_computations (struct dom_walk_data *,
282 static void record_equivalences_from_stmt (tree, int, stmt_ann_t);
283 static void thread_across_edge (struct dom_walk_data *, edge);
284 static void dom_opt_finalize_block (struct dom_walk_data *, basic_block);
285 static void dom_opt_initialize_block (struct dom_walk_data *, basic_block);
286 static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block);
287 static void remove_local_expressions_from_table (void);
288 static void restore_vars_to_original_value (void);
289 static edge single_incoming_edge_ignoring_loop_edges (basic_block);
290 static void restore_nonzero_vars_to_original_value (void);
291 static inline bool unsafe_associative_fp_binop (tree);
294 /* Local version of fold that doesn't introduce cruft. */
301 /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that
302 may have been added by fold, and "useless" type conversions that might
303 now be apparent due to propagation. */
304 STRIP_USELESS_TYPE_CONVERSION (t);
309 /* Allocate an EDGE_INFO for edge E and attach it to E.
310 Return the new EDGE_INFO structure. */
312 static struct edge_info *
313 allocate_edge_info (edge e)
315 struct edge_info *edge_info;
317 edge_info = xcalloc (1, sizeof (struct edge_info));
323 /* Free all EDGE_INFO structures associated with edges in the CFG.
324 If a particular edge can be threaded, copy the redirection
325 target from the EDGE_INFO structure into the edge's AUX field
326 as required by code to update the CFG and SSA graph for
330 free_all_edge_infos (void)
338 FOR_EACH_EDGE (e, ei, bb->preds)
340 struct edge_info *edge_info = e->aux;
344 e->aux = edge_info->redirection_target;
345 if (edge_info->cond_equivalences)
346 free (edge_info->cond_equivalences);
353 /* Jump threading, redundancy elimination and const/copy propagation.
355 This pass may expose new symbols that need to be renamed into SSA. For
356 every new symbol exposed, its corresponding bit will be set in
360 tree_ssa_dominator_optimize (void)
362 struct dom_walk_data walk_data;
364 struct loops loops_info;
366 memset (&opt_stats, 0, sizeof (opt_stats));
368 /* Create our hash tables. */
369 avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free);
370 vrp_data = htab_create (ceil_log2 (num_ssa_names), vrp_hash, vrp_eq, free);
371 avail_exprs_stack = VEC_alloc (tree, heap, 20);
372 const_and_copies_stack = VEC_alloc (tree, heap, 20);
373 nonzero_vars_stack = VEC_alloc (tree, heap, 20);
374 vrp_variables_stack = VEC_alloc (tree, heap, 20);
375 stmts_to_rescan = VEC_alloc (tree, heap, 20);
376 nonzero_vars = BITMAP_ALLOC (NULL);
377 threaded_blocks = BITMAP_ALLOC (NULL);
378 need_eh_cleanup = BITMAP_ALLOC (NULL);
380 /* Setup callbacks for the generic dominator tree walker. */
381 walk_data.walk_stmts_backward = false;
382 walk_data.dom_direction = CDI_DOMINATORS;
383 walk_data.initialize_block_local_data = NULL;
384 walk_data.before_dom_children_before_stmts = dom_opt_initialize_block;
385 walk_data.before_dom_children_walk_stmts = optimize_stmt;
386 walk_data.before_dom_children_after_stmts = propagate_to_outgoing_edges;
387 walk_data.after_dom_children_before_stmts = NULL;
388 walk_data.after_dom_children_walk_stmts = NULL;
389 walk_data.after_dom_children_after_stmts = dom_opt_finalize_block;
390 /* Right now we only attach a dummy COND_EXPR to the global data pointer.
391 When we attach more stuff we'll need to fill this out with a real
393 walk_data.global_data = NULL;
394 walk_data.block_local_data_size = 0;
395 walk_data.interesting_blocks = NULL;
397 /* Now initialize the dominator walker. */
398 init_walk_dominator_tree (&walk_data);
400 calculate_dominance_info (CDI_DOMINATORS);
402 /* We need to know which edges exit loops so that we can
403 aggressively thread through loop headers to an exit
405 flow_loops_find (&loops_info);
406 mark_loop_exit_edges (&loops_info);
407 flow_loops_free (&loops_info);
409 /* Clean up the CFG so that any forwarder blocks created by loop
410 canonicalization are removed. */
412 calculate_dominance_info (CDI_DOMINATORS);
414 /* If we prove certain blocks are unreachable, then we want to
415 repeat the dominator optimization process as PHI nodes may
416 have turned into copies which allows better propagation of
417 values. So we repeat until we do not identify any new unreachable
421 /* Optimize the dominator tree. */
424 /* We need accurate information regarding back edges in the CFG
425 for jump threading. */
426 mark_dfs_back_edges ();
428 /* Recursively walk the dominator tree optimizing statements. */
429 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
432 block_stmt_iterator bsi;
436 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
438 update_stmt_if_modified (bsi_stmt (bsi));
443 /* If we exposed any new variables, go ahead and put them into
444 SSA form now, before we handle jump threading. This simplifies
445 interactions between rewriting of _DECL nodes into SSA form
446 and rewriting SSA_NAME nodes into SSA form after block
447 duplication and CFG manipulation. */
448 update_ssa (TODO_update_ssa);
450 free_all_edge_infos ();
452 /* Thread jumps, creating duplicate blocks as needed. */
453 cfg_altered |= thread_through_all_blocks (threaded_blocks);
455 /* Removal of statements may make some EH edges dead. Purge
456 such edges from the CFG as needed. */
457 if (!bitmap_empty_p (need_eh_cleanup))
459 cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
460 bitmap_zero (need_eh_cleanup);
464 free_dominance_info (CDI_DOMINATORS);
466 cfg_altered = cleanup_tree_cfg ();
468 if (rediscover_loops_after_threading)
470 /* Rerun basic loop analysis to discover any newly
471 created loops and update the set of exit edges. */
472 rediscover_loops_after_threading = false;
473 flow_loops_find (&loops_info);
474 mark_loop_exit_edges (&loops_info);
475 flow_loops_free (&loops_info);
477 /* Remove any forwarder blocks inserted by loop
478 header canonicalization. */
482 calculate_dominance_info (CDI_DOMINATORS);
484 update_ssa (TODO_update_ssa);
486 /* Reinitialize the various tables. */
487 bitmap_clear (nonzero_vars);
488 bitmap_clear (threaded_blocks);
489 htab_empty (avail_exprs);
490 htab_empty (vrp_data);
492 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
494 This must be done before we iterate as we might have a
495 reference to an SSA_NAME which was removed by the call to
498 Long term we will be able to let everything in SSA_NAME_VALUE
499 persist. However, for now, we know this is the safe thing to do. */
500 for (i = 0; i < num_ssa_names; i++)
502 tree name = ssa_name (i);
508 value = SSA_NAME_VALUE (name);
509 if (value && !is_gimple_min_invariant (value))
510 SSA_NAME_VALUE (name) = NULL;
513 while (optimize > 1 && cfg_altered);
515 /* Debugging dumps. */
516 if (dump_file && (dump_flags & TDF_STATS))
517 dump_dominator_optimization_stats (dump_file);
519 /* We emptied the hash table earlier, now delete it completely. */
520 htab_delete (avail_exprs);
521 htab_delete (vrp_data);
523 /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
524 CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
525 of the do-while loop above. */
527 /* And finalize the dominator walker. */
528 fini_walk_dominator_tree (&walk_data);
530 /* Free nonzero_vars. */
531 BITMAP_FREE (nonzero_vars);
532 BITMAP_FREE (threaded_blocks);
533 BITMAP_FREE (need_eh_cleanup);
535 VEC_free (tree, heap, avail_exprs_stack);
536 VEC_free (tree, heap, const_and_copies_stack);
537 VEC_free (tree, heap, nonzero_vars_stack);
538 VEC_free (tree, heap, vrp_variables_stack);
539 VEC_free (tree, heap, stmts_to_rescan);
543 gate_dominator (void)
545 return flag_tree_dom != 0;
548 struct tree_opt_pass pass_dominator =
551 gate_dominator, /* gate */
552 tree_ssa_dominator_optimize, /* execute */
555 0, /* static_pass_number */
556 TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */
557 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
558 0, /* properties_provided */
559 0, /* properties_destroyed */
560 0, /* todo_flags_start */
563 | TODO_verify_ssa, /* todo_flags_finish */
568 /* We are exiting E->src, see if E->dest ends with a conditional
569 jump which has a known value when reached via E.
571 Special care is necessary if E is a back edge in the CFG as we
572 will have already recorded equivalences for E->dest into our
573 various tables, including the result of the conditional at
574 the end of E->dest. Threading opportunities are severely
575 limited in that case to avoid short-circuiting the loop
578 Note it is quite common for the first block inside a loop to
579 end with a conditional which is either always true or always
580 false when reached via the loop backedge. Thus we do not want
581 to blindly disable threading across a loop backedge. */
584 thread_across_edge (struct dom_walk_data *walk_data, edge e)
586 block_stmt_iterator bsi;
590 /* If E->dest does not end with a conditional, then there is
592 bsi = bsi_last (e->dest);
595 || (TREE_CODE (bsi_stmt (bsi)) != COND_EXPR
596 && TREE_CODE (bsi_stmt (bsi)) != GOTO_EXPR
597 && TREE_CODE (bsi_stmt (bsi)) != SWITCH_EXPR))
600 /* The basic idea here is to use whatever knowledge we have
601 from our dominator walk to simplify statements in E->dest,
602 with the ultimate goal being to simplify the conditional
603 at the end of E->dest.
605 Note that we must undo any changes we make to the underlying
606 statements as the simplifications we are making are control
607 flow sensitive (ie, the simplifications are valid when we
608 traverse E, but may not be valid on other paths to E->dest. */
610 /* Each PHI creates a temporary equivalence, record them. Again
611 these are context sensitive equivalences and will be removed
613 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
615 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
616 tree dst = PHI_RESULT (phi);
618 /* If the desired argument is not the same as this PHI's result
619 and it is set by a PHI in E->dest, then we can not thread
622 && TREE_CODE (src) == SSA_NAME
623 && TREE_CODE (SSA_NAME_DEF_STMT (src)) == PHI_NODE
624 && bb_for_stmt (SSA_NAME_DEF_STMT (src)) == e->dest)
627 record_const_or_copy (dst, src);
630 /* Try to simplify each statement in E->dest, ultimately leading to
631 a simplification of the COND_EXPR at the end of E->dest.
633 We might consider marking just those statements which ultimately
634 feed the COND_EXPR. It's not clear if the overhead of bookkeeping
635 would be recovered by trying to simplify fewer statements.
637 If we are able to simplify a statement into the form
638 SSA_NAME = (SSA_NAME | gimple invariant), then we can record
639 a context sensitive equivalency which may help us simplify
640 later statements in E->dest.
642 Failure to simplify into the form above merely means that the
643 statement provides no equivalences to help simplify later
644 statements. This does not prevent threading through E->dest. */
645 for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
649 stmt = bsi_stmt (bsi);
651 /* Ignore empty statements and labels. */
652 if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
655 /* Safely handle threading across loop backedges. This is
656 over conservative, but still allows us to capture the
657 majority of the cases where we can thread across a loop
659 if ((e->flags & EDGE_DFS_BACK) != 0
660 && TREE_CODE (stmt) != COND_EXPR
661 && TREE_CODE (stmt) != SWITCH_EXPR)
664 /* If the statement has volatile operands, then we assume we
665 can not thread through this block. This is overly
666 conservative in some ways. */
667 if (TREE_CODE (stmt) == ASM_EXPR && ASM_VOLATILE_P (stmt))
670 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
671 value, then do not try to simplify this statement as it will
672 not simplify in any way that is helpful for jump threading. */
673 if (TREE_CODE (stmt) != MODIFY_EXPR
674 || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
677 /* At this point we have a statement which assigns an RHS to an
678 SSA_VAR on the LHS. We want to try and simplify this statement
679 to expose more context sensitive equivalences which in turn may
680 allow us to simplify the condition at the end of the loop. */
681 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
682 cached_lhs = TREE_OPERAND (stmt, 1);
685 /* Copy the operands. */
689 unsigned int num, i = 0;
691 num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE));
692 copy = xcalloc (num, sizeof (tree));
694 /* Make a copy of the uses & vuses into USES_COPY, then cprop into
696 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
699 tree use = USE_FROM_PTR (use_p);
702 if (TREE_CODE (use) == SSA_NAME)
703 tmp = SSA_NAME_VALUE (use);
704 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
705 SET_USE (use_p, tmp);
708 /* Try to fold/lookup the new expression. Inserting the
709 expression into the hash table is unlikely to help
710 simplify anything later, so just query the hashtable. */
711 cached_lhs = fold (TREE_OPERAND (stmt, 1));
712 if (TREE_CODE (cached_lhs) != SSA_NAME
713 && !is_gimple_min_invariant (cached_lhs))
714 cached_lhs = lookup_avail_expr (stmt, false);
717 /* Restore the statement's original uses/defs. */
719 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
720 SET_USE (use_p, copy[i++]);
725 /* Record the context sensitive equivalence if we were able
726 to simplify this statement. */
728 && (TREE_CODE (cached_lhs) == SSA_NAME
729 || is_gimple_min_invariant (cached_lhs)))
730 record_const_or_copy (TREE_OPERAND (stmt, 0), cached_lhs);
733 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
736 && (TREE_CODE (stmt) == COND_EXPR
737 || TREE_CODE (stmt) == GOTO_EXPR
738 || TREE_CODE (stmt) == SWITCH_EXPR))
740 tree cond, cached_lhs;
742 /* Now temporarily cprop the operands and try to find the resulting
743 expression in the hash tables. */
744 if (TREE_CODE (stmt) == COND_EXPR)
745 cond = COND_EXPR_COND (stmt);
746 else if (TREE_CODE (stmt) == GOTO_EXPR)
747 cond = GOTO_DESTINATION (stmt);
749 cond = SWITCH_COND (stmt);
751 if (COMPARISON_CLASS_P (cond))
753 tree dummy_cond, op0, op1;
754 enum tree_code cond_code;
756 op0 = TREE_OPERAND (cond, 0);
757 op1 = TREE_OPERAND (cond, 1);
758 cond_code = TREE_CODE (cond);
760 /* Get the current value of both operands. */
761 if (TREE_CODE (op0) == SSA_NAME)
763 tree tmp = SSA_NAME_VALUE (op0);
764 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
768 if (TREE_CODE (op1) == SSA_NAME)
770 tree tmp = SSA_NAME_VALUE (op1);
771 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
775 /* Stuff the operator and operands into our dummy conditional
776 expression, creating the dummy conditional if necessary. */
777 dummy_cond = walk_data->global_data;
780 dummy_cond = build (cond_code, boolean_type_node, op0, op1);
781 dummy_cond = build (COND_EXPR, void_type_node,
782 dummy_cond, NULL, NULL);
783 walk_data->global_data = dummy_cond;
787 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
788 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
789 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
792 /* If the conditional folds to an invariant, then we are done,
793 otherwise look it up in the hash tables. */
794 cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
795 if (! is_gimple_min_invariant (cached_lhs))
797 cached_lhs = lookup_avail_expr (dummy_cond, false);
798 if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
799 cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
804 /* We can have conditionals which just test the state of a
805 variable rather than use a relational operator. These are
806 simpler to handle. */
807 else if (TREE_CODE (cond) == SSA_NAME)
810 cached_lhs = SSA_NAME_VALUE (cached_lhs);
811 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
815 cached_lhs = lookup_avail_expr (stmt, false);
819 edge taken_edge = find_taken_edge (e->dest, cached_lhs);
820 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
825 /* If we have a known destination for the conditional, then
826 we can perform this optimization, which saves at least one
827 conditional jump each time it applies since we get to
828 bypass the conditional at our original destination. */
831 struct edge_info *edge_info;
833 update_bb_profile_for_threading (e->dest, EDGE_FREQUENCY (e),
834 e->count, taken_edge);
838 edge_info = allocate_edge_info (e);
839 edge_info->redirection_target = taken_edge;
840 bitmap_set_bit (threaded_blocks, e->dest->index);
847 /* Initialize local stacks for this optimizer and record equivalences
848 upon entry to BB. Equivalences can come from the edge traversed to
849 reach BB or they may come from PHI nodes at the start of BB. */
852 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
855 if (dump_file && (dump_flags & TDF_DETAILS))
856 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
858 /* Push a marker on the stacks of local information so that we know how
859 far to unwind when we finalize this block. */
860 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
861 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
862 VEC_safe_push (tree, heap, nonzero_vars_stack, NULL_TREE);
863 VEC_safe_push (tree, heap, vrp_variables_stack, NULL_TREE);
865 record_equivalences_from_incoming_edge (bb);
867 /* PHI nodes can create equivalences too. */
868 record_equivalences_from_phis (bb);
871 /* Given an expression EXPR (a relational expression or a statement),
872 initialize the hash table element pointed by by ELEMENT. */
875 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
877 /* Hash table elements may be based on conditional expressions or statements.
879 For the former case, we have no annotation and we want to hash the
880 conditional expression. In the latter case we have an annotation and
881 we want to record the expression the statement evaluates. */
882 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
884 element->stmt = NULL;
887 else if (TREE_CODE (expr) == COND_EXPR)
889 element->stmt = expr;
890 element->rhs = COND_EXPR_COND (expr);
892 else if (TREE_CODE (expr) == SWITCH_EXPR)
894 element->stmt = expr;
895 element->rhs = SWITCH_COND (expr);
897 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
899 element->stmt = expr;
900 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
902 else if (TREE_CODE (expr) == GOTO_EXPR)
904 element->stmt = expr;
905 element->rhs = GOTO_DESTINATION (expr);
909 element->stmt = expr;
910 element->rhs = TREE_OPERAND (expr, 1);
914 element->hash = avail_expr_hash (element);
917 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
918 LIMIT entries left in LOCALs. */
921 remove_local_expressions_from_table (void)
923 /* Remove all the expressions made available in this block. */
924 while (VEC_length (tree, avail_exprs_stack) > 0)
926 struct expr_hash_elt element;
927 tree expr = VEC_pop (tree, avail_exprs_stack);
929 if (expr == NULL_TREE)
932 initialize_hash_element (expr, NULL, &element);
933 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
937 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
938 state, stopping when there are LIMIT entries left in LOCALs. */
941 restore_nonzero_vars_to_original_value (void)
943 while (VEC_length (tree, nonzero_vars_stack) > 0)
945 tree name = VEC_pop (tree, nonzero_vars_stack);
950 bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
954 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
955 CONST_AND_COPIES to its original state, stopping when we hit a
959 restore_vars_to_original_value (void)
961 while (VEC_length (tree, const_and_copies_stack) > 0)
963 tree prev_value, dest;
965 dest = VEC_pop (tree, const_and_copies_stack);
970 prev_value = VEC_pop (tree, const_and_copies_stack);
971 SSA_NAME_VALUE (dest) = prev_value;
975 /* We have finished processing the dominator children of BB, perform
976 any finalization actions in preparation for leaving this node in
977 the dominator tree. */
980 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
984 /* If we are at a leaf node in the dominator tree, see if we can thread
985 the edge from BB through its successor.
987 Do this before we remove entries from our equivalence tables. */
988 if (single_succ_p (bb)
989 && (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0
990 && (get_immediate_dominator (CDI_DOMINATORS, single_succ (bb)) != bb
991 || phi_nodes (single_succ (bb))))
994 thread_across_edge (walk_data, single_succ_edge (bb));
996 else if ((last = last_stmt (bb))
997 && TREE_CODE (last) == COND_EXPR
998 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
999 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
1000 && EDGE_COUNT (bb->succs) == 2
1001 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
1002 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
1004 edge true_edge, false_edge;
1006 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
1008 /* If the THEN arm is the end of a dominator tree or has PHI nodes,
1009 then try to thread through its edge. */
1010 if (get_immediate_dominator (CDI_DOMINATORS, true_edge->dest) != bb
1011 || phi_nodes (true_edge->dest))
1013 struct edge_info *edge_info;
1016 /* Push a marker onto the available expression stack so that we
1017 unwind any expressions related to the TRUE arm before processing
1018 the false arm below. */
1019 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
1020 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
1022 edge_info = true_edge->aux;
1024 /* If we have info associated with this edge, record it into
1025 our equivalency tables. */
1028 tree *cond_equivalences = edge_info->cond_equivalences;
1029 tree lhs = edge_info->lhs;
1030 tree rhs = edge_info->rhs;
1032 /* If we have a simple NAME = VALUE equivalency record it. */
1033 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1034 record_const_or_copy (lhs, rhs);
1036 /* If we have 0 = COND or 1 = COND equivalences, record them
1037 into our expression hash tables. */
1038 if (cond_equivalences)
1039 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1041 tree expr = cond_equivalences[i];
1042 tree value = cond_equivalences[i + 1];
1044 record_cond (expr, value);
1048 /* Now thread the edge. */
1049 thread_across_edge (walk_data, true_edge);
1051 /* And restore the various tables to their state before
1052 we threaded this edge. */
1053 remove_local_expressions_from_table ();
1054 restore_vars_to_original_value ();
1057 /* Similarly for the ELSE arm. */
1058 if (get_immediate_dominator (CDI_DOMINATORS, false_edge->dest) != bb
1059 || phi_nodes (false_edge->dest))
1061 struct edge_info *edge_info;
1064 edge_info = false_edge->aux;
1066 /* If we have info associated with this edge, record it into
1067 our equivalency tables. */
1070 tree *cond_equivalences = edge_info->cond_equivalences;
1071 tree lhs = edge_info->lhs;
1072 tree rhs = edge_info->rhs;
1074 /* If we have a simple NAME = VALUE equivalency record it. */
1075 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1076 record_const_or_copy (lhs, rhs);
1078 /* If we have 0 = COND or 1 = COND equivalences, record them
1079 into our expression hash tables. */
1080 if (cond_equivalences)
1081 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1083 tree expr = cond_equivalences[i];
1084 tree value = cond_equivalences[i + 1];
1086 record_cond (expr, value);
1090 thread_across_edge (walk_data, false_edge);
1092 /* No need to remove local expressions from our tables
1093 or restore vars to their original value as that will
1094 be done immediately below. */
1098 remove_local_expressions_from_table ();
1099 restore_nonzero_vars_to_original_value ();
1100 restore_vars_to_original_value ();
1102 /* Remove VRP records associated with this basic block. They are no
1105 To be efficient, we note which variables have had their values
1106 constrained in this block. So walk over each variable in the
1107 VRP_VARIABLEs array. */
1108 while (VEC_length (tree, vrp_variables_stack) > 0)
1110 tree var = VEC_pop (tree, vrp_variables_stack);
1111 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1114 /* Each variable has a stack of value range records. We want to
1115 invalidate those associated with our basic block. So we walk
1116 the array backwards popping off records associated with our
1117 block. Once we hit a record not associated with our block
1119 varray_type var_vrp_records;
1124 vrp_hash_elt.var = var;
1125 vrp_hash_elt.records = NULL;
1127 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1129 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1130 var_vrp_records = vrp_hash_elt_p->records;
1132 while (VARRAY_ACTIVE_SIZE (var_vrp_records) > 0)
1134 struct vrp_element *element
1135 = (struct vrp_element *)VARRAY_TOP_GENERIC_PTR (var_vrp_records);
1137 if (element->bb != bb)
1140 VARRAY_POP (var_vrp_records);
1144 /* If we queued any statements to rescan in this block, then
1145 go ahead and rescan them now. */
1146 while (VEC_length (tree, stmts_to_rescan) > 0)
1148 tree stmt = VEC_last (tree, stmts_to_rescan);
1149 basic_block stmt_bb = bb_for_stmt (stmt);
1154 VEC_pop (tree, stmts_to_rescan);
1155 mark_new_vars_to_rename (stmt);
1159 /* PHI nodes can create equivalences too.
1161 Ignoring any alternatives which are the same as the result, if
1162 all the alternatives are equal, then the PHI node creates an
1165 Additionally, if all the PHI alternatives are known to have a nonzero
1166 value, then the result of this PHI is known to have a nonzero value,
1167 even if we do not know its exact value. */
1170 record_equivalences_from_phis (basic_block bb)
1174 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1176 tree lhs = PHI_RESULT (phi);
1180 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1182 tree t = PHI_ARG_DEF (phi, i);
1184 /* Ignore alternatives which are the same as our LHS. Since
1185 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1186 can simply compare pointers. */
1190 /* If we have not processed an alternative yet, then set
1191 RHS to this alternative. */
1194 /* If we have processed an alternative (stored in RHS), then
1195 see if it is equal to this one. If it isn't, then stop
1197 else if (! operand_equal_for_phi_arg_p (rhs, t))
1201 /* If we had no interesting alternatives, then all the RHS alternatives
1202 must have been the same as LHS. */
1206 /* If we managed to iterate through each PHI alternative without
1207 breaking out of the loop, then we have a PHI which may create
1208 a useful equivalence. We do not need to record unwind data for
1209 this, since this is a true assignment and not an equivalence
1210 inferred from a comparison. All uses of this ssa name are dominated
1211 by this assignment, so unwinding just costs time and space. */
1212 if (i == PHI_NUM_ARGS (phi)
1213 && may_propagate_copy (lhs, rhs))
1214 SSA_NAME_VALUE (lhs) = rhs;
1216 /* Now see if we know anything about the nonzero property for the
1217 result of this PHI. */
1218 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1220 if (!PHI_ARG_NONZERO (phi, i))
1224 if (i == PHI_NUM_ARGS (phi))
1225 bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
1229 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1230 return that edge. Otherwise return NULL. */
1232 single_incoming_edge_ignoring_loop_edges (basic_block bb)
1238 FOR_EACH_EDGE (e, ei, bb->preds)
1240 /* A loop back edge can be identified by the destination of
1241 the edge dominating the source of the edge. */
1242 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
1245 /* If we have already seen a non-loop edge, then we must have
1246 multiple incoming non-loop edges and thus we return NULL. */
1250 /* This is the first non-loop incoming edge we have found. Record
1258 /* Record any equivalences created by the incoming edge to BB. If BB
1259 has more than one incoming edge, then no equivalence is created. */
1262 record_equivalences_from_incoming_edge (basic_block bb)
1266 struct edge_info *edge_info;
1268 /* If our parent block ended with a control statement, then we may be
1269 able to record some equivalences based on which outgoing edge from
1270 the parent was followed. */
1271 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
1273 e = single_incoming_edge_ignoring_loop_edges (bb);
1275 /* If we had a single incoming edge from our parent block, then enter
1276 any data associated with the edge into our tables. */
1277 if (e && e->src == parent)
1285 tree lhs = edge_info->lhs;
1286 tree rhs = edge_info->rhs;
1287 tree *cond_equivalences = edge_info->cond_equivalences;
1290 record_equality (lhs, rhs);
1292 if (cond_equivalences)
1294 bool recorded_range = false;
1295 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1297 tree expr = cond_equivalences[i];
1298 tree value = cond_equivalences[i + 1];
1300 record_cond (expr, value);
1302 /* For the first true equivalence, record range
1303 information. We only do this for the first
1304 true equivalence as it should dominate any
1305 later true equivalences. */
1306 if (! recorded_range
1307 && COMPARISON_CLASS_P (expr)
1308 && value == boolean_true_node
1309 && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
1311 record_range (expr, bb);
1312 recorded_range = true;
1320 /* Dump SSA statistics on FILE. */
1323 dump_dominator_optimization_stats (FILE *file)
1327 fprintf (file, "Total number of statements: %6ld\n\n",
1328 opt_stats.num_stmts);
1329 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
1330 opt_stats.num_exprs_considered);
1332 n_exprs = opt_stats.num_exprs_considered;
1336 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1337 opt_stats.num_re, PERCENT (opt_stats.num_re,
1339 fprintf (file, " Constants propagated: %6ld\n",
1340 opt_stats.num_const_prop);
1341 fprintf (file, " Copies propagated: %6ld\n",
1342 opt_stats.num_copy_prop);
1344 fprintf (file, "\nHash table statistics:\n");
1346 fprintf (file, " avail_exprs: ");
1347 htab_statistics (file, avail_exprs);
1351 /* Dump SSA statistics on stderr. */
1354 debug_dominator_optimization_stats (void)
1356 dump_dominator_optimization_stats (stderr);
1360 /* Dump statistics for the hash table HTAB. */
1363 htab_statistics (FILE *file, htab_t htab)
1365 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
1366 (long) htab_size (htab),
1367 (long) htab_elements (htab),
1368 htab_collisions (htab));
1371 /* Record the fact that VAR has a nonzero value, though we may not know
1372 its exact value. Note that if VAR is already known to have a nonzero
1373 value, then we do nothing. */
1376 record_var_is_nonzero (tree var)
1378 int indx = SSA_NAME_VERSION (var);
1380 if (bitmap_bit_p (nonzero_vars, indx))
1383 /* Mark it in the global table. */
1384 bitmap_set_bit (nonzero_vars, indx);
1386 /* Record this SSA_NAME so that we can reset the global table
1387 when we leave this block. */
1388 VEC_safe_push (tree, heap, nonzero_vars_stack, var);
1391 /* Enter a statement into the true/false expression hash table indicating
1392 that the condition COND has the value VALUE. */
1395 record_cond (tree cond, tree value)
1397 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
1400 initialize_hash_element (cond, value, element);
1402 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
1403 element->hash, INSERT);
1406 *slot = (void *) element;
1407 VEC_safe_push (tree, heap, avail_exprs_stack, cond);
1413 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1414 the new conditional into *p, then store a boolean_true_node
1418 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
1420 *p = build2 (new_code, boolean_type_node, op0, op1);
1422 *p = boolean_true_node;
1425 /* Record that COND is true and INVERTED is false into the edge information
1426 structure. Also record that any conditions dominated by COND are true
1429 For example, if a < b is true, then a <= b must also be true. */
1432 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
1436 if (!COMPARISON_CLASS_P (cond))
1439 op0 = TREE_OPERAND (cond, 0);
1440 op1 = TREE_OPERAND (cond, 1);
1442 switch (TREE_CODE (cond))
1446 edge_info->max_cond_equivalences = 12;
1447 edge_info->cond_equivalences = xmalloc (12 * sizeof (tree));
1448 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
1449 ? LE_EXPR : GE_EXPR),
1450 op0, op1, &edge_info->cond_equivalences[4]);
1451 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1452 &edge_info->cond_equivalences[6]);
1453 build_and_record_new_cond (NE_EXPR, op0, op1,
1454 &edge_info->cond_equivalences[8]);
1455 build_and_record_new_cond (LTGT_EXPR, op0, op1,
1456 &edge_info->cond_equivalences[10]);
1461 edge_info->max_cond_equivalences = 6;
1462 edge_info->cond_equivalences = xmalloc (6 * sizeof (tree));
1463 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1464 &edge_info->cond_equivalences[4]);
1468 edge_info->max_cond_equivalences = 10;
1469 edge_info->cond_equivalences = xmalloc (10 * sizeof (tree));
1470 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1471 &edge_info->cond_equivalences[4]);
1472 build_and_record_new_cond (LE_EXPR, op0, op1,
1473 &edge_info->cond_equivalences[6]);
1474 build_and_record_new_cond (GE_EXPR, op0, op1,
1475 &edge_info->cond_equivalences[8]);
1478 case UNORDERED_EXPR:
1479 edge_info->max_cond_equivalences = 16;
1480 edge_info->cond_equivalences = xmalloc (16 * sizeof (tree));
1481 build_and_record_new_cond (NE_EXPR, op0, op1,
1482 &edge_info->cond_equivalences[4]);
1483 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1484 &edge_info->cond_equivalences[6]);
1485 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1486 &edge_info->cond_equivalences[8]);
1487 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
1488 &edge_info->cond_equivalences[10]);
1489 build_and_record_new_cond (UNLT_EXPR, op0, op1,
1490 &edge_info->cond_equivalences[12]);
1491 build_and_record_new_cond (UNGT_EXPR, op0, op1,
1492 &edge_info->cond_equivalences[14]);
1497 edge_info->max_cond_equivalences = 8;
1498 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1499 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1500 ? UNLE_EXPR : UNGE_EXPR),
1501 op0, op1, &edge_info->cond_equivalences[4]);
1502 build_and_record_new_cond (NE_EXPR, op0, op1,
1503 &edge_info->cond_equivalences[6]);
1507 edge_info->max_cond_equivalences = 8;
1508 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1509 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1510 &edge_info->cond_equivalences[4]);
1511 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1512 &edge_info->cond_equivalences[6]);
1516 edge_info->max_cond_equivalences = 8;
1517 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1518 build_and_record_new_cond (NE_EXPR, op0, op1,
1519 &edge_info->cond_equivalences[4]);
1520 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1521 &edge_info->cond_equivalences[6]);
1525 edge_info->max_cond_equivalences = 4;
1526 edge_info->cond_equivalences = xmalloc (4 * sizeof (tree));
1530 /* Now store the original true and false conditions into the first
1532 edge_info->cond_equivalences[0] = cond;
1533 edge_info->cond_equivalences[1] = boolean_true_node;
1534 edge_info->cond_equivalences[2] = inverted;
1535 edge_info->cond_equivalences[3] = boolean_false_node;
1538 /* A helper function for record_const_or_copy and record_equality.
1539 Do the work of recording the value and undo info. */
1542 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1544 SSA_NAME_VALUE (x) = y;
1546 VEC_reserve (tree, heap, const_and_copies_stack, 2);
1547 VEC_quick_push (tree, const_and_copies_stack, prev_x);
1548 VEC_quick_push (tree, const_and_copies_stack, x);
1552 /* Return the loop depth of the basic block of the defining statement of X.
1553 This number should not be treated as absolutely correct because the loop
1554 information may not be completely up-to-date when dom runs. However, it
1555 will be relatively correct, and as more passes are taught to keep loop info
1556 up to date, the result will become more and more accurate. */
1559 loop_depth_of_name (tree x)
1564 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1565 if (TREE_CODE (x) != SSA_NAME)
1568 /* Otherwise return the loop depth of the defining statement's bb.
1569 Note that there may not actually be a bb for this statement, if the
1570 ssa_name is live on entry. */
1571 defstmt = SSA_NAME_DEF_STMT (x);
1572 defbb = bb_for_stmt (defstmt);
1576 return defbb->loop_depth;
1580 /* Record that X is equal to Y in const_and_copies. Record undo
1581 information in the block-local vector. */
1584 record_const_or_copy (tree x, tree y)
1586 tree prev_x = SSA_NAME_VALUE (x);
1588 if (TREE_CODE (y) == SSA_NAME)
1590 tree tmp = SSA_NAME_VALUE (y);
1595 record_const_or_copy_1 (x, y, prev_x);
1598 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1599 This constrains the cases in which we may treat this as assignment. */
1602 record_equality (tree x, tree y)
1604 tree prev_x = NULL, prev_y = NULL;
1606 if (TREE_CODE (x) == SSA_NAME)
1607 prev_x = SSA_NAME_VALUE (x);
1608 if (TREE_CODE (y) == SSA_NAME)
1609 prev_y = SSA_NAME_VALUE (y);
1611 /* If one of the previous values is invariant, or invariant in more loops
1612 (by depth), then use that.
1613 Otherwise it doesn't matter which value we choose, just so
1614 long as we canonicalize on one value. */
1615 if (TREE_INVARIANT (y))
1617 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1618 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1619 else if (prev_x && TREE_INVARIANT (prev_x))
1620 x = y, y = prev_x, prev_x = prev_y;
1621 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1624 /* After the swapping, we must have one SSA_NAME. */
1625 if (TREE_CODE (x) != SSA_NAME)
1628 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1629 variable compared against zero. If we're honoring signed zeros,
1630 then we cannot record this value unless we know that the value is
1632 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1633 && (TREE_CODE (y) != REAL_CST
1634 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1637 record_const_or_copy_1 (x, y, prev_x);
1640 /* Return true, if it is ok to do folding of an associative expression.
1641 EXP is the tree for the associative expression. */
1644 unsafe_associative_fp_binop (tree exp)
1646 enum tree_code code = TREE_CODE (exp);
1647 return !(!flag_unsafe_math_optimizations
1648 && (code == MULT_EXPR || code == PLUS_EXPR
1649 || code == MINUS_EXPR)
1650 && FLOAT_TYPE_P (TREE_TYPE (exp)));
1653 /* Returns true when STMT is a simple iv increment. It detects the
1654 following situation:
1656 i_1 = phi (..., i_2)
1657 i_2 = i_1 +/- ... */
1660 simple_iv_increment_p (tree stmt)
1662 tree lhs, rhs, preinc, phi;
1665 if (TREE_CODE (stmt) != MODIFY_EXPR)
1668 lhs = TREE_OPERAND (stmt, 0);
1669 if (TREE_CODE (lhs) != SSA_NAME)
1672 rhs = TREE_OPERAND (stmt, 1);
1674 if (TREE_CODE (rhs) != PLUS_EXPR
1675 && TREE_CODE (rhs) != MINUS_EXPR)
1678 preinc = TREE_OPERAND (rhs, 0);
1679 if (TREE_CODE (preinc) != SSA_NAME)
1682 phi = SSA_NAME_DEF_STMT (preinc);
1683 if (TREE_CODE (phi) != PHI_NODE)
1686 for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++)
1687 if (PHI_ARG_DEF (phi, i) == lhs)
1693 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1694 hash tables. Try to simplify the RHS using whatever equivalences
1695 we may have recorded.
1697 If we are able to simplify the RHS, then lookup the simplified form in
1698 the hash table and return the result. Otherwise return NULL. */
1701 simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *walk_data,
1702 tree stmt, int insert)
1704 tree rhs = TREE_OPERAND (stmt, 1);
1705 enum tree_code rhs_code = TREE_CODE (rhs);
1708 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1709 In which case we can change this statement to be lhs = y.
1710 Which can then be copy propagated.
1712 Similarly for negation. */
1713 if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR)
1714 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME)
1716 /* Get the definition statement for our RHS. */
1717 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1719 /* See if the RHS_DEF_STMT has the same form as our statement. */
1720 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR
1721 && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code)
1723 tree rhs_def_operand;
1725 rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0);
1727 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1728 if (TREE_CODE (rhs_def_operand) == SSA_NAME
1729 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand))
1730 result = update_rhs_and_lookup_avail_expr (stmt,
1736 /* If we have z = (x OP C1), see if we earlier had x = y OP C2.
1737 If OP is associative, create and fold (y OP C2) OP C1 which
1738 should result in (y OP C3), use that as the RHS for the
1739 assignment. Add minus to this, as we handle it specially below. */
1740 if ((associative_tree_code (rhs_code) || rhs_code == MINUS_EXPR)
1741 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME
1742 && is_gimple_min_invariant (TREE_OPERAND (rhs, 1)))
1744 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1746 /* If the statement defines an induction variable, do not propagate
1747 its value, so that we do not create overlapping life ranges. */
1748 if (simple_iv_increment_p (rhs_def_stmt))
1749 goto dont_fold_assoc;
1751 /* See if the RHS_DEF_STMT has the same form as our statement. */
1752 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR)
1754 tree rhs_def_rhs = TREE_OPERAND (rhs_def_stmt, 1);
1755 enum tree_code rhs_def_code = TREE_CODE (rhs_def_rhs);
1757 if ((rhs_code == rhs_def_code && unsafe_associative_fp_binop (rhs))
1758 || (rhs_code == PLUS_EXPR && rhs_def_code == MINUS_EXPR)
1759 || (rhs_code == MINUS_EXPR && rhs_def_code == PLUS_EXPR))
1761 tree def_stmt_op0 = TREE_OPERAND (rhs_def_rhs, 0);
1762 tree def_stmt_op1 = TREE_OPERAND (rhs_def_rhs, 1);
1764 if (TREE_CODE (def_stmt_op0) == SSA_NAME
1765 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0)
1766 && is_gimple_min_invariant (def_stmt_op1))
1768 tree outer_const = TREE_OPERAND (rhs, 1);
1769 tree type = TREE_TYPE (TREE_OPERAND (stmt, 0));
1772 /* If we care about correct floating point results, then
1773 don't fold x + c1 - c2. Note that we need to take both
1774 the codes and the signs to figure this out. */
1775 if (FLOAT_TYPE_P (type)
1776 && !flag_unsafe_math_optimizations
1777 && (rhs_def_code == PLUS_EXPR
1778 || rhs_def_code == MINUS_EXPR))
1782 neg ^= (rhs_code == MINUS_EXPR);
1783 neg ^= (rhs_def_code == MINUS_EXPR);
1784 neg ^= real_isneg (TREE_REAL_CST_PTR (outer_const));
1785 neg ^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1));
1788 goto dont_fold_assoc;
1791 /* Ho hum. So fold will only operate on the outermost
1792 thingy that we give it, so we have to build the new
1793 expression in two pieces. This requires that we handle
1794 combinations of plus and minus. */
1795 if (rhs_def_code != rhs_code)
1797 if (rhs_def_code == MINUS_EXPR)
1798 t = build (MINUS_EXPR, type, outer_const, def_stmt_op1);
1800 t = build (MINUS_EXPR, type, def_stmt_op1, outer_const);
1801 rhs_code = PLUS_EXPR;
1803 else if (rhs_def_code == MINUS_EXPR)
1804 t = build (PLUS_EXPR, type, def_stmt_op1, outer_const);
1806 t = build (rhs_def_code, type, def_stmt_op1, outer_const);
1808 t = build (rhs_code, type, def_stmt_op0, t);
1811 /* If the result is a suitable looking gimple expression,
1812 then use it instead of the original for STMT. */
1813 if (TREE_CODE (t) == SSA_NAME
1814 || (UNARY_CLASS_P (t)
1815 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME)
1816 || ((BINARY_CLASS_P (t) || COMPARISON_CLASS_P (t))
1817 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
1818 && is_gimple_val (TREE_OPERAND (t, 1))))
1819 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1826 /* Transform TRUNC_DIV_EXPR and TRUNC_MOD_EXPR into RSHIFT_EXPR
1827 and BIT_AND_EXPR respectively if the first operand is greater
1828 than zero and the second operand is an exact power of two. */
1829 if ((rhs_code == TRUNC_DIV_EXPR || rhs_code == TRUNC_MOD_EXPR)
1830 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0)))
1831 && integer_pow2p (TREE_OPERAND (rhs, 1)))
1834 tree op = TREE_OPERAND (rhs, 0);
1836 if (TYPE_UNSIGNED (TREE_TYPE (op)))
1838 val = integer_one_node;
1842 tree dummy_cond = walk_data->global_data;
1846 dummy_cond = build (GT_EXPR, boolean_type_node,
1847 op, integer_zero_node);
1848 dummy_cond = build (COND_EXPR, void_type_node,
1849 dummy_cond, NULL, NULL);
1850 walk_data->global_data = dummy_cond;
1854 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GT_EXPR);
1855 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1856 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1857 = integer_zero_node;
1859 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1862 if (val && integer_onep (val))
1865 tree op0 = TREE_OPERAND (rhs, 0);
1866 tree op1 = TREE_OPERAND (rhs, 1);
1868 if (rhs_code == TRUNC_DIV_EXPR)
1869 t = build (RSHIFT_EXPR, TREE_TYPE (op0), op0,
1870 build_int_cst (NULL_TREE, tree_log2 (op1)));
1872 t = build (BIT_AND_EXPR, TREE_TYPE (op0), op0,
1873 local_fold (build (MINUS_EXPR, TREE_TYPE (op1),
1874 op1, integer_one_node)));
1876 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1880 /* Transform ABS (X) into X or -X as appropriate. */
1881 if (rhs_code == ABS_EXPR
1882 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0))))
1885 tree op = TREE_OPERAND (rhs, 0);
1886 tree type = TREE_TYPE (op);
1888 if (TYPE_UNSIGNED (type))
1890 val = integer_zero_node;
1894 tree dummy_cond = walk_data->global_data;
1898 dummy_cond = build (LE_EXPR, boolean_type_node,
1899 op, integer_zero_node);
1900 dummy_cond = build (COND_EXPR, void_type_node,
1901 dummy_cond, NULL, NULL);
1902 walk_data->global_data = dummy_cond;
1906 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), LE_EXPR);
1907 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1908 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1909 = build_int_cst (type, 0);
1911 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1915 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GE_EXPR);
1916 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1917 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1918 = build_int_cst (type, 0);
1920 val = simplify_cond_and_lookup_avail_expr (dummy_cond,
1925 if (integer_zerop (val))
1926 val = integer_one_node;
1927 else if (integer_onep (val))
1928 val = integer_zero_node;
1934 && (integer_onep (val) || integer_zerop (val)))
1938 if (integer_onep (val))
1939 t = build1 (NEGATE_EXPR, TREE_TYPE (op), op);
1943 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1947 /* Optimize *"foo" into 'f'. This is done here rather than
1948 in fold to avoid problems with stuff like &*"foo". */
1949 if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF)
1951 tree t = fold_read_from_constant_string (rhs);
1954 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1960 /* COND is a condition of the form:
1962 x == const or x != const
1964 Look back to x's defining statement and see if x is defined as
1968 If const is unchanged if we convert it to type, then we can build
1969 the equivalent expression:
1972 y == const or y != const
1974 Which may allow further optimizations.
1976 Return the equivalent comparison or NULL if no such equivalent comparison
1980 find_equivalent_equality_comparison (tree cond)
1982 tree op0 = TREE_OPERAND (cond, 0);
1983 tree op1 = TREE_OPERAND (cond, 1);
1984 tree def_stmt = SSA_NAME_DEF_STMT (op0);
1986 /* OP0 might have been a parameter, so first make sure it
1987 was defined by a MODIFY_EXPR. */
1988 if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
1990 tree def_rhs = TREE_OPERAND (def_stmt, 1);
1992 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1993 if ((TREE_CODE (def_rhs) == NOP_EXPR
1994 || TREE_CODE (def_rhs) == CONVERT_EXPR)
1995 && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
1997 tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
1998 tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
2001 if (TYPE_PRECISION (def_rhs_inner_type)
2002 > TYPE_PRECISION (TREE_TYPE (def_rhs)))
2005 /* What we want to prove is that if we convert OP1 to
2006 the type of the object inside the NOP_EXPR that the
2007 result is still equivalent to SRC.
2009 If that is true, the build and return new equivalent
2010 condition which uses the source of the typecast and the
2011 new constant (which has only changed its type). */
2012 new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
2013 new = local_fold (new);
2014 if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
2015 return build (TREE_CODE (cond), TREE_TYPE (cond),
2016 def_rhs_inner, new);
2022 /* STMT is a COND_EXPR for which we could not trivially determine its
2023 result. This routine attempts to find equivalent forms of the
2024 condition which we may be able to optimize better. It also
2025 uses simple value range propagation to optimize conditionals. */
2028 simplify_cond_and_lookup_avail_expr (tree stmt,
2032 tree cond = COND_EXPR_COND (stmt);
2034 if (COMPARISON_CLASS_P (cond))
2036 tree op0 = TREE_OPERAND (cond, 0);
2037 tree op1 = TREE_OPERAND (cond, 1);
2039 if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
2042 tree low, high, cond_low, cond_high;
2043 int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
2044 varray_type vrp_records;
2045 struct vrp_element *element;
2046 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
2049 /* First see if we have test of an SSA_NAME against a constant
2050 where the SSA_NAME is defined by an earlier typecast which
2051 is irrelevant when performing tests against the given
2053 if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2055 tree new_cond = find_equivalent_equality_comparison (cond);
2059 /* Update the statement to use the new equivalent
2061 COND_EXPR_COND (stmt) = new_cond;
2063 /* If this is not a real stmt, ann will be NULL and we
2064 avoid processing the operands. */
2066 mark_stmt_modified (stmt);
2068 /* Lookup the condition and return its known value if it
2070 new_cond = lookup_avail_expr (stmt, insert);
2074 /* The operands have changed, so update op0 and op1. */
2075 op0 = TREE_OPERAND (cond, 0);
2076 op1 = TREE_OPERAND (cond, 1);
2080 /* Consult the value range records for this variable (if they exist)
2081 to see if we can eliminate or simplify this conditional.
2083 Note two tests are necessary to determine no records exist.
2084 First we have to see if the virtual array exists, if it
2085 exists, then we have to check its active size.
2087 Also note the vast majority of conditionals are not testing
2088 a variable which has had its range constrained by an earlier
2089 conditional. So this filter avoids a lot of unnecessary work. */
2090 vrp_hash_elt.var = op0;
2091 vrp_hash_elt.records = NULL;
2092 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
2096 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
2097 vrp_records = vrp_hash_elt_p->records;
2098 if (vrp_records == NULL)
2101 limit = VARRAY_ACTIVE_SIZE (vrp_records);
2103 /* If we have no value range records for this variable, or we are
2104 unable to extract a range for this condition, then there is
2107 || ! extract_range_from_cond (cond, &cond_high,
2108 &cond_low, &cond_inverted))
2111 /* We really want to avoid unnecessary computations of range
2112 info. So all ranges are computed lazily; this avoids a
2113 lot of unnecessary work. i.e., we record the conditional,
2114 but do not process how it constrains the variable's
2115 potential values until we know that processing the condition
2118 However, we do not want to have to walk a potentially long
2119 list of ranges, nor do we want to compute a variable's
2120 range more than once for a given path.
2122 Luckily, each time we encounter a conditional that can not
2123 be otherwise optimized we will end up here and we will
2124 compute the necessary range information for the variable
2125 used in this condition.
2127 Thus you can conclude that there will never be more than one
2128 conditional associated with a variable which has not been
2129 processed. So we never need to merge more than one new
2130 conditional into the current range.
2132 These properties also help us avoid unnecessary work. */
2134 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records, limit - 1);
2136 if (element->high && element->low)
2138 /* The last element has been processed, so there is no range
2139 merging to do, we can simply use the high/low values
2140 recorded in the last element. */
2142 high = element->high;
2146 tree tmp_high, tmp_low;
2149 /* The last element has not been processed. Process it now.
2150 record_range should ensure for cond inverted is not set.
2151 This call can only fail if cond is x < min or x > max,
2152 which fold should have optimized into false.
2153 If that doesn't happen, just pretend all values are
2155 if (! extract_range_from_cond (element->cond, &tmp_high,
2159 gcc_assert (dummy == 0);
2161 /* If this is the only element, then no merging is necessary,
2162 the high/low values from extract_range_from_cond are all
2171 /* Get the high/low value from the previous element. */
2172 struct vrp_element *prev
2173 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records,
2178 /* Merge in this element's range with the range from the
2181 The low value for the merged range is the maximum of
2182 the previous low value and the low value of this record.
2184 Similarly the high value for the merged range is the
2185 minimum of the previous high value and the high value of
2187 low = (low && tree_int_cst_compare (low, tmp_low) == 1
2189 high = (high && tree_int_cst_compare (high, tmp_high) == -1
2193 /* And record the computed range. */
2195 element->high = high;
2199 /* After we have constrained this variable's potential values,
2200 we try to determine the result of the given conditional.
2202 To simplify later tests, first determine if the current
2203 low value is the same low value as the conditional.
2204 Similarly for the current high value and the high value
2205 for the conditional. */
2206 lowequal = tree_int_cst_equal (low, cond_low);
2207 highequal = tree_int_cst_equal (high, cond_high);
2209 if (lowequal && highequal)
2210 return (cond_inverted ? boolean_false_node : boolean_true_node);
2212 /* To simplify the overlap/subset tests below we may want
2213 to swap the two ranges so that the larger of the two
2214 ranges occurs "first". */
2216 if (tree_int_cst_compare (low, cond_low) == 1
2218 && tree_int_cst_compare (cond_high, high) == 1))
2231 /* Now determine if there is no overlap in the ranges
2232 or if the second range is a subset of the first range. */
2233 no_overlap = tree_int_cst_lt (high, cond_low);
2234 subset = tree_int_cst_compare (cond_high, high) != 1;
2236 /* If there was no overlap in the ranges, then this conditional
2237 always has a false value (unless we had to invert this
2238 conditional, in which case it always has a true value). */
2240 return (cond_inverted ? boolean_true_node : boolean_false_node);
2242 /* If the current range is a subset of the condition's range,
2243 then this conditional always has a true value (unless we
2244 had to invert this conditional, in which case it always
2245 has a true value). */
2246 if (subset && swapped)
2247 return (cond_inverted ? boolean_false_node : boolean_true_node);
2249 /* We were unable to determine the result of the conditional.
2250 However, we may be able to simplify the conditional. First
2251 merge the ranges in the same manner as range merging above. */
2252 low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
2253 high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
2255 /* If the range has converged to a single point, then turn this
2256 into an equality comparison. */
2257 if (TREE_CODE (cond) != EQ_EXPR
2258 && TREE_CODE (cond) != NE_EXPR
2259 && tree_int_cst_equal (low, high))
2261 TREE_SET_CODE (cond, EQ_EXPR);
2262 TREE_OPERAND (cond, 1) = high;
2269 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2270 result. This routine attempts to find equivalent forms of the
2271 condition which we may be able to optimize better. */
2274 simplify_switch_and_lookup_avail_expr (tree stmt, int insert)
2276 tree cond = SWITCH_COND (stmt);
2279 /* The optimization that we really care about is removing unnecessary
2280 casts. That will let us do much better in propagating the inferred
2281 constant at the switch target. */
2282 if (TREE_CODE (cond) == SSA_NAME)
2284 def = SSA_NAME_DEF_STMT (cond);
2285 if (TREE_CODE (def) == MODIFY_EXPR)
2287 def = TREE_OPERAND (def, 1);
2288 if (TREE_CODE (def) == NOP_EXPR)
2293 def = TREE_OPERAND (def, 0);
2295 #ifdef ENABLE_CHECKING
2296 /* ??? Why was Jeff testing this? We are gimple... */
2297 gcc_assert (is_gimple_val (def));
2300 to = TREE_TYPE (cond);
2301 ti = TREE_TYPE (def);
2303 /* If we have an extension that preserves value, then we
2304 can copy the source value into the switch. */
2306 need_precision = TYPE_PRECISION (ti);
2308 if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti))
2310 else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti))
2311 need_precision += 1;
2312 if (TYPE_PRECISION (to) < need_precision)
2317 SWITCH_COND (stmt) = def;
2318 mark_stmt_modified (stmt);
2320 return lookup_avail_expr (stmt, insert);
2330 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2331 known value for that SSA_NAME (or NULL if no value is known).
2333 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2334 even if we don't know their precise value.
2336 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2337 nodes of the successors of BB. */
2340 cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
2345 FOR_EACH_EDGE (e, ei, bb->succs)
2350 /* If this is an abnormal edge, then we do not want to copy propagate
2351 into the PHI alternative associated with this edge. */
2352 if (e->flags & EDGE_ABNORMAL)
2355 phi = phi_nodes (e->dest);
2360 for ( ; phi; phi = PHI_CHAIN (phi))
2363 use_operand_p orig_p;
2366 /* The alternative may be associated with a constant, so verify
2367 it is an SSA_NAME before doing anything with it. */
2368 orig_p = PHI_ARG_DEF_PTR (phi, indx);
2369 orig = USE_FROM_PTR (orig_p);
2370 if (TREE_CODE (orig) != SSA_NAME)
2373 /* If the alternative is known to have a nonzero value, record
2374 that fact in the PHI node itself for future use. */
2375 if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
2376 PHI_ARG_NONZERO (phi, indx) = true;
2378 /* If we have *ORIG_P in our constant/copy table, then replace
2379 ORIG_P with its value in our constant/copy table. */
2380 new = SSA_NAME_VALUE (orig);
2383 && (TREE_CODE (new) == SSA_NAME
2384 || is_gimple_min_invariant (new))
2385 && may_propagate_copy (orig, new))
2386 propagate_value (orig_p, new);
2391 /* We have finished optimizing BB, record any information implied by
2392 taking a specific outgoing edge from BB. */
2395 record_edge_info (basic_block bb)
2397 block_stmt_iterator bsi = bsi_last (bb);
2398 struct edge_info *edge_info;
2400 if (! bsi_end_p (bsi))
2402 tree stmt = bsi_stmt (bsi);
2404 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
2406 tree cond = SWITCH_COND (stmt);
2408 if (TREE_CODE (cond) == SSA_NAME)
2410 tree labels = SWITCH_LABELS (stmt);
2411 int i, n_labels = TREE_VEC_LENGTH (labels);
2412 tree *info = xcalloc (n_basic_blocks, sizeof (tree));
2416 for (i = 0; i < n_labels; i++)
2418 tree label = TREE_VEC_ELT (labels, i);
2419 basic_block target_bb = label_to_block (CASE_LABEL (label));
2421 if (CASE_HIGH (label)
2422 || !CASE_LOW (label)
2423 || info[target_bb->index])
2424 info[target_bb->index] = error_mark_node;
2426 info[target_bb->index] = label;
2429 FOR_EACH_EDGE (e, ei, bb->succs)
2431 basic_block target_bb = e->dest;
2432 tree node = info[target_bb->index];
2434 if (node != NULL && node != error_mark_node)
2436 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
2437 edge_info = allocate_edge_info (e);
2438 edge_info->lhs = cond;
2446 /* A COND_EXPR may create equivalences too. */
2447 if (stmt && TREE_CODE (stmt) == COND_EXPR)
2449 tree cond = COND_EXPR_COND (stmt);
2453 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
2455 /* If the conditional is a single variable 'X', record 'X = 1'
2456 for the true edge and 'X = 0' on the false edge. */
2457 if (SSA_VAR_P (cond))
2459 struct edge_info *edge_info;
2461 edge_info = allocate_edge_info (true_edge);
2462 edge_info->lhs = cond;
2463 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
2465 edge_info = allocate_edge_info (false_edge);
2466 edge_info->lhs = cond;
2467 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
2469 /* Equality tests may create one or two equivalences. */
2470 else if (COMPARISON_CLASS_P (cond))
2472 tree op0 = TREE_OPERAND (cond, 0);
2473 tree op1 = TREE_OPERAND (cond, 1);
2475 /* Special case comparing booleans against a constant as we
2476 know the value of OP0 on both arms of the branch. i.e., we
2477 can record an equivalence for OP0 rather than COND. */
2478 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2479 && TREE_CODE (op0) == SSA_NAME
2480 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
2481 && is_gimple_min_invariant (op1))
2483 if (TREE_CODE (cond) == EQ_EXPR)
2485 edge_info = allocate_edge_info (true_edge);
2486 edge_info->lhs = op0;
2487 edge_info->rhs = (integer_zerop (op1)
2488 ? boolean_false_node
2489 : boolean_true_node);
2491 edge_info = allocate_edge_info (false_edge);
2492 edge_info->lhs = op0;
2493 edge_info->rhs = (integer_zerop (op1)
2495 : boolean_false_node);
2499 edge_info = allocate_edge_info (true_edge);
2500 edge_info->lhs = op0;
2501 edge_info->rhs = (integer_zerop (op1)
2503 : boolean_false_node);
2505 edge_info = allocate_edge_info (false_edge);
2506 edge_info->lhs = op0;
2507 edge_info->rhs = (integer_zerop (op1)
2508 ? boolean_false_node
2509 : boolean_true_node);
2513 else if (is_gimple_min_invariant (op0)
2514 && (TREE_CODE (op1) == SSA_NAME
2515 || is_gimple_min_invariant (op1)))
2517 tree inverted = invert_truthvalue (cond);
2518 struct edge_info *edge_info;
2520 edge_info = allocate_edge_info (true_edge);
2521 record_conditions (edge_info, cond, inverted);
2523 if (TREE_CODE (cond) == EQ_EXPR)
2525 edge_info->lhs = op1;
2526 edge_info->rhs = op0;
2529 edge_info = allocate_edge_info (false_edge);
2530 record_conditions (edge_info, inverted, cond);
2532 if (TREE_CODE (cond) == NE_EXPR)
2534 edge_info->lhs = op1;
2535 edge_info->rhs = op0;
2539 else if (TREE_CODE (op0) == SSA_NAME
2540 && (is_gimple_min_invariant (op1)
2541 || TREE_CODE (op1) == SSA_NAME))
2543 tree inverted = invert_truthvalue (cond);
2544 struct edge_info *edge_info;
2546 edge_info = allocate_edge_info (true_edge);
2547 record_conditions (edge_info, cond, inverted);
2549 if (TREE_CODE (cond) == EQ_EXPR)
2551 edge_info->lhs = op0;
2552 edge_info->rhs = op1;
2555 edge_info = allocate_edge_info (false_edge);
2556 record_conditions (edge_info, inverted, cond);
2558 if (TREE_CODE (cond) == NE_EXPR)
2560 edge_info->lhs = op0;
2561 edge_info->rhs = op1;
2566 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2571 /* Propagate information from BB to its outgoing edges.
2573 This can include equivalency information implied by control statements
2574 at the end of BB and const/copy propagation into PHIs in BB's
2575 successor blocks. */
2578 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2581 record_edge_info (bb);
2582 cprop_into_successor_phis (bb, nonzero_vars);
2585 /* Search for redundant computations in STMT. If any are found, then
2586 replace them with the variable holding the result of the computation.
2588 If safe, record this expression into the available expression hash
2592 eliminate_redundant_computations (struct dom_walk_data *walk_data,
2593 tree stmt, stmt_ann_t ann)
2595 tree *expr_p, def = NULL_TREE;
2598 bool retval = false;
2600 if (TREE_CODE (stmt) == MODIFY_EXPR)
2601 def = TREE_OPERAND (stmt, 0);
2603 /* Certain expressions on the RHS can be optimized away, but can not
2604 themselves be entered into the hash tables. */
2605 if (ann->makes_aliased_stores
2607 || TREE_CODE (def) != SSA_NAME
2608 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
2609 || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VMAYDEF)
2610 /* Do not record equivalences for increments of ivs. This would create
2611 overlapping live ranges for a very questionable gain. */
2612 || simple_iv_increment_p (stmt))
2615 /* Check if the expression has been computed before. */
2616 cached_lhs = lookup_avail_expr (stmt, insert);
2618 /* If this is an assignment and the RHS was not in the hash table,
2619 then try to simplify the RHS and lookup the new RHS in the
2621 if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR)
2622 cached_lhs = simplify_rhs_and_lookup_avail_expr (walk_data, stmt, insert);
2623 /* Similarly if this is a COND_EXPR and we did not find its
2624 expression in the hash table, simplify the condition and
2626 else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
2627 cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
2628 /* Similarly for a SWITCH_EXPR. */
2629 else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR)
2630 cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert);
2632 opt_stats.num_exprs_considered++;
2634 /* Get a pointer to the expression we are trying to optimize. */
2635 if (TREE_CODE (stmt) == COND_EXPR)
2636 expr_p = &COND_EXPR_COND (stmt);
2637 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2638 expr_p = &SWITCH_COND (stmt);
2639 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
2640 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
2642 expr_p = &TREE_OPERAND (stmt, 1);
2644 /* It is safe to ignore types here since we have already done
2645 type checking in the hashing and equality routines. In fact
2646 type checking here merely gets in the way of constant
2647 propagation. Also, make sure that it is safe to propagate
2648 CACHED_LHS into *EXPR_P. */
2650 && (TREE_CODE (cached_lhs) != SSA_NAME
2651 || may_propagate_copy (*expr_p, cached_lhs)))
2653 if (dump_file && (dump_flags & TDF_DETAILS))
2655 fprintf (dump_file, " Replaced redundant expr '");
2656 print_generic_expr (dump_file, *expr_p, dump_flags);
2657 fprintf (dump_file, "' with '");
2658 print_generic_expr (dump_file, cached_lhs, dump_flags);
2659 fprintf (dump_file, "'\n");
2664 #if defined ENABLE_CHECKING
2665 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
2666 || is_gimple_min_invariant (cached_lhs));
2669 if (TREE_CODE (cached_lhs) == ADDR_EXPR
2670 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
2671 && is_gimple_min_invariant (cached_lhs)))
2674 propagate_tree_value (expr_p, cached_lhs);
2675 mark_stmt_modified (stmt);
2680 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2681 the available expressions table or the const_and_copies table.
2682 Detect and record those equivalences. */
2685 record_equivalences_from_stmt (tree stmt,
2689 tree lhs = TREE_OPERAND (stmt, 0);
2690 enum tree_code lhs_code = TREE_CODE (lhs);
2693 if (lhs_code == SSA_NAME)
2695 tree rhs = TREE_OPERAND (stmt, 1);
2697 /* Strip away any useless type conversions. */
2698 STRIP_USELESS_TYPE_CONVERSION (rhs);
2700 /* If the RHS of the assignment is a constant or another variable that
2701 may be propagated, register it in the CONST_AND_COPIES table. We
2702 do not need to record unwind data for this, since this is a true
2703 assignment and not an equivalence inferred from a comparison. All
2704 uses of this ssa name are dominated by this assignment, so unwinding
2705 just costs time and space. */
2707 && (TREE_CODE (rhs) == SSA_NAME
2708 || is_gimple_min_invariant (rhs)))
2709 SSA_NAME_VALUE (lhs) = rhs;
2711 if (expr_computes_nonzero (rhs))
2712 record_var_is_nonzero (lhs);
2715 /* Look at both sides for pointer dereferences. If we find one, then
2716 the pointer must be nonnull and we can enter that equivalence into
2718 if (flag_delete_null_pointer_checks)
2719 for (i = 0; i < 2; i++)
2721 tree t = TREE_OPERAND (stmt, i);
2723 /* Strip away any COMPONENT_REFs. */
2724 while (TREE_CODE (t) == COMPONENT_REF)
2725 t = TREE_OPERAND (t, 0);
2727 /* Now see if this is a pointer dereference. */
2728 if (INDIRECT_REF_P (t))
2730 tree op = TREE_OPERAND (t, 0);
2732 /* If the pointer is a SSA variable, then enter new
2733 equivalences into the hash table. */
2734 while (TREE_CODE (op) == SSA_NAME)
2736 tree def = SSA_NAME_DEF_STMT (op);
2738 record_var_is_nonzero (op);
2740 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2741 which are known to have a nonzero value. */
2743 && TREE_CODE (def) == MODIFY_EXPR
2744 && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
2745 op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
2752 /* A memory store, even an aliased store, creates a useful
2753 equivalence. By exchanging the LHS and RHS, creating suitable
2754 vops and recording the result in the available expression table,
2755 we may be able to expose more redundant loads. */
2756 if (!ann->has_volatile_ops
2757 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
2758 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
2759 && !is_gimple_reg (lhs))
2761 tree rhs = TREE_OPERAND (stmt, 1);
2764 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2765 is a constant, we need to adjust the constant to fit into the
2766 type of the LHS. If the LHS is a bitfield and the RHS is not
2767 a constant, then we can not record any equivalences for this
2768 statement since we would need to represent the widening or
2769 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2770 and should not be necessary if GCC represented bitfields
2772 if (lhs_code == COMPONENT_REF
2773 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
2775 if (TREE_CONSTANT (rhs))
2776 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
2780 /* If the value overflowed, then we can not use this equivalence. */
2781 if (rhs && ! is_gimple_min_invariant (rhs))
2787 /* Build a new statement with the RHS and LHS exchanged. */
2788 new = build (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
2790 create_ssa_artficial_load_stmt (new, stmt);
2792 /* Finally enter the statement into the available expression
2794 lookup_avail_expr (new, true);
2799 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2800 CONST_AND_COPIES. */
2803 cprop_operand (tree stmt, use_operand_p op_p)
2805 bool may_have_exposed_new_symbols = false;
2807 tree op = USE_FROM_PTR (op_p);
2809 /* If the operand has a known constant value or it is known to be a
2810 copy of some other variable, use the value or copy stored in
2811 CONST_AND_COPIES. */
2812 val = SSA_NAME_VALUE (op);
2813 if (val && val != op && TREE_CODE (val) != VALUE_HANDLE)
2815 tree op_type, val_type;
2817 /* Do not change the base variable in the virtual operand
2818 tables. That would make it impossible to reconstruct
2819 the renamed virtual operand if we later modify this
2820 statement. Also only allow the new value to be an SSA_NAME
2821 for propagation into virtual operands. */
2822 if (!is_gimple_reg (op)
2823 && (TREE_CODE (val) != SSA_NAME
2824 || is_gimple_reg (val)
2825 || get_virtual_var (val) != get_virtual_var (op)))
2828 /* Do not replace hard register operands in asm statements. */
2829 if (TREE_CODE (stmt) == ASM_EXPR
2830 && !may_propagate_copy_into_asm (op))
2833 /* Get the toplevel type of each operand. */
2834 op_type = TREE_TYPE (op);
2835 val_type = TREE_TYPE (val);
2837 /* While both types are pointers, get the type of the object
2839 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
2841 op_type = TREE_TYPE (op_type);
2842 val_type = TREE_TYPE (val_type);
2845 /* Make sure underlying types match before propagating a constant by
2846 converting the constant to the proper type. Note that convert may
2847 return a non-gimple expression, in which case we ignore this
2848 propagation opportunity. */
2849 if (TREE_CODE (val) != SSA_NAME)
2851 if (!lang_hooks.types_compatible_p (op_type, val_type))
2853 val = fold_convert (TREE_TYPE (op), val);
2854 if (!is_gimple_min_invariant (val))
2859 /* Certain operands are not allowed to be copy propagated due
2860 to their interaction with exception handling and some GCC
2862 else if (!may_propagate_copy (op, val))
2865 /* Do not propagate copies if the propagated value is at a deeper loop
2866 depth than the propagatee. Otherwise, this may move loop variant
2867 variables outside of their loops and prevent coalescing
2868 opportunities. If the value was loop invariant, it will be hoisted
2869 by LICM and exposed for copy propagation. */
2870 if (loop_depth_of_name (val) > loop_depth_of_name (op))
2874 if (dump_file && (dump_flags & TDF_DETAILS))
2876 fprintf (dump_file, " Replaced '");
2877 print_generic_expr (dump_file, op, dump_flags);
2878 fprintf (dump_file, "' with %s '",
2879 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
2880 print_generic_expr (dump_file, val, dump_flags);
2881 fprintf (dump_file, "'\n");
2884 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2885 that we may have exposed a new symbol for SSA renaming. */
2886 if (TREE_CODE (val) == ADDR_EXPR
2887 || (POINTER_TYPE_P (TREE_TYPE (op))
2888 && is_gimple_min_invariant (val)))
2889 may_have_exposed_new_symbols = true;
2891 if (TREE_CODE (val) != SSA_NAME)
2892 opt_stats.num_const_prop++;
2894 opt_stats.num_copy_prop++;
2896 propagate_value (op_p, val);
2898 /* And note that we modified this statement. This is now
2899 safe, even if we changed virtual operands since we will
2900 rescan the statement and rewrite its operands again. */
2901 mark_stmt_modified (stmt);
2903 return may_have_exposed_new_symbols;
2906 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2907 known value for that SSA_NAME (or NULL if no value is known).
2909 Propagate values from CONST_AND_COPIES into the uses, vuses and
2910 v_may_def_ops of STMT. */
2913 cprop_into_stmt (tree stmt)
2915 bool may_have_exposed_new_symbols = false;
2919 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
2921 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
2922 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
2925 return may_have_exposed_new_symbols;
2929 /* Optimize the statement pointed by iterator SI.
2931 We try to perform some simplistic global redundancy elimination and
2932 constant propagation:
2934 1- To detect global redundancy, we keep track of expressions that have
2935 been computed in this block and its dominators. If we find that the
2936 same expression is computed more than once, we eliminate repeated
2937 computations by using the target of the first one.
2939 2- Constant values and copy assignments. This is used to do very
2940 simplistic constant and copy propagation. When a constant or copy
2941 assignment is found, we map the value on the RHS of the assignment to
2942 the variable in the LHS in the CONST_AND_COPIES table. */
2945 optimize_stmt (struct dom_walk_data *walk_data, basic_block bb,
2946 block_stmt_iterator si)
2949 tree stmt, old_stmt;
2950 bool may_optimize_p;
2951 bool may_have_exposed_new_symbols = false;
2953 old_stmt = stmt = bsi_stmt (si);
2955 update_stmt_if_modified (stmt);
2956 ann = stmt_ann (stmt);
2957 opt_stats.num_stmts++;
2958 may_have_exposed_new_symbols = false;
2960 if (dump_file && (dump_flags & TDF_DETAILS))
2962 fprintf (dump_file, "Optimizing statement ");
2963 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2966 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2967 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
2969 /* If the statement has been modified with constant replacements,
2970 fold its RHS before checking for redundant computations. */
2975 /* Try to fold the statement making sure that STMT is kept
2977 if (fold_stmt (bsi_stmt_ptr (si)))
2979 stmt = bsi_stmt (si);
2980 ann = stmt_ann (stmt);
2982 if (dump_file && (dump_flags & TDF_DETAILS))
2984 fprintf (dump_file, " Folded to: ");
2985 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2989 rhs = get_rhs (stmt);
2990 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
2991 recompute_tree_invarant_for_addr_expr (rhs);
2993 /* Constant/copy propagation above may change the set of
2994 virtual operands associated with this statement. Folding
2995 may remove the need for some virtual operands.
2997 Indicate we will need to rescan and rewrite the statement. */
2998 may_have_exposed_new_symbols = true;
3001 /* Check for redundant computations. Do this optimization only
3002 for assignments that have no volatile ops and conditionals. */
3003 may_optimize_p = (!ann->has_volatile_ops
3004 && ((TREE_CODE (stmt) == RETURN_EXPR
3005 && TREE_OPERAND (stmt, 0)
3006 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
3007 && ! (TREE_SIDE_EFFECTS
3008 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
3009 || (TREE_CODE (stmt) == MODIFY_EXPR
3010 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
3011 || TREE_CODE (stmt) == COND_EXPR
3012 || TREE_CODE (stmt) == SWITCH_EXPR));
3015 may_have_exposed_new_symbols
3016 |= eliminate_redundant_computations (walk_data, stmt, ann);
3018 /* Record any additional equivalences created by this statement. */
3019 if (TREE_CODE (stmt) == MODIFY_EXPR)
3020 record_equivalences_from_stmt (stmt,
3024 /* If STMT is a COND_EXPR and it was modified, then we may know
3025 where it goes. If that is the case, then mark the CFG as altered.
3027 This will cause us to later call remove_unreachable_blocks and
3028 cleanup_tree_cfg when it is safe to do so. It is not safe to
3029 clean things up here since removal of edges and such can trigger
3030 the removal of PHI nodes, which in turn can release SSA_NAMEs to
3033 That's all fine and good, except that once SSA_NAMEs are released
3034 to the manager, we must not call create_ssa_name until all references
3035 to released SSA_NAMEs have been eliminated.
3037 All references to the deleted SSA_NAMEs can not be eliminated until
3038 we remove unreachable blocks.
3040 We can not remove unreachable blocks until after we have completed
3041 any queued jump threading.
3043 We can not complete any queued jump threads until we have taken
3044 appropriate variables out of SSA form. Taking variables out of
3045 SSA form can call create_ssa_name and thus we lose.
3047 Ultimately I suspect we're going to need to change the interface
3048 into the SSA_NAME manager. */
3054 if (TREE_CODE (stmt) == COND_EXPR)
3055 val = COND_EXPR_COND (stmt);
3056 else if (TREE_CODE (stmt) == SWITCH_EXPR)
3057 val = SWITCH_COND (stmt);
3059 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
3062 /* If we simplified a statement in such a way as to be shown that it
3063 cannot trap, update the eh information and the cfg to match. */
3064 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
3066 bitmap_set_bit (need_eh_cleanup, bb->index);
3067 if (dump_file && (dump_flags & TDF_DETAILS))
3068 fprintf (dump_file, " Flagged to clear EH edges.\n");
3072 if (may_have_exposed_new_symbols)
3073 VEC_safe_push (tree, heap, stmts_to_rescan, bsi_stmt (si));
3076 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
3077 available expression hashtable, then return the LHS from the hash
3080 If INSERT is true, then we also update the available expression
3081 hash table to account for the changes made to STMT. */
3084 update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert)
3086 tree cached_lhs = NULL;
3088 /* Remove the old entry from the hash table. */
3091 struct expr_hash_elt element;
3093 initialize_hash_element (stmt, NULL, &element);
3094 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
3097 /* Now update the RHS of the assignment. */
3098 TREE_OPERAND (stmt, 1) = new_rhs;
3100 /* Now lookup the updated statement in the hash table. */
3101 cached_lhs = lookup_avail_expr (stmt, insert);
3103 /* We have now called lookup_avail_expr twice with two different
3104 versions of this same statement, once in optimize_stmt, once here.
3106 We know the call in optimize_stmt did not find an existing entry
3107 in the hash table, so a new entry was created. At the same time
3108 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
3110 If this call failed to find an existing entry on the hash table,
3111 then the new version of this statement was entered into the
3112 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3113 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3115 If this call succeeded, we still have one copy of this statement
3116 on the BLOCK_AVAIL_EXPRs vector.
3118 For both cases, we need to pop the most recent entry off the
3119 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3120 statement in the hash tables, that will leave precisely one
3121 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3122 we found a copy of this statement in the second hash table lookup
3123 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3125 VEC_pop (tree, avail_exprs_stack);
3127 /* And make sure we record the fact that we modified this
3129 mark_stmt_modified (stmt);
3134 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3135 found, return its LHS. Otherwise insert STMT in the table and return
3138 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3139 is also added to the stack pointed by BLOCK_AVAIL_EXPRS_P, so that they
3140 can be removed when we finish processing this block and its children.
3142 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3143 contains no CALL_EXPR on its RHS and makes no volatile nor
3144 aliased references. */
3147 lookup_avail_expr (tree stmt, bool insert)
3152 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
3154 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
3156 initialize_hash_element (stmt, lhs, element);
3158 /* Don't bother remembering constant assignments and copy operations.
3159 Constants and copy operations are handled by the constant/copy propagator
3160 in optimize_stmt. */
3161 if (TREE_CODE (element->rhs) == SSA_NAME
3162 || is_gimple_min_invariant (element->rhs))
3168 /* If this is an equality test against zero, see if we have recorded a
3169 nonzero value for the variable in question. */
3170 if ((TREE_CODE (element->rhs) == EQ_EXPR
3171 || TREE_CODE (element->rhs) == NE_EXPR)
3172 && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
3173 && integer_zerop (TREE_OPERAND (element->rhs, 1)))
3175 int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
3177 if (bitmap_bit_p (nonzero_vars, indx))
3179 tree t = element->rhs;
3182 if (TREE_CODE (t) == EQ_EXPR)
3183 return boolean_false_node;
3185 return boolean_true_node;
3189 /* Finally try to find the expression in the main expression hash table. */
3190 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
3191 (insert ? INSERT : NO_INSERT));
3200 *slot = (void *) element;
3201 VEC_safe_push (tree, heap, avail_exprs_stack,
3202 stmt ? stmt : element->rhs);
3206 /* Extract the LHS of the assignment so that it can be used as the current
3207 definition of another variable. */
3208 lhs = ((struct expr_hash_elt *)*slot)->lhs;
3210 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3211 use the value from the const_and_copies table. */
3212 if (TREE_CODE (lhs) == SSA_NAME)
3214 temp = SSA_NAME_VALUE (lhs);
3215 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
3223 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3224 range of values that result in the conditional having a true value.
3226 Return true if we are successful in extracting a range from COND and
3227 false if we are unsuccessful. */
3230 extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
3232 tree op1 = TREE_OPERAND (cond, 1);
3233 tree high, low, type;
3236 type = TREE_TYPE (op1);
3238 /* Experiments have shown that it's rarely, if ever useful to
3239 record ranges for enumerations. Presumably this is due to
3240 the fact that they're rarely used directly. They are typically
3241 cast into an integer type and used that way. */
3242 if (TREE_CODE (type) != INTEGER_TYPE
3243 /* We don't know how to deal with types with variable bounds. */
3244 || TREE_CODE (TYPE_MIN_VALUE (type)) != INTEGER_CST
3245 || TREE_CODE (TYPE_MAX_VALUE (type)) != INTEGER_CST)
3248 switch (TREE_CODE (cond))
3262 high = TYPE_MAX_VALUE (type);
3267 high = TYPE_MAX_VALUE (type);
3268 if (!tree_int_cst_lt (op1, high))
3270 low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
3276 low = TYPE_MIN_VALUE (type);
3281 low = TYPE_MIN_VALUE (type);
3282 if (!tree_int_cst_lt (low, op1))
3284 high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
3294 *inverted_p = inverted;
3298 /* Record a range created by COND for basic block BB. */
3301 record_range (tree cond, basic_block bb)
3303 enum tree_code code = TREE_CODE (cond);
3305 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3306 They rarely allow for meaningful range optimizations and significantly
3307 complicate the implementation. */
3308 if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
3309 || code == GE_EXPR || code == EQ_EXPR)
3310 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
3312 struct vrp_hash_elt *vrp_hash_elt;
3313 struct vrp_element *element;
3314 varray_type *vrp_records_p;
3318 vrp_hash_elt = xmalloc (sizeof (struct vrp_hash_elt));
3319 vrp_hash_elt->var = TREE_OPERAND (cond, 0);
3320 vrp_hash_elt->records = NULL;
3321 slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
3324 *slot = (void *) vrp_hash_elt;
3326 free (vrp_hash_elt);
3328 vrp_hash_elt = (struct vrp_hash_elt *) *slot;
3329 vrp_records_p = &vrp_hash_elt->records;
3331 element = ggc_alloc (sizeof (struct vrp_element));
3332 element->low = NULL;
3333 element->high = NULL;
3334 element->cond = cond;
3337 if (*vrp_records_p == NULL)
3338 VARRAY_GENERIC_PTR_INIT (*vrp_records_p, 2, "vrp records");
3340 VARRAY_PUSH_GENERIC_PTR (*vrp_records_p, element);
3341 VEC_safe_push (tree, heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
3345 /* Hashing and equality functions for VRP_DATA.
3347 Since this hash table is addressed by SSA_NAMEs, we can hash on
3348 their version number and equality can be determined with a
3349 pointer comparison. */
3352 vrp_hash (const void *p)
3354 tree var = ((struct vrp_hash_elt *)p)->var;
3356 return SSA_NAME_VERSION (var);
3360 vrp_eq (const void *p1, const void *p2)
3362 tree var1 = ((struct vrp_hash_elt *)p1)->var;
3363 tree var2 = ((struct vrp_hash_elt *)p2)->var;
3365 return var1 == var2;
3368 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3369 MODIFY_EXPR statements. We compute a value number for expressions using
3370 the code of the expression and the SSA numbers of its operands. */
3373 avail_expr_hash (const void *p)
3375 tree stmt = ((struct expr_hash_elt *)p)->stmt;
3376 tree rhs = ((struct expr_hash_elt *)p)->rhs;
3381 /* iterative_hash_expr knows how to deal with any expression and
3382 deals with commutative operators as well, so just use it instead
3383 of duplicating such complexities here. */
3384 val = iterative_hash_expr (rhs, val);
3386 /* If the hash table entry is not associated with a statement, then we
3387 can just hash the expression and not worry about virtual operands
3389 if (!stmt || !stmt_ann (stmt))
3392 /* Add the SSA version numbers of every vuse operand. This is important
3393 because compound variables like arrays are not renamed in the
3394 operands. Rather, the rename is done on the virtual variable
3395 representing all the elements of the array. */
3396 FOR_EACH_SSA_TREE_OPERAND (vuse, stmt, iter, SSA_OP_VUSE)
3397 val = iterative_hash_expr (vuse, val);
3403 real_avail_expr_hash (const void *p)
3405 return ((const struct expr_hash_elt *)p)->hash;
3409 avail_expr_eq (const void *p1, const void *p2)
3411 tree stmt1 = ((struct expr_hash_elt *)p1)->stmt;
3412 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
3413 tree stmt2 = ((struct expr_hash_elt *)p2)->stmt;
3414 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
3416 /* If they are the same physical expression, return true. */
3417 if (rhs1 == rhs2 && stmt1 == stmt2)
3420 /* If their codes are not equal, then quit now. */
3421 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
3424 /* In case of a collision, both RHS have to be identical and have the
3425 same VUSE operands. */
3426 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
3427 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
3428 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
3430 bool ret = compare_ssa_operands_equal (stmt1, stmt2, SSA_OP_VUSE);
3431 gcc_assert (!ret || ((struct expr_hash_elt *)p1)->hash
3432 == ((struct expr_hash_elt *)p2)->hash);