1 /* SSA Dominator optimizations for trees
2 Copyright (C) 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
3 Contributed by Diego Novillo <dnovillo@redhat.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
24 #include "coretypes.h"
31 #include "basic-block.h"
36 #include "diagnostic.h"
38 #include "tree-dump.h"
39 #include "tree-flow.h"
42 #include "tree-pass.h"
43 #include "tree-ssa-propagate.h"
44 #include "langhooks.h"
46 /* This file implements optimizations on the dominator tree. */
49 /* Structure for recording edge equivalences as well as any pending
50 edge redirections during the dominator optimizer.
52 Computing and storing the edge equivalences instead of creating
53 them on-demand can save significant amounts of time, particularly
54 for pathological cases involving switch statements.
56 These structures live for a single iteration of the dominator
57 optimizer in the edge's AUX field. At the end of an iteration we
58 free each of these structures and update the AUX field to point
59 to any requested redirection target (the code for updating the
60 CFG and SSA graph for edge redirection expects redirection edge
61 targets to be in the AUX field for each edge. */
65 /* If this edge creates a simple equivalence, the LHS and RHS of
66 the equivalence will be stored here. */
70 /* Traversing an edge may also indicate one or more particular conditions
71 are true or false. The number of recorded conditions can vary, but
72 can be determined by the condition's code. So we have an array
73 and its maximum index rather than use a varray. */
74 tree
*cond_equivalences
;
75 unsigned int max_cond_equivalences
;
77 /* If we can thread this edge this field records the new target. */
78 edge redirection_target
;
82 /* Hash table with expressions made available during the renaming process.
83 When an assignment of the form X_i = EXPR is found, the statement is
84 stored in this table. If the same expression EXPR is later found on the
85 RHS of another statement, it is replaced with X_i (thus performing
86 global redundancy elimination). Similarly as we pass through conditionals
87 we record the conditional itself as having either a true or false value
89 static htab_t avail_exprs
;
91 /* Stack of available expressions in AVAIL_EXPRs. Each block pushes any
92 expressions it enters into the hash table along with a marker entry
93 (null). When we finish processing the block, we pop off entries and
94 remove the expressions from the global hash table until we hit the
96 static VEC(tree_on_heap
) *avail_exprs_stack
;
98 /* Stack of trees used to restore the global currdefs to its original
99 state after completing optimization of a block and its dominator children.
101 An SSA_NAME indicates that the current definition of the underlying
102 variable should be set to the given SSA_NAME.
104 A _DECL node indicates that the underlying variable has no current
107 A NULL node is used to mark the last node associated with the
109 static VEC(tree_on_heap
) *block_defs_stack
;
111 /* Stack of statements we need to rescan during finalization for newly
114 Statement rescanning must occur after the current block's available
115 expressions are removed from AVAIL_EXPRS. Else we may change the
116 hash code for an expression and be unable to find/remove it from
118 static VEC(tree_on_heap
) *stmts_to_rescan
;
120 /* Structure for entries in the expression hash table.
122 This requires more memory for the hash table entries, but allows us
123 to avoid creating silly tree nodes and annotations for conditionals,
124 eliminates 2 global hash tables and two block local varrays.
126 It also allows us to reduce the number of hash table lookups we
127 have to perform in lookup_avail_expr and finally it allows us to
128 significantly reduce the number of calls into the hashing routine
133 /* The value (lhs) of this expression. */
136 /* The expression (rhs) we want to record. */
139 /* The annotation if this element corresponds to a statement. */
142 /* The hash value for RHS/ann. */
146 /* Stack of dest,src pairs that need to be restored during finalization.
148 A NULL entry is used to mark the end of pairs which need to be
149 restored during finalization of this block. */
150 static VEC(tree_on_heap
) *const_and_copies_stack
;
152 /* Bitmap of SSA_NAMEs known to have a nonzero value, even if we do not
153 know their exact value. */
154 static bitmap nonzero_vars
;
156 /* Stack of SSA_NAMEs which need their NONZERO_VARS property cleared
157 when the current block is finalized.
159 A NULL entry is used to mark the end of names needing their
160 entry in NONZERO_VARS cleared during finalization of this block. */
161 static VEC(tree_on_heap
) *nonzero_vars_stack
;
163 /* Track whether or not we have changed the control flow graph. */
164 static bool cfg_altered
;
166 /* Bitmap of blocks that have had EH statements cleaned. We should
167 remove their dead edges eventually. */
168 static bitmap need_eh_cleanup
;
170 /* Statistics for dominator optimizations. */
174 long num_exprs_considered
;
178 static struct opt_stats_d opt_stats
;
180 /* Value range propagation record. Each time we encounter a conditional
181 of the form SSA_NAME COND CONST we create a new vrp_element to record
182 how the condition affects the possible values SSA_NAME may have.
184 Each record contains the condition tested (COND), and the range of
185 values the variable may legitimately have if COND is true. Note the
186 range of values may be a smaller range than COND specifies if we have
187 recorded other ranges for this variable. Each record also contains the
188 block in which the range was recorded for invalidation purposes.
190 Note that the current known range is computed lazily. This allows us
191 to avoid the overhead of computing ranges which are never queried.
193 When we encounter a conditional, we look for records which constrain
194 the SSA_NAME used in the condition. In some cases those records allow
195 us to determine the condition's result at compile time. In other cases
196 they may allow us to simplify the condition.
198 We also use value ranges to do things like transform signed div/mod
199 operations into unsigned div/mod or to simplify ABS_EXPRs.
201 Simple experiments have shown these optimizations to not be all that
202 useful on switch statements (much to my surprise). So switch statement
203 optimizations are not performed.
205 Note carefully we do not propagate information through each statement
206 in the block. i.e., if we know variable X has a value defined of
207 [0, 25] and we encounter Y = X + 1, we do not track a value range
208 for Y (which would be [1, 26] if we cared). Similarly we do not
209 constrain values as we encounter narrowing typecasts, etc. */
213 /* The highest and lowest values the variable in COND may contain when
214 COND is true. Note this may not necessarily be the same values
215 tested by COND if the same variable was used in earlier conditionals.
217 Note this is computed lazily and thus can be NULL indicating that
218 the values have not been computed yet. */
222 /* The actual conditional we recorded. This is needed since we compute
226 /* The basic block where this record was created. We use this to determine
227 when to remove records. */
231 /* A hash table holding value range records (VRP_ELEMENTs) for a given
232 SSA_NAME. We used to use a varray indexed by SSA_NAME_VERSION, but
233 that gets awful wasteful, particularly since the density objects
234 with useful information is very low. */
235 static htab_t vrp_data
;
237 /* An entry in the VRP_DATA hash table. We record the variable and a
238 varray of VRP_ELEMENT records associated with that variable. */
245 /* Array of variables which have their values constrained by operations
246 in this basic block. We use this during finalization to know
247 which variables need their VRP data updated. */
249 /* Stack of SSA_NAMEs which had their values 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_on_heap
) *vrp_variables_stack
;
262 /* Local functions. */
263 static void optimize_stmt (struct dom_walk_data
*,
265 block_stmt_iterator
);
266 static tree
lookup_avail_expr (tree
, bool);
267 static hashval_t
vrp_hash (const void *);
268 static int vrp_eq (const void *, const void *);
269 static hashval_t
avail_expr_hash (const void *);
270 static hashval_t
real_avail_expr_hash (const void *);
271 static int avail_expr_eq (const void *, const void *);
272 static void htab_statistics (FILE *, htab_t
);
273 static void record_cond (tree
, tree
);
274 static void record_const_or_copy (tree
, tree
);
275 static void record_equality (tree
, tree
);
276 static tree
update_rhs_and_lookup_avail_expr (tree
, tree
, bool);
277 static tree
simplify_rhs_and_lookup_avail_expr (struct dom_walk_data
*,
279 static tree
simplify_cond_and_lookup_avail_expr (tree
, stmt_ann_t
, int);
280 static tree
simplify_switch_and_lookup_avail_expr (tree
, int);
281 static tree
find_equivalent_equality_comparison (tree
);
282 static void record_range (tree
, basic_block
);
283 static bool extract_range_from_cond (tree
, tree
*, tree
*, int *);
284 static void record_equivalences_from_phis (basic_block
);
285 static void record_equivalences_from_incoming_edge (basic_block
);
286 static bool eliminate_redundant_computations (struct dom_walk_data
*,
288 static void record_equivalences_from_stmt (tree
, int, stmt_ann_t
);
289 static void thread_across_edge (struct dom_walk_data
*, edge
);
290 static void dom_opt_finalize_block (struct dom_walk_data
*, basic_block
);
291 static void dom_opt_initialize_block (struct dom_walk_data
*, basic_block
);
292 static void propagate_to_outgoing_edges (struct dom_walk_data
*, basic_block
);
293 static void remove_local_expressions_from_table (void);
294 static void restore_vars_to_original_value (void);
295 static void restore_currdefs_to_original_value (void);
296 static void register_definitions_for_stmt (tree
);
297 static edge
single_incoming_edge_ignoring_loop_edges (basic_block
);
298 static void restore_nonzero_vars_to_original_value (void);
299 static inline bool unsafe_associative_fp_binop (tree
);
301 /* Local version of fold that doesn't introduce cruft. */
308 /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that
309 may have been added by fold, and "useless" type conversions that might
310 now be apparent due to propagation. */
311 STRIP_USELESS_TYPE_CONVERSION (t
);
316 /* Allocate an EDGE_INFO for edge E and attach it to E.
317 Return the new EDGE_INFO structure. */
319 static struct edge_info
*
320 allocate_edge_info (edge e
)
322 struct edge_info
*edge_info
;
324 edge_info
= xcalloc (1, sizeof (struct edge_info
));
330 /* Free all EDGE_INFO structures associated with edges in the CFG.
331 If a particular edge can be threaded, copy the redirection
332 target from the EDGE_INFO structure into the edge's AUX field
333 as required by code to update the CFG and SSA graph for
337 free_all_edge_infos (void)
345 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
347 struct edge_info
*edge_info
= e
->aux
;
351 e
->aux
= edge_info
->redirection_target
;
352 if (edge_info
->cond_equivalences
)
353 free (edge_info
->cond_equivalences
);
360 /* Jump threading, redundancy elimination and const/copy propagation.
362 This pass may expose new symbols that need to be renamed into SSA. For
363 every new symbol exposed, its corresponding bit will be set in
367 tree_ssa_dominator_optimize (void)
369 struct dom_walk_data walk_data
;
372 memset (&opt_stats
, 0, sizeof (opt_stats
));
374 for (i
= 0; i
< num_referenced_vars
; i
++)
375 var_ann (referenced_var (i
))->current_def
= NULL
;
377 /* Create our hash tables. */
378 avail_exprs
= htab_create (1024, real_avail_expr_hash
, avail_expr_eq
, free
);
379 vrp_data
= htab_create (ceil_log2 (num_ssa_names
), vrp_hash
, vrp_eq
, free
);
380 avail_exprs_stack
= VEC_alloc (tree_on_heap
, 20);
381 block_defs_stack
= VEC_alloc (tree_on_heap
, 20);
382 const_and_copies_stack
= VEC_alloc (tree_on_heap
, 20);
383 nonzero_vars_stack
= VEC_alloc (tree_on_heap
, 20);
384 vrp_variables_stack
= VEC_alloc (tree_on_heap
, 20);
385 stmts_to_rescan
= VEC_alloc (tree_on_heap
, 20);
386 nonzero_vars
= BITMAP_ALLOC (NULL
);
387 need_eh_cleanup
= BITMAP_ALLOC (NULL
);
389 /* Setup callbacks for the generic dominator tree walker. */
390 walk_data
.walk_stmts_backward
= false;
391 walk_data
.dom_direction
= CDI_DOMINATORS
;
392 walk_data
.initialize_block_local_data
= NULL
;
393 walk_data
.before_dom_children_before_stmts
= dom_opt_initialize_block
;
394 walk_data
.before_dom_children_walk_stmts
= optimize_stmt
;
395 walk_data
.before_dom_children_after_stmts
= propagate_to_outgoing_edges
;
396 walk_data
.after_dom_children_before_stmts
= NULL
;
397 walk_data
.after_dom_children_walk_stmts
= NULL
;
398 walk_data
.after_dom_children_after_stmts
= dom_opt_finalize_block
;
399 /* Right now we only attach a dummy COND_EXPR to the global data pointer.
400 When we attach more stuff we'll need to fill this out with a real
402 walk_data
.global_data
= NULL
;
403 walk_data
.block_local_data_size
= 0;
405 /* Now initialize the dominator walker. */
406 init_walk_dominator_tree (&walk_data
);
408 calculate_dominance_info (CDI_DOMINATORS
);
410 /* If we prove certain blocks are unreachable, then we want to
411 repeat the dominator optimization process as PHI nodes may
412 have turned into copies which allows better propagation of
413 values. So we repeat until we do not identify any new unreachable
417 /* Optimize the dominator tree. */
420 /* Recursively walk the dominator tree optimizing statements. */
421 walk_dominator_tree (&walk_data
, ENTRY_BLOCK_PTR
);
423 /* If we exposed any new variables, go ahead and put them into
424 SSA form now, before we handle jump threading. This simplifies
425 interactions between rewriting of _DECL nodes into SSA form
426 and rewriting SSA_NAME nodes into SSA form after block
427 duplication and CFG manipulation. */
428 if (!bitmap_empty_p (vars_to_rename
))
430 rewrite_into_ssa (false);
431 bitmap_clear (vars_to_rename
);
434 free_all_edge_infos ();
436 /* Thread jumps, creating duplicate blocks as needed. */
437 cfg_altered
|= thread_through_all_blocks ();
439 /* Removal of statements may make some EH edges dead. Purge
440 such edges from the CFG as needed. */
441 if (!bitmap_empty_p (need_eh_cleanup
))
443 cfg_altered
|= tree_purge_all_dead_eh_edges (need_eh_cleanup
);
444 bitmap_zero (need_eh_cleanup
);
448 free_dominance_info (CDI_DOMINATORS
);
449 cfg_altered
|= cleanup_tree_cfg ();
450 calculate_dominance_info (CDI_DOMINATORS
);
452 rewrite_ssa_into_ssa ();
454 /* Reinitialize the various tables. */
455 bitmap_clear (nonzero_vars
);
456 htab_empty (avail_exprs
);
457 htab_empty (vrp_data
);
459 for (i
= 0; i
< num_referenced_vars
; i
++)
460 var_ann (referenced_var (i
))->current_def
= NULL
;
462 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
464 This must be done before we iterate as we might have a
465 reference to an SSA_NAME which was removed by the call to
466 rewrite_ssa_into_ssa.
468 Long term we will be able to let everything in SSA_NAME_VALUE
469 persist. However, for now, we know this is the safe thing to do. */
470 for (i
= 0; i
< num_ssa_names
; i
++)
472 tree name
= ssa_name (i
);
478 value
= SSA_NAME_VALUE (name
);
479 if (value
&& !is_gimple_min_invariant (value
))
480 SSA_NAME_VALUE (name
) = NULL
;
483 while (optimize
> 1 && cfg_altered
);
485 /* Debugging dumps. */
486 if (dump_file
&& (dump_flags
& TDF_STATS
))
487 dump_dominator_optimization_stats (dump_file
);
489 /* We emptied the hash table earlier, now delete it completely. */
490 htab_delete (avail_exprs
);
491 htab_delete (vrp_data
);
493 /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
494 CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
495 of the do-while loop above. */
497 /* And finalize the dominator walker. */
498 fini_walk_dominator_tree (&walk_data
);
500 /* Free nonzero_vars. */
501 BITMAP_FREE (nonzero_vars
);
502 BITMAP_FREE (need_eh_cleanup
);
504 VEC_free (tree_on_heap
, block_defs_stack
);
505 VEC_free (tree_on_heap
, avail_exprs_stack
);
506 VEC_free (tree_on_heap
, const_and_copies_stack
);
507 VEC_free (tree_on_heap
, nonzero_vars_stack
);
508 VEC_free (tree_on_heap
, vrp_variables_stack
);
509 VEC_free (tree_on_heap
, stmts_to_rescan
);
513 gate_dominator (void)
515 return flag_tree_dom
!= 0;
518 struct tree_opt_pass pass_dominator
=
521 gate_dominator
, /* gate */
522 tree_ssa_dominator_optimize
, /* execute */
525 0, /* static_pass_number */
526 TV_TREE_SSA_DOMINATOR_OPTS
, /* tv_id */
527 PROP_cfg
| PROP_ssa
| PROP_alias
, /* properties_required */
528 0, /* properties_provided */
529 0, /* properties_destroyed */
530 0, /* todo_flags_start */
531 TODO_dump_func
| TODO_rename_vars
532 | TODO_verify_ssa
, /* todo_flags_finish */
537 /* We are exiting BB, see if the target block begins with a conditional
538 jump which has a known value when reached via BB. */
541 thread_across_edge (struct dom_walk_data
*walk_data
, edge e
)
543 block_stmt_iterator bsi
;
547 /* Each PHI creates a temporary equivalence, record them. */
548 for (phi
= phi_nodes (e
->dest
); phi
; phi
= PHI_CHAIN (phi
))
550 tree src
= PHI_ARG_DEF_FROM_EDGE (phi
, e
);
551 tree dst
= PHI_RESULT (phi
);
553 /* If the desired argument is not the same as this PHI's result
554 and it is set by a PHI in this block, then we can not thread
555 through this block. */
557 && TREE_CODE (src
) == SSA_NAME
558 && TREE_CODE (SSA_NAME_DEF_STMT (src
)) == PHI_NODE
559 && bb_for_stmt (SSA_NAME_DEF_STMT (src
)) == e
->dest
)
562 record_const_or_copy (dst
, src
);
563 register_new_def (dst
, &block_defs_stack
);
566 for (bsi
= bsi_start (e
->dest
); ! bsi_end_p (bsi
); bsi_next (&bsi
))
568 tree lhs
, cached_lhs
;
570 stmt
= bsi_stmt (bsi
);
572 /* Ignore empty statements and labels. */
573 if (IS_EMPTY_STMT (stmt
) || TREE_CODE (stmt
) == LABEL_EXPR
)
576 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
577 value, then stop our search here. Ideally when we stop a
578 search we stop on a COND_EXPR or SWITCH_EXPR. */
579 if (TREE_CODE (stmt
) != MODIFY_EXPR
580 || TREE_CODE (TREE_OPERAND (stmt
, 0)) != SSA_NAME
)
583 /* At this point we have a statement which assigns an RHS to an
584 SSA_VAR on the LHS. We want to prove that the RHS is already
585 available and that its value is held in the current definition
586 of the LHS -- meaning that this assignment is a NOP when
587 reached via edge E. */
588 if (TREE_CODE (TREE_OPERAND (stmt
, 1)) == SSA_NAME
)
589 cached_lhs
= TREE_OPERAND (stmt
, 1);
591 cached_lhs
= lookup_avail_expr (stmt
, false);
593 lhs
= TREE_OPERAND (stmt
, 0);
595 /* This can happen if we thread around to the start of a loop. */
596 if (lhs
== cached_lhs
)
599 /* If we did not find RHS in the hash table, then try again after
600 temporarily const/copy propagating the operands. */
603 /* Copy the operands. */
604 stmt_ann_t ann
= stmt_ann (stmt
);
605 use_optype uses
= USE_OPS (ann
);
606 vuse_optype vuses
= VUSE_OPS (ann
);
607 tree
*uses_copy
= xmalloc (NUM_USES (uses
) * sizeof (tree
));
608 tree
*vuses_copy
= xmalloc (NUM_VUSES (vuses
) * sizeof (tree
));
611 /* Make a copy of the uses into USES_COPY, then cprop into
613 for (i
= 0; i
< NUM_USES (uses
); i
++)
617 uses_copy
[i
] = USE_OP (uses
, i
);
618 if (TREE_CODE (USE_OP (uses
, i
)) == SSA_NAME
)
619 tmp
= SSA_NAME_VALUE (USE_OP (uses
, i
));
620 if (tmp
&& TREE_CODE (tmp
) != VALUE_HANDLE
)
621 SET_USE_OP (uses
, i
, tmp
);
624 /* Similarly for virtual uses. */
625 for (i
= 0; i
< NUM_VUSES (vuses
); i
++)
629 vuses_copy
[i
] = VUSE_OP (vuses
, i
);
630 if (TREE_CODE (VUSE_OP (vuses
, i
)) == SSA_NAME
)
631 tmp
= SSA_NAME_VALUE (VUSE_OP (vuses
, i
));
632 if (tmp
&& TREE_CODE (tmp
) != VALUE_HANDLE
)
633 SET_VUSE_OP (vuses
, i
, tmp
);
636 /* Try to lookup the new expression. */
637 cached_lhs
= lookup_avail_expr (stmt
, false);
639 /* Restore the statement's original uses/defs. */
640 for (i
= 0; i
< NUM_USES (uses
); i
++)
641 SET_USE_OP (uses
, i
, uses_copy
[i
]);
643 for (i
= 0; i
< NUM_VUSES (vuses
); i
++)
644 SET_VUSE_OP (vuses
, i
, vuses_copy
[i
]);
649 /* If we still did not find the expression in the hash table,
650 then we can not ignore this statement. */
655 /* If the expression in the hash table was not assigned to an
656 SSA_NAME, then we can not ignore this statement. */
657 if (TREE_CODE (cached_lhs
) != SSA_NAME
)
660 /* If we have different underlying variables, then we can not
661 ignore this statement. */
662 if (SSA_NAME_VAR (cached_lhs
) != SSA_NAME_VAR (lhs
))
665 /* If CACHED_LHS does not represent the current value of the underlying
666 variable in CACHED_LHS/LHS, then we can not ignore this statement. */
667 if (var_ann (SSA_NAME_VAR (lhs
))->current_def
!= cached_lhs
)
670 /* If we got here, then we can ignore this statement and continue
671 walking through the statements in the block looking for a threadable
674 We want to record an equivalence lhs = cache_lhs so that if
675 the result of this statement is used later we can copy propagate
677 record_const_or_copy (lhs
, cached_lhs
);
678 register_new_def (lhs
, &block_defs_stack
);
681 /* If we stopped at a COND_EXPR or SWITCH_EXPR, then see if we know which
682 arm will be taken. */
684 && (TREE_CODE (stmt
) == COND_EXPR
685 || TREE_CODE (stmt
) == SWITCH_EXPR
))
687 tree cond
, cached_lhs
;
689 /* Now temporarily cprop the operands and try to find the resulting
690 expression in the hash tables. */
691 if (TREE_CODE (stmt
) == COND_EXPR
)
692 cond
= COND_EXPR_COND (stmt
);
694 cond
= SWITCH_COND (stmt
);
696 if (COMPARISON_CLASS_P (cond
))
698 tree dummy_cond
, op0
, op1
;
699 enum tree_code cond_code
;
701 op0
= TREE_OPERAND (cond
, 0);
702 op1
= TREE_OPERAND (cond
, 1);
703 cond_code
= TREE_CODE (cond
);
705 /* Get the current value of both operands. */
706 if (TREE_CODE (op0
) == SSA_NAME
)
708 tree tmp
= SSA_NAME_VALUE (op0
);
709 if (tmp
&& TREE_CODE (tmp
) != VALUE_HANDLE
)
713 if (TREE_CODE (op1
) == SSA_NAME
)
715 tree tmp
= SSA_NAME_VALUE (op1
);
716 if (tmp
&& TREE_CODE (tmp
) != VALUE_HANDLE
)
720 /* Stuff the operator and operands into our dummy conditional
721 expression, creating the dummy conditional if necessary. */
722 dummy_cond
= walk_data
->global_data
;
725 dummy_cond
= build (cond_code
, boolean_type_node
, op0
, op1
);
726 dummy_cond
= build (COND_EXPR
, void_type_node
,
727 dummy_cond
, NULL
, NULL
);
728 walk_data
->global_data
= dummy_cond
;
732 TREE_SET_CODE (COND_EXPR_COND (dummy_cond
), cond_code
);
733 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 0) = op0
;
734 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 1) = op1
;
737 /* If the conditional folds to an invariant, then we are done,
738 otherwise look it up in the hash tables. */
739 cached_lhs
= local_fold (COND_EXPR_COND (dummy_cond
));
740 if (! is_gimple_min_invariant (cached_lhs
))
742 cached_lhs
= lookup_avail_expr (dummy_cond
, false);
743 if (!cached_lhs
|| ! is_gimple_min_invariant (cached_lhs
))
744 cached_lhs
= simplify_cond_and_lookup_avail_expr (dummy_cond
,
749 /* We can have conditionals which just test the state of a
750 variable rather than use a relational operator. These are
751 simpler to handle. */
752 else if (TREE_CODE (cond
) == SSA_NAME
)
755 cached_lhs
= SSA_NAME_VALUE (cached_lhs
);
756 if (cached_lhs
&& ! is_gimple_min_invariant (cached_lhs
))
760 cached_lhs
= lookup_avail_expr (stmt
, false);
764 edge taken_edge
= find_taken_edge (e
->dest
, cached_lhs
);
765 basic_block dest
= (taken_edge
? taken_edge
->dest
: NULL
);
770 /* If we have a known destination for the conditional, then
771 we can perform this optimization, which saves at least one
772 conditional jump each time it applies since we get to
773 bypass the conditional at our original destination. */
776 struct edge_info
*edge_info
;
778 update_bb_profile_for_threading (e
->dest
, EDGE_FREQUENCY (e
),
779 e
->count
, taken_edge
);
783 edge_info
= allocate_edge_info (e
);
784 edge_info
->redirection_target
= taken_edge
;
785 bb_ann (e
->dest
)->incoming_edge_threaded
= true;
792 /* Initialize local stacks for this optimizer and record equivalences
793 upon entry to BB. Equivalences can come from the edge traversed to
794 reach BB or they may come from PHI nodes at the start of BB. */
797 dom_opt_initialize_block (struct dom_walk_data
*walk_data ATTRIBUTE_UNUSED
,
800 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
801 fprintf (dump_file
, "\n\nOptimizing block #%d\n\n", bb
->index
);
803 /* Push a marker on the stacks of local information so that we know how
804 far to unwind when we finalize this block. */
805 VEC_safe_push (tree_on_heap
, avail_exprs_stack
, NULL_TREE
);
806 VEC_safe_push (tree_on_heap
, block_defs_stack
, NULL_TREE
);
807 VEC_safe_push (tree_on_heap
, const_and_copies_stack
, NULL_TREE
);
808 VEC_safe_push (tree_on_heap
, nonzero_vars_stack
, NULL_TREE
);
809 VEC_safe_push (tree_on_heap
, vrp_variables_stack
, NULL_TREE
);
811 record_equivalences_from_incoming_edge (bb
);
813 /* PHI nodes can create equivalences too. */
814 record_equivalences_from_phis (bb
);
817 /* Given an expression EXPR (a relational expression or a statement),
818 initialize the hash table element pointed by by ELEMENT. */
821 initialize_hash_element (tree expr
, tree lhs
, struct expr_hash_elt
*element
)
823 /* Hash table elements may be based on conditional expressions or statements.
825 For the former case, we have no annotation and we want to hash the
826 conditional expression. In the latter case we have an annotation and
827 we want to record the expression the statement evaluates. */
828 if (COMPARISON_CLASS_P (expr
) || TREE_CODE (expr
) == TRUTH_NOT_EXPR
)
833 else if (TREE_CODE (expr
) == COND_EXPR
)
835 element
->ann
= stmt_ann (expr
);
836 element
->rhs
= COND_EXPR_COND (expr
);
838 else if (TREE_CODE (expr
) == SWITCH_EXPR
)
840 element
->ann
= stmt_ann (expr
);
841 element
->rhs
= SWITCH_COND (expr
);
843 else if (TREE_CODE (expr
) == RETURN_EXPR
&& TREE_OPERAND (expr
, 0))
845 element
->ann
= stmt_ann (expr
);
846 element
->rhs
= TREE_OPERAND (TREE_OPERAND (expr
, 0), 1);
850 element
->ann
= stmt_ann (expr
);
851 element
->rhs
= TREE_OPERAND (expr
, 1);
855 element
->hash
= avail_expr_hash (element
);
858 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
859 LIMIT entries left in LOCALs. */
862 remove_local_expressions_from_table (void)
864 /* Remove all the expressions made available in this block. */
865 while (VEC_length (tree_on_heap
, avail_exprs_stack
) > 0)
867 struct expr_hash_elt element
;
868 tree expr
= VEC_pop (tree_on_heap
, avail_exprs_stack
);
870 if (expr
== NULL_TREE
)
873 initialize_hash_element (expr
, NULL
, &element
);
874 htab_remove_elt_with_hash (avail_exprs
, &element
, element
.hash
);
878 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
879 state, stopping when there are LIMIT entries left in LOCALs. */
882 restore_nonzero_vars_to_original_value (void)
884 while (VEC_length (tree_on_heap
, nonzero_vars_stack
) > 0)
886 tree name
= VEC_pop (tree_on_heap
, nonzero_vars_stack
);
891 bitmap_clear_bit (nonzero_vars
, SSA_NAME_VERSION (name
));
895 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
896 CONST_AND_COPIES to its original state, stopping when we hit a
900 restore_vars_to_original_value (void)
902 while (VEC_length (tree_on_heap
, const_and_copies_stack
) > 0)
904 tree prev_value
, dest
;
906 dest
= VEC_pop (tree_on_heap
, const_and_copies_stack
);
911 prev_value
= VEC_pop (tree_on_heap
, const_and_copies_stack
);
912 SSA_NAME_VALUE (dest
) = prev_value
;
916 /* Similar to restore_vars_to_original_value, except that it restores
917 CURRDEFS to its original value. */
919 restore_currdefs_to_original_value (void)
921 /* Restore CURRDEFS to its original state. */
922 while (VEC_length (tree_on_heap
, block_defs_stack
) > 0)
924 tree tmp
= VEC_pop (tree_on_heap
, block_defs_stack
);
927 if (tmp
== NULL_TREE
)
930 /* If we recorded an SSA_NAME, then make the SSA_NAME the current
931 definition of its underlying variable. If we recorded anything
932 else, it must have been an _DECL node and its current reaching
933 definition must have been NULL. */
934 if (TREE_CODE (tmp
) == SSA_NAME
)
937 var
= SSA_NAME_VAR (saved_def
);
945 var_ann (var
)->current_def
= saved_def
;
949 /* We have finished processing the dominator children of BB, perform
950 any finalization actions in preparation for leaving this node in
951 the dominator tree. */
954 dom_opt_finalize_block (struct dom_walk_data
*walk_data
, basic_block bb
)
958 /* If we are at a leaf node in the dominator tree, see if we can thread
959 the edge from BB through its successor.
961 Do this before we remove entries from our equivalence tables. */
962 if (EDGE_COUNT (bb
->succs
) == 1
963 && (EDGE_SUCC (bb
, 0)->flags
& EDGE_ABNORMAL
) == 0
964 && (get_immediate_dominator (CDI_DOMINATORS
, EDGE_SUCC (bb
, 0)->dest
) != bb
965 || phi_nodes (EDGE_SUCC (bb
, 0)->dest
)))
968 thread_across_edge (walk_data
, EDGE_SUCC (bb
, 0));
970 else if ((last
= last_stmt (bb
))
971 && TREE_CODE (last
) == COND_EXPR
972 && (COMPARISON_CLASS_P (COND_EXPR_COND (last
))
973 || TREE_CODE (COND_EXPR_COND (last
)) == SSA_NAME
)
974 && EDGE_COUNT (bb
->succs
) == 2
975 && (EDGE_SUCC (bb
, 0)->flags
& EDGE_ABNORMAL
) == 0
976 && (EDGE_SUCC (bb
, 1)->flags
& EDGE_ABNORMAL
) == 0)
978 edge true_edge
, false_edge
;
980 extract_true_false_edges_from_block (bb
, &true_edge
, &false_edge
);
982 /* If the THEN arm is the end of a dominator tree or has PHI nodes,
983 then try to thread through its edge. */
984 if (get_immediate_dominator (CDI_DOMINATORS
, true_edge
->dest
) != bb
985 || phi_nodes (true_edge
->dest
))
987 struct edge_info
*edge_info
;
990 /* Push a marker onto the available expression stack so that we
991 unwind any expressions related to the TRUE arm before processing
992 the false arm below. */
993 VEC_safe_push (tree_on_heap
, avail_exprs_stack
, NULL_TREE
);
994 VEC_safe_push (tree_on_heap
, block_defs_stack
, NULL_TREE
);
995 VEC_safe_push (tree_on_heap
, const_and_copies_stack
, NULL_TREE
);
997 edge_info
= true_edge
->aux
;
999 /* If we have info associated with this edge, record it into
1000 our equivalency tables. */
1003 tree
*cond_equivalences
= edge_info
->cond_equivalences
;
1004 tree lhs
= edge_info
->lhs
;
1005 tree rhs
= edge_info
->rhs
;
1007 /* If we have a simple NAME = VALUE equivalency record it.
1008 Until the jump threading selection code improves, only
1009 do this if both the name and value are SSA_NAMEs with
1010 the same underlying variable to avoid missing threading
1013 && TREE_CODE (COND_EXPR_COND (last
)) == SSA_NAME
1014 && TREE_CODE (edge_info
->rhs
) == SSA_NAME
1015 && SSA_NAME_VAR (lhs
) == SSA_NAME_VAR (rhs
))
1016 record_const_or_copy (lhs
, rhs
);
1018 /* If we have 0 = COND or 1 = COND equivalences, record them
1019 into our expression hash tables. */
1020 if (cond_equivalences
)
1021 for (i
= 0; i
< edge_info
->max_cond_equivalences
; i
+= 2)
1023 tree expr
= cond_equivalences
[i
];
1024 tree value
= cond_equivalences
[i
+ 1];
1026 record_cond (expr
, value
);
1030 /* Now thread the edge. */
1031 thread_across_edge (walk_data
, true_edge
);
1033 /* And restore the various tables to their state before
1034 we threaded this edge. */
1035 remove_local_expressions_from_table ();
1036 restore_vars_to_original_value ();
1037 restore_currdefs_to_original_value ();
1040 /* Similarly for the ELSE arm. */
1041 if (get_immediate_dominator (CDI_DOMINATORS
, false_edge
->dest
) != bb
1042 || phi_nodes (false_edge
->dest
))
1044 struct edge_info
*edge_info
;
1047 edge_info
= false_edge
->aux
;
1049 /* If we have info associated with this edge, record it into
1050 our equivalency tables. */
1053 tree
*cond_equivalences
= edge_info
->cond_equivalences
;
1054 tree lhs
= edge_info
->lhs
;
1055 tree rhs
= edge_info
->rhs
;
1057 /* If we have a simple NAME = VALUE equivalency record it.
1058 Until the jump threading selection code improves, only
1059 do this if both the name and value are SSA_NAMEs with
1060 the same underlying variable to avoid missing threading
1063 && TREE_CODE (COND_EXPR_COND (last
)) == SSA_NAME
)
1064 record_const_or_copy (lhs
, rhs
);
1066 /* If we have 0 = COND or 1 = COND equivalences, record them
1067 into our expression hash tables. */
1068 if (cond_equivalences
)
1069 for (i
= 0; i
< edge_info
->max_cond_equivalences
; i
+= 2)
1071 tree expr
= cond_equivalences
[i
];
1072 tree value
= cond_equivalences
[i
+ 1];
1074 record_cond (expr
, value
);
1078 thread_across_edge (walk_data
, false_edge
);
1080 /* No need to remove local expressions from our tables
1081 or restore vars to their original value as that will
1082 be done immediately below. */
1086 remove_local_expressions_from_table ();
1087 restore_nonzero_vars_to_original_value ();
1088 restore_vars_to_original_value ();
1089 restore_currdefs_to_original_value ();
1091 /* Remove VRP records associated with this basic block. They are no
1094 To be efficient, we note which variables have had their values
1095 constrained in this block. So walk over each variable in the
1096 VRP_VARIABLEs array. */
1097 while (VEC_length (tree_on_heap
, vrp_variables_stack
) > 0)
1099 tree var
= VEC_pop (tree_on_heap
, vrp_variables_stack
);
1100 struct vrp_hash_elt vrp_hash_elt
, *vrp_hash_elt_p
;
1103 /* Each variable has a stack of value range records. We want to
1104 invalidate those associated with our basic block. So we walk
1105 the array backwards popping off records associated with our
1106 block. Once we hit a record not associated with our block
1108 varray_type var_vrp_records
;
1113 vrp_hash_elt
.var
= var
;
1114 vrp_hash_elt
.records
= NULL
;
1116 slot
= htab_find_slot (vrp_data
, &vrp_hash_elt
, NO_INSERT
);
1118 vrp_hash_elt_p
= (struct vrp_hash_elt
*) *slot
;
1119 var_vrp_records
= vrp_hash_elt_p
->records
;
1121 while (VARRAY_ACTIVE_SIZE (var_vrp_records
) > 0)
1123 struct vrp_element
*element
1124 = (struct vrp_element
*)VARRAY_TOP_GENERIC_PTR (var_vrp_records
);
1126 if (element
->bb
!= bb
)
1129 VARRAY_POP (var_vrp_records
);
1133 /* If we queued any statements to rescan in this block, then
1134 go ahead and rescan them now. */
1135 while (VEC_length (tree_on_heap
, stmts_to_rescan
) > 0)
1137 tree stmt
= VEC_last (tree_on_heap
, stmts_to_rescan
);
1138 basic_block stmt_bb
= bb_for_stmt (stmt
);
1143 VEC_pop (tree_on_heap
, stmts_to_rescan
);
1144 mark_new_vars_to_rename (stmt
, vars_to_rename
);
1148 /* PHI nodes can create equivalences too.
1150 Ignoring any alternatives which are the same as the result, if
1151 all the alternatives are equal, then the PHI node creates an
1154 Additionally, if all the PHI alternatives are known to have a nonzero
1155 value, then the result of this PHI is known to have a nonzero value,
1156 even if we do not know its exact value. */
1159 record_equivalences_from_phis (basic_block bb
)
1163 for (phi
= phi_nodes (bb
); phi
; phi
= PHI_CHAIN (phi
))
1165 tree lhs
= PHI_RESULT (phi
);
1169 for (i
= 0; i
< PHI_NUM_ARGS (phi
); i
++)
1171 tree t
= PHI_ARG_DEF (phi
, i
);
1173 /* Ignore alternatives which are the same as our LHS. Since
1174 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1175 can simply compare pointers. */
1179 /* If we have not processed an alternative yet, then set
1180 RHS to this alternative. */
1183 /* If we have processed an alternative (stored in RHS), then
1184 see if it is equal to this one. If it isn't, then stop
1186 else if (! operand_equal_for_phi_arg_p (rhs
, t
))
1190 /* If we had no interesting alternatives, then all the RHS alternatives
1191 must have been the same as LHS. */
1195 /* If we managed to iterate through each PHI alternative without
1196 breaking out of the loop, then we have a PHI which may create
1197 a useful equivalence. We do not need to record unwind data for
1198 this, since this is a true assignment and not an equivalence
1199 inferred from a comparison. All uses of this ssa name are dominated
1200 by this assignment, so unwinding just costs time and space. */
1201 if (i
== PHI_NUM_ARGS (phi
)
1202 && may_propagate_copy (lhs
, rhs
))
1203 SSA_NAME_VALUE (lhs
) = rhs
;
1205 /* Now see if we know anything about the nonzero property for the
1206 result of this PHI. */
1207 for (i
= 0; i
< PHI_NUM_ARGS (phi
); i
++)
1209 if (!PHI_ARG_NONZERO (phi
, i
))
1213 if (i
== PHI_NUM_ARGS (phi
))
1214 bitmap_set_bit (nonzero_vars
, SSA_NAME_VERSION (PHI_RESULT (phi
)));
1216 register_new_def (lhs
, &block_defs_stack
);
1220 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1221 return that edge. Otherwise return NULL. */
1223 single_incoming_edge_ignoring_loop_edges (basic_block bb
)
1229 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1231 /* A loop back edge can be identified by the destination of
1232 the edge dominating the source of the edge. */
1233 if (dominated_by_p (CDI_DOMINATORS
, e
->src
, e
->dest
))
1236 /* If we have already seen a non-loop edge, then we must have
1237 multiple incoming non-loop edges and thus we return NULL. */
1241 /* This is the first non-loop incoming edge we have found. Record
1249 /* Record any equivalences created by the incoming edge to BB. If BB
1250 has more than one incoming edge, then no equivalence is created. */
1253 record_equivalences_from_incoming_edge (basic_block bb
)
1257 struct edge_info
*edge_info
;
1259 /* If our parent block ended with a control statement, then we may be
1260 able to record some equivalences based on which outgoing edge from
1261 the parent was followed. */
1262 parent
= get_immediate_dominator (CDI_DOMINATORS
, bb
);
1264 e
= single_incoming_edge_ignoring_loop_edges (bb
);
1266 /* If we had a single incoming edge from our parent block, then enter
1267 any data associated with the edge into our tables. */
1268 if (e
&& e
->src
== parent
)
1276 tree lhs
= edge_info
->lhs
;
1277 tree rhs
= edge_info
->rhs
;
1278 tree
*cond_equivalences
= edge_info
->cond_equivalences
;
1281 record_equality (lhs
, rhs
);
1283 if (cond_equivalences
)
1285 bool recorded_range
= false;
1286 for (i
= 0; i
< edge_info
->max_cond_equivalences
; i
+= 2)
1288 tree expr
= cond_equivalences
[i
];
1289 tree value
= cond_equivalences
[i
+ 1];
1291 record_cond (expr
, value
);
1293 /* For the first true equivalence, record range
1294 information. We only do this for the first
1295 true equivalence as it should dominate any
1296 later true equivalences. */
1297 if (! recorded_range
1298 && COMPARISON_CLASS_P (expr
)
1299 && value
== boolean_true_node
1300 && TREE_CONSTANT (TREE_OPERAND (expr
, 1)))
1302 record_range (expr
, bb
);
1303 recorded_range
= true;
1311 /* Dump SSA statistics on FILE. */
1314 dump_dominator_optimization_stats (FILE *file
)
1318 fprintf (file
, "Total number of statements: %6ld\n\n",
1319 opt_stats
.num_stmts
);
1320 fprintf (file
, "Exprs considered for dominator optimizations: %6ld\n",
1321 opt_stats
.num_exprs_considered
);
1323 n_exprs
= opt_stats
.num_exprs_considered
;
1327 fprintf (file
, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1328 opt_stats
.num_re
, PERCENT (opt_stats
.num_re
,
1331 fprintf (file
, "\nHash table statistics:\n");
1333 fprintf (file
, " avail_exprs: ");
1334 htab_statistics (file
, avail_exprs
);
1338 /* Dump SSA statistics on stderr. */
1341 debug_dominator_optimization_stats (void)
1343 dump_dominator_optimization_stats (stderr
);
1347 /* Dump statistics for the hash table HTAB. */
1350 htab_statistics (FILE *file
, htab_t htab
)
1352 fprintf (file
, "size %ld, %ld elements, %f collision/search ratio\n",
1353 (long) htab_size (htab
),
1354 (long) htab_elements (htab
),
1355 htab_collisions (htab
));
1358 /* Record the fact that VAR has a nonzero value, though we may not know
1359 its exact value. Note that if VAR is already known to have a nonzero
1360 value, then we do nothing. */
1363 record_var_is_nonzero (tree var
)
1365 int indx
= SSA_NAME_VERSION (var
);
1367 if (bitmap_bit_p (nonzero_vars
, indx
))
1370 /* Mark it in the global table. */
1371 bitmap_set_bit (nonzero_vars
, indx
);
1373 /* Record this SSA_NAME so that we can reset the global table
1374 when we leave this block. */
1375 VEC_safe_push (tree_on_heap
, nonzero_vars_stack
, var
);
1378 /* Enter a statement into the true/false expression hash table indicating
1379 that the condition COND has the value VALUE. */
1382 record_cond (tree cond
, tree value
)
1384 struct expr_hash_elt
*element
= xmalloc (sizeof (struct expr_hash_elt
));
1387 initialize_hash_element (cond
, value
, element
);
1389 slot
= htab_find_slot_with_hash (avail_exprs
, (void *)element
,
1390 element
->hash
, INSERT
);
1393 *slot
= (void *) element
;
1394 VEC_safe_push (tree_on_heap
, avail_exprs_stack
, cond
);
1400 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1401 the new conditional into *p, then store a boolean_true_node
1405 build_and_record_new_cond (enum tree_code new_code
, tree op0
, tree op1
, tree
*p
)
1407 *p
= build2 (new_code
, boolean_type_node
, op0
, op1
);
1409 *p
= boolean_true_node
;
1412 /* Record that COND is true and INVERTED is false into the edge information
1413 structure. Also record that any conditions dominated by COND are true
1416 For example, if a < b is true, then a <= b must also be true. */
1419 record_conditions (struct edge_info
*edge_info
, tree cond
, tree inverted
)
1423 if (!COMPARISON_CLASS_P (cond
))
1426 op0
= TREE_OPERAND (cond
, 0);
1427 op1
= TREE_OPERAND (cond
, 1);
1429 switch (TREE_CODE (cond
))
1433 edge_info
->max_cond_equivalences
= 12;
1434 edge_info
->cond_equivalences
= xmalloc (12 * sizeof (tree
));
1435 build_and_record_new_cond ((TREE_CODE (cond
) == LT_EXPR
1436 ? LE_EXPR
: GE_EXPR
),
1437 op0
, op1
, &edge_info
->cond_equivalences
[4]);
1438 build_and_record_new_cond (ORDERED_EXPR
, op0
, op1
,
1439 &edge_info
->cond_equivalences
[6]);
1440 build_and_record_new_cond (NE_EXPR
, op0
, op1
,
1441 &edge_info
->cond_equivalences
[8]);
1442 build_and_record_new_cond (LTGT_EXPR
, op0
, op1
,
1443 &edge_info
->cond_equivalences
[10]);
1448 edge_info
->max_cond_equivalences
= 6;
1449 edge_info
->cond_equivalences
= xmalloc (6 * sizeof (tree
));
1450 build_and_record_new_cond (ORDERED_EXPR
, op0
, op1
,
1451 &edge_info
->cond_equivalences
[4]);
1455 edge_info
->max_cond_equivalences
= 10;
1456 edge_info
->cond_equivalences
= xmalloc (10 * sizeof (tree
));
1457 build_and_record_new_cond (ORDERED_EXPR
, op0
, op1
,
1458 &edge_info
->cond_equivalences
[4]);
1459 build_and_record_new_cond (LE_EXPR
, op0
, op1
,
1460 &edge_info
->cond_equivalences
[6]);
1461 build_and_record_new_cond (GE_EXPR
, op0
, op1
,
1462 &edge_info
->cond_equivalences
[8]);
1465 case UNORDERED_EXPR
:
1466 edge_info
->max_cond_equivalences
= 16;
1467 edge_info
->cond_equivalences
= xmalloc (16 * sizeof (tree
));
1468 build_and_record_new_cond (NE_EXPR
, op0
, op1
,
1469 &edge_info
->cond_equivalences
[4]);
1470 build_and_record_new_cond (UNLE_EXPR
, op0
, op1
,
1471 &edge_info
->cond_equivalences
[6]);
1472 build_and_record_new_cond (UNGE_EXPR
, op0
, op1
,
1473 &edge_info
->cond_equivalences
[8]);
1474 build_and_record_new_cond (UNEQ_EXPR
, op0
, op1
,
1475 &edge_info
->cond_equivalences
[10]);
1476 build_and_record_new_cond (UNLT_EXPR
, op0
, op1
,
1477 &edge_info
->cond_equivalences
[12]);
1478 build_and_record_new_cond (UNGT_EXPR
, op0
, op1
,
1479 &edge_info
->cond_equivalences
[14]);
1484 edge_info
->max_cond_equivalences
= 8;
1485 edge_info
->cond_equivalences
= xmalloc (8 * sizeof (tree
));
1486 build_and_record_new_cond ((TREE_CODE (cond
) == UNLT_EXPR
1487 ? UNLE_EXPR
: UNGE_EXPR
),
1488 op0
, op1
, &edge_info
->cond_equivalences
[4]);
1489 build_and_record_new_cond (NE_EXPR
, op0
, op1
,
1490 &edge_info
->cond_equivalences
[6]);
1494 edge_info
->max_cond_equivalences
= 8;
1495 edge_info
->cond_equivalences
= xmalloc (8 * sizeof (tree
));
1496 build_and_record_new_cond (UNLE_EXPR
, op0
, op1
,
1497 &edge_info
->cond_equivalences
[4]);
1498 build_and_record_new_cond (UNGE_EXPR
, op0
, op1
,
1499 &edge_info
->cond_equivalences
[6]);
1503 edge_info
->max_cond_equivalences
= 8;
1504 edge_info
->cond_equivalences
= xmalloc (8 * sizeof (tree
));
1505 build_and_record_new_cond (NE_EXPR
, op0
, op1
,
1506 &edge_info
->cond_equivalences
[4]);
1507 build_and_record_new_cond (ORDERED_EXPR
, op0
, op1
,
1508 &edge_info
->cond_equivalences
[6]);
1512 edge_info
->max_cond_equivalences
= 4;
1513 edge_info
->cond_equivalences
= xmalloc (4 * sizeof (tree
));
1517 /* Now store the original true and false conditions into the first
1519 edge_info
->cond_equivalences
[0] = cond
;
1520 edge_info
->cond_equivalences
[1] = boolean_true_node
;
1521 edge_info
->cond_equivalences
[2] = inverted
;
1522 edge_info
->cond_equivalences
[3] = boolean_false_node
;
1525 /* A helper function for record_const_or_copy and record_equality.
1526 Do the work of recording the value and undo info. */
1529 record_const_or_copy_1 (tree x
, tree y
, tree prev_x
)
1531 SSA_NAME_VALUE (x
) = y
;
1533 VEC_safe_push (tree_on_heap
, const_and_copies_stack
, prev_x
);
1534 VEC_safe_push (tree_on_heap
, const_and_copies_stack
, x
);
1538 /* Return the loop depth of the basic block of the defining statement of X.
1539 This number should not be treated as absolutely correct because the loop
1540 information may not be completely up-to-date when dom runs. However, it
1541 will be relatively correct, and as more passes are taught to keep loop info
1542 up to date, the result will become more and more accurate. */
1545 loop_depth_of_name (tree x
)
1550 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1551 if (TREE_CODE (x
) != SSA_NAME
)
1554 /* Otherwise return the loop depth of the defining statement's bb.
1555 Note that there may not actually be a bb for this statement, if the
1556 ssa_name is live on entry. */
1557 defstmt
= SSA_NAME_DEF_STMT (x
);
1558 defbb
= bb_for_stmt (defstmt
);
1562 return defbb
->loop_depth
;
1566 /* Record that X is equal to Y in const_and_copies. Record undo
1567 information in the block-local vector. */
1570 record_const_or_copy (tree x
, tree y
)
1572 tree prev_x
= SSA_NAME_VALUE (x
);
1574 if (TREE_CODE (y
) == SSA_NAME
)
1576 tree tmp
= SSA_NAME_VALUE (y
);
1581 record_const_or_copy_1 (x
, y
, prev_x
);
1584 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1585 This constrains the cases in which we may treat this as assignment. */
1588 record_equality (tree x
, tree y
)
1590 tree prev_x
= NULL
, prev_y
= NULL
;
1592 if (TREE_CODE (x
) == SSA_NAME
)
1593 prev_x
= SSA_NAME_VALUE (x
);
1594 if (TREE_CODE (y
) == SSA_NAME
)
1595 prev_y
= SSA_NAME_VALUE (y
);
1597 /* If one of the previous values is invariant, or invariant in more loops
1598 (by depth), then use that.
1599 Otherwise it doesn't matter which value we choose, just so
1600 long as we canonicalize on one value. */
1601 if (TREE_INVARIANT (y
))
1603 else if (TREE_INVARIANT (x
) || (loop_depth_of_name (x
) <= loop_depth_of_name (y
)))
1604 prev_x
= x
, x
= y
, y
= prev_x
, prev_x
= prev_y
;
1605 else if (prev_x
&& TREE_INVARIANT (prev_x
))
1606 x
= y
, y
= prev_x
, prev_x
= prev_y
;
1607 else if (prev_y
&& TREE_CODE (prev_y
) != VALUE_HANDLE
)
1610 /* After the swapping, we must have one SSA_NAME. */
1611 if (TREE_CODE (x
) != SSA_NAME
)
1614 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1615 variable compared against zero. If we're honoring signed zeros,
1616 then we cannot record this value unless we know that the value is
1618 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x
)))
1619 && (TREE_CODE (y
) != REAL_CST
1620 || REAL_VALUES_EQUAL (dconst0
, TREE_REAL_CST (y
))))
1623 record_const_or_copy_1 (x
, y
, prev_x
);
1626 /* Return true, if it is ok to do folding of an associative expression.
1627 EXP is the tree for the associative expression. */
1630 unsafe_associative_fp_binop (tree exp
)
1632 enum tree_code code
= TREE_CODE (exp
);
1633 return !(!flag_unsafe_math_optimizations
1634 && (code
== MULT_EXPR
|| code
== PLUS_EXPR
1635 || code
== MINUS_EXPR
)
1636 && FLOAT_TYPE_P (TREE_TYPE (exp
)));
1639 /* Returns true when STMT is a simple iv increment. It detects the
1640 following situation:
1642 i_1 = phi (..., i_2)
1643 i_2 = i_1 +/- ... */
1646 simple_iv_increment_p (tree stmt
)
1648 tree lhs
, rhs
, preinc
, phi
;
1651 if (TREE_CODE (stmt
) != MODIFY_EXPR
)
1654 lhs
= TREE_OPERAND (stmt
, 0);
1655 if (TREE_CODE (lhs
) != SSA_NAME
)
1658 rhs
= TREE_OPERAND (stmt
, 1);
1660 if (TREE_CODE (rhs
) != PLUS_EXPR
1661 && TREE_CODE (rhs
) != MINUS_EXPR
)
1664 preinc
= TREE_OPERAND (rhs
, 0);
1665 if (TREE_CODE (preinc
) != SSA_NAME
)
1668 phi
= SSA_NAME_DEF_STMT (preinc
);
1669 if (TREE_CODE (phi
) != PHI_NODE
)
1672 for (i
= 0; i
< (unsigned) PHI_NUM_ARGS (phi
); i
++)
1673 if (PHI_ARG_DEF (phi
, i
) == lhs
)
1679 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1680 hash tables. Try to simplify the RHS using whatever equivalences
1681 we may have recorded.
1683 If we are able to simplify the RHS, then lookup the simplified form in
1684 the hash table and return the result. Otherwise return NULL. */
1687 simplify_rhs_and_lookup_avail_expr (struct dom_walk_data
*walk_data
,
1688 tree stmt
, int insert
)
1690 tree rhs
= TREE_OPERAND (stmt
, 1);
1691 enum tree_code rhs_code
= TREE_CODE (rhs
);
1694 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1695 In which case we can change this statement to be lhs = y.
1696 Which can then be copy propagated.
1698 Similarly for negation. */
1699 if ((rhs_code
== BIT_NOT_EXPR
|| rhs_code
== NEGATE_EXPR
)
1700 && TREE_CODE (TREE_OPERAND (rhs
, 0)) == SSA_NAME
)
1702 /* Get the definition statement for our RHS. */
1703 tree rhs_def_stmt
= SSA_NAME_DEF_STMT (TREE_OPERAND (rhs
, 0));
1705 /* See if the RHS_DEF_STMT has the same form as our statement. */
1706 if (TREE_CODE (rhs_def_stmt
) == MODIFY_EXPR
1707 && TREE_CODE (TREE_OPERAND (rhs_def_stmt
, 1)) == rhs_code
)
1709 tree rhs_def_operand
;
1711 rhs_def_operand
= TREE_OPERAND (TREE_OPERAND (rhs_def_stmt
, 1), 0);
1713 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1714 if (TREE_CODE (rhs_def_operand
) == SSA_NAME
1715 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand
))
1716 result
= update_rhs_and_lookup_avail_expr (stmt
,
1722 /* If we have z = (x OP C1), see if we earlier had x = y OP C2.
1723 If OP is associative, create and fold (y OP C2) OP C1 which
1724 should result in (y OP C3), use that as the RHS for the
1725 assignment. Add minus to this, as we handle it specially below. */
1726 if ((associative_tree_code (rhs_code
) || rhs_code
== MINUS_EXPR
)
1727 && TREE_CODE (TREE_OPERAND (rhs
, 0)) == SSA_NAME
1728 && is_gimple_min_invariant (TREE_OPERAND (rhs
, 1)))
1730 tree rhs_def_stmt
= SSA_NAME_DEF_STMT (TREE_OPERAND (rhs
, 0));
1732 /* If the statement defines an induction variable, do not propagate
1733 its value, so that we do not create overlapping life ranges. */
1734 if (simple_iv_increment_p (rhs_def_stmt
))
1735 goto dont_fold_assoc
;
1737 /* See if the RHS_DEF_STMT has the same form as our statement. */
1738 if (TREE_CODE (rhs_def_stmt
) == MODIFY_EXPR
)
1740 tree rhs_def_rhs
= TREE_OPERAND (rhs_def_stmt
, 1);
1741 enum tree_code rhs_def_code
= TREE_CODE (rhs_def_rhs
);
1743 if ((rhs_code
== rhs_def_code
&& unsafe_associative_fp_binop (rhs
))
1744 || (rhs_code
== PLUS_EXPR
&& rhs_def_code
== MINUS_EXPR
)
1745 || (rhs_code
== MINUS_EXPR
&& rhs_def_code
== PLUS_EXPR
))
1747 tree def_stmt_op0
= TREE_OPERAND (rhs_def_rhs
, 0);
1748 tree def_stmt_op1
= TREE_OPERAND (rhs_def_rhs
, 1);
1750 if (TREE_CODE (def_stmt_op0
) == SSA_NAME
1751 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0
)
1752 && is_gimple_min_invariant (def_stmt_op1
))
1754 tree outer_const
= TREE_OPERAND (rhs
, 1);
1755 tree type
= TREE_TYPE (TREE_OPERAND (stmt
, 0));
1758 /* If we care about correct floating point results, then
1759 don't fold x + c1 - c2. Note that we need to take both
1760 the codes and the signs to figure this out. */
1761 if (FLOAT_TYPE_P (type
)
1762 && !flag_unsafe_math_optimizations
1763 && (rhs_def_code
== PLUS_EXPR
1764 || rhs_def_code
== MINUS_EXPR
))
1768 neg
^= (rhs_code
== MINUS_EXPR
);
1769 neg
^= (rhs_def_code
== MINUS_EXPR
);
1770 neg
^= real_isneg (TREE_REAL_CST_PTR (outer_const
));
1771 neg
^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1
));
1774 goto dont_fold_assoc
;
1777 /* Ho hum. So fold will only operate on the outermost
1778 thingy that we give it, so we have to build the new
1779 expression in two pieces. This requires that we handle
1780 combinations of plus and minus. */
1781 if (rhs_def_code
!= rhs_code
)
1783 if (rhs_def_code
== MINUS_EXPR
)
1784 t
= build (MINUS_EXPR
, type
, outer_const
, def_stmt_op1
);
1786 t
= build (MINUS_EXPR
, type
, def_stmt_op1
, outer_const
);
1787 rhs_code
= PLUS_EXPR
;
1789 else if (rhs_def_code
== MINUS_EXPR
)
1790 t
= build (PLUS_EXPR
, type
, def_stmt_op1
, outer_const
);
1792 t
= build (rhs_def_code
, type
, def_stmt_op1
, outer_const
);
1794 t
= build (rhs_code
, type
, def_stmt_op0
, t
);
1797 /* If the result is a suitable looking gimple expression,
1798 then use it instead of the original for STMT. */
1799 if (TREE_CODE (t
) == SSA_NAME
1800 || (UNARY_CLASS_P (t
)
1801 && TREE_CODE (TREE_OPERAND (t
, 0)) == SSA_NAME
)
1802 || ((BINARY_CLASS_P (t
) || COMPARISON_CLASS_P (t
))
1803 && TREE_CODE (TREE_OPERAND (t
, 0)) == SSA_NAME
1804 && is_gimple_val (TREE_OPERAND (t
, 1))))
1805 result
= update_rhs_and_lookup_avail_expr (stmt
, t
, insert
);
1812 /* Transform TRUNC_DIV_EXPR and TRUNC_MOD_EXPR into RSHIFT_EXPR
1813 and BIT_AND_EXPR respectively if the first operand is greater
1814 than zero and the second operand is an exact power of two. */
1815 if ((rhs_code
== TRUNC_DIV_EXPR
|| rhs_code
== TRUNC_MOD_EXPR
)
1816 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs
, 0)))
1817 && integer_pow2p (TREE_OPERAND (rhs
, 1)))
1820 tree op
= TREE_OPERAND (rhs
, 0);
1822 if (TYPE_UNSIGNED (TREE_TYPE (op
)))
1824 val
= integer_one_node
;
1828 tree dummy_cond
= walk_data
->global_data
;
1832 dummy_cond
= build (GT_EXPR
, boolean_type_node
,
1833 op
, integer_zero_node
);
1834 dummy_cond
= build (COND_EXPR
, void_type_node
,
1835 dummy_cond
, NULL
, NULL
);
1836 walk_data
->global_data
= dummy_cond
;
1840 TREE_SET_CODE (COND_EXPR_COND (dummy_cond
), GT_EXPR
);
1841 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 0) = op
;
1842 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 1)
1843 = integer_zero_node
;
1845 val
= simplify_cond_and_lookup_avail_expr (dummy_cond
, NULL
, false);
1848 if (val
&& integer_onep (val
))
1851 tree op0
= TREE_OPERAND (rhs
, 0);
1852 tree op1
= TREE_OPERAND (rhs
, 1);
1854 if (rhs_code
== TRUNC_DIV_EXPR
)
1855 t
= build (RSHIFT_EXPR
, TREE_TYPE (op0
), op0
,
1856 build_int_cst (NULL_TREE
, tree_log2 (op1
)));
1858 t
= build (BIT_AND_EXPR
, TREE_TYPE (op0
), op0
,
1859 local_fold (build (MINUS_EXPR
, TREE_TYPE (op1
),
1860 op1
, integer_one_node
)));
1862 result
= update_rhs_and_lookup_avail_expr (stmt
, t
, insert
);
1866 /* Transform ABS (X) into X or -X as appropriate. */
1867 if (rhs_code
== ABS_EXPR
1868 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs
, 0))))
1871 tree op
= TREE_OPERAND (rhs
, 0);
1872 tree type
= TREE_TYPE (op
);
1874 if (TYPE_UNSIGNED (type
))
1876 val
= integer_zero_node
;
1880 tree dummy_cond
= walk_data
->global_data
;
1884 dummy_cond
= build (LE_EXPR
, boolean_type_node
,
1885 op
, integer_zero_node
);
1886 dummy_cond
= build (COND_EXPR
, void_type_node
,
1887 dummy_cond
, NULL
, NULL
);
1888 walk_data
->global_data
= dummy_cond
;
1892 TREE_SET_CODE (COND_EXPR_COND (dummy_cond
), LE_EXPR
);
1893 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 0) = op
;
1894 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 1)
1895 = build_int_cst (type
, 0);
1897 val
= simplify_cond_and_lookup_avail_expr (dummy_cond
, NULL
, false);
1901 TREE_SET_CODE (COND_EXPR_COND (dummy_cond
), GE_EXPR
);
1902 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 0) = op
;
1903 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 1)
1904 = build_int_cst (type
, 0);
1906 val
= simplify_cond_and_lookup_avail_expr (dummy_cond
,
1911 if (integer_zerop (val
))
1912 val
= integer_one_node
;
1913 else if (integer_onep (val
))
1914 val
= integer_zero_node
;
1920 && (integer_onep (val
) || integer_zerop (val
)))
1924 if (integer_onep (val
))
1925 t
= build1 (NEGATE_EXPR
, TREE_TYPE (op
), op
);
1929 result
= update_rhs_and_lookup_avail_expr (stmt
, t
, insert
);
1933 /* Optimize *"foo" into 'f'. This is done here rather than
1934 in fold to avoid problems with stuff like &*"foo". */
1935 if (TREE_CODE (rhs
) == INDIRECT_REF
|| TREE_CODE (rhs
) == ARRAY_REF
)
1937 tree t
= fold_read_from_constant_string (rhs
);
1940 result
= update_rhs_and_lookup_avail_expr (stmt
, t
, insert
);
1946 /* COND is a condition of the form:
1948 x == const or x != const
1950 Look back to x's defining statement and see if x is defined as
1954 If const is unchanged if we convert it to type, then we can build
1955 the equivalent expression:
1958 y == const or y != const
1960 Which may allow further optimizations.
1962 Return the equivalent comparison or NULL if no such equivalent comparison
1966 find_equivalent_equality_comparison (tree cond
)
1968 tree op0
= TREE_OPERAND (cond
, 0);
1969 tree op1
= TREE_OPERAND (cond
, 1);
1970 tree def_stmt
= SSA_NAME_DEF_STMT (op0
);
1972 /* OP0 might have been a parameter, so first make sure it
1973 was defined by a MODIFY_EXPR. */
1974 if (def_stmt
&& TREE_CODE (def_stmt
) == MODIFY_EXPR
)
1976 tree def_rhs
= TREE_OPERAND (def_stmt
, 1);
1978 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1979 if ((TREE_CODE (def_rhs
) == NOP_EXPR
1980 || TREE_CODE (def_rhs
) == CONVERT_EXPR
)
1981 && TREE_CODE (TREE_OPERAND (def_rhs
, 0)) == SSA_NAME
)
1983 tree def_rhs_inner
= TREE_OPERAND (def_rhs
, 0);
1984 tree def_rhs_inner_type
= TREE_TYPE (def_rhs_inner
);
1987 if (TYPE_PRECISION (def_rhs_inner_type
)
1988 > TYPE_PRECISION (TREE_TYPE (def_rhs
)))
1991 /* What we want to prove is that if we convert OP1 to
1992 the type of the object inside the NOP_EXPR that the
1993 result is still equivalent to SRC.
1995 If that is true, the build and return new equivalent
1996 condition which uses the source of the typecast and the
1997 new constant (which has only changed its type). */
1998 new = build1 (TREE_CODE (def_rhs
), def_rhs_inner_type
, op1
);
1999 new = local_fold (new);
2000 if (is_gimple_val (new) && tree_int_cst_equal (new, op1
))
2001 return build (TREE_CODE (cond
), TREE_TYPE (cond
),
2002 def_rhs_inner
, new);
2008 /* STMT is a COND_EXPR for which we could not trivially determine its
2009 result. This routine attempts to find equivalent forms of the
2010 condition which we may be able to optimize better. It also
2011 uses simple value range propagation to optimize conditionals. */
2014 simplify_cond_and_lookup_avail_expr (tree stmt
,
2018 tree cond
= COND_EXPR_COND (stmt
);
2020 if (COMPARISON_CLASS_P (cond
))
2022 tree op0
= TREE_OPERAND (cond
, 0);
2023 tree op1
= TREE_OPERAND (cond
, 1);
2025 if (TREE_CODE (op0
) == SSA_NAME
&& is_gimple_min_invariant (op1
))
2028 tree low
, high
, cond_low
, cond_high
;
2029 int lowequal
, highequal
, swapped
, no_overlap
, subset
, cond_inverted
;
2030 varray_type vrp_records
;
2031 struct vrp_element
*element
;
2032 struct vrp_hash_elt vrp_hash_elt
, *vrp_hash_elt_p
;
2035 /* First see if we have test of an SSA_NAME against a constant
2036 where the SSA_NAME is defined by an earlier typecast which
2037 is irrelevant when performing tests against the given
2039 if (TREE_CODE (cond
) == EQ_EXPR
|| TREE_CODE (cond
) == NE_EXPR
)
2041 tree new_cond
= find_equivalent_equality_comparison (cond
);
2045 /* Update the statement to use the new equivalent
2047 COND_EXPR_COND (stmt
) = new_cond
;
2049 /* If this is not a real stmt, ann will be NULL and we
2050 avoid processing the operands. */
2054 /* Lookup the condition and return its known value if it
2056 new_cond
= lookup_avail_expr (stmt
, insert
);
2060 /* The operands have changed, so update op0 and op1. */
2061 op0
= TREE_OPERAND (cond
, 0);
2062 op1
= TREE_OPERAND (cond
, 1);
2066 /* Consult the value range records for this variable (if they exist)
2067 to see if we can eliminate or simplify this conditional.
2069 Note two tests are necessary to determine no records exist.
2070 First we have to see if the virtual array exists, if it
2071 exists, then we have to check its active size.
2073 Also note the vast majority of conditionals are not testing
2074 a variable which has had its range constrained by an earlier
2075 conditional. So this filter avoids a lot of unnecessary work. */
2076 vrp_hash_elt
.var
= op0
;
2077 vrp_hash_elt
.records
= NULL
;
2078 slot
= htab_find_slot (vrp_data
, &vrp_hash_elt
, NO_INSERT
);
2082 vrp_hash_elt_p
= (struct vrp_hash_elt
*) *slot
;
2083 vrp_records
= vrp_hash_elt_p
->records
;
2084 if (vrp_records
== NULL
)
2087 limit
= VARRAY_ACTIVE_SIZE (vrp_records
);
2089 /* If we have no value range records for this variable, or we are
2090 unable to extract a range for this condition, then there is
2093 || ! extract_range_from_cond (cond
, &cond_high
,
2094 &cond_low
, &cond_inverted
))
2097 /* We really want to avoid unnecessary computations of range
2098 info. So all ranges are computed lazily; this avoids a
2099 lot of unnecessary work. i.e., we record the conditional,
2100 but do not process how it constrains the variable's
2101 potential values until we know that processing the condition
2104 However, we do not want to have to walk a potentially long
2105 list of ranges, nor do we want to compute a variable's
2106 range more than once for a given path.
2108 Luckily, each time we encounter a conditional that can not
2109 be otherwise optimized we will end up here and we will
2110 compute the necessary range information for the variable
2111 used in this condition.
2113 Thus you can conclude that there will never be more than one
2114 conditional associated with a variable which has not been
2115 processed. So we never need to merge more than one new
2116 conditional into the current range.
2118 These properties also help us avoid unnecessary work. */
2120 = (struct vrp_element
*)VARRAY_GENERIC_PTR (vrp_records
, limit
- 1);
2122 if (element
->high
&& element
->low
)
2124 /* The last element has been processed, so there is no range
2125 merging to do, we can simply use the high/low values
2126 recorded in the last element. */
2128 high
= element
->high
;
2132 tree tmp_high
, tmp_low
;
2135 /* The last element has not been processed. Process it now.
2136 record_range should ensure for cond inverted is not set.
2137 This call can only fail if cond is x < min or x > max,
2138 which fold should have optimized into false.
2139 If that doesn't happen, just pretend all values are
2141 if (! extract_range_from_cond (element
->cond
, &tmp_high
,
2145 gcc_assert (dummy
== 0);
2147 /* If this is the only element, then no merging is necessary,
2148 the high/low values from extract_range_from_cond are all
2157 /* Get the high/low value from the previous element. */
2158 struct vrp_element
*prev
2159 = (struct vrp_element
*)VARRAY_GENERIC_PTR (vrp_records
,
2164 /* Merge in this element's range with the range from the
2167 The low value for the merged range is the maximum of
2168 the previous low value and the low value of this record.
2170 Similarly the high value for the merged range is the
2171 minimum of the previous high value and the high value of
2173 low
= (tree_int_cst_compare (low
, tmp_low
) == 1
2175 high
= (tree_int_cst_compare (high
, tmp_high
) == -1
2179 /* And record the computed range. */
2181 element
->high
= high
;
2185 /* After we have constrained this variable's potential values,
2186 we try to determine the result of the given conditional.
2188 To simplify later tests, first determine if the current
2189 low value is the same low value as the conditional.
2190 Similarly for the current high value and the high value
2191 for the conditional. */
2192 lowequal
= tree_int_cst_equal (low
, cond_low
);
2193 highequal
= tree_int_cst_equal (high
, cond_high
);
2195 if (lowequal
&& highequal
)
2196 return (cond_inverted
? boolean_false_node
: boolean_true_node
);
2198 /* To simplify the overlap/subset tests below we may want
2199 to swap the two ranges so that the larger of the two
2200 ranges occurs "first". */
2202 if (tree_int_cst_compare (low
, cond_low
) == 1
2204 && tree_int_cst_compare (cond_high
, high
) == 1))
2217 /* Now determine if there is no overlap in the ranges
2218 or if the second range is a subset of the first range. */
2219 no_overlap
= tree_int_cst_lt (high
, cond_low
);
2220 subset
= tree_int_cst_compare (cond_high
, high
) != 1;
2222 /* If there was no overlap in the ranges, then this conditional
2223 always has a false value (unless we had to invert this
2224 conditional, in which case it always has a true value). */
2226 return (cond_inverted
? boolean_true_node
: boolean_false_node
);
2228 /* If the current range is a subset of the condition's range,
2229 then this conditional always has a true value (unless we
2230 had to invert this conditional, in which case it always
2231 has a true value). */
2232 if (subset
&& swapped
)
2233 return (cond_inverted
? boolean_false_node
: boolean_true_node
);
2235 /* We were unable to determine the result of the conditional.
2236 However, we may be able to simplify the conditional. First
2237 merge the ranges in the same manner as range merging above. */
2238 low
= tree_int_cst_compare (low
, cond_low
) == 1 ? low
: cond_low
;
2239 high
= tree_int_cst_compare (high
, cond_high
) == -1 ? high
: cond_high
;
2241 /* If the range has converged to a single point, then turn this
2242 into an equality comparison. */
2243 if (TREE_CODE (cond
) != EQ_EXPR
2244 && TREE_CODE (cond
) != NE_EXPR
2245 && tree_int_cst_equal (low
, high
))
2247 TREE_SET_CODE (cond
, EQ_EXPR
);
2248 TREE_OPERAND (cond
, 1) = high
;
2255 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2256 result. This routine attempts to find equivalent forms of the
2257 condition which we may be able to optimize better. */
2260 simplify_switch_and_lookup_avail_expr (tree stmt
, int insert
)
2262 tree cond
= SWITCH_COND (stmt
);
2265 /* The optimization that we really care about is removing unnecessary
2266 casts. That will let us do much better in propagating the inferred
2267 constant at the switch target. */
2268 if (TREE_CODE (cond
) == SSA_NAME
)
2270 def
= SSA_NAME_DEF_STMT (cond
);
2271 if (TREE_CODE (def
) == MODIFY_EXPR
)
2273 def
= TREE_OPERAND (def
, 1);
2274 if (TREE_CODE (def
) == NOP_EXPR
)
2279 def
= TREE_OPERAND (def
, 0);
2281 #ifdef ENABLE_CHECKING
2282 /* ??? Why was Jeff testing this? We are gimple... */
2283 gcc_assert (is_gimple_val (def
));
2286 to
= TREE_TYPE (cond
);
2287 ti
= TREE_TYPE (def
);
2289 /* If we have an extension that preserves value, then we
2290 can copy the source value into the switch. */
2292 need_precision
= TYPE_PRECISION (ti
);
2294 if (TYPE_UNSIGNED (to
) && !TYPE_UNSIGNED (ti
))
2296 else if (!TYPE_UNSIGNED (to
) && TYPE_UNSIGNED (ti
))
2297 need_precision
+= 1;
2298 if (TYPE_PRECISION (to
) < need_precision
)
2303 SWITCH_COND (stmt
) = def
;
2306 return lookup_avail_expr (stmt
, insert
);
2316 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2317 known value for that SSA_NAME (or NULL if no value is known).
2319 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2320 even if we don't know their precise value.
2322 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2323 nodes of the successors of BB. */
2326 cprop_into_successor_phis (basic_block bb
, bitmap nonzero_vars
)
2331 /* This can get rather expensive if the implementation is naive in
2332 how it finds the phi alternative associated with a particular edge. */
2333 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2338 /* If this is an abnormal edge, then we do not want to copy propagate
2339 into the PHI alternative associated with this edge. */
2340 if (e
->flags
& EDGE_ABNORMAL
)
2343 phi
= phi_nodes (e
->dest
);
2348 for ( ; phi
; phi
= PHI_CHAIN (phi
))
2351 use_operand_p orig_p
;
2354 /* The alternative may be associated with a constant, so verify
2355 it is an SSA_NAME before doing anything with it. */
2356 orig_p
= PHI_ARG_DEF_PTR (phi
, indx
);
2357 orig
= USE_FROM_PTR (orig_p
);
2358 if (TREE_CODE (orig
) != SSA_NAME
)
2361 /* If the alternative is known to have a nonzero value, record
2362 that fact in the PHI node itself for future use. */
2363 if (bitmap_bit_p (nonzero_vars
, SSA_NAME_VERSION (orig
)))
2364 PHI_ARG_NONZERO (phi
, indx
) = true;
2366 /* If we have *ORIG_P in our constant/copy table, then replace
2367 ORIG_P with its value in our constant/copy table. */
2368 new = SSA_NAME_VALUE (orig
);
2370 && (TREE_CODE (new) == SSA_NAME
2371 || is_gimple_min_invariant (new))
2372 && may_propagate_copy (orig
, new))
2374 propagate_value (orig_p
, new);
2380 /* We have finished optimizing BB, record any information implied by
2381 taking a specific outgoing edge from BB. */
2384 record_edge_info (basic_block bb
)
2386 block_stmt_iterator bsi
= bsi_last (bb
);
2387 struct edge_info
*edge_info
;
2389 if (! bsi_end_p (bsi
))
2391 tree stmt
= bsi_stmt (bsi
);
2393 if (stmt
&& TREE_CODE (stmt
) == SWITCH_EXPR
)
2395 tree cond
= SWITCH_COND (stmt
);
2397 if (TREE_CODE (cond
) == SSA_NAME
)
2399 tree labels
= SWITCH_LABELS (stmt
);
2400 int i
, n_labels
= TREE_VEC_LENGTH (labels
);
2401 tree
*info
= xcalloc (n_basic_blocks
, sizeof (tree
));
2405 for (i
= 0; i
< n_labels
; i
++)
2407 tree label
= TREE_VEC_ELT (labels
, i
);
2408 basic_block target_bb
= label_to_block (CASE_LABEL (label
));
2410 if (CASE_HIGH (label
)
2411 || !CASE_LOW (label
)
2412 || info
[target_bb
->index
])
2413 info
[target_bb
->index
] = error_mark_node
;
2415 info
[target_bb
->index
] = label
;
2418 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2420 basic_block target_bb
= e
->dest
;
2421 tree node
= info
[target_bb
->index
];
2423 if (node
!= NULL
&& node
!= error_mark_node
)
2425 tree x
= fold_convert (TREE_TYPE (cond
), CASE_LOW (node
));
2426 edge_info
= allocate_edge_info (e
);
2427 edge_info
->lhs
= cond
;
2435 /* A COND_EXPR may create equivalences too. */
2436 if (stmt
&& TREE_CODE (stmt
) == COND_EXPR
)
2438 tree cond
= COND_EXPR_COND (stmt
);
2442 extract_true_false_edges_from_block (bb
, &true_edge
, &false_edge
);
2444 /* If the conditional is a single variable 'X', record 'X = 1'
2445 for the true edge and 'X = 0' on the false edge. */
2446 if (SSA_VAR_P (cond
))
2448 struct edge_info
*edge_info
;
2450 edge_info
= allocate_edge_info (true_edge
);
2451 edge_info
->lhs
= cond
;
2452 edge_info
->rhs
= constant_boolean_node (1, TREE_TYPE (cond
));
2454 edge_info
= allocate_edge_info (false_edge
);
2455 edge_info
->lhs
= cond
;
2456 edge_info
->rhs
= constant_boolean_node (0, TREE_TYPE (cond
));
2458 /* Equality tests may create one or two equivalences. */
2459 else if (COMPARISON_CLASS_P (cond
))
2461 tree op0
= TREE_OPERAND (cond
, 0);
2462 tree op1
= TREE_OPERAND (cond
, 1);
2464 /* Special case comparing booleans against a constant as we
2465 know the value of OP0 on both arms of the branch. i.e., we
2466 can record an equivalence for OP0 rather than COND. */
2467 if ((TREE_CODE (cond
) == EQ_EXPR
|| TREE_CODE (cond
) == NE_EXPR
)
2468 && TREE_CODE (op0
) == SSA_NAME
2469 && TREE_CODE (TREE_TYPE (op0
)) == BOOLEAN_TYPE
2470 && is_gimple_min_invariant (op1
))
2472 if (TREE_CODE (cond
) == EQ_EXPR
)
2474 edge_info
= allocate_edge_info (true_edge
);
2475 edge_info
->lhs
= op0
;
2476 edge_info
->rhs
= (integer_zerop (op1
)
2477 ? boolean_false_node
2478 : boolean_true_node
);
2480 edge_info
= allocate_edge_info (false_edge
);
2481 edge_info
->lhs
= op0
;
2482 edge_info
->rhs
= (integer_zerop (op1
)
2484 : boolean_false_node
);
2488 edge_info
= allocate_edge_info (true_edge
);
2489 edge_info
->lhs
= op0
;
2490 edge_info
->rhs
= (integer_zerop (op1
)
2492 : boolean_false_node
);
2494 edge_info
= allocate_edge_info (false_edge
);
2495 edge_info
->lhs
= op0
;
2496 edge_info
->rhs
= (integer_zerop (op1
)
2497 ? boolean_false_node
2498 : boolean_true_node
);
2502 else if (is_gimple_min_invariant (op0
)
2503 && (TREE_CODE (op1
) == SSA_NAME
2504 || is_gimple_min_invariant (op1
)))
2506 tree inverted
= invert_truthvalue (cond
);
2507 struct edge_info
*edge_info
;
2509 edge_info
= allocate_edge_info (true_edge
);
2510 record_conditions (edge_info
, cond
, inverted
);
2512 if (TREE_CODE (cond
) == EQ_EXPR
)
2514 edge_info
->lhs
= op1
;
2515 edge_info
->rhs
= op0
;
2518 edge_info
= allocate_edge_info (false_edge
);
2519 record_conditions (edge_info
, inverted
, cond
);
2521 if (TREE_CODE (cond
) == NE_EXPR
)
2523 edge_info
->lhs
= op1
;
2524 edge_info
->rhs
= op0
;
2528 else if (TREE_CODE (op0
) == SSA_NAME
2529 && (is_gimple_min_invariant (op1
)
2530 || TREE_CODE (op1
) == SSA_NAME
))
2532 tree inverted
= invert_truthvalue (cond
);
2533 struct edge_info
*edge_info
;
2535 edge_info
= allocate_edge_info (true_edge
);
2536 record_conditions (edge_info
, cond
, inverted
);
2538 if (TREE_CODE (cond
) == EQ_EXPR
)
2540 edge_info
->lhs
= op0
;
2541 edge_info
->rhs
= op1
;
2544 edge_info
= allocate_edge_info (false_edge
);
2545 record_conditions (edge_info
, inverted
, cond
);
2547 if (TREE_CODE (cond
) == NE_EXPR
)
2549 edge_info
->lhs
= op0
;
2550 edge_info
->rhs
= op1
;
2555 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2560 /* Propagate information from BB to its outgoing edges.
2562 This can include equivalency information implied by control statements
2563 at the end of BB and const/copy propagation into PHIs in BB's
2564 successor blocks. */
2567 propagate_to_outgoing_edges (struct dom_walk_data
*walk_data ATTRIBUTE_UNUSED
,
2571 record_edge_info (bb
);
2572 cprop_into_successor_phis (bb
, nonzero_vars
);
2575 /* Search for redundant computations in STMT. If any are found, then
2576 replace them with the variable holding the result of the computation.
2578 If safe, record this expression into the available expression hash
2582 eliminate_redundant_computations (struct dom_walk_data
*walk_data
,
2583 tree stmt
, stmt_ann_t ann
)
2585 v_may_def_optype v_may_defs
= V_MAY_DEF_OPS (ann
);
2586 tree
*expr_p
, def
= NULL_TREE
;
2589 bool retval
= false;
2591 if (TREE_CODE (stmt
) == MODIFY_EXPR
)
2592 def
= TREE_OPERAND (stmt
, 0);
2594 /* Certain expressions on the RHS can be optimized away, but can not
2595 themselves be entered into the hash tables. */
2596 if (ann
->makes_aliased_stores
2598 || TREE_CODE (def
) != SSA_NAME
2599 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def
)
2600 || NUM_V_MAY_DEFS (v_may_defs
) != 0
2601 /* Do not record equivalences for increments of ivs. This would create
2602 overlapping live ranges for a very questionable gain. */
2603 || simple_iv_increment_p (stmt
))
2606 /* Check if the expression has been computed before. */
2607 cached_lhs
= lookup_avail_expr (stmt
, insert
);
2609 /* If this is an assignment and the RHS was not in the hash table,
2610 then try to simplify the RHS and lookup the new RHS in the
2612 if (! cached_lhs
&& TREE_CODE (stmt
) == MODIFY_EXPR
)
2613 cached_lhs
= simplify_rhs_and_lookup_avail_expr (walk_data
, stmt
, insert
);
2614 /* Similarly if this is a COND_EXPR and we did not find its
2615 expression in the hash table, simplify the condition and
2617 else if (! cached_lhs
&& TREE_CODE (stmt
) == COND_EXPR
)
2618 cached_lhs
= simplify_cond_and_lookup_avail_expr (stmt
, ann
, insert
);
2619 /* Similarly for a SWITCH_EXPR. */
2620 else if (!cached_lhs
&& TREE_CODE (stmt
) == SWITCH_EXPR
)
2621 cached_lhs
= simplify_switch_and_lookup_avail_expr (stmt
, insert
);
2623 opt_stats
.num_exprs_considered
++;
2625 /* Get a pointer to the expression we are trying to optimize. */
2626 if (TREE_CODE (stmt
) == COND_EXPR
)
2627 expr_p
= &COND_EXPR_COND (stmt
);
2628 else if (TREE_CODE (stmt
) == SWITCH_EXPR
)
2629 expr_p
= &SWITCH_COND (stmt
);
2630 else if (TREE_CODE (stmt
) == RETURN_EXPR
&& TREE_OPERAND (stmt
, 0))
2631 expr_p
= &TREE_OPERAND (TREE_OPERAND (stmt
, 0), 1);
2633 expr_p
= &TREE_OPERAND (stmt
, 1);
2635 /* It is safe to ignore types here since we have already done
2636 type checking in the hashing and equality routines. In fact
2637 type checking here merely gets in the way of constant
2638 propagation. Also, make sure that it is safe to propagate
2639 CACHED_LHS into *EXPR_P. */
2641 && (TREE_CODE (cached_lhs
) != SSA_NAME
2642 || may_propagate_copy (*expr_p
, cached_lhs
)))
2644 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2646 fprintf (dump_file
, " Replaced redundant expr '");
2647 print_generic_expr (dump_file
, *expr_p
, dump_flags
);
2648 fprintf (dump_file
, "' with '");
2649 print_generic_expr (dump_file
, cached_lhs
, dump_flags
);
2650 fprintf (dump_file
, "'\n");
2655 #if defined ENABLE_CHECKING
2656 gcc_assert (TREE_CODE (cached_lhs
) == SSA_NAME
2657 || is_gimple_min_invariant (cached_lhs
));
2660 if (TREE_CODE (cached_lhs
) == ADDR_EXPR
2661 || (POINTER_TYPE_P (TREE_TYPE (*expr_p
))
2662 && is_gimple_min_invariant (cached_lhs
)))
2665 propagate_tree_value (expr_p
, cached_lhs
);
2671 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2672 the available expressions table or the const_and_copies table.
2673 Detect and record those equivalences. */
2676 record_equivalences_from_stmt (tree stmt
,
2680 tree lhs
= TREE_OPERAND (stmt
, 0);
2681 enum tree_code lhs_code
= TREE_CODE (lhs
);
2684 if (lhs_code
== SSA_NAME
)
2686 tree rhs
= TREE_OPERAND (stmt
, 1);
2688 /* Strip away any useless type conversions. */
2689 STRIP_USELESS_TYPE_CONVERSION (rhs
);
2691 /* If the RHS of the assignment is a constant or another variable that
2692 may be propagated, register it in the CONST_AND_COPIES table. We
2693 do not need to record unwind data for this, since this is a true
2694 assignment and not an equivalence inferred from a comparison. All
2695 uses of this ssa name are dominated by this assignment, so unwinding
2696 just costs time and space. */
2698 && (TREE_CODE (rhs
) == SSA_NAME
2699 || is_gimple_min_invariant (rhs
)))
2700 SSA_NAME_VALUE (lhs
) = rhs
;
2702 /* alloca never returns zero and the address of a non-weak symbol
2703 is never zero. NOP_EXPRs and CONVERT_EXPRs can be completely
2704 stripped as they do not affect this equivalence. */
2705 while (TREE_CODE (rhs
) == NOP_EXPR
2706 || TREE_CODE (rhs
) == CONVERT_EXPR
)
2707 rhs
= TREE_OPERAND (rhs
, 0);
2709 if (alloca_call_p (rhs
)
2710 || (TREE_CODE (rhs
) == ADDR_EXPR
2711 && DECL_P (TREE_OPERAND (rhs
, 0))
2712 && ! DECL_WEAK (TREE_OPERAND (rhs
, 0))))
2713 record_var_is_nonzero (lhs
);
2715 /* IOR of any value with a nonzero value will result in a nonzero
2716 value. Even if we do not know the exact result recording that
2717 the result is nonzero is worth the effort. */
2718 if (TREE_CODE (rhs
) == BIT_IOR_EXPR
2719 && integer_nonzerop (TREE_OPERAND (rhs
, 1)))
2720 record_var_is_nonzero (lhs
);
2723 /* Look at both sides for pointer dereferences. If we find one, then
2724 the pointer must be nonnull and we can enter that equivalence into
2726 if (flag_delete_null_pointer_checks
)
2727 for (i
= 0; i
< 2; i
++)
2729 tree t
= TREE_OPERAND (stmt
, i
);
2731 /* Strip away any COMPONENT_REFs. */
2732 while (TREE_CODE (t
) == COMPONENT_REF
)
2733 t
= TREE_OPERAND (t
, 0);
2735 /* Now see if this is a pointer dereference. */
2736 if (INDIRECT_REF_P (t
))
2738 tree op
= TREE_OPERAND (t
, 0);
2740 /* If the pointer is a SSA variable, then enter new
2741 equivalences into the hash table. */
2742 while (TREE_CODE (op
) == SSA_NAME
)
2744 tree def
= SSA_NAME_DEF_STMT (op
);
2746 record_var_is_nonzero (op
);
2748 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2749 which are known to have a nonzero value. */
2751 && TREE_CODE (def
) == MODIFY_EXPR
2752 && TREE_CODE (TREE_OPERAND (def
, 1)) == NOP_EXPR
)
2753 op
= TREE_OPERAND (TREE_OPERAND (def
, 1), 0);
2760 /* A memory store, even an aliased store, creates a useful
2761 equivalence. By exchanging the LHS and RHS, creating suitable
2762 vops and recording the result in the available expression table,
2763 we may be able to expose more redundant loads. */
2764 if (!ann
->has_volatile_ops
2765 && (TREE_CODE (TREE_OPERAND (stmt
, 1)) == SSA_NAME
2766 || is_gimple_min_invariant (TREE_OPERAND (stmt
, 1)))
2767 && !is_gimple_reg (lhs
))
2769 tree rhs
= TREE_OPERAND (stmt
, 1);
2772 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2773 is a constant, we need to adjust the constant to fit into the
2774 type of the LHS. If the LHS is a bitfield and the RHS is not
2775 a constant, then we can not record any equivalences for this
2776 statement since we would need to represent the widening or
2777 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2778 and should not be necessary if GCC represented bitfields
2780 if (lhs_code
== COMPONENT_REF
2781 && DECL_BIT_FIELD (TREE_OPERAND (lhs
, 1)))
2783 if (TREE_CONSTANT (rhs
))
2784 rhs
= widen_bitfield (rhs
, TREE_OPERAND (lhs
, 1), lhs
);
2788 /* If the value overflowed, then we can not use this equivalence. */
2789 if (rhs
&& ! is_gimple_min_invariant (rhs
))
2795 /* Build a new statement with the RHS and LHS exchanged. */
2796 new = build (MODIFY_EXPR
, TREE_TYPE (stmt
), rhs
, lhs
);
2798 create_ssa_artficial_load_stmt (&(ann
->operands
), new);
2800 /* Finally enter the statement into the available expression
2802 lookup_avail_expr (new, true);
2807 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2808 CONST_AND_COPIES. */
2811 cprop_operand (tree stmt
, use_operand_p op_p
)
2813 bool may_have_exposed_new_symbols
= false;
2815 tree op
= USE_FROM_PTR (op_p
);
2817 /* If the operand has a known constant value or it is known to be a
2818 copy of some other variable, use the value or copy stored in
2819 CONST_AND_COPIES. */
2820 val
= SSA_NAME_VALUE (op
);
2821 if (val
&& TREE_CODE (val
) != VALUE_HANDLE
)
2823 tree op_type
, val_type
;
2825 /* Do not change the base variable in the virtual operand
2826 tables. That would make it impossible to reconstruct
2827 the renamed virtual operand if we later modify this
2828 statement. Also only allow the new value to be an SSA_NAME
2829 for propagation into virtual operands. */
2830 if (!is_gimple_reg (op
)
2831 && (get_virtual_var (val
) != get_virtual_var (op
)
2832 || TREE_CODE (val
) != SSA_NAME
))
2835 /* Do not replace hard register operands in asm statements. */
2836 if (TREE_CODE (stmt
) == ASM_EXPR
2837 && !may_propagate_copy_into_asm (op
))
2840 /* Get the toplevel type of each operand. */
2841 op_type
= TREE_TYPE (op
);
2842 val_type
= TREE_TYPE (val
);
2844 /* While both types are pointers, get the type of the object
2846 while (POINTER_TYPE_P (op_type
) && POINTER_TYPE_P (val_type
))
2848 op_type
= TREE_TYPE (op_type
);
2849 val_type
= TREE_TYPE (val_type
);
2852 /* Make sure underlying types match before propagating a constant by
2853 converting the constant to the proper type. Note that convert may
2854 return a non-gimple expression, in which case we ignore this
2855 propagation opportunity. */
2856 if (TREE_CODE (val
) != SSA_NAME
)
2858 if (!lang_hooks
.types_compatible_p (op_type
, val_type
))
2860 val
= fold_convert (TREE_TYPE (op
), val
);
2861 if (!is_gimple_min_invariant (val
))
2866 /* Certain operands are not allowed to be copy propagated due
2867 to their interaction with exception handling and some GCC
2869 else if (!may_propagate_copy (op
, val
))
2872 /* Do not propagate copies if the propagated value is at a deeper loop
2873 depth than the propagatee. Otherwise, this may move loop variant
2874 variables outside of their loops and prevent coalescing
2875 opportunities. If the value was loop invariant, it will be hoisted
2876 by LICM and exposed for copy propagation. */
2877 if (loop_depth_of_name (val
) > loop_depth_of_name (op
))
2881 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2883 fprintf (dump_file
, " Replaced '");
2884 print_generic_expr (dump_file
, op
, dump_flags
);
2885 fprintf (dump_file
, "' with %s '",
2886 (TREE_CODE (val
) != SSA_NAME
? "constant" : "variable"));
2887 print_generic_expr (dump_file
, val
, dump_flags
);
2888 fprintf (dump_file
, "'\n");
2891 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2892 that we may have exposed a new symbol for SSA renaming. */
2893 if (TREE_CODE (val
) == ADDR_EXPR
2894 || (POINTER_TYPE_P (TREE_TYPE (op
))
2895 && is_gimple_min_invariant (val
)))
2896 may_have_exposed_new_symbols
= true;
2898 propagate_value (op_p
, val
);
2900 /* And note that we modified this statement. This is now
2901 safe, even if we changed virtual operands since we will
2902 rescan the statement and rewrite its operands again. */
2905 return may_have_exposed_new_symbols
;
2908 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2909 known value for that SSA_NAME (or NULL if no value is known).
2911 Propagate values from CONST_AND_COPIES into the uses, vuses and
2912 v_may_def_ops of STMT. */
2915 cprop_into_stmt (tree stmt
)
2917 bool may_have_exposed_new_symbols
= false;
2922 FOR_EACH_SSA_USE_OPERAND (op_p
, stmt
, iter
, SSA_OP_ALL_USES
)
2924 if (TREE_CODE (USE_FROM_PTR (op_p
)) == SSA_NAME
)
2925 may_have_exposed_new_symbols
|= cprop_operand (stmt
, op_p
);
2928 if (may_have_exposed_new_symbols
)
2930 rhs
= get_rhs (stmt
);
2931 if (rhs
&& TREE_CODE (rhs
) == ADDR_EXPR
)
2932 recompute_tree_invarant_for_addr_expr (rhs
);
2935 return may_have_exposed_new_symbols
;
2939 /* Optimize the statement pointed by iterator SI.
2941 We try to perform some simplistic global redundancy elimination and
2942 constant propagation:
2944 1- To detect global redundancy, we keep track of expressions that have
2945 been computed in this block and its dominators. If we find that the
2946 same expression is computed more than once, we eliminate repeated
2947 computations by using the target of the first one.
2949 2- Constant values and copy assignments. This is used to do very
2950 simplistic constant and copy propagation. When a constant or copy
2951 assignment is found, we map the value on the RHS of the assignment to
2952 the variable in the LHS in the CONST_AND_COPIES table. */
2955 optimize_stmt (struct dom_walk_data
*walk_data
, basic_block bb
,
2956 block_stmt_iterator si
)
2960 bool may_optimize_p
;
2961 bool may_have_exposed_new_symbols
= false;
2963 stmt
= bsi_stmt (si
);
2965 get_stmt_operands (stmt
);
2966 ann
= stmt_ann (stmt
);
2967 opt_stats
.num_stmts
++;
2968 may_have_exposed_new_symbols
= false;
2970 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2972 fprintf (dump_file
, "Optimizing statement ");
2973 print_generic_stmt (dump_file
, stmt
, TDF_SLIM
);
2976 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2977 may_have_exposed_new_symbols
= cprop_into_stmt (stmt
);
2979 /* If the statement has been modified with constant replacements,
2980 fold its RHS before checking for redundant computations. */
2983 /* Try to fold the statement making sure that STMT is kept
2985 if (fold_stmt (bsi_stmt_ptr (si
)))
2987 stmt
= bsi_stmt (si
);
2988 ann
= stmt_ann (stmt
);
2990 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2992 fprintf (dump_file
, " Folded to: ");
2993 print_generic_stmt (dump_file
, stmt
, TDF_SLIM
);
2997 /* Constant/copy propagation above may change the set of
2998 virtual operands associated with this statement. Folding
2999 may remove the need for some virtual operands.
3001 Indicate we will need to rescan and rewrite the statement. */
3002 may_have_exposed_new_symbols
= true;
3005 /* Check for redundant computations. Do this optimization only
3006 for assignments that have no volatile ops and conditionals. */
3007 may_optimize_p
= (!ann
->has_volatile_ops
3008 && ((TREE_CODE (stmt
) == RETURN_EXPR
3009 && TREE_OPERAND (stmt
, 0)
3010 && TREE_CODE (TREE_OPERAND (stmt
, 0)) == MODIFY_EXPR
3011 && ! (TREE_SIDE_EFFECTS
3012 (TREE_OPERAND (TREE_OPERAND (stmt
, 0), 1))))
3013 || (TREE_CODE (stmt
) == MODIFY_EXPR
3014 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt
, 1)))
3015 || TREE_CODE (stmt
) == COND_EXPR
3016 || TREE_CODE (stmt
) == SWITCH_EXPR
));
3019 may_have_exposed_new_symbols
3020 |= eliminate_redundant_computations (walk_data
, stmt
, ann
);
3022 /* Record any additional equivalences created by this statement. */
3023 if (TREE_CODE (stmt
) == MODIFY_EXPR
)
3024 record_equivalences_from_stmt (stmt
,
3028 register_definitions_for_stmt (stmt
);
3030 /* If STMT is a COND_EXPR and it was modified, then we may know
3031 where it goes. If that is the case, then mark the CFG as altered.
3033 This will cause us to later call remove_unreachable_blocks and
3034 cleanup_tree_cfg when it is safe to do so. It is not safe to
3035 clean things up here since removal of edges and such can trigger
3036 the removal of PHI nodes, which in turn can release SSA_NAMEs to
3039 That's all fine and good, except that once SSA_NAMEs are released
3040 to the manager, we must not call create_ssa_name until all references
3041 to released SSA_NAMEs have been eliminated.
3043 All references to the deleted SSA_NAMEs can not be eliminated until
3044 we remove unreachable blocks.
3046 We can not remove unreachable blocks until after we have completed
3047 any queued jump threading.
3049 We can not complete any queued jump threads until we have taken
3050 appropriate variables out of SSA form. Taking variables out of
3051 SSA form can call create_ssa_name and thus we lose.
3053 Ultimately I suspect we're going to need to change the interface
3054 into the SSA_NAME manager. */
3060 if (TREE_CODE (stmt
) == COND_EXPR
)
3061 val
= COND_EXPR_COND (stmt
);
3062 else if (TREE_CODE (stmt
) == SWITCH_EXPR
)
3063 val
= SWITCH_COND (stmt
);
3065 if (val
&& TREE_CODE (val
) == INTEGER_CST
&& find_taken_edge (bb
, val
))
3068 /* If we simplified a statement in such a way as to be shown that it
3069 cannot trap, update the eh information and the cfg to match. */
3070 if (maybe_clean_eh_stmt (stmt
))
3072 bitmap_set_bit (need_eh_cleanup
, bb
->index
);
3073 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3074 fprintf (dump_file
, " Flagged to clear EH edges.\n");
3078 if (may_have_exposed_new_symbols
)
3079 VEC_safe_push (tree_on_heap
, stmts_to_rescan
, bsi_stmt (si
));
3082 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
3083 available expression hashtable, then return the LHS from the hash
3086 If INSERT is true, then we also update the available expression
3087 hash table to account for the changes made to STMT. */
3090 update_rhs_and_lookup_avail_expr (tree stmt
, tree new_rhs
, bool insert
)
3092 tree cached_lhs
= NULL
;
3094 /* Remove the old entry from the hash table. */
3097 struct expr_hash_elt element
;
3099 initialize_hash_element (stmt
, NULL
, &element
);
3100 htab_remove_elt_with_hash (avail_exprs
, &element
, element
.hash
);
3103 /* Now update the RHS of the assignment. */
3104 TREE_OPERAND (stmt
, 1) = new_rhs
;
3106 /* Now lookup the updated statement in the hash table. */
3107 cached_lhs
= lookup_avail_expr (stmt
, insert
);
3109 /* We have now called lookup_avail_expr twice with two different
3110 versions of this same statement, once in optimize_stmt, once here.
3112 We know the call in optimize_stmt did not find an existing entry
3113 in the hash table, so a new entry was created. At the same time
3114 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
3116 If this call failed to find an existing entry on the hash table,
3117 then the new version of this statement was entered into the
3118 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3119 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3121 If this call succeeded, we still have one copy of this statement
3122 on the BLOCK_AVAIL_EXPRs vector.
3124 For both cases, we need to pop the most recent entry off the
3125 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3126 statement in the hash tables, that will leave precisely one
3127 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3128 we found a copy of this statement in the second hash table lookup
3129 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3131 VEC_pop (tree_on_heap
, avail_exprs_stack
);
3133 /* And make sure we record the fact that we modified this
3140 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3141 found, return its LHS. Otherwise insert STMT in the table and return
3144 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3145 is also added to the stack pointed by BLOCK_AVAIL_EXPRS_P, so that they
3146 can be removed when we finish processing this block and its children.
3148 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3149 contains no CALL_EXPR on its RHS and makes no volatile nor
3150 aliased references. */
3153 lookup_avail_expr (tree stmt
, bool insert
)
3158 struct expr_hash_elt
*element
= xmalloc (sizeof (struct expr_hash_elt
));
3160 lhs
= TREE_CODE (stmt
) == MODIFY_EXPR
? TREE_OPERAND (stmt
, 0) : NULL
;
3162 initialize_hash_element (stmt
, lhs
, element
);
3164 /* Don't bother remembering constant assignments and copy operations.
3165 Constants and copy operations are handled by the constant/copy propagator
3166 in optimize_stmt. */
3167 if (TREE_CODE (element
->rhs
) == SSA_NAME
3168 || is_gimple_min_invariant (element
->rhs
))
3174 /* If this is an equality test against zero, see if we have recorded a
3175 nonzero value for the variable in question. */
3176 if ((TREE_CODE (element
->rhs
) == EQ_EXPR
3177 || TREE_CODE (element
->rhs
) == NE_EXPR
)
3178 && TREE_CODE (TREE_OPERAND (element
->rhs
, 0)) == SSA_NAME
3179 && integer_zerop (TREE_OPERAND (element
->rhs
, 1)))
3181 int indx
= SSA_NAME_VERSION (TREE_OPERAND (element
->rhs
, 0));
3183 if (bitmap_bit_p (nonzero_vars
, indx
))
3185 tree t
= element
->rhs
;
3188 if (TREE_CODE (t
) == EQ_EXPR
)
3189 return boolean_false_node
;
3191 return boolean_true_node
;
3195 /* Finally try to find the expression in the main expression hash table. */
3196 slot
= htab_find_slot_with_hash (avail_exprs
, element
, element
->hash
,
3197 (insert
? INSERT
: NO_INSERT
));
3206 *slot
= (void *) element
;
3207 VEC_safe_push (tree_on_heap
, avail_exprs_stack
,
3208 stmt
? stmt
: element
->rhs
);
3212 /* Extract the LHS of the assignment so that it can be used as the current
3213 definition of another variable. */
3214 lhs
= ((struct expr_hash_elt
*)*slot
)->lhs
;
3216 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3217 use the value from the const_and_copies table. */
3218 if (TREE_CODE (lhs
) == SSA_NAME
)
3220 temp
= SSA_NAME_VALUE (lhs
);
3221 if (temp
&& TREE_CODE (temp
) != VALUE_HANDLE
)
3229 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3230 range of values that result in the conditional having a true value.
3232 Return true if we are successful in extracting a range from COND and
3233 false if we are unsuccessful. */
3236 extract_range_from_cond (tree cond
, tree
*hi_p
, tree
*lo_p
, int *inverted_p
)
3238 tree op1
= TREE_OPERAND (cond
, 1);
3239 tree high
, low
, type
;
3242 type
= TREE_TYPE (op1
);
3244 /* Experiments have shown that it's rarely, if ever useful to
3245 record ranges for enumerations. Presumably this is due to
3246 the fact that they're rarely used directly. They are typically
3247 cast into an integer type and used that way. */
3248 if (TREE_CODE (type
) != INTEGER_TYPE
3249 /* We don't know how to deal with types with variable bounds. */
3250 || TREE_CODE (TYPE_MIN_VALUE (type
)) != INTEGER_CST
3251 || TREE_CODE (TYPE_MAX_VALUE (type
)) != INTEGER_CST
)
3254 switch (TREE_CODE (cond
))
3268 high
= TYPE_MAX_VALUE (type
);
3273 high
= TYPE_MAX_VALUE (type
);
3274 if (!tree_int_cst_lt (op1
, high
))
3276 low
= int_const_binop (PLUS_EXPR
, op1
, integer_one_node
, 1);
3282 low
= TYPE_MIN_VALUE (type
);
3287 low
= TYPE_MIN_VALUE (type
);
3288 if (!tree_int_cst_lt (low
, op1
))
3290 high
= int_const_binop (MINUS_EXPR
, op1
, integer_one_node
, 1);
3300 *inverted_p
= inverted
;
3304 /* Record a range created by COND for basic block BB. */
3307 record_range (tree cond
, basic_block bb
)
3309 enum tree_code code
= TREE_CODE (cond
);
3311 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3312 They rarely allow for meaningful range optimizations and significantly
3313 complicate the implementation. */
3314 if ((code
== LT_EXPR
|| code
== LE_EXPR
|| code
== GT_EXPR
3315 || code
== GE_EXPR
|| code
== EQ_EXPR
)
3316 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond
, 1))) == INTEGER_TYPE
)
3318 struct vrp_hash_elt
*vrp_hash_elt
;
3319 struct vrp_element
*element
;
3320 varray_type
*vrp_records_p
;
3324 vrp_hash_elt
= xmalloc (sizeof (struct vrp_hash_elt
));
3325 vrp_hash_elt
->var
= TREE_OPERAND (cond
, 0);
3326 vrp_hash_elt
->records
= NULL
;
3327 slot
= htab_find_slot (vrp_data
, vrp_hash_elt
, INSERT
);
3330 *slot
= (void *) vrp_hash_elt
;
3332 free (vrp_hash_elt
);
3334 vrp_hash_elt
= (struct vrp_hash_elt
*) *slot
;
3335 vrp_records_p
= &vrp_hash_elt
->records
;
3337 element
= ggc_alloc (sizeof (struct vrp_element
));
3338 element
->low
= NULL
;
3339 element
->high
= NULL
;
3340 element
->cond
= cond
;
3343 if (*vrp_records_p
== NULL
)
3344 VARRAY_GENERIC_PTR_INIT (*vrp_records_p
, 2, "vrp records");
3346 VARRAY_PUSH_GENERIC_PTR (*vrp_records_p
, element
);
3347 VEC_safe_push (tree_on_heap
, vrp_variables_stack
, TREE_OPERAND (cond
, 0));
3351 /* Hashing and equality functions for VRP_DATA.
3353 Since this hash table is addressed by SSA_NAMEs, we can hash on
3354 their version number and equality can be determined with a
3355 pointer comparison. */
3358 vrp_hash (const void *p
)
3360 tree var
= ((struct vrp_hash_elt
*)p
)->var
;
3362 return SSA_NAME_VERSION (var
);
3366 vrp_eq (const void *p1
, const void *p2
)
3368 tree var1
= ((struct vrp_hash_elt
*)p1
)->var
;
3369 tree var2
= ((struct vrp_hash_elt
*)p2
)->var
;
3371 return var1
== var2
;
3374 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3375 MODIFY_EXPR statements. We compute a value number for expressions using
3376 the code of the expression and the SSA numbers of its operands. */
3379 avail_expr_hash (const void *p
)
3381 stmt_ann_t ann
= ((struct expr_hash_elt
*)p
)->ann
;
3382 tree rhs
= ((struct expr_hash_elt
*)p
)->rhs
;
3387 /* iterative_hash_expr knows how to deal with any expression and
3388 deals with commutative operators as well, so just use it instead
3389 of duplicating such complexities here. */
3390 val
= iterative_hash_expr (rhs
, val
);
3392 /* If the hash table entry is not associated with a statement, then we
3393 can just hash the expression and not worry about virtual operands
3398 /* Add the SSA version numbers of every vuse operand. This is important
3399 because compound variables like arrays are not renamed in the
3400 operands. Rather, the rename is done on the virtual variable
3401 representing all the elements of the array. */
3402 vuses
= VUSE_OPS (ann
);
3403 for (i
= 0; i
< NUM_VUSES (vuses
); i
++)
3404 val
= iterative_hash_expr (VUSE_OP (vuses
, i
), val
);
3410 real_avail_expr_hash (const void *p
)
3412 return ((const struct expr_hash_elt
*)p
)->hash
;
3416 avail_expr_eq (const void *p1
, const void *p2
)
3418 stmt_ann_t ann1
= ((struct expr_hash_elt
*)p1
)->ann
;
3419 tree rhs1
= ((struct expr_hash_elt
*)p1
)->rhs
;
3420 stmt_ann_t ann2
= ((struct expr_hash_elt
*)p2
)->ann
;
3421 tree rhs2
= ((struct expr_hash_elt
*)p2
)->rhs
;
3423 /* If they are the same physical expression, return true. */
3424 if (rhs1
== rhs2
&& ann1
== ann2
)
3427 /* If their codes are not equal, then quit now. */
3428 if (TREE_CODE (rhs1
) != TREE_CODE (rhs2
))
3431 /* In case of a collision, both RHS have to be identical and have the
3432 same VUSE operands. */
3433 if ((TREE_TYPE (rhs1
) == TREE_TYPE (rhs2
)
3434 || lang_hooks
.types_compatible_p (TREE_TYPE (rhs1
), TREE_TYPE (rhs2
)))
3435 && operand_equal_p (rhs1
, rhs2
, OEP_PURE_SAME
))
3437 vuse_optype ops1
= NULL
;
3438 vuse_optype ops2
= NULL
;
3439 size_t num_ops1
= 0;
3440 size_t num_ops2
= 0;
3445 ops1
= VUSE_OPS (ann1
);
3446 num_ops1
= NUM_VUSES (ops1
);
3451 ops2
= VUSE_OPS (ann2
);
3452 num_ops2
= NUM_VUSES (ops2
);
3455 /* If the number of virtual uses is different, then we consider
3457 if (num_ops1
!= num_ops2
)
3460 for (i
= 0; i
< num_ops1
; i
++)
3461 if (VUSE_OP (ops1
, i
) != VUSE_OP (ops2
, i
))
3464 gcc_assert (((struct expr_hash_elt
*)p1
)->hash
3465 == ((struct expr_hash_elt
*)p2
)->hash
);
3472 /* Given STMT and a pointer to the block local definitions BLOCK_DEFS_P,
3473 register register all objects set by this statement into BLOCK_DEFS_P
3477 register_definitions_for_stmt (tree stmt
)
3482 FOR_EACH_SSA_TREE_OPERAND (def
, stmt
, iter
, SSA_OP_ALL_DEFS
)
3485 /* FIXME: We shouldn't be registering new defs if the variable
3486 doesn't need to be renamed. */
3487 register_new_def (def
, &block_defs_stack
);