* opt-functions.awk (var_type): New function.
[official-gcc.git] / gcc / tree-ssa-dom.c
blob278d27a9c7a88412623fdd15e5f1e461ed6cdee8
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, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, 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 "errors.h"
35 #include "expr.h"
36 #include "function.h"
37 #include "diagnostic.h"
38 #include "timevar.h"
39 #include "tree-dump.h"
40 #include "tree-flow.h"
41 #include "domwalk.h"
42 #include "real.h"
43 #include "tree-pass.h"
44 #include "tree-ssa-propagate.h"
45 #include "langhooks.h"
47 /* This file implements optimizations on the dominator tree. */
50 /* Structure for recording edge equivalences as well as any pending
51 edge redirections during the dominator optimizer.
53 Computing and storing the edge equivalences instead of creating
54 them on-demand can save significant amounts of time, particularly
55 for pathological cases involving switch statements.
57 These structures live for a single iteration of the dominator
58 optimizer in the edge's AUX field. At the end of an iteration we
59 free each of these structures and update the AUX field to point
60 to any requested redirection target (the code for updating the
61 CFG and SSA graph for edge redirection expects redirection edge
62 targets to be in the AUX field for each edge. */
64 struct edge_info
66 /* If this edge creates a simple equivalence, the LHS and RHS of
67 the equivalence will be stored here. */
68 tree lhs;
69 tree rhs;
71 /* Traversing an edge may also indicate one or more particular conditions
72 are true or false. The number of recorded conditions can vary, but
73 can be determined by the condition's code. So we have an array
74 and its maximum index rather than use a varray. */
75 tree *cond_equivalences;
76 unsigned int max_cond_equivalences;
78 /* If we can thread this edge this field records the new target. */
79 edge redirection_target;
83 /* Hash table with expressions made available during the renaming process.
84 When an assignment of the form X_i = EXPR is found, the statement is
85 stored in this table. If the same expression EXPR is later found on the
86 RHS of another statement, it is replaced with X_i (thus performing
87 global redundancy elimination). Similarly as we pass through conditionals
88 we record the conditional itself as having either a true or false value
89 in this table. */
90 static htab_t avail_exprs;
92 /* Stack of available expressions in AVAIL_EXPRs. Each block pushes any
93 expressions it enters into the hash table along with a marker entry
94 (null). When we finish processing the block, we pop off entries and
95 remove the expressions from the global hash table until we hit the
96 marker. */
97 static VEC(tree,heap) *avail_exprs_stack;
99 /* Stack of statements we need to rescan during finalization for newly
100 exposed variables.
102 Statement rescanning must occur after the current block's available
103 expressions are removed from AVAIL_EXPRS. Else we may change the
104 hash code for an expression and be unable to find/remove it from
105 AVAIL_EXPRS. */
106 static VEC(tree,heap) *stmts_to_rescan;
108 /* Structure for entries in the expression hash table.
110 This requires more memory for the hash table entries, but allows us
111 to avoid creating silly tree nodes and annotations for conditionals,
112 eliminates 2 global hash tables and two block local varrays.
114 It also allows us to reduce the number of hash table lookups we
115 have to perform in lookup_avail_expr and finally it allows us to
116 significantly reduce the number of calls into the hashing routine
117 itself. */
119 struct expr_hash_elt
121 /* The value (lhs) of this expression. */
122 tree lhs;
124 /* The expression (rhs) we want to record. */
125 tree rhs;
127 /* The stmt pointer if this element corresponds to a statement. */
128 tree stmt;
130 /* The hash value for RHS/ann. */
131 hashval_t hash;
134 /* Stack of dest,src pairs that need to be restored during finalization.
136 A NULL entry is used to mark the end of pairs which need to be
137 restored during finalization of this block. */
138 static VEC(tree,heap) *const_and_copies_stack;
140 /* Bitmap of SSA_NAMEs known to have a nonzero value, even if we do not
141 know their exact value. */
142 static bitmap nonzero_vars;
144 /* Stack of SSA_NAMEs which need their NONZERO_VARS property cleared
145 when the current block is finalized.
147 A NULL entry is used to mark the end of names needing their
148 entry in NONZERO_VARS cleared during finalization of this block. */
149 static VEC(tree,heap) *nonzero_vars_stack;
151 /* Track whether or not we have changed the control flow graph. */
152 static bool cfg_altered;
154 /* Bitmap of blocks that have had EH statements cleaned. We should
155 remove their dead edges eventually. */
156 static bitmap need_eh_cleanup;
158 /* Statistics for dominator optimizations. */
159 struct opt_stats_d
161 long num_stmts;
162 long num_exprs_considered;
163 long num_re;
164 long num_const_prop;
165 long num_copy_prop;
168 static struct opt_stats_d opt_stats;
170 /* Value range propagation record. Each time we encounter a conditional
171 of the form SSA_NAME COND CONST we create a new vrp_element to record
172 how the condition affects the possible values SSA_NAME may have.
174 Each record contains the condition tested (COND), and the range of
175 values the variable may legitimately have if COND is true. Note the
176 range of values may be a smaller range than COND specifies if we have
177 recorded other ranges for this variable. Each record also contains the
178 block in which the range was recorded for invalidation purposes.
180 Note that the current known range is computed lazily. This allows us
181 to avoid the overhead of computing ranges which are never queried.
183 When we encounter a conditional, we look for records which constrain
184 the SSA_NAME used in the condition. In some cases those records allow
185 us to determine the condition's result at compile time. In other cases
186 they may allow us to simplify the condition.
188 We also use value ranges to do things like transform signed div/mod
189 operations into unsigned div/mod or to simplify ABS_EXPRs.
191 Simple experiments have shown these optimizations to not be all that
192 useful on switch statements (much to my surprise). So switch statement
193 optimizations are not performed.
195 Note carefully we do not propagate information through each statement
196 in the block. i.e., if we know variable X has a value defined of
197 [0, 25] and we encounter Y = X + 1, we do not track a value range
198 for Y (which would be [1, 26] if we cared). Similarly we do not
199 constrain values as we encounter narrowing typecasts, etc. */
201 struct vrp_element
203 /* The highest and lowest values the variable in COND may contain when
204 COND is true. Note this may not necessarily be the same values
205 tested by COND if the same variable was used in earlier conditionals.
207 Note this is computed lazily and thus can be NULL indicating that
208 the values have not been computed yet. */
209 tree low;
210 tree high;
212 /* The actual conditional we recorded. This is needed since we compute
213 ranges lazily. */
214 tree cond;
216 /* The basic block where this record was created. We use this to determine
217 when to remove records. */
218 basic_block bb;
221 /* A hash table holding value range records (VRP_ELEMENTs) for a given
222 SSA_NAME. We used to use a varray indexed by SSA_NAME_VERSION, but
223 that gets awful wasteful, particularly since the density objects
224 with useful information is very low. */
225 static htab_t vrp_data;
227 /* An entry in the VRP_DATA hash table. We record the variable and a
228 varray of VRP_ELEMENT records associated with that variable. */
229 struct vrp_hash_elt
231 tree var;
232 varray_type records;
235 /* Array of variables which have their values constrained by operations
236 in this basic block. We use this during finalization to know
237 which variables need their VRP data updated. */
239 /* Stack of SSA_NAMEs which had their values constrained by operations
240 in this basic block. During finalization of this block we use this
241 list to determine which variables need their VRP data updated.
243 A NULL entry marks the end of the SSA_NAMEs associated with this block. */
244 static VEC(tree,heap) *vrp_variables_stack;
246 struct eq_expr_value
248 tree src;
249 tree dst;
252 /* Local functions. */
253 static void optimize_stmt (struct dom_walk_data *,
254 basic_block bb,
255 block_stmt_iterator);
256 static tree lookup_avail_expr (tree, bool);
257 static hashval_t vrp_hash (const void *);
258 static int vrp_eq (const void *, const void *);
259 static hashval_t avail_expr_hash (const void *);
260 static hashval_t real_avail_expr_hash (const void *);
261 static int avail_expr_eq (const void *, const void *);
262 static void htab_statistics (FILE *, htab_t);
263 static void record_cond (tree, tree);
264 static void record_const_or_copy (tree, tree);
265 static void record_equality (tree, tree);
266 static tree update_rhs_and_lookup_avail_expr (tree, tree, bool);
267 static tree simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *,
268 tree, int);
269 static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int);
270 static tree simplify_switch_and_lookup_avail_expr (tree, int);
271 static tree find_equivalent_equality_comparison (tree);
272 static void record_range (tree, basic_block);
273 static bool extract_range_from_cond (tree, tree *, tree *, int *);
274 static void record_equivalences_from_phis (basic_block);
275 static void record_equivalences_from_incoming_edge (basic_block);
276 static bool eliminate_redundant_computations (struct dom_walk_data *,
277 tree, stmt_ann_t);
278 static void record_equivalences_from_stmt (tree, int, stmt_ann_t);
279 static void thread_across_edge (struct dom_walk_data *, edge);
280 static void dom_opt_finalize_block (struct dom_walk_data *, basic_block);
281 static void dom_opt_initialize_block (struct dom_walk_data *, basic_block);
282 static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block);
283 static void remove_local_expressions_from_table (void);
284 static void restore_vars_to_original_value (void);
285 static edge single_incoming_edge_ignoring_loop_edges (basic_block);
286 static void restore_nonzero_vars_to_original_value (void);
287 static inline bool unsafe_associative_fp_binop (tree);
290 /* Local version of fold that doesn't introduce cruft. */
292 static tree
293 local_fold (tree t)
295 t = fold (t);
297 /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that
298 may have been added by fold, and "useless" type conversions that might
299 now be apparent due to propagation. */
300 STRIP_USELESS_TYPE_CONVERSION (t);
302 return t;
305 /* Allocate an EDGE_INFO for edge E and attach it to E.
306 Return the new EDGE_INFO structure. */
308 static struct edge_info *
309 allocate_edge_info (edge e)
311 struct edge_info *edge_info;
313 edge_info = xcalloc (1, sizeof (struct edge_info));
315 e->aux = edge_info;
316 return edge_info;
319 /* Free all EDGE_INFO structures associated with edges in the CFG.
320 If a particular edge can be threaded, copy the redirection
321 target from the EDGE_INFO structure into the edge's AUX field
322 as required by code to update the CFG and SSA graph for
323 jump threading. */
325 static void
326 free_all_edge_infos (void)
328 basic_block bb;
329 edge_iterator ei;
330 edge e;
332 FOR_EACH_BB (bb)
334 FOR_EACH_EDGE (e, ei, bb->preds)
336 struct edge_info *edge_info = e->aux;
338 if (edge_info)
340 e->aux = edge_info->redirection_target;
341 if (edge_info->cond_equivalences)
342 free (edge_info->cond_equivalences);
343 free (edge_info);
349 /* Jump threading, redundancy elimination and const/copy propagation.
351 This pass may expose new symbols that need to be renamed into SSA. For
352 every new symbol exposed, its corresponding bit will be set in
353 VARS_TO_RENAME. */
355 static void
356 tree_ssa_dominator_optimize (void)
358 struct dom_walk_data walk_data;
359 unsigned int i;
360 struct loops loops_info;
362 memset (&opt_stats, 0, sizeof (opt_stats));
364 /* Create our hash tables. */
365 avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free);
366 vrp_data = htab_create (ceil_log2 (num_ssa_names), vrp_hash, vrp_eq, free);
367 avail_exprs_stack = VEC_alloc (tree, heap, 20);
368 const_and_copies_stack = VEC_alloc (tree, heap, 20);
369 nonzero_vars_stack = VEC_alloc (tree, heap, 20);
370 vrp_variables_stack = VEC_alloc (tree, heap, 20);
371 stmts_to_rescan = VEC_alloc (tree, heap, 20);
372 nonzero_vars = BITMAP_ALLOC (NULL);
373 need_eh_cleanup = BITMAP_ALLOC (NULL);
375 /* Setup callbacks for the generic dominator tree walker. */
376 walk_data.walk_stmts_backward = false;
377 walk_data.dom_direction = CDI_DOMINATORS;
378 walk_data.initialize_block_local_data = NULL;
379 walk_data.before_dom_children_before_stmts = dom_opt_initialize_block;
380 walk_data.before_dom_children_walk_stmts = optimize_stmt;
381 walk_data.before_dom_children_after_stmts = propagate_to_outgoing_edges;
382 walk_data.after_dom_children_before_stmts = NULL;
383 walk_data.after_dom_children_walk_stmts = NULL;
384 walk_data.after_dom_children_after_stmts = dom_opt_finalize_block;
385 /* Right now we only attach a dummy COND_EXPR to the global data pointer.
386 When we attach more stuff we'll need to fill this out with a real
387 structure. */
388 walk_data.global_data = NULL;
389 walk_data.block_local_data_size = 0;
390 walk_data.interesting_blocks = NULL;
392 /* Now initialize the dominator walker. */
393 init_walk_dominator_tree (&walk_data);
395 calculate_dominance_info (CDI_DOMINATORS);
397 /* We need to know which edges exit loops so that we can
398 aggressively thread through loop headers to an exit
399 edge. */
400 flow_loops_find (&loops_info);
401 mark_loop_exit_edges (&loops_info);
402 flow_loops_free (&loops_info);
404 /* Clean up the CFG so that any forwarder blocks created by loop
405 canonicalization are removed. */
406 cleanup_tree_cfg ();
407 calculate_dominance_info (CDI_DOMINATORS);
409 /* If we prove certain blocks are unreachable, then we want to
410 repeat the dominator optimization process as PHI nodes may
411 have turned into copies which allows better propagation of
412 values. So we repeat until we do not identify any new unreachable
413 blocks. */
416 /* Optimize the dominator tree. */
417 cfg_altered = false;
419 /* We need accurate information regarding back edges in the CFG
420 for jump threading. */
421 mark_dfs_back_edges ();
423 /* Recursively walk the dominator tree optimizing statements. */
424 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
426 /* If we exposed any new variables, go ahead and put them into
427 SSA form now, before we handle jump threading. This simplifies
428 interactions between rewriting of _DECL nodes into SSA form
429 and rewriting SSA_NAME nodes into SSA form after block
430 duplication and CFG manipulation. */
431 update_ssa (TODO_update_ssa);
433 free_all_edge_infos ();
436 block_stmt_iterator bsi;
437 basic_block bb;
438 FOR_EACH_BB (bb)
440 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
442 update_stmt_if_modified (bsi_stmt (bsi));
447 /* Thread jumps, creating duplicate blocks as needed. */
448 cfg_altered |= thread_through_all_blocks ();
450 /* Removal of statements may make some EH edges dead. Purge
451 such edges from the CFG as needed. */
452 if (!bitmap_empty_p (need_eh_cleanup))
454 cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
455 bitmap_zero (need_eh_cleanup);
458 if (cfg_altered)
459 free_dominance_info (CDI_DOMINATORS);
461 cfg_altered = cleanup_tree_cfg ();
463 if (rediscover_loops_after_threading)
465 /* Rerun basic loop analysis to discover any newly
466 created loops and update the set of exit edges. */
467 rediscover_loops_after_threading = false;
468 flow_loops_find (&loops_info);
469 mark_loop_exit_edges (&loops_info);
470 flow_loops_free (&loops_info);
472 /* Remove any forwarder blocks inserted by loop
473 header canonicalization. */
474 cleanup_tree_cfg ();
477 calculate_dominance_info (CDI_DOMINATORS);
479 update_ssa (TODO_update_ssa);
481 /* Reinitialize the various tables. */
482 bitmap_clear (nonzero_vars);
483 htab_empty (avail_exprs);
484 htab_empty (vrp_data);
486 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
488 This must be done before we iterate as we might have a
489 reference to an SSA_NAME which was removed by the call to
490 update_ssa.
492 Long term we will be able to let everything in SSA_NAME_VALUE
493 persist. However, for now, we know this is the safe thing to do. */
494 for (i = 0; i < num_ssa_names; i++)
496 tree name = ssa_name (i);
497 tree value;
499 if (!name)
500 continue;
502 value = SSA_NAME_VALUE (name);
503 if (value && !is_gimple_min_invariant (value))
504 SSA_NAME_VALUE (name) = NULL;
507 while (optimize > 1 && cfg_altered);
509 /* Debugging dumps. */
510 if (dump_file && (dump_flags & TDF_STATS))
511 dump_dominator_optimization_stats (dump_file);
513 /* We emptied the hash table earlier, now delete it completely. */
514 htab_delete (avail_exprs);
515 htab_delete (vrp_data);
517 /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
518 CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
519 of the do-while loop above. */
521 /* And finalize the dominator walker. */
522 fini_walk_dominator_tree (&walk_data);
524 /* Free nonzero_vars. */
525 BITMAP_FREE (nonzero_vars);
526 BITMAP_FREE (need_eh_cleanup);
528 VEC_free (tree, heap, avail_exprs_stack);
529 VEC_free (tree, heap, const_and_copies_stack);
530 VEC_free (tree, heap, nonzero_vars_stack);
531 VEC_free (tree, heap, vrp_variables_stack);
532 VEC_free (tree, heap, stmts_to_rescan);
535 static bool
536 gate_dominator (void)
538 return flag_tree_dom != 0;
541 struct tree_opt_pass pass_dominator =
543 "dom", /* name */
544 gate_dominator, /* gate */
545 tree_ssa_dominator_optimize, /* execute */
546 NULL, /* sub */
547 NULL, /* next */
548 0, /* static_pass_number */
549 TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */
550 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
551 0, /* properties_provided */
552 0, /* properties_destroyed */
553 0, /* todo_flags_start */
554 TODO_dump_func
555 | TODO_update_ssa
556 | TODO_verify_ssa, /* todo_flags_finish */
557 0 /* letter */
561 /* We are exiting E->src, see if E->dest ends with a conditional
562 jump which has a known value when reached via E.
564 Special care is necessary if E is a back edge in the CFG as we
565 will have already recorded equivalences for E->dest into our
566 various tables, including the result of the conditional at
567 the end of E->dest. Threading opportunities are severely
568 limited in that case to avoid short-circuiting the loop
569 incorrectly.
571 Note it is quite common for the first block inside a loop to
572 end with a conditional which is either always true or always
573 false when reached via the loop backedge. Thus we do not want
574 to blindly disable threading across a loop backedge. */
576 static void
577 thread_across_edge (struct dom_walk_data *walk_data, edge e)
579 block_stmt_iterator bsi;
580 tree stmt = NULL;
581 tree phi;
583 /* If E->dest does not end with a conditional, then there is
584 nothing to do. */
585 bsi = bsi_last (e->dest);
586 if (bsi_end_p (bsi)
587 || ! bsi_stmt (bsi)
588 || (TREE_CODE (bsi_stmt (bsi)) != COND_EXPR
589 && TREE_CODE (bsi_stmt (bsi)) != GOTO_EXPR
590 && TREE_CODE (bsi_stmt (bsi)) != SWITCH_EXPR))
591 return;
593 /* The basic idea here is to use whatever knowledge we have
594 from our dominator walk to simplify statements in E->dest,
595 with the ultimate goal being to simplify the conditional
596 at the end of E->dest.
598 Note that we must undo any changes we make to the underlying
599 statements as the simplifications we are making are control
600 flow sensitive (ie, the simplifications are valid when we
601 traverse E, but may not be valid on other paths to E->dest. */
603 /* Each PHI creates a temporary equivalence, record them. Again
604 these are context sensitive equivalences and will be removed
605 by our caller. */
606 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
608 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
609 tree dst = PHI_RESULT (phi);
611 /* If the desired argument is not the same as this PHI's result
612 and it is set by a PHI in E->dest, then we can not thread
613 through E->dest. */
614 if (src != dst
615 && TREE_CODE (src) == SSA_NAME
616 && TREE_CODE (SSA_NAME_DEF_STMT (src)) == PHI_NODE
617 && bb_for_stmt (SSA_NAME_DEF_STMT (src)) == e->dest)
618 return;
620 record_const_or_copy (dst, src);
623 /* Try to simplify each statement in E->dest, ultimately leading to
624 a simplification of the COND_EXPR at the end of E->dest.
626 We might consider marking just those statements which ultimately
627 feed the COND_EXPR. It's not clear if the overhead of bookkeeping
628 would be recovered by trying to simplify fewer statements.
630 If we are able to simplify a statement into the form
631 SSA_NAME = (SSA_NAME | gimple invariant), then we can record
632 a context sensitive equivalency which may help us simplify
633 later statements in E->dest.
635 Failure to simplify into the form above merely means that the
636 statement provides no equivalences to help simplify later
637 statements. This does not prevent threading through E->dest. */
638 for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
640 tree cached_lhs;
642 stmt = bsi_stmt (bsi);
644 /* Ignore empty statements and labels. */
645 if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
646 continue;
648 /* Safely handle threading across loop backedges. This is
649 over conservative, but still allows us to capture the
650 majority of the cases where we can thread across a loop
651 backedge. */
652 if ((e->flags & EDGE_DFS_BACK) != 0
653 && TREE_CODE (stmt) != COND_EXPR
654 && TREE_CODE (stmt) != SWITCH_EXPR)
655 return;
657 /* If the statement has volatile operands, then we assume we
658 can not thread through this block. This is overly
659 conservative in some ways. */
660 if (TREE_CODE (stmt) == ASM_EXPR && ASM_VOLATILE_P (stmt))
661 return;
663 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
664 value, then do not try to simplify this statement as it will
665 not simplify in any way that is helpful for jump threading. */
666 if (TREE_CODE (stmt) != MODIFY_EXPR
667 || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
668 continue;
670 /* At this point we have a statement which assigns an RHS to an
671 SSA_VAR on the LHS. We want to try and simplify this statement
672 to expose more context sensitive equivalences which in turn may
673 allow us to simplify the condition at the end of the loop. */
674 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
675 cached_lhs = TREE_OPERAND (stmt, 1);
676 else
678 /* Copy the operands. */
679 tree *copy;
680 ssa_op_iter iter;
681 use_operand_p use_p;
682 unsigned int num, i = 0;
684 num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE));
685 copy = xcalloc (num, sizeof (tree));
687 /* Make a copy of the uses & vuses into USES_COPY, then cprop into
688 the operands. */
689 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
691 tree tmp = NULL;
692 tree use = USE_FROM_PTR (use_p);
694 copy[i++] = use;
695 if (TREE_CODE (use) == SSA_NAME)
696 tmp = SSA_NAME_VALUE (use);
697 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
698 SET_USE (use_p, tmp);
701 /* Try to fold/lookup the new expression. Inserting the
702 expression into the hash table is unlikely to help
703 simplify anything later, so just query the hashtable. */
704 cached_lhs = fold (TREE_OPERAND (stmt, 1));
705 if (TREE_CODE (cached_lhs) != SSA_NAME
706 && !is_gimple_min_invariant (cached_lhs))
707 cached_lhs = lookup_avail_expr (stmt, false);
710 /* Restore the statement's original uses/defs. */
711 i = 0;
712 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
713 SET_USE (use_p, copy[i++]);
715 free (copy);
718 /* Record the context sensitive equivalence if we were able
719 to simplify this statement. */
720 if (cached_lhs
721 && (TREE_CODE (cached_lhs) == SSA_NAME
722 || is_gimple_min_invariant (cached_lhs)))
723 record_const_or_copy (TREE_OPERAND (stmt, 0), cached_lhs);
726 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
727 will be taken. */
728 if (stmt
729 && (TREE_CODE (stmt) == COND_EXPR
730 || TREE_CODE (stmt) == GOTO_EXPR
731 || TREE_CODE (stmt) == SWITCH_EXPR))
733 tree cond, cached_lhs;
735 /* Now temporarily cprop the operands and try to find the resulting
736 expression in the hash tables. */
737 if (TREE_CODE (stmt) == COND_EXPR)
738 cond = COND_EXPR_COND (stmt);
739 else if (TREE_CODE (stmt) == GOTO_EXPR)
740 cond = GOTO_DESTINATION (stmt);
741 else
742 cond = SWITCH_COND (stmt);
744 if (COMPARISON_CLASS_P (cond))
746 tree dummy_cond, op0, op1;
747 enum tree_code cond_code;
749 op0 = TREE_OPERAND (cond, 0);
750 op1 = TREE_OPERAND (cond, 1);
751 cond_code = TREE_CODE (cond);
753 /* Get the current value of both operands. */
754 if (TREE_CODE (op0) == SSA_NAME)
756 tree tmp = SSA_NAME_VALUE (op0);
757 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
758 op0 = tmp;
761 if (TREE_CODE (op1) == SSA_NAME)
763 tree tmp = SSA_NAME_VALUE (op1);
764 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
765 op1 = tmp;
768 /* Stuff the operator and operands into our dummy conditional
769 expression, creating the dummy conditional if necessary. */
770 dummy_cond = walk_data->global_data;
771 if (! dummy_cond)
773 dummy_cond = build (cond_code, boolean_type_node, op0, op1);
774 dummy_cond = build (COND_EXPR, void_type_node,
775 dummy_cond, NULL, NULL);
776 walk_data->global_data = dummy_cond;
778 else
780 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
781 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
782 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
785 /* If the conditional folds to an invariant, then we are done,
786 otherwise look it up in the hash tables. */
787 cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
788 if (! is_gimple_min_invariant (cached_lhs))
790 cached_lhs = lookup_avail_expr (dummy_cond, false);
791 if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
792 cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
793 NULL,
794 false);
797 /* We can have conditionals which just test the state of a
798 variable rather than use a relational operator. These are
799 simpler to handle. */
800 else if (TREE_CODE (cond) == SSA_NAME)
802 cached_lhs = cond;
803 cached_lhs = SSA_NAME_VALUE (cached_lhs);
804 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
805 cached_lhs = NULL;
807 else
808 cached_lhs = lookup_avail_expr (stmt, false);
810 if (cached_lhs)
812 edge taken_edge = find_taken_edge (e->dest, cached_lhs);
813 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
815 if (dest == e->dest)
816 return;
818 /* If we have a known destination for the conditional, then
819 we can perform this optimization, which saves at least one
820 conditional jump each time it applies since we get to
821 bypass the conditional at our original destination. */
822 if (dest)
824 struct edge_info *edge_info;
826 update_bb_profile_for_threading (e->dest, EDGE_FREQUENCY (e),
827 e->count, taken_edge);
828 if (e->aux)
829 edge_info = e->aux;
830 else
831 edge_info = allocate_edge_info (e);
832 edge_info->redirection_target = taken_edge;
833 bb_ann (e->dest)->incoming_edge_threaded = true;
840 /* Initialize local stacks for this optimizer and record equivalences
841 upon entry to BB. Equivalences can come from the edge traversed to
842 reach BB or they may come from PHI nodes at the start of BB. */
844 static void
845 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
846 basic_block bb)
848 if (dump_file && (dump_flags & TDF_DETAILS))
849 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
851 /* Push a marker on the stacks of local information so that we know how
852 far to unwind when we finalize this block. */
853 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
854 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
855 VEC_safe_push (tree, heap, nonzero_vars_stack, NULL_TREE);
856 VEC_safe_push (tree, heap, vrp_variables_stack, NULL_TREE);
858 record_equivalences_from_incoming_edge (bb);
860 /* PHI nodes can create equivalences too. */
861 record_equivalences_from_phis (bb);
864 /* Given an expression EXPR (a relational expression or a statement),
865 initialize the hash table element pointed by by ELEMENT. */
867 static void
868 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
870 /* Hash table elements may be based on conditional expressions or statements.
872 For the former case, we have no annotation and we want to hash the
873 conditional expression. In the latter case we have an annotation and
874 we want to record the expression the statement evaluates. */
875 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
877 element->stmt = NULL;
878 element->rhs = expr;
880 else if (TREE_CODE (expr) == COND_EXPR)
882 element->stmt = expr;
883 element->rhs = COND_EXPR_COND (expr);
885 else if (TREE_CODE (expr) == SWITCH_EXPR)
887 element->stmt = expr;
888 element->rhs = SWITCH_COND (expr);
890 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
892 element->stmt = expr;
893 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
895 else if (TREE_CODE (expr) == GOTO_EXPR)
897 element->stmt = expr;
898 element->rhs = GOTO_DESTINATION (expr);
900 else
902 element->stmt = expr;
903 element->rhs = TREE_OPERAND (expr, 1);
906 element->lhs = lhs;
907 element->hash = avail_expr_hash (element);
910 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
911 LIMIT entries left in LOCALs. */
913 static void
914 remove_local_expressions_from_table (void)
916 /* Remove all the expressions made available in this block. */
917 while (VEC_length (tree, avail_exprs_stack) > 0)
919 struct expr_hash_elt element;
920 tree expr = VEC_pop (tree, avail_exprs_stack);
922 if (expr == NULL_TREE)
923 break;
925 initialize_hash_element (expr, NULL, &element);
926 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
930 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
931 state, stopping when there are LIMIT entries left in LOCALs. */
933 static void
934 restore_nonzero_vars_to_original_value (void)
936 while (VEC_length (tree, nonzero_vars_stack) > 0)
938 tree name = VEC_pop (tree, nonzero_vars_stack);
940 if (name == NULL)
941 break;
943 bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
947 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
948 CONST_AND_COPIES to its original state, stopping when we hit a
949 NULL marker. */
951 static void
952 restore_vars_to_original_value (void)
954 while (VEC_length (tree, const_and_copies_stack) > 0)
956 tree prev_value, dest;
958 dest = VEC_pop (tree, const_and_copies_stack);
960 if (dest == NULL)
961 break;
963 prev_value = VEC_pop (tree, const_and_copies_stack);
964 SSA_NAME_VALUE (dest) = prev_value;
968 /* We have finished processing the dominator children of BB, perform
969 any finalization actions in preparation for leaving this node in
970 the dominator tree. */
972 static void
973 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
975 tree last;
977 /* If we are at a leaf node in the dominator tree, see if we can thread
978 the edge from BB through its successor.
980 Do this before we remove entries from our equivalence tables. */
981 if (single_succ_p (bb)
982 && (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0
983 && (get_immediate_dominator (CDI_DOMINATORS, single_succ (bb)) != bb
984 || phi_nodes (single_succ (bb))))
987 thread_across_edge (walk_data, single_succ_edge (bb));
989 else if ((last = last_stmt (bb))
990 && TREE_CODE (last) == COND_EXPR
991 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
992 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
993 && EDGE_COUNT (bb->succs) == 2
994 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
995 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
997 edge true_edge, false_edge;
999 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
1001 /* If the THEN arm is the end of a dominator tree or has PHI nodes,
1002 then try to thread through its edge. */
1003 if (get_immediate_dominator (CDI_DOMINATORS, true_edge->dest) != bb
1004 || phi_nodes (true_edge->dest))
1006 struct edge_info *edge_info;
1007 unsigned int i;
1009 /* Push a marker onto the available expression stack so that we
1010 unwind any expressions related to the TRUE arm before processing
1011 the false arm below. */
1012 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
1013 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
1015 edge_info = true_edge->aux;
1017 /* If we have info associated with this edge, record it into
1018 our equivalency tables. */
1019 if (edge_info)
1021 tree *cond_equivalences = edge_info->cond_equivalences;
1022 tree lhs = edge_info->lhs;
1023 tree rhs = edge_info->rhs;
1025 /* If we have a simple NAME = VALUE equivalency record it. */
1026 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1027 record_const_or_copy (lhs, rhs);
1029 /* If we have 0 = COND or 1 = COND equivalences, record them
1030 into our expression hash tables. */
1031 if (cond_equivalences)
1032 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1034 tree expr = cond_equivalences[i];
1035 tree value = cond_equivalences[i + 1];
1037 record_cond (expr, value);
1041 /* Now thread the edge. */
1042 thread_across_edge (walk_data, true_edge);
1044 /* And restore the various tables to their state before
1045 we threaded this edge. */
1046 remove_local_expressions_from_table ();
1047 restore_vars_to_original_value ();
1050 /* Similarly for the ELSE arm. */
1051 if (get_immediate_dominator (CDI_DOMINATORS, false_edge->dest) != bb
1052 || phi_nodes (false_edge->dest))
1054 struct edge_info *edge_info;
1055 unsigned int i;
1057 edge_info = false_edge->aux;
1059 /* If we have info associated with this edge, record it into
1060 our equivalency tables. */
1061 if (edge_info)
1063 tree *cond_equivalences = edge_info->cond_equivalences;
1064 tree lhs = edge_info->lhs;
1065 tree rhs = edge_info->rhs;
1067 /* If we have a simple NAME = VALUE equivalency record it. */
1068 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1069 record_const_or_copy (lhs, rhs);
1071 /* If we have 0 = COND or 1 = COND equivalences, record them
1072 into our expression hash tables. */
1073 if (cond_equivalences)
1074 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1076 tree expr = cond_equivalences[i];
1077 tree value = cond_equivalences[i + 1];
1079 record_cond (expr, value);
1083 thread_across_edge (walk_data, false_edge);
1085 /* No need to remove local expressions from our tables
1086 or restore vars to their original value as that will
1087 be done immediately below. */
1091 remove_local_expressions_from_table ();
1092 restore_nonzero_vars_to_original_value ();
1093 restore_vars_to_original_value ();
1095 /* Remove VRP records associated with this basic block. They are no
1096 longer valid.
1098 To be efficient, we note which variables have had their values
1099 constrained in this block. So walk over each variable in the
1100 VRP_VARIABLEs array. */
1101 while (VEC_length (tree, vrp_variables_stack) > 0)
1103 tree var = VEC_pop (tree, vrp_variables_stack);
1104 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1105 void **slot;
1107 /* Each variable has a stack of value range records. We want to
1108 invalidate those associated with our basic block. So we walk
1109 the array backwards popping off records associated with our
1110 block. Once we hit a record not associated with our block
1111 we are done. */
1112 varray_type var_vrp_records;
1114 if (var == NULL)
1115 break;
1117 vrp_hash_elt.var = var;
1118 vrp_hash_elt.records = NULL;
1120 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1122 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1123 var_vrp_records = vrp_hash_elt_p->records;
1125 while (VARRAY_ACTIVE_SIZE (var_vrp_records) > 0)
1127 struct vrp_element *element
1128 = (struct vrp_element *)VARRAY_TOP_GENERIC_PTR (var_vrp_records);
1130 if (element->bb != bb)
1131 break;
1133 VARRAY_POP (var_vrp_records);
1137 /* If we queued any statements to rescan in this block, then
1138 go ahead and rescan them now. */
1139 while (VEC_length (tree, stmts_to_rescan) > 0)
1141 tree stmt = VEC_last (tree, stmts_to_rescan);
1142 basic_block stmt_bb = bb_for_stmt (stmt);
1144 if (stmt_bb != bb)
1145 break;
1147 VEC_pop (tree, stmts_to_rescan);
1148 mark_new_vars_to_rename (stmt);
1152 /* PHI nodes can create equivalences too.
1154 Ignoring any alternatives which are the same as the result, if
1155 all the alternatives are equal, then the PHI node creates an
1156 equivalence.
1158 Additionally, if all the PHI alternatives are known to have a nonzero
1159 value, then the result of this PHI is known to have a nonzero value,
1160 even if we do not know its exact value. */
1162 static void
1163 record_equivalences_from_phis (basic_block bb)
1165 tree phi;
1167 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1169 tree lhs = PHI_RESULT (phi);
1170 tree rhs = NULL;
1171 int i;
1173 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1175 tree t = PHI_ARG_DEF (phi, i);
1177 /* Ignore alternatives which are the same as our LHS. Since
1178 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1179 can simply compare pointers. */
1180 if (lhs == t)
1181 continue;
1183 /* If we have not processed an alternative yet, then set
1184 RHS to this alternative. */
1185 if (rhs == NULL)
1186 rhs = t;
1187 /* If we have processed an alternative (stored in RHS), then
1188 see if it is equal to this one. If it isn't, then stop
1189 the search. */
1190 else if (! operand_equal_for_phi_arg_p (rhs, t))
1191 break;
1194 /* If we had no interesting alternatives, then all the RHS alternatives
1195 must have been the same as LHS. */
1196 if (!rhs)
1197 rhs = lhs;
1199 /* If we managed to iterate through each PHI alternative without
1200 breaking out of the loop, then we have a PHI which may create
1201 a useful equivalence. We do not need to record unwind data for
1202 this, since this is a true assignment and not an equivalence
1203 inferred from a comparison. All uses of this ssa name are dominated
1204 by this assignment, so unwinding just costs time and space. */
1205 if (i == PHI_NUM_ARGS (phi)
1206 && may_propagate_copy (lhs, rhs))
1207 SSA_NAME_VALUE (lhs) = rhs;
1209 /* Now see if we know anything about the nonzero property for the
1210 result of this PHI. */
1211 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1213 if (!PHI_ARG_NONZERO (phi, i))
1214 break;
1217 if (i == PHI_NUM_ARGS (phi))
1218 bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
1222 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1223 return that edge. Otherwise return NULL. */
1224 static edge
1225 single_incoming_edge_ignoring_loop_edges (basic_block bb)
1227 edge retval = NULL;
1228 edge e;
1229 edge_iterator ei;
1231 FOR_EACH_EDGE (e, ei, bb->preds)
1233 /* A loop back edge can be identified by the destination of
1234 the edge dominating the source of the edge. */
1235 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
1236 continue;
1238 /* If we have already seen a non-loop edge, then we must have
1239 multiple incoming non-loop edges and thus we return NULL. */
1240 if (retval)
1241 return NULL;
1243 /* This is the first non-loop incoming edge we have found. Record
1244 it. */
1245 retval = e;
1248 return retval;
1251 /* Record any equivalences created by the incoming edge to BB. If BB
1252 has more than one incoming edge, then no equivalence is created. */
1254 static void
1255 record_equivalences_from_incoming_edge (basic_block bb)
1257 edge e;
1258 basic_block parent;
1259 struct edge_info *edge_info;
1261 /* If our parent block ended with a control statement, then we may be
1262 able to record some equivalences based on which outgoing edge from
1263 the parent was followed. */
1264 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
1266 e = single_incoming_edge_ignoring_loop_edges (bb);
1268 /* If we had a single incoming edge from our parent block, then enter
1269 any data associated with the edge into our tables. */
1270 if (e && e->src == parent)
1272 unsigned int i;
1274 edge_info = e->aux;
1276 if (edge_info)
1278 tree lhs = edge_info->lhs;
1279 tree rhs = edge_info->rhs;
1280 tree *cond_equivalences = edge_info->cond_equivalences;
1282 if (lhs)
1283 record_equality (lhs, rhs);
1285 if (cond_equivalences)
1287 bool recorded_range = false;
1288 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1290 tree expr = cond_equivalences[i];
1291 tree value = cond_equivalences[i + 1];
1293 record_cond (expr, value);
1295 /* For the first true equivalence, record range
1296 information. We only do this for the first
1297 true equivalence as it should dominate any
1298 later true equivalences. */
1299 if (! recorded_range
1300 && COMPARISON_CLASS_P (expr)
1301 && value == boolean_true_node
1302 && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
1304 record_range (expr, bb);
1305 recorded_range = true;
1313 /* Dump SSA statistics on FILE. */
1315 void
1316 dump_dominator_optimization_stats (FILE *file)
1318 long n_exprs;
1320 fprintf (file, "Total number of statements: %6ld\n\n",
1321 opt_stats.num_stmts);
1322 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
1323 opt_stats.num_exprs_considered);
1325 n_exprs = opt_stats.num_exprs_considered;
1326 if (n_exprs == 0)
1327 n_exprs = 1;
1329 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1330 opt_stats.num_re, PERCENT (opt_stats.num_re,
1331 n_exprs));
1332 fprintf (file, " Constants propagated: %6ld\n",
1333 opt_stats.num_const_prop);
1334 fprintf (file, " Copies propagated: %6ld\n",
1335 opt_stats.num_copy_prop);
1337 fprintf (file, "\nHash table statistics:\n");
1339 fprintf (file, " avail_exprs: ");
1340 htab_statistics (file, avail_exprs);
1344 /* Dump SSA statistics on stderr. */
1346 void
1347 debug_dominator_optimization_stats (void)
1349 dump_dominator_optimization_stats (stderr);
1353 /* Dump statistics for the hash table HTAB. */
1355 static void
1356 htab_statistics (FILE *file, htab_t htab)
1358 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
1359 (long) htab_size (htab),
1360 (long) htab_elements (htab),
1361 htab_collisions (htab));
1364 /* Record the fact that VAR has a nonzero value, though we may not know
1365 its exact value. Note that if VAR is already known to have a nonzero
1366 value, then we do nothing. */
1368 static void
1369 record_var_is_nonzero (tree var)
1371 int indx = SSA_NAME_VERSION (var);
1373 if (bitmap_bit_p (nonzero_vars, indx))
1374 return;
1376 /* Mark it in the global table. */
1377 bitmap_set_bit (nonzero_vars, indx);
1379 /* Record this SSA_NAME so that we can reset the global table
1380 when we leave this block. */
1381 VEC_safe_push (tree, heap, nonzero_vars_stack, var);
1384 /* Enter a statement into the true/false expression hash table indicating
1385 that the condition COND has the value VALUE. */
1387 static void
1388 record_cond (tree cond, tree value)
1390 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
1391 void **slot;
1393 initialize_hash_element (cond, value, element);
1395 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
1396 element->hash, INSERT);
1397 if (*slot == NULL)
1399 *slot = (void *) element;
1400 VEC_safe_push (tree, heap, avail_exprs_stack, cond);
1402 else
1403 free (element);
1406 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1407 the new conditional into *p, then store a boolean_true_node
1408 into *(p + 1). */
1410 static void
1411 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
1413 *p = build2 (new_code, boolean_type_node, op0, op1);
1414 p++;
1415 *p = boolean_true_node;
1418 /* Record that COND is true and INVERTED is false into the edge information
1419 structure. Also record that any conditions dominated by COND are true
1420 as well.
1422 For example, if a < b is true, then a <= b must also be true. */
1424 static void
1425 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
1427 tree op0, op1;
1429 if (!COMPARISON_CLASS_P (cond))
1430 return;
1432 op0 = TREE_OPERAND (cond, 0);
1433 op1 = TREE_OPERAND (cond, 1);
1435 switch (TREE_CODE (cond))
1437 case LT_EXPR:
1438 case GT_EXPR:
1439 edge_info->max_cond_equivalences = 12;
1440 edge_info->cond_equivalences = xmalloc (12 * sizeof (tree));
1441 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
1442 ? LE_EXPR : GE_EXPR),
1443 op0, op1, &edge_info->cond_equivalences[4]);
1444 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1445 &edge_info->cond_equivalences[6]);
1446 build_and_record_new_cond (NE_EXPR, op0, op1,
1447 &edge_info->cond_equivalences[8]);
1448 build_and_record_new_cond (LTGT_EXPR, op0, op1,
1449 &edge_info->cond_equivalences[10]);
1450 break;
1452 case GE_EXPR:
1453 case LE_EXPR:
1454 edge_info->max_cond_equivalences = 6;
1455 edge_info->cond_equivalences = xmalloc (6 * sizeof (tree));
1456 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1457 &edge_info->cond_equivalences[4]);
1458 break;
1460 case EQ_EXPR:
1461 edge_info->max_cond_equivalences = 10;
1462 edge_info->cond_equivalences = xmalloc (10 * sizeof (tree));
1463 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1464 &edge_info->cond_equivalences[4]);
1465 build_and_record_new_cond (LE_EXPR, op0, op1,
1466 &edge_info->cond_equivalences[6]);
1467 build_and_record_new_cond (GE_EXPR, op0, op1,
1468 &edge_info->cond_equivalences[8]);
1469 break;
1471 case UNORDERED_EXPR:
1472 edge_info->max_cond_equivalences = 16;
1473 edge_info->cond_equivalences = xmalloc (16 * sizeof (tree));
1474 build_and_record_new_cond (NE_EXPR, op0, op1,
1475 &edge_info->cond_equivalences[4]);
1476 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1477 &edge_info->cond_equivalences[6]);
1478 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1479 &edge_info->cond_equivalences[8]);
1480 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
1481 &edge_info->cond_equivalences[10]);
1482 build_and_record_new_cond (UNLT_EXPR, op0, op1,
1483 &edge_info->cond_equivalences[12]);
1484 build_and_record_new_cond (UNGT_EXPR, op0, op1,
1485 &edge_info->cond_equivalences[14]);
1486 break;
1488 case UNLT_EXPR:
1489 case UNGT_EXPR:
1490 edge_info->max_cond_equivalences = 8;
1491 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1492 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1493 ? UNLE_EXPR : UNGE_EXPR),
1494 op0, op1, &edge_info->cond_equivalences[4]);
1495 build_and_record_new_cond (NE_EXPR, op0, op1,
1496 &edge_info->cond_equivalences[6]);
1497 break;
1499 case UNEQ_EXPR:
1500 edge_info->max_cond_equivalences = 8;
1501 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1502 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1503 &edge_info->cond_equivalences[4]);
1504 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1505 &edge_info->cond_equivalences[6]);
1506 break;
1508 case LTGT_EXPR:
1509 edge_info->max_cond_equivalences = 8;
1510 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1511 build_and_record_new_cond (NE_EXPR, op0, op1,
1512 &edge_info->cond_equivalences[4]);
1513 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1514 &edge_info->cond_equivalences[6]);
1515 break;
1517 default:
1518 edge_info->max_cond_equivalences = 4;
1519 edge_info->cond_equivalences = xmalloc (4 * sizeof (tree));
1520 break;
1523 /* Now store the original true and false conditions into the first
1524 two slots. */
1525 edge_info->cond_equivalences[0] = cond;
1526 edge_info->cond_equivalences[1] = boolean_true_node;
1527 edge_info->cond_equivalences[2] = inverted;
1528 edge_info->cond_equivalences[3] = boolean_false_node;
1531 /* A helper function for record_const_or_copy and record_equality.
1532 Do the work of recording the value and undo info. */
1534 static void
1535 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1537 SSA_NAME_VALUE (x) = y;
1539 VEC_reserve (tree, heap, const_and_copies_stack, 2);
1540 VEC_quick_push (tree, const_and_copies_stack, prev_x);
1541 VEC_quick_push (tree, const_and_copies_stack, x);
1545 /* Return the loop depth of the basic block of the defining statement of X.
1546 This number should not be treated as absolutely correct because the loop
1547 information may not be completely up-to-date when dom runs. However, it
1548 will be relatively correct, and as more passes are taught to keep loop info
1549 up to date, the result will become more and more accurate. */
1552 loop_depth_of_name (tree x)
1554 tree defstmt;
1555 basic_block defbb;
1557 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1558 if (TREE_CODE (x) != SSA_NAME)
1559 return 0;
1561 /* Otherwise return the loop depth of the defining statement's bb.
1562 Note that there may not actually be a bb for this statement, if the
1563 ssa_name is live on entry. */
1564 defstmt = SSA_NAME_DEF_STMT (x);
1565 defbb = bb_for_stmt (defstmt);
1566 if (!defbb)
1567 return 0;
1569 return defbb->loop_depth;
1573 /* Record that X is equal to Y in const_and_copies. Record undo
1574 information in the block-local vector. */
1576 static void
1577 record_const_or_copy (tree x, tree y)
1579 tree prev_x = SSA_NAME_VALUE (x);
1581 if (TREE_CODE (y) == SSA_NAME)
1583 tree tmp = SSA_NAME_VALUE (y);
1584 if (tmp)
1585 y = tmp;
1588 record_const_or_copy_1 (x, y, prev_x);
1591 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1592 This constrains the cases in which we may treat this as assignment. */
1594 static void
1595 record_equality (tree x, tree y)
1597 tree prev_x = NULL, prev_y = NULL;
1599 if (TREE_CODE (x) == SSA_NAME)
1600 prev_x = SSA_NAME_VALUE (x);
1601 if (TREE_CODE (y) == SSA_NAME)
1602 prev_y = SSA_NAME_VALUE (y);
1604 /* If one of the previous values is invariant, or invariant in more loops
1605 (by depth), then use that.
1606 Otherwise it doesn't matter which value we choose, just so
1607 long as we canonicalize on one value. */
1608 if (TREE_INVARIANT (y))
1610 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1611 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1612 else if (prev_x && TREE_INVARIANT (prev_x))
1613 x = y, y = prev_x, prev_x = prev_y;
1614 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1615 y = prev_y;
1617 /* After the swapping, we must have one SSA_NAME. */
1618 if (TREE_CODE (x) != SSA_NAME)
1619 return;
1621 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1622 variable compared against zero. If we're honoring signed zeros,
1623 then we cannot record this value unless we know that the value is
1624 nonzero. */
1625 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1626 && (TREE_CODE (y) != REAL_CST
1627 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1628 return;
1630 record_const_or_copy_1 (x, y, prev_x);
1633 /* Return true, if it is ok to do folding of an associative expression.
1634 EXP is the tree for the associative expression. */
1636 static inline bool
1637 unsafe_associative_fp_binop (tree exp)
1639 enum tree_code code = TREE_CODE (exp);
1640 return !(!flag_unsafe_math_optimizations
1641 && (code == MULT_EXPR || code == PLUS_EXPR
1642 || code == MINUS_EXPR)
1643 && FLOAT_TYPE_P (TREE_TYPE (exp)));
1646 /* Returns true when STMT is a simple iv increment. It detects the
1647 following situation:
1649 i_1 = phi (..., i_2)
1650 i_2 = i_1 +/- ... */
1652 static bool
1653 simple_iv_increment_p (tree stmt)
1655 tree lhs, rhs, preinc, phi;
1656 unsigned i;
1658 if (TREE_CODE (stmt) != MODIFY_EXPR)
1659 return false;
1661 lhs = TREE_OPERAND (stmt, 0);
1662 if (TREE_CODE (lhs) != SSA_NAME)
1663 return false;
1665 rhs = TREE_OPERAND (stmt, 1);
1667 if (TREE_CODE (rhs) != PLUS_EXPR
1668 && TREE_CODE (rhs) != MINUS_EXPR)
1669 return false;
1671 preinc = TREE_OPERAND (rhs, 0);
1672 if (TREE_CODE (preinc) != SSA_NAME)
1673 return false;
1675 phi = SSA_NAME_DEF_STMT (preinc);
1676 if (TREE_CODE (phi) != PHI_NODE)
1677 return false;
1679 for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++)
1680 if (PHI_ARG_DEF (phi, i) == lhs)
1681 return true;
1683 return false;
1686 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1687 hash tables. Try to simplify the RHS using whatever equivalences
1688 we may have recorded.
1690 If we are able to simplify the RHS, then lookup the simplified form in
1691 the hash table and return the result. Otherwise return NULL. */
1693 static tree
1694 simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *walk_data,
1695 tree stmt, int insert)
1697 tree rhs = TREE_OPERAND (stmt, 1);
1698 enum tree_code rhs_code = TREE_CODE (rhs);
1699 tree result = NULL;
1701 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1702 In which case we can change this statement to be lhs = y.
1703 Which can then be copy propagated.
1705 Similarly for negation. */
1706 if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR)
1707 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME)
1709 /* Get the definition statement for our RHS. */
1710 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1712 /* See if the RHS_DEF_STMT has the same form as our statement. */
1713 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR
1714 && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code)
1716 tree rhs_def_operand;
1718 rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0);
1720 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1721 if (TREE_CODE (rhs_def_operand) == SSA_NAME
1722 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand))
1723 result = update_rhs_and_lookup_avail_expr (stmt,
1724 rhs_def_operand,
1725 insert);
1729 /* If we have z = (x OP C1), see if we earlier had x = y OP C2.
1730 If OP is associative, create and fold (y OP C2) OP C1 which
1731 should result in (y OP C3), use that as the RHS for the
1732 assignment. Add minus to this, as we handle it specially below. */
1733 if ((associative_tree_code (rhs_code) || rhs_code == MINUS_EXPR)
1734 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME
1735 && is_gimple_min_invariant (TREE_OPERAND (rhs, 1)))
1737 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1739 /* If the statement defines an induction variable, do not propagate
1740 its value, so that we do not create overlapping life ranges. */
1741 if (simple_iv_increment_p (rhs_def_stmt))
1742 goto dont_fold_assoc;
1744 /* See if the RHS_DEF_STMT has the same form as our statement. */
1745 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR)
1747 tree rhs_def_rhs = TREE_OPERAND (rhs_def_stmt, 1);
1748 enum tree_code rhs_def_code = TREE_CODE (rhs_def_rhs);
1750 if ((rhs_code == rhs_def_code && unsafe_associative_fp_binop (rhs))
1751 || (rhs_code == PLUS_EXPR && rhs_def_code == MINUS_EXPR)
1752 || (rhs_code == MINUS_EXPR && rhs_def_code == PLUS_EXPR))
1754 tree def_stmt_op0 = TREE_OPERAND (rhs_def_rhs, 0);
1755 tree def_stmt_op1 = TREE_OPERAND (rhs_def_rhs, 1);
1757 if (TREE_CODE (def_stmt_op0) == SSA_NAME
1758 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0)
1759 && is_gimple_min_invariant (def_stmt_op1))
1761 tree outer_const = TREE_OPERAND (rhs, 1);
1762 tree type = TREE_TYPE (TREE_OPERAND (stmt, 0));
1763 tree t;
1765 /* If we care about correct floating point results, then
1766 don't fold x + c1 - c2. Note that we need to take both
1767 the codes and the signs to figure this out. */
1768 if (FLOAT_TYPE_P (type)
1769 && !flag_unsafe_math_optimizations
1770 && (rhs_def_code == PLUS_EXPR
1771 || rhs_def_code == MINUS_EXPR))
1773 bool neg = false;
1775 neg ^= (rhs_code == MINUS_EXPR);
1776 neg ^= (rhs_def_code == MINUS_EXPR);
1777 neg ^= real_isneg (TREE_REAL_CST_PTR (outer_const));
1778 neg ^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1));
1780 if (neg)
1781 goto dont_fold_assoc;
1784 /* Ho hum. So fold will only operate on the outermost
1785 thingy that we give it, so we have to build the new
1786 expression in two pieces. This requires that we handle
1787 combinations of plus and minus. */
1788 if (rhs_def_code != rhs_code)
1790 if (rhs_def_code == MINUS_EXPR)
1791 t = build (MINUS_EXPR, type, outer_const, def_stmt_op1);
1792 else
1793 t = build (MINUS_EXPR, type, def_stmt_op1, outer_const);
1794 rhs_code = PLUS_EXPR;
1796 else if (rhs_def_code == MINUS_EXPR)
1797 t = build (PLUS_EXPR, type, def_stmt_op1, outer_const);
1798 else
1799 t = build (rhs_def_code, type, def_stmt_op1, outer_const);
1800 t = local_fold (t);
1801 t = build (rhs_code, type, def_stmt_op0, t);
1802 t = local_fold (t);
1804 /* If the result is a suitable looking gimple expression,
1805 then use it instead of the original for STMT. */
1806 if (TREE_CODE (t) == SSA_NAME
1807 || (UNARY_CLASS_P (t)
1808 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME)
1809 || ((BINARY_CLASS_P (t) || COMPARISON_CLASS_P (t))
1810 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
1811 && is_gimple_val (TREE_OPERAND (t, 1))))
1812 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1816 dont_fold_assoc:;
1819 /* Transform TRUNC_DIV_EXPR and TRUNC_MOD_EXPR into RSHIFT_EXPR
1820 and BIT_AND_EXPR respectively if the first operand is greater
1821 than zero and the second operand is an exact power of two. */
1822 if ((rhs_code == TRUNC_DIV_EXPR || rhs_code == TRUNC_MOD_EXPR)
1823 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0)))
1824 && integer_pow2p (TREE_OPERAND (rhs, 1)))
1826 tree val;
1827 tree op = TREE_OPERAND (rhs, 0);
1829 if (TYPE_UNSIGNED (TREE_TYPE (op)))
1831 val = integer_one_node;
1833 else
1835 tree dummy_cond = walk_data->global_data;
1837 if (! dummy_cond)
1839 dummy_cond = build (GT_EXPR, boolean_type_node,
1840 op, integer_zero_node);
1841 dummy_cond = build (COND_EXPR, void_type_node,
1842 dummy_cond, NULL, NULL);
1843 walk_data->global_data = dummy_cond;
1845 else
1847 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GT_EXPR);
1848 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1849 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1850 = integer_zero_node;
1852 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1855 if (val && integer_onep (val))
1857 tree t;
1858 tree op0 = TREE_OPERAND (rhs, 0);
1859 tree op1 = TREE_OPERAND (rhs, 1);
1861 if (rhs_code == TRUNC_DIV_EXPR)
1862 t = build (RSHIFT_EXPR, TREE_TYPE (op0), op0,
1863 build_int_cst (NULL_TREE, tree_log2 (op1)));
1864 else
1865 t = build (BIT_AND_EXPR, TREE_TYPE (op0), op0,
1866 local_fold (build (MINUS_EXPR, TREE_TYPE (op1),
1867 op1, integer_one_node)));
1869 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1873 /* Transform ABS (X) into X or -X as appropriate. */
1874 if (rhs_code == ABS_EXPR
1875 && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0))))
1877 tree val;
1878 tree op = TREE_OPERAND (rhs, 0);
1879 tree type = TREE_TYPE (op);
1881 if (TYPE_UNSIGNED (type))
1883 val = integer_zero_node;
1885 else
1887 tree dummy_cond = walk_data->global_data;
1889 if (! dummy_cond)
1891 dummy_cond = build (LE_EXPR, boolean_type_node,
1892 op, integer_zero_node);
1893 dummy_cond = build (COND_EXPR, void_type_node,
1894 dummy_cond, NULL, NULL);
1895 walk_data->global_data = dummy_cond;
1897 else
1899 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), LE_EXPR);
1900 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1901 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1902 = build_int_cst (type, 0);
1904 val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false);
1906 if (!val)
1908 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), GE_EXPR);
1909 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op;
1910 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1)
1911 = build_int_cst (type, 0);
1913 val = simplify_cond_and_lookup_avail_expr (dummy_cond,
1914 NULL, false);
1916 if (val)
1918 if (integer_zerop (val))
1919 val = integer_one_node;
1920 else if (integer_onep (val))
1921 val = integer_zero_node;
1926 if (val
1927 && (integer_onep (val) || integer_zerop (val)))
1929 tree t;
1931 if (integer_onep (val))
1932 t = build1 (NEGATE_EXPR, TREE_TYPE (op), op);
1933 else
1934 t = op;
1936 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1940 /* Optimize *"foo" into 'f'. This is done here rather than
1941 in fold to avoid problems with stuff like &*"foo". */
1942 if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF)
1944 tree t = fold_read_from_constant_string (rhs);
1946 if (t)
1947 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1950 return result;
1953 /* COND is a condition of the form:
1955 x == const or x != const
1957 Look back to x's defining statement and see if x is defined as
1959 x = (type) y;
1961 If const is unchanged if we convert it to type, then we can build
1962 the equivalent expression:
1965 y == const or y != const
1967 Which may allow further optimizations.
1969 Return the equivalent comparison or NULL if no such equivalent comparison
1970 was found. */
1972 static tree
1973 find_equivalent_equality_comparison (tree cond)
1975 tree op0 = TREE_OPERAND (cond, 0);
1976 tree op1 = TREE_OPERAND (cond, 1);
1977 tree def_stmt = SSA_NAME_DEF_STMT (op0);
1979 /* OP0 might have been a parameter, so first make sure it
1980 was defined by a MODIFY_EXPR. */
1981 if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
1983 tree def_rhs = TREE_OPERAND (def_stmt, 1);
1985 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1986 if ((TREE_CODE (def_rhs) == NOP_EXPR
1987 || TREE_CODE (def_rhs) == CONVERT_EXPR)
1988 && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
1990 tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
1991 tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
1992 tree new;
1994 if (TYPE_PRECISION (def_rhs_inner_type)
1995 > TYPE_PRECISION (TREE_TYPE (def_rhs)))
1996 return NULL;
1998 /* What we want to prove is that if we convert OP1 to
1999 the type of the object inside the NOP_EXPR that the
2000 result is still equivalent to SRC.
2002 If that is true, the build and return new equivalent
2003 condition which uses the source of the typecast and the
2004 new constant (which has only changed its type). */
2005 new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
2006 new = local_fold (new);
2007 if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
2008 return build (TREE_CODE (cond), TREE_TYPE (cond),
2009 def_rhs_inner, new);
2012 return NULL;
2015 /* STMT is a COND_EXPR for which we could not trivially determine its
2016 result. This routine attempts to find equivalent forms of the
2017 condition which we may be able to optimize better. It also
2018 uses simple value range propagation to optimize conditionals. */
2020 static tree
2021 simplify_cond_and_lookup_avail_expr (tree stmt,
2022 stmt_ann_t ann,
2023 int insert)
2025 tree cond = COND_EXPR_COND (stmt);
2027 if (COMPARISON_CLASS_P (cond))
2029 tree op0 = TREE_OPERAND (cond, 0);
2030 tree op1 = TREE_OPERAND (cond, 1);
2032 if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
2034 int limit;
2035 tree low, high, cond_low, cond_high;
2036 int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
2037 varray_type vrp_records;
2038 struct vrp_element *element;
2039 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
2040 void **slot;
2042 /* First see if we have test of an SSA_NAME against a constant
2043 where the SSA_NAME is defined by an earlier typecast which
2044 is irrelevant when performing tests against the given
2045 constant. */
2046 if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2048 tree new_cond = find_equivalent_equality_comparison (cond);
2050 if (new_cond)
2052 /* Update the statement to use the new equivalent
2053 condition. */
2054 COND_EXPR_COND (stmt) = new_cond;
2056 /* If this is not a real stmt, ann will be NULL and we
2057 avoid processing the operands. */
2058 if (ann)
2059 mark_stmt_modified (stmt);
2061 /* Lookup the condition and return its known value if it
2062 exists. */
2063 new_cond = lookup_avail_expr (stmt, insert);
2064 if (new_cond)
2065 return new_cond;
2067 /* The operands have changed, so update op0 and op1. */
2068 op0 = TREE_OPERAND (cond, 0);
2069 op1 = TREE_OPERAND (cond, 1);
2073 /* Consult the value range records for this variable (if they exist)
2074 to see if we can eliminate or simplify this conditional.
2076 Note two tests are necessary to determine no records exist.
2077 First we have to see if the virtual array exists, if it
2078 exists, then we have to check its active size.
2080 Also note the vast majority of conditionals are not testing
2081 a variable which has had its range constrained by an earlier
2082 conditional. So this filter avoids a lot of unnecessary work. */
2083 vrp_hash_elt.var = op0;
2084 vrp_hash_elt.records = NULL;
2085 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
2086 if (slot == NULL)
2087 return NULL;
2089 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
2090 vrp_records = vrp_hash_elt_p->records;
2091 if (vrp_records == NULL)
2092 return NULL;
2094 limit = VARRAY_ACTIVE_SIZE (vrp_records);
2096 /* If we have no value range records for this variable, or we are
2097 unable to extract a range for this condition, then there is
2098 nothing to do. */
2099 if (limit == 0
2100 || ! extract_range_from_cond (cond, &cond_high,
2101 &cond_low, &cond_inverted))
2102 return NULL;
2104 /* We really want to avoid unnecessary computations of range
2105 info. So all ranges are computed lazily; this avoids a
2106 lot of unnecessary work. i.e., we record the conditional,
2107 but do not process how it constrains the variable's
2108 potential values until we know that processing the condition
2109 could be helpful.
2111 However, we do not want to have to walk a potentially long
2112 list of ranges, nor do we want to compute a variable's
2113 range more than once for a given path.
2115 Luckily, each time we encounter a conditional that can not
2116 be otherwise optimized we will end up here and we will
2117 compute the necessary range information for the variable
2118 used in this condition.
2120 Thus you can conclude that there will never be more than one
2121 conditional associated with a variable which has not been
2122 processed. So we never need to merge more than one new
2123 conditional into the current range.
2125 These properties also help us avoid unnecessary work. */
2126 element
2127 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records, limit - 1);
2129 if (element->high && element->low)
2131 /* The last element has been processed, so there is no range
2132 merging to do, we can simply use the high/low values
2133 recorded in the last element. */
2134 low = element->low;
2135 high = element->high;
2137 else
2139 tree tmp_high, tmp_low;
2140 int dummy;
2142 /* The last element has not been processed. Process it now.
2143 record_range should ensure for cond inverted is not set.
2144 This call can only fail if cond is x < min or x > max,
2145 which fold should have optimized into false.
2146 If that doesn't happen, just pretend all values are
2147 in the range. */
2148 if (! extract_range_from_cond (element->cond, &tmp_high,
2149 &tmp_low, &dummy))
2150 gcc_unreachable ();
2151 else
2152 gcc_assert (dummy == 0);
2154 /* If this is the only element, then no merging is necessary,
2155 the high/low values from extract_range_from_cond are all
2156 we need. */
2157 if (limit == 1)
2159 low = tmp_low;
2160 high = tmp_high;
2162 else
2164 /* Get the high/low value from the previous element. */
2165 struct vrp_element *prev
2166 = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records,
2167 limit - 2);
2168 low = prev->low;
2169 high = prev->high;
2171 /* Merge in this element's range with the range from the
2172 previous element.
2174 The low value for the merged range is the maximum of
2175 the previous low value and the low value of this record.
2177 Similarly the high value for the merged range is the
2178 minimum of the previous high value and the high value of
2179 this record. */
2180 low = (low && tree_int_cst_compare (low, tmp_low) == 1
2181 ? low : tmp_low);
2182 high = (high && tree_int_cst_compare (high, tmp_high) == -1
2183 ? high : tmp_high);
2186 /* And record the computed range. */
2187 element->low = low;
2188 element->high = high;
2192 /* After we have constrained this variable's potential values,
2193 we try to determine the result of the given conditional.
2195 To simplify later tests, first determine if the current
2196 low value is the same low value as the conditional.
2197 Similarly for the current high value and the high value
2198 for the conditional. */
2199 lowequal = tree_int_cst_equal (low, cond_low);
2200 highequal = tree_int_cst_equal (high, cond_high);
2202 if (lowequal && highequal)
2203 return (cond_inverted ? boolean_false_node : boolean_true_node);
2205 /* To simplify the overlap/subset tests below we may want
2206 to swap the two ranges so that the larger of the two
2207 ranges occurs "first". */
2208 swapped = 0;
2209 if (tree_int_cst_compare (low, cond_low) == 1
2210 || (lowequal
2211 && tree_int_cst_compare (cond_high, high) == 1))
2213 tree temp;
2215 swapped = 1;
2216 temp = low;
2217 low = cond_low;
2218 cond_low = temp;
2219 temp = high;
2220 high = cond_high;
2221 cond_high = temp;
2224 /* Now determine if there is no overlap in the ranges
2225 or if the second range is a subset of the first range. */
2226 no_overlap = tree_int_cst_lt (high, cond_low);
2227 subset = tree_int_cst_compare (cond_high, high) != 1;
2229 /* If there was no overlap in the ranges, then this conditional
2230 always has a false value (unless we had to invert this
2231 conditional, in which case it always has a true value). */
2232 if (no_overlap)
2233 return (cond_inverted ? boolean_true_node : boolean_false_node);
2235 /* If the current range is a subset of the condition's range,
2236 then this conditional always has a true value (unless we
2237 had to invert this conditional, in which case it always
2238 has a true value). */
2239 if (subset && swapped)
2240 return (cond_inverted ? boolean_false_node : boolean_true_node);
2242 /* We were unable to determine the result of the conditional.
2243 However, we may be able to simplify the conditional. First
2244 merge the ranges in the same manner as range merging above. */
2245 low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
2246 high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
2248 /* If the range has converged to a single point, then turn this
2249 into an equality comparison. */
2250 if (TREE_CODE (cond) != EQ_EXPR
2251 && TREE_CODE (cond) != NE_EXPR
2252 && tree_int_cst_equal (low, high))
2254 TREE_SET_CODE (cond, EQ_EXPR);
2255 TREE_OPERAND (cond, 1) = high;
2259 return 0;
2262 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2263 result. This routine attempts to find equivalent forms of the
2264 condition which we may be able to optimize better. */
2266 static tree
2267 simplify_switch_and_lookup_avail_expr (tree stmt, int insert)
2269 tree cond = SWITCH_COND (stmt);
2270 tree def, to, ti;
2272 /* The optimization that we really care about is removing unnecessary
2273 casts. That will let us do much better in propagating the inferred
2274 constant at the switch target. */
2275 if (TREE_CODE (cond) == SSA_NAME)
2277 def = SSA_NAME_DEF_STMT (cond);
2278 if (TREE_CODE (def) == MODIFY_EXPR)
2280 def = TREE_OPERAND (def, 1);
2281 if (TREE_CODE (def) == NOP_EXPR)
2283 int need_precision;
2284 bool fail;
2286 def = TREE_OPERAND (def, 0);
2288 #ifdef ENABLE_CHECKING
2289 /* ??? Why was Jeff testing this? We are gimple... */
2290 gcc_assert (is_gimple_val (def));
2291 #endif
2293 to = TREE_TYPE (cond);
2294 ti = TREE_TYPE (def);
2296 /* If we have an extension that preserves value, then we
2297 can copy the source value into the switch. */
2299 need_precision = TYPE_PRECISION (ti);
2300 fail = false;
2301 if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti))
2302 fail = true;
2303 else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti))
2304 need_precision += 1;
2305 if (TYPE_PRECISION (to) < need_precision)
2306 fail = true;
2308 if (!fail)
2310 SWITCH_COND (stmt) = def;
2311 mark_stmt_modified (stmt);
2313 return lookup_avail_expr (stmt, insert);
2319 return 0;
2323 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2324 known value for that SSA_NAME (or NULL if no value is known).
2326 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2327 even if we don't know their precise value.
2329 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2330 nodes of the successors of BB. */
2332 static void
2333 cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
2335 edge e;
2336 edge_iterator ei;
2338 FOR_EACH_EDGE (e, ei, bb->succs)
2340 tree phi;
2341 int indx;
2343 /* If this is an abnormal edge, then we do not want to copy propagate
2344 into the PHI alternative associated with this edge. */
2345 if (e->flags & EDGE_ABNORMAL)
2346 continue;
2348 phi = phi_nodes (e->dest);
2349 if (! phi)
2350 continue;
2352 indx = e->dest_idx;
2353 for ( ; phi; phi = PHI_CHAIN (phi))
2355 tree new;
2356 use_operand_p orig_p;
2357 tree orig;
2359 /* The alternative may be associated with a constant, so verify
2360 it is an SSA_NAME before doing anything with it. */
2361 orig_p = PHI_ARG_DEF_PTR (phi, indx);
2362 orig = USE_FROM_PTR (orig_p);
2363 if (TREE_CODE (orig) != SSA_NAME)
2364 continue;
2366 /* If the alternative is known to have a nonzero value, record
2367 that fact in the PHI node itself for future use. */
2368 if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
2369 PHI_ARG_NONZERO (phi, indx) = true;
2371 /* If we have *ORIG_P in our constant/copy table, then replace
2372 ORIG_P with its value in our constant/copy table. */
2373 new = SSA_NAME_VALUE (orig);
2374 if (new
2375 && new != orig
2376 && (TREE_CODE (new) == SSA_NAME
2377 || is_gimple_min_invariant (new))
2378 && may_propagate_copy (orig, new))
2379 propagate_value (orig_p, new);
2384 /* We have finished optimizing BB, record any information implied by
2385 taking a specific outgoing edge from BB. */
2387 static void
2388 record_edge_info (basic_block bb)
2390 block_stmt_iterator bsi = bsi_last (bb);
2391 struct edge_info *edge_info;
2393 if (! bsi_end_p (bsi))
2395 tree stmt = bsi_stmt (bsi);
2397 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
2399 tree cond = SWITCH_COND (stmt);
2401 if (TREE_CODE (cond) == SSA_NAME)
2403 tree labels = SWITCH_LABELS (stmt);
2404 int i, n_labels = TREE_VEC_LENGTH (labels);
2405 tree *info = xcalloc (n_basic_blocks, sizeof (tree));
2406 edge e;
2407 edge_iterator ei;
2409 for (i = 0; i < n_labels; i++)
2411 tree label = TREE_VEC_ELT (labels, i);
2412 basic_block target_bb = label_to_block (CASE_LABEL (label));
2414 if (CASE_HIGH (label)
2415 || !CASE_LOW (label)
2416 || info[target_bb->index])
2417 info[target_bb->index] = error_mark_node;
2418 else
2419 info[target_bb->index] = label;
2422 FOR_EACH_EDGE (e, ei, bb->succs)
2424 basic_block target_bb = e->dest;
2425 tree node = info[target_bb->index];
2427 if (node != NULL && node != error_mark_node)
2429 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
2430 edge_info = allocate_edge_info (e);
2431 edge_info->lhs = cond;
2432 edge_info->rhs = x;
2435 free (info);
2439 /* A COND_EXPR may create equivalences too. */
2440 if (stmt && TREE_CODE (stmt) == COND_EXPR)
2442 tree cond = COND_EXPR_COND (stmt);
2443 edge true_edge;
2444 edge false_edge;
2446 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
2448 /* If the conditional is a single variable 'X', record 'X = 1'
2449 for the true edge and 'X = 0' on the false edge. */
2450 if (SSA_VAR_P (cond))
2452 struct edge_info *edge_info;
2454 edge_info = allocate_edge_info (true_edge);
2455 edge_info->lhs = cond;
2456 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
2458 edge_info = allocate_edge_info (false_edge);
2459 edge_info->lhs = cond;
2460 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
2462 /* Equality tests may create one or two equivalences. */
2463 else if (COMPARISON_CLASS_P (cond))
2465 tree op0 = TREE_OPERAND (cond, 0);
2466 tree op1 = TREE_OPERAND (cond, 1);
2468 /* Special case comparing booleans against a constant as we
2469 know the value of OP0 on both arms of the branch. i.e., we
2470 can record an equivalence for OP0 rather than COND. */
2471 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2472 && TREE_CODE (op0) == SSA_NAME
2473 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
2474 && is_gimple_min_invariant (op1))
2476 if (TREE_CODE (cond) == EQ_EXPR)
2478 edge_info = allocate_edge_info (true_edge);
2479 edge_info->lhs = op0;
2480 edge_info->rhs = (integer_zerop (op1)
2481 ? boolean_false_node
2482 : boolean_true_node);
2484 edge_info = allocate_edge_info (false_edge);
2485 edge_info->lhs = op0;
2486 edge_info->rhs = (integer_zerop (op1)
2487 ? boolean_true_node
2488 : boolean_false_node);
2490 else
2492 edge_info = allocate_edge_info (true_edge);
2493 edge_info->lhs = op0;
2494 edge_info->rhs = (integer_zerop (op1)
2495 ? boolean_true_node
2496 : boolean_false_node);
2498 edge_info = allocate_edge_info (false_edge);
2499 edge_info->lhs = op0;
2500 edge_info->rhs = (integer_zerop (op1)
2501 ? boolean_false_node
2502 : boolean_true_node);
2506 else if (is_gimple_min_invariant (op0)
2507 && (TREE_CODE (op1) == SSA_NAME
2508 || is_gimple_min_invariant (op1)))
2510 tree inverted = invert_truthvalue (cond);
2511 struct edge_info *edge_info;
2513 edge_info = allocate_edge_info (true_edge);
2514 record_conditions (edge_info, cond, inverted);
2516 if (TREE_CODE (cond) == EQ_EXPR)
2518 edge_info->lhs = op1;
2519 edge_info->rhs = op0;
2522 edge_info = allocate_edge_info (false_edge);
2523 record_conditions (edge_info, inverted, cond);
2525 if (TREE_CODE (cond) == NE_EXPR)
2527 edge_info->lhs = op1;
2528 edge_info->rhs = op0;
2532 else if (TREE_CODE (op0) == SSA_NAME
2533 && (is_gimple_min_invariant (op1)
2534 || TREE_CODE (op1) == SSA_NAME))
2536 tree inverted = invert_truthvalue (cond);
2537 struct edge_info *edge_info;
2539 edge_info = allocate_edge_info (true_edge);
2540 record_conditions (edge_info, cond, inverted);
2542 if (TREE_CODE (cond) == EQ_EXPR)
2544 edge_info->lhs = op0;
2545 edge_info->rhs = op1;
2548 edge_info = allocate_edge_info (false_edge);
2549 record_conditions (edge_info, inverted, cond);
2551 if (TREE_CODE (cond) == NE_EXPR)
2553 edge_info->lhs = op0;
2554 edge_info->rhs = op1;
2559 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2564 /* Propagate information from BB to its outgoing edges.
2566 This can include equivalency information implied by control statements
2567 at the end of BB and const/copy propagation into PHIs in BB's
2568 successor blocks. */
2570 static void
2571 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2572 basic_block bb)
2574 record_edge_info (bb);
2575 cprop_into_successor_phis (bb, nonzero_vars);
2578 /* Search for redundant computations in STMT. If any are found, then
2579 replace them with the variable holding the result of the computation.
2581 If safe, record this expression into the available expression hash
2582 table. */
2584 static bool
2585 eliminate_redundant_computations (struct dom_walk_data *walk_data,
2586 tree stmt, stmt_ann_t ann)
2588 tree *expr_p, def = NULL_TREE;
2589 bool insert = true;
2590 tree cached_lhs;
2591 bool retval = false;
2593 if (TREE_CODE (stmt) == MODIFY_EXPR)
2594 def = TREE_OPERAND (stmt, 0);
2596 /* Certain expressions on the RHS can be optimized away, but can not
2597 themselves be entered into the hash tables. */
2598 if (ann->makes_aliased_stores
2599 || ! def
2600 || TREE_CODE (def) != SSA_NAME
2601 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
2602 || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VMAYDEF)
2603 /* Do not record equivalences for increments of ivs. This would create
2604 overlapping live ranges for a very questionable gain. */
2605 || simple_iv_increment_p (stmt))
2606 insert = false;
2608 /* Check if the expression has been computed before. */
2609 cached_lhs = lookup_avail_expr (stmt, insert);
2611 /* If this is an assignment and the RHS was not in the hash table,
2612 then try to simplify the RHS and lookup the new RHS in the
2613 hash table. */
2614 if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR)
2615 cached_lhs = simplify_rhs_and_lookup_avail_expr (walk_data, stmt, insert);
2616 /* Similarly if this is a COND_EXPR and we did not find its
2617 expression in the hash table, simplify the condition and
2618 try again. */
2619 else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
2620 cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
2621 /* Similarly for a SWITCH_EXPR. */
2622 else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR)
2623 cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert);
2625 opt_stats.num_exprs_considered++;
2627 /* Get a pointer to the expression we are trying to optimize. */
2628 if (TREE_CODE (stmt) == COND_EXPR)
2629 expr_p = &COND_EXPR_COND (stmt);
2630 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2631 expr_p = &SWITCH_COND (stmt);
2632 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
2633 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
2634 else
2635 expr_p = &TREE_OPERAND (stmt, 1);
2637 /* It is safe to ignore types here since we have already done
2638 type checking in the hashing and equality routines. In fact
2639 type checking here merely gets in the way of constant
2640 propagation. Also, make sure that it is safe to propagate
2641 CACHED_LHS into *EXPR_P. */
2642 if (cached_lhs
2643 && (TREE_CODE (cached_lhs) != SSA_NAME
2644 || may_propagate_copy (*expr_p, cached_lhs)))
2646 if (dump_file && (dump_flags & TDF_DETAILS))
2648 fprintf (dump_file, " Replaced redundant expr '");
2649 print_generic_expr (dump_file, *expr_p, dump_flags);
2650 fprintf (dump_file, "' with '");
2651 print_generic_expr (dump_file, cached_lhs, dump_flags);
2652 fprintf (dump_file, "'\n");
2655 opt_stats.num_re++;
2657 #if defined ENABLE_CHECKING
2658 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
2659 || is_gimple_min_invariant (cached_lhs));
2660 #endif
2662 if (TREE_CODE (cached_lhs) == ADDR_EXPR
2663 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
2664 && is_gimple_min_invariant (cached_lhs)))
2665 retval = true;
2667 propagate_tree_value (expr_p, cached_lhs);
2668 mark_stmt_modified (stmt);
2670 return retval;
2673 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2674 the available expressions table or the const_and_copies table.
2675 Detect and record those equivalences. */
2677 static void
2678 record_equivalences_from_stmt (tree stmt,
2679 int may_optimize_p,
2680 stmt_ann_t ann)
2682 tree lhs = TREE_OPERAND (stmt, 0);
2683 enum tree_code lhs_code = TREE_CODE (lhs);
2684 int i;
2686 if (lhs_code == SSA_NAME)
2688 tree rhs = TREE_OPERAND (stmt, 1);
2690 /* Strip away any useless type conversions. */
2691 STRIP_USELESS_TYPE_CONVERSION (rhs);
2693 /* If the RHS of the assignment is a constant or another variable that
2694 may be propagated, register it in the CONST_AND_COPIES table. We
2695 do not need to record unwind data for this, since this is a true
2696 assignment and not an equivalence inferred from a comparison. All
2697 uses of this ssa name are dominated by this assignment, so unwinding
2698 just costs time and space. */
2699 if (may_optimize_p
2700 && (TREE_CODE (rhs) == SSA_NAME
2701 || is_gimple_min_invariant (rhs)))
2702 SSA_NAME_VALUE (lhs) = rhs;
2704 if (expr_computes_nonzero (rhs))
2705 record_var_is_nonzero (lhs);
2708 /* Look at both sides for pointer dereferences. If we find one, then
2709 the pointer must be nonnull and we can enter that equivalence into
2710 the hash tables. */
2711 if (flag_delete_null_pointer_checks)
2712 for (i = 0; i < 2; i++)
2714 tree t = TREE_OPERAND (stmt, i);
2716 /* Strip away any COMPONENT_REFs. */
2717 while (TREE_CODE (t) == COMPONENT_REF)
2718 t = TREE_OPERAND (t, 0);
2720 /* Now see if this is a pointer dereference. */
2721 if (INDIRECT_REF_P (t))
2723 tree op = TREE_OPERAND (t, 0);
2725 /* If the pointer is a SSA variable, then enter new
2726 equivalences into the hash table. */
2727 while (TREE_CODE (op) == SSA_NAME)
2729 tree def = SSA_NAME_DEF_STMT (op);
2731 record_var_is_nonzero (op);
2733 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2734 which are known to have a nonzero value. */
2735 if (def
2736 && TREE_CODE (def) == MODIFY_EXPR
2737 && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
2738 op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
2739 else
2740 break;
2745 /* A memory store, even an aliased store, creates a useful
2746 equivalence. By exchanging the LHS and RHS, creating suitable
2747 vops and recording the result in the available expression table,
2748 we may be able to expose more redundant loads. */
2749 if (!ann->has_volatile_ops
2750 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
2751 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
2752 && !is_gimple_reg (lhs))
2754 tree rhs = TREE_OPERAND (stmt, 1);
2755 tree new;
2757 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2758 is a constant, we need to adjust the constant to fit into the
2759 type of the LHS. If the LHS is a bitfield and the RHS is not
2760 a constant, then we can not record any equivalences for this
2761 statement since we would need to represent the widening or
2762 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2763 and should not be necessary if GCC represented bitfields
2764 properly. */
2765 if (lhs_code == COMPONENT_REF
2766 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
2768 if (TREE_CONSTANT (rhs))
2769 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
2770 else
2771 rhs = NULL;
2773 /* If the value overflowed, then we can not use this equivalence. */
2774 if (rhs && ! is_gimple_min_invariant (rhs))
2775 rhs = NULL;
2778 if (rhs)
2780 /* Build a new statement with the RHS and LHS exchanged. */
2781 new = build (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
2783 create_ssa_artficial_load_stmt (new, stmt);
2785 /* Finally enter the statement into the available expression
2786 table. */
2787 lookup_avail_expr (new, true);
2792 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2793 CONST_AND_COPIES. */
2795 static bool
2796 cprop_operand (tree stmt, use_operand_p op_p)
2798 bool may_have_exposed_new_symbols = false;
2799 tree val;
2800 tree op = USE_FROM_PTR (op_p);
2802 /* If the operand has a known constant value or it is known to be a
2803 copy of some other variable, use the value or copy stored in
2804 CONST_AND_COPIES. */
2805 val = SSA_NAME_VALUE (op);
2806 if (val && val != op && TREE_CODE (val) != VALUE_HANDLE)
2808 tree op_type, val_type;
2810 /* Do not change the base variable in the virtual operand
2811 tables. That would make it impossible to reconstruct
2812 the renamed virtual operand if we later modify this
2813 statement. Also only allow the new value to be an SSA_NAME
2814 for propagation into virtual operands. */
2815 if (!is_gimple_reg (op)
2816 && (TREE_CODE (val) != SSA_NAME
2817 || is_gimple_reg (val)
2818 || get_virtual_var (val) != get_virtual_var (op)))
2819 return false;
2821 /* Do not replace hard register operands in asm statements. */
2822 if (TREE_CODE (stmt) == ASM_EXPR
2823 && !may_propagate_copy_into_asm (op))
2824 return false;
2826 /* Get the toplevel type of each operand. */
2827 op_type = TREE_TYPE (op);
2828 val_type = TREE_TYPE (val);
2830 /* While both types are pointers, get the type of the object
2831 pointed to. */
2832 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
2834 op_type = TREE_TYPE (op_type);
2835 val_type = TREE_TYPE (val_type);
2838 /* Make sure underlying types match before propagating a constant by
2839 converting the constant to the proper type. Note that convert may
2840 return a non-gimple expression, in which case we ignore this
2841 propagation opportunity. */
2842 if (TREE_CODE (val) != SSA_NAME)
2844 if (!lang_hooks.types_compatible_p (op_type, val_type))
2846 val = fold_convert (TREE_TYPE (op), val);
2847 if (!is_gimple_min_invariant (val))
2848 return false;
2852 /* Certain operands are not allowed to be copy propagated due
2853 to their interaction with exception handling and some GCC
2854 extensions. */
2855 else if (!may_propagate_copy (op, val))
2856 return false;
2858 /* Do not propagate copies if the propagated value is at a deeper loop
2859 depth than the propagatee. Otherwise, this may move loop variant
2860 variables outside of their loops and prevent coalescing
2861 opportunities. If the value was loop invariant, it will be hoisted
2862 by LICM and exposed for copy propagation. */
2863 if (loop_depth_of_name (val) > loop_depth_of_name (op))
2864 return false;
2866 /* Dump details. */
2867 if (dump_file && (dump_flags & TDF_DETAILS))
2869 fprintf (dump_file, " Replaced '");
2870 print_generic_expr (dump_file, op, dump_flags);
2871 fprintf (dump_file, "' with %s '",
2872 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
2873 print_generic_expr (dump_file, val, dump_flags);
2874 fprintf (dump_file, "'\n");
2877 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2878 that we may have exposed a new symbol for SSA renaming. */
2879 if (TREE_CODE (val) == ADDR_EXPR
2880 || (POINTER_TYPE_P (TREE_TYPE (op))
2881 && is_gimple_min_invariant (val)))
2882 may_have_exposed_new_symbols = true;
2884 if (TREE_CODE (val) != SSA_NAME)
2885 opt_stats.num_const_prop++;
2886 else
2887 opt_stats.num_copy_prop++;
2889 propagate_value (op_p, val);
2891 /* And note that we modified this statement. This is now
2892 safe, even if we changed virtual operands since we will
2893 rescan the statement and rewrite its operands again. */
2894 mark_stmt_modified (stmt);
2896 return may_have_exposed_new_symbols;
2899 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2900 known value for that SSA_NAME (or NULL if no value is known).
2902 Propagate values from CONST_AND_COPIES into the uses, vuses and
2903 v_may_def_ops of STMT. */
2905 static bool
2906 cprop_into_stmt (tree stmt)
2908 bool may_have_exposed_new_symbols = false;
2909 use_operand_p op_p;
2910 ssa_op_iter iter;
2912 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
2914 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
2915 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
2918 return may_have_exposed_new_symbols;
2922 /* Optimize the statement pointed by iterator SI.
2924 We try to perform some simplistic global redundancy elimination and
2925 constant propagation:
2927 1- To detect global redundancy, we keep track of expressions that have
2928 been computed in this block and its dominators. If we find that the
2929 same expression is computed more than once, we eliminate repeated
2930 computations by using the target of the first one.
2932 2- Constant values and copy assignments. This is used to do very
2933 simplistic constant and copy propagation. When a constant or copy
2934 assignment is found, we map the value on the RHS of the assignment to
2935 the variable in the LHS in the CONST_AND_COPIES table. */
2937 static void
2938 optimize_stmt (struct dom_walk_data *walk_data, basic_block bb,
2939 block_stmt_iterator si)
2941 stmt_ann_t ann;
2942 tree stmt, old_stmt;
2943 bool may_optimize_p;
2944 bool may_have_exposed_new_symbols = false;
2946 old_stmt = stmt = bsi_stmt (si);
2948 update_stmt_if_modified (stmt);
2949 ann = stmt_ann (stmt);
2950 opt_stats.num_stmts++;
2951 may_have_exposed_new_symbols = false;
2953 if (dump_file && (dump_flags & TDF_DETAILS))
2955 fprintf (dump_file, "Optimizing statement ");
2956 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2959 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2960 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
2962 /* If the statement has been modified with constant replacements,
2963 fold its RHS before checking for redundant computations. */
2964 if (ann->modified)
2966 tree rhs;
2968 /* Try to fold the statement making sure that STMT is kept
2969 up to date. */
2970 if (fold_stmt (bsi_stmt_ptr (si)))
2972 stmt = bsi_stmt (si);
2973 ann = stmt_ann (stmt);
2975 if (dump_file && (dump_flags & TDF_DETAILS))
2977 fprintf (dump_file, " Folded to: ");
2978 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2982 rhs = get_rhs (stmt);
2983 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
2984 recompute_tree_invarant_for_addr_expr (rhs);
2986 /* Constant/copy propagation above may change the set of
2987 virtual operands associated with this statement. Folding
2988 may remove the need for some virtual operands.
2990 Indicate we will need to rescan and rewrite the statement. */
2991 may_have_exposed_new_symbols = true;
2994 /* Check for redundant computations. Do this optimization only
2995 for assignments that have no volatile ops and conditionals. */
2996 may_optimize_p = (!ann->has_volatile_ops
2997 && ((TREE_CODE (stmt) == RETURN_EXPR
2998 && TREE_OPERAND (stmt, 0)
2999 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
3000 && ! (TREE_SIDE_EFFECTS
3001 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
3002 || (TREE_CODE (stmt) == MODIFY_EXPR
3003 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
3004 || TREE_CODE (stmt) == COND_EXPR
3005 || TREE_CODE (stmt) == SWITCH_EXPR));
3007 if (may_optimize_p)
3008 may_have_exposed_new_symbols
3009 |= eliminate_redundant_computations (walk_data, stmt, ann);
3011 /* Record any additional equivalences created by this statement. */
3012 if (TREE_CODE (stmt) == MODIFY_EXPR)
3013 record_equivalences_from_stmt (stmt,
3014 may_optimize_p,
3015 ann);
3017 /* If STMT is a COND_EXPR and it was modified, then we may know
3018 where it goes. If that is the case, then mark the CFG as altered.
3020 This will cause us to later call remove_unreachable_blocks and
3021 cleanup_tree_cfg when it is safe to do so. It is not safe to
3022 clean things up here since removal of edges and such can trigger
3023 the removal of PHI nodes, which in turn can release SSA_NAMEs to
3024 the manager.
3026 That's all fine and good, except that once SSA_NAMEs are released
3027 to the manager, we must not call create_ssa_name until all references
3028 to released SSA_NAMEs have been eliminated.
3030 All references to the deleted SSA_NAMEs can not be eliminated until
3031 we remove unreachable blocks.
3033 We can not remove unreachable blocks until after we have completed
3034 any queued jump threading.
3036 We can not complete any queued jump threads until we have taken
3037 appropriate variables out of SSA form. Taking variables out of
3038 SSA form can call create_ssa_name and thus we lose.
3040 Ultimately I suspect we're going to need to change the interface
3041 into the SSA_NAME manager. */
3043 if (ann->modified)
3045 tree val = NULL;
3047 if (TREE_CODE (stmt) == COND_EXPR)
3048 val = COND_EXPR_COND (stmt);
3049 else if (TREE_CODE (stmt) == SWITCH_EXPR)
3050 val = SWITCH_COND (stmt);
3052 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
3053 cfg_altered = true;
3055 /* If we simplified a statement in such a way as to be shown that it
3056 cannot trap, update the eh information and the cfg to match. */
3057 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
3059 bitmap_set_bit (need_eh_cleanup, bb->index);
3060 if (dump_file && (dump_flags & TDF_DETAILS))
3061 fprintf (dump_file, " Flagged to clear EH edges.\n");
3065 if (may_have_exposed_new_symbols)
3066 VEC_safe_push (tree, heap, stmts_to_rescan, bsi_stmt (si));
3069 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
3070 available expression hashtable, then return the LHS from the hash
3071 table.
3073 If INSERT is true, then we also update the available expression
3074 hash table to account for the changes made to STMT. */
3076 static tree
3077 update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert)
3079 tree cached_lhs = NULL;
3081 /* Remove the old entry from the hash table. */
3082 if (insert)
3084 struct expr_hash_elt element;
3086 initialize_hash_element (stmt, NULL, &element);
3087 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
3090 /* Now update the RHS of the assignment. */
3091 TREE_OPERAND (stmt, 1) = new_rhs;
3093 /* Now lookup the updated statement in the hash table. */
3094 cached_lhs = lookup_avail_expr (stmt, insert);
3096 /* We have now called lookup_avail_expr twice with two different
3097 versions of this same statement, once in optimize_stmt, once here.
3099 We know the call in optimize_stmt did not find an existing entry
3100 in the hash table, so a new entry was created. At the same time
3101 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
3103 If this call failed to find an existing entry on the hash table,
3104 then the new version of this statement was entered into the
3105 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3106 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3108 If this call succeeded, we still have one copy of this statement
3109 on the BLOCK_AVAIL_EXPRs vector.
3111 For both cases, we need to pop the most recent entry off the
3112 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3113 statement in the hash tables, that will leave precisely one
3114 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3115 we found a copy of this statement in the second hash table lookup
3116 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3117 if (insert)
3118 VEC_pop (tree, avail_exprs_stack);
3120 /* And make sure we record the fact that we modified this
3121 statement. */
3122 mark_stmt_modified (stmt);
3124 return cached_lhs;
3127 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3128 found, return its LHS. Otherwise insert STMT in the table and return
3129 NULL_TREE.
3131 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3132 is also added to the stack pointed by BLOCK_AVAIL_EXPRS_P, so that they
3133 can be removed when we finish processing this block and its children.
3135 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3136 contains no CALL_EXPR on its RHS and makes no volatile nor
3137 aliased references. */
3139 static tree
3140 lookup_avail_expr (tree stmt, bool insert)
3142 void **slot;
3143 tree lhs;
3144 tree temp;
3145 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
3147 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
3149 initialize_hash_element (stmt, lhs, element);
3151 /* Don't bother remembering constant assignments and copy operations.
3152 Constants and copy operations are handled by the constant/copy propagator
3153 in optimize_stmt. */
3154 if (TREE_CODE (element->rhs) == SSA_NAME
3155 || is_gimple_min_invariant (element->rhs))
3157 free (element);
3158 return NULL_TREE;
3161 /* If this is an equality test against zero, see if we have recorded a
3162 nonzero value for the variable in question. */
3163 if ((TREE_CODE (element->rhs) == EQ_EXPR
3164 || TREE_CODE (element->rhs) == NE_EXPR)
3165 && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
3166 && integer_zerop (TREE_OPERAND (element->rhs, 1)))
3168 int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
3170 if (bitmap_bit_p (nonzero_vars, indx))
3172 tree t = element->rhs;
3173 free (element);
3175 if (TREE_CODE (t) == EQ_EXPR)
3176 return boolean_false_node;
3177 else
3178 return boolean_true_node;
3182 /* Finally try to find the expression in the main expression hash table. */
3183 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
3184 (insert ? INSERT : NO_INSERT));
3185 if (slot == NULL)
3187 free (element);
3188 return NULL_TREE;
3191 if (*slot == NULL)
3193 *slot = (void *) element;
3194 VEC_safe_push (tree, heap, avail_exprs_stack,
3195 stmt ? stmt : element->rhs);
3196 return NULL_TREE;
3199 /* Extract the LHS of the assignment so that it can be used as the current
3200 definition of another variable. */
3201 lhs = ((struct expr_hash_elt *)*slot)->lhs;
3203 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3204 use the value from the const_and_copies table. */
3205 if (TREE_CODE (lhs) == SSA_NAME)
3207 temp = SSA_NAME_VALUE (lhs);
3208 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
3209 lhs = temp;
3212 free (element);
3213 return lhs;
3216 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3217 range of values that result in the conditional having a true value.
3219 Return true if we are successful in extracting a range from COND and
3220 false if we are unsuccessful. */
3222 static bool
3223 extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
3225 tree op1 = TREE_OPERAND (cond, 1);
3226 tree high, low, type;
3227 int inverted;
3229 type = TREE_TYPE (op1);
3231 /* Experiments have shown that it's rarely, if ever useful to
3232 record ranges for enumerations. Presumably this is due to
3233 the fact that they're rarely used directly. They are typically
3234 cast into an integer type and used that way. */
3235 if (TREE_CODE (type) != INTEGER_TYPE
3236 /* We don't know how to deal with types with variable bounds. */
3237 || TREE_CODE (TYPE_MIN_VALUE (type)) != INTEGER_CST
3238 || TREE_CODE (TYPE_MAX_VALUE (type)) != INTEGER_CST)
3239 return 0;
3241 switch (TREE_CODE (cond))
3243 case EQ_EXPR:
3244 high = low = op1;
3245 inverted = 0;
3246 break;
3248 case NE_EXPR:
3249 high = low = op1;
3250 inverted = 1;
3251 break;
3253 case GE_EXPR:
3254 low = op1;
3255 high = TYPE_MAX_VALUE (type);
3256 inverted = 0;
3257 break;
3259 case GT_EXPR:
3260 high = TYPE_MAX_VALUE (type);
3261 if (!tree_int_cst_lt (op1, high))
3262 return 0;
3263 low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
3264 inverted = 0;
3265 break;
3267 case LE_EXPR:
3268 high = op1;
3269 low = TYPE_MIN_VALUE (type);
3270 inverted = 0;
3271 break;
3273 case LT_EXPR:
3274 low = TYPE_MIN_VALUE (type);
3275 if (!tree_int_cst_lt (low, op1))
3276 return 0;
3277 high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
3278 inverted = 0;
3279 break;
3281 default:
3282 return 0;
3285 *hi_p = high;
3286 *lo_p = low;
3287 *inverted_p = inverted;
3288 return 1;
3291 /* Record a range created by COND for basic block BB. */
3293 static void
3294 record_range (tree cond, basic_block bb)
3296 enum tree_code code = TREE_CODE (cond);
3298 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3299 They rarely allow for meaningful range optimizations and significantly
3300 complicate the implementation. */
3301 if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
3302 || code == GE_EXPR || code == EQ_EXPR)
3303 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
3305 struct vrp_hash_elt *vrp_hash_elt;
3306 struct vrp_element *element;
3307 varray_type *vrp_records_p;
3308 void **slot;
3311 vrp_hash_elt = xmalloc (sizeof (struct vrp_hash_elt));
3312 vrp_hash_elt->var = TREE_OPERAND (cond, 0);
3313 vrp_hash_elt->records = NULL;
3314 slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
3316 if (*slot == NULL)
3317 *slot = (void *) vrp_hash_elt;
3318 else
3319 free (vrp_hash_elt);
3321 vrp_hash_elt = (struct vrp_hash_elt *) *slot;
3322 vrp_records_p = &vrp_hash_elt->records;
3324 element = ggc_alloc (sizeof (struct vrp_element));
3325 element->low = NULL;
3326 element->high = NULL;
3327 element->cond = cond;
3328 element->bb = bb;
3330 if (*vrp_records_p == NULL)
3331 VARRAY_GENERIC_PTR_INIT (*vrp_records_p, 2, "vrp records");
3333 VARRAY_PUSH_GENERIC_PTR (*vrp_records_p, element);
3334 VEC_safe_push (tree, heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
3338 /* Hashing and equality functions for VRP_DATA.
3340 Since this hash table is addressed by SSA_NAMEs, we can hash on
3341 their version number and equality can be determined with a
3342 pointer comparison. */
3344 static hashval_t
3345 vrp_hash (const void *p)
3347 tree var = ((struct vrp_hash_elt *)p)->var;
3349 return SSA_NAME_VERSION (var);
3352 static int
3353 vrp_eq (const void *p1, const void *p2)
3355 tree var1 = ((struct vrp_hash_elt *)p1)->var;
3356 tree var2 = ((struct vrp_hash_elt *)p2)->var;
3358 return var1 == var2;
3361 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3362 MODIFY_EXPR statements. We compute a value number for expressions using
3363 the code of the expression and the SSA numbers of its operands. */
3365 static hashval_t
3366 avail_expr_hash (const void *p)
3368 tree stmt = ((struct expr_hash_elt *)p)->stmt;
3369 tree rhs = ((struct expr_hash_elt *)p)->rhs;
3370 tree vuse;
3371 ssa_op_iter iter;
3372 hashval_t val = 0;
3374 /* iterative_hash_expr knows how to deal with any expression and
3375 deals with commutative operators as well, so just use it instead
3376 of duplicating such complexities here. */
3377 val = iterative_hash_expr (rhs, val);
3379 /* If the hash table entry is not associated with a statement, then we
3380 can just hash the expression and not worry about virtual operands
3381 and such. */
3382 if (!stmt || !stmt_ann (stmt))
3383 return val;
3385 /* Add the SSA version numbers of every vuse operand. This is important
3386 because compound variables like arrays are not renamed in the
3387 operands. Rather, the rename is done on the virtual variable
3388 representing all the elements of the array. */
3389 FOR_EACH_SSA_TREE_OPERAND (vuse, stmt, iter, SSA_OP_VUSE)
3390 val = iterative_hash_expr (vuse, val);
3392 return val;
3395 static hashval_t
3396 real_avail_expr_hash (const void *p)
3398 return ((const struct expr_hash_elt *)p)->hash;
3401 static int
3402 avail_expr_eq (const void *p1, const void *p2)
3404 tree stmt1 = ((struct expr_hash_elt *)p1)->stmt;
3405 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
3406 tree stmt2 = ((struct expr_hash_elt *)p2)->stmt;
3407 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
3409 /* If they are the same physical expression, return true. */
3410 if (rhs1 == rhs2 && stmt1 == stmt2)
3411 return true;
3413 /* If their codes are not equal, then quit now. */
3414 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
3415 return false;
3417 /* In case of a collision, both RHS have to be identical and have the
3418 same VUSE operands. */
3419 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
3420 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
3421 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
3423 bool ret = compare_ssa_operands_equal (stmt1, stmt2, SSA_OP_VUSE);
3424 gcc_assert (!ret || ((struct expr_hash_elt *)p1)->hash
3425 == ((struct expr_hash_elt *)p2)->hash);
3426 return ret;
3429 return false;