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, 51 Franklin Street, Fifth Floor,
20 Boston, MA 02110-1301, USA. */
24 #include "coretypes.h"
31 #include "basic-block.h"
36 #include "diagnostic.h"
38 #include "tree-dump.h"
39 #include "tree-flow.h"
42 #include "tree-pass.h"
43 #include "tree-ssa-propagate.h"
44 #include "langhooks.h"
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;
173 static struct opt_stats_d opt_stats;
175 /* Value range propagation record. Each time we encounter a conditional
176 of the form SSA_NAME COND CONST we create a new vrp_element to record
177 how the condition affects the possible values SSA_NAME may have.
179 Each record contains the condition tested (COND), and the range of
180 values the variable may legitimately have if COND is true. Note the
181 range of values may be a smaller range than COND specifies if we have
182 recorded other ranges for this variable. Each record also contains the
183 block in which the range was recorded for invalidation purposes.
185 Note that the current known range is computed lazily. This allows us
186 to avoid the overhead of computing ranges which are never queried.
188 When we encounter a conditional, we look for records which constrain
189 the SSA_NAME used in the condition. In some cases those records allow
190 us to determine the condition's result at compile time. In other cases
191 they may allow us to simplify the condition.
193 We also use value ranges to do things like transform signed div/mod
194 operations into unsigned div/mod or to simplify ABS_EXPRs.
196 Simple experiments have shown these optimizations to not be all that
197 useful on switch statements (much to my surprise). So switch statement
198 optimizations are not performed.
200 Note carefully we do not propagate information through each statement
201 in the block. i.e., if we know variable X has a value defined of
202 [0, 25] and we encounter Y = X + 1, we do not track a value range
203 for Y (which would be [1, 26] if we cared). Similarly we do not
204 constrain values as we encounter narrowing typecasts, etc. */
208 /* The highest and lowest values the variable in COND may contain when
209 COND is true. Note this may not necessarily be the same values
210 tested by COND if the same variable was used in earlier conditionals.
212 Note this is computed lazily and thus can be NULL indicating that
213 the values have not been computed yet. */
217 /* The actual conditional we recorded. This is needed since we compute
221 /* The basic block where this record was created. We use this to determine
222 when to remove records. */
226 /* A hash table holding value range records (VRP_ELEMENTs) for a given
227 SSA_NAME. We used to use a varray indexed by SSA_NAME_VERSION, but
228 that gets awful wasteful, particularly since the density objects
229 with useful information is very low. */
230 static htab_t vrp_data;
232 typedef struct vrp_element *vrp_element_p;
234 DEF_VEC_P(vrp_element_p);
235 DEF_VEC_ALLOC_P(vrp_element_p,heap);
237 /* An entry in the VRP_DATA hash table. We record the variable and a
238 varray of VRP_ELEMENT records associated with that variable. */
242 VEC(vrp_element_p,heap) *records;
245 /* Array of variables which have their values constrained by operations
246 in this basic block. We use this during finalization to know
247 which variables need their VRP data updated. */
249 /* Stack of SSA_NAMEs which had their values constrained by operations
250 in this basic block. During finalization of this block we use this
251 list to determine which variables need their VRP data updated.
253 A NULL entry marks the end of the SSA_NAMEs associated with this block. */
254 static VEC(tree,heap) *vrp_variables_stack;
262 /* Local functions. */
263 static void optimize_stmt (struct dom_walk_data *,
265 block_stmt_iterator);
266 static tree lookup_avail_expr (tree, bool);
267 static hashval_t vrp_hash (const void *);
268 static int vrp_eq (const void *, const void *);
269 static hashval_t avail_expr_hash (const void *);
270 static hashval_t real_avail_expr_hash (const void *);
271 static int avail_expr_eq (const void *, const void *);
272 static void htab_statistics (FILE *, htab_t);
273 static void record_cond (tree, tree);
274 static void record_const_or_copy (tree, tree);
275 static void record_equality (tree, tree);
276 static tree update_rhs_and_lookup_avail_expr (tree, tree, bool);
277 static tree simplify_rhs_and_lookup_avail_expr (tree, int);
278 static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int);
279 static tree simplify_switch_and_lookup_avail_expr (tree, int);
280 static tree find_equivalent_equality_comparison (tree);
281 static void record_range (tree, basic_block);
282 static bool extract_range_from_cond (tree, tree *, tree *, int *);
283 static void record_equivalences_from_phis (basic_block);
284 static void record_equivalences_from_incoming_edge (basic_block);
285 static bool eliminate_redundant_computations (tree, stmt_ann_t);
286 static void record_equivalences_from_stmt (tree, int, stmt_ann_t);
287 static void thread_across_edge (struct dom_walk_data *, edge);
288 static void dom_opt_finalize_block (struct dom_walk_data *, basic_block);
289 static void dom_opt_initialize_block (struct dom_walk_data *, basic_block);
290 static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block);
291 static void remove_local_expressions_from_table (void);
292 static void restore_vars_to_original_value (void);
293 static edge single_incoming_edge_ignoring_loop_edges (basic_block);
294 static void restore_nonzero_vars_to_original_value (void);
295 static inline bool unsafe_associative_fp_binop (tree);
298 /* Local version of fold that doesn't introduce cruft. */
305 /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that
306 may have been added by fold, and "useless" type conversions that might
307 now be apparent due to propagation. */
308 STRIP_USELESS_TYPE_CONVERSION (t);
313 /* Allocate an EDGE_INFO for edge E and attach it to E.
314 Return the new EDGE_INFO structure. */
316 static struct edge_info *
317 allocate_edge_info (edge e)
319 struct edge_info *edge_info;
321 edge_info = xcalloc (1, sizeof (struct edge_info));
327 /* Free all EDGE_INFO structures associated with edges in the CFG.
328 If a particular edge can be threaded, copy the redirection
329 target from the EDGE_INFO structure into the edge's AUX field
330 as required by code to update the CFG and SSA graph for
334 free_all_edge_infos (void)
342 FOR_EACH_EDGE (e, ei, bb->preds)
344 struct edge_info *edge_info = e->aux;
348 e->aux = edge_info->redirection_target;
349 if (edge_info->cond_equivalences)
350 free (edge_info->cond_equivalences);
357 /* Free an instance of vrp_hash_elt. */
360 vrp_free (void *data)
362 struct vrp_hash_elt *elt = data;
363 struct VEC(vrp_element_p,heap) **vrp_elt = &elt->records;
365 VEC_free (vrp_element_p, heap, *vrp_elt);
369 /* Jump threading, redundancy elimination and const/copy propagation.
371 This pass may expose new symbols that need to be renamed into SSA. For
372 every new symbol exposed, its corresponding bit will be set in
376 tree_ssa_dominator_optimize (void)
378 struct dom_walk_data walk_data;
380 struct loops loops_info;
382 memset (&opt_stats, 0, sizeof (opt_stats));
384 /* Create our hash tables. */
385 avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free);
386 vrp_data = htab_create (ceil_log2 (num_ssa_names), vrp_hash, vrp_eq,
388 avail_exprs_stack = VEC_alloc (tree, heap, 20);
389 const_and_copies_stack = VEC_alloc (tree, heap, 20);
390 nonzero_vars_stack = VEC_alloc (tree, heap, 20);
391 vrp_variables_stack = VEC_alloc (tree, heap, 20);
392 stmts_to_rescan = VEC_alloc (tree, heap, 20);
393 nonzero_vars = BITMAP_ALLOC (NULL);
394 threaded_blocks = BITMAP_ALLOC (NULL);
395 need_eh_cleanup = BITMAP_ALLOC (NULL);
397 /* Setup callbacks for the generic dominator tree walker. */
398 walk_data.walk_stmts_backward = false;
399 walk_data.dom_direction = CDI_DOMINATORS;
400 walk_data.initialize_block_local_data = NULL;
401 walk_data.before_dom_children_before_stmts = dom_opt_initialize_block;
402 walk_data.before_dom_children_walk_stmts = optimize_stmt;
403 walk_data.before_dom_children_after_stmts = propagate_to_outgoing_edges;
404 walk_data.after_dom_children_before_stmts = NULL;
405 walk_data.after_dom_children_walk_stmts = NULL;
406 walk_data.after_dom_children_after_stmts = dom_opt_finalize_block;
407 /* Right now we only attach a dummy COND_EXPR to the global data pointer.
408 When we attach more stuff we'll need to fill this out with a real
410 walk_data.global_data = NULL;
411 walk_data.block_local_data_size = 0;
412 walk_data.interesting_blocks = NULL;
414 /* Now initialize the dominator walker. */
415 init_walk_dominator_tree (&walk_data);
417 calculate_dominance_info (CDI_DOMINATORS);
419 /* We need to know which edges exit loops so that we can
420 aggressively thread through loop headers to an exit
422 flow_loops_find (&loops_info);
423 mark_loop_exit_edges (&loops_info);
424 flow_loops_free (&loops_info);
426 /* Clean up the CFG so that any forwarder blocks created by loop
427 canonicalization are removed. */
429 calculate_dominance_info (CDI_DOMINATORS);
431 /* If we prove certain blocks are unreachable, then we want to
432 repeat the dominator optimization process as PHI nodes may
433 have turned into copies which allows better propagation of
434 values. So we repeat until we do not identify any new unreachable
438 /* Optimize the dominator tree. */
441 /* We need accurate information regarding back edges in the CFG
442 for jump threading. */
443 mark_dfs_back_edges ();
445 /* Recursively walk the dominator tree optimizing statements. */
446 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
449 block_stmt_iterator bsi;
453 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
455 update_stmt_if_modified (bsi_stmt (bsi));
460 /* If we exposed any new variables, go ahead and put them into
461 SSA form now, before we handle jump threading. This simplifies
462 interactions between rewriting of _DECL nodes into SSA form
463 and rewriting SSA_NAME nodes into SSA form after block
464 duplication and CFG manipulation. */
465 update_ssa (TODO_update_ssa);
467 free_all_edge_infos ();
469 /* Thread jumps, creating duplicate blocks as needed. */
470 cfg_altered |= thread_through_all_blocks (threaded_blocks);
472 /* Removal of statements may make some EH edges dead. Purge
473 such edges from the CFG as needed. */
474 if (!bitmap_empty_p (need_eh_cleanup))
476 cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
477 bitmap_zero (need_eh_cleanup);
481 free_dominance_info (CDI_DOMINATORS);
483 /* Only iterate if we threaded jumps AND the CFG cleanup did
484 something interesting. Other cases generate far fewer
485 optimization opportunities and thus are not worth another
486 full DOM iteration. */
487 cfg_altered &= cleanup_tree_cfg ();
489 if (rediscover_loops_after_threading)
491 /* Rerun basic loop analysis to discover any newly
492 created loops and update the set of exit edges. */
493 rediscover_loops_after_threading = false;
494 flow_loops_find (&loops_info);
495 mark_loop_exit_edges (&loops_info);
496 flow_loops_free (&loops_info);
498 /* Remove any forwarder blocks inserted by loop
499 header canonicalization. */
503 calculate_dominance_info (CDI_DOMINATORS);
505 update_ssa (TODO_update_ssa);
507 /* Reinitialize the various tables. */
508 bitmap_clear (nonzero_vars);
509 bitmap_clear (threaded_blocks);
510 htab_empty (avail_exprs);
511 htab_empty (vrp_data);
513 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
515 This must be done before we iterate as we might have a
516 reference to an SSA_NAME which was removed by the call to
519 Long term we will be able to let everything in SSA_NAME_VALUE
520 persist. However, for now, we know this is the safe thing to do. */
521 for (i = 0; i < num_ssa_names; i++)
523 tree name = ssa_name (i);
529 value = SSA_NAME_VALUE (name);
530 if (value && !is_gimple_min_invariant (value))
531 SSA_NAME_VALUE (name) = NULL;
534 opt_stats.num_iterations++;
536 while (optimize > 1 && cfg_altered);
538 /* Debugging dumps. */
539 if (dump_file && (dump_flags & TDF_STATS))
540 dump_dominator_optimization_stats (dump_file);
542 /* We emptied the hash table earlier, now delete it completely. */
543 htab_delete (avail_exprs);
544 htab_delete (vrp_data);
546 /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
547 CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
548 of the do-while loop above. */
550 /* And finalize the dominator walker. */
551 fini_walk_dominator_tree (&walk_data);
553 /* Free nonzero_vars. */
554 BITMAP_FREE (nonzero_vars);
555 BITMAP_FREE (threaded_blocks);
556 BITMAP_FREE (need_eh_cleanup);
558 VEC_free (tree, heap, avail_exprs_stack);
559 VEC_free (tree, heap, const_and_copies_stack);
560 VEC_free (tree, heap, nonzero_vars_stack);
561 VEC_free (tree, heap, vrp_variables_stack);
562 VEC_free (tree, heap, stmts_to_rescan);
566 gate_dominator (void)
568 return flag_tree_dom != 0;
571 struct tree_opt_pass pass_dominator =
574 gate_dominator, /* gate */
575 tree_ssa_dominator_optimize, /* execute */
578 0, /* static_pass_number */
579 TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */
580 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
581 0, /* properties_provided */
582 0, /* properties_destroyed */
583 0, /* todo_flags_start */
586 | TODO_verify_ssa, /* todo_flags_finish */
591 /* Given a stmt CONDSTMT containing a COND_EXPR, canonicalize the
592 COND_EXPR into a canonical form. */
595 canonicalize_comparison (tree condstmt)
597 tree cond = COND_EXPR_COND (condstmt);
600 enum tree_code code = TREE_CODE (cond);
602 if (!COMPARISON_CLASS_P (cond))
605 op0 = TREE_OPERAND (cond, 0);
606 op1 = TREE_OPERAND (cond, 1);
608 /* If it would be profitable to swap the operands, then do so to
609 canonicalize the statement, enabling better optimization.
611 By placing canonicalization of such expressions here we
612 transparently keep statements in canonical form, even
613 when the statement is modified. */
614 if (tree_swap_operands_p (op0, op1, false))
616 /* For relationals we need to swap the operands
617 and change the code. */
623 TREE_SET_CODE (cond, swap_tree_comparison (code));
624 swap_tree_operands (condstmt,
625 &TREE_OPERAND (cond, 0),
626 &TREE_OPERAND (cond, 1));
627 /* If one operand was in the operand cache, but the other is
628 not, because it is a constant, this is a case that the
629 internal updating code of swap_tree_operands can't handle
631 if (TREE_CODE_CLASS (TREE_CODE (op0))
632 != TREE_CODE_CLASS (TREE_CODE (op1)))
633 update_stmt (condstmt);
637 /* We are exiting E->src, see if E->dest ends with a conditional
638 jump which has a known value when reached via E.
640 Special care is necessary if E is a back edge in the CFG as we
641 will have already recorded equivalences for E->dest into our
642 various tables, including the result of the conditional at
643 the end of E->dest. Threading opportunities are severely
644 limited in that case to avoid short-circuiting the loop
647 Note it is quite common for the first block inside a loop to
648 end with a conditional which is either always true or always
649 false when reached via the loop backedge. Thus we do not want
650 to blindly disable threading across a loop backedge. */
653 thread_across_edge (struct dom_walk_data *walk_data, edge e)
655 block_stmt_iterator bsi;
662 /* If E->dest does not end with a conditional, then there is
664 bsi = bsi_last (e->dest);
667 || (TREE_CODE (bsi_stmt (bsi)) != COND_EXPR
668 && TREE_CODE (bsi_stmt (bsi)) != GOTO_EXPR
669 && TREE_CODE (bsi_stmt (bsi)) != SWITCH_EXPR))
672 /* The basic idea here is to use whatever knowledge we have
673 from our dominator walk to simplify statements in E->dest,
674 with the ultimate goal being to simplify the conditional
675 at the end of E->dest.
677 Note that we must undo any changes we make to the underlying
678 statements as the simplifications we are making are control
679 flow sensitive (ie, the simplifications are valid when we
680 traverse E, but may not be valid on other paths to E->dest. */
682 /* Each PHI creates a temporary equivalence, record them. Again
683 these are context sensitive equivalences and will be removed
685 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
687 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
688 tree dst = PHI_RESULT (phi);
690 /* Do not include virtual PHIs in our statement count as
691 they never generate code. */
692 if (is_gimple_reg (dst))
695 /* If the desired argument is not the same as this PHI's result
696 and it is set by a PHI in E->dest, then we can not thread
699 && TREE_CODE (src) == SSA_NAME
700 && TREE_CODE (SSA_NAME_DEF_STMT (src)) == PHI_NODE
701 && bb_for_stmt (SSA_NAME_DEF_STMT (src)) == e->dest)
704 record_const_or_copy (dst, src);
707 /* Try to simplify each statement in E->dest, ultimately leading to
708 a simplification of the COND_EXPR at the end of E->dest.
710 We might consider marking just those statements which ultimately
711 feed the COND_EXPR. It's not clear if the overhead of bookkeeping
712 would be recovered by trying to simplify fewer statements.
714 If we are able to simplify a statement into the form
715 SSA_NAME = (SSA_NAME | gimple invariant), then we can record
716 a context sensitive equivalency which may help us simplify
717 later statements in E->dest.
719 Failure to simplify into the form above merely means that the
720 statement provides no equivalences to help simplify later
721 statements. This does not prevent threading through E->dest. */
722 max_stmt_count = PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS);
723 for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
725 tree cached_lhs = NULL;
727 stmt = bsi_stmt (bsi);
729 /* Ignore empty statements and labels. */
730 if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
733 /* If duplicating this block is going to cause too much code
734 expansion, then do not thread through this block. */
736 if (stmt_count > max_stmt_count)
739 /* Safely handle threading across loop backedges. This is
740 over conservative, but still allows us to capture the
741 majority of the cases where we can thread across a loop
743 if ((e->flags & EDGE_DFS_BACK) != 0
744 && TREE_CODE (stmt) != COND_EXPR
745 && TREE_CODE (stmt) != SWITCH_EXPR)
748 /* If the statement has volatile operands, then we assume we
749 can not thread through this block. This is overly
750 conservative in some ways. */
751 if (TREE_CODE (stmt) == ASM_EXPR && ASM_VOLATILE_P (stmt))
754 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
755 value, then do not try to simplify this statement as it will
756 not simplify in any way that is helpful for jump threading. */
757 if (TREE_CODE (stmt) != MODIFY_EXPR
758 || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
761 /* At this point we have a statement which assigns an RHS to an
762 SSA_VAR on the LHS. We want to try and simplify this statement
763 to expose more context sensitive equivalences which in turn may
764 allow us to simplify the condition at the end of the loop. */
765 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
766 cached_lhs = TREE_OPERAND (stmt, 1);
769 /* Copy the operands. */
770 tree *copy, pre_fold_expr;
773 unsigned int num, i = 0;
775 num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE));
776 copy = xcalloc (num, sizeof (tree));
778 /* Make a copy of the uses & vuses into USES_COPY, then cprop into
780 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
783 tree use = USE_FROM_PTR (use_p);
786 if (TREE_CODE (use) == SSA_NAME)
787 tmp = SSA_NAME_VALUE (use);
788 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
789 SET_USE (use_p, tmp);
792 /* Try to fold/lookup the new expression. Inserting the
793 expression into the hash table is unlikely to help
794 Sadly, we have to handle conditional assignments specially
795 here, because fold expects all the operands of an expression
796 to be folded before the expression itself is folded, but we
797 can't just substitute the folded condition here. */
798 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == COND_EXPR)
800 tree cond = COND_EXPR_COND (TREE_OPERAND (stmt, 1));
802 if (cond == boolean_true_node)
803 pre_fold_expr = COND_EXPR_THEN (TREE_OPERAND (stmt, 1));
804 else if (cond == boolean_false_node)
805 pre_fold_expr = COND_EXPR_ELSE (TREE_OPERAND (stmt, 1));
807 pre_fold_expr = TREE_OPERAND (stmt, 1);
810 pre_fold_expr = TREE_OPERAND (stmt, 1);
814 cached_lhs = fold (pre_fold_expr);
815 if (TREE_CODE (cached_lhs) != SSA_NAME
816 && !is_gimple_min_invariant (cached_lhs))
817 cached_lhs = lookup_avail_expr (stmt, false);
820 /* Restore the statement's original uses/defs. */
822 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
823 SET_USE (use_p, copy[i++]);
828 /* Record the context sensitive equivalence if we were able
829 to simplify this statement. */
831 && (TREE_CODE (cached_lhs) == SSA_NAME
832 || is_gimple_min_invariant (cached_lhs)))
833 record_const_or_copy (TREE_OPERAND (stmt, 0), cached_lhs);
836 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
839 && (TREE_CODE (stmt) == COND_EXPR
840 || TREE_CODE (stmt) == GOTO_EXPR
841 || TREE_CODE (stmt) == SWITCH_EXPR))
843 tree cond, cached_lhs;
845 /* Now temporarily cprop the operands and try to find the resulting
846 expression in the hash tables. */
847 if (TREE_CODE (stmt) == COND_EXPR)
849 canonicalize_comparison (stmt);
850 cond = COND_EXPR_COND (stmt);
852 else if (TREE_CODE (stmt) == GOTO_EXPR)
853 cond = GOTO_DESTINATION (stmt);
855 cond = SWITCH_COND (stmt);
857 if (COMPARISON_CLASS_P (cond))
859 tree dummy_cond, op0, op1;
860 enum tree_code cond_code;
862 op0 = TREE_OPERAND (cond, 0);
863 op1 = TREE_OPERAND (cond, 1);
864 cond_code = TREE_CODE (cond);
866 /* Get the current value of both operands. */
867 if (TREE_CODE (op0) == SSA_NAME)
869 tree tmp = SSA_NAME_VALUE (op0);
870 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
874 if (TREE_CODE (op1) == SSA_NAME)
876 tree tmp = SSA_NAME_VALUE (op1);
877 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
881 /* Stuff the operator and operands into our dummy conditional
882 expression, creating the dummy conditional if necessary. */
883 dummy_cond = walk_data->global_data;
886 dummy_cond = build2 (cond_code, boolean_type_node, op0, op1);
887 dummy_cond = build3 (COND_EXPR, void_type_node,
888 dummy_cond, NULL_TREE, NULL_TREE);
889 walk_data->global_data = dummy_cond;
893 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
894 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
895 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
898 /* If the conditional folds to an invariant, then we are done,
899 otherwise look it up in the hash tables. */
900 cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
901 if (! is_gimple_min_invariant (cached_lhs))
903 cached_lhs = lookup_avail_expr (dummy_cond, false);
904 if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
905 cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
910 /* We can have conditionals which just test the state of a
911 variable rather than use a relational operator. These are
912 simpler to handle. */
913 else if (TREE_CODE (cond) == SSA_NAME)
916 cached_lhs = SSA_NAME_VALUE (cached_lhs);
917 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
921 cached_lhs = lookup_avail_expr (stmt, false);
925 edge taken_edge = find_taken_edge (e->dest, cached_lhs);
926 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
931 /* If we have a known destination for the conditional, then
932 we can perform this optimization, which saves at least one
933 conditional jump each time it applies since we get to
934 bypass the conditional at our original destination. */
937 struct edge_info *edge_info;
942 edge_info = allocate_edge_info (e);
943 edge_info->redirection_target = taken_edge;
944 bitmap_set_bit (threaded_blocks, e->dest->index);
951 /* Initialize local stacks for this optimizer and record equivalences
952 upon entry to BB. Equivalences can come from the edge traversed to
953 reach BB or they may come from PHI nodes at the start of BB. */
956 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
959 if (dump_file && (dump_flags & TDF_DETAILS))
960 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
962 /* Push a marker on the stacks of local information so that we know how
963 far to unwind when we finalize this block. */
964 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
965 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
966 VEC_safe_push (tree, heap, nonzero_vars_stack, NULL_TREE);
967 VEC_safe_push (tree, heap, vrp_variables_stack, NULL_TREE);
969 record_equivalences_from_incoming_edge (bb);
971 /* PHI nodes can create equivalences too. */
972 record_equivalences_from_phis (bb);
975 /* Given an expression EXPR (a relational expression or a statement),
976 initialize the hash table element pointed to by ELEMENT. */
979 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
981 /* Hash table elements may be based on conditional expressions or statements.
983 For the former case, we have no annotation and we want to hash the
984 conditional expression. In the latter case we have an annotation and
985 we want to record the expression the statement evaluates. */
986 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
988 element->stmt = NULL;
991 else if (TREE_CODE (expr) == COND_EXPR)
993 element->stmt = expr;
994 element->rhs = COND_EXPR_COND (expr);
996 else if (TREE_CODE (expr) == SWITCH_EXPR)
998 element->stmt = expr;
999 element->rhs = SWITCH_COND (expr);
1001 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
1003 element->stmt = expr;
1004 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
1006 else if (TREE_CODE (expr) == GOTO_EXPR)
1008 element->stmt = expr;
1009 element->rhs = GOTO_DESTINATION (expr);
1013 element->stmt = expr;
1014 element->rhs = TREE_OPERAND (expr, 1);
1018 element->hash = avail_expr_hash (element);
1021 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
1022 LIMIT entries left in LOCALs. */
1025 remove_local_expressions_from_table (void)
1027 /* Remove all the expressions made available in this block. */
1028 while (VEC_length (tree, avail_exprs_stack) > 0)
1030 struct expr_hash_elt element;
1031 tree expr = VEC_pop (tree, avail_exprs_stack);
1033 if (expr == NULL_TREE)
1036 initialize_hash_element (expr, NULL, &element);
1037 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
1041 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
1042 state, stopping when there are LIMIT entries left in LOCALs. */
1045 restore_nonzero_vars_to_original_value (void)
1047 while (VEC_length (tree, nonzero_vars_stack) > 0)
1049 tree name = VEC_pop (tree, nonzero_vars_stack);
1054 bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
1058 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
1059 CONST_AND_COPIES to its original state, stopping when we hit a
1063 restore_vars_to_original_value (void)
1065 while (VEC_length (tree, const_and_copies_stack) > 0)
1067 tree prev_value, dest;
1069 dest = VEC_pop (tree, const_and_copies_stack);
1074 prev_value = VEC_pop (tree, const_and_copies_stack);
1075 SSA_NAME_VALUE (dest) = prev_value;
1079 /* We have finished processing the dominator children of BB, perform
1080 any finalization actions in preparation for leaving this node in
1081 the dominator tree. */
1084 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
1088 /* If we have an outgoing edge to a block with multiple incoming and
1089 outgoing edges, then we may be able to thread the edge. ie, we
1090 may be able to statically determine which of the outgoing edges
1091 will be traversed when the incoming edge from BB is traversed. */
1092 if (single_succ_p (bb)
1093 && (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0
1094 && !single_pred_p (single_succ (bb))
1095 && !single_succ_p (single_succ (bb)))
1098 thread_across_edge (walk_data, single_succ_edge (bb));
1100 else if ((last = last_stmt (bb))
1101 && TREE_CODE (last) == COND_EXPR
1102 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
1103 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
1104 && EDGE_COUNT (bb->succs) == 2
1105 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
1106 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
1108 edge true_edge, false_edge;
1110 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
1112 /* Only try to thread the edge if it reaches a target block with
1113 more than one predecessor and more than one successor. */
1114 if (!single_pred_p (true_edge->dest) && !single_succ_p (true_edge->dest))
1116 struct edge_info *edge_info;
1119 /* Push a marker onto the available expression stack so that we
1120 unwind any expressions related to the TRUE arm before processing
1121 the false arm below. */
1122 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
1123 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
1125 edge_info = true_edge->aux;
1127 /* If we have info associated with this edge, record it into
1128 our equivalency tables. */
1131 tree *cond_equivalences = edge_info->cond_equivalences;
1132 tree lhs = edge_info->lhs;
1133 tree rhs = edge_info->rhs;
1135 /* If we have a simple NAME = VALUE equivalency record it. */
1136 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1137 record_const_or_copy (lhs, rhs);
1139 /* If we have 0 = COND or 1 = COND equivalences, record them
1140 into our expression hash tables. */
1141 if (cond_equivalences)
1142 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1144 tree expr = cond_equivalences[i];
1145 tree value = cond_equivalences[i + 1];
1147 record_cond (expr, value);
1151 /* Now thread the edge. */
1152 thread_across_edge (walk_data, true_edge);
1154 /* And restore the various tables to their state before
1155 we threaded this edge. */
1156 remove_local_expressions_from_table ();
1157 restore_vars_to_original_value ();
1160 /* Similarly for the ELSE arm. */
1161 if (!single_pred_p (false_edge->dest) && !single_succ_p (false_edge->dest))
1163 struct edge_info *edge_info;
1166 edge_info = false_edge->aux;
1168 /* If we have info associated with this edge, record it into
1169 our equivalency tables. */
1172 tree *cond_equivalences = edge_info->cond_equivalences;
1173 tree lhs = edge_info->lhs;
1174 tree rhs = edge_info->rhs;
1176 /* If we have a simple NAME = VALUE equivalency record it. */
1177 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1178 record_const_or_copy (lhs, rhs);
1180 /* If we have 0 = COND or 1 = COND equivalences, record them
1181 into our expression hash tables. */
1182 if (cond_equivalences)
1183 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1185 tree expr = cond_equivalences[i];
1186 tree value = cond_equivalences[i + 1];
1188 record_cond (expr, value);
1192 thread_across_edge (walk_data, false_edge);
1194 /* No need to remove local expressions from our tables
1195 or restore vars to their original value as that will
1196 be done immediately below. */
1200 remove_local_expressions_from_table ();
1201 restore_nonzero_vars_to_original_value ();
1202 restore_vars_to_original_value ();
1204 /* Remove VRP records associated with this basic block. They are no
1207 To be efficient, we note which variables have had their values
1208 constrained in this block. So walk over each variable in the
1209 VRP_VARIABLEs array. */
1210 while (VEC_length (tree, vrp_variables_stack) > 0)
1212 tree var = VEC_pop (tree, vrp_variables_stack);
1213 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1216 /* Each variable has a stack of value range records. We want to
1217 invalidate those associated with our basic block. So we walk
1218 the array backwards popping off records associated with our
1219 block. Once we hit a record not associated with our block
1221 VEC(vrp_element_p,heap) **var_vrp_records;
1226 vrp_hash_elt.var = var;
1227 vrp_hash_elt.records = NULL;
1229 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1231 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1232 var_vrp_records = &vrp_hash_elt_p->records;
1234 while (VEC_length (vrp_element_p, *var_vrp_records) > 0)
1236 struct vrp_element *element
1237 = VEC_last (vrp_element_p, *var_vrp_records);
1239 if (element->bb != bb)
1242 VEC_pop (vrp_element_p, *var_vrp_records);
1246 /* If we queued any statements to rescan in this block, then
1247 go ahead and rescan them now. */
1248 while (VEC_length (tree, stmts_to_rescan) > 0)
1250 tree stmt = VEC_last (tree, stmts_to_rescan);
1251 basic_block stmt_bb = bb_for_stmt (stmt);
1256 VEC_pop (tree, stmts_to_rescan);
1257 mark_new_vars_to_rename (stmt);
1261 /* PHI nodes can create equivalences too.
1263 Ignoring any alternatives which are the same as the result, if
1264 all the alternatives are equal, then the PHI node creates an
1267 Additionally, if all the PHI alternatives are known to have a nonzero
1268 value, then the result of this PHI is known to have a nonzero value,
1269 even if we do not know its exact value. */
1272 record_equivalences_from_phis (basic_block bb)
1276 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1278 tree lhs = PHI_RESULT (phi);
1282 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1284 tree t = PHI_ARG_DEF (phi, i);
1286 /* Ignore alternatives which are the same as our LHS. Since
1287 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1288 can simply compare pointers. */
1292 /* If we have not processed an alternative yet, then set
1293 RHS to this alternative. */
1296 /* If we have processed an alternative (stored in RHS), then
1297 see if it is equal to this one. If it isn't, then stop
1299 else if (! operand_equal_for_phi_arg_p (rhs, t))
1303 /* If we had no interesting alternatives, then all the RHS alternatives
1304 must have been the same as LHS. */
1308 /* If we managed to iterate through each PHI alternative without
1309 breaking out of the loop, then we have a PHI which may create
1310 a useful equivalence. We do not need to record unwind data for
1311 this, since this is a true assignment and not an equivalence
1312 inferred from a comparison. All uses of this ssa name are dominated
1313 by this assignment, so unwinding just costs time and space. */
1314 if (i == PHI_NUM_ARGS (phi)
1315 && may_propagate_copy (lhs, rhs))
1316 SSA_NAME_VALUE (lhs) = rhs;
1318 /* Now see if we know anything about the nonzero property for the
1319 result of this PHI. */
1320 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1322 if (!PHI_ARG_NONZERO (phi, i))
1326 if (i == PHI_NUM_ARGS (phi))
1327 bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
1331 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1332 return that edge. Otherwise return NULL. */
1334 single_incoming_edge_ignoring_loop_edges (basic_block bb)
1340 FOR_EACH_EDGE (e, ei, bb->preds)
1342 /* A loop back edge can be identified by the destination of
1343 the edge dominating the source of the edge. */
1344 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
1347 /* If we have already seen a non-loop edge, then we must have
1348 multiple incoming non-loop edges and thus we return NULL. */
1352 /* This is the first non-loop incoming edge we have found. Record
1360 /* Record any equivalences created by the incoming edge to BB. If BB
1361 has more than one incoming edge, then no equivalence is created. */
1364 record_equivalences_from_incoming_edge (basic_block bb)
1368 struct edge_info *edge_info;
1370 /* If our parent block ended with a control statement, then we may be
1371 able to record some equivalences based on which outgoing edge from
1372 the parent was followed. */
1373 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
1375 e = single_incoming_edge_ignoring_loop_edges (bb);
1377 /* If we had a single incoming edge from our parent block, then enter
1378 any data associated with the edge into our tables. */
1379 if (e && e->src == parent)
1387 tree lhs = edge_info->lhs;
1388 tree rhs = edge_info->rhs;
1389 tree *cond_equivalences = edge_info->cond_equivalences;
1392 record_equality (lhs, rhs);
1394 if (cond_equivalences)
1396 bool recorded_range = false;
1397 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1399 tree expr = cond_equivalences[i];
1400 tree value = cond_equivalences[i + 1];
1402 record_cond (expr, value);
1404 /* For the first true equivalence, record range
1405 information. We only do this for the first
1406 true equivalence as it should dominate any
1407 later true equivalences. */
1408 if (! recorded_range
1409 && COMPARISON_CLASS_P (expr)
1410 && value == boolean_true_node
1411 && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
1413 record_range (expr, bb);
1414 recorded_range = true;
1422 /* Dump SSA statistics on FILE. */
1425 dump_dominator_optimization_stats (FILE *file)
1429 fprintf (file, "Total number of statements: %6ld\n\n",
1430 opt_stats.num_stmts);
1431 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
1432 opt_stats.num_exprs_considered);
1434 n_exprs = opt_stats.num_exprs_considered;
1438 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1439 opt_stats.num_re, PERCENT (opt_stats.num_re,
1441 fprintf (file, " Constants propagated: %6ld\n",
1442 opt_stats.num_const_prop);
1443 fprintf (file, " Copies propagated: %6ld\n",
1444 opt_stats.num_copy_prop);
1446 fprintf (file, "\nTotal number of DOM iterations: %6ld\n",
1447 opt_stats.num_iterations);
1449 fprintf (file, "\nHash table statistics:\n");
1451 fprintf (file, " avail_exprs: ");
1452 htab_statistics (file, avail_exprs);
1456 /* Dump SSA statistics on stderr. */
1459 debug_dominator_optimization_stats (void)
1461 dump_dominator_optimization_stats (stderr);
1465 /* Dump statistics for the hash table HTAB. */
1468 htab_statistics (FILE *file, htab_t htab)
1470 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
1471 (long) htab_size (htab),
1472 (long) htab_elements (htab),
1473 htab_collisions (htab));
1476 /* Record the fact that VAR has a nonzero value, though we may not know
1477 its exact value. Note that if VAR is already known to have a nonzero
1478 value, then we do nothing. */
1481 record_var_is_nonzero (tree var)
1483 int indx = SSA_NAME_VERSION (var);
1485 if (bitmap_bit_p (nonzero_vars, indx))
1488 /* Mark it in the global table. */
1489 bitmap_set_bit (nonzero_vars, indx);
1491 /* Record this SSA_NAME so that we can reset the global table
1492 when we leave this block. */
1493 VEC_safe_push (tree, heap, nonzero_vars_stack, var);
1496 /* Enter a statement into the true/false expression hash table indicating
1497 that the condition COND has the value VALUE. */
1500 record_cond (tree cond, tree value)
1502 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
1505 initialize_hash_element (cond, value, element);
1507 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
1508 element->hash, INSERT);
1511 *slot = (void *) element;
1512 VEC_safe_push (tree, heap, avail_exprs_stack, cond);
1518 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1519 the new conditional into *p, then store a boolean_true_node
1523 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
1525 *p = build2 (new_code, boolean_type_node, op0, op1);
1527 *p = boolean_true_node;
1530 /* Record that COND is true and INVERTED is false into the edge information
1531 structure. Also record that any conditions dominated by COND are true
1534 For example, if a < b is true, then a <= b must also be true. */
1537 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
1541 if (!COMPARISON_CLASS_P (cond))
1544 op0 = TREE_OPERAND (cond, 0);
1545 op1 = TREE_OPERAND (cond, 1);
1547 switch (TREE_CODE (cond))
1551 edge_info->max_cond_equivalences = 12;
1552 edge_info->cond_equivalences = xmalloc (12 * sizeof (tree));
1553 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
1554 ? LE_EXPR : GE_EXPR),
1555 op0, op1, &edge_info->cond_equivalences[4]);
1556 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1557 &edge_info->cond_equivalences[6]);
1558 build_and_record_new_cond (NE_EXPR, op0, op1,
1559 &edge_info->cond_equivalences[8]);
1560 build_and_record_new_cond (LTGT_EXPR, op0, op1,
1561 &edge_info->cond_equivalences[10]);
1566 edge_info->max_cond_equivalences = 6;
1567 edge_info->cond_equivalences = xmalloc (6 * sizeof (tree));
1568 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1569 &edge_info->cond_equivalences[4]);
1573 edge_info->max_cond_equivalences = 10;
1574 edge_info->cond_equivalences = xmalloc (10 * sizeof (tree));
1575 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1576 &edge_info->cond_equivalences[4]);
1577 build_and_record_new_cond (LE_EXPR, op0, op1,
1578 &edge_info->cond_equivalences[6]);
1579 build_and_record_new_cond (GE_EXPR, op0, op1,
1580 &edge_info->cond_equivalences[8]);
1583 case UNORDERED_EXPR:
1584 edge_info->max_cond_equivalences = 16;
1585 edge_info->cond_equivalences = xmalloc (16 * sizeof (tree));
1586 build_and_record_new_cond (NE_EXPR, op0, op1,
1587 &edge_info->cond_equivalences[4]);
1588 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1589 &edge_info->cond_equivalences[6]);
1590 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1591 &edge_info->cond_equivalences[8]);
1592 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
1593 &edge_info->cond_equivalences[10]);
1594 build_and_record_new_cond (UNLT_EXPR, op0, op1,
1595 &edge_info->cond_equivalences[12]);
1596 build_and_record_new_cond (UNGT_EXPR, op0, op1,
1597 &edge_info->cond_equivalences[14]);
1602 edge_info->max_cond_equivalences = 8;
1603 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1604 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1605 ? UNLE_EXPR : UNGE_EXPR),
1606 op0, op1, &edge_info->cond_equivalences[4]);
1607 build_and_record_new_cond (NE_EXPR, op0, op1,
1608 &edge_info->cond_equivalences[6]);
1612 edge_info->max_cond_equivalences = 8;
1613 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1614 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1615 &edge_info->cond_equivalences[4]);
1616 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1617 &edge_info->cond_equivalences[6]);
1621 edge_info->max_cond_equivalences = 8;
1622 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1623 build_and_record_new_cond (NE_EXPR, op0, op1,
1624 &edge_info->cond_equivalences[4]);
1625 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1626 &edge_info->cond_equivalences[6]);
1630 edge_info->max_cond_equivalences = 4;
1631 edge_info->cond_equivalences = xmalloc (4 * sizeof (tree));
1635 /* Now store the original true and false conditions into the first
1637 edge_info->cond_equivalences[0] = cond;
1638 edge_info->cond_equivalences[1] = boolean_true_node;
1639 edge_info->cond_equivalences[2] = inverted;
1640 edge_info->cond_equivalences[3] = boolean_false_node;
1643 /* A helper function for record_const_or_copy and record_equality.
1644 Do the work of recording the value and undo info. */
1647 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1649 SSA_NAME_VALUE (x) = y;
1651 VEC_reserve (tree, heap, const_and_copies_stack, 2);
1652 VEC_quick_push (tree, const_and_copies_stack, prev_x);
1653 VEC_quick_push (tree, const_and_copies_stack, x);
1657 /* Return the loop depth of the basic block of the defining statement of X.
1658 This number should not be treated as absolutely correct because the loop
1659 information may not be completely up-to-date when dom runs. However, it
1660 will be relatively correct, and as more passes are taught to keep loop info
1661 up to date, the result will become more and more accurate. */
1664 loop_depth_of_name (tree x)
1669 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1670 if (TREE_CODE (x) != SSA_NAME)
1673 /* Otherwise return the loop depth of the defining statement's bb.
1674 Note that there may not actually be a bb for this statement, if the
1675 ssa_name is live on entry. */
1676 defstmt = SSA_NAME_DEF_STMT (x);
1677 defbb = bb_for_stmt (defstmt);
1681 return defbb->loop_depth;
1685 /* Record that X is equal to Y in const_and_copies. Record undo
1686 information in the block-local vector. */
1689 record_const_or_copy (tree x, tree y)
1691 tree prev_x = SSA_NAME_VALUE (x);
1693 if (TREE_CODE (y) == SSA_NAME)
1695 tree tmp = SSA_NAME_VALUE (y);
1700 record_const_or_copy_1 (x, y, prev_x);
1703 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1704 This constrains the cases in which we may treat this as assignment. */
1707 record_equality (tree x, tree y)
1709 tree prev_x = NULL, prev_y = NULL;
1711 if (TREE_CODE (x) == SSA_NAME)
1712 prev_x = SSA_NAME_VALUE (x);
1713 if (TREE_CODE (y) == SSA_NAME)
1714 prev_y = SSA_NAME_VALUE (y);
1716 /* If one of the previous values is invariant, or invariant in more loops
1717 (by depth), then use that.
1718 Otherwise it doesn't matter which value we choose, just so
1719 long as we canonicalize on one value. */
1720 if (TREE_INVARIANT (y))
1722 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1723 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1724 else if (prev_x && TREE_INVARIANT (prev_x))
1725 x = y, y = prev_x, prev_x = prev_y;
1726 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1729 /* After the swapping, we must have one SSA_NAME. */
1730 if (TREE_CODE (x) != SSA_NAME)
1733 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1734 variable compared against zero. If we're honoring signed zeros,
1735 then we cannot record this value unless we know that the value is
1737 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1738 && (TREE_CODE (y) != REAL_CST
1739 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1742 record_const_or_copy_1 (x, y, prev_x);
1745 /* Return true, if it is ok to do folding of an associative expression.
1746 EXP is the tree for the associative expression. */
1749 unsafe_associative_fp_binop (tree exp)
1751 enum tree_code code = TREE_CODE (exp);
1752 return !(!flag_unsafe_math_optimizations
1753 && (code == MULT_EXPR || code == PLUS_EXPR
1754 || code == MINUS_EXPR)
1755 && FLOAT_TYPE_P (TREE_TYPE (exp)));
1758 /* Returns true when STMT is a simple iv increment. It detects the
1759 following situation:
1761 i_1 = phi (..., i_2)
1762 i_2 = i_1 +/- ... */
1765 simple_iv_increment_p (tree stmt)
1767 tree lhs, rhs, preinc, phi;
1770 if (TREE_CODE (stmt) != MODIFY_EXPR)
1773 lhs = TREE_OPERAND (stmt, 0);
1774 if (TREE_CODE (lhs) != SSA_NAME)
1777 rhs = TREE_OPERAND (stmt, 1);
1779 if (TREE_CODE (rhs) != PLUS_EXPR
1780 && TREE_CODE (rhs) != MINUS_EXPR)
1783 preinc = TREE_OPERAND (rhs, 0);
1784 if (TREE_CODE (preinc) != SSA_NAME)
1787 phi = SSA_NAME_DEF_STMT (preinc);
1788 if (TREE_CODE (phi) != PHI_NODE)
1791 for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++)
1792 if (PHI_ARG_DEF (phi, i) == lhs)
1798 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1799 hash tables. Try to simplify the RHS using whatever equivalences
1800 we may have recorded.
1802 If we are able to simplify the RHS, then lookup the simplified form in
1803 the hash table and return the result. Otherwise return NULL. */
1806 simplify_rhs_and_lookup_avail_expr (tree stmt, int insert)
1808 tree rhs = TREE_OPERAND (stmt, 1);
1809 enum tree_code rhs_code = TREE_CODE (rhs);
1812 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1813 In which case we can change this statement to be lhs = y.
1814 Which can then be copy propagated.
1816 Similarly for negation. */
1817 if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR)
1818 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME)
1820 /* Get the definition statement for our RHS. */
1821 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1823 /* See if the RHS_DEF_STMT has the same form as our statement. */
1824 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR
1825 && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code)
1827 tree rhs_def_operand;
1829 rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0);
1831 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1832 if (TREE_CODE (rhs_def_operand) == SSA_NAME
1833 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand))
1834 result = update_rhs_and_lookup_avail_expr (stmt,
1840 /* Optimize *"foo" into 'f'. This is done here rather than
1841 in fold to avoid problems with stuff like &*"foo". */
1842 if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF)
1844 tree t = fold_read_from_constant_string (rhs);
1847 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1853 /* COND is a condition of the form:
1855 x == const or x != const
1857 Look back to x's defining statement and see if x is defined as
1861 If const is unchanged if we convert it to type, then we can build
1862 the equivalent expression:
1865 y == const or y != const
1867 Which may allow further optimizations.
1869 Return the equivalent comparison or NULL if no such equivalent comparison
1873 find_equivalent_equality_comparison (tree cond)
1875 tree op0 = TREE_OPERAND (cond, 0);
1876 tree op1 = TREE_OPERAND (cond, 1);
1877 tree def_stmt = SSA_NAME_DEF_STMT (op0);
1879 /* OP0 might have been a parameter, so first make sure it
1880 was defined by a MODIFY_EXPR. */
1881 if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
1883 tree def_rhs = TREE_OPERAND (def_stmt, 1);
1886 /* If either operand to the comparison is a pointer to
1887 a function, then we can not apply this optimization
1888 as some targets require function pointers to be
1889 canonicalized and in this case this optimization would
1890 eliminate a necessary canonicalization. */
1891 if ((POINTER_TYPE_P (TREE_TYPE (op0))
1892 && TREE_CODE (TREE_TYPE (TREE_TYPE (op0))) == FUNCTION_TYPE)
1893 || (POINTER_TYPE_P (TREE_TYPE (op1))
1894 && TREE_CODE (TREE_TYPE (TREE_TYPE (op1))) == FUNCTION_TYPE))
1897 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1898 if ((TREE_CODE (def_rhs) == NOP_EXPR
1899 || TREE_CODE (def_rhs) == CONVERT_EXPR)
1900 && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
1902 tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
1903 tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
1906 if (TYPE_PRECISION (def_rhs_inner_type)
1907 > TYPE_PRECISION (TREE_TYPE (def_rhs)))
1910 /* If the inner type of the conversion is a pointer to
1911 a function, then we can not apply this optimization
1912 as some targets require function pointers to be
1913 canonicalized. This optimization would result in
1914 canonicalization of the pointer when it was not originally
1916 if (POINTER_TYPE_P (def_rhs_inner_type)
1917 && TREE_CODE (TREE_TYPE (def_rhs_inner_type)) == FUNCTION_TYPE)
1920 /* What we want to prove is that if we convert OP1 to
1921 the type of the object inside the NOP_EXPR that the
1922 result is still equivalent to SRC.
1924 If that is true, the build and return new equivalent
1925 condition which uses the source of the typecast and the
1926 new constant (which has only changed its type). */
1927 new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
1928 new = local_fold (new);
1929 if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
1930 return build2 (TREE_CODE (cond), TREE_TYPE (cond),
1931 def_rhs_inner, new);
1937 /* STMT is a COND_EXPR for which we could not trivially determine its
1938 result. This routine attempts to find equivalent forms of the
1939 condition which we may be able to optimize better. It also
1940 uses simple value range propagation to optimize conditionals. */
1943 simplify_cond_and_lookup_avail_expr (tree stmt,
1947 tree cond = COND_EXPR_COND (stmt);
1949 if (COMPARISON_CLASS_P (cond))
1951 tree op0 = TREE_OPERAND (cond, 0);
1952 tree op1 = TREE_OPERAND (cond, 1);
1954 if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
1957 tree low, high, cond_low, cond_high;
1958 int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
1959 VEC(vrp_element_p,heap) **vrp_records;
1960 struct vrp_element *element;
1961 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1964 /* First see if we have test of an SSA_NAME against a constant
1965 where the SSA_NAME is defined by an earlier typecast which
1966 is irrelevant when performing tests against the given
1968 if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
1970 tree new_cond = find_equivalent_equality_comparison (cond);
1974 /* Update the statement to use the new equivalent
1976 COND_EXPR_COND (stmt) = new_cond;
1978 /* If this is not a real stmt, ann will be NULL and we
1979 avoid processing the operands. */
1981 mark_stmt_modified (stmt);
1983 /* Lookup the condition and return its known value if it
1985 new_cond = lookup_avail_expr (stmt, insert);
1989 /* The operands have changed, so update op0 and op1. */
1990 op0 = TREE_OPERAND (cond, 0);
1991 op1 = TREE_OPERAND (cond, 1);
1995 /* Consult the value range records for this variable (if they exist)
1996 to see if we can eliminate or simplify this conditional.
1998 Note two tests are necessary to determine no records exist.
1999 First we have to see if the virtual array exists, if it
2000 exists, then we have to check its active size.
2002 Also note the vast majority of conditionals are not testing
2003 a variable which has had its range constrained by an earlier
2004 conditional. So this filter avoids a lot of unnecessary work. */
2005 vrp_hash_elt.var = op0;
2006 vrp_hash_elt.records = NULL;
2007 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
2011 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
2012 vrp_records = &vrp_hash_elt_p->records;
2014 limit = VEC_length (vrp_element_p, *vrp_records);
2016 /* If we have no value range records for this variable, or we are
2017 unable to extract a range for this condition, then there is
2020 || ! extract_range_from_cond (cond, &cond_high,
2021 &cond_low, &cond_inverted))
2024 /* We really want to avoid unnecessary computations of range
2025 info. So all ranges are computed lazily; this avoids a
2026 lot of unnecessary work. i.e., we record the conditional,
2027 but do not process how it constrains the variable's
2028 potential values until we know that processing the condition
2031 However, we do not want to have to walk a potentially long
2032 list of ranges, nor do we want to compute a variable's
2033 range more than once for a given path.
2035 Luckily, each time we encounter a conditional that can not
2036 be otherwise optimized we will end up here and we will
2037 compute the necessary range information for the variable
2038 used in this condition.
2040 Thus you can conclude that there will never be more than one
2041 conditional associated with a variable which has not been
2042 processed. So we never need to merge more than one new
2043 conditional into the current range.
2045 These properties also help us avoid unnecessary work. */
2046 element = VEC_last (vrp_element_p, *vrp_records);
2048 if (element->high && element->low)
2050 /* The last element has been processed, so there is no range
2051 merging to do, we can simply use the high/low values
2052 recorded in the last element. */
2054 high = element->high;
2058 tree tmp_high, tmp_low;
2061 /* The last element has not been processed. Process it now.
2062 record_range should ensure for cond inverted is not set.
2063 This call can only fail if cond is x < min or x > max,
2064 which fold should have optimized into false.
2065 If that doesn't happen, just pretend all values are
2067 if (! extract_range_from_cond (element->cond, &tmp_high,
2071 gcc_assert (dummy == 0);
2073 /* If this is the only element, then no merging is necessary,
2074 the high/low values from extract_range_from_cond are all
2083 /* Get the high/low value from the previous element. */
2084 struct vrp_element *prev
2085 = VEC_index (vrp_element_p, *vrp_records, limit - 2);
2089 /* Merge in this element's range with the range from the
2092 The low value for the merged range is the maximum of
2093 the previous low value and the low value of this record.
2095 Similarly the high value for the merged range is the
2096 minimum of the previous high value and the high value of
2098 low = (low && tree_int_cst_compare (low, tmp_low) == 1
2100 high = (high && tree_int_cst_compare (high, tmp_high) == -1
2104 /* And record the computed range. */
2106 element->high = high;
2110 /* After we have constrained this variable's potential values,
2111 we try to determine the result of the given conditional.
2113 To simplify later tests, first determine if the current
2114 low value is the same low value as the conditional.
2115 Similarly for the current high value and the high value
2116 for the conditional. */
2117 lowequal = tree_int_cst_equal (low, cond_low);
2118 highequal = tree_int_cst_equal (high, cond_high);
2120 if (lowequal && highequal)
2121 return (cond_inverted ? boolean_false_node : boolean_true_node);
2123 /* To simplify the overlap/subset tests below we may want
2124 to swap the two ranges so that the larger of the two
2125 ranges occurs "first". */
2127 if (tree_int_cst_compare (low, cond_low) == 1
2129 && tree_int_cst_compare (cond_high, high) == 1))
2142 /* Now determine if there is no overlap in the ranges
2143 or if the second range is a subset of the first range. */
2144 no_overlap = tree_int_cst_lt (high, cond_low);
2145 subset = tree_int_cst_compare (cond_high, high) != 1;
2147 /* If there was no overlap in the ranges, then this conditional
2148 always has a false value (unless we had to invert this
2149 conditional, in which case it always has a true value). */
2151 return (cond_inverted ? boolean_true_node : boolean_false_node);
2153 /* If the current range is a subset of the condition's range,
2154 then this conditional always has a true value (unless we
2155 had to invert this conditional, in which case it always
2156 has a true value). */
2157 if (subset && swapped)
2158 return (cond_inverted ? boolean_false_node : boolean_true_node);
2160 /* We were unable to determine the result of the conditional.
2161 However, we may be able to simplify the conditional. First
2162 merge the ranges in the same manner as range merging above. */
2163 low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
2164 high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
2166 /* If the range has converged to a single point, then turn this
2167 into an equality comparison. */
2168 if (TREE_CODE (cond) != EQ_EXPR
2169 && TREE_CODE (cond) != NE_EXPR
2170 && tree_int_cst_equal (low, high))
2172 TREE_SET_CODE (cond, EQ_EXPR);
2173 TREE_OPERAND (cond, 1) = high;
2180 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2181 result. This routine attempts to find equivalent forms of the
2182 condition which we may be able to optimize better. */
2185 simplify_switch_and_lookup_avail_expr (tree stmt, int insert)
2187 tree cond = SWITCH_COND (stmt);
2190 /* The optimization that we really care about is removing unnecessary
2191 casts. That will let us do much better in propagating the inferred
2192 constant at the switch target. */
2193 if (TREE_CODE (cond) == SSA_NAME)
2195 def = SSA_NAME_DEF_STMT (cond);
2196 if (TREE_CODE (def) == MODIFY_EXPR)
2198 def = TREE_OPERAND (def, 1);
2199 if (TREE_CODE (def) == NOP_EXPR)
2204 def = TREE_OPERAND (def, 0);
2206 #ifdef ENABLE_CHECKING
2207 /* ??? Why was Jeff testing this? We are gimple... */
2208 gcc_assert (is_gimple_val (def));
2211 to = TREE_TYPE (cond);
2212 ti = TREE_TYPE (def);
2214 /* If we have an extension that preserves value, then we
2215 can copy the source value into the switch. */
2217 need_precision = TYPE_PRECISION (ti);
2219 if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti))
2221 else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti))
2222 need_precision += 1;
2223 if (TYPE_PRECISION (to) < need_precision)
2228 SWITCH_COND (stmt) = def;
2229 mark_stmt_modified (stmt);
2231 return lookup_avail_expr (stmt, insert);
2241 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2242 known value for that SSA_NAME (or NULL if no value is known).
2244 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2245 even if we don't know their precise value.
2247 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2248 nodes of the successors of BB. */
2251 cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
2256 FOR_EACH_EDGE (e, ei, bb->succs)
2261 /* If this is an abnormal edge, then we do not want to copy propagate
2262 into the PHI alternative associated with this edge. */
2263 if (e->flags & EDGE_ABNORMAL)
2266 phi = phi_nodes (e->dest);
2271 for ( ; phi; phi = PHI_CHAIN (phi))
2274 use_operand_p orig_p;
2277 /* The alternative may be associated with a constant, so verify
2278 it is an SSA_NAME before doing anything with it. */
2279 orig_p = PHI_ARG_DEF_PTR (phi, indx);
2280 orig = USE_FROM_PTR (orig_p);
2281 if (TREE_CODE (orig) != SSA_NAME)
2284 /* If the alternative is known to have a nonzero value, record
2285 that fact in the PHI node itself for future use. */
2286 if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
2287 PHI_ARG_NONZERO (phi, indx) = true;
2289 /* If we have *ORIG_P in our constant/copy table, then replace
2290 ORIG_P with its value in our constant/copy table. */
2291 new = SSA_NAME_VALUE (orig);
2294 && (TREE_CODE (new) == SSA_NAME
2295 || is_gimple_min_invariant (new))
2296 && may_propagate_copy (orig, new))
2297 propagate_value (orig_p, new);
2302 /* We have finished optimizing BB, record any information implied by
2303 taking a specific outgoing edge from BB. */
2306 record_edge_info (basic_block bb)
2308 block_stmt_iterator bsi = bsi_last (bb);
2309 struct edge_info *edge_info;
2311 if (! bsi_end_p (bsi))
2313 tree stmt = bsi_stmt (bsi);
2315 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
2317 tree cond = SWITCH_COND (stmt);
2319 if (TREE_CODE (cond) == SSA_NAME)
2321 tree labels = SWITCH_LABELS (stmt);
2322 int i, n_labels = TREE_VEC_LENGTH (labels);
2323 tree *info = xcalloc (last_basic_block, sizeof (tree));
2327 for (i = 0; i < n_labels; i++)
2329 tree label = TREE_VEC_ELT (labels, i);
2330 basic_block target_bb = label_to_block (CASE_LABEL (label));
2332 if (CASE_HIGH (label)
2333 || !CASE_LOW (label)
2334 || info[target_bb->index])
2335 info[target_bb->index] = error_mark_node;
2337 info[target_bb->index] = label;
2340 FOR_EACH_EDGE (e, ei, bb->succs)
2342 basic_block target_bb = e->dest;
2343 tree node = info[target_bb->index];
2345 if (node != NULL && node != error_mark_node)
2347 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
2348 edge_info = allocate_edge_info (e);
2349 edge_info->lhs = cond;
2357 /* A COND_EXPR may create equivalences too. */
2358 if (stmt && TREE_CODE (stmt) == COND_EXPR)
2360 tree cond = COND_EXPR_COND (stmt);
2364 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
2366 /* If the conditional is a single variable 'X', record 'X = 1'
2367 for the true edge and 'X = 0' on the false edge. */
2368 if (SSA_VAR_P (cond))
2370 struct edge_info *edge_info;
2372 edge_info = allocate_edge_info (true_edge);
2373 edge_info->lhs = cond;
2374 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
2376 edge_info = allocate_edge_info (false_edge);
2377 edge_info->lhs = cond;
2378 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
2380 /* Equality tests may create one or two equivalences. */
2381 else if (COMPARISON_CLASS_P (cond))
2383 tree op0 = TREE_OPERAND (cond, 0);
2384 tree op1 = TREE_OPERAND (cond, 1);
2386 /* Special case comparing booleans against a constant as we
2387 know the value of OP0 on both arms of the branch. i.e., we
2388 can record an equivalence for OP0 rather than COND. */
2389 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2390 && TREE_CODE (op0) == SSA_NAME
2391 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
2392 && is_gimple_min_invariant (op1))
2394 if (TREE_CODE (cond) == EQ_EXPR)
2396 edge_info = allocate_edge_info (true_edge);
2397 edge_info->lhs = op0;
2398 edge_info->rhs = (integer_zerop (op1)
2399 ? boolean_false_node
2400 : boolean_true_node);
2402 edge_info = allocate_edge_info (false_edge);
2403 edge_info->lhs = op0;
2404 edge_info->rhs = (integer_zerop (op1)
2406 : boolean_false_node);
2410 edge_info = allocate_edge_info (true_edge);
2411 edge_info->lhs = op0;
2412 edge_info->rhs = (integer_zerop (op1)
2414 : boolean_false_node);
2416 edge_info = allocate_edge_info (false_edge);
2417 edge_info->lhs = op0;
2418 edge_info->rhs = (integer_zerop (op1)
2419 ? boolean_false_node
2420 : boolean_true_node);
2424 else if (is_gimple_min_invariant (op0)
2425 && (TREE_CODE (op1) == SSA_NAME
2426 || is_gimple_min_invariant (op1)))
2428 tree inverted = invert_truthvalue (cond);
2429 struct edge_info *edge_info;
2431 edge_info = allocate_edge_info (true_edge);
2432 record_conditions (edge_info, cond, inverted);
2434 if (TREE_CODE (cond) == EQ_EXPR)
2436 edge_info->lhs = op1;
2437 edge_info->rhs = op0;
2440 edge_info = allocate_edge_info (false_edge);
2441 record_conditions (edge_info, inverted, cond);
2443 if (TREE_CODE (cond) == NE_EXPR)
2445 edge_info->lhs = op1;
2446 edge_info->rhs = op0;
2450 else if (TREE_CODE (op0) == SSA_NAME
2451 && (is_gimple_min_invariant (op1)
2452 || TREE_CODE (op1) == SSA_NAME))
2454 tree inverted = invert_truthvalue (cond);
2455 struct edge_info *edge_info;
2457 edge_info = allocate_edge_info (true_edge);
2458 record_conditions (edge_info, cond, inverted);
2460 if (TREE_CODE (cond) == EQ_EXPR)
2462 edge_info->lhs = op0;
2463 edge_info->rhs = op1;
2466 edge_info = allocate_edge_info (false_edge);
2467 record_conditions (edge_info, inverted, cond);
2469 if (TREE_CODE (cond) == NE_EXPR)
2471 edge_info->lhs = op0;
2472 edge_info->rhs = op1;
2477 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2482 /* Propagate information from BB to its outgoing edges.
2484 This can include equivalency information implied by control statements
2485 at the end of BB and const/copy propagation into PHIs in BB's
2486 successor blocks. */
2489 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2492 record_edge_info (bb);
2493 cprop_into_successor_phis (bb, nonzero_vars);
2496 /* Search for redundant computations in STMT. If any are found, then
2497 replace them with the variable holding the result of the computation.
2499 If safe, record this expression into the available expression hash
2503 eliminate_redundant_computations (tree stmt, stmt_ann_t ann)
2505 tree *expr_p, def = NULL_TREE;
2508 bool retval = false;
2509 bool modify_expr_p = false;
2511 if (TREE_CODE (stmt) == MODIFY_EXPR)
2512 def = TREE_OPERAND (stmt, 0);
2514 /* Certain expressions on the RHS can be optimized away, but can not
2515 themselves be entered into the hash tables. */
2516 if (ann->makes_aliased_stores
2518 || TREE_CODE (def) != SSA_NAME
2519 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
2520 || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VMAYDEF)
2521 /* Do not record equivalences for increments of ivs. This would create
2522 overlapping live ranges for a very questionable gain. */
2523 || simple_iv_increment_p (stmt))
2526 /* Check if the expression has been computed before. */
2527 cached_lhs = lookup_avail_expr (stmt, insert);
2529 /* If this is an assignment and the RHS was not in the hash table,
2530 then try to simplify the RHS and lookup the new RHS in the
2532 if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR)
2533 cached_lhs = simplify_rhs_and_lookup_avail_expr (stmt, insert);
2534 /* Similarly if this is a COND_EXPR and we did not find its
2535 expression in the hash table, simplify the condition and
2537 else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
2538 cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
2539 /* Similarly for a SWITCH_EXPR. */
2540 else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR)
2541 cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert);
2543 opt_stats.num_exprs_considered++;
2545 /* Get a pointer to the expression we are trying to optimize. */
2546 if (TREE_CODE (stmt) == COND_EXPR)
2547 expr_p = &COND_EXPR_COND (stmt);
2548 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2549 expr_p = &SWITCH_COND (stmt);
2550 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
2552 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
2553 modify_expr_p = true;
2557 expr_p = &TREE_OPERAND (stmt, 1);
2558 modify_expr_p = true;
2561 /* It is safe to ignore types here since we have already done
2562 type checking in the hashing and equality routines. In fact
2563 type checking here merely gets in the way of constant
2564 propagation. Also, make sure that it is safe to propagate
2565 CACHED_LHS into *EXPR_P. */
2567 && ((TREE_CODE (cached_lhs) != SSA_NAME
2569 || tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
2570 TREE_TYPE (cached_lhs))))
2571 || may_propagate_copy (*expr_p, cached_lhs)))
2573 if (dump_file && (dump_flags & TDF_DETAILS))
2575 fprintf (dump_file, " Replaced redundant expr '");
2576 print_generic_expr (dump_file, *expr_p, dump_flags);
2577 fprintf (dump_file, "' with '");
2578 print_generic_expr (dump_file, cached_lhs, dump_flags);
2579 fprintf (dump_file, "'\n");
2584 #if defined ENABLE_CHECKING
2585 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
2586 || is_gimple_min_invariant (cached_lhs));
2589 if (TREE_CODE (cached_lhs) == ADDR_EXPR
2590 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
2591 && is_gimple_min_invariant (cached_lhs)))
2595 && !tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
2596 TREE_TYPE (cached_lhs)))
2597 cached_lhs = fold_convert (TREE_TYPE (*expr_p), cached_lhs);
2599 propagate_tree_value (expr_p, cached_lhs);
2600 mark_stmt_modified (stmt);
2605 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2606 the available expressions table or the const_and_copies table.
2607 Detect and record those equivalences. */
2610 record_equivalences_from_stmt (tree stmt,
2614 tree lhs = TREE_OPERAND (stmt, 0);
2615 enum tree_code lhs_code = TREE_CODE (lhs);
2618 if (lhs_code == SSA_NAME)
2620 tree rhs = TREE_OPERAND (stmt, 1);
2622 /* Strip away any useless type conversions. */
2623 STRIP_USELESS_TYPE_CONVERSION (rhs);
2625 /* If the RHS of the assignment is a constant or another variable that
2626 may be propagated, register it in the CONST_AND_COPIES table. We
2627 do not need to record unwind data for this, since this is a true
2628 assignment and not an equivalence inferred from a comparison. All
2629 uses of this ssa name are dominated by this assignment, so unwinding
2630 just costs time and space. */
2632 && (TREE_CODE (rhs) == SSA_NAME
2633 || is_gimple_min_invariant (rhs)))
2634 SSA_NAME_VALUE (lhs) = rhs;
2636 if (tree_expr_nonzero_p (rhs))
2637 record_var_is_nonzero (lhs);
2640 /* Look at both sides for pointer dereferences. If we find one, then
2641 the pointer must be nonnull and we can enter that equivalence into
2643 if (flag_delete_null_pointer_checks)
2644 for (i = 0; i < 2; i++)
2646 tree t = TREE_OPERAND (stmt, i);
2648 /* Strip away any COMPONENT_REFs. */
2649 while (TREE_CODE (t) == COMPONENT_REF)
2650 t = TREE_OPERAND (t, 0);
2652 /* Now see if this is a pointer dereference. */
2653 if (INDIRECT_REF_P (t))
2655 tree op = TREE_OPERAND (t, 0);
2657 /* If the pointer is a SSA variable, then enter new
2658 equivalences into the hash table. */
2659 while (TREE_CODE (op) == SSA_NAME)
2661 tree def = SSA_NAME_DEF_STMT (op);
2663 record_var_is_nonzero (op);
2665 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2666 which are known to have a nonzero value. */
2668 && TREE_CODE (def) == MODIFY_EXPR
2669 && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
2670 op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
2677 /* A memory store, even an aliased store, creates a useful
2678 equivalence. By exchanging the LHS and RHS, creating suitable
2679 vops and recording the result in the available expression table,
2680 we may be able to expose more redundant loads. */
2681 if (!ann->has_volatile_ops
2682 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
2683 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
2684 && !is_gimple_reg (lhs))
2686 tree rhs = TREE_OPERAND (stmt, 1);
2689 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2690 is a constant, we need to adjust the constant to fit into the
2691 type of the LHS. If the LHS is a bitfield and the RHS is not
2692 a constant, then we can not record any equivalences for this
2693 statement since we would need to represent the widening or
2694 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2695 and should not be necessary if GCC represented bitfields
2697 if (lhs_code == COMPONENT_REF
2698 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
2700 if (TREE_CONSTANT (rhs))
2701 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
2705 /* If the value overflowed, then we can not use this equivalence. */
2706 if (rhs && ! is_gimple_min_invariant (rhs))
2712 /* Build a new statement with the RHS and LHS exchanged. */
2713 new = build2 (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
2715 create_ssa_artficial_load_stmt (new, stmt);
2717 /* Finally enter the statement into the available expression
2719 lookup_avail_expr (new, true);
2724 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2725 CONST_AND_COPIES. */
2728 cprop_operand (tree stmt, use_operand_p op_p)
2730 bool may_have_exposed_new_symbols = false;
2732 tree op = USE_FROM_PTR (op_p);
2734 /* If the operand has a known constant value or it is known to be a
2735 copy of some other variable, use the value or copy stored in
2736 CONST_AND_COPIES. */
2737 val = SSA_NAME_VALUE (op);
2738 if (val && val != op && TREE_CODE (val) != VALUE_HANDLE)
2740 tree op_type, val_type;
2742 /* Do not change the base variable in the virtual operand
2743 tables. That would make it impossible to reconstruct
2744 the renamed virtual operand if we later modify this
2745 statement. Also only allow the new value to be an SSA_NAME
2746 for propagation into virtual operands. */
2747 if (!is_gimple_reg (op)
2748 && (TREE_CODE (val) != SSA_NAME
2749 || is_gimple_reg (val)
2750 || get_virtual_var (val) != get_virtual_var (op)))
2753 /* Do not replace hard register operands in asm statements. */
2754 if (TREE_CODE (stmt) == ASM_EXPR
2755 && !may_propagate_copy_into_asm (op))
2758 /* Get the toplevel type of each operand. */
2759 op_type = TREE_TYPE (op);
2760 val_type = TREE_TYPE (val);
2762 /* While both types are pointers, get the type of the object
2764 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
2766 op_type = TREE_TYPE (op_type);
2767 val_type = TREE_TYPE (val_type);
2770 /* Make sure underlying types match before propagating a constant by
2771 converting the constant to the proper type. Note that convert may
2772 return a non-gimple expression, in which case we ignore this
2773 propagation opportunity. */
2774 if (TREE_CODE (val) != SSA_NAME)
2776 if (!lang_hooks.types_compatible_p (op_type, val_type))
2778 val = fold_convert (TREE_TYPE (op), val);
2779 if (!is_gimple_min_invariant (val))
2784 /* Certain operands are not allowed to be copy propagated due
2785 to their interaction with exception handling and some GCC
2787 else if (!may_propagate_copy (op, val))
2790 /* Do not propagate copies if the propagated value is at a deeper loop
2791 depth than the propagatee. Otherwise, this may move loop variant
2792 variables outside of their loops and prevent coalescing
2793 opportunities. If the value was loop invariant, it will be hoisted
2794 by LICM and exposed for copy propagation. */
2795 if (loop_depth_of_name (val) > loop_depth_of_name (op))
2799 if (dump_file && (dump_flags & TDF_DETAILS))
2801 fprintf (dump_file, " Replaced '");
2802 print_generic_expr (dump_file, op, dump_flags);
2803 fprintf (dump_file, "' with %s '",
2804 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
2805 print_generic_expr (dump_file, val, dump_flags);
2806 fprintf (dump_file, "'\n");
2809 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2810 that we may have exposed a new symbol for SSA renaming. */
2811 if (TREE_CODE (val) == ADDR_EXPR
2812 || (POINTER_TYPE_P (TREE_TYPE (op))
2813 && is_gimple_min_invariant (val)))
2814 may_have_exposed_new_symbols = true;
2816 if (TREE_CODE (val) != SSA_NAME)
2817 opt_stats.num_const_prop++;
2819 opt_stats.num_copy_prop++;
2821 propagate_value (op_p, val);
2823 /* And note that we modified this statement. This is now
2824 safe, even if we changed virtual operands since we will
2825 rescan the statement and rewrite its operands again. */
2826 mark_stmt_modified (stmt);
2828 return may_have_exposed_new_symbols;
2831 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2832 known value for that SSA_NAME (or NULL if no value is known).
2834 Propagate values from CONST_AND_COPIES into the uses, vuses and
2835 v_may_def_ops of STMT. */
2838 cprop_into_stmt (tree stmt)
2840 bool may_have_exposed_new_symbols = false;
2844 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
2846 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
2847 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
2850 return may_have_exposed_new_symbols;
2854 /* Optimize the statement pointed to by iterator SI.
2856 We try to perform some simplistic global redundancy elimination and
2857 constant propagation:
2859 1- To detect global redundancy, we keep track of expressions that have
2860 been computed in this block and its dominators. If we find that the
2861 same expression is computed more than once, we eliminate repeated
2862 computations by using the target of the first one.
2864 2- Constant values and copy assignments. This is used to do very
2865 simplistic constant and copy propagation. When a constant or copy
2866 assignment is found, we map the value on the RHS of the assignment to
2867 the variable in the LHS in the CONST_AND_COPIES table. */
2870 optimize_stmt (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2871 basic_block bb, block_stmt_iterator si)
2874 tree stmt, old_stmt;
2875 bool may_optimize_p;
2876 bool may_have_exposed_new_symbols = false;
2878 old_stmt = stmt = bsi_stmt (si);
2880 if (TREE_CODE (stmt) == COND_EXPR)
2881 canonicalize_comparison (stmt);
2883 update_stmt_if_modified (stmt);
2884 ann = stmt_ann (stmt);
2885 opt_stats.num_stmts++;
2886 may_have_exposed_new_symbols = false;
2888 if (dump_file && (dump_flags & TDF_DETAILS))
2890 fprintf (dump_file, "Optimizing statement ");
2891 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2894 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2895 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
2897 /* If the statement has been modified with constant replacements,
2898 fold its RHS before checking for redundant computations. */
2903 /* Try to fold the statement making sure that STMT is kept
2905 if (fold_stmt (bsi_stmt_ptr (si)))
2907 stmt = bsi_stmt (si);
2908 ann = stmt_ann (stmt);
2910 if (dump_file && (dump_flags & TDF_DETAILS))
2912 fprintf (dump_file, " Folded to: ");
2913 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2917 rhs = get_rhs (stmt);
2918 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
2919 recompute_tree_invariant_for_addr_expr (rhs);
2921 /* Constant/copy propagation above may change the set of
2922 virtual operands associated with this statement. Folding
2923 may remove the need for some virtual operands.
2925 Indicate we will need to rescan and rewrite the statement. */
2926 may_have_exposed_new_symbols = true;
2929 /* Check for redundant computations. Do this optimization only
2930 for assignments that have no volatile ops and conditionals. */
2931 may_optimize_p = (!ann->has_volatile_ops
2932 && ((TREE_CODE (stmt) == RETURN_EXPR
2933 && TREE_OPERAND (stmt, 0)
2934 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
2935 && ! (TREE_SIDE_EFFECTS
2936 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
2937 || (TREE_CODE (stmt) == MODIFY_EXPR
2938 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
2939 || TREE_CODE (stmt) == COND_EXPR
2940 || TREE_CODE (stmt) == SWITCH_EXPR));
2943 may_have_exposed_new_symbols
2944 |= eliminate_redundant_computations (stmt, ann);
2946 /* Record any additional equivalences created by this statement. */
2947 if (TREE_CODE (stmt) == MODIFY_EXPR)
2948 record_equivalences_from_stmt (stmt,
2952 /* If STMT is a COND_EXPR and it was modified, then we may know
2953 where it goes. If that is the case, then mark the CFG as altered.
2955 This will cause us to later call remove_unreachable_blocks and
2956 cleanup_tree_cfg when it is safe to do so. It is not safe to
2957 clean things up here since removal of edges and such can trigger
2958 the removal of PHI nodes, which in turn can release SSA_NAMEs to
2961 That's all fine and good, except that once SSA_NAMEs are released
2962 to the manager, we must not call create_ssa_name until all references
2963 to released SSA_NAMEs have been eliminated.
2965 All references to the deleted SSA_NAMEs can not be eliminated until
2966 we remove unreachable blocks.
2968 We can not remove unreachable blocks until after we have completed
2969 any queued jump threading.
2971 We can not complete any queued jump threads until we have taken
2972 appropriate variables out of SSA form. Taking variables out of
2973 SSA form can call create_ssa_name and thus we lose.
2975 Ultimately I suspect we're going to need to change the interface
2976 into the SSA_NAME manager. */
2982 if (TREE_CODE (stmt) == COND_EXPR)
2983 val = COND_EXPR_COND (stmt);
2984 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2985 val = SWITCH_COND (stmt);
2987 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
2990 /* If we simplified a statement in such a way as to be shown that it
2991 cannot trap, update the eh information and the cfg to match. */
2992 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
2994 bitmap_set_bit (need_eh_cleanup, bb->index);
2995 if (dump_file && (dump_flags & TDF_DETAILS))
2996 fprintf (dump_file, " Flagged to clear EH edges.\n");
3000 if (may_have_exposed_new_symbols)
3001 VEC_safe_push (tree, heap, stmts_to_rescan, bsi_stmt (si));
3004 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
3005 available expression hashtable, then return the LHS from the hash
3008 If INSERT is true, then we also update the available expression
3009 hash table to account for the changes made to STMT. */
3012 update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert)
3014 tree cached_lhs = NULL;
3016 /* Remove the old entry from the hash table. */
3019 struct expr_hash_elt element;
3021 initialize_hash_element (stmt, NULL, &element);
3022 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
3025 /* Now update the RHS of the assignment. */
3026 TREE_OPERAND (stmt, 1) = new_rhs;
3028 /* Now lookup the updated statement in the hash table. */
3029 cached_lhs = lookup_avail_expr (stmt, insert);
3031 /* We have now called lookup_avail_expr twice with two different
3032 versions of this same statement, once in optimize_stmt, once here.
3034 We know the call in optimize_stmt did not find an existing entry
3035 in the hash table, so a new entry was created. At the same time
3036 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
3038 If this call failed to find an existing entry on the hash table,
3039 then the new version of this statement was entered into the
3040 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3041 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3043 If this call succeeded, we still have one copy of this statement
3044 on the BLOCK_AVAIL_EXPRs vector.
3046 For both cases, we need to pop the most recent entry off the
3047 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3048 statement in the hash tables, that will leave precisely one
3049 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3050 we found a copy of this statement in the second hash table lookup
3051 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3053 VEC_pop (tree, avail_exprs_stack);
3055 /* And make sure we record the fact that we modified this
3057 mark_stmt_modified (stmt);
3062 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3063 found, return its LHS. Otherwise insert STMT in the table and return
3066 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3067 is also added to the stack pointed to by BLOCK_AVAIL_EXPRS_P, so that they
3068 can be removed when we finish processing this block and its children.
3070 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3071 contains no CALL_EXPR on its RHS and makes no volatile nor
3072 aliased references. */
3075 lookup_avail_expr (tree stmt, bool insert)
3080 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
3082 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
3084 initialize_hash_element (stmt, lhs, element);
3086 /* Don't bother remembering constant assignments and copy operations.
3087 Constants and copy operations are handled by the constant/copy propagator
3088 in optimize_stmt. */
3089 if (TREE_CODE (element->rhs) == SSA_NAME
3090 || is_gimple_min_invariant (element->rhs))
3096 /* If this is an equality test against zero, see if we have recorded a
3097 nonzero value for the variable in question. */
3098 if ((TREE_CODE (element->rhs) == EQ_EXPR
3099 || TREE_CODE (element->rhs) == NE_EXPR)
3100 && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
3101 && integer_zerop (TREE_OPERAND (element->rhs, 1)))
3103 int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
3105 if (bitmap_bit_p (nonzero_vars, indx))
3107 tree t = element->rhs;
3109 return constant_boolean_node (TREE_CODE (t) != EQ_EXPR,
3114 /* Finally try to find the expression in the main expression hash table. */
3115 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
3116 (insert ? INSERT : NO_INSERT));
3125 *slot = (void *) element;
3126 VEC_safe_push (tree, heap, avail_exprs_stack,
3127 stmt ? stmt : element->rhs);
3131 /* Extract the LHS of the assignment so that it can be used as the current
3132 definition of another variable. */
3133 lhs = ((struct expr_hash_elt *)*slot)->lhs;
3135 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3136 use the value from the const_and_copies table. */
3137 if (TREE_CODE (lhs) == SSA_NAME)
3139 temp = SSA_NAME_VALUE (lhs);
3140 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
3148 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3149 range of values that result in the conditional having a true value.
3151 Return true if we are successful in extracting a range from COND and
3152 false if we are unsuccessful. */
3155 extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
3157 tree op1 = TREE_OPERAND (cond, 1);
3158 tree high, low, type;
3161 type = TREE_TYPE (op1);
3163 /* Experiments have shown that it's rarely, if ever useful to
3164 record ranges for enumerations. Presumably this is due to
3165 the fact that they're rarely used directly. They are typically
3166 cast into an integer type and used that way. */
3167 if (TREE_CODE (type) != INTEGER_TYPE)
3170 switch (TREE_CODE (cond))
3185 /* Get the highest value of the type. If not a constant, use that
3186 of its base type, if it has one. */
3187 high = TYPE_MAX_VALUE (type);
3188 if (TREE_CODE (high) != INTEGER_CST && TREE_TYPE (type))
3189 high = TYPE_MAX_VALUE (TREE_TYPE (type));
3194 high = TYPE_MAX_VALUE (type);
3195 if (TREE_CODE (high) != INTEGER_CST && TREE_TYPE (type))
3196 high = TYPE_MAX_VALUE (TREE_TYPE (type));
3197 if (!tree_int_cst_lt (op1, high))
3199 low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
3205 low = TYPE_MIN_VALUE (type);
3206 if (TREE_CODE (low) != INTEGER_CST && TREE_TYPE (type))
3207 low = TYPE_MIN_VALUE (TREE_TYPE (type));
3212 low = TYPE_MIN_VALUE (type);
3213 if (TREE_CODE (low) != INTEGER_CST && TREE_TYPE (type))
3214 low = TYPE_MIN_VALUE (TREE_TYPE (type));
3215 if (!tree_int_cst_lt (low, op1))
3217 high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
3227 *inverted_p = inverted;
3231 /* Record a range created by COND for basic block BB. */
3234 record_range (tree cond, basic_block bb)
3236 enum tree_code code = TREE_CODE (cond);
3238 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3239 They rarely allow for meaningful range optimizations and significantly
3240 complicate the implementation. */
3241 if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
3242 || code == GE_EXPR || code == EQ_EXPR)
3243 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
3245 struct vrp_hash_elt *vrp_hash_elt;
3246 struct vrp_element *element;
3247 VEC(vrp_element_p,heap) **vrp_records_p;
3251 vrp_hash_elt = xmalloc (sizeof (struct vrp_hash_elt));
3252 vrp_hash_elt->var = TREE_OPERAND (cond, 0);
3253 vrp_hash_elt->records = NULL;
3254 slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
3257 *slot = (void *) vrp_hash_elt;
3259 vrp_free (vrp_hash_elt);
3261 vrp_hash_elt = (struct vrp_hash_elt *) *slot;
3262 vrp_records_p = &vrp_hash_elt->records;
3264 element = ggc_alloc (sizeof (struct vrp_element));
3265 element->low = NULL;
3266 element->high = NULL;
3267 element->cond = cond;
3270 VEC_safe_push (vrp_element_p, heap, *vrp_records_p, element);
3271 VEC_safe_push (tree, heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
3275 /* Hashing and equality functions for VRP_DATA.
3277 Since this hash table is addressed by SSA_NAMEs, we can hash on
3278 their version number and equality can be determined with a
3279 pointer comparison. */
3282 vrp_hash (const void *p)
3284 tree var = ((struct vrp_hash_elt *)p)->var;
3286 return SSA_NAME_VERSION (var);
3290 vrp_eq (const void *p1, const void *p2)
3292 tree var1 = ((struct vrp_hash_elt *)p1)->var;
3293 tree var2 = ((struct vrp_hash_elt *)p2)->var;
3295 return var1 == var2;
3298 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3299 MODIFY_EXPR statements. We compute a value number for expressions using
3300 the code of the expression and the SSA numbers of its operands. */
3303 avail_expr_hash (const void *p)
3305 tree stmt = ((struct expr_hash_elt *)p)->stmt;
3306 tree rhs = ((struct expr_hash_elt *)p)->rhs;
3311 /* iterative_hash_expr knows how to deal with any expression and
3312 deals with commutative operators as well, so just use it instead
3313 of duplicating such complexities here. */
3314 val = iterative_hash_expr (rhs, val);
3316 /* If the hash table entry is not associated with a statement, then we
3317 can just hash the expression and not worry about virtual operands
3319 if (!stmt || !stmt_ann (stmt))
3322 /* Add the SSA version numbers of every vuse operand. This is important
3323 because compound variables like arrays are not renamed in the
3324 operands. Rather, the rename is done on the virtual variable
3325 representing all the elements of the array. */
3326 FOR_EACH_SSA_TREE_OPERAND (vuse, stmt, iter, SSA_OP_VUSE)
3327 val = iterative_hash_expr (vuse, val);
3333 real_avail_expr_hash (const void *p)
3335 return ((const struct expr_hash_elt *)p)->hash;
3339 avail_expr_eq (const void *p1, const void *p2)
3341 tree stmt1 = ((struct expr_hash_elt *)p1)->stmt;
3342 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
3343 tree stmt2 = ((struct expr_hash_elt *)p2)->stmt;
3344 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
3346 /* If they are the same physical expression, return true. */
3347 if (rhs1 == rhs2 && stmt1 == stmt2)
3350 /* If their codes are not equal, then quit now. */
3351 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
3354 /* In case of a collision, both RHS have to be identical and have the
3355 same VUSE operands. */
3356 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
3357 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
3358 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
3360 bool ret = compare_ssa_operands_equal (stmt1, stmt2, SSA_OP_VUSE);
3361 gcc_assert (!ret || ((struct expr_hash_elt *)p1)->hash
3362 == ((struct expr_hash_elt *)p2)->hash);