2005-06-30 J. D. Johnston <jjohnst@us.ibm.com>
[official-gcc.git] / gcc / tree-ssa-dom.c
blob86da07b437f3c8578a1669a98c9ae1b105cfc79e
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)
10 any later version.
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to
19 the Free Software Foundation, 51 Franklin Street, Fifth Floor,
20 Boston, MA 02110-1301, USA. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "tree.h"
27 #include "flags.h"
28 #include "rtl.h"
29 #include "tm_p.h"
30 #include "ggc.h"
31 #include "basic-block.h"
32 #include "cfgloop.h"
33 #include "output.h"
34 #include "expr.h"
35 #include "function.h"
36 #include "diagnostic.h"
37 #include "timevar.h"
38 #include "tree-dump.h"
39 #include "tree-flow.h"
40 #include "domwalk.h"
41 #include "real.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. */
63 struct edge_info
65 /* If this edge creates a simple equivalence, the LHS and RHS of
66 the equivalence will be stored here. */
67 tree lhs;
68 tree rhs;
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
88 in this table. */
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
95 marker. */
96 static VEC(tree,heap) *avail_exprs_stack;
98 /* Stack of statements we need to rescan during finalization for newly
99 exposed variables.
101 Statement rescanning must occur after the current block's available
102 expressions are removed from AVAIL_EXPRS. Else we may change the
103 hash code for an expression and be unable to find/remove it from
104 AVAIL_EXPRS. */
105 static VEC(tree,heap) *stmts_to_rescan;
107 /* Structure for entries in the expression hash table.
109 This requires more memory for the hash table entries, but allows us
110 to avoid creating silly tree nodes and annotations for conditionals,
111 eliminates 2 global hash tables and two block local varrays.
113 It also allows us to reduce the number of hash table lookups we
114 have to perform in lookup_avail_expr and finally it allows us to
115 significantly reduce the number of calls into the hashing routine
116 itself. */
118 struct expr_hash_elt
120 /* The value (lhs) of this expression. */
121 tree lhs;
123 /* The expression (rhs) we want to record. */
124 tree rhs;
126 /* The stmt pointer if this element corresponds to a statement. */
127 tree stmt;
129 /* The hash value for RHS/ann. */
130 hashval_t hash;
133 /* Stack of dest,src pairs that need to be restored during finalization.
135 A NULL entry is used to mark the end of pairs which need to be
136 restored during finalization of this block. */
137 static VEC(tree,heap) *const_and_copies_stack;
139 /* Bitmap of SSA_NAMEs known to have a nonzero value, even if we do not
140 know their exact value. */
141 static bitmap nonzero_vars;
143 /* Bitmap of blocks that are scheduled to be threaded through. This
144 is used to communicate with thread_through_blocks. */
145 static bitmap threaded_blocks;
147 /* Stack of SSA_NAMEs which need their NONZERO_VARS property cleared
148 when the current block is finalized.
150 A NULL entry is used to mark the end of names needing their
151 entry in NONZERO_VARS cleared during finalization of this block. */
152 static VEC(tree,heap) *nonzero_vars_stack;
154 /* Track whether or not we have changed the control flow graph. */
155 static bool cfg_altered;
157 /* Bitmap of blocks that have had EH statements cleaned. We should
158 remove their dead edges eventually. */
159 static bitmap need_eh_cleanup;
161 /* Statistics for dominator optimizations. */
162 struct opt_stats_d
164 long num_stmts;
165 long num_exprs_considered;
166 long num_re;
167 long num_const_prop;
168 long num_copy_prop;
171 static struct opt_stats_d opt_stats;
173 /* Value range propagation record. Each time we encounter a conditional
174 of the form SSA_NAME COND CONST we create a new vrp_element to record
175 how the condition affects the possible values SSA_NAME may have.
177 Each record contains the condition tested (COND), and the range of
178 values the variable may legitimately have if COND is true. Note the
179 range of values may be a smaller range than COND specifies if we have
180 recorded other ranges for this variable. Each record also contains the
181 block in which the range was recorded for invalidation purposes.
183 Note that the current known range is computed lazily. This allows us
184 to avoid the overhead of computing ranges which are never queried.
186 When we encounter a conditional, we look for records which constrain
187 the SSA_NAME used in the condition. In some cases those records allow
188 us to determine the condition's result at compile time. In other cases
189 they may allow us to simplify the condition.
191 We also use value ranges to do things like transform signed div/mod
192 operations into unsigned div/mod or to simplify ABS_EXPRs.
194 Simple experiments have shown these optimizations to not be all that
195 useful on switch statements (much to my surprise). So switch statement
196 optimizations are not performed.
198 Note carefully we do not propagate information through each statement
199 in the block. i.e., if we know variable X has a value defined of
200 [0, 25] and we encounter Y = X + 1, we do not track a value range
201 for Y (which would be [1, 26] if we cared). Similarly we do not
202 constrain values as we encounter narrowing typecasts, etc. */
204 struct vrp_element
206 /* The highest and lowest values the variable in COND may contain when
207 COND is true. Note this may not necessarily be the same values
208 tested by COND if the same variable was used in earlier conditionals.
210 Note this is computed lazily and thus can be NULL indicating that
211 the values have not been computed yet. */
212 tree low;
213 tree high;
215 /* The actual conditional we recorded. This is needed since we compute
216 ranges lazily. */
217 tree cond;
219 /* The basic block where this record was created. We use this to determine
220 when to remove records. */
221 basic_block bb;
224 /* A hash table holding value range records (VRP_ELEMENTs) for a given
225 SSA_NAME. We used to use a varray indexed by SSA_NAME_VERSION, but
226 that gets awful wasteful, particularly since the density objects
227 with useful information is very low. */
228 static htab_t vrp_data;
230 typedef struct vrp_element *vrp_element_p;
232 DEF_VEC_P(vrp_element_p);
233 DEF_VEC_ALLOC_P(vrp_element_p,heap);
235 /* An entry in the VRP_DATA hash table. We record the variable and a
236 varray of VRP_ELEMENT records associated with that variable. */
237 struct vrp_hash_elt
239 tree var;
240 VEC(vrp_element_p,heap) *records;
243 /* Array of variables which have their values constrained by operations
244 in this basic block. We use this during finalization to know
245 which variables need their VRP data updated. */
247 /* Stack of SSA_NAMEs which had their values constrained by operations
248 in this basic block. During finalization of this block we use this
249 list to determine which variables need their VRP data updated.
251 A NULL entry marks the end of the SSA_NAMEs associated with this block. */
252 static VEC(tree,heap) *vrp_variables_stack;
254 struct eq_expr_value
256 tree src;
257 tree dst;
260 /* Local functions. */
261 static void optimize_stmt (struct dom_walk_data *,
262 basic_block bb,
263 block_stmt_iterator);
264 static tree lookup_avail_expr (tree, bool);
265 static hashval_t vrp_hash (const void *);
266 static int vrp_eq (const void *, const void *);
267 static hashval_t avail_expr_hash (const void *);
268 static hashval_t real_avail_expr_hash (const void *);
269 static int avail_expr_eq (const void *, const void *);
270 static void htab_statistics (FILE *, htab_t);
271 static void record_cond (tree, tree);
272 static void record_const_or_copy (tree, tree);
273 static void record_equality (tree, tree);
274 static tree update_rhs_and_lookup_avail_expr (tree, tree, bool);
275 static tree simplify_rhs_and_lookup_avail_expr (tree, int);
276 static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int);
277 static tree simplify_switch_and_lookup_avail_expr (tree, int);
278 static tree find_equivalent_equality_comparison (tree);
279 static void record_range (tree, basic_block);
280 static bool extract_range_from_cond (tree, tree *, tree *, int *);
281 static void record_equivalences_from_phis (basic_block);
282 static void record_equivalences_from_incoming_edge (basic_block);
283 static bool eliminate_redundant_computations (tree, stmt_ann_t);
284 static void record_equivalences_from_stmt (tree, int, stmt_ann_t);
285 static void thread_across_edge (struct dom_walk_data *, edge);
286 static void dom_opt_finalize_block (struct dom_walk_data *, basic_block);
287 static void dom_opt_initialize_block (struct dom_walk_data *, basic_block);
288 static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block);
289 static void remove_local_expressions_from_table (void);
290 static void restore_vars_to_original_value (void);
291 static edge single_incoming_edge_ignoring_loop_edges (basic_block);
292 static void restore_nonzero_vars_to_original_value (void);
293 static inline bool unsafe_associative_fp_binop (tree);
296 /* Local version of fold that doesn't introduce cruft. */
298 static tree
299 local_fold (tree t)
301 t = fold (t);
303 /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that
304 may have been added by fold, and "useless" type conversions that might
305 now be apparent due to propagation. */
306 STRIP_USELESS_TYPE_CONVERSION (t);
308 return t;
311 /* Allocate an EDGE_INFO for edge E and attach it to E.
312 Return the new EDGE_INFO structure. */
314 static struct edge_info *
315 allocate_edge_info (edge e)
317 struct edge_info *edge_info;
319 edge_info = xcalloc (1, sizeof (struct edge_info));
321 e->aux = edge_info;
322 return edge_info;
325 /* Free all EDGE_INFO structures associated with edges in the CFG.
326 If a particular edge can be threaded, copy the redirection
327 target from the EDGE_INFO structure into the edge's AUX field
328 as required by code to update the CFG and SSA graph for
329 jump threading. */
331 static void
332 free_all_edge_infos (void)
334 basic_block bb;
335 edge_iterator ei;
336 edge e;
338 FOR_EACH_BB (bb)
340 FOR_EACH_EDGE (e, ei, bb->preds)
342 struct edge_info *edge_info = e->aux;
344 if (edge_info)
346 e->aux = edge_info->redirection_target;
347 if (edge_info->cond_equivalences)
348 free (edge_info->cond_equivalences);
349 free (edge_info);
355 /* Free an instance of vrp_hash_elt. */
357 static void
358 vrp_free (void *data)
360 struct vrp_hash_elt *elt = data;
361 struct VEC(vrp_element_p,heap) **vrp_elt = &elt->records;
363 VEC_free (vrp_element_p, heap, *vrp_elt);
364 free (elt);
367 /* Jump threading, redundancy elimination and const/copy propagation.
369 This pass may expose new symbols that need to be renamed into SSA. For
370 every new symbol exposed, its corresponding bit will be set in
371 VARS_TO_RENAME. */
373 static void
374 tree_ssa_dominator_optimize (void)
376 struct dom_walk_data walk_data;
377 unsigned int i;
378 struct loops loops_info;
380 memset (&opt_stats, 0, sizeof (opt_stats));
382 /* Create our hash tables. */
383 avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free);
384 vrp_data = htab_create (ceil_log2 (num_ssa_names), vrp_hash, vrp_eq,
385 vrp_free);
386 avail_exprs_stack = VEC_alloc (tree, heap, 20);
387 const_and_copies_stack = VEC_alloc (tree, heap, 20);
388 nonzero_vars_stack = VEC_alloc (tree, heap, 20);
389 vrp_variables_stack = VEC_alloc (tree, heap, 20);
390 stmts_to_rescan = VEC_alloc (tree, heap, 20);
391 nonzero_vars = BITMAP_ALLOC (NULL);
392 threaded_blocks = BITMAP_ALLOC (NULL);
393 need_eh_cleanup = BITMAP_ALLOC (NULL);
395 /* Setup callbacks for the generic dominator tree walker. */
396 walk_data.walk_stmts_backward = false;
397 walk_data.dom_direction = CDI_DOMINATORS;
398 walk_data.initialize_block_local_data = NULL;
399 walk_data.before_dom_children_before_stmts = dom_opt_initialize_block;
400 walk_data.before_dom_children_walk_stmts = optimize_stmt;
401 walk_data.before_dom_children_after_stmts = propagate_to_outgoing_edges;
402 walk_data.after_dom_children_before_stmts = NULL;
403 walk_data.after_dom_children_walk_stmts = NULL;
404 walk_data.after_dom_children_after_stmts = dom_opt_finalize_block;
405 /* Right now we only attach a dummy COND_EXPR to the global data pointer.
406 When we attach more stuff we'll need to fill this out with a real
407 structure. */
408 walk_data.global_data = NULL;
409 walk_data.block_local_data_size = 0;
410 walk_data.interesting_blocks = NULL;
412 /* Now initialize the dominator walker. */
413 init_walk_dominator_tree (&walk_data);
415 calculate_dominance_info (CDI_DOMINATORS);
417 /* We need to know which edges exit loops so that we can
418 aggressively thread through loop headers to an exit
419 edge. */
420 flow_loops_find (&loops_info);
421 mark_loop_exit_edges (&loops_info);
422 flow_loops_free (&loops_info);
424 /* Clean up the CFG so that any forwarder blocks created by loop
425 canonicalization are removed. */
426 cleanup_tree_cfg ();
427 calculate_dominance_info (CDI_DOMINATORS);
429 /* If we prove certain blocks are unreachable, then we want to
430 repeat the dominator optimization process as PHI nodes may
431 have turned into copies which allows better propagation of
432 values. So we repeat until we do not identify any new unreachable
433 blocks. */
436 /* Optimize the dominator tree. */
437 cfg_altered = false;
439 /* We need accurate information regarding back edges in the CFG
440 for jump threading. */
441 mark_dfs_back_edges ();
443 /* Recursively walk the dominator tree optimizing statements. */
444 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
447 block_stmt_iterator bsi;
448 basic_block bb;
449 FOR_EACH_BB (bb)
451 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
453 update_stmt_if_modified (bsi_stmt (bsi));
458 /* If we exposed any new variables, go ahead and put them into
459 SSA form now, before we handle jump threading. This simplifies
460 interactions between rewriting of _DECL nodes into SSA form
461 and rewriting SSA_NAME nodes into SSA form after block
462 duplication and CFG manipulation. */
463 update_ssa (TODO_update_ssa);
465 free_all_edge_infos ();
467 /* Thread jumps, creating duplicate blocks as needed. */
468 cfg_altered |= thread_through_all_blocks (threaded_blocks);
470 /* Removal of statements may make some EH edges dead. Purge
471 such edges from the CFG as needed. */
472 if (!bitmap_empty_p (need_eh_cleanup))
474 cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
475 bitmap_zero (need_eh_cleanup);
478 if (cfg_altered)
479 free_dominance_info (CDI_DOMINATORS);
481 cfg_altered = cleanup_tree_cfg ();
483 if (rediscover_loops_after_threading)
485 /* Rerun basic loop analysis to discover any newly
486 created loops and update the set of exit edges. */
487 rediscover_loops_after_threading = false;
488 flow_loops_find (&loops_info);
489 mark_loop_exit_edges (&loops_info);
490 flow_loops_free (&loops_info);
492 /* Remove any forwarder blocks inserted by loop
493 header canonicalization. */
494 cleanup_tree_cfg ();
497 calculate_dominance_info (CDI_DOMINATORS);
499 update_ssa (TODO_update_ssa);
501 /* Reinitialize the various tables. */
502 bitmap_clear (nonzero_vars);
503 bitmap_clear (threaded_blocks);
504 htab_empty (avail_exprs);
505 htab_empty (vrp_data);
507 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
509 This must be done before we iterate as we might have a
510 reference to an SSA_NAME which was removed by the call to
511 update_ssa.
513 Long term we will be able to let everything in SSA_NAME_VALUE
514 persist. However, for now, we know this is the safe thing to do. */
515 for (i = 0; i < num_ssa_names; i++)
517 tree name = ssa_name (i);
518 tree value;
520 if (!name)
521 continue;
523 value = SSA_NAME_VALUE (name);
524 if (value && !is_gimple_min_invariant (value))
525 SSA_NAME_VALUE (name) = NULL;
528 while (optimize > 1 && cfg_altered);
530 /* Debugging dumps. */
531 if (dump_file && (dump_flags & TDF_STATS))
532 dump_dominator_optimization_stats (dump_file);
534 /* We emptied the hash table earlier, now delete it completely. */
535 htab_delete (avail_exprs);
536 htab_delete (vrp_data);
538 /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
539 CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
540 of the do-while loop above. */
542 /* And finalize the dominator walker. */
543 fini_walk_dominator_tree (&walk_data);
545 /* Free nonzero_vars. */
546 BITMAP_FREE (nonzero_vars);
547 BITMAP_FREE (threaded_blocks);
548 BITMAP_FREE (need_eh_cleanup);
550 VEC_free (tree, heap, avail_exprs_stack);
551 VEC_free (tree, heap, const_and_copies_stack);
552 VEC_free (tree, heap, nonzero_vars_stack);
553 VEC_free (tree, heap, vrp_variables_stack);
554 VEC_free (tree, heap, stmts_to_rescan);
557 static bool
558 gate_dominator (void)
560 return flag_tree_dom != 0;
563 struct tree_opt_pass pass_dominator =
565 "dom", /* name */
566 gate_dominator, /* gate */
567 tree_ssa_dominator_optimize, /* execute */
568 NULL, /* sub */
569 NULL, /* next */
570 0, /* static_pass_number */
571 TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */
572 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
573 0, /* properties_provided */
574 0, /* properties_destroyed */
575 0, /* todo_flags_start */
576 TODO_dump_func
577 | TODO_update_ssa
578 | TODO_verify_ssa, /* todo_flags_finish */
579 0 /* letter */
583 /* We are exiting E->src, see if E->dest ends with a conditional
584 jump which has a known value when reached via E.
586 Special care is necessary if E is a back edge in the CFG as we
587 will have already recorded equivalences for E->dest into our
588 various tables, including the result of the conditional at
589 the end of E->dest. Threading opportunities are severely
590 limited in that case to avoid short-circuiting the loop
591 incorrectly.
593 Note it is quite common for the first block inside a loop to
594 end with a conditional which is either always true or always
595 false when reached via the loop backedge. Thus we do not want
596 to blindly disable threading across a loop backedge. */
598 static void
599 thread_across_edge (struct dom_walk_data *walk_data, edge e)
601 block_stmt_iterator bsi;
602 tree stmt = NULL;
603 tree phi;
605 /* If E->dest does not end with a conditional, then there is
606 nothing to do. */
607 bsi = bsi_last (e->dest);
608 if (bsi_end_p (bsi)
609 || ! bsi_stmt (bsi)
610 || (TREE_CODE (bsi_stmt (bsi)) != COND_EXPR
611 && TREE_CODE (bsi_stmt (bsi)) != GOTO_EXPR
612 && TREE_CODE (bsi_stmt (bsi)) != SWITCH_EXPR))
613 return;
615 /* The basic idea here is to use whatever knowledge we have
616 from our dominator walk to simplify statements in E->dest,
617 with the ultimate goal being to simplify the conditional
618 at the end of E->dest.
620 Note that we must undo any changes we make to the underlying
621 statements as the simplifications we are making are control
622 flow sensitive (ie, the simplifications are valid when we
623 traverse E, but may not be valid on other paths to E->dest. */
625 /* Each PHI creates a temporary equivalence, record them. Again
626 these are context sensitive equivalences and will be removed
627 by our caller. */
628 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
630 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
631 tree dst = PHI_RESULT (phi);
633 /* If the desired argument is not the same as this PHI's result
634 and it is set by a PHI in E->dest, then we can not thread
635 through E->dest. */
636 if (src != dst
637 && TREE_CODE (src) == SSA_NAME
638 && TREE_CODE (SSA_NAME_DEF_STMT (src)) == PHI_NODE
639 && bb_for_stmt (SSA_NAME_DEF_STMT (src)) == e->dest)
640 return;
642 record_const_or_copy (dst, src);
645 /* Try to simplify each statement in E->dest, ultimately leading to
646 a simplification of the COND_EXPR at the end of E->dest.
648 We might consider marking just those statements which ultimately
649 feed the COND_EXPR. It's not clear if the overhead of bookkeeping
650 would be recovered by trying to simplify fewer statements.
652 If we are able to simplify a statement into the form
653 SSA_NAME = (SSA_NAME | gimple invariant), then we can record
654 a context sensitive equivalency which may help us simplify
655 later statements in E->dest.
657 Failure to simplify into the form above merely means that the
658 statement provides no equivalences to help simplify later
659 statements. This does not prevent threading through E->dest. */
660 for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
662 tree cached_lhs;
664 stmt = bsi_stmt (bsi);
666 /* Ignore empty statements and labels. */
667 if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
668 continue;
670 /* Safely handle threading across loop backedges. This is
671 over conservative, but still allows us to capture the
672 majority of the cases where we can thread across a loop
673 backedge. */
674 if ((e->flags & EDGE_DFS_BACK) != 0
675 && TREE_CODE (stmt) != COND_EXPR
676 && TREE_CODE (stmt) != SWITCH_EXPR)
677 return;
679 /* If the statement has volatile operands, then we assume we
680 can not thread through this block. This is overly
681 conservative in some ways. */
682 if (TREE_CODE (stmt) == ASM_EXPR && ASM_VOLATILE_P (stmt))
683 return;
685 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
686 value, then do not try to simplify this statement as it will
687 not simplify in any way that is helpful for jump threading. */
688 if (TREE_CODE (stmt) != MODIFY_EXPR
689 || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
690 continue;
692 /* At this point we have a statement which assigns an RHS to an
693 SSA_VAR on the LHS. We want to try and simplify this statement
694 to expose more context sensitive equivalences which in turn may
695 allow us to simplify the condition at the end of the loop. */
696 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
697 cached_lhs = TREE_OPERAND (stmt, 1);
698 else
700 /* Copy the operands. */
701 tree *copy;
702 ssa_op_iter iter;
703 use_operand_p use_p;
704 unsigned int num, i = 0;
706 num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE));
707 copy = xcalloc (num, sizeof (tree));
709 /* Make a copy of the uses & vuses into USES_COPY, then cprop into
710 the operands. */
711 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
713 tree tmp = NULL;
714 tree use = USE_FROM_PTR (use_p);
716 copy[i++] = use;
717 if (TREE_CODE (use) == SSA_NAME)
718 tmp = SSA_NAME_VALUE (use);
719 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
720 SET_USE (use_p, tmp);
723 /* Try to fold/lookup the new expression. Inserting the
724 expression into the hash table is unlikely to help
725 simplify anything later, so just query the hashtable. */
726 cached_lhs = fold (TREE_OPERAND (stmt, 1));
727 if (TREE_CODE (cached_lhs) != SSA_NAME
728 && !is_gimple_min_invariant (cached_lhs))
729 cached_lhs = lookup_avail_expr (stmt, false);
732 /* Restore the statement's original uses/defs. */
733 i = 0;
734 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
735 SET_USE (use_p, copy[i++]);
737 free (copy);
740 /* Record the context sensitive equivalence if we were able
741 to simplify this statement. */
742 if (cached_lhs
743 && (TREE_CODE (cached_lhs) == SSA_NAME
744 || is_gimple_min_invariant (cached_lhs)))
745 record_const_or_copy (TREE_OPERAND (stmt, 0), cached_lhs);
748 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
749 will be taken. */
750 if (stmt
751 && (TREE_CODE (stmt) == COND_EXPR
752 || TREE_CODE (stmt) == GOTO_EXPR
753 || TREE_CODE (stmt) == SWITCH_EXPR))
755 tree cond, cached_lhs;
757 /* Now temporarily cprop the operands and try to find the resulting
758 expression in the hash tables. */
759 if (TREE_CODE (stmt) == COND_EXPR)
760 cond = COND_EXPR_COND (stmt);
761 else if (TREE_CODE (stmt) == GOTO_EXPR)
762 cond = GOTO_DESTINATION (stmt);
763 else
764 cond = SWITCH_COND (stmt);
766 if (COMPARISON_CLASS_P (cond))
768 tree dummy_cond, op0, op1;
769 enum tree_code cond_code;
771 op0 = TREE_OPERAND (cond, 0);
772 op1 = TREE_OPERAND (cond, 1);
773 cond_code = TREE_CODE (cond);
775 /* Get the current value of both operands. */
776 if (TREE_CODE (op0) == SSA_NAME)
778 tree tmp = SSA_NAME_VALUE (op0);
779 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
780 op0 = tmp;
783 if (TREE_CODE (op1) == SSA_NAME)
785 tree tmp = SSA_NAME_VALUE (op1);
786 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
787 op1 = tmp;
790 /* Stuff the operator and operands into our dummy conditional
791 expression, creating the dummy conditional if necessary. */
792 dummy_cond = walk_data->global_data;
793 if (! dummy_cond)
795 dummy_cond = build (cond_code, boolean_type_node, op0, op1);
796 dummy_cond = build (COND_EXPR, void_type_node,
797 dummy_cond, NULL, NULL);
798 walk_data->global_data = dummy_cond;
800 else
802 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
803 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
804 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
807 /* If the conditional folds to an invariant, then we are done,
808 otherwise look it up in the hash tables. */
809 cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
810 if (! is_gimple_min_invariant (cached_lhs))
812 cached_lhs = lookup_avail_expr (dummy_cond, false);
813 if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
814 cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
815 NULL,
816 false);
819 /* We can have conditionals which just test the state of a
820 variable rather than use a relational operator. These are
821 simpler to handle. */
822 else if (TREE_CODE (cond) == SSA_NAME)
824 cached_lhs = cond;
825 cached_lhs = SSA_NAME_VALUE (cached_lhs);
826 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
827 cached_lhs = NULL;
829 else
830 cached_lhs = lookup_avail_expr (stmt, false);
832 if (cached_lhs)
834 edge taken_edge = find_taken_edge (e->dest, cached_lhs);
835 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
837 if (dest == e->dest)
838 return;
840 /* If we have a known destination for the conditional, then
841 we can perform this optimization, which saves at least one
842 conditional jump each time it applies since we get to
843 bypass the conditional at our original destination. */
844 if (dest)
846 struct edge_info *edge_info;
848 update_bb_profile_for_threading (e->dest, EDGE_FREQUENCY (e),
849 e->count, taken_edge);
850 if (e->aux)
851 edge_info = e->aux;
852 else
853 edge_info = allocate_edge_info (e);
854 edge_info->redirection_target = taken_edge;
855 bitmap_set_bit (threaded_blocks, e->dest->index);
862 /* Initialize local stacks for this optimizer and record equivalences
863 upon entry to BB. Equivalences can come from the edge traversed to
864 reach BB or they may come from PHI nodes at the start of BB. */
866 static void
867 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
868 basic_block bb)
870 if (dump_file && (dump_flags & TDF_DETAILS))
871 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
873 /* Push a marker on the stacks of local information so that we know how
874 far to unwind when we finalize this block. */
875 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
876 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
877 VEC_safe_push (tree, heap, nonzero_vars_stack, NULL_TREE);
878 VEC_safe_push (tree, heap, vrp_variables_stack, NULL_TREE);
880 record_equivalences_from_incoming_edge (bb);
882 /* PHI nodes can create equivalences too. */
883 record_equivalences_from_phis (bb);
886 /* Given an expression EXPR (a relational expression or a statement),
887 initialize the hash table element pointed by by ELEMENT. */
889 static void
890 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
892 /* Hash table elements may be based on conditional expressions or statements.
894 For the former case, we have no annotation and we want to hash the
895 conditional expression. In the latter case we have an annotation and
896 we want to record the expression the statement evaluates. */
897 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
899 element->stmt = NULL;
900 element->rhs = expr;
902 else if (TREE_CODE (expr) == COND_EXPR)
904 element->stmt = expr;
905 element->rhs = COND_EXPR_COND (expr);
907 else if (TREE_CODE (expr) == SWITCH_EXPR)
909 element->stmt = expr;
910 element->rhs = SWITCH_COND (expr);
912 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
914 element->stmt = expr;
915 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
917 else if (TREE_CODE (expr) == GOTO_EXPR)
919 element->stmt = expr;
920 element->rhs = GOTO_DESTINATION (expr);
922 else
924 element->stmt = expr;
925 element->rhs = TREE_OPERAND (expr, 1);
928 element->lhs = lhs;
929 element->hash = avail_expr_hash (element);
932 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
933 LIMIT entries left in LOCALs. */
935 static void
936 remove_local_expressions_from_table (void)
938 /* Remove all the expressions made available in this block. */
939 while (VEC_length (tree, avail_exprs_stack) > 0)
941 struct expr_hash_elt element;
942 tree expr = VEC_pop (tree, avail_exprs_stack);
944 if (expr == NULL_TREE)
945 break;
947 initialize_hash_element (expr, NULL, &element);
948 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
952 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
953 state, stopping when there are LIMIT entries left in LOCALs. */
955 static void
956 restore_nonzero_vars_to_original_value (void)
958 while (VEC_length (tree, nonzero_vars_stack) > 0)
960 tree name = VEC_pop (tree, nonzero_vars_stack);
962 if (name == NULL)
963 break;
965 bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
969 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
970 CONST_AND_COPIES to its original state, stopping when we hit a
971 NULL marker. */
973 static void
974 restore_vars_to_original_value (void)
976 while (VEC_length (tree, const_and_copies_stack) > 0)
978 tree prev_value, dest;
980 dest = VEC_pop (tree, const_and_copies_stack);
982 if (dest == NULL)
983 break;
985 prev_value = VEC_pop (tree, const_and_copies_stack);
986 SSA_NAME_VALUE (dest) = prev_value;
990 /* We have finished processing the dominator children of BB, perform
991 any finalization actions in preparation for leaving this node in
992 the dominator tree. */
994 static void
995 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
997 tree last;
999 /* If we are at a leaf node in the dominator tree, see if we can thread
1000 the edge from BB through its successor.
1002 Do this before we remove entries from our equivalence tables. */
1003 if (single_succ_p (bb)
1004 && (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0
1005 && (get_immediate_dominator (CDI_DOMINATORS, single_succ (bb)) != bb
1006 || phi_nodes (single_succ (bb))))
1009 thread_across_edge (walk_data, single_succ_edge (bb));
1011 else if ((last = last_stmt (bb))
1012 && TREE_CODE (last) == COND_EXPR
1013 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
1014 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
1015 && EDGE_COUNT (bb->succs) == 2
1016 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
1017 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
1019 edge true_edge, false_edge;
1021 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
1023 /* If the THEN arm is the end of a dominator tree or has PHI nodes,
1024 then try to thread through its edge. */
1025 if (get_immediate_dominator (CDI_DOMINATORS, true_edge->dest) != bb
1026 || phi_nodes (true_edge->dest))
1028 struct edge_info *edge_info;
1029 unsigned int i;
1031 /* Push a marker onto the available expression stack so that we
1032 unwind any expressions related to the TRUE arm before processing
1033 the false arm below. */
1034 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
1035 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
1037 edge_info = true_edge->aux;
1039 /* If we have info associated with this edge, record it into
1040 our equivalency tables. */
1041 if (edge_info)
1043 tree *cond_equivalences = edge_info->cond_equivalences;
1044 tree lhs = edge_info->lhs;
1045 tree rhs = edge_info->rhs;
1047 /* If we have a simple NAME = VALUE equivalency record it. */
1048 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1049 record_const_or_copy (lhs, rhs);
1051 /* If we have 0 = COND or 1 = COND equivalences, record them
1052 into our expression hash tables. */
1053 if (cond_equivalences)
1054 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1056 tree expr = cond_equivalences[i];
1057 tree value = cond_equivalences[i + 1];
1059 record_cond (expr, value);
1063 /* Now thread the edge. */
1064 thread_across_edge (walk_data, true_edge);
1066 /* And restore the various tables to their state before
1067 we threaded this edge. */
1068 remove_local_expressions_from_table ();
1069 restore_vars_to_original_value ();
1072 /* Similarly for the ELSE arm. */
1073 if (get_immediate_dominator (CDI_DOMINATORS, false_edge->dest) != bb
1074 || phi_nodes (false_edge->dest))
1076 struct edge_info *edge_info;
1077 unsigned int i;
1079 edge_info = false_edge->aux;
1081 /* If we have info associated with this edge, record it into
1082 our equivalency tables. */
1083 if (edge_info)
1085 tree *cond_equivalences = edge_info->cond_equivalences;
1086 tree lhs = edge_info->lhs;
1087 tree rhs = edge_info->rhs;
1089 /* If we have a simple NAME = VALUE equivalency record it. */
1090 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1091 record_const_or_copy (lhs, rhs);
1093 /* If we have 0 = COND or 1 = COND equivalences, record them
1094 into our expression hash tables. */
1095 if (cond_equivalences)
1096 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1098 tree expr = cond_equivalences[i];
1099 tree value = cond_equivalences[i + 1];
1101 record_cond (expr, value);
1105 thread_across_edge (walk_data, false_edge);
1107 /* No need to remove local expressions from our tables
1108 or restore vars to their original value as that will
1109 be done immediately below. */
1113 remove_local_expressions_from_table ();
1114 restore_nonzero_vars_to_original_value ();
1115 restore_vars_to_original_value ();
1117 /* Remove VRP records associated with this basic block. They are no
1118 longer valid.
1120 To be efficient, we note which variables have had their values
1121 constrained in this block. So walk over each variable in the
1122 VRP_VARIABLEs array. */
1123 while (VEC_length (tree, vrp_variables_stack) > 0)
1125 tree var = VEC_pop (tree, vrp_variables_stack);
1126 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1127 void **slot;
1129 /* Each variable has a stack of value range records. We want to
1130 invalidate those associated with our basic block. So we walk
1131 the array backwards popping off records associated with our
1132 block. Once we hit a record not associated with our block
1133 we are done. */
1134 VEC(vrp_element_p,heap) **var_vrp_records;
1136 if (var == NULL)
1137 break;
1139 vrp_hash_elt.var = var;
1140 vrp_hash_elt.records = NULL;
1142 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1144 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1145 var_vrp_records = &vrp_hash_elt_p->records;
1147 while (VEC_length (vrp_element_p, *var_vrp_records) > 0)
1149 struct vrp_element *element
1150 = VEC_last (vrp_element_p, *var_vrp_records);
1152 if (element->bb != bb)
1153 break;
1155 VEC_pop (vrp_element_p, *var_vrp_records);
1159 /* If we queued any statements to rescan in this block, then
1160 go ahead and rescan them now. */
1161 while (VEC_length (tree, stmts_to_rescan) > 0)
1163 tree stmt = VEC_last (tree, stmts_to_rescan);
1164 basic_block stmt_bb = bb_for_stmt (stmt);
1166 if (stmt_bb != bb)
1167 break;
1169 VEC_pop (tree, stmts_to_rescan);
1170 mark_new_vars_to_rename (stmt);
1174 /* PHI nodes can create equivalences too.
1176 Ignoring any alternatives which are the same as the result, if
1177 all the alternatives are equal, then the PHI node creates an
1178 equivalence.
1180 Additionally, if all the PHI alternatives are known to have a nonzero
1181 value, then the result of this PHI is known to have a nonzero value,
1182 even if we do not know its exact value. */
1184 static void
1185 record_equivalences_from_phis (basic_block bb)
1187 tree phi;
1189 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1191 tree lhs = PHI_RESULT (phi);
1192 tree rhs = NULL;
1193 int i;
1195 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1197 tree t = PHI_ARG_DEF (phi, i);
1199 /* Ignore alternatives which are the same as our LHS. Since
1200 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1201 can simply compare pointers. */
1202 if (lhs == t)
1203 continue;
1205 /* If we have not processed an alternative yet, then set
1206 RHS to this alternative. */
1207 if (rhs == NULL)
1208 rhs = t;
1209 /* If we have processed an alternative (stored in RHS), then
1210 see if it is equal to this one. If it isn't, then stop
1211 the search. */
1212 else if (! operand_equal_for_phi_arg_p (rhs, t))
1213 break;
1216 /* If we had no interesting alternatives, then all the RHS alternatives
1217 must have been the same as LHS. */
1218 if (!rhs)
1219 rhs = lhs;
1221 /* If we managed to iterate through each PHI alternative without
1222 breaking out of the loop, then we have a PHI which may create
1223 a useful equivalence. We do not need to record unwind data for
1224 this, since this is a true assignment and not an equivalence
1225 inferred from a comparison. All uses of this ssa name are dominated
1226 by this assignment, so unwinding just costs time and space. */
1227 if (i == PHI_NUM_ARGS (phi)
1228 && may_propagate_copy (lhs, rhs))
1229 SSA_NAME_VALUE (lhs) = rhs;
1231 /* Now see if we know anything about the nonzero property for the
1232 result of this PHI. */
1233 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1235 if (!PHI_ARG_NONZERO (phi, i))
1236 break;
1239 if (i == PHI_NUM_ARGS (phi))
1240 bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
1244 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1245 return that edge. Otherwise return NULL. */
1246 static edge
1247 single_incoming_edge_ignoring_loop_edges (basic_block bb)
1249 edge retval = NULL;
1250 edge e;
1251 edge_iterator ei;
1253 FOR_EACH_EDGE (e, ei, bb->preds)
1255 /* A loop back edge can be identified by the destination of
1256 the edge dominating the source of the edge. */
1257 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
1258 continue;
1260 /* If we have already seen a non-loop edge, then we must have
1261 multiple incoming non-loop edges and thus we return NULL. */
1262 if (retval)
1263 return NULL;
1265 /* This is the first non-loop incoming edge we have found. Record
1266 it. */
1267 retval = e;
1270 return retval;
1273 /* Record any equivalences created by the incoming edge to BB. If BB
1274 has more than one incoming edge, then no equivalence is created. */
1276 static void
1277 record_equivalences_from_incoming_edge (basic_block bb)
1279 edge e;
1280 basic_block parent;
1281 struct edge_info *edge_info;
1283 /* If our parent block ended with a control statement, then we may be
1284 able to record some equivalences based on which outgoing edge from
1285 the parent was followed. */
1286 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
1288 e = single_incoming_edge_ignoring_loop_edges (bb);
1290 /* If we had a single incoming edge from our parent block, then enter
1291 any data associated with the edge into our tables. */
1292 if (e && e->src == parent)
1294 unsigned int i;
1296 edge_info = e->aux;
1298 if (edge_info)
1300 tree lhs = edge_info->lhs;
1301 tree rhs = edge_info->rhs;
1302 tree *cond_equivalences = edge_info->cond_equivalences;
1304 if (lhs)
1305 record_equality (lhs, rhs);
1307 if (cond_equivalences)
1309 bool recorded_range = false;
1310 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1312 tree expr = cond_equivalences[i];
1313 tree value = cond_equivalences[i + 1];
1315 record_cond (expr, value);
1317 /* For the first true equivalence, record range
1318 information. We only do this for the first
1319 true equivalence as it should dominate any
1320 later true equivalences. */
1321 if (! recorded_range
1322 && COMPARISON_CLASS_P (expr)
1323 && value == boolean_true_node
1324 && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
1326 record_range (expr, bb);
1327 recorded_range = true;
1335 /* Dump SSA statistics on FILE. */
1337 void
1338 dump_dominator_optimization_stats (FILE *file)
1340 long n_exprs;
1342 fprintf (file, "Total number of statements: %6ld\n\n",
1343 opt_stats.num_stmts);
1344 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
1345 opt_stats.num_exprs_considered);
1347 n_exprs = opt_stats.num_exprs_considered;
1348 if (n_exprs == 0)
1349 n_exprs = 1;
1351 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1352 opt_stats.num_re, PERCENT (opt_stats.num_re,
1353 n_exprs));
1354 fprintf (file, " Constants propagated: %6ld\n",
1355 opt_stats.num_const_prop);
1356 fprintf (file, " Copies propagated: %6ld\n",
1357 opt_stats.num_copy_prop);
1359 fprintf (file, "\nHash table statistics:\n");
1361 fprintf (file, " avail_exprs: ");
1362 htab_statistics (file, avail_exprs);
1366 /* Dump SSA statistics on stderr. */
1368 void
1369 debug_dominator_optimization_stats (void)
1371 dump_dominator_optimization_stats (stderr);
1375 /* Dump statistics for the hash table HTAB. */
1377 static void
1378 htab_statistics (FILE *file, htab_t htab)
1380 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
1381 (long) htab_size (htab),
1382 (long) htab_elements (htab),
1383 htab_collisions (htab));
1386 /* Record the fact that VAR has a nonzero value, though we may not know
1387 its exact value. Note that if VAR is already known to have a nonzero
1388 value, then we do nothing. */
1390 static void
1391 record_var_is_nonzero (tree var)
1393 int indx = SSA_NAME_VERSION (var);
1395 if (bitmap_bit_p (nonzero_vars, indx))
1396 return;
1398 /* Mark it in the global table. */
1399 bitmap_set_bit (nonzero_vars, indx);
1401 /* Record this SSA_NAME so that we can reset the global table
1402 when we leave this block. */
1403 VEC_safe_push (tree, heap, nonzero_vars_stack, var);
1406 /* Enter a statement into the true/false expression hash table indicating
1407 that the condition COND has the value VALUE. */
1409 static void
1410 record_cond (tree cond, tree value)
1412 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
1413 void **slot;
1415 initialize_hash_element (cond, value, element);
1417 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
1418 element->hash, INSERT);
1419 if (*slot == NULL)
1421 *slot = (void *) element;
1422 VEC_safe_push (tree, heap, avail_exprs_stack, cond);
1424 else
1425 free (element);
1428 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1429 the new conditional into *p, then store a boolean_true_node
1430 into *(p + 1). */
1432 static void
1433 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
1435 *p = build2 (new_code, boolean_type_node, op0, op1);
1436 p++;
1437 *p = boolean_true_node;
1440 /* Record that COND is true and INVERTED is false into the edge information
1441 structure. Also record that any conditions dominated by COND are true
1442 as well.
1444 For example, if a < b is true, then a <= b must also be true. */
1446 static void
1447 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
1449 tree op0, op1;
1451 if (!COMPARISON_CLASS_P (cond))
1452 return;
1454 op0 = TREE_OPERAND (cond, 0);
1455 op1 = TREE_OPERAND (cond, 1);
1457 switch (TREE_CODE (cond))
1459 case LT_EXPR:
1460 case GT_EXPR:
1461 edge_info->max_cond_equivalences = 12;
1462 edge_info->cond_equivalences = xmalloc (12 * sizeof (tree));
1463 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
1464 ? LE_EXPR : GE_EXPR),
1465 op0, op1, &edge_info->cond_equivalences[4]);
1466 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1467 &edge_info->cond_equivalences[6]);
1468 build_and_record_new_cond (NE_EXPR, op0, op1,
1469 &edge_info->cond_equivalences[8]);
1470 build_and_record_new_cond (LTGT_EXPR, op0, op1,
1471 &edge_info->cond_equivalences[10]);
1472 break;
1474 case GE_EXPR:
1475 case LE_EXPR:
1476 edge_info->max_cond_equivalences = 6;
1477 edge_info->cond_equivalences = xmalloc (6 * sizeof (tree));
1478 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1479 &edge_info->cond_equivalences[4]);
1480 break;
1482 case EQ_EXPR:
1483 edge_info->max_cond_equivalences = 10;
1484 edge_info->cond_equivalences = xmalloc (10 * sizeof (tree));
1485 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1486 &edge_info->cond_equivalences[4]);
1487 build_and_record_new_cond (LE_EXPR, op0, op1,
1488 &edge_info->cond_equivalences[6]);
1489 build_and_record_new_cond (GE_EXPR, op0, op1,
1490 &edge_info->cond_equivalences[8]);
1491 break;
1493 case UNORDERED_EXPR:
1494 edge_info->max_cond_equivalences = 16;
1495 edge_info->cond_equivalences = xmalloc (16 * sizeof (tree));
1496 build_and_record_new_cond (NE_EXPR, op0, op1,
1497 &edge_info->cond_equivalences[4]);
1498 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1499 &edge_info->cond_equivalences[6]);
1500 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1501 &edge_info->cond_equivalences[8]);
1502 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
1503 &edge_info->cond_equivalences[10]);
1504 build_and_record_new_cond (UNLT_EXPR, op0, op1,
1505 &edge_info->cond_equivalences[12]);
1506 build_and_record_new_cond (UNGT_EXPR, op0, op1,
1507 &edge_info->cond_equivalences[14]);
1508 break;
1510 case UNLT_EXPR:
1511 case UNGT_EXPR:
1512 edge_info->max_cond_equivalences = 8;
1513 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1514 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1515 ? UNLE_EXPR : UNGE_EXPR),
1516 op0, op1, &edge_info->cond_equivalences[4]);
1517 build_and_record_new_cond (NE_EXPR, op0, op1,
1518 &edge_info->cond_equivalences[6]);
1519 break;
1521 case UNEQ_EXPR:
1522 edge_info->max_cond_equivalences = 8;
1523 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1524 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1525 &edge_info->cond_equivalences[4]);
1526 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1527 &edge_info->cond_equivalences[6]);
1528 break;
1530 case LTGT_EXPR:
1531 edge_info->max_cond_equivalences = 8;
1532 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1533 build_and_record_new_cond (NE_EXPR, op0, op1,
1534 &edge_info->cond_equivalences[4]);
1535 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1536 &edge_info->cond_equivalences[6]);
1537 break;
1539 default:
1540 edge_info->max_cond_equivalences = 4;
1541 edge_info->cond_equivalences = xmalloc (4 * sizeof (tree));
1542 break;
1545 /* Now store the original true and false conditions into the first
1546 two slots. */
1547 edge_info->cond_equivalences[0] = cond;
1548 edge_info->cond_equivalences[1] = boolean_true_node;
1549 edge_info->cond_equivalences[2] = inverted;
1550 edge_info->cond_equivalences[3] = boolean_false_node;
1553 /* A helper function for record_const_or_copy and record_equality.
1554 Do the work of recording the value and undo info. */
1556 static void
1557 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1559 SSA_NAME_VALUE (x) = y;
1561 VEC_reserve (tree, heap, const_and_copies_stack, 2);
1562 VEC_quick_push (tree, const_and_copies_stack, prev_x);
1563 VEC_quick_push (tree, const_and_copies_stack, x);
1567 /* Return the loop depth of the basic block of the defining statement of X.
1568 This number should not be treated as absolutely correct because the loop
1569 information may not be completely up-to-date when dom runs. However, it
1570 will be relatively correct, and as more passes are taught to keep loop info
1571 up to date, the result will become more and more accurate. */
1574 loop_depth_of_name (tree x)
1576 tree defstmt;
1577 basic_block defbb;
1579 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1580 if (TREE_CODE (x) != SSA_NAME)
1581 return 0;
1583 /* Otherwise return the loop depth of the defining statement's bb.
1584 Note that there may not actually be a bb for this statement, if the
1585 ssa_name is live on entry. */
1586 defstmt = SSA_NAME_DEF_STMT (x);
1587 defbb = bb_for_stmt (defstmt);
1588 if (!defbb)
1589 return 0;
1591 return defbb->loop_depth;
1595 /* Record that X is equal to Y in const_and_copies. Record undo
1596 information in the block-local vector. */
1598 static void
1599 record_const_or_copy (tree x, tree y)
1601 tree prev_x = SSA_NAME_VALUE (x);
1603 if (TREE_CODE (y) == SSA_NAME)
1605 tree tmp = SSA_NAME_VALUE (y);
1606 if (tmp)
1607 y = tmp;
1610 record_const_or_copy_1 (x, y, prev_x);
1613 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1614 This constrains the cases in which we may treat this as assignment. */
1616 static void
1617 record_equality (tree x, tree y)
1619 tree prev_x = NULL, prev_y = NULL;
1621 if (TREE_CODE (x) == SSA_NAME)
1622 prev_x = SSA_NAME_VALUE (x);
1623 if (TREE_CODE (y) == SSA_NAME)
1624 prev_y = SSA_NAME_VALUE (y);
1626 /* If one of the previous values is invariant, or invariant in more loops
1627 (by depth), then use that.
1628 Otherwise it doesn't matter which value we choose, just so
1629 long as we canonicalize on one value. */
1630 if (TREE_INVARIANT (y))
1632 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1633 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1634 else if (prev_x && TREE_INVARIANT (prev_x))
1635 x = y, y = prev_x, prev_x = prev_y;
1636 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1637 y = prev_y;
1639 /* After the swapping, we must have one SSA_NAME. */
1640 if (TREE_CODE (x) != SSA_NAME)
1641 return;
1643 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1644 variable compared against zero. If we're honoring signed zeros,
1645 then we cannot record this value unless we know that the value is
1646 nonzero. */
1647 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1648 && (TREE_CODE (y) != REAL_CST
1649 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1650 return;
1652 record_const_or_copy_1 (x, y, prev_x);
1655 /* Return true, if it is ok to do folding of an associative expression.
1656 EXP is the tree for the associative expression. */
1658 static inline bool
1659 unsafe_associative_fp_binop (tree exp)
1661 enum tree_code code = TREE_CODE (exp);
1662 return !(!flag_unsafe_math_optimizations
1663 && (code == MULT_EXPR || code == PLUS_EXPR
1664 || code == MINUS_EXPR)
1665 && FLOAT_TYPE_P (TREE_TYPE (exp)));
1668 /* Returns true when STMT is a simple iv increment. It detects the
1669 following situation:
1671 i_1 = phi (..., i_2)
1672 i_2 = i_1 +/- ... */
1674 static bool
1675 simple_iv_increment_p (tree stmt)
1677 tree lhs, rhs, preinc, phi;
1678 unsigned i;
1680 if (TREE_CODE (stmt) != MODIFY_EXPR)
1681 return false;
1683 lhs = TREE_OPERAND (stmt, 0);
1684 if (TREE_CODE (lhs) != SSA_NAME)
1685 return false;
1687 rhs = TREE_OPERAND (stmt, 1);
1689 if (TREE_CODE (rhs) != PLUS_EXPR
1690 && TREE_CODE (rhs) != MINUS_EXPR)
1691 return false;
1693 preinc = TREE_OPERAND (rhs, 0);
1694 if (TREE_CODE (preinc) != SSA_NAME)
1695 return false;
1697 phi = SSA_NAME_DEF_STMT (preinc);
1698 if (TREE_CODE (phi) != PHI_NODE)
1699 return false;
1701 for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++)
1702 if (PHI_ARG_DEF (phi, i) == lhs)
1703 return true;
1705 return false;
1708 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1709 hash tables. Try to simplify the RHS using whatever equivalences
1710 we may have recorded.
1712 If we are able to simplify the RHS, then lookup the simplified form in
1713 the hash table and return the result. Otherwise return NULL. */
1715 static tree
1716 simplify_rhs_and_lookup_avail_expr (tree stmt, int insert)
1718 tree rhs = TREE_OPERAND (stmt, 1);
1719 enum tree_code rhs_code = TREE_CODE (rhs);
1720 tree result = NULL;
1722 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1723 In which case we can change this statement to be lhs = y.
1724 Which can then be copy propagated.
1726 Similarly for negation. */
1727 if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR)
1728 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME)
1730 /* Get the definition statement for our RHS. */
1731 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1733 /* See if the RHS_DEF_STMT has the same form as our statement. */
1734 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR
1735 && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code)
1737 tree rhs_def_operand;
1739 rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0);
1741 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1742 if (TREE_CODE (rhs_def_operand) == SSA_NAME
1743 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand))
1744 result = update_rhs_and_lookup_avail_expr (stmt,
1745 rhs_def_operand,
1746 insert);
1750 /* If we have z = (x OP C1), see if we earlier had x = y OP C2.
1751 If OP is associative, create and fold (y OP C2) OP C1 which
1752 should result in (y OP C3), use that as the RHS for the
1753 assignment. Add minus to this, as we handle it specially below. */
1754 if ((associative_tree_code (rhs_code) || rhs_code == MINUS_EXPR)
1755 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME
1756 && is_gimple_min_invariant (TREE_OPERAND (rhs, 1)))
1758 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1760 /* If the statement defines an induction variable, do not propagate
1761 its value, so that we do not create overlapping life ranges. */
1762 if (simple_iv_increment_p (rhs_def_stmt))
1763 goto dont_fold_assoc;
1765 /* See if the RHS_DEF_STMT has the same form as our statement. */
1766 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR)
1768 tree rhs_def_rhs = TREE_OPERAND (rhs_def_stmt, 1);
1769 enum tree_code rhs_def_code = TREE_CODE (rhs_def_rhs);
1771 if ((rhs_code == rhs_def_code && unsafe_associative_fp_binop (rhs))
1772 || (rhs_code == PLUS_EXPR && rhs_def_code == MINUS_EXPR)
1773 || (rhs_code == MINUS_EXPR && rhs_def_code == PLUS_EXPR))
1775 tree def_stmt_op0 = TREE_OPERAND (rhs_def_rhs, 0);
1776 tree def_stmt_op1 = TREE_OPERAND (rhs_def_rhs, 1);
1778 if (TREE_CODE (def_stmt_op0) == SSA_NAME
1779 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0)
1780 && is_gimple_min_invariant (def_stmt_op1))
1782 tree outer_const = TREE_OPERAND (rhs, 1);
1783 tree type = TREE_TYPE (TREE_OPERAND (stmt, 0));
1784 tree t;
1786 /* If we care about correct floating point results, then
1787 don't fold x + c1 - c2. Note that we need to take both
1788 the codes and the signs to figure this out. */
1789 if (FLOAT_TYPE_P (type)
1790 && !flag_unsafe_math_optimizations
1791 && (rhs_def_code == PLUS_EXPR
1792 || rhs_def_code == MINUS_EXPR))
1794 bool neg = false;
1796 neg ^= (rhs_code == MINUS_EXPR);
1797 neg ^= (rhs_def_code == MINUS_EXPR);
1798 neg ^= real_isneg (TREE_REAL_CST_PTR (outer_const));
1799 neg ^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1));
1801 if (neg)
1802 goto dont_fold_assoc;
1805 /* Ho hum. So fold will only operate on the outermost
1806 thingy that we give it, so we have to build the new
1807 expression in two pieces. This requires that we handle
1808 combinations of plus and minus. */
1809 if (rhs_def_code != rhs_code)
1811 if (rhs_def_code == MINUS_EXPR)
1812 t = build (MINUS_EXPR, type, outer_const, def_stmt_op1);
1813 else
1814 t = build (MINUS_EXPR, type, def_stmt_op1, outer_const);
1815 rhs_code = PLUS_EXPR;
1817 else if (rhs_def_code == MINUS_EXPR)
1818 t = build (PLUS_EXPR, type, def_stmt_op1, outer_const);
1819 else
1820 t = build (rhs_def_code, type, def_stmt_op1, outer_const);
1821 t = local_fold (t);
1822 t = build (rhs_code, type, def_stmt_op0, t);
1823 t = local_fold (t);
1825 /* If the result is a suitable looking gimple expression,
1826 then use it instead of the original for STMT. */
1827 if (TREE_CODE (t) == SSA_NAME
1828 || (UNARY_CLASS_P (t)
1829 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME)
1830 || ((BINARY_CLASS_P (t) || COMPARISON_CLASS_P (t))
1831 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
1832 && is_gimple_val (TREE_OPERAND (t, 1))))
1833 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1837 dont_fold_assoc:;
1840 /* Optimize *"foo" into 'f'. This is done here rather than
1841 in fold to avoid problems with stuff like &*"foo". */
1842 if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF)
1844 tree t = fold_read_from_constant_string (rhs);
1846 if (t)
1847 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1850 return result;
1853 /* COND is a condition of the form:
1855 x == const or x != const
1857 Look back to x's defining statement and see if x is defined as
1859 x = (type) y;
1861 If const is unchanged if we convert it to type, then we can build
1862 the equivalent expression:
1865 y == const or y != const
1867 Which may allow further optimizations.
1869 Return the equivalent comparison or NULL if no such equivalent comparison
1870 was found. */
1872 static tree
1873 find_equivalent_equality_comparison (tree cond)
1875 tree op0 = TREE_OPERAND (cond, 0);
1876 tree op1 = TREE_OPERAND (cond, 1);
1877 tree def_stmt = SSA_NAME_DEF_STMT (op0);
1879 /* OP0 might have been a parameter, so first make sure it
1880 was defined by a MODIFY_EXPR. */
1881 if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
1883 tree def_rhs = TREE_OPERAND (def_stmt, 1);
1885 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1886 if ((TREE_CODE (def_rhs) == NOP_EXPR
1887 || TREE_CODE (def_rhs) == CONVERT_EXPR)
1888 && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
1890 tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
1891 tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
1892 tree new;
1894 if (TYPE_PRECISION (def_rhs_inner_type)
1895 > TYPE_PRECISION (TREE_TYPE (def_rhs)))
1896 return NULL;
1898 /* What we want to prove is that if we convert OP1 to
1899 the type of the object inside the NOP_EXPR that the
1900 result is still equivalent to SRC.
1902 If that is true, the build and return new equivalent
1903 condition which uses the source of the typecast and the
1904 new constant (which has only changed its type). */
1905 new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
1906 new = local_fold (new);
1907 if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
1908 return build (TREE_CODE (cond), TREE_TYPE (cond),
1909 def_rhs_inner, new);
1912 return NULL;
1915 /* STMT is a COND_EXPR for which we could not trivially determine its
1916 result. This routine attempts to find equivalent forms of the
1917 condition which we may be able to optimize better. It also
1918 uses simple value range propagation to optimize conditionals. */
1920 static tree
1921 simplify_cond_and_lookup_avail_expr (tree stmt,
1922 stmt_ann_t ann,
1923 int insert)
1925 tree cond = COND_EXPR_COND (stmt);
1927 if (COMPARISON_CLASS_P (cond))
1929 tree op0 = TREE_OPERAND (cond, 0);
1930 tree op1 = TREE_OPERAND (cond, 1);
1932 if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
1934 int limit;
1935 tree low, high, cond_low, cond_high;
1936 int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
1937 VEC(vrp_element_p,heap) **vrp_records;
1938 struct vrp_element *element;
1939 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1940 void **slot;
1942 /* First see if we have test of an SSA_NAME against a constant
1943 where the SSA_NAME is defined by an earlier typecast which
1944 is irrelevant when performing tests against the given
1945 constant. */
1946 if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
1948 tree new_cond = find_equivalent_equality_comparison (cond);
1950 if (new_cond)
1952 /* Update the statement to use the new equivalent
1953 condition. */
1954 COND_EXPR_COND (stmt) = new_cond;
1956 /* If this is not a real stmt, ann will be NULL and we
1957 avoid processing the operands. */
1958 if (ann)
1959 mark_stmt_modified (stmt);
1961 /* Lookup the condition and return its known value if it
1962 exists. */
1963 new_cond = lookup_avail_expr (stmt, insert);
1964 if (new_cond)
1965 return new_cond;
1967 /* The operands have changed, so update op0 and op1. */
1968 op0 = TREE_OPERAND (cond, 0);
1969 op1 = TREE_OPERAND (cond, 1);
1973 /* Consult the value range records for this variable (if they exist)
1974 to see if we can eliminate or simplify this conditional.
1976 Note two tests are necessary to determine no records exist.
1977 First we have to see if the virtual array exists, if it
1978 exists, then we have to check its active size.
1980 Also note the vast majority of conditionals are not testing
1981 a variable which has had its range constrained by an earlier
1982 conditional. So this filter avoids a lot of unnecessary work. */
1983 vrp_hash_elt.var = op0;
1984 vrp_hash_elt.records = NULL;
1985 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1986 if (slot == NULL)
1987 return NULL;
1989 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1990 vrp_records = &vrp_hash_elt_p->records;
1992 limit = VEC_length (vrp_element_p, *vrp_records);
1994 /* If we have no value range records for this variable, or we are
1995 unable to extract a range for this condition, then there is
1996 nothing to do. */
1997 if (limit == 0
1998 || ! extract_range_from_cond (cond, &cond_high,
1999 &cond_low, &cond_inverted))
2000 return NULL;
2002 /* We really want to avoid unnecessary computations of range
2003 info. So all ranges are computed lazily; this avoids a
2004 lot of unnecessary work. i.e., we record the conditional,
2005 but do not process how it constrains the variable's
2006 potential values until we know that processing the condition
2007 could be helpful.
2009 However, we do not want to have to walk a potentially long
2010 list of ranges, nor do we want to compute a variable's
2011 range more than once for a given path.
2013 Luckily, each time we encounter a conditional that can not
2014 be otherwise optimized we will end up here and we will
2015 compute the necessary range information for the variable
2016 used in this condition.
2018 Thus you can conclude that there will never be more than one
2019 conditional associated with a variable which has not been
2020 processed. So we never need to merge more than one new
2021 conditional into the current range.
2023 These properties also help us avoid unnecessary work. */
2024 element = VEC_last (vrp_element_p, *vrp_records);
2026 if (element->high && element->low)
2028 /* The last element has been processed, so there is no range
2029 merging to do, we can simply use the high/low values
2030 recorded in the last element. */
2031 low = element->low;
2032 high = element->high;
2034 else
2036 tree tmp_high, tmp_low;
2037 int dummy;
2039 /* The last element has not been processed. Process it now.
2040 record_range should ensure for cond inverted is not set.
2041 This call can only fail if cond is x < min or x > max,
2042 which fold should have optimized into false.
2043 If that doesn't happen, just pretend all values are
2044 in the range. */
2045 if (! extract_range_from_cond (element->cond, &tmp_high,
2046 &tmp_low, &dummy))
2047 gcc_unreachable ();
2048 else
2049 gcc_assert (dummy == 0);
2051 /* If this is the only element, then no merging is necessary,
2052 the high/low values from extract_range_from_cond are all
2053 we need. */
2054 if (limit == 1)
2056 low = tmp_low;
2057 high = tmp_high;
2059 else
2061 /* Get the high/low value from the previous element. */
2062 struct vrp_element *prev
2063 = VEC_index (vrp_element_p, *vrp_records, limit - 2);
2064 low = prev->low;
2065 high = prev->high;
2067 /* Merge in this element's range with the range from the
2068 previous element.
2070 The low value for the merged range is the maximum of
2071 the previous low value and the low value of this record.
2073 Similarly the high value for the merged range is the
2074 minimum of the previous high value and the high value of
2075 this record. */
2076 low = (low && tree_int_cst_compare (low, tmp_low) == 1
2077 ? low : tmp_low);
2078 high = (high && tree_int_cst_compare (high, tmp_high) == -1
2079 ? high : tmp_high);
2082 /* And record the computed range. */
2083 element->low = low;
2084 element->high = high;
2088 /* After we have constrained this variable's potential values,
2089 we try to determine the result of the given conditional.
2091 To simplify later tests, first determine if the current
2092 low value is the same low value as the conditional.
2093 Similarly for the current high value and the high value
2094 for the conditional. */
2095 lowequal = tree_int_cst_equal (low, cond_low);
2096 highequal = tree_int_cst_equal (high, cond_high);
2098 if (lowequal && highequal)
2099 return (cond_inverted ? boolean_false_node : boolean_true_node);
2101 /* To simplify the overlap/subset tests below we may want
2102 to swap the two ranges so that the larger of the two
2103 ranges occurs "first". */
2104 swapped = 0;
2105 if (tree_int_cst_compare (low, cond_low) == 1
2106 || (lowequal
2107 && tree_int_cst_compare (cond_high, high) == 1))
2109 tree temp;
2111 swapped = 1;
2112 temp = low;
2113 low = cond_low;
2114 cond_low = temp;
2115 temp = high;
2116 high = cond_high;
2117 cond_high = temp;
2120 /* Now determine if there is no overlap in the ranges
2121 or if the second range is a subset of the first range. */
2122 no_overlap = tree_int_cst_lt (high, cond_low);
2123 subset = tree_int_cst_compare (cond_high, high) != 1;
2125 /* If there was no overlap in the ranges, then this conditional
2126 always has a false value (unless we had to invert this
2127 conditional, in which case it always has a true value). */
2128 if (no_overlap)
2129 return (cond_inverted ? boolean_true_node : boolean_false_node);
2131 /* If the current range is a subset of the condition's range,
2132 then this conditional always has a true value (unless we
2133 had to invert this conditional, in which case it always
2134 has a true value). */
2135 if (subset && swapped)
2136 return (cond_inverted ? boolean_false_node : boolean_true_node);
2138 /* We were unable to determine the result of the conditional.
2139 However, we may be able to simplify the conditional. First
2140 merge the ranges in the same manner as range merging above. */
2141 low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
2142 high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
2144 /* If the range has converged to a single point, then turn this
2145 into an equality comparison. */
2146 if (TREE_CODE (cond) != EQ_EXPR
2147 && TREE_CODE (cond) != NE_EXPR
2148 && tree_int_cst_equal (low, high))
2150 TREE_SET_CODE (cond, EQ_EXPR);
2151 TREE_OPERAND (cond, 1) = high;
2155 return 0;
2158 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2159 result. This routine attempts to find equivalent forms of the
2160 condition which we may be able to optimize better. */
2162 static tree
2163 simplify_switch_and_lookup_avail_expr (tree stmt, int insert)
2165 tree cond = SWITCH_COND (stmt);
2166 tree def, to, ti;
2168 /* The optimization that we really care about is removing unnecessary
2169 casts. That will let us do much better in propagating the inferred
2170 constant at the switch target. */
2171 if (TREE_CODE (cond) == SSA_NAME)
2173 def = SSA_NAME_DEF_STMT (cond);
2174 if (TREE_CODE (def) == MODIFY_EXPR)
2176 def = TREE_OPERAND (def, 1);
2177 if (TREE_CODE (def) == NOP_EXPR)
2179 int need_precision;
2180 bool fail;
2182 def = TREE_OPERAND (def, 0);
2184 #ifdef ENABLE_CHECKING
2185 /* ??? Why was Jeff testing this? We are gimple... */
2186 gcc_assert (is_gimple_val (def));
2187 #endif
2189 to = TREE_TYPE (cond);
2190 ti = TREE_TYPE (def);
2192 /* If we have an extension that preserves value, then we
2193 can copy the source value into the switch. */
2195 need_precision = TYPE_PRECISION (ti);
2196 fail = false;
2197 if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti))
2198 fail = true;
2199 else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti))
2200 need_precision += 1;
2201 if (TYPE_PRECISION (to) < need_precision)
2202 fail = true;
2204 if (!fail)
2206 SWITCH_COND (stmt) = def;
2207 mark_stmt_modified (stmt);
2209 return lookup_avail_expr (stmt, insert);
2215 return 0;
2219 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2220 known value for that SSA_NAME (or NULL if no value is known).
2222 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2223 even if we don't know their precise value.
2225 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2226 nodes of the successors of BB. */
2228 static void
2229 cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
2231 edge e;
2232 edge_iterator ei;
2234 FOR_EACH_EDGE (e, ei, bb->succs)
2236 tree phi;
2237 int indx;
2239 /* If this is an abnormal edge, then we do not want to copy propagate
2240 into the PHI alternative associated with this edge. */
2241 if (e->flags & EDGE_ABNORMAL)
2242 continue;
2244 phi = phi_nodes (e->dest);
2245 if (! phi)
2246 continue;
2248 indx = e->dest_idx;
2249 for ( ; phi; phi = PHI_CHAIN (phi))
2251 tree new;
2252 use_operand_p orig_p;
2253 tree orig;
2255 /* The alternative may be associated with a constant, so verify
2256 it is an SSA_NAME before doing anything with it. */
2257 orig_p = PHI_ARG_DEF_PTR (phi, indx);
2258 orig = USE_FROM_PTR (orig_p);
2259 if (TREE_CODE (orig) != SSA_NAME)
2260 continue;
2262 /* If the alternative is known to have a nonzero value, record
2263 that fact in the PHI node itself for future use. */
2264 if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
2265 PHI_ARG_NONZERO (phi, indx) = true;
2267 /* If we have *ORIG_P in our constant/copy table, then replace
2268 ORIG_P with its value in our constant/copy table. */
2269 new = SSA_NAME_VALUE (orig);
2270 if (new
2271 && new != orig
2272 && (TREE_CODE (new) == SSA_NAME
2273 || is_gimple_min_invariant (new))
2274 && may_propagate_copy (orig, new))
2275 propagate_value (orig_p, new);
2280 /* We have finished optimizing BB, record any information implied by
2281 taking a specific outgoing edge from BB. */
2283 static void
2284 record_edge_info (basic_block bb)
2286 block_stmt_iterator bsi = bsi_last (bb);
2287 struct edge_info *edge_info;
2289 if (! bsi_end_p (bsi))
2291 tree stmt = bsi_stmt (bsi);
2293 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
2295 tree cond = SWITCH_COND (stmt);
2297 if (TREE_CODE (cond) == SSA_NAME)
2299 tree labels = SWITCH_LABELS (stmt);
2300 int i, n_labels = TREE_VEC_LENGTH (labels);
2301 tree *info = xcalloc (last_basic_block, sizeof (tree));
2302 edge e;
2303 edge_iterator ei;
2305 for (i = 0; i < n_labels; i++)
2307 tree label = TREE_VEC_ELT (labels, i);
2308 basic_block target_bb = label_to_block (CASE_LABEL (label));
2310 if (CASE_HIGH (label)
2311 || !CASE_LOW (label)
2312 || info[target_bb->index])
2313 info[target_bb->index] = error_mark_node;
2314 else
2315 info[target_bb->index] = label;
2318 FOR_EACH_EDGE (e, ei, bb->succs)
2320 basic_block target_bb = e->dest;
2321 tree node = info[target_bb->index];
2323 if (node != NULL && node != error_mark_node)
2325 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
2326 edge_info = allocate_edge_info (e);
2327 edge_info->lhs = cond;
2328 edge_info->rhs = x;
2331 free (info);
2335 /* A COND_EXPR may create equivalences too. */
2336 if (stmt && TREE_CODE (stmt) == COND_EXPR)
2338 tree cond = COND_EXPR_COND (stmt);
2339 edge true_edge;
2340 edge false_edge;
2342 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
2344 /* If the conditional is a single variable 'X', record 'X = 1'
2345 for the true edge and 'X = 0' on the false edge. */
2346 if (SSA_VAR_P (cond))
2348 struct edge_info *edge_info;
2350 edge_info = allocate_edge_info (true_edge);
2351 edge_info->lhs = cond;
2352 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
2354 edge_info = allocate_edge_info (false_edge);
2355 edge_info->lhs = cond;
2356 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
2358 /* Equality tests may create one or two equivalences. */
2359 else if (COMPARISON_CLASS_P (cond))
2361 tree op0 = TREE_OPERAND (cond, 0);
2362 tree op1 = TREE_OPERAND (cond, 1);
2364 /* Special case comparing booleans against a constant as we
2365 know the value of OP0 on both arms of the branch. i.e., we
2366 can record an equivalence for OP0 rather than COND. */
2367 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2368 && TREE_CODE (op0) == SSA_NAME
2369 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
2370 && is_gimple_min_invariant (op1))
2372 if (TREE_CODE (cond) == EQ_EXPR)
2374 edge_info = allocate_edge_info (true_edge);
2375 edge_info->lhs = op0;
2376 edge_info->rhs = (integer_zerop (op1)
2377 ? boolean_false_node
2378 : boolean_true_node);
2380 edge_info = allocate_edge_info (false_edge);
2381 edge_info->lhs = op0;
2382 edge_info->rhs = (integer_zerop (op1)
2383 ? boolean_true_node
2384 : boolean_false_node);
2386 else
2388 edge_info = allocate_edge_info (true_edge);
2389 edge_info->lhs = op0;
2390 edge_info->rhs = (integer_zerop (op1)
2391 ? boolean_true_node
2392 : boolean_false_node);
2394 edge_info = allocate_edge_info (false_edge);
2395 edge_info->lhs = op0;
2396 edge_info->rhs = (integer_zerop (op1)
2397 ? boolean_false_node
2398 : boolean_true_node);
2402 else if (is_gimple_min_invariant (op0)
2403 && (TREE_CODE (op1) == SSA_NAME
2404 || is_gimple_min_invariant (op1)))
2406 tree inverted = invert_truthvalue (cond);
2407 struct edge_info *edge_info;
2409 edge_info = allocate_edge_info (true_edge);
2410 record_conditions (edge_info, cond, inverted);
2412 if (TREE_CODE (cond) == EQ_EXPR)
2414 edge_info->lhs = op1;
2415 edge_info->rhs = op0;
2418 edge_info = allocate_edge_info (false_edge);
2419 record_conditions (edge_info, inverted, cond);
2421 if (TREE_CODE (cond) == NE_EXPR)
2423 edge_info->lhs = op1;
2424 edge_info->rhs = op0;
2428 else if (TREE_CODE (op0) == SSA_NAME
2429 && (is_gimple_min_invariant (op1)
2430 || TREE_CODE (op1) == SSA_NAME))
2432 tree inverted = invert_truthvalue (cond);
2433 struct edge_info *edge_info;
2435 edge_info = allocate_edge_info (true_edge);
2436 record_conditions (edge_info, cond, inverted);
2438 if (TREE_CODE (cond) == EQ_EXPR)
2440 edge_info->lhs = op0;
2441 edge_info->rhs = op1;
2444 edge_info = allocate_edge_info (false_edge);
2445 record_conditions (edge_info, inverted, cond);
2447 if (TREE_CODE (cond) == NE_EXPR)
2449 edge_info->lhs = op0;
2450 edge_info->rhs = op1;
2455 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2460 /* Propagate information from BB to its outgoing edges.
2462 This can include equivalency information implied by control statements
2463 at the end of BB and const/copy propagation into PHIs in BB's
2464 successor blocks. */
2466 static void
2467 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2468 basic_block bb)
2470 record_edge_info (bb);
2471 cprop_into_successor_phis (bb, nonzero_vars);
2474 /* Search for redundant computations in STMT. If any are found, then
2475 replace them with the variable holding the result of the computation.
2477 If safe, record this expression into the available expression hash
2478 table. */
2480 static bool
2481 eliminate_redundant_computations (tree stmt, stmt_ann_t ann)
2483 tree *expr_p, def = NULL_TREE;
2484 bool insert = true;
2485 tree cached_lhs;
2486 bool retval = false;
2488 if (TREE_CODE (stmt) == MODIFY_EXPR)
2489 def = TREE_OPERAND (stmt, 0);
2491 /* Certain expressions on the RHS can be optimized away, but can not
2492 themselves be entered into the hash tables. */
2493 if (ann->makes_aliased_stores
2494 || ! def
2495 || TREE_CODE (def) != SSA_NAME
2496 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
2497 || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VMAYDEF)
2498 /* Do not record equivalences for increments of ivs. This would create
2499 overlapping live ranges for a very questionable gain. */
2500 || simple_iv_increment_p (stmt))
2501 insert = false;
2503 /* Check if the expression has been computed before. */
2504 cached_lhs = lookup_avail_expr (stmt, insert);
2506 /* If this is an assignment and the RHS was not in the hash table,
2507 then try to simplify the RHS and lookup the new RHS in the
2508 hash table. */
2509 if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR)
2510 cached_lhs = simplify_rhs_and_lookup_avail_expr (stmt, insert);
2511 /* Similarly if this is a COND_EXPR and we did not find its
2512 expression in the hash table, simplify the condition and
2513 try again. */
2514 else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
2515 cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
2516 /* Similarly for a SWITCH_EXPR. */
2517 else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR)
2518 cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert);
2520 opt_stats.num_exprs_considered++;
2522 /* Get a pointer to the expression we are trying to optimize. */
2523 if (TREE_CODE (stmt) == COND_EXPR)
2524 expr_p = &COND_EXPR_COND (stmt);
2525 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2526 expr_p = &SWITCH_COND (stmt);
2527 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
2528 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
2529 else
2530 expr_p = &TREE_OPERAND (stmt, 1);
2532 /* It is safe to ignore types here since we have already done
2533 type checking in the hashing and equality routines. In fact
2534 type checking here merely gets in the way of constant
2535 propagation. Also, make sure that it is safe to propagate
2536 CACHED_LHS into *EXPR_P. */
2537 if (cached_lhs
2538 && (TREE_CODE (cached_lhs) != SSA_NAME
2539 || may_propagate_copy (*expr_p, cached_lhs)))
2541 if (dump_file && (dump_flags & TDF_DETAILS))
2543 fprintf (dump_file, " Replaced redundant expr '");
2544 print_generic_expr (dump_file, *expr_p, dump_flags);
2545 fprintf (dump_file, "' with '");
2546 print_generic_expr (dump_file, cached_lhs, dump_flags);
2547 fprintf (dump_file, "'\n");
2550 opt_stats.num_re++;
2552 #if defined ENABLE_CHECKING
2553 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
2554 || is_gimple_min_invariant (cached_lhs));
2555 #endif
2557 if (TREE_CODE (cached_lhs) == ADDR_EXPR
2558 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
2559 && is_gimple_min_invariant (cached_lhs)))
2560 retval = true;
2562 propagate_tree_value (expr_p, cached_lhs);
2563 mark_stmt_modified (stmt);
2565 return retval;
2568 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2569 the available expressions table or the const_and_copies table.
2570 Detect and record those equivalences. */
2572 static void
2573 record_equivalences_from_stmt (tree stmt,
2574 int may_optimize_p,
2575 stmt_ann_t ann)
2577 tree lhs = TREE_OPERAND (stmt, 0);
2578 enum tree_code lhs_code = TREE_CODE (lhs);
2579 int i;
2581 if (lhs_code == SSA_NAME)
2583 tree rhs = TREE_OPERAND (stmt, 1);
2585 /* Strip away any useless type conversions. */
2586 STRIP_USELESS_TYPE_CONVERSION (rhs);
2588 /* If the RHS of the assignment is a constant or another variable that
2589 may be propagated, register it in the CONST_AND_COPIES table. We
2590 do not need to record unwind data for this, since this is a true
2591 assignment and not an equivalence inferred from a comparison. All
2592 uses of this ssa name are dominated by this assignment, so unwinding
2593 just costs time and space. */
2594 if (may_optimize_p
2595 && (TREE_CODE (rhs) == SSA_NAME
2596 || is_gimple_min_invariant (rhs)))
2597 SSA_NAME_VALUE (lhs) = rhs;
2599 if (expr_computes_nonzero (rhs))
2600 record_var_is_nonzero (lhs);
2603 /* Look at both sides for pointer dereferences. If we find one, then
2604 the pointer must be nonnull and we can enter that equivalence into
2605 the hash tables. */
2606 if (flag_delete_null_pointer_checks)
2607 for (i = 0; i < 2; i++)
2609 tree t = TREE_OPERAND (stmt, i);
2611 /* Strip away any COMPONENT_REFs. */
2612 while (TREE_CODE (t) == COMPONENT_REF)
2613 t = TREE_OPERAND (t, 0);
2615 /* Now see if this is a pointer dereference. */
2616 if (INDIRECT_REF_P (t))
2618 tree op = TREE_OPERAND (t, 0);
2620 /* If the pointer is a SSA variable, then enter new
2621 equivalences into the hash table. */
2622 while (TREE_CODE (op) == SSA_NAME)
2624 tree def = SSA_NAME_DEF_STMT (op);
2626 record_var_is_nonzero (op);
2628 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2629 which are known to have a nonzero value. */
2630 if (def
2631 && TREE_CODE (def) == MODIFY_EXPR
2632 && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
2633 op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
2634 else
2635 break;
2640 /* A memory store, even an aliased store, creates a useful
2641 equivalence. By exchanging the LHS and RHS, creating suitable
2642 vops and recording the result in the available expression table,
2643 we may be able to expose more redundant loads. */
2644 if (!ann->has_volatile_ops
2645 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
2646 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
2647 && !is_gimple_reg (lhs))
2649 tree rhs = TREE_OPERAND (stmt, 1);
2650 tree new;
2652 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2653 is a constant, we need to adjust the constant to fit into the
2654 type of the LHS. If the LHS is a bitfield and the RHS is not
2655 a constant, then we can not record any equivalences for this
2656 statement since we would need to represent the widening or
2657 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2658 and should not be necessary if GCC represented bitfields
2659 properly. */
2660 if (lhs_code == COMPONENT_REF
2661 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
2663 if (TREE_CONSTANT (rhs))
2664 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
2665 else
2666 rhs = NULL;
2668 /* If the value overflowed, then we can not use this equivalence. */
2669 if (rhs && ! is_gimple_min_invariant (rhs))
2670 rhs = NULL;
2673 if (rhs)
2675 /* Build a new statement with the RHS and LHS exchanged. */
2676 new = build (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
2678 create_ssa_artficial_load_stmt (new, stmt);
2680 /* Finally enter the statement into the available expression
2681 table. */
2682 lookup_avail_expr (new, true);
2687 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2688 CONST_AND_COPIES. */
2690 static bool
2691 cprop_operand (tree stmt, use_operand_p op_p)
2693 bool may_have_exposed_new_symbols = false;
2694 tree val;
2695 tree op = USE_FROM_PTR (op_p);
2697 /* If the operand has a known constant value or it is known to be a
2698 copy of some other variable, use the value or copy stored in
2699 CONST_AND_COPIES. */
2700 val = SSA_NAME_VALUE (op);
2701 if (val && val != op && TREE_CODE (val) != VALUE_HANDLE)
2703 tree op_type, val_type;
2705 /* Do not change the base variable in the virtual operand
2706 tables. That would make it impossible to reconstruct
2707 the renamed virtual operand if we later modify this
2708 statement. Also only allow the new value to be an SSA_NAME
2709 for propagation into virtual operands. */
2710 if (!is_gimple_reg (op)
2711 && (TREE_CODE (val) != SSA_NAME
2712 || is_gimple_reg (val)
2713 || get_virtual_var (val) != get_virtual_var (op)))
2714 return false;
2716 /* Do not replace hard register operands in asm statements. */
2717 if (TREE_CODE (stmt) == ASM_EXPR
2718 && !may_propagate_copy_into_asm (op))
2719 return false;
2721 /* Get the toplevel type of each operand. */
2722 op_type = TREE_TYPE (op);
2723 val_type = TREE_TYPE (val);
2725 /* While both types are pointers, get the type of the object
2726 pointed to. */
2727 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
2729 op_type = TREE_TYPE (op_type);
2730 val_type = TREE_TYPE (val_type);
2733 /* Make sure underlying types match before propagating a constant by
2734 converting the constant to the proper type. Note that convert may
2735 return a non-gimple expression, in which case we ignore this
2736 propagation opportunity. */
2737 if (TREE_CODE (val) != SSA_NAME)
2739 if (!lang_hooks.types_compatible_p (op_type, val_type))
2741 val = fold_convert (TREE_TYPE (op), val);
2742 if (!is_gimple_min_invariant (val))
2743 return false;
2747 /* Certain operands are not allowed to be copy propagated due
2748 to their interaction with exception handling and some GCC
2749 extensions. */
2750 else if (!may_propagate_copy (op, val))
2751 return false;
2753 /* Do not propagate copies if the propagated value is at a deeper loop
2754 depth than the propagatee. Otherwise, this may move loop variant
2755 variables outside of their loops and prevent coalescing
2756 opportunities. If the value was loop invariant, it will be hoisted
2757 by LICM and exposed for copy propagation. */
2758 if (loop_depth_of_name (val) > loop_depth_of_name (op))
2759 return false;
2761 /* Dump details. */
2762 if (dump_file && (dump_flags & TDF_DETAILS))
2764 fprintf (dump_file, " Replaced '");
2765 print_generic_expr (dump_file, op, dump_flags);
2766 fprintf (dump_file, "' with %s '",
2767 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
2768 print_generic_expr (dump_file, val, dump_flags);
2769 fprintf (dump_file, "'\n");
2772 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2773 that we may have exposed a new symbol for SSA renaming. */
2774 if (TREE_CODE (val) == ADDR_EXPR
2775 || (POINTER_TYPE_P (TREE_TYPE (op))
2776 && is_gimple_min_invariant (val)))
2777 may_have_exposed_new_symbols = true;
2779 if (TREE_CODE (val) != SSA_NAME)
2780 opt_stats.num_const_prop++;
2781 else
2782 opt_stats.num_copy_prop++;
2784 propagate_value (op_p, val);
2786 /* And note that we modified this statement. This is now
2787 safe, even if we changed virtual operands since we will
2788 rescan the statement and rewrite its operands again. */
2789 mark_stmt_modified (stmt);
2791 return may_have_exposed_new_symbols;
2794 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2795 known value for that SSA_NAME (or NULL if no value is known).
2797 Propagate values from CONST_AND_COPIES into the uses, vuses and
2798 v_may_def_ops of STMT. */
2800 static bool
2801 cprop_into_stmt (tree stmt)
2803 bool may_have_exposed_new_symbols = false;
2804 use_operand_p op_p;
2805 ssa_op_iter iter;
2807 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
2809 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
2810 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
2813 return may_have_exposed_new_symbols;
2817 /* Optimize the statement pointed by iterator SI.
2819 We try to perform some simplistic global redundancy elimination and
2820 constant propagation:
2822 1- To detect global redundancy, we keep track of expressions that have
2823 been computed in this block and its dominators. If we find that the
2824 same expression is computed more than once, we eliminate repeated
2825 computations by using the target of the first one.
2827 2- Constant values and copy assignments. This is used to do very
2828 simplistic constant and copy propagation. When a constant or copy
2829 assignment is found, we map the value on the RHS of the assignment to
2830 the variable in the LHS in the CONST_AND_COPIES table. */
2832 static void
2833 optimize_stmt (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2834 basic_block bb, block_stmt_iterator si)
2836 stmt_ann_t ann;
2837 tree stmt, old_stmt;
2838 bool may_optimize_p;
2839 bool may_have_exposed_new_symbols = false;
2841 old_stmt = stmt = bsi_stmt (si);
2843 update_stmt_if_modified (stmt);
2844 ann = stmt_ann (stmt);
2845 opt_stats.num_stmts++;
2846 may_have_exposed_new_symbols = false;
2848 if (dump_file && (dump_flags & TDF_DETAILS))
2850 fprintf (dump_file, "Optimizing statement ");
2851 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2854 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2855 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
2857 /* If the statement has been modified with constant replacements,
2858 fold its RHS before checking for redundant computations. */
2859 if (ann->modified)
2861 tree rhs;
2863 /* Try to fold the statement making sure that STMT is kept
2864 up to date. */
2865 if (fold_stmt (bsi_stmt_ptr (si)))
2867 stmt = bsi_stmt (si);
2868 ann = stmt_ann (stmt);
2870 if (dump_file && (dump_flags & TDF_DETAILS))
2872 fprintf (dump_file, " Folded to: ");
2873 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2877 rhs = get_rhs (stmt);
2878 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
2879 recompute_tree_invarant_for_addr_expr (rhs);
2881 /* Constant/copy propagation above may change the set of
2882 virtual operands associated with this statement. Folding
2883 may remove the need for some virtual operands.
2885 Indicate we will need to rescan and rewrite the statement. */
2886 may_have_exposed_new_symbols = true;
2889 /* Check for redundant computations. Do this optimization only
2890 for assignments that have no volatile ops and conditionals. */
2891 may_optimize_p = (!ann->has_volatile_ops
2892 && ((TREE_CODE (stmt) == RETURN_EXPR
2893 && TREE_OPERAND (stmt, 0)
2894 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
2895 && ! (TREE_SIDE_EFFECTS
2896 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
2897 || (TREE_CODE (stmt) == MODIFY_EXPR
2898 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
2899 || TREE_CODE (stmt) == COND_EXPR
2900 || TREE_CODE (stmt) == SWITCH_EXPR));
2902 if (may_optimize_p)
2903 may_have_exposed_new_symbols
2904 |= eliminate_redundant_computations (stmt, ann);
2906 /* Record any additional equivalences created by this statement. */
2907 if (TREE_CODE (stmt) == MODIFY_EXPR)
2908 record_equivalences_from_stmt (stmt,
2909 may_optimize_p,
2910 ann);
2912 /* If STMT is a COND_EXPR and it was modified, then we may know
2913 where it goes. If that is the case, then mark the CFG as altered.
2915 This will cause us to later call remove_unreachable_blocks and
2916 cleanup_tree_cfg when it is safe to do so. It is not safe to
2917 clean things up here since removal of edges and such can trigger
2918 the removal of PHI nodes, which in turn can release SSA_NAMEs to
2919 the manager.
2921 That's all fine and good, except that once SSA_NAMEs are released
2922 to the manager, we must not call create_ssa_name until all references
2923 to released SSA_NAMEs have been eliminated.
2925 All references to the deleted SSA_NAMEs can not be eliminated until
2926 we remove unreachable blocks.
2928 We can not remove unreachable blocks until after we have completed
2929 any queued jump threading.
2931 We can not complete any queued jump threads until we have taken
2932 appropriate variables out of SSA form. Taking variables out of
2933 SSA form can call create_ssa_name and thus we lose.
2935 Ultimately I suspect we're going to need to change the interface
2936 into the SSA_NAME manager. */
2938 if (ann->modified)
2940 tree val = NULL;
2942 if (TREE_CODE (stmt) == COND_EXPR)
2943 val = COND_EXPR_COND (stmt);
2944 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2945 val = SWITCH_COND (stmt);
2947 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
2948 cfg_altered = true;
2950 /* If we simplified a statement in such a way as to be shown that it
2951 cannot trap, update the eh information and the cfg to match. */
2952 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
2954 bitmap_set_bit (need_eh_cleanup, bb->index);
2955 if (dump_file && (dump_flags & TDF_DETAILS))
2956 fprintf (dump_file, " Flagged to clear EH edges.\n");
2960 if (may_have_exposed_new_symbols)
2961 VEC_safe_push (tree, heap, stmts_to_rescan, bsi_stmt (si));
2964 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
2965 available expression hashtable, then return the LHS from the hash
2966 table.
2968 If INSERT is true, then we also update the available expression
2969 hash table to account for the changes made to STMT. */
2971 static tree
2972 update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert)
2974 tree cached_lhs = NULL;
2976 /* Remove the old entry from the hash table. */
2977 if (insert)
2979 struct expr_hash_elt element;
2981 initialize_hash_element (stmt, NULL, &element);
2982 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
2985 /* Now update the RHS of the assignment. */
2986 TREE_OPERAND (stmt, 1) = new_rhs;
2988 /* Now lookup the updated statement in the hash table. */
2989 cached_lhs = lookup_avail_expr (stmt, insert);
2991 /* We have now called lookup_avail_expr twice with two different
2992 versions of this same statement, once in optimize_stmt, once here.
2994 We know the call in optimize_stmt did not find an existing entry
2995 in the hash table, so a new entry was created. At the same time
2996 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
2998 If this call failed to find an existing entry on the hash table,
2999 then the new version of this statement was entered into the
3000 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3001 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3003 If this call succeeded, we still have one copy of this statement
3004 on the BLOCK_AVAIL_EXPRs vector.
3006 For both cases, we need to pop the most recent entry off the
3007 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3008 statement in the hash tables, that will leave precisely one
3009 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3010 we found a copy of this statement in the second hash table lookup
3011 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3012 if (insert)
3013 VEC_pop (tree, avail_exprs_stack);
3015 /* And make sure we record the fact that we modified this
3016 statement. */
3017 mark_stmt_modified (stmt);
3019 return cached_lhs;
3022 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3023 found, return its LHS. Otherwise insert STMT in the table and return
3024 NULL_TREE.
3026 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3027 is also added to the stack pointed by BLOCK_AVAIL_EXPRS_P, so that they
3028 can be removed when we finish processing this block and its children.
3030 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3031 contains no CALL_EXPR on its RHS and makes no volatile nor
3032 aliased references. */
3034 static tree
3035 lookup_avail_expr (tree stmt, bool insert)
3037 void **slot;
3038 tree lhs;
3039 tree temp;
3040 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
3042 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
3044 initialize_hash_element (stmt, lhs, element);
3046 /* Don't bother remembering constant assignments and copy operations.
3047 Constants and copy operations are handled by the constant/copy propagator
3048 in optimize_stmt. */
3049 if (TREE_CODE (element->rhs) == SSA_NAME
3050 || is_gimple_min_invariant (element->rhs))
3052 free (element);
3053 return NULL_TREE;
3056 /* If this is an equality test against zero, see if we have recorded a
3057 nonzero value for the variable in question. */
3058 if ((TREE_CODE (element->rhs) == EQ_EXPR
3059 || TREE_CODE (element->rhs) == NE_EXPR)
3060 && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
3061 && integer_zerop (TREE_OPERAND (element->rhs, 1)))
3063 int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
3065 if (bitmap_bit_p (nonzero_vars, indx))
3067 tree t = element->rhs;
3068 free (element);
3070 if (TREE_CODE (t) == EQ_EXPR)
3071 return boolean_false_node;
3072 else
3073 return boolean_true_node;
3077 /* Finally try to find the expression in the main expression hash table. */
3078 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
3079 (insert ? INSERT : NO_INSERT));
3080 if (slot == NULL)
3082 free (element);
3083 return NULL_TREE;
3086 if (*slot == NULL)
3088 *slot = (void *) element;
3089 VEC_safe_push (tree, heap, avail_exprs_stack,
3090 stmt ? stmt : element->rhs);
3091 return NULL_TREE;
3094 /* Extract the LHS of the assignment so that it can be used as the current
3095 definition of another variable. */
3096 lhs = ((struct expr_hash_elt *)*slot)->lhs;
3098 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3099 use the value from the const_and_copies table. */
3100 if (TREE_CODE (lhs) == SSA_NAME)
3102 temp = SSA_NAME_VALUE (lhs);
3103 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
3104 lhs = temp;
3107 free (element);
3108 return lhs;
3111 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3112 range of values that result in the conditional having a true value.
3114 Return true if we are successful in extracting a range from COND and
3115 false if we are unsuccessful. */
3117 static bool
3118 extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
3120 tree op1 = TREE_OPERAND (cond, 1);
3121 tree high, low, type;
3122 int inverted;
3124 type = TREE_TYPE (op1);
3126 /* Experiments have shown that it's rarely, if ever useful to
3127 record ranges for enumerations. Presumably this is due to
3128 the fact that they're rarely used directly. They are typically
3129 cast into an integer type and used that way. */
3130 if (TREE_CODE (type) != INTEGER_TYPE
3131 /* We don't know how to deal with types with variable bounds. */
3132 || TREE_CODE (TYPE_MIN_VALUE (type)) != INTEGER_CST
3133 || TREE_CODE (TYPE_MAX_VALUE (type)) != INTEGER_CST)
3134 return 0;
3136 switch (TREE_CODE (cond))
3138 case EQ_EXPR:
3139 high = low = op1;
3140 inverted = 0;
3141 break;
3143 case NE_EXPR:
3144 high = low = op1;
3145 inverted = 1;
3146 break;
3148 case GE_EXPR:
3149 low = op1;
3150 high = TYPE_MAX_VALUE (type);
3151 inverted = 0;
3152 break;
3154 case GT_EXPR:
3155 high = TYPE_MAX_VALUE (type);
3156 if (!tree_int_cst_lt (op1, high))
3157 return 0;
3158 low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
3159 inverted = 0;
3160 break;
3162 case LE_EXPR:
3163 high = op1;
3164 low = TYPE_MIN_VALUE (type);
3165 inverted = 0;
3166 break;
3168 case LT_EXPR:
3169 low = TYPE_MIN_VALUE (type);
3170 if (!tree_int_cst_lt (low, op1))
3171 return 0;
3172 high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
3173 inverted = 0;
3174 break;
3176 default:
3177 return 0;
3180 *hi_p = high;
3181 *lo_p = low;
3182 *inverted_p = inverted;
3183 return 1;
3186 /* Record a range created by COND for basic block BB. */
3188 static void
3189 record_range (tree cond, basic_block bb)
3191 enum tree_code code = TREE_CODE (cond);
3193 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3194 They rarely allow for meaningful range optimizations and significantly
3195 complicate the implementation. */
3196 if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
3197 || code == GE_EXPR || code == EQ_EXPR)
3198 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
3200 struct vrp_hash_elt *vrp_hash_elt;
3201 struct vrp_element *element;
3202 VEC(vrp_element_p,heap) **vrp_records_p;
3203 void **slot;
3206 vrp_hash_elt = xmalloc (sizeof (struct vrp_hash_elt));
3207 vrp_hash_elt->var = TREE_OPERAND (cond, 0);
3208 vrp_hash_elt->records = NULL;
3209 slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
3211 if (*slot == NULL)
3212 *slot = (void *) vrp_hash_elt;
3213 else
3214 vrp_free (vrp_hash_elt);
3216 vrp_hash_elt = (struct vrp_hash_elt *) *slot;
3217 vrp_records_p = &vrp_hash_elt->records;
3219 element = ggc_alloc (sizeof (struct vrp_element));
3220 element->low = NULL;
3221 element->high = NULL;
3222 element->cond = cond;
3223 element->bb = bb;
3225 VEC_safe_push (vrp_element_p, heap, *vrp_records_p, element);
3226 VEC_safe_push (tree, heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
3230 /* Hashing and equality functions for VRP_DATA.
3232 Since this hash table is addressed by SSA_NAMEs, we can hash on
3233 their version number and equality can be determined with a
3234 pointer comparison. */
3236 static hashval_t
3237 vrp_hash (const void *p)
3239 tree var = ((struct vrp_hash_elt *)p)->var;
3241 return SSA_NAME_VERSION (var);
3244 static int
3245 vrp_eq (const void *p1, const void *p2)
3247 tree var1 = ((struct vrp_hash_elt *)p1)->var;
3248 tree var2 = ((struct vrp_hash_elt *)p2)->var;
3250 return var1 == var2;
3253 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3254 MODIFY_EXPR statements. We compute a value number for expressions using
3255 the code of the expression and the SSA numbers of its operands. */
3257 static hashval_t
3258 avail_expr_hash (const void *p)
3260 tree stmt = ((struct expr_hash_elt *)p)->stmt;
3261 tree rhs = ((struct expr_hash_elt *)p)->rhs;
3262 tree vuse;
3263 ssa_op_iter iter;
3264 hashval_t val = 0;
3266 /* iterative_hash_expr knows how to deal with any expression and
3267 deals with commutative operators as well, so just use it instead
3268 of duplicating such complexities here. */
3269 val = iterative_hash_expr (rhs, val);
3271 /* If the hash table entry is not associated with a statement, then we
3272 can just hash the expression and not worry about virtual operands
3273 and such. */
3274 if (!stmt || !stmt_ann (stmt))
3275 return val;
3277 /* Add the SSA version numbers of every vuse operand. This is important
3278 because compound variables like arrays are not renamed in the
3279 operands. Rather, the rename is done on the virtual variable
3280 representing all the elements of the array. */
3281 FOR_EACH_SSA_TREE_OPERAND (vuse, stmt, iter, SSA_OP_VUSE)
3282 val = iterative_hash_expr (vuse, val);
3284 return val;
3287 static hashval_t
3288 real_avail_expr_hash (const void *p)
3290 return ((const struct expr_hash_elt *)p)->hash;
3293 static int
3294 avail_expr_eq (const void *p1, const void *p2)
3296 tree stmt1 = ((struct expr_hash_elt *)p1)->stmt;
3297 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
3298 tree stmt2 = ((struct expr_hash_elt *)p2)->stmt;
3299 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
3301 /* If they are the same physical expression, return true. */
3302 if (rhs1 == rhs2 && stmt1 == stmt2)
3303 return true;
3305 /* If their codes are not equal, then quit now. */
3306 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
3307 return false;
3309 /* In case of a collision, both RHS have to be identical and have the
3310 same VUSE operands. */
3311 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
3312 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
3313 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
3315 bool ret = compare_ssa_operands_equal (stmt1, stmt2, SSA_OP_VUSE);
3316 gcc_assert (!ret || ((struct expr_hash_elt *)p1)->hash
3317 == ((struct expr_hash_elt *)p2)->hash);
3318 return ret;
3321 return false;