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
);
447 free_dominance_info (CDI_DOMINATORS
);
448 cfg_altered
= cleanup_tree_cfg ();
449 calculate_dominance_info (CDI_DOMINATORS
);
451 rewrite_ssa_into_ssa ();
453 /* Reinitialize the various tables. */
454 bitmap_clear (nonzero_vars
);
455 htab_empty (avail_exprs
);
456 htab_empty (vrp_data
);
458 for (i
= 0; i
< num_referenced_vars
; i
++)
459 var_ann (referenced_var (i
))->current_def
= NULL
;
461 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
463 This must be done before we iterate as we might have a
464 reference to an SSA_NAME which was removed by the call to
465 rewrite_ssa_into_ssa.
467 Long term we will be able to let everything in SSA_NAME_VALUE
468 persist. However, for now, we know this is the safe thing to do. */
469 for (i
= 0; i
< num_ssa_names
; i
++)
471 tree name
= ssa_name (i
);
477 value
= SSA_NAME_VALUE (name
);
478 if (value
&& !is_gimple_min_invariant (value
))
479 SSA_NAME_VALUE (name
) = NULL
;
482 while (optimize
> 1 && cfg_altered
);
484 /* Debugging dumps. */
485 if (dump_file
&& (dump_flags
& TDF_STATS
))
486 dump_dominator_optimization_stats (dump_file
);
488 /* We emptied the hash table earlier, now delete it completely. */
489 htab_delete (avail_exprs
);
490 htab_delete (vrp_data
);
492 /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
493 CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
494 of the do-while loop above. */
496 /* And finalize the dominator walker. */
497 fini_walk_dominator_tree (&walk_data
);
499 /* Free nonzero_vars. */
500 BITMAP_FREE (nonzero_vars
);
501 BITMAP_FREE (need_eh_cleanup
);
503 VEC_free (tree_on_heap
, block_defs_stack
);
504 VEC_free (tree_on_heap
, avail_exprs_stack
);
505 VEC_free (tree_on_heap
, const_and_copies_stack
);
506 VEC_free (tree_on_heap
, nonzero_vars_stack
);
507 VEC_free (tree_on_heap
, vrp_variables_stack
);
508 VEC_free (tree_on_heap
, stmts_to_rescan
);
512 gate_dominator (void)
514 return flag_tree_dom
!= 0;
517 struct tree_opt_pass pass_dominator
=
520 gate_dominator
, /* gate */
521 tree_ssa_dominator_optimize
, /* execute */
524 0, /* static_pass_number */
525 TV_TREE_SSA_DOMINATOR_OPTS
, /* tv_id */
526 PROP_cfg
| PROP_ssa
| PROP_alias
, /* properties_required */
527 0, /* properties_provided */
528 0, /* properties_destroyed */
529 0, /* todo_flags_start */
530 TODO_dump_func
| TODO_rename_vars
531 | TODO_verify_ssa
, /* todo_flags_finish */
536 /* We are exiting BB, see if the target block begins with a conditional
537 jump which has a known value when reached via BB. */
540 thread_across_edge (struct dom_walk_data
*walk_data
, edge e
)
542 block_stmt_iterator bsi
;
546 /* Each PHI creates a temporary equivalence, record them. */
547 for (phi
= phi_nodes (e
->dest
); phi
; phi
= PHI_CHAIN (phi
))
549 tree src
= PHI_ARG_DEF_FROM_EDGE (phi
, e
);
550 tree dst
= PHI_RESULT (phi
);
552 /* If the desired argument is not the same as this PHI's result
553 and it is set by a PHI in this block, then we can not thread
554 through this block. */
556 && TREE_CODE (src
) == SSA_NAME
557 && TREE_CODE (SSA_NAME_DEF_STMT (src
)) == PHI_NODE
558 && bb_for_stmt (SSA_NAME_DEF_STMT (src
)) == e
->dest
)
561 record_const_or_copy (dst
, src
);
562 register_new_def (dst
, &block_defs_stack
);
565 for (bsi
= bsi_start (e
->dest
); ! bsi_end_p (bsi
); bsi_next (&bsi
))
567 tree lhs
, cached_lhs
;
569 stmt
= bsi_stmt (bsi
);
571 /* Ignore empty statements and labels. */
572 if (IS_EMPTY_STMT (stmt
) || TREE_CODE (stmt
) == LABEL_EXPR
)
575 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
576 value, then stop our search here. Ideally when we stop a
577 search we stop on a COND_EXPR or SWITCH_EXPR. */
578 if (TREE_CODE (stmt
) != MODIFY_EXPR
579 || TREE_CODE (TREE_OPERAND (stmt
, 0)) != SSA_NAME
)
582 /* At this point we have a statement which assigns an RHS to an
583 SSA_VAR on the LHS. We want to prove that the RHS is already
584 available and that its value is held in the current definition
585 of the LHS -- meaning that this assignment is a NOP when
586 reached via edge E. */
587 if (TREE_CODE (TREE_OPERAND (stmt
, 1)) == SSA_NAME
)
588 cached_lhs
= TREE_OPERAND (stmt
, 1);
590 cached_lhs
= lookup_avail_expr (stmt
, false);
592 lhs
= TREE_OPERAND (stmt
, 0);
594 /* This can happen if we thread around to the start of a loop. */
595 if (lhs
== cached_lhs
)
598 /* If we did not find RHS in the hash table, then try again after
599 temporarily const/copy propagating the operands. */
602 /* Copy the operands. */
603 stmt_ann_t ann
= stmt_ann (stmt
);
604 use_optype uses
= USE_OPS (ann
);
605 vuse_optype vuses
= VUSE_OPS (ann
);
606 tree
*uses_copy
= xmalloc (NUM_USES (uses
) * sizeof (tree
));
607 tree
*vuses_copy
= xmalloc (NUM_VUSES (vuses
) * sizeof (tree
));
610 /* Make a copy of the uses into USES_COPY, then cprop into
612 for (i
= 0; i
< NUM_USES (uses
); i
++)
616 uses_copy
[i
] = USE_OP (uses
, i
);
617 if (TREE_CODE (USE_OP (uses
, i
)) == SSA_NAME
)
618 tmp
= SSA_NAME_VALUE (USE_OP (uses
, i
));
619 if (tmp
&& TREE_CODE (tmp
) != VALUE_HANDLE
)
620 SET_USE_OP (uses
, i
, tmp
);
623 /* Similarly for virtual uses. */
624 for (i
= 0; i
< NUM_VUSES (vuses
); i
++)
628 vuses_copy
[i
] = VUSE_OP (vuses
, i
);
629 if (TREE_CODE (VUSE_OP (vuses
, i
)) == SSA_NAME
)
630 tmp
= SSA_NAME_VALUE (VUSE_OP (vuses
, i
));
631 if (tmp
&& TREE_CODE (tmp
) != VALUE_HANDLE
)
632 SET_VUSE_OP (vuses
, i
, tmp
);
635 /* Try to lookup the new expression. */
636 cached_lhs
= lookup_avail_expr (stmt
, false);
638 /* Restore the statement's original uses/defs. */
639 for (i
= 0; i
< NUM_USES (uses
); i
++)
640 SET_USE_OP (uses
, i
, uses_copy
[i
]);
642 for (i
= 0; i
< NUM_VUSES (vuses
); i
++)
643 SET_VUSE_OP (vuses
, i
, vuses_copy
[i
]);
648 /* If we still did not find the expression in the hash table,
649 then we can not ignore this statement. */
654 /* If the expression in the hash table was not assigned to an
655 SSA_NAME, then we can not ignore this statement. */
656 if (TREE_CODE (cached_lhs
) != SSA_NAME
)
659 /* If we have different underlying variables, then we can not
660 ignore this statement. */
661 if (SSA_NAME_VAR (cached_lhs
) != SSA_NAME_VAR (lhs
))
664 /* If CACHED_LHS does not represent the current value of the underlying
665 variable in CACHED_LHS/LHS, then we can not ignore this statement. */
666 if (var_ann (SSA_NAME_VAR (lhs
))->current_def
!= cached_lhs
)
669 /* If we got here, then we can ignore this statement and continue
670 walking through the statements in the block looking for a threadable
673 We want to record an equivalence lhs = cache_lhs so that if
674 the result of this statement is used later we can copy propagate
676 record_const_or_copy (lhs
, cached_lhs
);
677 register_new_def (lhs
, &block_defs_stack
);
680 /* If we stopped at a COND_EXPR or SWITCH_EXPR, then see if we know which
681 arm will be taken. */
683 && (TREE_CODE (stmt
) == COND_EXPR
684 || TREE_CODE (stmt
) == SWITCH_EXPR
))
686 tree cond
, cached_lhs
;
688 /* Now temporarily cprop the operands and try to find the resulting
689 expression in the hash tables. */
690 if (TREE_CODE (stmt
) == COND_EXPR
)
691 cond
= COND_EXPR_COND (stmt
);
693 cond
= SWITCH_COND (stmt
);
695 if (COMPARISON_CLASS_P (cond
))
697 tree dummy_cond
, op0
, op1
;
698 enum tree_code cond_code
;
700 op0
= TREE_OPERAND (cond
, 0);
701 op1
= TREE_OPERAND (cond
, 1);
702 cond_code
= TREE_CODE (cond
);
704 /* Get the current value of both operands. */
705 if (TREE_CODE (op0
) == SSA_NAME
)
707 tree tmp
= SSA_NAME_VALUE (op0
);
708 if (tmp
&& TREE_CODE (tmp
) != VALUE_HANDLE
)
712 if (TREE_CODE (op1
) == SSA_NAME
)
714 tree tmp
= SSA_NAME_VALUE (op1
);
715 if (tmp
&& TREE_CODE (tmp
) != VALUE_HANDLE
)
719 /* Stuff the operator and operands into our dummy conditional
720 expression, creating the dummy conditional if necessary. */
721 dummy_cond
= walk_data
->global_data
;
724 dummy_cond
= build (cond_code
, boolean_type_node
, op0
, op1
);
725 dummy_cond
= build (COND_EXPR
, void_type_node
,
726 dummy_cond
, NULL
, NULL
);
727 walk_data
->global_data
= dummy_cond
;
731 TREE_SET_CODE (COND_EXPR_COND (dummy_cond
), cond_code
);
732 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 0) = op0
;
733 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 1) = op1
;
736 /* If the conditional folds to an invariant, then we are done,
737 otherwise look it up in the hash tables. */
738 cached_lhs
= local_fold (COND_EXPR_COND (dummy_cond
));
739 if (! is_gimple_min_invariant (cached_lhs
))
741 cached_lhs
= lookup_avail_expr (dummy_cond
, false);
742 if (!cached_lhs
|| ! is_gimple_min_invariant (cached_lhs
))
743 cached_lhs
= simplify_cond_and_lookup_avail_expr (dummy_cond
,
748 /* We can have conditionals which just test the state of a
749 variable rather than use a relational operator. These are
750 simpler to handle. */
751 else if (TREE_CODE (cond
) == SSA_NAME
)
754 cached_lhs
= SSA_NAME_VALUE (cached_lhs
);
755 if (cached_lhs
&& ! is_gimple_min_invariant (cached_lhs
))
759 cached_lhs
= lookup_avail_expr (stmt
, false);
763 edge taken_edge
= find_taken_edge (e
->dest
, cached_lhs
);
764 basic_block dest
= (taken_edge
? taken_edge
->dest
: NULL
);
769 /* If we have a known destination for the conditional, then
770 we can perform this optimization, which saves at least one
771 conditional jump each time it applies since we get to
772 bypass the conditional at our original destination. */
775 struct edge_info
*edge_info
;
777 update_bb_profile_for_threading (e
->dest
, EDGE_FREQUENCY (e
),
778 e
->count
, taken_edge
);
782 edge_info
= allocate_edge_info (e
);
783 edge_info
->redirection_target
= taken_edge
;
784 bb_ann (e
->dest
)->incoming_edge_threaded
= true;
791 /* Initialize local stacks for this optimizer and record equivalences
792 upon entry to BB. Equivalences can come from the edge traversed to
793 reach BB or they may come from PHI nodes at the start of BB. */
796 dom_opt_initialize_block (struct dom_walk_data
*walk_data ATTRIBUTE_UNUSED
,
799 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
800 fprintf (dump_file
, "\n\nOptimizing block #%d\n\n", bb
->index
);
802 /* Push a marker on the stacks of local information so that we know how
803 far to unwind when we finalize this block. */
804 VEC_safe_push (tree_on_heap
, avail_exprs_stack
, NULL_TREE
);
805 VEC_safe_push (tree_on_heap
, block_defs_stack
, NULL_TREE
);
806 VEC_safe_push (tree_on_heap
, const_and_copies_stack
, NULL_TREE
);
807 VEC_safe_push (tree_on_heap
, nonzero_vars_stack
, NULL_TREE
);
808 VEC_safe_push (tree_on_heap
, vrp_variables_stack
, NULL_TREE
);
810 record_equivalences_from_incoming_edge (bb
);
812 /* PHI nodes can create equivalences too. */
813 record_equivalences_from_phis (bb
);
816 /* Given an expression EXPR (a relational expression or a statement),
817 initialize the hash table element pointed by by ELEMENT. */
820 initialize_hash_element (tree expr
, tree lhs
, struct expr_hash_elt
*element
)
822 /* Hash table elements may be based on conditional expressions or statements.
824 For the former case, we have no annotation and we want to hash the
825 conditional expression. In the latter case we have an annotation and
826 we want to record the expression the statement evaluates. */
827 if (COMPARISON_CLASS_P (expr
) || TREE_CODE (expr
) == TRUTH_NOT_EXPR
)
832 else if (TREE_CODE (expr
) == COND_EXPR
)
834 element
->ann
= stmt_ann (expr
);
835 element
->rhs
= COND_EXPR_COND (expr
);
837 else if (TREE_CODE (expr
) == SWITCH_EXPR
)
839 element
->ann
= stmt_ann (expr
);
840 element
->rhs
= SWITCH_COND (expr
);
842 else if (TREE_CODE (expr
) == RETURN_EXPR
&& TREE_OPERAND (expr
, 0))
844 element
->ann
= stmt_ann (expr
);
845 element
->rhs
= TREE_OPERAND (TREE_OPERAND (expr
, 0), 1);
849 element
->ann
= stmt_ann (expr
);
850 element
->rhs
= TREE_OPERAND (expr
, 1);
854 element
->hash
= avail_expr_hash (element
);
857 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
858 LIMIT entries left in LOCALs. */
861 remove_local_expressions_from_table (void)
863 /* Remove all the expressions made available in this block. */
864 while (VEC_length (tree_on_heap
, avail_exprs_stack
) > 0)
866 struct expr_hash_elt element
;
867 tree expr
= VEC_pop (tree_on_heap
, avail_exprs_stack
);
869 if (expr
== NULL_TREE
)
872 initialize_hash_element (expr
, NULL
, &element
);
873 htab_remove_elt_with_hash (avail_exprs
, &element
, element
.hash
);
877 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
878 state, stopping when there are LIMIT entries left in LOCALs. */
881 restore_nonzero_vars_to_original_value (void)
883 while (VEC_length (tree_on_heap
, nonzero_vars_stack
) > 0)
885 tree name
= VEC_pop (tree_on_heap
, nonzero_vars_stack
);
890 bitmap_clear_bit (nonzero_vars
, SSA_NAME_VERSION (name
));
894 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
895 CONST_AND_COPIES to its original state, stopping when we hit a
899 restore_vars_to_original_value (void)
901 while (VEC_length (tree_on_heap
, const_and_copies_stack
) > 0)
903 tree prev_value
, dest
;
905 dest
= VEC_pop (tree_on_heap
, const_and_copies_stack
);
910 prev_value
= VEC_pop (tree_on_heap
, const_and_copies_stack
);
911 SSA_NAME_VALUE (dest
) = prev_value
;
915 /* Similar to restore_vars_to_original_value, except that it restores
916 CURRDEFS to its original value. */
918 restore_currdefs_to_original_value (void)
920 /* Restore CURRDEFS to its original state. */
921 while (VEC_length (tree_on_heap
, block_defs_stack
) > 0)
923 tree tmp
= VEC_pop (tree_on_heap
, block_defs_stack
);
926 if (tmp
== NULL_TREE
)
929 /* If we recorded an SSA_NAME, then make the SSA_NAME the current
930 definition of its underlying variable. If we recorded anything
931 else, it must have been an _DECL node and its current reaching
932 definition must have been NULL. */
933 if (TREE_CODE (tmp
) == SSA_NAME
)
936 var
= SSA_NAME_VAR (saved_def
);
944 var_ann (var
)->current_def
= saved_def
;
948 /* We have finished processing the dominator children of BB, perform
949 any finalization actions in preparation for leaving this node in
950 the dominator tree. */
953 dom_opt_finalize_block (struct dom_walk_data
*walk_data
, basic_block bb
)
957 /* If we are at a leaf node in the dominator tree, see if we can thread
958 the edge from BB through its successor.
960 Do this before we remove entries from our equivalence tables. */
961 if (EDGE_COUNT (bb
->succs
) == 1
962 && (EDGE_SUCC (bb
, 0)->flags
& EDGE_ABNORMAL
) == 0
963 && (get_immediate_dominator (CDI_DOMINATORS
, EDGE_SUCC (bb
, 0)->dest
) != bb
964 || phi_nodes (EDGE_SUCC (bb
, 0)->dest
)))
967 thread_across_edge (walk_data
, EDGE_SUCC (bb
, 0));
969 else if ((last
= last_stmt (bb
))
970 && TREE_CODE (last
) == COND_EXPR
971 && (COMPARISON_CLASS_P (COND_EXPR_COND (last
))
972 || TREE_CODE (COND_EXPR_COND (last
)) == SSA_NAME
)
973 && EDGE_COUNT (bb
->succs
) == 2
974 && (EDGE_SUCC (bb
, 0)->flags
& EDGE_ABNORMAL
) == 0
975 && (EDGE_SUCC (bb
, 1)->flags
& EDGE_ABNORMAL
) == 0)
977 edge true_edge
, false_edge
;
979 extract_true_false_edges_from_block (bb
, &true_edge
, &false_edge
);
981 /* If the THEN arm is the end of a dominator tree or has PHI nodes,
982 then try to thread through its edge. */
983 if (get_immediate_dominator (CDI_DOMINATORS
, true_edge
->dest
) != bb
984 || phi_nodes (true_edge
->dest
))
986 struct edge_info
*edge_info
;
989 /* Push a marker onto the available expression stack so that we
990 unwind any expressions related to the TRUE arm before processing
991 the false arm below. */
992 VEC_safe_push (tree_on_heap
, avail_exprs_stack
, NULL_TREE
);
993 VEC_safe_push (tree_on_heap
, block_defs_stack
, NULL_TREE
);
994 VEC_safe_push (tree_on_heap
, const_and_copies_stack
, NULL_TREE
);
996 edge_info
= true_edge
->aux
;
998 /* If we have info associated with this edge, record it into
999 our equivalency tables. */
1002 tree
*cond_equivalences
= edge_info
->cond_equivalences
;
1003 tree lhs
= edge_info
->lhs
;
1004 tree rhs
= edge_info
->rhs
;
1006 /* If we have a simple NAME = VALUE equivalency record it.
1007 Until the jump threading selection code improves, only
1008 do this if both the name and value are SSA_NAMEs with
1009 the same underlying variable to avoid missing threading
1012 && TREE_CODE (COND_EXPR_COND (last
)) == SSA_NAME
1013 && TREE_CODE (edge_info
->rhs
) == SSA_NAME
1014 && SSA_NAME_VAR (lhs
) == SSA_NAME_VAR (rhs
))
1015 record_const_or_copy (lhs
, rhs
);
1017 /* If we have 0 = COND or 1 = COND equivalences, record them
1018 into our expression hash tables. */
1019 if (cond_equivalences
)
1020 for (i
= 0; i
< edge_info
->max_cond_equivalences
; i
+= 2)
1022 tree expr
= cond_equivalences
[i
];
1023 tree value
= cond_equivalences
[i
+ 1];
1025 record_cond (expr
, value
);
1029 /* Now thread the edge. */
1030 thread_across_edge (walk_data
, true_edge
);
1032 /* And restore the various tables to their state before
1033 we threaded this edge. */
1034 remove_local_expressions_from_table ();
1035 restore_vars_to_original_value ();
1036 restore_currdefs_to_original_value ();
1039 /* Similarly for the ELSE arm. */
1040 if (get_immediate_dominator (CDI_DOMINATORS
, false_edge
->dest
) != bb
1041 || phi_nodes (false_edge
->dest
))
1043 struct edge_info
*edge_info
;
1046 edge_info
= false_edge
->aux
;
1048 /* If we have info associated with this edge, record it into
1049 our equivalency tables. */
1052 tree
*cond_equivalences
= edge_info
->cond_equivalences
;
1053 tree lhs
= edge_info
->lhs
;
1054 tree rhs
= edge_info
->rhs
;
1056 /* If we have a simple NAME = VALUE equivalency record it.
1057 Until the jump threading selection code improves, only
1058 do this if both the name and value are SSA_NAMEs with
1059 the same underlying variable to avoid missing threading
1062 && TREE_CODE (COND_EXPR_COND (last
)) == SSA_NAME
)
1063 record_const_or_copy (lhs
, rhs
);
1065 /* If we have 0 = COND or 1 = COND equivalences, record them
1066 into our expression hash tables. */
1067 if (cond_equivalences
)
1068 for (i
= 0; i
< edge_info
->max_cond_equivalences
; i
+= 2)
1070 tree expr
= cond_equivalences
[i
];
1071 tree value
= cond_equivalences
[i
+ 1];
1073 record_cond (expr
, value
);
1077 thread_across_edge (walk_data
, false_edge
);
1079 /* No need to remove local expressions from our tables
1080 or restore vars to their original value as that will
1081 be done immediately below. */
1085 remove_local_expressions_from_table ();
1086 restore_nonzero_vars_to_original_value ();
1087 restore_vars_to_original_value ();
1088 restore_currdefs_to_original_value ();
1090 /* Remove VRP records associated with this basic block. They are no
1093 To be efficient, we note which variables have had their values
1094 constrained in this block. So walk over each variable in the
1095 VRP_VARIABLEs array. */
1096 while (VEC_length (tree_on_heap
, vrp_variables_stack
) > 0)
1098 tree var
= VEC_pop (tree_on_heap
, vrp_variables_stack
);
1099 struct vrp_hash_elt vrp_hash_elt
, *vrp_hash_elt_p
;
1102 /* Each variable has a stack of value range records. We want to
1103 invalidate those associated with our basic block. So we walk
1104 the array backwards popping off records associated with our
1105 block. Once we hit a record not associated with our block
1107 varray_type var_vrp_records
;
1112 vrp_hash_elt
.var
= var
;
1113 vrp_hash_elt
.records
= NULL
;
1115 slot
= htab_find_slot (vrp_data
, &vrp_hash_elt
, NO_INSERT
);
1117 vrp_hash_elt_p
= (struct vrp_hash_elt
*) *slot
;
1118 var_vrp_records
= vrp_hash_elt_p
->records
;
1120 while (VARRAY_ACTIVE_SIZE (var_vrp_records
) > 0)
1122 struct vrp_element
*element
1123 = (struct vrp_element
*)VARRAY_TOP_GENERIC_PTR (var_vrp_records
);
1125 if (element
->bb
!= bb
)
1128 VARRAY_POP (var_vrp_records
);
1132 /* If we queued any statements to rescan in this block, then
1133 go ahead and rescan them now. */
1134 while (VEC_length (tree_on_heap
, stmts_to_rescan
) > 0)
1136 tree stmt
= VEC_last (tree_on_heap
, stmts_to_rescan
);
1137 basic_block stmt_bb
= bb_for_stmt (stmt
);
1142 VEC_pop (tree_on_heap
, stmts_to_rescan
);
1143 mark_new_vars_to_rename (stmt
, vars_to_rename
);
1147 /* PHI nodes can create equivalences too.
1149 Ignoring any alternatives which are the same as the result, if
1150 all the alternatives are equal, then the PHI node creates an
1153 Additionally, if all the PHI alternatives are known to have a nonzero
1154 value, then the result of this PHI is known to have a nonzero value,
1155 even if we do not know its exact value. */
1158 record_equivalences_from_phis (basic_block bb
)
1162 for (phi
= phi_nodes (bb
); phi
; phi
= PHI_CHAIN (phi
))
1164 tree lhs
= PHI_RESULT (phi
);
1168 for (i
= 0; i
< PHI_NUM_ARGS (phi
); i
++)
1170 tree t
= PHI_ARG_DEF (phi
, i
);
1172 /* Ignore alternatives which are the same as our LHS. Since
1173 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1174 can simply compare pointers. */
1178 /* If we have not processed an alternative yet, then set
1179 RHS to this alternative. */
1182 /* If we have processed an alternative (stored in RHS), then
1183 see if it is equal to this one. If it isn't, then stop
1185 else if (! operand_equal_for_phi_arg_p (rhs
, t
))
1189 /* If we had no interesting alternatives, then all the RHS alternatives
1190 must have been the same as LHS. */
1194 /* If we managed to iterate through each PHI alternative without
1195 breaking out of the loop, then we have a PHI which may create
1196 a useful equivalence. We do not need to record unwind data for
1197 this, since this is a true assignment and not an equivalence
1198 inferred from a comparison. All uses of this ssa name are dominated
1199 by this assignment, so unwinding just costs time and space. */
1200 if (i
== PHI_NUM_ARGS (phi
)
1201 && may_propagate_copy (lhs
, rhs
))
1202 SSA_NAME_VALUE (lhs
) = rhs
;
1204 /* Now see if we know anything about the nonzero property for the
1205 result of this PHI. */
1206 for (i
= 0; i
< PHI_NUM_ARGS (phi
); i
++)
1208 if (!PHI_ARG_NONZERO (phi
, i
))
1212 if (i
== PHI_NUM_ARGS (phi
))
1213 bitmap_set_bit (nonzero_vars
, SSA_NAME_VERSION (PHI_RESULT (phi
)));
1215 register_new_def (lhs
, &block_defs_stack
);
1219 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1220 return that edge. Otherwise return NULL. */
1222 single_incoming_edge_ignoring_loop_edges (basic_block bb
)
1228 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1230 /* A loop back edge can be identified by the destination of
1231 the edge dominating the source of the edge. */
1232 if (dominated_by_p (CDI_DOMINATORS
, e
->src
, e
->dest
))
1235 /* If we have already seen a non-loop edge, then we must have
1236 multiple incoming non-loop edges and thus we return NULL. */
1240 /* This is the first non-loop incoming edge we have found. Record
1248 /* Record any equivalences created by the incoming edge to BB. If BB
1249 has more than one incoming edge, then no equivalence is created. */
1252 record_equivalences_from_incoming_edge (basic_block bb
)
1256 struct edge_info
*edge_info
;
1258 /* If our parent block ended with a control statement, then we may be
1259 able to record some equivalences based on which outgoing edge from
1260 the parent was followed. */
1261 parent
= get_immediate_dominator (CDI_DOMINATORS
, bb
);
1263 e
= single_incoming_edge_ignoring_loop_edges (bb
);
1265 /* If we had a single incoming edge from our parent block, then enter
1266 any data associated with the edge into our tables. */
1267 if (e
&& e
->src
== parent
)
1275 tree lhs
= edge_info
->lhs
;
1276 tree rhs
= edge_info
->rhs
;
1277 tree
*cond_equivalences
= edge_info
->cond_equivalences
;
1280 record_equality (lhs
, rhs
);
1282 if (cond_equivalences
)
1284 bool recorded_range
= false;
1285 for (i
= 0; i
< edge_info
->max_cond_equivalences
; i
+= 2)
1287 tree expr
= cond_equivalences
[i
];
1288 tree value
= cond_equivalences
[i
+ 1];
1290 record_cond (expr
, value
);
1292 /* For the first true equivalence, record range
1293 information. We only do this for the first
1294 true equivalence as it should dominate any
1295 later true equivalences. */
1296 if (! recorded_range
1297 && COMPARISON_CLASS_P (expr
)
1298 && value
== boolean_true_node
1299 && TREE_CONSTANT (TREE_OPERAND (expr
, 1)))
1301 record_range (expr
, bb
);
1302 recorded_range
= true;
1310 /* Dump SSA statistics on FILE. */
1313 dump_dominator_optimization_stats (FILE *file
)
1317 fprintf (file
, "Total number of statements: %6ld\n\n",
1318 opt_stats
.num_stmts
);
1319 fprintf (file
, "Exprs considered for dominator optimizations: %6ld\n",
1320 opt_stats
.num_exprs_considered
);
1322 n_exprs
= opt_stats
.num_exprs_considered
;
1326 fprintf (file
, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1327 opt_stats
.num_re
, PERCENT (opt_stats
.num_re
,
1330 fprintf (file
, "\nHash table statistics:\n");
1332 fprintf (file
, " avail_exprs: ");
1333 htab_statistics (file
, avail_exprs
);
1337 /* Dump SSA statistics on stderr. */
1340 debug_dominator_optimization_stats (void)
1342 dump_dominator_optimization_stats (stderr
);
1346 /* Dump statistics for the hash table HTAB. */
1349 htab_statistics (FILE *file
, htab_t htab
)
1351 fprintf (file
, "size %ld, %ld elements, %f collision/search ratio\n",
1352 (long) htab_size (htab
),
1353 (long) htab_elements (htab
),
1354 htab_collisions (htab
));
1357 /* Record the fact that VAR has a nonzero value, though we may not know
1358 its exact value. Note that if VAR is already known to have a nonzero
1359 value, then we do nothing. */
1362 record_var_is_nonzero (tree var
)
1364 int indx
= SSA_NAME_VERSION (var
);
1366 if (bitmap_bit_p (nonzero_vars
, indx
))
1369 /* Mark it in the global table. */
1370 bitmap_set_bit (nonzero_vars
, indx
);
1372 /* Record this SSA_NAME so that we can reset the global table
1373 when we leave this block. */
1374 VEC_safe_push (tree_on_heap
, nonzero_vars_stack
, var
);
1377 /* Enter a statement into the true/false expression hash table indicating
1378 that the condition COND has the value VALUE. */
1381 record_cond (tree cond
, tree value
)
1383 struct expr_hash_elt
*element
= xmalloc (sizeof (struct expr_hash_elt
));
1386 initialize_hash_element (cond
, value
, element
);
1388 slot
= htab_find_slot_with_hash (avail_exprs
, (void *)element
,
1389 element
->hash
, INSERT
);
1392 *slot
= (void *) element
;
1393 VEC_safe_push (tree_on_heap
, avail_exprs_stack
, cond
);
1399 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1400 the new conditional into *p, then store a boolean_true_node
1404 build_and_record_new_cond (enum tree_code new_code
, tree op0
, tree op1
, tree
*p
)
1406 *p
= build2 (new_code
, boolean_type_node
, op0
, op1
);
1408 *p
= boolean_true_node
;
1411 /* Record that COND is true and INVERTED is false into the edge information
1412 structure. Also record that any conditions dominated by COND are true
1415 For example, if a < b is true, then a <= b must also be true. */
1418 record_conditions (struct edge_info
*edge_info
, tree cond
, tree inverted
)
1422 if (!COMPARISON_CLASS_P (cond
))
1425 op0
= TREE_OPERAND (cond
, 0);
1426 op1
= TREE_OPERAND (cond
, 1);
1428 switch (TREE_CODE (cond
))
1432 edge_info
->max_cond_equivalences
= 12;
1433 edge_info
->cond_equivalences
= xmalloc (12 * sizeof (tree
));
1434 build_and_record_new_cond ((TREE_CODE (cond
) == LT_EXPR
1435 ? LE_EXPR
: GE_EXPR
),
1436 op0
, op1
, &edge_info
->cond_equivalences
[4]);
1437 build_and_record_new_cond (ORDERED_EXPR
, op0
, op1
,
1438 &edge_info
->cond_equivalences
[6]);
1439 build_and_record_new_cond (NE_EXPR
, op0
, op1
,
1440 &edge_info
->cond_equivalences
[8]);
1441 build_and_record_new_cond (LTGT_EXPR
, op0
, op1
,
1442 &edge_info
->cond_equivalences
[10]);
1447 edge_info
->max_cond_equivalences
= 6;
1448 edge_info
->cond_equivalences
= xmalloc (6 * sizeof (tree
));
1449 build_and_record_new_cond (ORDERED_EXPR
, op0
, op1
,
1450 &edge_info
->cond_equivalences
[4]);
1454 edge_info
->max_cond_equivalences
= 10;
1455 edge_info
->cond_equivalences
= xmalloc (10 * sizeof (tree
));
1456 build_and_record_new_cond (ORDERED_EXPR
, op0
, op1
,
1457 &edge_info
->cond_equivalences
[4]);
1458 build_and_record_new_cond (LE_EXPR
, op0
, op1
,
1459 &edge_info
->cond_equivalences
[6]);
1460 build_and_record_new_cond (GE_EXPR
, op0
, op1
,
1461 &edge_info
->cond_equivalences
[8]);
1464 case UNORDERED_EXPR
:
1465 edge_info
->max_cond_equivalences
= 16;
1466 edge_info
->cond_equivalences
= xmalloc (16 * sizeof (tree
));
1467 build_and_record_new_cond (NE_EXPR
, op0
, op1
,
1468 &edge_info
->cond_equivalences
[4]);
1469 build_and_record_new_cond (UNLE_EXPR
, op0
, op1
,
1470 &edge_info
->cond_equivalences
[6]);
1471 build_and_record_new_cond (UNGE_EXPR
, op0
, op1
,
1472 &edge_info
->cond_equivalences
[8]);
1473 build_and_record_new_cond (UNEQ_EXPR
, op0
, op1
,
1474 &edge_info
->cond_equivalences
[10]);
1475 build_and_record_new_cond (UNLT_EXPR
, op0
, op1
,
1476 &edge_info
->cond_equivalences
[12]);
1477 build_and_record_new_cond (UNGT_EXPR
, op0
, op1
,
1478 &edge_info
->cond_equivalences
[14]);
1483 edge_info
->max_cond_equivalences
= 8;
1484 edge_info
->cond_equivalences
= xmalloc (8 * sizeof (tree
));
1485 build_and_record_new_cond ((TREE_CODE (cond
) == UNLT_EXPR
1486 ? UNLE_EXPR
: UNGE_EXPR
),
1487 op0
, op1
, &edge_info
->cond_equivalences
[4]);
1488 build_and_record_new_cond (NE_EXPR
, op0
, op1
,
1489 &edge_info
->cond_equivalences
[6]);
1493 edge_info
->max_cond_equivalences
= 8;
1494 edge_info
->cond_equivalences
= xmalloc (8 * sizeof (tree
));
1495 build_and_record_new_cond (UNLE_EXPR
, op0
, op1
,
1496 &edge_info
->cond_equivalences
[4]);
1497 build_and_record_new_cond (UNGE_EXPR
, op0
, op1
,
1498 &edge_info
->cond_equivalences
[6]);
1502 edge_info
->max_cond_equivalences
= 8;
1503 edge_info
->cond_equivalences
= xmalloc (8 * sizeof (tree
));
1504 build_and_record_new_cond (NE_EXPR
, op0
, op1
,
1505 &edge_info
->cond_equivalences
[4]);
1506 build_and_record_new_cond (ORDERED_EXPR
, op0
, op1
,
1507 &edge_info
->cond_equivalences
[6]);
1511 edge_info
->max_cond_equivalences
= 4;
1512 edge_info
->cond_equivalences
= xmalloc (4 * sizeof (tree
));
1516 /* Now store the original true and false conditions into the first
1518 edge_info
->cond_equivalences
[0] = cond
;
1519 edge_info
->cond_equivalences
[1] = boolean_true_node
;
1520 edge_info
->cond_equivalences
[2] = inverted
;
1521 edge_info
->cond_equivalences
[3] = boolean_false_node
;
1524 /* A helper function for record_const_or_copy and record_equality.
1525 Do the work of recording the value and undo info. */
1528 record_const_or_copy_1 (tree x
, tree y
, tree prev_x
)
1530 SSA_NAME_VALUE (x
) = y
;
1532 VEC_safe_push (tree_on_heap
, const_and_copies_stack
, prev_x
);
1533 VEC_safe_push (tree_on_heap
, const_and_copies_stack
, x
);
1537 /* Return the loop depth of the basic block of the defining statement of X.
1538 This number should not be treated as absolutely correct because the loop
1539 information may not be completely up-to-date when dom runs. However, it
1540 will be relatively correct, and as more passes are taught to keep loop info
1541 up to date, the result will become more and more accurate. */
1544 loop_depth_of_name (tree x
)
1549 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1550 if (TREE_CODE (x
) != SSA_NAME
)
1553 /* Otherwise return the loop depth of the defining statement's bb.
1554 Note that there may not actually be a bb for this statement, if the
1555 ssa_name is live on entry. */
1556 defstmt
= SSA_NAME_DEF_STMT (x
);
1557 defbb
= bb_for_stmt (defstmt
);
1561 return defbb
->loop_depth
;
1565 /* Record that X is equal to Y in const_and_copies. Record undo
1566 information in the block-local vector. */
1569 record_const_or_copy (tree x
, tree y
)
1571 tree prev_x
= SSA_NAME_VALUE (x
);
1573 if (TREE_CODE (y
) == SSA_NAME
)
1575 tree tmp
= SSA_NAME_VALUE (y
);
1580 record_const_or_copy_1 (x
, y
, prev_x
);
1583 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1584 This constrains the cases in which we may treat this as assignment. */
1587 record_equality (tree x
, tree y
)
1589 tree prev_x
= NULL
, prev_y
= NULL
;
1591 if (TREE_CODE (x
) == SSA_NAME
)
1592 prev_x
= SSA_NAME_VALUE (x
);
1593 if (TREE_CODE (y
) == SSA_NAME
)
1594 prev_y
= SSA_NAME_VALUE (y
);
1596 /* If one of the previous values is invariant, or invariant in more loops
1597 (by depth), then use that.
1598 Otherwise it doesn't matter which value we choose, just so
1599 long as we canonicalize on one value. */
1600 if (TREE_INVARIANT (y
))
1602 else if (TREE_INVARIANT (x
) || (loop_depth_of_name (x
) <= loop_depth_of_name (y
)))
1603 prev_x
= x
, x
= y
, y
= prev_x
, prev_x
= prev_y
;
1604 else if (prev_x
&& TREE_INVARIANT (prev_x
))
1605 x
= y
, y
= prev_x
, prev_x
= prev_y
;
1606 else if (prev_y
&& TREE_CODE (prev_y
) != VALUE_HANDLE
)
1609 /* After the swapping, we must have one SSA_NAME. */
1610 if (TREE_CODE (x
) != SSA_NAME
)
1613 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1614 variable compared against zero. If we're honoring signed zeros,
1615 then we cannot record this value unless we know that the value is
1617 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x
)))
1618 && (TREE_CODE (y
) != REAL_CST
1619 || REAL_VALUES_EQUAL (dconst0
, TREE_REAL_CST (y
))))
1622 record_const_or_copy_1 (x
, y
, prev_x
);
1625 /* Return true, if it is ok to do folding of an associative expression.
1626 EXP is the tree for the associative expression. */
1629 unsafe_associative_fp_binop (tree exp
)
1631 enum tree_code code
= TREE_CODE (exp
);
1632 return !(!flag_unsafe_math_optimizations
1633 && (code
== MULT_EXPR
|| code
== PLUS_EXPR
1634 || code
== MINUS_EXPR
)
1635 && FLOAT_TYPE_P (TREE_TYPE (exp
)));
1638 /* Returns true when STMT is a simple iv increment. It detects the
1639 following situation:
1641 i_1 = phi (..., i_2)
1642 i_2 = i_1 +/- ... */
1645 simple_iv_increment_p (tree stmt
)
1647 tree lhs
, rhs
, preinc
, phi
;
1650 if (TREE_CODE (stmt
) != MODIFY_EXPR
)
1653 lhs
= TREE_OPERAND (stmt
, 0);
1654 if (TREE_CODE (lhs
) != SSA_NAME
)
1657 rhs
= TREE_OPERAND (stmt
, 1);
1659 if (TREE_CODE (rhs
) != PLUS_EXPR
1660 && TREE_CODE (rhs
) != MINUS_EXPR
)
1663 preinc
= TREE_OPERAND (rhs
, 0);
1664 if (TREE_CODE (preinc
) != SSA_NAME
)
1667 phi
= SSA_NAME_DEF_STMT (preinc
);
1668 if (TREE_CODE (phi
) != PHI_NODE
)
1671 for (i
= 0; i
< (unsigned) PHI_NUM_ARGS (phi
); i
++)
1672 if (PHI_ARG_DEF (phi
, i
) == lhs
)
1678 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1679 hash tables. Try to simplify the RHS using whatever equivalences
1680 we may have recorded.
1682 If we are able to simplify the RHS, then lookup the simplified form in
1683 the hash table and return the result. Otherwise return NULL. */
1686 simplify_rhs_and_lookup_avail_expr (struct dom_walk_data
*walk_data
,
1687 tree stmt
, int insert
)
1689 tree rhs
= TREE_OPERAND (stmt
, 1);
1690 enum tree_code rhs_code
= TREE_CODE (rhs
);
1693 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1694 In which case we can change this statement to be lhs = y.
1695 Which can then be copy propagated.
1697 Similarly for negation. */
1698 if ((rhs_code
== BIT_NOT_EXPR
|| rhs_code
== NEGATE_EXPR
)
1699 && TREE_CODE (TREE_OPERAND (rhs
, 0)) == SSA_NAME
)
1701 /* Get the definition statement for our RHS. */
1702 tree rhs_def_stmt
= SSA_NAME_DEF_STMT (TREE_OPERAND (rhs
, 0));
1704 /* See if the RHS_DEF_STMT has the same form as our statement. */
1705 if (TREE_CODE (rhs_def_stmt
) == MODIFY_EXPR
1706 && TREE_CODE (TREE_OPERAND (rhs_def_stmt
, 1)) == rhs_code
)
1708 tree rhs_def_operand
;
1710 rhs_def_operand
= TREE_OPERAND (TREE_OPERAND (rhs_def_stmt
, 1), 0);
1712 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1713 if (TREE_CODE (rhs_def_operand
) == SSA_NAME
1714 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand
))
1715 result
= update_rhs_and_lookup_avail_expr (stmt
,
1721 /* If we have z = (x OP C1), see if we earlier had x = y OP C2.
1722 If OP is associative, create and fold (y OP C2) OP C1 which
1723 should result in (y OP C3), use that as the RHS for the
1724 assignment. Add minus to this, as we handle it specially below. */
1725 if ((associative_tree_code (rhs_code
) || rhs_code
== MINUS_EXPR
)
1726 && TREE_CODE (TREE_OPERAND (rhs
, 0)) == SSA_NAME
1727 && is_gimple_min_invariant (TREE_OPERAND (rhs
, 1)))
1729 tree rhs_def_stmt
= SSA_NAME_DEF_STMT (TREE_OPERAND (rhs
, 0));
1731 /* If the statement defines an induction variable, do not propagate
1732 its value, so that we do not create overlapping life ranges. */
1733 if (simple_iv_increment_p (rhs_def_stmt
))
1734 goto dont_fold_assoc
;
1736 /* See if the RHS_DEF_STMT has the same form as our statement. */
1737 if (TREE_CODE (rhs_def_stmt
) == MODIFY_EXPR
)
1739 tree rhs_def_rhs
= TREE_OPERAND (rhs_def_stmt
, 1);
1740 enum tree_code rhs_def_code
= TREE_CODE (rhs_def_rhs
);
1742 if ((rhs_code
== rhs_def_code
&& unsafe_associative_fp_binop (rhs
))
1743 || (rhs_code
== PLUS_EXPR
&& rhs_def_code
== MINUS_EXPR
)
1744 || (rhs_code
== MINUS_EXPR
&& rhs_def_code
== PLUS_EXPR
))
1746 tree def_stmt_op0
= TREE_OPERAND (rhs_def_rhs
, 0);
1747 tree def_stmt_op1
= TREE_OPERAND (rhs_def_rhs
, 1);
1749 if (TREE_CODE (def_stmt_op0
) == SSA_NAME
1750 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0
)
1751 && is_gimple_min_invariant (def_stmt_op1
))
1753 tree outer_const
= TREE_OPERAND (rhs
, 1);
1754 tree type
= TREE_TYPE (TREE_OPERAND (stmt
, 0));
1757 /* If we care about correct floating point results, then
1758 don't fold x + c1 - c2. Note that we need to take both
1759 the codes and the signs to figure this out. */
1760 if (FLOAT_TYPE_P (type
)
1761 && !flag_unsafe_math_optimizations
1762 && (rhs_def_code
== PLUS_EXPR
1763 || rhs_def_code
== MINUS_EXPR
))
1767 neg
^= (rhs_code
== MINUS_EXPR
);
1768 neg
^= (rhs_def_code
== MINUS_EXPR
);
1769 neg
^= real_isneg (TREE_REAL_CST_PTR (outer_const
));
1770 neg
^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1
));
1773 goto dont_fold_assoc
;
1776 /* Ho hum. So fold will only operate on the outermost
1777 thingy that we give it, so we have to build the new
1778 expression in two pieces. This requires that we handle
1779 combinations of plus and minus. */
1780 if (rhs_def_code
!= rhs_code
)
1782 if (rhs_def_code
== MINUS_EXPR
)
1783 t
= build (MINUS_EXPR
, type
, outer_const
, def_stmt_op1
);
1785 t
= build (MINUS_EXPR
, type
, def_stmt_op1
, outer_const
);
1786 rhs_code
= PLUS_EXPR
;
1788 else if (rhs_def_code
== MINUS_EXPR
)
1789 t
= build (PLUS_EXPR
, type
, def_stmt_op1
, outer_const
);
1791 t
= build (rhs_def_code
, type
, def_stmt_op1
, outer_const
);
1793 t
= build (rhs_code
, type
, def_stmt_op0
, t
);
1796 /* If the result is a suitable looking gimple expression,
1797 then use it instead of the original for STMT. */
1798 if (TREE_CODE (t
) == SSA_NAME
1799 || (UNARY_CLASS_P (t
)
1800 && TREE_CODE (TREE_OPERAND (t
, 0)) == SSA_NAME
)
1801 || ((BINARY_CLASS_P (t
) || COMPARISON_CLASS_P (t
))
1802 && TREE_CODE (TREE_OPERAND (t
, 0)) == SSA_NAME
1803 && is_gimple_val (TREE_OPERAND (t
, 1))))
1804 result
= update_rhs_and_lookup_avail_expr (stmt
, t
, insert
);
1811 /* Transform TRUNC_DIV_EXPR and TRUNC_MOD_EXPR into RSHIFT_EXPR
1812 and BIT_AND_EXPR respectively if the first operand is greater
1813 than zero and the second operand is an exact power of two. */
1814 if ((rhs_code
== TRUNC_DIV_EXPR
|| rhs_code
== TRUNC_MOD_EXPR
)
1815 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs
, 0)))
1816 && integer_pow2p (TREE_OPERAND (rhs
, 1)))
1819 tree op
= TREE_OPERAND (rhs
, 0);
1821 if (TYPE_UNSIGNED (TREE_TYPE (op
)))
1823 val
= integer_one_node
;
1827 tree dummy_cond
= walk_data
->global_data
;
1831 dummy_cond
= build (GT_EXPR
, boolean_type_node
,
1832 op
, integer_zero_node
);
1833 dummy_cond
= build (COND_EXPR
, void_type_node
,
1834 dummy_cond
, NULL
, NULL
);
1835 walk_data
->global_data
= dummy_cond
;
1839 TREE_SET_CODE (COND_EXPR_COND (dummy_cond
), GT_EXPR
);
1840 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 0) = op
;
1841 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 1)
1842 = integer_zero_node
;
1844 val
= simplify_cond_and_lookup_avail_expr (dummy_cond
, NULL
, false);
1847 if (val
&& integer_onep (val
))
1850 tree op0
= TREE_OPERAND (rhs
, 0);
1851 tree op1
= TREE_OPERAND (rhs
, 1);
1853 if (rhs_code
== TRUNC_DIV_EXPR
)
1854 t
= build (RSHIFT_EXPR
, TREE_TYPE (op0
), op0
,
1855 build_int_cst (NULL_TREE
, tree_log2 (op1
)));
1857 t
= build (BIT_AND_EXPR
, TREE_TYPE (op0
), op0
,
1858 local_fold (build (MINUS_EXPR
, TREE_TYPE (op1
),
1859 op1
, integer_one_node
)));
1861 result
= update_rhs_and_lookup_avail_expr (stmt
, t
, insert
);
1865 /* Transform ABS (X) into X or -X as appropriate. */
1866 if (rhs_code
== ABS_EXPR
1867 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs
, 0))))
1870 tree op
= TREE_OPERAND (rhs
, 0);
1871 tree type
= TREE_TYPE (op
);
1873 if (TYPE_UNSIGNED (type
))
1875 val
= integer_zero_node
;
1879 tree dummy_cond
= walk_data
->global_data
;
1883 dummy_cond
= build (LE_EXPR
, boolean_type_node
,
1884 op
, integer_zero_node
);
1885 dummy_cond
= build (COND_EXPR
, void_type_node
,
1886 dummy_cond
, NULL
, NULL
);
1887 walk_data
->global_data
= dummy_cond
;
1891 TREE_SET_CODE (COND_EXPR_COND (dummy_cond
), LE_EXPR
);
1892 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 0) = op
;
1893 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 1)
1894 = build_int_cst (type
, 0);
1896 val
= simplify_cond_and_lookup_avail_expr (dummy_cond
, NULL
, false);
1900 TREE_SET_CODE (COND_EXPR_COND (dummy_cond
), GE_EXPR
);
1901 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 0) = op
;
1902 TREE_OPERAND (COND_EXPR_COND (dummy_cond
), 1)
1903 = build_int_cst (type
, 0);
1905 val
= simplify_cond_and_lookup_avail_expr (dummy_cond
,
1910 if (integer_zerop (val
))
1911 val
= integer_one_node
;
1912 else if (integer_onep (val
))
1913 val
= integer_zero_node
;
1919 && (integer_onep (val
) || integer_zerop (val
)))
1923 if (integer_onep (val
))
1924 t
= build1 (NEGATE_EXPR
, TREE_TYPE (op
), op
);
1928 result
= update_rhs_and_lookup_avail_expr (stmt
, t
, insert
);
1932 /* Optimize *"foo" into 'f'. This is done here rather than
1933 in fold to avoid problems with stuff like &*"foo". */
1934 if (TREE_CODE (rhs
) == INDIRECT_REF
|| TREE_CODE (rhs
) == ARRAY_REF
)
1936 tree t
= fold_read_from_constant_string (rhs
);
1939 result
= update_rhs_and_lookup_avail_expr (stmt
, t
, insert
);
1945 /* COND is a condition of the form:
1947 x == const or x != const
1949 Look back to x's defining statement and see if x is defined as
1953 If const is unchanged if we convert it to type, then we can build
1954 the equivalent expression:
1957 y == const or y != const
1959 Which may allow further optimizations.
1961 Return the equivalent comparison or NULL if no such equivalent comparison
1965 find_equivalent_equality_comparison (tree cond
)
1967 tree op0
= TREE_OPERAND (cond
, 0);
1968 tree op1
= TREE_OPERAND (cond
, 1);
1969 tree def_stmt
= SSA_NAME_DEF_STMT (op0
);
1971 /* OP0 might have been a parameter, so first make sure it
1972 was defined by a MODIFY_EXPR. */
1973 if (def_stmt
&& TREE_CODE (def_stmt
) == MODIFY_EXPR
)
1975 tree def_rhs
= TREE_OPERAND (def_stmt
, 1);
1977 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1978 if ((TREE_CODE (def_rhs
) == NOP_EXPR
1979 || TREE_CODE (def_rhs
) == CONVERT_EXPR
)
1980 && TREE_CODE (TREE_OPERAND (def_rhs
, 0)) == SSA_NAME
)
1982 tree def_rhs_inner
= TREE_OPERAND (def_rhs
, 0);
1983 tree def_rhs_inner_type
= TREE_TYPE (def_rhs_inner
);
1986 if (TYPE_PRECISION (def_rhs_inner_type
)
1987 > TYPE_PRECISION (TREE_TYPE (def_rhs
)))
1990 /* What we want to prove is that if we convert OP1 to
1991 the type of the object inside the NOP_EXPR that the
1992 result is still equivalent to SRC.
1994 If that is true, the build and return new equivalent
1995 condition which uses the source of the typecast and the
1996 new constant (which has only changed its type). */
1997 new = build1 (TREE_CODE (def_rhs
), def_rhs_inner_type
, op1
);
1998 new = local_fold (new);
1999 if (is_gimple_val (new) && tree_int_cst_equal (new, op1
))
2000 return build (TREE_CODE (cond
), TREE_TYPE (cond
),
2001 def_rhs_inner
, new);
2007 /* STMT is a COND_EXPR for which we could not trivially determine its
2008 result. This routine attempts to find equivalent forms of the
2009 condition which we may be able to optimize better. It also
2010 uses simple value range propagation to optimize conditionals. */
2013 simplify_cond_and_lookup_avail_expr (tree stmt
,
2017 tree cond
= COND_EXPR_COND (stmt
);
2019 if (COMPARISON_CLASS_P (cond
))
2021 tree op0
= TREE_OPERAND (cond
, 0);
2022 tree op1
= TREE_OPERAND (cond
, 1);
2024 if (TREE_CODE (op0
) == SSA_NAME
&& is_gimple_min_invariant (op1
))
2027 tree low
, high
, cond_low
, cond_high
;
2028 int lowequal
, highequal
, swapped
, no_overlap
, subset
, cond_inverted
;
2029 varray_type vrp_records
;
2030 struct vrp_element
*element
;
2031 struct vrp_hash_elt vrp_hash_elt
, *vrp_hash_elt_p
;
2034 /* First see if we have test of an SSA_NAME against a constant
2035 where the SSA_NAME is defined by an earlier typecast which
2036 is irrelevant when performing tests against the given
2038 if (TREE_CODE (cond
) == EQ_EXPR
|| TREE_CODE (cond
) == NE_EXPR
)
2040 tree new_cond
= find_equivalent_equality_comparison (cond
);
2044 /* Update the statement to use the new equivalent
2046 COND_EXPR_COND (stmt
) = new_cond
;
2048 /* If this is not a real stmt, ann will be NULL and we
2049 avoid processing the operands. */
2053 /* Lookup the condition and return its known value if it
2055 new_cond
= lookup_avail_expr (stmt
, insert
);
2059 /* The operands have changed, so update op0 and op1. */
2060 op0
= TREE_OPERAND (cond
, 0);
2061 op1
= TREE_OPERAND (cond
, 1);
2065 /* Consult the value range records for this variable (if they exist)
2066 to see if we can eliminate or simplify this conditional.
2068 Note two tests are necessary to determine no records exist.
2069 First we have to see if the virtual array exists, if it
2070 exists, then we have to check its active size.
2072 Also note the vast majority of conditionals are not testing
2073 a variable which has had its range constrained by an earlier
2074 conditional. So this filter avoids a lot of unnecessary work. */
2075 vrp_hash_elt
.var
= op0
;
2076 vrp_hash_elt
.records
= NULL
;
2077 slot
= htab_find_slot (vrp_data
, &vrp_hash_elt
, NO_INSERT
);
2081 vrp_hash_elt_p
= (struct vrp_hash_elt
*) *slot
;
2082 vrp_records
= vrp_hash_elt_p
->records
;
2083 if (vrp_records
== NULL
)
2086 limit
= VARRAY_ACTIVE_SIZE (vrp_records
);
2088 /* If we have no value range records for this variable, or we are
2089 unable to extract a range for this condition, then there is
2092 || ! extract_range_from_cond (cond
, &cond_high
,
2093 &cond_low
, &cond_inverted
))
2096 /* We really want to avoid unnecessary computations of range
2097 info. So all ranges are computed lazily; this avoids a
2098 lot of unnecessary work. i.e., we record the conditional,
2099 but do not process how it constrains the variable's
2100 potential values until we know that processing the condition
2103 However, we do not want to have to walk a potentially long
2104 list of ranges, nor do we want to compute a variable's
2105 range more than once for a given path.
2107 Luckily, each time we encounter a conditional that can not
2108 be otherwise optimized we will end up here and we will
2109 compute the necessary range information for the variable
2110 used in this condition.
2112 Thus you can conclude that there will never be more than one
2113 conditional associated with a variable which has not been
2114 processed. So we never need to merge more than one new
2115 conditional into the current range.
2117 These properties also help us avoid unnecessary work. */
2119 = (struct vrp_element
*)VARRAY_GENERIC_PTR (vrp_records
, limit
- 1);
2121 if (element
->high
&& element
->low
)
2123 /* The last element has been processed, so there is no range
2124 merging to do, we can simply use the high/low values
2125 recorded in the last element. */
2127 high
= element
->high
;
2131 tree tmp_high
, tmp_low
;
2134 /* The last element has not been processed. Process it now.
2135 record_range should ensure for cond inverted is not set.
2136 This call can only fail if cond is x < min or x > max,
2137 which fold should have optimized into false.
2138 If that doesn't happen, just pretend all values are
2140 if (! extract_range_from_cond (element
->cond
, &tmp_high
,
2144 gcc_assert (dummy
== 0);
2146 /* If this is the only element, then no merging is necessary,
2147 the high/low values from extract_range_from_cond are all
2156 /* Get the high/low value from the previous element. */
2157 struct vrp_element
*prev
2158 = (struct vrp_element
*)VARRAY_GENERIC_PTR (vrp_records
,
2163 /* Merge in this element's range with the range from the
2166 The low value for the merged range is the maximum of
2167 the previous low value and the low value of this record.
2169 Similarly the high value for the merged range is the
2170 minimum of the previous high value and the high value of
2172 low
= (tree_int_cst_compare (low
, tmp_low
) == 1
2174 high
= (tree_int_cst_compare (high
, tmp_high
) == -1
2178 /* And record the computed range. */
2180 element
->high
= high
;
2184 /* After we have constrained this variable's potential values,
2185 we try to determine the result of the given conditional.
2187 To simplify later tests, first determine if the current
2188 low value is the same low value as the conditional.
2189 Similarly for the current high value and the high value
2190 for the conditional. */
2191 lowequal
= tree_int_cst_equal (low
, cond_low
);
2192 highequal
= tree_int_cst_equal (high
, cond_high
);
2194 if (lowequal
&& highequal
)
2195 return (cond_inverted
? boolean_false_node
: boolean_true_node
);
2197 /* To simplify the overlap/subset tests below we may want
2198 to swap the two ranges so that the larger of the two
2199 ranges occurs "first". */
2201 if (tree_int_cst_compare (low
, cond_low
) == 1
2203 && tree_int_cst_compare (cond_high
, high
) == 1))
2216 /* Now determine if there is no overlap in the ranges
2217 or if the second range is a subset of the first range. */
2218 no_overlap
= tree_int_cst_lt (high
, cond_low
);
2219 subset
= tree_int_cst_compare (cond_high
, high
) != 1;
2221 /* If there was no overlap in the ranges, then this conditional
2222 always has a false value (unless we had to invert this
2223 conditional, in which case it always has a true value). */
2225 return (cond_inverted
? boolean_true_node
: boolean_false_node
);
2227 /* If the current range is a subset of the condition's range,
2228 then this conditional always has a true value (unless we
2229 had to invert this conditional, in which case it always
2230 has a true value). */
2231 if (subset
&& swapped
)
2232 return (cond_inverted
? boolean_false_node
: boolean_true_node
);
2234 /* We were unable to determine the result of the conditional.
2235 However, we may be able to simplify the conditional. First
2236 merge the ranges in the same manner as range merging above. */
2237 low
= tree_int_cst_compare (low
, cond_low
) == 1 ? low
: cond_low
;
2238 high
= tree_int_cst_compare (high
, cond_high
) == -1 ? high
: cond_high
;
2240 /* If the range has converged to a single point, then turn this
2241 into an equality comparison. */
2242 if (TREE_CODE (cond
) != EQ_EXPR
2243 && TREE_CODE (cond
) != NE_EXPR
2244 && tree_int_cst_equal (low
, high
))
2246 TREE_SET_CODE (cond
, EQ_EXPR
);
2247 TREE_OPERAND (cond
, 1) = high
;
2254 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2255 result. This routine attempts to find equivalent forms of the
2256 condition which we may be able to optimize better. */
2259 simplify_switch_and_lookup_avail_expr (tree stmt
, int insert
)
2261 tree cond
= SWITCH_COND (stmt
);
2264 /* The optimization that we really care about is removing unnecessary
2265 casts. That will let us do much better in propagating the inferred
2266 constant at the switch target. */
2267 if (TREE_CODE (cond
) == SSA_NAME
)
2269 def
= SSA_NAME_DEF_STMT (cond
);
2270 if (TREE_CODE (def
) == MODIFY_EXPR
)
2272 def
= TREE_OPERAND (def
, 1);
2273 if (TREE_CODE (def
) == NOP_EXPR
)
2278 def
= TREE_OPERAND (def
, 0);
2280 #ifdef ENABLE_CHECKING
2281 /* ??? Why was Jeff testing this? We are gimple... */
2282 gcc_assert (is_gimple_val (def
));
2285 to
= TREE_TYPE (cond
);
2286 ti
= TREE_TYPE (def
);
2288 /* If we have an extension that preserves value, then we
2289 can copy the source value into the switch. */
2291 need_precision
= TYPE_PRECISION (ti
);
2293 if (TYPE_UNSIGNED (to
) && !TYPE_UNSIGNED (ti
))
2295 else if (!TYPE_UNSIGNED (to
) && TYPE_UNSIGNED (ti
))
2296 need_precision
+= 1;
2297 if (TYPE_PRECISION (to
) < need_precision
)
2302 SWITCH_COND (stmt
) = def
;
2305 return lookup_avail_expr (stmt
, insert
);
2315 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2316 known value for that SSA_NAME (or NULL if no value is known).
2318 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2319 even if we don't know their precise value.
2321 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2322 nodes of the successors of BB. */
2325 cprop_into_successor_phis (basic_block bb
, bitmap nonzero_vars
)
2330 /* This can get rather expensive if the implementation is naive in
2331 how it finds the phi alternative associated with a particular edge. */
2332 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2337 /* If this is an abnormal edge, then we do not want to copy propagate
2338 into the PHI alternative associated with this edge. */
2339 if (e
->flags
& EDGE_ABNORMAL
)
2342 phi
= phi_nodes (e
->dest
);
2347 for ( ; phi
; phi
= PHI_CHAIN (phi
))
2350 use_operand_p orig_p
;
2353 /* The alternative may be associated with a constant, so verify
2354 it is an SSA_NAME before doing anything with it. */
2355 orig_p
= PHI_ARG_DEF_PTR (phi
, indx
);
2356 orig
= USE_FROM_PTR (orig_p
);
2357 if (TREE_CODE (orig
) != SSA_NAME
)
2360 /* If the alternative is known to have a nonzero value, record
2361 that fact in the PHI node itself for future use. */
2362 if (bitmap_bit_p (nonzero_vars
, SSA_NAME_VERSION (orig
)))
2363 PHI_ARG_NONZERO (phi
, indx
) = true;
2365 /* If we have *ORIG_P in our constant/copy table, then replace
2366 ORIG_P with its value in our constant/copy table. */
2367 new = SSA_NAME_VALUE (orig
);
2369 && (TREE_CODE (new) == SSA_NAME
2370 || is_gimple_min_invariant (new))
2371 && may_propagate_copy (orig
, new))
2373 propagate_value (orig_p
, new);
2379 /* We have finished optimizing BB, record any information implied by
2380 taking a specific outgoing edge from BB. */
2383 record_edge_info (basic_block bb
)
2385 block_stmt_iterator bsi
= bsi_last (bb
);
2386 struct edge_info
*edge_info
;
2388 if (! bsi_end_p (bsi
))
2390 tree stmt
= bsi_stmt (bsi
);
2392 if (stmt
&& TREE_CODE (stmt
) == SWITCH_EXPR
)
2394 tree cond
= SWITCH_COND (stmt
);
2396 if (TREE_CODE (cond
) == SSA_NAME
)
2398 tree labels
= SWITCH_LABELS (stmt
);
2399 int i
, n_labels
= TREE_VEC_LENGTH (labels
);
2400 tree
*info
= xcalloc (n_basic_blocks
, sizeof (tree
));
2404 for (i
= 0; i
< n_labels
; i
++)
2406 tree label
= TREE_VEC_ELT (labels
, i
);
2407 basic_block target_bb
= label_to_block (CASE_LABEL (label
));
2409 if (CASE_HIGH (label
)
2410 || !CASE_LOW (label
)
2411 || info
[target_bb
->index
])
2412 info
[target_bb
->index
] = error_mark_node
;
2414 info
[target_bb
->index
] = label
;
2417 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2419 basic_block target_bb
= e
->dest
;
2420 tree node
= info
[target_bb
->index
];
2422 if (node
!= NULL
&& node
!= error_mark_node
)
2424 tree x
= fold_convert (TREE_TYPE (cond
), CASE_LOW (node
));
2425 edge_info
= allocate_edge_info (e
);
2426 edge_info
->lhs
= cond
;
2434 /* A COND_EXPR may create equivalences too. */
2435 if (stmt
&& TREE_CODE (stmt
) == COND_EXPR
)
2437 tree cond
= COND_EXPR_COND (stmt
);
2441 extract_true_false_edges_from_block (bb
, &true_edge
, &false_edge
);
2443 /* If the conditional is a single variable 'X', record 'X = 1'
2444 for the true edge and 'X = 0' on the false edge. */
2445 if (SSA_VAR_P (cond
))
2447 struct edge_info
*edge_info
;
2449 edge_info
= allocate_edge_info (true_edge
);
2450 edge_info
->lhs
= cond
;
2451 edge_info
->rhs
= constant_boolean_node (1, TREE_TYPE (cond
));
2453 edge_info
= allocate_edge_info (false_edge
);
2454 edge_info
->lhs
= cond
;
2455 edge_info
->rhs
= constant_boolean_node (0, TREE_TYPE (cond
));
2457 /* Equality tests may create one or two equivalences. */
2458 else if (COMPARISON_CLASS_P (cond
))
2460 tree op0
= TREE_OPERAND (cond
, 0);
2461 tree op1
= TREE_OPERAND (cond
, 1);
2463 /* Special case comparing booleans against a constant as we
2464 know the value of OP0 on both arms of the branch. i.e., we
2465 can record an equivalence for OP0 rather than COND. */
2466 if ((TREE_CODE (cond
) == EQ_EXPR
|| TREE_CODE (cond
) == NE_EXPR
)
2467 && TREE_CODE (op0
) == SSA_NAME
2468 && TREE_CODE (TREE_TYPE (op0
)) == BOOLEAN_TYPE
2469 && is_gimple_min_invariant (op1
))
2471 if (TREE_CODE (cond
) == EQ_EXPR
)
2473 edge_info
= allocate_edge_info (true_edge
);
2474 edge_info
->lhs
= op0
;
2475 edge_info
->rhs
= (integer_zerop (op1
)
2476 ? boolean_false_node
2477 : boolean_true_node
);
2479 edge_info
= allocate_edge_info (false_edge
);
2480 edge_info
->lhs
= op0
;
2481 edge_info
->rhs
= (integer_zerop (op1
)
2483 : boolean_false_node
);
2487 edge_info
= allocate_edge_info (true_edge
);
2488 edge_info
->lhs
= op0
;
2489 edge_info
->rhs
= (integer_zerop (op1
)
2491 : boolean_false_node
);
2493 edge_info
= allocate_edge_info (false_edge
);
2494 edge_info
->lhs
= op0
;
2495 edge_info
->rhs
= (integer_zerop (op1
)
2496 ? boolean_false_node
2497 : boolean_true_node
);
2501 else if (is_gimple_min_invariant (op0
)
2502 && (TREE_CODE (op1
) == SSA_NAME
2503 || is_gimple_min_invariant (op1
)))
2505 tree inverted
= invert_truthvalue (cond
);
2506 struct edge_info
*edge_info
;
2508 edge_info
= allocate_edge_info (true_edge
);
2509 record_conditions (edge_info
, cond
, inverted
);
2511 if (TREE_CODE (cond
) == EQ_EXPR
)
2513 edge_info
->lhs
= op1
;
2514 edge_info
->rhs
= op0
;
2517 edge_info
= allocate_edge_info (false_edge
);
2518 record_conditions (edge_info
, inverted
, cond
);
2520 if (TREE_CODE (cond
) == NE_EXPR
)
2522 edge_info
->lhs
= op1
;
2523 edge_info
->rhs
= op0
;
2527 else if (TREE_CODE (op0
) == SSA_NAME
2528 && (is_gimple_min_invariant (op1
)
2529 || TREE_CODE (op1
) == SSA_NAME
))
2531 tree inverted
= invert_truthvalue (cond
);
2532 struct edge_info
*edge_info
;
2534 edge_info
= allocate_edge_info (true_edge
);
2535 record_conditions (edge_info
, cond
, inverted
);
2537 if (TREE_CODE (cond
) == EQ_EXPR
)
2539 edge_info
->lhs
= op0
;
2540 edge_info
->rhs
= op1
;
2543 edge_info
= allocate_edge_info (false_edge
);
2544 record_conditions (edge_info
, inverted
, cond
);
2546 if (TREE_CODE (cond
) == NE_EXPR
)
2548 edge_info
->lhs
= op0
;
2549 edge_info
->rhs
= op1
;
2554 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2559 /* Propagate information from BB to its outgoing edges.
2561 This can include equivalency information implied by control statements
2562 at the end of BB and const/copy propagation into PHIs in BB's
2563 successor blocks. */
2566 propagate_to_outgoing_edges (struct dom_walk_data
*walk_data ATTRIBUTE_UNUSED
,
2570 record_edge_info (bb
);
2571 cprop_into_successor_phis (bb
, nonzero_vars
);
2574 /* Search for redundant computations in STMT. If any are found, then
2575 replace them with the variable holding the result of the computation.
2577 If safe, record this expression into the available expression hash
2581 eliminate_redundant_computations (struct dom_walk_data
*walk_data
,
2582 tree stmt
, stmt_ann_t ann
)
2584 v_may_def_optype v_may_defs
= V_MAY_DEF_OPS (ann
);
2585 tree
*expr_p
, def
= NULL_TREE
;
2588 bool retval
= false;
2590 if (TREE_CODE (stmt
) == MODIFY_EXPR
)
2591 def
= TREE_OPERAND (stmt
, 0);
2593 /* Certain expressions on the RHS can be optimized away, but can not
2594 themselves be entered into the hash tables. */
2595 if (ann
->makes_aliased_stores
2597 || TREE_CODE (def
) != SSA_NAME
2598 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def
)
2599 || NUM_V_MAY_DEFS (v_may_defs
) != 0
2600 /* Do not record equivalences for increments of ivs. This would create
2601 overlapping live ranges for a very questionable gain. */
2602 || simple_iv_increment_p (stmt
))
2605 /* Check if the expression has been computed before. */
2606 cached_lhs
= lookup_avail_expr (stmt
, insert
);
2608 /* If this is an assignment and the RHS was not in the hash table,
2609 then try to simplify the RHS and lookup the new RHS in the
2611 if (! cached_lhs
&& TREE_CODE (stmt
) == MODIFY_EXPR
)
2612 cached_lhs
= simplify_rhs_and_lookup_avail_expr (walk_data
, stmt
, insert
);
2613 /* Similarly if this is a COND_EXPR and we did not find its
2614 expression in the hash table, simplify the condition and
2616 else if (! cached_lhs
&& TREE_CODE (stmt
) == COND_EXPR
)
2617 cached_lhs
= simplify_cond_and_lookup_avail_expr (stmt
, ann
, insert
);
2618 /* Similarly for a SWITCH_EXPR. */
2619 else if (!cached_lhs
&& TREE_CODE (stmt
) == SWITCH_EXPR
)
2620 cached_lhs
= simplify_switch_and_lookup_avail_expr (stmt
, insert
);
2622 opt_stats
.num_exprs_considered
++;
2624 /* Get a pointer to the expression we are trying to optimize. */
2625 if (TREE_CODE (stmt
) == COND_EXPR
)
2626 expr_p
= &COND_EXPR_COND (stmt
);
2627 else if (TREE_CODE (stmt
) == SWITCH_EXPR
)
2628 expr_p
= &SWITCH_COND (stmt
);
2629 else if (TREE_CODE (stmt
) == RETURN_EXPR
&& TREE_OPERAND (stmt
, 0))
2630 expr_p
= &TREE_OPERAND (TREE_OPERAND (stmt
, 0), 1);
2632 expr_p
= &TREE_OPERAND (stmt
, 1);
2634 /* It is safe to ignore types here since we have already done
2635 type checking in the hashing and equality routines. In fact
2636 type checking here merely gets in the way of constant
2637 propagation. Also, make sure that it is safe to propagate
2638 CACHED_LHS into *EXPR_P. */
2640 && (TREE_CODE (cached_lhs
) != SSA_NAME
2641 || may_propagate_copy (*expr_p
, cached_lhs
)))
2643 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2645 fprintf (dump_file
, " Replaced redundant expr '");
2646 print_generic_expr (dump_file
, *expr_p
, dump_flags
);
2647 fprintf (dump_file
, "' with '");
2648 print_generic_expr (dump_file
, cached_lhs
, dump_flags
);
2649 fprintf (dump_file
, "'\n");
2654 #if defined ENABLE_CHECKING
2655 gcc_assert (TREE_CODE (cached_lhs
) == SSA_NAME
2656 || is_gimple_min_invariant (cached_lhs
));
2659 if (TREE_CODE (cached_lhs
) == ADDR_EXPR
2660 || (POINTER_TYPE_P (TREE_TYPE (*expr_p
))
2661 && is_gimple_min_invariant (cached_lhs
)))
2664 propagate_tree_value (expr_p
, cached_lhs
);
2670 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2671 the available expressions table or the const_and_copies table.
2672 Detect and record those equivalences. */
2675 record_equivalences_from_stmt (tree stmt
,
2679 tree lhs
= TREE_OPERAND (stmt
, 0);
2680 enum tree_code lhs_code
= TREE_CODE (lhs
);
2683 if (lhs_code
== SSA_NAME
)
2685 tree rhs
= TREE_OPERAND (stmt
, 1);
2687 /* Strip away any useless type conversions. */
2688 STRIP_USELESS_TYPE_CONVERSION (rhs
);
2690 /* If the RHS of the assignment is a constant or another variable that
2691 may be propagated, register it in the CONST_AND_COPIES table. We
2692 do not need to record unwind data for this, since this is a true
2693 assignment and not an equivalence inferred from a comparison. All
2694 uses of this ssa name are dominated by this assignment, so unwinding
2695 just costs time and space. */
2697 && (TREE_CODE (rhs
) == SSA_NAME
2698 || is_gimple_min_invariant (rhs
)))
2699 SSA_NAME_VALUE (lhs
) = rhs
;
2701 /* alloca never returns zero and the address of a non-weak symbol
2702 is never zero. NOP_EXPRs and CONVERT_EXPRs can be completely
2703 stripped as they do not affect this equivalence. */
2704 while (TREE_CODE (rhs
) == NOP_EXPR
2705 || TREE_CODE (rhs
) == CONVERT_EXPR
)
2706 rhs
= TREE_OPERAND (rhs
, 0);
2708 if (alloca_call_p (rhs
)
2709 || (TREE_CODE (rhs
) == ADDR_EXPR
2710 && DECL_P (TREE_OPERAND (rhs
, 0))
2711 && ! DECL_WEAK (TREE_OPERAND (rhs
, 0))))
2712 record_var_is_nonzero (lhs
);
2714 /* IOR of any value with a nonzero value will result in a nonzero
2715 value. Even if we do not know the exact result recording that
2716 the result is nonzero is worth the effort. */
2717 if (TREE_CODE (rhs
) == BIT_IOR_EXPR
2718 && integer_nonzerop (TREE_OPERAND (rhs
, 1)))
2719 record_var_is_nonzero (lhs
);
2722 /* Look at both sides for pointer dereferences. If we find one, then
2723 the pointer must be nonnull and we can enter that equivalence into
2725 if (flag_delete_null_pointer_checks
)
2726 for (i
= 0; i
< 2; i
++)
2728 tree t
= TREE_OPERAND (stmt
, i
);
2730 /* Strip away any COMPONENT_REFs. */
2731 while (TREE_CODE (t
) == COMPONENT_REF
)
2732 t
= TREE_OPERAND (t
, 0);
2734 /* Now see if this is a pointer dereference. */
2735 if (INDIRECT_REF_P (t
))
2737 tree op
= TREE_OPERAND (t
, 0);
2739 /* If the pointer is a SSA variable, then enter new
2740 equivalences into the hash table. */
2741 while (TREE_CODE (op
) == SSA_NAME
)
2743 tree def
= SSA_NAME_DEF_STMT (op
);
2745 record_var_is_nonzero (op
);
2747 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2748 which are known to have a nonzero value. */
2750 && TREE_CODE (def
) == MODIFY_EXPR
2751 && TREE_CODE (TREE_OPERAND (def
, 1)) == NOP_EXPR
)
2752 op
= TREE_OPERAND (TREE_OPERAND (def
, 1), 0);
2759 /* A memory store, even an aliased store, creates a useful
2760 equivalence. By exchanging the LHS and RHS, creating suitable
2761 vops and recording the result in the available expression table,
2762 we may be able to expose more redundant loads. */
2763 if (!ann
->has_volatile_ops
2764 && (TREE_CODE (TREE_OPERAND (stmt
, 1)) == SSA_NAME
2765 || is_gimple_min_invariant (TREE_OPERAND (stmt
, 1)))
2766 && !is_gimple_reg (lhs
))
2768 tree rhs
= TREE_OPERAND (stmt
, 1);
2771 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2772 is a constant, we need to adjust the constant to fit into the
2773 type of the LHS. If the LHS is a bitfield and the RHS is not
2774 a constant, then we can not record any equivalences for this
2775 statement since we would need to represent the widening or
2776 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2777 and should not be necessary if GCC represented bitfields
2779 if (lhs_code
== COMPONENT_REF
2780 && DECL_BIT_FIELD (TREE_OPERAND (lhs
, 1)))
2782 if (TREE_CONSTANT (rhs
))
2783 rhs
= widen_bitfield (rhs
, TREE_OPERAND (lhs
, 1), lhs
);
2787 /* If the value overflowed, then we can not use this equivalence. */
2788 if (rhs
&& ! is_gimple_min_invariant (rhs
))
2794 /* Build a new statement with the RHS and LHS exchanged. */
2795 new = build (MODIFY_EXPR
, TREE_TYPE (stmt
), rhs
, lhs
);
2797 create_ssa_artficial_load_stmt (&(ann
->operands
), new);
2799 /* Finally enter the statement into the available expression
2801 lookup_avail_expr (new, true);
2806 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2807 CONST_AND_COPIES. */
2810 cprop_operand (tree stmt
, use_operand_p op_p
)
2812 bool may_have_exposed_new_symbols
= false;
2814 tree op
= USE_FROM_PTR (op_p
);
2816 /* If the operand has a known constant value or it is known to be a
2817 copy of some other variable, use the value or copy stored in
2818 CONST_AND_COPIES. */
2819 val
= SSA_NAME_VALUE (op
);
2820 if (val
&& TREE_CODE (val
) != VALUE_HANDLE
)
2822 tree op_type
, val_type
;
2824 /* Do not change the base variable in the virtual operand
2825 tables. That would make it impossible to reconstruct
2826 the renamed virtual operand if we later modify this
2827 statement. Also only allow the new value to be an SSA_NAME
2828 for propagation into virtual operands. */
2829 if (!is_gimple_reg (op
)
2830 && (get_virtual_var (val
) != get_virtual_var (op
)
2831 || TREE_CODE (val
) != SSA_NAME
))
2834 /* Do not replace hard register operands in asm statements. */
2835 if (TREE_CODE (stmt
) == ASM_EXPR
2836 && !may_propagate_copy_into_asm (op
))
2839 /* Get the toplevel type of each operand. */
2840 op_type
= TREE_TYPE (op
);
2841 val_type
= TREE_TYPE (val
);
2843 /* While both types are pointers, get the type of the object
2845 while (POINTER_TYPE_P (op_type
) && POINTER_TYPE_P (val_type
))
2847 op_type
= TREE_TYPE (op_type
);
2848 val_type
= TREE_TYPE (val_type
);
2851 /* Make sure underlying types match before propagating a constant by
2852 converting the constant to the proper type. Note that convert may
2853 return a non-gimple expression, in which case we ignore this
2854 propagation opportunity. */
2855 if (TREE_CODE (val
) != SSA_NAME
)
2857 if (!lang_hooks
.types_compatible_p (op_type
, val_type
))
2859 val
= fold_convert (TREE_TYPE (op
), val
);
2860 if (!is_gimple_min_invariant (val
))
2865 /* Certain operands are not allowed to be copy propagated due
2866 to their interaction with exception handling and some GCC
2868 else if (!may_propagate_copy (op
, val
))
2871 /* Do not propagate copies if the propagated value is at a deeper loop
2872 depth than the propagatee. Otherwise, this may move loop variant
2873 variables outside of their loops and prevent coalescing
2874 opportunities. If the value was loop invariant, it will be hoisted
2875 by LICM and exposed for copy propagation. */
2876 if (loop_depth_of_name (val
) > loop_depth_of_name (op
))
2880 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2882 fprintf (dump_file
, " Replaced '");
2883 print_generic_expr (dump_file
, op
, dump_flags
);
2884 fprintf (dump_file
, "' with %s '",
2885 (TREE_CODE (val
) != SSA_NAME
? "constant" : "variable"));
2886 print_generic_expr (dump_file
, val
, dump_flags
);
2887 fprintf (dump_file
, "'\n");
2890 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2891 that we may have exposed a new symbol for SSA renaming. */
2892 if (TREE_CODE (val
) == ADDR_EXPR
2893 || (POINTER_TYPE_P (TREE_TYPE (op
))
2894 && is_gimple_min_invariant (val
)))
2895 may_have_exposed_new_symbols
= true;
2897 propagate_value (op_p
, val
);
2899 /* And note that we modified this statement. This is now
2900 safe, even if we changed virtual operands since we will
2901 rescan the statement and rewrite its operands again. */
2904 return may_have_exposed_new_symbols
;
2907 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2908 known value for that SSA_NAME (or NULL if no value is known).
2910 Propagate values from CONST_AND_COPIES into the uses, vuses and
2911 v_may_def_ops of STMT. */
2914 cprop_into_stmt (tree stmt
)
2916 bool may_have_exposed_new_symbols
= false;
2921 FOR_EACH_SSA_USE_OPERAND (op_p
, stmt
, iter
, SSA_OP_ALL_USES
)
2923 if (TREE_CODE (USE_FROM_PTR (op_p
)) == SSA_NAME
)
2924 may_have_exposed_new_symbols
|= cprop_operand (stmt
, op_p
);
2927 if (may_have_exposed_new_symbols
)
2929 rhs
= get_rhs (stmt
);
2930 if (rhs
&& TREE_CODE (rhs
) == ADDR_EXPR
)
2931 recompute_tree_invarant_for_addr_expr (rhs
);
2934 return may_have_exposed_new_symbols
;
2938 /* Optimize the statement pointed by iterator SI.
2940 We try to perform some simplistic global redundancy elimination and
2941 constant propagation:
2943 1- To detect global redundancy, we keep track of expressions that have
2944 been computed in this block and its dominators. If we find that the
2945 same expression is computed more than once, we eliminate repeated
2946 computations by using the target of the first one.
2948 2- Constant values and copy assignments. This is used to do very
2949 simplistic constant and copy propagation. When a constant or copy
2950 assignment is found, we map the value on the RHS of the assignment to
2951 the variable in the LHS in the CONST_AND_COPIES table. */
2954 optimize_stmt (struct dom_walk_data
*walk_data
, basic_block bb
,
2955 block_stmt_iterator si
)
2959 bool may_optimize_p
;
2960 bool may_have_exposed_new_symbols
= false;
2962 stmt
= bsi_stmt (si
);
2964 get_stmt_operands (stmt
);
2965 ann
= stmt_ann (stmt
);
2966 opt_stats
.num_stmts
++;
2967 may_have_exposed_new_symbols
= false;
2969 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2971 fprintf (dump_file
, "Optimizing statement ");
2972 print_generic_stmt (dump_file
, stmt
, TDF_SLIM
);
2975 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2976 may_have_exposed_new_symbols
= cprop_into_stmt (stmt
);
2978 /* If the statement has been modified with constant replacements,
2979 fold its RHS before checking for redundant computations. */
2982 /* Try to fold the statement making sure that STMT is kept
2984 if (fold_stmt (bsi_stmt_ptr (si
)))
2986 stmt
= bsi_stmt (si
);
2987 ann
= stmt_ann (stmt
);
2989 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2991 fprintf (dump_file
, " Folded to: ");
2992 print_generic_stmt (dump_file
, stmt
, TDF_SLIM
);
2996 /* Constant/copy propagation above may change the set of
2997 virtual operands associated with this statement. Folding
2998 may remove the need for some virtual operands.
3000 Indicate we will need to rescan and rewrite the statement. */
3001 may_have_exposed_new_symbols
= true;
3004 /* Check for redundant computations. Do this optimization only
3005 for assignments that have no volatile ops and conditionals. */
3006 may_optimize_p
= (!ann
->has_volatile_ops
3007 && ((TREE_CODE (stmt
) == RETURN_EXPR
3008 && TREE_OPERAND (stmt
, 0)
3009 && TREE_CODE (TREE_OPERAND (stmt
, 0)) == MODIFY_EXPR
3010 && ! (TREE_SIDE_EFFECTS
3011 (TREE_OPERAND (TREE_OPERAND (stmt
, 0), 1))))
3012 || (TREE_CODE (stmt
) == MODIFY_EXPR
3013 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt
, 1)))
3014 || TREE_CODE (stmt
) == COND_EXPR
3015 || TREE_CODE (stmt
) == SWITCH_EXPR
));
3018 may_have_exposed_new_symbols
3019 |= eliminate_redundant_computations (walk_data
, stmt
, ann
);
3021 /* Record any additional equivalences created by this statement. */
3022 if (TREE_CODE (stmt
) == MODIFY_EXPR
)
3023 record_equivalences_from_stmt (stmt
,
3027 register_definitions_for_stmt (stmt
);
3029 /* If STMT is a COND_EXPR and it was modified, then we may know
3030 where it goes. If that is the case, then mark the CFG as altered.
3032 This will cause us to later call remove_unreachable_blocks and
3033 cleanup_tree_cfg when it is safe to do so. It is not safe to
3034 clean things up here since removal of edges and such can trigger
3035 the removal of PHI nodes, which in turn can release SSA_NAMEs to
3038 That's all fine and good, except that once SSA_NAMEs are released
3039 to the manager, we must not call create_ssa_name until all references
3040 to released SSA_NAMEs have been eliminated.
3042 All references to the deleted SSA_NAMEs can not be eliminated until
3043 we remove unreachable blocks.
3045 We can not remove unreachable blocks until after we have completed
3046 any queued jump threading.
3048 We can not complete any queued jump threads until we have taken
3049 appropriate variables out of SSA form. Taking variables out of
3050 SSA form can call create_ssa_name and thus we lose.
3052 Ultimately I suspect we're going to need to change the interface
3053 into the SSA_NAME manager. */
3059 if (TREE_CODE (stmt
) == COND_EXPR
)
3060 val
= COND_EXPR_COND (stmt
);
3061 else if (TREE_CODE (stmt
) == SWITCH_EXPR
)
3062 val
= SWITCH_COND (stmt
);
3064 if (val
&& TREE_CODE (val
) == INTEGER_CST
&& find_taken_edge (bb
, val
))
3067 /* If we simplified a statement in such a way as to be shown that it
3068 cannot trap, update the eh information and the cfg to match. */
3069 if (maybe_clean_eh_stmt (stmt
))
3071 bitmap_set_bit (need_eh_cleanup
, bb
->index
);
3072 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3073 fprintf (dump_file
, " Flagged to clear EH edges.\n");
3077 if (may_have_exposed_new_symbols
)
3078 VEC_safe_push (tree_on_heap
, stmts_to_rescan
, bsi_stmt (si
));
3081 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
3082 available expression hashtable, then return the LHS from the hash
3085 If INSERT is true, then we also update the available expression
3086 hash table to account for the changes made to STMT. */
3089 update_rhs_and_lookup_avail_expr (tree stmt
, tree new_rhs
, bool insert
)
3091 tree cached_lhs
= NULL
;
3093 /* Remove the old entry from the hash table. */
3096 struct expr_hash_elt element
;
3098 initialize_hash_element (stmt
, NULL
, &element
);
3099 htab_remove_elt_with_hash (avail_exprs
, &element
, element
.hash
);
3102 /* Now update the RHS of the assignment. */
3103 TREE_OPERAND (stmt
, 1) = new_rhs
;
3105 /* Now lookup the updated statement in the hash table. */
3106 cached_lhs
= lookup_avail_expr (stmt
, insert
);
3108 /* We have now called lookup_avail_expr twice with two different
3109 versions of this same statement, once in optimize_stmt, once here.
3111 We know the call in optimize_stmt did not find an existing entry
3112 in the hash table, so a new entry was created. At the same time
3113 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
3115 If this call failed to find an existing entry on the hash table,
3116 then the new version of this statement was entered into the
3117 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3118 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3120 If this call succeeded, we still have one copy of this statement
3121 on the BLOCK_AVAIL_EXPRs vector.
3123 For both cases, we need to pop the most recent entry off the
3124 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3125 statement in the hash tables, that will leave precisely one
3126 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3127 we found a copy of this statement in the second hash table lookup
3128 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3130 VEC_pop (tree_on_heap
, avail_exprs_stack
);
3132 /* And make sure we record the fact that we modified this
3139 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3140 found, return its LHS. Otherwise insert STMT in the table and return
3143 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3144 is also added to the stack pointed by BLOCK_AVAIL_EXPRS_P, so that they
3145 can be removed when we finish processing this block and its children.
3147 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3148 contains no CALL_EXPR on its RHS and makes no volatile nor
3149 aliased references. */
3152 lookup_avail_expr (tree stmt
, bool insert
)
3157 struct expr_hash_elt
*element
= xmalloc (sizeof (struct expr_hash_elt
));
3159 lhs
= TREE_CODE (stmt
) == MODIFY_EXPR
? TREE_OPERAND (stmt
, 0) : NULL
;
3161 initialize_hash_element (stmt
, lhs
, element
);
3163 /* Don't bother remembering constant assignments and copy operations.
3164 Constants and copy operations are handled by the constant/copy propagator
3165 in optimize_stmt. */
3166 if (TREE_CODE (element
->rhs
) == SSA_NAME
3167 || is_gimple_min_invariant (element
->rhs
))
3173 /* If this is an equality test against zero, see if we have recorded a
3174 nonzero value for the variable in question. */
3175 if ((TREE_CODE (element
->rhs
) == EQ_EXPR
3176 || TREE_CODE (element
->rhs
) == NE_EXPR
)
3177 && TREE_CODE (TREE_OPERAND (element
->rhs
, 0)) == SSA_NAME
3178 && integer_zerop (TREE_OPERAND (element
->rhs
, 1)))
3180 int indx
= SSA_NAME_VERSION (TREE_OPERAND (element
->rhs
, 0));
3182 if (bitmap_bit_p (nonzero_vars
, indx
))
3184 tree t
= element
->rhs
;
3187 if (TREE_CODE (t
) == EQ_EXPR
)
3188 return boolean_false_node
;
3190 return boolean_true_node
;
3194 /* Finally try to find the expression in the main expression hash table. */
3195 slot
= htab_find_slot_with_hash (avail_exprs
, element
, element
->hash
,
3196 (insert
? INSERT
: NO_INSERT
));
3205 *slot
= (void *) element
;
3206 VEC_safe_push (tree_on_heap
, avail_exprs_stack
,
3207 stmt
? stmt
: element
->rhs
);
3211 /* Extract the LHS of the assignment so that it can be used as the current
3212 definition of another variable. */
3213 lhs
= ((struct expr_hash_elt
*)*slot
)->lhs
;
3215 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3216 use the value from the const_and_copies table. */
3217 if (TREE_CODE (lhs
) == SSA_NAME
)
3219 temp
= SSA_NAME_VALUE (lhs
);
3220 if (temp
&& TREE_CODE (temp
) != VALUE_HANDLE
)
3228 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3229 range of values that result in the conditional having a true value.
3231 Return true if we are successful in extracting a range from COND and
3232 false if we are unsuccessful. */
3235 extract_range_from_cond (tree cond
, tree
*hi_p
, tree
*lo_p
, int *inverted_p
)
3237 tree op1
= TREE_OPERAND (cond
, 1);
3238 tree high
, low
, type
;
3241 type
= TREE_TYPE (op1
);
3243 /* Experiments have shown that it's rarely, if ever useful to
3244 record ranges for enumerations. Presumably this is due to
3245 the fact that they're rarely used directly. They are typically
3246 cast into an integer type and used that way. */
3247 if (TREE_CODE (type
) != INTEGER_TYPE
3248 /* We don't know how to deal with types with variable bounds. */
3249 || TREE_CODE (TYPE_MIN_VALUE (type
)) != INTEGER_CST
3250 || TREE_CODE (TYPE_MAX_VALUE (type
)) != INTEGER_CST
)
3253 switch (TREE_CODE (cond
))
3267 high
= TYPE_MAX_VALUE (type
);
3272 high
= TYPE_MAX_VALUE (type
);
3273 if (!tree_int_cst_lt (op1
, high
))
3275 low
= int_const_binop (PLUS_EXPR
, op1
, integer_one_node
, 1);
3281 low
= TYPE_MIN_VALUE (type
);
3286 low
= TYPE_MIN_VALUE (type
);
3287 if (!tree_int_cst_lt (low
, op1
))
3289 high
= int_const_binop (MINUS_EXPR
, op1
, integer_one_node
, 1);
3299 *inverted_p
= inverted
;
3303 /* Record a range created by COND for basic block BB. */
3306 record_range (tree cond
, basic_block bb
)
3308 enum tree_code code
= TREE_CODE (cond
);
3310 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3311 They rarely allow for meaningful range optimizations and significantly
3312 complicate the implementation. */
3313 if ((code
== LT_EXPR
|| code
== LE_EXPR
|| code
== GT_EXPR
3314 || code
== GE_EXPR
|| code
== EQ_EXPR
)
3315 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond
, 1))) == INTEGER_TYPE
)
3317 struct vrp_hash_elt
*vrp_hash_elt
;
3318 struct vrp_element
*element
;
3319 varray_type
*vrp_records_p
;
3323 vrp_hash_elt
= xmalloc (sizeof (struct vrp_hash_elt
));
3324 vrp_hash_elt
->var
= TREE_OPERAND (cond
, 0);
3325 vrp_hash_elt
->records
= NULL
;
3326 slot
= htab_find_slot (vrp_data
, vrp_hash_elt
, INSERT
);
3329 *slot
= (void *) vrp_hash_elt
;
3331 free (vrp_hash_elt
);
3333 vrp_hash_elt
= (struct vrp_hash_elt
*) *slot
;
3334 vrp_records_p
= &vrp_hash_elt
->records
;
3336 element
= ggc_alloc (sizeof (struct vrp_element
));
3337 element
->low
= NULL
;
3338 element
->high
= NULL
;
3339 element
->cond
= cond
;
3342 if (*vrp_records_p
== NULL
)
3343 VARRAY_GENERIC_PTR_INIT (*vrp_records_p
, 2, "vrp records");
3345 VARRAY_PUSH_GENERIC_PTR (*vrp_records_p
, element
);
3346 VEC_safe_push (tree_on_heap
, vrp_variables_stack
, TREE_OPERAND (cond
, 0));
3350 /* Hashing and equality functions for VRP_DATA.
3352 Since this hash table is addressed by SSA_NAMEs, we can hash on
3353 their version number and equality can be determined with a
3354 pointer comparison. */
3357 vrp_hash (const void *p
)
3359 tree var
= ((struct vrp_hash_elt
*)p
)->var
;
3361 return SSA_NAME_VERSION (var
);
3365 vrp_eq (const void *p1
, const void *p2
)
3367 tree var1
= ((struct vrp_hash_elt
*)p1
)->var
;
3368 tree var2
= ((struct vrp_hash_elt
*)p2
)->var
;
3370 return var1
== var2
;
3373 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3374 MODIFY_EXPR statements. We compute a value number for expressions using
3375 the code of the expression and the SSA numbers of its operands. */
3378 avail_expr_hash (const void *p
)
3380 stmt_ann_t ann
= ((struct expr_hash_elt
*)p
)->ann
;
3381 tree rhs
= ((struct expr_hash_elt
*)p
)->rhs
;
3386 /* iterative_hash_expr knows how to deal with any expression and
3387 deals with commutative operators as well, so just use it instead
3388 of duplicating such complexities here. */
3389 val
= iterative_hash_expr (rhs
, val
);
3391 /* If the hash table entry is not associated with a statement, then we
3392 can just hash the expression and not worry about virtual operands
3397 /* Add the SSA version numbers of every vuse operand. This is important
3398 because compound variables like arrays are not renamed in the
3399 operands. Rather, the rename is done on the virtual variable
3400 representing all the elements of the array. */
3401 vuses
= VUSE_OPS (ann
);
3402 for (i
= 0; i
< NUM_VUSES (vuses
); i
++)
3403 val
= iterative_hash_expr (VUSE_OP (vuses
, i
), val
);
3409 real_avail_expr_hash (const void *p
)
3411 return ((const struct expr_hash_elt
*)p
)->hash
;
3415 avail_expr_eq (const void *p1
, const void *p2
)
3417 stmt_ann_t ann1
= ((struct expr_hash_elt
*)p1
)->ann
;
3418 tree rhs1
= ((struct expr_hash_elt
*)p1
)->rhs
;
3419 stmt_ann_t ann2
= ((struct expr_hash_elt
*)p2
)->ann
;
3420 tree rhs2
= ((struct expr_hash_elt
*)p2
)->rhs
;
3422 /* If they are the same physical expression, return true. */
3423 if (rhs1
== rhs2
&& ann1
== ann2
)
3426 /* If their codes are not equal, then quit now. */
3427 if (TREE_CODE (rhs1
) != TREE_CODE (rhs2
))
3430 /* In case of a collision, both RHS have to be identical and have the
3431 same VUSE operands. */
3432 if ((TREE_TYPE (rhs1
) == TREE_TYPE (rhs2
)
3433 || lang_hooks
.types_compatible_p (TREE_TYPE (rhs1
), TREE_TYPE (rhs2
)))
3434 && operand_equal_p (rhs1
, rhs2
, OEP_PURE_SAME
))
3436 vuse_optype ops1
= NULL
;
3437 vuse_optype ops2
= NULL
;
3438 size_t num_ops1
= 0;
3439 size_t num_ops2
= 0;
3444 ops1
= VUSE_OPS (ann1
);
3445 num_ops1
= NUM_VUSES (ops1
);
3450 ops2
= VUSE_OPS (ann2
);
3451 num_ops2
= NUM_VUSES (ops2
);
3454 /* If the number of virtual uses is different, then we consider
3456 if (num_ops1
!= num_ops2
)
3459 for (i
= 0; i
< num_ops1
; i
++)
3460 if (VUSE_OP (ops1
, i
) != VUSE_OP (ops2
, i
))
3463 gcc_assert (((struct expr_hash_elt
*)p1
)->hash
3464 == ((struct expr_hash_elt
*)p2
)->hash
);
3471 /* Given STMT and a pointer to the block local definitions BLOCK_DEFS_P,
3472 register register all objects set by this statement into BLOCK_DEFS_P
3476 register_definitions_for_stmt (tree stmt
)
3481 FOR_EACH_SSA_TREE_OPERAND (def
, stmt
, iter
, SSA_OP_ALL_DEFS
)
3484 /* FIXME: We shouldn't be registering new defs if the variable
3485 doesn't need to be renamed. */
3486 register_new_def (def
, &block_defs_stack
);