* tree-flow.h (struct stmt_ann_d): Remove
[official-gcc.git] / gcc / tree-ssa-dom.c
blobb6257d33383f07d0c831b235ae3dec869d3d7905
1 /* SSA Dominator optimizations for trees
2 Copyright (C) 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
3 Contributed by Diego Novillo <dnovillo@redhat.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
10 any later version.
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to
19 the Free Software Foundation, 51 Franklin Street, Fifth Floor,
20 Boston, MA 02110-1301, USA. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "tree.h"
27 #include "flags.h"
28 #include "rtl.h"
29 #include "tm_p.h"
30 #include "ggc.h"
31 #include "basic-block.h"
32 #include "cfgloop.h"
33 #include "output.h"
34 #include "expr.h"
35 #include "function.h"
36 #include "diagnostic.h"
37 #include "timevar.h"
38 #include "tree-dump.h"
39 #include "tree-flow.h"
40 #include "domwalk.h"
41 #include "real.h"
42 #include "tree-pass.h"
43 #include "tree-ssa-propagate.h"
44 #include "langhooks.h"
45 #include "params.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 /* Bitmap of blocks that are scheduled to be threaded through. This
145 is used to communicate with thread_through_blocks. */
146 static bitmap threaded_blocks;
148 /* Stack of SSA_NAMEs which need their NONZERO_VARS property cleared
149 when the current block is finalized.
151 A NULL entry is used to mark the end of names needing their
152 entry in NONZERO_VARS cleared during finalization of this block. */
153 static VEC(tree,heap) *nonzero_vars_stack;
155 /* Track whether or not we have changed the control flow graph. */
156 static bool cfg_altered;
158 /* Bitmap of blocks that have had EH statements cleaned. We should
159 remove their dead edges eventually. */
160 static bitmap need_eh_cleanup;
162 /* Statistics for dominator optimizations. */
163 struct opt_stats_d
165 long num_stmts;
166 long num_exprs_considered;
167 long num_re;
168 long num_const_prop;
169 long num_copy_prop;
170 long num_iterations;
173 static struct opt_stats_d opt_stats;
175 /* Value range propagation record. Each time we encounter a conditional
176 of the form SSA_NAME COND CONST we create a new vrp_element to record
177 how the condition affects the possible values SSA_NAME may have.
179 Each record contains the condition tested (COND), and the range of
180 values the variable may legitimately have if COND is true. Note the
181 range of values may be a smaller range than COND specifies if we have
182 recorded other ranges for this variable. Each record also contains the
183 block in which the range was recorded for invalidation purposes.
185 Note that the current known range is computed lazily. This allows us
186 to avoid the overhead of computing ranges which are never queried.
188 When we encounter a conditional, we look for records which constrain
189 the SSA_NAME used in the condition. In some cases those records allow
190 us to determine the condition's result at compile time. In other cases
191 they may allow us to simplify the condition.
193 We also use value ranges to do things like transform signed div/mod
194 operations into unsigned div/mod or to simplify ABS_EXPRs.
196 Simple experiments have shown these optimizations to not be all that
197 useful on switch statements (much to my surprise). So switch statement
198 optimizations are not performed.
200 Note carefully we do not propagate information through each statement
201 in the block. i.e., if we know variable X has a value defined of
202 [0, 25] and we encounter Y = X + 1, we do not track a value range
203 for Y (which would be [1, 26] if we cared). Similarly we do not
204 constrain values as we encounter narrowing typecasts, etc. */
206 struct vrp_element
208 /* The highest and lowest values the variable in COND may contain when
209 COND is true. Note this may not necessarily be the same values
210 tested by COND if the same variable was used in earlier conditionals.
212 Note this is computed lazily and thus can be NULL indicating that
213 the values have not been computed yet. */
214 tree low;
215 tree high;
217 /* The actual conditional we recorded. This is needed since we compute
218 ranges lazily. */
219 tree cond;
221 /* The basic block where this record was created. We use this to determine
222 when to remove records. */
223 basic_block bb;
226 /* A hash table holding value range records (VRP_ELEMENTs) for a given
227 SSA_NAME. We used to use a varray indexed by SSA_NAME_VERSION, but
228 that gets awful wasteful, particularly since the density objects
229 with useful information is very low. */
230 static htab_t vrp_data;
232 typedef struct vrp_element *vrp_element_p;
234 DEF_VEC_P(vrp_element_p);
235 DEF_VEC_ALLOC_P(vrp_element_p,heap);
237 /* An entry in the VRP_DATA hash table. We record the variable and a
238 varray of VRP_ELEMENT records associated with that variable. */
239 struct vrp_hash_elt
241 tree var;
242 VEC(vrp_element_p,heap) *records;
245 /* Array of variables which have their values constrained by operations
246 in this basic block. We use this during finalization to know
247 which variables need their VRP data updated. */
249 /* Stack of SSA_NAMEs which had their values constrained by operations
250 in this basic block. During finalization of this block we use this
251 list to determine which variables need their VRP data updated.
253 A NULL entry marks the end of the SSA_NAMEs associated with this block. */
254 static VEC(tree,heap) *vrp_variables_stack;
256 struct eq_expr_value
258 tree src;
259 tree dst;
262 /* Local functions. */
263 static void optimize_stmt (struct dom_walk_data *,
264 basic_block bb,
265 block_stmt_iterator);
266 static tree lookup_avail_expr (tree, bool);
267 static hashval_t vrp_hash (const void *);
268 static int vrp_eq (const void *, const void *);
269 static hashval_t avail_expr_hash (const void *);
270 static hashval_t real_avail_expr_hash (const void *);
271 static int avail_expr_eq (const void *, const void *);
272 static void htab_statistics (FILE *, htab_t);
273 static void record_cond (tree, tree);
274 static void record_const_or_copy (tree, tree);
275 static void record_equality (tree, tree);
276 static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int);
277 static tree find_equivalent_equality_comparison (tree);
278 static void record_range (tree, basic_block);
279 static bool extract_range_from_cond (tree, tree *, tree *, int *);
280 static void record_equivalences_from_phis (basic_block);
281 static void record_equivalences_from_incoming_edge (basic_block);
282 static bool eliminate_redundant_computations (tree, stmt_ann_t);
283 static void record_equivalences_from_stmt (tree, int, stmt_ann_t);
284 static void thread_across_edge (struct dom_walk_data *, edge);
285 static void dom_opt_finalize_block (struct dom_walk_data *, basic_block);
286 static void dom_opt_initialize_block (struct dom_walk_data *, basic_block);
287 static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block);
288 static void remove_local_expressions_from_table (void);
289 static void restore_vars_to_original_value (void);
290 static edge single_incoming_edge_ignoring_loop_edges (basic_block);
291 static void restore_nonzero_vars_to_original_value (void);
292 static inline bool unsafe_associative_fp_binop (tree);
295 /* Local version of fold that doesn't introduce cruft. */
297 static tree
298 local_fold (tree t)
300 t = fold (t);
302 /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that
303 may have been added by fold, and "useless" type conversions that might
304 now be apparent due to propagation. */
305 STRIP_USELESS_TYPE_CONVERSION (t);
307 return t;
310 /* Allocate an EDGE_INFO for edge E and attach it to E.
311 Return the new EDGE_INFO structure. */
313 static struct edge_info *
314 allocate_edge_info (edge e)
316 struct edge_info *edge_info;
318 edge_info = XCNEW (struct edge_info);
320 e->aux = edge_info;
321 return edge_info;
324 /* Free all EDGE_INFO structures associated with edges in the CFG.
325 If a particular edge can be threaded, copy the redirection
326 target from the EDGE_INFO structure into the edge's AUX field
327 as required by code to update the CFG and SSA graph for
328 jump threading. */
330 static void
331 free_all_edge_infos (void)
333 basic_block bb;
334 edge_iterator ei;
335 edge e;
337 FOR_EACH_BB (bb)
339 FOR_EACH_EDGE (e, ei, bb->preds)
341 struct edge_info *edge_info = (struct edge_info *) e->aux;
343 if (edge_info)
345 e->aux = edge_info->redirection_target;
346 if (edge_info->cond_equivalences)
347 free (edge_info->cond_equivalences);
348 free (edge_info);
354 /* Free an instance of vrp_hash_elt. */
356 static void
357 vrp_free (void *data)
359 struct vrp_hash_elt *elt = (struct vrp_hash_elt *) data;
360 struct VEC(vrp_element_p,heap) **vrp_elt = &elt->records;
362 VEC_free (vrp_element_p, heap, *vrp_elt);
363 free (elt);
366 /* Jump threading, redundancy elimination and const/copy propagation.
368 This pass may expose new symbols that need to be renamed into SSA. For
369 every new symbol exposed, its corresponding bit will be set in
370 VARS_TO_RENAME. */
372 static void
373 tree_ssa_dominator_optimize (void)
375 struct dom_walk_data walk_data;
376 unsigned int i;
377 struct loops loops_info;
379 memset (&opt_stats, 0, sizeof (opt_stats));
381 /* Create our hash tables. */
382 avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free);
383 vrp_data = htab_create (ceil_log2 (num_ssa_names), vrp_hash, vrp_eq,
384 vrp_free);
385 avail_exprs_stack = VEC_alloc (tree, heap, 20);
386 const_and_copies_stack = VEC_alloc (tree, heap, 20);
387 nonzero_vars_stack = VEC_alloc (tree, heap, 20);
388 vrp_variables_stack = VEC_alloc (tree, heap, 20);
389 stmts_to_rescan = VEC_alloc (tree, heap, 20);
390 nonzero_vars = BITMAP_ALLOC (NULL);
391 threaded_blocks = BITMAP_ALLOC (NULL);
392 need_eh_cleanup = BITMAP_ALLOC (NULL);
394 /* Setup callbacks for the generic dominator tree walker. */
395 walk_data.walk_stmts_backward = false;
396 walk_data.dom_direction = CDI_DOMINATORS;
397 walk_data.initialize_block_local_data = NULL;
398 walk_data.before_dom_children_before_stmts = dom_opt_initialize_block;
399 walk_data.before_dom_children_walk_stmts = optimize_stmt;
400 walk_data.before_dom_children_after_stmts = propagate_to_outgoing_edges;
401 walk_data.after_dom_children_before_stmts = NULL;
402 walk_data.after_dom_children_walk_stmts = NULL;
403 walk_data.after_dom_children_after_stmts = dom_opt_finalize_block;
404 /* Right now we only attach a dummy COND_EXPR to the global data pointer.
405 When we attach more stuff we'll need to fill this out with a real
406 structure. */
407 walk_data.global_data = NULL;
408 walk_data.block_local_data_size = 0;
409 walk_data.interesting_blocks = NULL;
411 /* Now initialize the dominator walker. */
412 init_walk_dominator_tree (&walk_data);
414 calculate_dominance_info (CDI_DOMINATORS);
416 /* We need to know which edges exit loops so that we can
417 aggressively thread through loop headers to an exit
418 edge. */
419 flow_loops_find (&loops_info);
420 mark_loop_exit_edges (&loops_info);
421 flow_loops_free (&loops_info);
423 /* Clean up the CFG so that any forwarder blocks created by loop
424 canonicalization are removed. */
425 cleanup_tree_cfg ();
426 calculate_dominance_info (CDI_DOMINATORS);
428 /* If we prove certain blocks are unreachable, then we want to
429 repeat the dominator optimization process as PHI nodes may
430 have turned into copies which allows better propagation of
431 values. So we repeat until we do not identify any new unreachable
432 blocks. */
435 /* Optimize the dominator tree. */
436 cfg_altered = false;
438 /* We need accurate information regarding back edges in the CFG
439 for jump threading. */
440 mark_dfs_back_edges ();
442 /* Recursively walk the dominator tree optimizing statements. */
443 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
446 block_stmt_iterator bsi;
447 basic_block bb;
448 FOR_EACH_BB (bb)
450 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
452 update_stmt_if_modified (bsi_stmt (bsi));
457 /* If we exposed any new variables, go ahead and put them into
458 SSA form now, before we handle jump threading. This simplifies
459 interactions between rewriting of _DECL nodes into SSA form
460 and rewriting SSA_NAME nodes into SSA form after block
461 duplication and CFG manipulation. */
462 update_ssa (TODO_update_ssa);
464 free_all_edge_infos ();
466 /* Thread jumps, creating duplicate blocks as needed. */
467 cfg_altered |= thread_through_all_blocks (threaded_blocks);
469 /* Removal of statements may make some EH edges dead. Purge
470 such edges from the CFG as needed. */
471 if (!bitmap_empty_p (need_eh_cleanup))
473 cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
474 bitmap_zero (need_eh_cleanup);
477 if (cfg_altered)
478 free_dominance_info (CDI_DOMINATORS);
480 /* Only iterate if we threaded jumps AND the CFG cleanup did
481 something interesting. Other cases generate far fewer
482 optimization opportunities and thus are not worth another
483 full DOM iteration. */
484 cfg_altered &= cleanup_tree_cfg ();
486 if (rediscover_loops_after_threading)
488 /* Rerun basic loop analysis to discover any newly
489 created loops and update the set of exit edges. */
490 rediscover_loops_after_threading = false;
491 flow_loops_find (&loops_info);
492 mark_loop_exit_edges (&loops_info);
493 flow_loops_free (&loops_info);
495 /* Remove any forwarder blocks inserted by loop
496 header canonicalization. */
497 cleanup_tree_cfg ();
500 calculate_dominance_info (CDI_DOMINATORS);
502 update_ssa (TODO_update_ssa);
504 /* Reinitialize the various tables. */
505 bitmap_clear (nonzero_vars);
506 bitmap_clear (threaded_blocks);
507 htab_empty (avail_exprs);
508 htab_empty (vrp_data);
510 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
512 This must be done before we iterate as we might have a
513 reference to an SSA_NAME which was removed by the call to
514 update_ssa.
516 Long term we will be able to let everything in SSA_NAME_VALUE
517 persist. However, for now, we know this is the safe thing to do. */
518 for (i = 0; i < num_ssa_names; i++)
520 tree name = ssa_name (i);
521 tree value;
523 if (!name)
524 continue;
526 value = SSA_NAME_VALUE (name);
527 if (value && !is_gimple_min_invariant (value))
528 SSA_NAME_VALUE (name) = NULL;
531 opt_stats.num_iterations++;
533 while (optimize > 1 && cfg_altered);
535 /* Debugging dumps. */
536 if (dump_file && (dump_flags & TDF_STATS))
537 dump_dominator_optimization_stats (dump_file);
539 /* We emptied the hash table earlier, now delete it completely. */
540 htab_delete (avail_exprs);
541 htab_delete (vrp_data);
543 /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
544 CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
545 of the do-while loop above. */
547 /* And finalize the dominator walker. */
548 fini_walk_dominator_tree (&walk_data);
550 /* Free nonzero_vars. */
551 BITMAP_FREE (nonzero_vars);
552 BITMAP_FREE (threaded_blocks);
553 BITMAP_FREE (need_eh_cleanup);
555 VEC_free (tree, heap, avail_exprs_stack);
556 VEC_free (tree, heap, const_and_copies_stack);
557 VEC_free (tree, heap, nonzero_vars_stack);
558 VEC_free (tree, heap, vrp_variables_stack);
559 VEC_free (tree, heap, stmts_to_rescan);
562 static bool
563 gate_dominator (void)
565 return flag_tree_dom != 0;
568 struct tree_opt_pass pass_dominator =
570 "dom", /* name */
571 gate_dominator, /* gate */
572 tree_ssa_dominator_optimize, /* execute */
573 NULL, /* sub */
574 NULL, /* next */
575 0, /* static_pass_number */
576 TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */
577 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
578 0, /* properties_provided */
579 0, /* properties_destroyed */
580 0, /* todo_flags_start */
581 TODO_dump_func
582 | TODO_update_ssa
583 | TODO_verify_ssa, /* todo_flags_finish */
584 0 /* letter */
588 /* Given a stmt CONDSTMT containing a COND_EXPR, canonicalize the
589 COND_EXPR into a canonical form. */
591 static void
592 canonicalize_comparison (tree condstmt)
594 tree cond = COND_EXPR_COND (condstmt);
595 tree op0;
596 tree op1;
597 enum tree_code code = TREE_CODE (cond);
599 if (!COMPARISON_CLASS_P (cond))
600 return;
602 op0 = TREE_OPERAND (cond, 0);
603 op1 = TREE_OPERAND (cond, 1);
605 /* If it would be profitable to swap the operands, then do so to
606 canonicalize the statement, enabling better optimization.
608 By placing canonicalization of such expressions here we
609 transparently keep statements in canonical form, even
610 when the statement is modified. */
611 if (tree_swap_operands_p (op0, op1, false))
613 /* For relationals we need to swap the operands
614 and change the code. */
615 if (code == LT_EXPR
616 || code == GT_EXPR
617 || code == LE_EXPR
618 || code == GE_EXPR)
620 TREE_SET_CODE (cond, swap_tree_comparison (code));
621 swap_tree_operands (condstmt,
622 &TREE_OPERAND (cond, 0),
623 &TREE_OPERAND (cond, 1));
624 /* If one operand was in the operand cache, but the other is
625 not, because it is a constant, this is a case that the
626 internal updating code of swap_tree_operands can't handle
627 properly. */
628 if (TREE_CODE_CLASS (TREE_CODE (op0))
629 != TREE_CODE_CLASS (TREE_CODE (op1)))
630 update_stmt (condstmt);
634 /* We are exiting E->src, see if E->dest ends with a conditional
635 jump which has a known value when reached via E.
637 Special care is necessary if E is a back edge in the CFG as we
638 will have already recorded equivalences for E->dest into our
639 various tables, including the result of the conditional at
640 the end of E->dest. Threading opportunities are severely
641 limited in that case to avoid short-circuiting the loop
642 incorrectly.
644 Note it is quite common for the first block inside a loop to
645 end with a conditional which is either always true or always
646 false when reached via the loop backedge. Thus we do not want
647 to blindly disable threading across a loop backedge. */
649 static void
650 thread_across_edge (struct dom_walk_data *walk_data, edge e)
652 block_stmt_iterator bsi;
653 tree stmt = NULL;
654 tree phi;
655 int stmt_count = 0;
656 int max_stmt_count;
659 /* If E->dest does not end with a conditional, then there is
660 nothing to do. */
661 bsi = bsi_last (e->dest);
662 if (bsi_end_p (bsi)
663 || ! bsi_stmt (bsi)
664 || (TREE_CODE (bsi_stmt (bsi)) != COND_EXPR
665 && TREE_CODE (bsi_stmt (bsi)) != GOTO_EXPR
666 && TREE_CODE (bsi_stmt (bsi)) != SWITCH_EXPR))
667 return;
669 /* The basic idea here is to use whatever knowledge we have
670 from our dominator walk to simplify statements in E->dest,
671 with the ultimate goal being to simplify the conditional
672 at the end of E->dest.
674 Note that we must undo any changes we make to the underlying
675 statements as the simplifications we are making are control
676 flow sensitive (ie, the simplifications are valid when we
677 traverse E, but may not be valid on other paths to E->dest. */
679 /* Each PHI creates a temporary equivalence, record them. Again
680 these are context sensitive equivalences and will be removed
681 by our caller. */
682 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
684 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
685 tree dst = PHI_RESULT (phi);
687 /* Do not include virtual PHIs in our statement count as
688 they never generate code. */
689 if (is_gimple_reg (dst))
690 stmt_count++;
692 /* If the desired argument is not the same as this PHI's result
693 and it is set by a PHI in E->dest, then we can not thread
694 through E->dest. */
695 if (src != dst
696 && TREE_CODE (src) == SSA_NAME
697 && TREE_CODE (SSA_NAME_DEF_STMT (src)) == PHI_NODE
698 && bb_for_stmt (SSA_NAME_DEF_STMT (src)) == e->dest)
699 return;
701 record_const_or_copy (dst, src);
704 /* Try to simplify each statement in E->dest, ultimately leading to
705 a simplification of the COND_EXPR at the end of E->dest.
707 We might consider marking just those statements which ultimately
708 feed the COND_EXPR. It's not clear if the overhead of bookkeeping
709 would be recovered by trying to simplify fewer statements.
711 If we are able to simplify a statement into the form
712 SSA_NAME = (SSA_NAME | gimple invariant), then we can record
713 a context sensitive equivalency which may help us simplify
714 later statements in E->dest.
716 Failure to simplify into the form above merely means that the
717 statement provides no equivalences to help simplify later
718 statements. This does not prevent threading through E->dest. */
719 max_stmt_count = PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS);
720 for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
722 tree cached_lhs = NULL;
724 stmt = bsi_stmt (bsi);
726 /* Ignore empty statements and labels. */
727 if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
728 continue;
730 /* If duplicating this block is going to cause too much code
731 expansion, then do not thread through this block. */
732 stmt_count++;
733 if (stmt_count > max_stmt_count)
734 return;
736 /* Safely handle threading across loop backedges. This is
737 over conservative, but still allows us to capture the
738 majority of the cases where we can thread across a loop
739 backedge. */
740 if ((e->flags & EDGE_DFS_BACK) != 0
741 && TREE_CODE (stmt) != COND_EXPR
742 && TREE_CODE (stmt) != SWITCH_EXPR)
743 return;
745 /* If the statement has volatile operands, then we assume we
746 can not thread through this block. This is overly
747 conservative in some ways. */
748 if (TREE_CODE (stmt) == ASM_EXPR && ASM_VOLATILE_P (stmt))
749 return;
751 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
752 value, then do not try to simplify this statement as it will
753 not simplify in any way that is helpful for jump threading. */
754 if (TREE_CODE (stmt) != MODIFY_EXPR
755 || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
756 continue;
758 /* At this point we have a statement which assigns an RHS to an
759 SSA_VAR on the LHS. We want to try and simplify this statement
760 to expose more context sensitive equivalences which in turn may
761 allow us to simplify the condition at the end of the loop. */
762 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
763 cached_lhs = TREE_OPERAND (stmt, 1);
764 else
766 /* Copy the operands. */
767 tree *copy, pre_fold_expr;
768 ssa_op_iter iter;
769 use_operand_p use_p;
770 unsigned int num, i = 0;
772 num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE));
773 copy = XCNEWVEC (tree, num);
775 /* Make a copy of the uses & vuses into USES_COPY, then cprop into
776 the operands. */
777 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
779 tree tmp = NULL;
780 tree use = USE_FROM_PTR (use_p);
782 copy[i++] = use;
783 if (TREE_CODE (use) == SSA_NAME)
784 tmp = SSA_NAME_VALUE (use);
785 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
786 SET_USE (use_p, tmp);
789 /* Try to fold/lookup the new expression. Inserting the
790 expression into the hash table is unlikely to help
791 Sadly, we have to handle conditional assignments specially
792 here, because fold expects all the operands of an expression
793 to be folded before the expression itself is folded, but we
794 can't just substitute the folded condition here. */
795 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == COND_EXPR)
797 tree cond = COND_EXPR_COND (TREE_OPERAND (stmt, 1));
798 cond = fold (cond);
799 if (cond == boolean_true_node)
800 pre_fold_expr = COND_EXPR_THEN (TREE_OPERAND (stmt, 1));
801 else if (cond == boolean_false_node)
802 pre_fold_expr = COND_EXPR_ELSE (TREE_OPERAND (stmt, 1));
803 else
804 pre_fold_expr = TREE_OPERAND (stmt, 1);
806 else
807 pre_fold_expr = TREE_OPERAND (stmt, 1);
809 if (pre_fold_expr)
811 cached_lhs = fold (pre_fold_expr);
812 if (TREE_CODE (cached_lhs) != SSA_NAME
813 && !is_gimple_min_invariant (cached_lhs))
814 cached_lhs = lookup_avail_expr (stmt, false);
817 /* Restore the statement's original uses/defs. */
818 i = 0;
819 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
820 SET_USE (use_p, copy[i++]);
822 free (copy);
825 /* Record the context sensitive equivalence if we were able
826 to simplify this statement. */
827 if (cached_lhs
828 && (TREE_CODE (cached_lhs) == SSA_NAME
829 || is_gimple_min_invariant (cached_lhs)))
830 record_const_or_copy (TREE_OPERAND (stmt, 0), cached_lhs);
833 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
834 will be taken. */
835 if (stmt
836 && (TREE_CODE (stmt) == COND_EXPR
837 || TREE_CODE (stmt) == GOTO_EXPR
838 || TREE_CODE (stmt) == SWITCH_EXPR))
840 tree cond, cached_lhs;
842 /* Now temporarily cprop the operands and try to find the resulting
843 expression in the hash tables. */
844 if (TREE_CODE (stmt) == COND_EXPR)
846 canonicalize_comparison (stmt);
847 cond = COND_EXPR_COND (stmt);
849 else if (TREE_CODE (stmt) == GOTO_EXPR)
850 cond = GOTO_DESTINATION (stmt);
851 else
852 cond = SWITCH_COND (stmt);
854 if (COMPARISON_CLASS_P (cond))
856 tree dummy_cond, op0, op1;
857 enum tree_code cond_code;
859 op0 = TREE_OPERAND (cond, 0);
860 op1 = TREE_OPERAND (cond, 1);
861 cond_code = TREE_CODE (cond);
863 /* Get the current value of both operands. */
864 if (TREE_CODE (op0) == SSA_NAME)
866 tree tmp = SSA_NAME_VALUE (op0);
867 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
868 op0 = tmp;
871 if (TREE_CODE (op1) == SSA_NAME)
873 tree tmp = SSA_NAME_VALUE (op1);
874 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
875 op1 = tmp;
878 /* Stuff the operator and operands into our dummy conditional
879 expression, creating the dummy conditional if necessary. */
880 dummy_cond = (tree) walk_data->global_data;
881 if (! dummy_cond)
883 dummy_cond = build2 (cond_code, boolean_type_node, op0, op1);
884 dummy_cond = build3 (COND_EXPR, void_type_node,
885 dummy_cond, NULL_TREE, NULL_TREE);
886 walk_data->global_data = dummy_cond;
888 else
890 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
891 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
892 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
895 /* If the conditional folds to an invariant, then we are done,
896 otherwise look it up in the hash tables. */
897 cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
898 if (! is_gimple_min_invariant (cached_lhs))
900 cached_lhs = lookup_avail_expr (dummy_cond, false);
901 if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
902 cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
903 NULL,
904 false);
907 /* We can have conditionals which just test the state of a
908 variable rather than use a relational operator. These are
909 simpler to handle. */
910 else if (TREE_CODE (cond) == SSA_NAME)
912 cached_lhs = cond;
913 cached_lhs = SSA_NAME_VALUE (cached_lhs);
914 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
915 cached_lhs = NULL;
917 else
918 cached_lhs = lookup_avail_expr (stmt, false);
920 if (cached_lhs)
922 edge taken_edge = find_taken_edge (e->dest, cached_lhs);
923 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
925 if (dest == e->dest)
926 return;
928 /* If we have a known destination for the conditional, then
929 we can perform this optimization, which saves at least one
930 conditional jump each time it applies since we get to
931 bypass the conditional at our original destination. */
932 if (dest)
934 struct edge_info *edge_info;
936 if (e->aux)
937 edge_info = (struct edge_info *) e->aux;
938 else
939 edge_info = allocate_edge_info (e);
940 edge_info->redirection_target = taken_edge;
941 bitmap_set_bit (threaded_blocks, e->dest->index);
948 /* Initialize local stacks for this optimizer and record equivalences
949 upon entry to BB. Equivalences can come from the edge traversed to
950 reach BB or they may come from PHI nodes at the start of BB. */
952 static void
953 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
954 basic_block bb)
956 if (dump_file && (dump_flags & TDF_DETAILS))
957 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
959 /* Push a marker on the stacks of local information so that we know how
960 far to unwind when we finalize this block. */
961 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
962 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
963 VEC_safe_push (tree, heap, nonzero_vars_stack, NULL_TREE);
964 VEC_safe_push (tree, heap, vrp_variables_stack, NULL_TREE);
966 record_equivalences_from_incoming_edge (bb);
968 /* PHI nodes can create equivalences too. */
969 record_equivalences_from_phis (bb);
972 /* Given an expression EXPR (a relational expression or a statement),
973 initialize the hash table element pointed to by ELEMENT. */
975 static void
976 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
978 /* Hash table elements may be based on conditional expressions or statements.
980 For the former case, we have no annotation and we want to hash the
981 conditional expression. In the latter case we have an annotation and
982 we want to record the expression the statement evaluates. */
983 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
985 element->stmt = NULL;
986 element->rhs = expr;
988 else if (TREE_CODE (expr) == COND_EXPR)
990 element->stmt = expr;
991 element->rhs = COND_EXPR_COND (expr);
993 else if (TREE_CODE (expr) == SWITCH_EXPR)
995 element->stmt = expr;
996 element->rhs = SWITCH_COND (expr);
998 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
1000 element->stmt = expr;
1001 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
1003 else if (TREE_CODE (expr) == GOTO_EXPR)
1005 element->stmt = expr;
1006 element->rhs = GOTO_DESTINATION (expr);
1008 else
1010 element->stmt = expr;
1011 element->rhs = TREE_OPERAND (expr, 1);
1014 element->lhs = lhs;
1015 element->hash = avail_expr_hash (element);
1018 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
1019 LIMIT entries left in LOCALs. */
1021 static void
1022 remove_local_expressions_from_table (void)
1024 /* Remove all the expressions made available in this block. */
1025 while (VEC_length (tree, avail_exprs_stack) > 0)
1027 struct expr_hash_elt element;
1028 tree expr = VEC_pop (tree, avail_exprs_stack);
1030 if (expr == NULL_TREE)
1031 break;
1033 initialize_hash_element (expr, NULL, &element);
1034 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
1038 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
1039 state, stopping when there are LIMIT entries left in LOCALs. */
1041 static void
1042 restore_nonzero_vars_to_original_value (void)
1044 while (VEC_length (tree, nonzero_vars_stack) > 0)
1046 tree name = VEC_pop (tree, nonzero_vars_stack);
1048 if (name == NULL)
1049 break;
1051 bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
1055 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
1056 CONST_AND_COPIES to its original state, stopping when we hit a
1057 NULL marker. */
1059 static void
1060 restore_vars_to_original_value (void)
1062 while (VEC_length (tree, const_and_copies_stack) > 0)
1064 tree prev_value, dest;
1066 dest = VEC_pop (tree, const_and_copies_stack);
1068 if (dest == NULL)
1069 break;
1071 prev_value = VEC_pop (tree, const_and_copies_stack);
1072 SSA_NAME_VALUE (dest) = prev_value;
1076 /* We have finished processing the dominator children of BB, perform
1077 any finalization actions in preparation for leaving this node in
1078 the dominator tree. */
1080 static void
1081 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
1083 tree last;
1085 /* If we have an outgoing edge to a block with multiple incoming and
1086 outgoing edges, then we may be able to thread the edge. ie, we
1087 may be able to statically determine which of the outgoing edges
1088 will be traversed when the incoming edge from BB is traversed. */
1089 if (single_succ_p (bb)
1090 && (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0
1091 && !single_pred_p (single_succ (bb))
1092 && !single_succ_p (single_succ (bb)))
1095 thread_across_edge (walk_data, single_succ_edge (bb));
1097 else if ((last = last_stmt (bb))
1098 && TREE_CODE (last) == COND_EXPR
1099 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
1100 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
1101 && EDGE_COUNT (bb->succs) == 2
1102 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
1103 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
1105 edge true_edge, false_edge;
1107 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
1109 /* Only try to thread the edge if it reaches a target block with
1110 more than one predecessor and more than one successor. */
1111 if (!single_pred_p (true_edge->dest) && !single_succ_p (true_edge->dest))
1113 struct edge_info *edge_info;
1114 unsigned int i;
1116 /* Push a marker onto the available expression stack so that we
1117 unwind any expressions related to the TRUE arm before processing
1118 the false arm below. */
1119 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
1120 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
1122 edge_info = (struct edge_info *) true_edge->aux;
1124 /* If we have info associated with this edge, record it into
1125 our equivalency tables. */
1126 if (edge_info)
1128 tree *cond_equivalences = edge_info->cond_equivalences;
1129 tree lhs = edge_info->lhs;
1130 tree rhs = edge_info->rhs;
1132 /* If we have a simple NAME = VALUE equivalency record it. */
1133 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1134 record_const_or_copy (lhs, rhs);
1136 /* If we have 0 = COND or 1 = COND equivalences, record them
1137 into our expression hash tables. */
1138 if (cond_equivalences)
1139 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1141 tree expr = cond_equivalences[i];
1142 tree value = cond_equivalences[i + 1];
1144 record_cond (expr, value);
1148 /* Now thread the edge. */
1149 thread_across_edge (walk_data, true_edge);
1151 /* And restore the various tables to their state before
1152 we threaded this edge. */
1153 remove_local_expressions_from_table ();
1154 restore_vars_to_original_value ();
1157 /* Similarly for the ELSE arm. */
1158 if (!single_pred_p (false_edge->dest) && !single_succ_p (false_edge->dest))
1160 struct edge_info *edge_info;
1161 unsigned int i;
1163 edge_info = (struct edge_info *) false_edge->aux;
1165 /* If we have info associated with this edge, record it into
1166 our equivalency tables. */
1167 if (edge_info)
1169 tree *cond_equivalences = edge_info->cond_equivalences;
1170 tree lhs = edge_info->lhs;
1171 tree rhs = edge_info->rhs;
1173 /* If we have a simple NAME = VALUE equivalency record it. */
1174 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1175 record_const_or_copy (lhs, rhs);
1177 /* If we have 0 = COND or 1 = COND equivalences, record them
1178 into our expression hash tables. */
1179 if (cond_equivalences)
1180 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1182 tree expr = cond_equivalences[i];
1183 tree value = cond_equivalences[i + 1];
1185 record_cond (expr, value);
1189 thread_across_edge (walk_data, false_edge);
1191 /* No need to remove local expressions from our tables
1192 or restore vars to their original value as that will
1193 be done immediately below. */
1197 remove_local_expressions_from_table ();
1198 restore_nonzero_vars_to_original_value ();
1199 restore_vars_to_original_value ();
1201 /* Remove VRP records associated with this basic block. They are no
1202 longer valid.
1204 To be efficient, we note which variables have had their values
1205 constrained in this block. So walk over each variable in the
1206 VRP_VARIABLEs array. */
1207 while (VEC_length (tree, vrp_variables_stack) > 0)
1209 tree var = VEC_pop (tree, vrp_variables_stack);
1210 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1211 void **slot;
1213 /* Each variable has a stack of value range records. We want to
1214 invalidate those associated with our basic block. So we walk
1215 the array backwards popping off records associated with our
1216 block. Once we hit a record not associated with our block
1217 we are done. */
1218 VEC(vrp_element_p,heap) **var_vrp_records;
1220 if (var == NULL)
1221 break;
1223 vrp_hash_elt.var = var;
1224 vrp_hash_elt.records = NULL;
1226 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1228 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1229 var_vrp_records = &vrp_hash_elt_p->records;
1231 while (VEC_length (vrp_element_p, *var_vrp_records) > 0)
1233 struct vrp_element *element
1234 = VEC_last (vrp_element_p, *var_vrp_records);
1236 if (element->bb != bb)
1237 break;
1239 VEC_pop (vrp_element_p, *var_vrp_records);
1243 /* If we queued any statements to rescan in this block, then
1244 go ahead and rescan them now. */
1245 while (VEC_length (tree, stmts_to_rescan) > 0)
1247 tree stmt = VEC_last (tree, stmts_to_rescan);
1248 basic_block stmt_bb = bb_for_stmt (stmt);
1250 if (stmt_bb != bb)
1251 break;
1253 VEC_pop (tree, stmts_to_rescan);
1254 mark_new_vars_to_rename (stmt);
1258 /* PHI nodes can create equivalences too.
1260 Ignoring any alternatives which are the same as the result, if
1261 all the alternatives are equal, then the PHI node creates an
1262 equivalence.
1264 Additionally, if all the PHI alternatives are known to have a nonzero
1265 value, then the result of this PHI is known to have a nonzero value,
1266 even if we do not know its exact value. */
1268 static void
1269 record_equivalences_from_phis (basic_block bb)
1271 tree phi;
1273 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1275 tree lhs = PHI_RESULT (phi);
1276 tree rhs = NULL;
1277 int i;
1279 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1281 tree t = PHI_ARG_DEF (phi, i);
1283 /* Ignore alternatives which are the same as our LHS. Since
1284 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1285 can simply compare pointers. */
1286 if (lhs == t)
1287 continue;
1289 /* If we have not processed an alternative yet, then set
1290 RHS to this alternative. */
1291 if (rhs == NULL)
1292 rhs = t;
1293 /* If we have processed an alternative (stored in RHS), then
1294 see if it is equal to this one. If it isn't, then stop
1295 the search. */
1296 else if (! operand_equal_for_phi_arg_p (rhs, t))
1297 break;
1300 /* If we had no interesting alternatives, then all the RHS alternatives
1301 must have been the same as LHS. */
1302 if (!rhs)
1303 rhs = lhs;
1305 /* If we managed to iterate through each PHI alternative without
1306 breaking out of the loop, then we have a PHI which may create
1307 a useful equivalence. We do not need to record unwind data for
1308 this, since this is a true assignment and not an equivalence
1309 inferred from a comparison. All uses of this ssa name are dominated
1310 by this assignment, so unwinding just costs time and space. */
1311 if (i == PHI_NUM_ARGS (phi)
1312 && may_propagate_copy (lhs, rhs))
1313 SSA_NAME_VALUE (lhs) = rhs;
1315 /* Now see if we know anything about the nonzero property for the
1316 result of this PHI. */
1317 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1319 if (!PHI_ARG_NONZERO (phi, i))
1320 break;
1323 if (i == PHI_NUM_ARGS (phi))
1324 bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
1328 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1329 return that edge. Otherwise return NULL. */
1330 static edge
1331 single_incoming_edge_ignoring_loop_edges (basic_block bb)
1333 edge retval = NULL;
1334 edge e;
1335 edge_iterator ei;
1337 FOR_EACH_EDGE (e, ei, bb->preds)
1339 /* A loop back edge can be identified by the destination of
1340 the edge dominating the source of the edge. */
1341 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
1342 continue;
1344 /* If we have already seen a non-loop edge, then we must have
1345 multiple incoming non-loop edges and thus we return NULL. */
1346 if (retval)
1347 return NULL;
1349 /* This is the first non-loop incoming edge we have found. Record
1350 it. */
1351 retval = e;
1354 return retval;
1357 /* Record any equivalences created by the incoming edge to BB. If BB
1358 has more than one incoming edge, then no equivalence is created. */
1360 static void
1361 record_equivalences_from_incoming_edge (basic_block bb)
1363 edge e;
1364 basic_block parent;
1365 struct edge_info *edge_info;
1367 /* If our parent block ended with a control statement, then we may be
1368 able to record some equivalences based on which outgoing edge from
1369 the parent was followed. */
1370 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
1372 e = single_incoming_edge_ignoring_loop_edges (bb);
1374 /* If we had a single incoming edge from our parent block, then enter
1375 any data associated with the edge into our tables. */
1376 if (e && e->src == parent)
1378 unsigned int i;
1380 edge_info = (struct edge_info *) e->aux;
1382 if (edge_info)
1384 tree lhs = edge_info->lhs;
1385 tree rhs = edge_info->rhs;
1386 tree *cond_equivalences = edge_info->cond_equivalences;
1388 if (lhs)
1389 record_equality (lhs, rhs);
1391 if (cond_equivalences)
1393 bool recorded_range = false;
1394 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1396 tree expr = cond_equivalences[i];
1397 tree value = cond_equivalences[i + 1];
1399 record_cond (expr, value);
1401 /* For the first true equivalence, record range
1402 information. We only do this for the first
1403 true equivalence as it should dominate any
1404 later true equivalences. */
1405 if (! recorded_range
1406 && COMPARISON_CLASS_P (expr)
1407 && value == boolean_true_node
1408 && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
1410 record_range (expr, bb);
1411 recorded_range = true;
1419 /* Dump SSA statistics on FILE. */
1421 void
1422 dump_dominator_optimization_stats (FILE *file)
1424 long n_exprs;
1426 fprintf (file, "Total number of statements: %6ld\n\n",
1427 opt_stats.num_stmts);
1428 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
1429 opt_stats.num_exprs_considered);
1431 n_exprs = opt_stats.num_exprs_considered;
1432 if (n_exprs == 0)
1433 n_exprs = 1;
1435 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1436 opt_stats.num_re, PERCENT (opt_stats.num_re,
1437 n_exprs));
1438 fprintf (file, " Constants propagated: %6ld\n",
1439 opt_stats.num_const_prop);
1440 fprintf (file, " Copies propagated: %6ld\n",
1441 opt_stats.num_copy_prop);
1443 fprintf (file, "\nTotal number of DOM iterations: %6ld\n",
1444 opt_stats.num_iterations);
1446 fprintf (file, "\nHash table statistics:\n");
1448 fprintf (file, " avail_exprs: ");
1449 htab_statistics (file, avail_exprs);
1453 /* Dump SSA statistics on stderr. */
1455 void
1456 debug_dominator_optimization_stats (void)
1458 dump_dominator_optimization_stats (stderr);
1462 /* Dump statistics for the hash table HTAB. */
1464 static void
1465 htab_statistics (FILE *file, htab_t htab)
1467 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
1468 (long) htab_size (htab),
1469 (long) htab_elements (htab),
1470 htab_collisions (htab));
1473 /* Record the fact that VAR has a nonzero value, though we may not know
1474 its exact value. Note that if VAR is already known to have a nonzero
1475 value, then we do nothing. */
1477 static void
1478 record_var_is_nonzero (tree var)
1480 int indx = SSA_NAME_VERSION (var);
1482 if (bitmap_bit_p (nonzero_vars, indx))
1483 return;
1485 /* Mark it in the global table. */
1486 bitmap_set_bit (nonzero_vars, indx);
1488 /* Record this SSA_NAME so that we can reset the global table
1489 when we leave this block. */
1490 VEC_safe_push (tree, heap, nonzero_vars_stack, var);
1493 /* Enter a statement into the true/false expression hash table indicating
1494 that the condition COND has the value VALUE. */
1496 static void
1497 record_cond (tree cond, tree value)
1499 struct expr_hash_elt *element = XCNEW (struct expr_hash_elt);
1500 void **slot;
1502 initialize_hash_element (cond, value, element);
1504 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
1505 element->hash, INSERT);
1506 if (*slot == NULL)
1508 *slot = (void *) element;
1509 VEC_safe_push (tree, heap, avail_exprs_stack, cond);
1511 else
1512 free (element);
1515 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1516 the new conditional into *p, then store a boolean_true_node
1517 into *(p + 1). */
1519 static void
1520 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
1522 *p = build2 (new_code, boolean_type_node, op0, op1);
1523 p++;
1524 *p = boolean_true_node;
1527 /* Record that COND is true and INVERTED is false into the edge information
1528 structure. Also record that any conditions dominated by COND are true
1529 as well.
1531 For example, if a < b is true, then a <= b must also be true. */
1533 static void
1534 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
1536 tree op0, op1;
1538 if (!COMPARISON_CLASS_P (cond))
1539 return;
1541 op0 = TREE_OPERAND (cond, 0);
1542 op1 = TREE_OPERAND (cond, 1);
1544 switch (TREE_CODE (cond))
1546 case LT_EXPR:
1547 case GT_EXPR:
1548 edge_info->max_cond_equivalences = 12;
1549 edge_info->cond_equivalences = XNEWVEC (tree, 12);
1550 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
1551 ? LE_EXPR : GE_EXPR),
1552 op0, op1, &edge_info->cond_equivalences[4]);
1553 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1554 &edge_info->cond_equivalences[6]);
1555 build_and_record_new_cond (NE_EXPR, op0, op1,
1556 &edge_info->cond_equivalences[8]);
1557 build_and_record_new_cond (LTGT_EXPR, op0, op1,
1558 &edge_info->cond_equivalences[10]);
1559 break;
1561 case GE_EXPR:
1562 case LE_EXPR:
1563 edge_info->max_cond_equivalences = 6;
1564 edge_info->cond_equivalences = XNEWVEC (tree, 6);
1565 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1566 &edge_info->cond_equivalences[4]);
1567 break;
1569 case EQ_EXPR:
1570 edge_info->max_cond_equivalences = 10;
1571 edge_info->cond_equivalences = XNEWVEC (tree, 10);
1572 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1573 &edge_info->cond_equivalences[4]);
1574 build_and_record_new_cond (LE_EXPR, op0, op1,
1575 &edge_info->cond_equivalences[6]);
1576 build_and_record_new_cond (GE_EXPR, op0, op1,
1577 &edge_info->cond_equivalences[8]);
1578 break;
1580 case UNORDERED_EXPR:
1581 edge_info->max_cond_equivalences = 16;
1582 edge_info->cond_equivalences = XNEWVEC (tree, 16);
1583 build_and_record_new_cond (NE_EXPR, op0, op1,
1584 &edge_info->cond_equivalences[4]);
1585 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1586 &edge_info->cond_equivalences[6]);
1587 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1588 &edge_info->cond_equivalences[8]);
1589 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
1590 &edge_info->cond_equivalences[10]);
1591 build_and_record_new_cond (UNLT_EXPR, op0, op1,
1592 &edge_info->cond_equivalences[12]);
1593 build_and_record_new_cond (UNGT_EXPR, op0, op1,
1594 &edge_info->cond_equivalences[14]);
1595 break;
1597 case UNLT_EXPR:
1598 case UNGT_EXPR:
1599 edge_info->max_cond_equivalences = 8;
1600 edge_info->cond_equivalences = XNEWVEC (tree, 8);
1601 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1602 ? UNLE_EXPR : UNGE_EXPR),
1603 op0, op1, &edge_info->cond_equivalences[4]);
1604 build_and_record_new_cond (NE_EXPR, op0, op1,
1605 &edge_info->cond_equivalences[6]);
1606 break;
1608 case UNEQ_EXPR:
1609 edge_info->max_cond_equivalences = 8;
1610 edge_info->cond_equivalences = XNEWVEC (tree, 8);
1611 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1612 &edge_info->cond_equivalences[4]);
1613 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1614 &edge_info->cond_equivalences[6]);
1615 break;
1617 case LTGT_EXPR:
1618 edge_info->max_cond_equivalences = 8;
1619 edge_info->cond_equivalences = XNEWVEC (tree, 8);
1620 build_and_record_new_cond (NE_EXPR, op0, op1,
1621 &edge_info->cond_equivalences[4]);
1622 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1623 &edge_info->cond_equivalences[6]);
1624 break;
1626 default:
1627 edge_info->max_cond_equivalences = 4;
1628 edge_info->cond_equivalences = XNEWVEC (tree, 4);
1629 break;
1632 /* Now store the original true and false conditions into the first
1633 two slots. */
1634 edge_info->cond_equivalences[0] = cond;
1635 edge_info->cond_equivalences[1] = boolean_true_node;
1636 edge_info->cond_equivalences[2] = inverted;
1637 edge_info->cond_equivalences[3] = boolean_false_node;
1640 /* A helper function for record_const_or_copy and record_equality.
1641 Do the work of recording the value and undo info. */
1643 static void
1644 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1646 SSA_NAME_VALUE (x) = y;
1648 VEC_reserve (tree, heap, const_and_copies_stack, 2);
1649 VEC_quick_push (tree, const_and_copies_stack, prev_x);
1650 VEC_quick_push (tree, const_and_copies_stack, x);
1654 /* Return the loop depth of the basic block of the defining statement of X.
1655 This number should not be treated as absolutely correct because the loop
1656 information may not be completely up-to-date when dom runs. However, it
1657 will be relatively correct, and as more passes are taught to keep loop info
1658 up to date, the result will become more and more accurate. */
1661 loop_depth_of_name (tree x)
1663 tree defstmt;
1664 basic_block defbb;
1666 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1667 if (TREE_CODE (x) != SSA_NAME)
1668 return 0;
1670 /* Otherwise return the loop depth of the defining statement's bb.
1671 Note that there may not actually be a bb for this statement, if the
1672 ssa_name is live on entry. */
1673 defstmt = SSA_NAME_DEF_STMT (x);
1674 defbb = bb_for_stmt (defstmt);
1675 if (!defbb)
1676 return 0;
1678 return defbb->loop_depth;
1682 /* Record that X is equal to Y in const_and_copies. Record undo
1683 information in the block-local vector. */
1685 static void
1686 record_const_or_copy (tree x, tree y)
1688 tree prev_x = SSA_NAME_VALUE (x);
1690 if (TREE_CODE (y) == SSA_NAME)
1692 tree tmp = SSA_NAME_VALUE (y);
1693 if (tmp)
1694 y = tmp;
1697 record_const_or_copy_1 (x, y, prev_x);
1700 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1701 This constrains the cases in which we may treat this as assignment. */
1703 static void
1704 record_equality (tree x, tree y)
1706 tree prev_x = NULL, prev_y = NULL;
1708 if (TREE_CODE (x) == SSA_NAME)
1709 prev_x = SSA_NAME_VALUE (x);
1710 if (TREE_CODE (y) == SSA_NAME)
1711 prev_y = SSA_NAME_VALUE (y);
1713 /* If one of the previous values is invariant, or invariant in more loops
1714 (by depth), then use that.
1715 Otherwise it doesn't matter which value we choose, just so
1716 long as we canonicalize on one value. */
1717 if (TREE_INVARIANT (y))
1719 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1720 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1721 else if (prev_x && TREE_INVARIANT (prev_x))
1722 x = y, y = prev_x, prev_x = prev_y;
1723 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1724 y = prev_y;
1726 /* After the swapping, we must have one SSA_NAME. */
1727 if (TREE_CODE (x) != SSA_NAME)
1728 return;
1730 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1731 variable compared against zero. If we're honoring signed zeros,
1732 then we cannot record this value unless we know that the value is
1733 nonzero. */
1734 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1735 && (TREE_CODE (y) != REAL_CST
1736 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1737 return;
1739 record_const_or_copy_1 (x, y, prev_x);
1742 /* Return true, if it is ok to do folding of an associative expression.
1743 EXP is the tree for the associative expression. */
1745 static inline bool
1746 unsafe_associative_fp_binop (tree exp)
1748 enum tree_code code = TREE_CODE (exp);
1749 return !(!flag_unsafe_math_optimizations
1750 && (code == MULT_EXPR || code == PLUS_EXPR
1751 || code == MINUS_EXPR)
1752 && FLOAT_TYPE_P (TREE_TYPE (exp)));
1755 /* Returns true when STMT is a simple iv increment. It detects the
1756 following situation:
1758 i_1 = phi (..., i_2)
1759 i_2 = i_1 +/- ... */
1761 static bool
1762 simple_iv_increment_p (tree stmt)
1764 tree lhs, rhs, preinc, phi;
1765 unsigned i;
1767 if (TREE_CODE (stmt) != MODIFY_EXPR)
1768 return false;
1770 lhs = TREE_OPERAND (stmt, 0);
1771 if (TREE_CODE (lhs) != SSA_NAME)
1772 return false;
1774 rhs = TREE_OPERAND (stmt, 1);
1776 if (TREE_CODE (rhs) != PLUS_EXPR
1777 && TREE_CODE (rhs) != MINUS_EXPR)
1778 return false;
1780 preinc = TREE_OPERAND (rhs, 0);
1781 if (TREE_CODE (preinc) != SSA_NAME)
1782 return false;
1784 phi = SSA_NAME_DEF_STMT (preinc);
1785 if (TREE_CODE (phi) != PHI_NODE)
1786 return false;
1788 for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++)
1789 if (PHI_ARG_DEF (phi, i) == lhs)
1790 return true;
1792 return false;
1795 /* COND is a condition of the form:
1797 x == const or x != const
1799 Look back to x's defining statement and see if x is defined as
1801 x = (type) y;
1803 If const is unchanged if we convert it to type, then we can build
1804 the equivalent expression:
1807 y == const or y != const
1809 Which may allow further optimizations.
1811 Return the equivalent comparison or NULL if no such equivalent comparison
1812 was found. */
1814 static tree
1815 find_equivalent_equality_comparison (tree cond)
1817 tree op0 = TREE_OPERAND (cond, 0);
1818 tree op1 = TREE_OPERAND (cond, 1);
1819 tree def_stmt = SSA_NAME_DEF_STMT (op0);
1821 /* OP0 might have been a parameter, so first make sure it
1822 was defined by a MODIFY_EXPR. */
1823 if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
1825 tree def_rhs = TREE_OPERAND (def_stmt, 1);
1828 /* If either operand to the comparison is a pointer to
1829 a function, then we can not apply this optimization
1830 as some targets require function pointers to be
1831 canonicalized and in this case this optimization would
1832 eliminate a necessary canonicalization. */
1833 if ((POINTER_TYPE_P (TREE_TYPE (op0))
1834 && TREE_CODE (TREE_TYPE (TREE_TYPE (op0))) == FUNCTION_TYPE)
1835 || (POINTER_TYPE_P (TREE_TYPE (op1))
1836 && TREE_CODE (TREE_TYPE (TREE_TYPE (op1))) == FUNCTION_TYPE))
1837 return NULL;
1839 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1840 if ((TREE_CODE (def_rhs) == NOP_EXPR
1841 || TREE_CODE (def_rhs) == CONVERT_EXPR)
1842 && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
1844 tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
1845 tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
1846 tree new;
1848 if (TYPE_PRECISION (def_rhs_inner_type)
1849 > TYPE_PRECISION (TREE_TYPE (def_rhs)))
1850 return NULL;
1852 /* If the inner type of the conversion is a pointer to
1853 a function, then we can not apply this optimization
1854 as some targets require function pointers to be
1855 canonicalized. This optimization would result in
1856 canonicalization of the pointer when it was not originally
1857 needed/intended. */
1858 if (POINTER_TYPE_P (def_rhs_inner_type)
1859 && TREE_CODE (TREE_TYPE (def_rhs_inner_type)) == FUNCTION_TYPE)
1860 return NULL;
1862 /* What we want to prove is that if we convert OP1 to
1863 the type of the object inside the NOP_EXPR that the
1864 result is still equivalent to SRC.
1866 If that is true, the build and return new equivalent
1867 condition which uses the source of the typecast and the
1868 new constant (which has only changed its type). */
1869 new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
1870 new = local_fold (new);
1871 if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
1872 return build2 (TREE_CODE (cond), TREE_TYPE (cond),
1873 def_rhs_inner, new);
1876 return NULL;
1879 /* STMT is a COND_EXPR for which we could not trivially determine its
1880 result. This routine attempts to find equivalent forms of the
1881 condition which we may be able to optimize better. It also
1882 uses simple value range propagation to optimize conditionals. */
1884 static tree
1885 simplify_cond_and_lookup_avail_expr (tree stmt,
1886 stmt_ann_t ann,
1887 int insert)
1889 tree cond = COND_EXPR_COND (stmt);
1891 if (COMPARISON_CLASS_P (cond))
1893 tree op0 = TREE_OPERAND (cond, 0);
1894 tree op1 = TREE_OPERAND (cond, 1);
1896 if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
1898 int limit;
1899 tree low, high, cond_low, cond_high;
1900 int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
1901 VEC(vrp_element_p,heap) **vrp_records;
1902 struct vrp_element *element;
1903 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1904 void **slot;
1906 /* First see if we have test of an SSA_NAME against a constant
1907 where the SSA_NAME is defined by an earlier typecast which
1908 is irrelevant when performing tests against the given
1909 constant. */
1910 if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
1912 tree new_cond = find_equivalent_equality_comparison (cond);
1914 if (new_cond)
1916 /* Update the statement to use the new equivalent
1917 condition. */
1918 COND_EXPR_COND (stmt) = new_cond;
1920 /* If this is not a real stmt, ann will be NULL and we
1921 avoid processing the operands. */
1922 if (ann)
1923 mark_stmt_modified (stmt);
1925 /* Lookup the condition and return its known value if it
1926 exists. */
1927 new_cond = lookup_avail_expr (stmt, insert);
1928 if (new_cond)
1929 return new_cond;
1931 /* The operands have changed, so update op0 and op1. */
1932 op0 = TREE_OPERAND (cond, 0);
1933 op1 = TREE_OPERAND (cond, 1);
1937 /* Consult the value range records for this variable (if they exist)
1938 to see if we can eliminate or simplify this conditional.
1940 Note two tests are necessary to determine no records exist.
1941 First we have to see if the virtual array exists, if it
1942 exists, then we have to check its active size.
1944 Also note the vast majority of conditionals are not testing
1945 a variable which has had its range constrained by an earlier
1946 conditional. So this filter avoids a lot of unnecessary work. */
1947 vrp_hash_elt.var = op0;
1948 vrp_hash_elt.records = NULL;
1949 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1950 if (slot == NULL)
1951 return NULL;
1953 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1954 vrp_records = &vrp_hash_elt_p->records;
1956 limit = VEC_length (vrp_element_p, *vrp_records);
1958 /* If we have no value range records for this variable, or we are
1959 unable to extract a range for this condition, then there is
1960 nothing to do. */
1961 if (limit == 0
1962 || ! extract_range_from_cond (cond, &cond_high,
1963 &cond_low, &cond_inverted))
1964 return NULL;
1966 /* We really want to avoid unnecessary computations of range
1967 info. So all ranges are computed lazily; this avoids a
1968 lot of unnecessary work. i.e., we record the conditional,
1969 but do not process how it constrains the variable's
1970 potential values until we know that processing the condition
1971 could be helpful.
1973 However, we do not want to have to walk a potentially long
1974 list of ranges, nor do we want to compute a variable's
1975 range more than once for a given path.
1977 Luckily, each time we encounter a conditional that can not
1978 be otherwise optimized we will end up here and we will
1979 compute the necessary range information for the variable
1980 used in this condition.
1982 Thus you can conclude that there will never be more than one
1983 conditional associated with a variable which has not been
1984 processed. So we never need to merge more than one new
1985 conditional into the current range.
1987 These properties also help us avoid unnecessary work. */
1988 element = VEC_last (vrp_element_p, *vrp_records);
1990 if (element->high && element->low)
1992 /* The last element has been processed, so there is no range
1993 merging to do, we can simply use the high/low values
1994 recorded in the last element. */
1995 low = element->low;
1996 high = element->high;
1998 else
2000 tree tmp_high, tmp_low;
2001 int dummy;
2003 /* The last element has not been processed. Process it now.
2004 record_range should ensure for cond inverted is not set.
2005 This call can only fail if cond is x < min or x > max,
2006 which fold should have optimized into false.
2007 If that doesn't happen, just pretend all values are
2008 in the range. */
2009 if (! extract_range_from_cond (element->cond, &tmp_high,
2010 &tmp_low, &dummy))
2011 gcc_unreachable ();
2012 else
2013 gcc_assert (dummy == 0);
2015 /* If this is the only element, then no merging is necessary,
2016 the high/low values from extract_range_from_cond are all
2017 we need. */
2018 if (limit == 1)
2020 low = tmp_low;
2021 high = tmp_high;
2023 else
2025 /* Get the high/low value from the previous element. */
2026 struct vrp_element *prev
2027 = VEC_index (vrp_element_p, *vrp_records, limit - 2);
2028 low = prev->low;
2029 high = prev->high;
2031 /* Merge in this element's range with the range from the
2032 previous element.
2034 The low value for the merged range is the maximum of
2035 the previous low value and the low value of this record.
2037 Similarly the high value for the merged range is the
2038 minimum of the previous high value and the high value of
2039 this record. */
2040 low = (low && tree_int_cst_compare (low, tmp_low) == 1
2041 ? low : tmp_low);
2042 high = (high && tree_int_cst_compare (high, tmp_high) == -1
2043 ? high : tmp_high);
2046 /* And record the computed range. */
2047 element->low = low;
2048 element->high = high;
2052 /* After we have constrained this variable's potential values,
2053 we try to determine the result of the given conditional.
2055 To simplify later tests, first determine if the current
2056 low value is the same low value as the conditional.
2057 Similarly for the current high value and the high value
2058 for the conditional. */
2059 lowequal = tree_int_cst_equal (low, cond_low);
2060 highequal = tree_int_cst_equal (high, cond_high);
2062 if (lowequal && highequal)
2063 return (cond_inverted ? boolean_false_node : boolean_true_node);
2065 /* To simplify the overlap/subset tests below we may want
2066 to swap the two ranges so that the larger of the two
2067 ranges occurs "first". */
2068 swapped = 0;
2069 if (tree_int_cst_compare (low, cond_low) == 1
2070 || (lowequal
2071 && tree_int_cst_compare (cond_high, high) == 1))
2073 tree temp;
2075 swapped = 1;
2076 temp = low;
2077 low = cond_low;
2078 cond_low = temp;
2079 temp = high;
2080 high = cond_high;
2081 cond_high = temp;
2084 /* Now determine if there is no overlap in the ranges
2085 or if the second range is a subset of the first range. */
2086 no_overlap = tree_int_cst_lt (high, cond_low);
2087 subset = tree_int_cst_compare (cond_high, high) != 1;
2089 /* If there was no overlap in the ranges, then this conditional
2090 always has a false value (unless we had to invert this
2091 conditional, in which case it always has a true value). */
2092 if (no_overlap)
2093 return (cond_inverted ? boolean_true_node : boolean_false_node);
2095 /* If the current range is a subset of the condition's range,
2096 then this conditional always has a true value (unless we
2097 had to invert this conditional, in which case it always
2098 has a true value). */
2099 if (subset && swapped)
2100 return (cond_inverted ? boolean_false_node : boolean_true_node);
2102 /* We were unable to determine the result of the conditional.
2103 However, we may be able to simplify the conditional. First
2104 merge the ranges in the same manner as range merging above. */
2105 low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
2106 high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
2108 /* If the range has converged to a single point, then turn this
2109 into an equality comparison. */
2110 if (TREE_CODE (cond) != EQ_EXPR
2111 && TREE_CODE (cond) != NE_EXPR
2112 && tree_int_cst_equal (low, high))
2114 TREE_SET_CODE (cond, EQ_EXPR);
2115 TREE_OPERAND (cond, 1) = high;
2119 return 0;
2122 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2123 known value for that SSA_NAME (or NULL if no value is known).
2125 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2126 even if we don't know their precise value.
2128 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2129 nodes of the successors of BB. */
2131 static void
2132 cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
2134 edge e;
2135 edge_iterator ei;
2137 FOR_EACH_EDGE (e, ei, bb->succs)
2139 tree phi;
2140 int indx;
2142 /* If this is an abnormal edge, then we do not want to copy propagate
2143 into the PHI alternative associated with this edge. */
2144 if (e->flags & EDGE_ABNORMAL)
2145 continue;
2147 phi = phi_nodes (e->dest);
2148 if (! phi)
2149 continue;
2151 indx = e->dest_idx;
2152 for ( ; phi; phi = PHI_CHAIN (phi))
2154 tree new;
2155 use_operand_p orig_p;
2156 tree orig;
2158 /* The alternative may be associated with a constant, so verify
2159 it is an SSA_NAME before doing anything with it. */
2160 orig_p = PHI_ARG_DEF_PTR (phi, indx);
2161 orig = USE_FROM_PTR (orig_p);
2162 if (TREE_CODE (orig) != SSA_NAME)
2163 continue;
2165 /* If the alternative is known to have a nonzero value, record
2166 that fact in the PHI node itself for future use. */
2167 if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
2168 PHI_ARG_NONZERO (phi, indx) = true;
2170 /* If we have *ORIG_P in our constant/copy table, then replace
2171 ORIG_P with its value in our constant/copy table. */
2172 new = SSA_NAME_VALUE (orig);
2173 if (new
2174 && new != orig
2175 && (TREE_CODE (new) == SSA_NAME
2176 || is_gimple_min_invariant (new))
2177 && may_propagate_copy (orig, new))
2178 propagate_value (orig_p, new);
2183 /* We have finished optimizing BB, record any information implied by
2184 taking a specific outgoing edge from BB. */
2186 static void
2187 record_edge_info (basic_block bb)
2189 block_stmt_iterator bsi = bsi_last (bb);
2190 struct edge_info *edge_info;
2192 if (! bsi_end_p (bsi))
2194 tree stmt = bsi_stmt (bsi);
2196 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
2198 tree cond = SWITCH_COND (stmt);
2200 if (TREE_CODE (cond) == SSA_NAME)
2202 tree labels = SWITCH_LABELS (stmt);
2203 int i, n_labels = TREE_VEC_LENGTH (labels);
2204 tree *info = XCNEWVEC (tree, last_basic_block);
2205 edge e;
2206 edge_iterator ei;
2208 for (i = 0; i < n_labels; i++)
2210 tree label = TREE_VEC_ELT (labels, i);
2211 basic_block target_bb = label_to_block (CASE_LABEL (label));
2213 if (CASE_HIGH (label)
2214 || !CASE_LOW (label)
2215 || info[target_bb->index])
2216 info[target_bb->index] = error_mark_node;
2217 else
2218 info[target_bb->index] = label;
2221 FOR_EACH_EDGE (e, ei, bb->succs)
2223 basic_block target_bb = e->dest;
2224 tree node = info[target_bb->index];
2226 if (node != NULL && node != error_mark_node)
2228 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
2229 edge_info = allocate_edge_info (e);
2230 edge_info->lhs = cond;
2231 edge_info->rhs = x;
2234 free (info);
2238 /* A COND_EXPR may create equivalences too. */
2239 if (stmt && TREE_CODE (stmt) == COND_EXPR)
2241 tree cond = COND_EXPR_COND (stmt);
2242 edge true_edge;
2243 edge false_edge;
2245 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
2247 /* If the conditional is a single variable 'X', record 'X = 1'
2248 for the true edge and 'X = 0' on the false edge. */
2249 if (SSA_VAR_P (cond))
2251 struct edge_info *edge_info;
2253 edge_info = allocate_edge_info (true_edge);
2254 edge_info->lhs = cond;
2255 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
2257 edge_info = allocate_edge_info (false_edge);
2258 edge_info->lhs = cond;
2259 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
2261 /* Equality tests may create one or two equivalences. */
2262 else if (COMPARISON_CLASS_P (cond))
2264 tree op0 = TREE_OPERAND (cond, 0);
2265 tree op1 = TREE_OPERAND (cond, 1);
2267 /* Special case comparing booleans against a constant as we
2268 know the value of OP0 on both arms of the branch. i.e., we
2269 can record an equivalence for OP0 rather than COND. */
2270 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2271 && TREE_CODE (op0) == SSA_NAME
2272 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
2273 && is_gimple_min_invariant (op1))
2275 if (TREE_CODE (cond) == EQ_EXPR)
2277 edge_info = allocate_edge_info (true_edge);
2278 edge_info->lhs = op0;
2279 edge_info->rhs = (integer_zerop (op1)
2280 ? boolean_false_node
2281 : boolean_true_node);
2283 edge_info = allocate_edge_info (false_edge);
2284 edge_info->lhs = op0;
2285 edge_info->rhs = (integer_zerop (op1)
2286 ? boolean_true_node
2287 : boolean_false_node);
2289 else
2291 edge_info = allocate_edge_info (true_edge);
2292 edge_info->lhs = op0;
2293 edge_info->rhs = (integer_zerop (op1)
2294 ? boolean_true_node
2295 : boolean_false_node);
2297 edge_info = allocate_edge_info (false_edge);
2298 edge_info->lhs = op0;
2299 edge_info->rhs = (integer_zerop (op1)
2300 ? boolean_false_node
2301 : boolean_true_node);
2305 else if (is_gimple_min_invariant (op0)
2306 && (TREE_CODE (op1) == SSA_NAME
2307 || is_gimple_min_invariant (op1)))
2309 tree inverted = invert_truthvalue (cond);
2310 struct edge_info *edge_info;
2312 edge_info = allocate_edge_info (true_edge);
2313 record_conditions (edge_info, cond, inverted);
2315 if (TREE_CODE (cond) == EQ_EXPR)
2317 edge_info->lhs = op1;
2318 edge_info->rhs = op0;
2321 edge_info = allocate_edge_info (false_edge);
2322 record_conditions (edge_info, inverted, cond);
2324 if (TREE_CODE (cond) == NE_EXPR)
2326 edge_info->lhs = op1;
2327 edge_info->rhs = op0;
2331 else if (TREE_CODE (op0) == SSA_NAME
2332 && (is_gimple_min_invariant (op1)
2333 || TREE_CODE (op1) == SSA_NAME))
2335 tree inverted = invert_truthvalue (cond);
2336 struct edge_info *edge_info;
2338 edge_info = allocate_edge_info (true_edge);
2339 record_conditions (edge_info, cond, inverted);
2341 if (TREE_CODE (cond) == EQ_EXPR)
2343 edge_info->lhs = op0;
2344 edge_info->rhs = op1;
2347 edge_info = allocate_edge_info (false_edge);
2348 record_conditions (edge_info, inverted, cond);
2350 if (TREE_CODE (cond) == NE_EXPR)
2352 edge_info->lhs = op0;
2353 edge_info->rhs = op1;
2358 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2363 /* Propagate information from BB to its outgoing edges.
2365 This can include equivalency information implied by control statements
2366 at the end of BB and const/copy propagation into PHIs in BB's
2367 successor blocks. */
2369 static void
2370 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2371 basic_block bb)
2373 record_edge_info (bb);
2374 cprop_into_successor_phis (bb, nonzero_vars);
2377 /* Search for redundant computations in STMT. If any are found, then
2378 replace them with the variable holding the result of the computation.
2380 If safe, record this expression into the available expression hash
2381 table. */
2383 static bool
2384 eliminate_redundant_computations (tree stmt, stmt_ann_t ann)
2386 tree *expr_p, def = NULL_TREE;
2387 bool insert = true;
2388 tree cached_lhs;
2389 bool retval = false;
2390 bool modify_expr_p = false;
2392 if (TREE_CODE (stmt) == MODIFY_EXPR)
2393 def = TREE_OPERAND (stmt, 0);
2395 /* Certain expressions on the RHS can be optimized away, but can not
2396 themselves be entered into the hash tables. */
2397 if (! def
2398 || TREE_CODE (def) != SSA_NAME
2399 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
2400 || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VMAYDEF)
2401 /* Do not record equivalences for increments of ivs. This would create
2402 overlapping live ranges for a very questionable gain. */
2403 || simple_iv_increment_p (stmt))
2404 insert = false;
2406 /* Check if the expression has been computed before. */
2407 cached_lhs = lookup_avail_expr (stmt, insert);
2409 /* If this is a COND_EXPR and we did not find its expression in
2410 the hash table, simplify the condition and try again. */
2411 if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
2412 cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
2414 opt_stats.num_exprs_considered++;
2416 /* Get a pointer to the expression we are trying to optimize. */
2417 if (TREE_CODE (stmt) == COND_EXPR)
2418 expr_p = &COND_EXPR_COND (stmt);
2419 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2420 expr_p = &SWITCH_COND (stmt);
2421 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
2423 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
2424 modify_expr_p = true;
2426 else
2428 expr_p = &TREE_OPERAND (stmt, 1);
2429 modify_expr_p = true;
2432 /* It is safe to ignore types here since we have already done
2433 type checking in the hashing and equality routines. In fact
2434 type checking here merely gets in the way of constant
2435 propagation. Also, make sure that it is safe to propagate
2436 CACHED_LHS into *EXPR_P. */
2437 if (cached_lhs
2438 && ((TREE_CODE (cached_lhs) != SSA_NAME
2439 && (modify_expr_p
2440 || tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
2441 TREE_TYPE (cached_lhs))))
2442 || may_propagate_copy (*expr_p, cached_lhs)))
2444 if (dump_file && (dump_flags & TDF_DETAILS))
2446 fprintf (dump_file, " Replaced redundant expr '");
2447 print_generic_expr (dump_file, *expr_p, dump_flags);
2448 fprintf (dump_file, "' with '");
2449 print_generic_expr (dump_file, cached_lhs, dump_flags);
2450 fprintf (dump_file, "'\n");
2453 opt_stats.num_re++;
2455 #if defined ENABLE_CHECKING
2456 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
2457 || is_gimple_min_invariant (cached_lhs));
2458 #endif
2460 if (TREE_CODE (cached_lhs) == ADDR_EXPR
2461 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
2462 && is_gimple_min_invariant (cached_lhs)))
2463 retval = true;
2465 if (modify_expr_p
2466 && !tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
2467 TREE_TYPE (cached_lhs)))
2468 cached_lhs = fold_convert (TREE_TYPE (*expr_p), cached_lhs);
2470 propagate_tree_value (expr_p, cached_lhs);
2471 mark_stmt_modified (stmt);
2473 return retval;
2476 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2477 the available expressions table or the const_and_copies table.
2478 Detect and record those equivalences. */
2480 static void
2481 record_equivalences_from_stmt (tree stmt,
2482 int may_optimize_p,
2483 stmt_ann_t ann)
2485 tree lhs = TREE_OPERAND (stmt, 0);
2486 enum tree_code lhs_code = TREE_CODE (lhs);
2487 int i;
2489 if (lhs_code == SSA_NAME)
2491 tree rhs = TREE_OPERAND (stmt, 1);
2493 /* Strip away any useless type conversions. */
2494 STRIP_USELESS_TYPE_CONVERSION (rhs);
2496 /* If the RHS of the assignment is a constant or another variable that
2497 may be propagated, register it in the CONST_AND_COPIES table. We
2498 do not need to record unwind data for this, since this is a true
2499 assignment and not an equivalence inferred from a comparison. All
2500 uses of this ssa name are dominated by this assignment, so unwinding
2501 just costs time and space. */
2502 if (may_optimize_p
2503 && (TREE_CODE (rhs) == SSA_NAME
2504 || is_gimple_min_invariant (rhs)))
2505 SSA_NAME_VALUE (lhs) = rhs;
2507 if (tree_expr_nonzero_p (rhs))
2508 record_var_is_nonzero (lhs);
2511 /* Look at both sides for pointer dereferences. If we find one, then
2512 the pointer must be nonnull and we can enter that equivalence into
2513 the hash tables. */
2514 if (flag_delete_null_pointer_checks)
2515 for (i = 0; i < 2; i++)
2517 tree t = TREE_OPERAND (stmt, i);
2519 /* Strip away any COMPONENT_REFs. */
2520 while (TREE_CODE (t) == COMPONENT_REF)
2521 t = TREE_OPERAND (t, 0);
2523 /* Now see if this is a pointer dereference. */
2524 if (INDIRECT_REF_P (t))
2526 tree op = TREE_OPERAND (t, 0);
2528 /* If the pointer is a SSA variable, then enter new
2529 equivalences into the hash table. */
2530 while (TREE_CODE (op) == SSA_NAME)
2532 tree def = SSA_NAME_DEF_STMT (op);
2534 record_var_is_nonzero (op);
2536 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2537 which are known to have a nonzero value. */
2538 if (def
2539 && TREE_CODE (def) == MODIFY_EXPR
2540 && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
2541 op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
2542 else
2543 break;
2548 /* A memory store, even an aliased store, creates a useful
2549 equivalence. By exchanging the LHS and RHS, creating suitable
2550 vops and recording the result in the available expression table,
2551 we may be able to expose more redundant loads. */
2552 if (!ann->has_volatile_ops
2553 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
2554 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
2555 && !is_gimple_reg (lhs))
2557 tree rhs = TREE_OPERAND (stmt, 1);
2558 tree new;
2560 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2561 is a constant, we need to adjust the constant to fit into the
2562 type of the LHS. If the LHS is a bitfield and the RHS is not
2563 a constant, then we can not record any equivalences for this
2564 statement since we would need to represent the widening or
2565 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2566 and should not be necessary if GCC represented bitfields
2567 properly. */
2568 if (lhs_code == COMPONENT_REF
2569 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
2571 if (TREE_CONSTANT (rhs))
2572 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
2573 else
2574 rhs = NULL;
2576 /* If the value overflowed, then we can not use this equivalence. */
2577 if (rhs && ! is_gimple_min_invariant (rhs))
2578 rhs = NULL;
2581 if (rhs)
2583 /* Build a new statement with the RHS and LHS exchanged. */
2584 new = build2 (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
2586 create_ssa_artficial_load_stmt (new, stmt);
2588 /* Finally enter the statement into the available expression
2589 table. */
2590 lookup_avail_expr (new, true);
2595 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2596 CONST_AND_COPIES. */
2598 static bool
2599 cprop_operand (tree stmt, use_operand_p op_p)
2601 bool may_have_exposed_new_symbols = false;
2602 tree val;
2603 tree op = USE_FROM_PTR (op_p);
2605 /* If the operand has a known constant value or it is known to be a
2606 copy of some other variable, use the value or copy stored in
2607 CONST_AND_COPIES. */
2608 val = SSA_NAME_VALUE (op);
2609 if (val && val != op && TREE_CODE (val) != VALUE_HANDLE)
2611 tree op_type, val_type;
2613 /* Do not change the base variable in the virtual operand
2614 tables. That would make it impossible to reconstruct
2615 the renamed virtual operand if we later modify this
2616 statement. Also only allow the new value to be an SSA_NAME
2617 for propagation into virtual operands. */
2618 if (!is_gimple_reg (op)
2619 && (TREE_CODE (val) != SSA_NAME
2620 || is_gimple_reg (val)
2621 || get_virtual_var (val) != get_virtual_var (op)))
2622 return false;
2624 /* Do not replace hard register operands in asm statements. */
2625 if (TREE_CODE (stmt) == ASM_EXPR
2626 && !may_propagate_copy_into_asm (op))
2627 return false;
2629 /* Get the toplevel type of each operand. */
2630 op_type = TREE_TYPE (op);
2631 val_type = TREE_TYPE (val);
2633 /* While both types are pointers, get the type of the object
2634 pointed to. */
2635 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
2637 op_type = TREE_TYPE (op_type);
2638 val_type = TREE_TYPE (val_type);
2641 /* Make sure underlying types match before propagating a constant by
2642 converting the constant to the proper type. Note that convert may
2643 return a non-gimple expression, in which case we ignore this
2644 propagation opportunity. */
2645 if (TREE_CODE (val) != SSA_NAME)
2647 if (!lang_hooks.types_compatible_p (op_type, val_type))
2649 val = fold_convert (TREE_TYPE (op), val);
2650 if (!is_gimple_min_invariant (val))
2651 return false;
2655 /* Certain operands are not allowed to be copy propagated due
2656 to their interaction with exception handling and some GCC
2657 extensions. */
2658 else if (!may_propagate_copy (op, val))
2659 return false;
2661 /* Do not propagate copies if the propagated value is at a deeper loop
2662 depth than the propagatee. Otherwise, this may move loop variant
2663 variables outside of their loops and prevent coalescing
2664 opportunities. If the value was loop invariant, it will be hoisted
2665 by LICM and exposed for copy propagation. */
2666 if (loop_depth_of_name (val) > loop_depth_of_name (op))
2667 return false;
2669 /* Dump details. */
2670 if (dump_file && (dump_flags & TDF_DETAILS))
2672 fprintf (dump_file, " Replaced '");
2673 print_generic_expr (dump_file, op, dump_flags);
2674 fprintf (dump_file, "' with %s '",
2675 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
2676 print_generic_expr (dump_file, val, dump_flags);
2677 fprintf (dump_file, "'\n");
2680 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2681 that we may have exposed a new symbol for SSA renaming. */
2682 if (TREE_CODE (val) == ADDR_EXPR
2683 || (POINTER_TYPE_P (TREE_TYPE (op))
2684 && is_gimple_min_invariant (val)))
2685 may_have_exposed_new_symbols = true;
2687 if (TREE_CODE (val) != SSA_NAME)
2688 opt_stats.num_const_prop++;
2689 else
2690 opt_stats.num_copy_prop++;
2692 propagate_value (op_p, val);
2694 /* And note that we modified this statement. This is now
2695 safe, even if we changed virtual operands since we will
2696 rescan the statement and rewrite its operands again. */
2697 mark_stmt_modified (stmt);
2699 return may_have_exposed_new_symbols;
2702 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2703 known value for that SSA_NAME (or NULL if no value is known).
2705 Propagate values from CONST_AND_COPIES into the uses, vuses and
2706 v_may_def_ops of STMT. */
2708 static bool
2709 cprop_into_stmt (tree stmt)
2711 bool may_have_exposed_new_symbols = false;
2712 use_operand_p op_p;
2713 ssa_op_iter iter;
2715 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
2717 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
2718 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
2721 return may_have_exposed_new_symbols;
2725 /* Optimize the statement pointed to by iterator SI.
2727 We try to perform some simplistic global redundancy elimination and
2728 constant propagation:
2730 1- To detect global redundancy, we keep track of expressions that have
2731 been computed in this block and its dominators. If we find that the
2732 same expression is computed more than once, we eliminate repeated
2733 computations by using the target of the first one.
2735 2- Constant values and copy assignments. This is used to do very
2736 simplistic constant and copy propagation. When a constant or copy
2737 assignment is found, we map the value on the RHS of the assignment to
2738 the variable in the LHS in the CONST_AND_COPIES table. */
2740 static void
2741 optimize_stmt (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2742 basic_block bb, block_stmt_iterator si)
2744 stmt_ann_t ann;
2745 tree stmt, old_stmt;
2746 bool may_optimize_p;
2747 bool may_have_exposed_new_symbols = false;
2749 old_stmt = stmt = bsi_stmt (si);
2751 if (TREE_CODE (stmt) == COND_EXPR)
2752 canonicalize_comparison (stmt);
2754 update_stmt_if_modified (stmt);
2755 ann = stmt_ann (stmt);
2756 opt_stats.num_stmts++;
2757 may_have_exposed_new_symbols = false;
2759 if (dump_file && (dump_flags & TDF_DETAILS))
2761 fprintf (dump_file, "Optimizing statement ");
2762 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2765 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2766 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
2768 /* If the statement has been modified with constant replacements,
2769 fold its RHS before checking for redundant computations. */
2770 if (ann->modified)
2772 tree rhs;
2774 /* Try to fold the statement making sure that STMT is kept
2775 up to date. */
2776 if (fold_stmt (bsi_stmt_ptr (si)))
2778 stmt = bsi_stmt (si);
2779 ann = stmt_ann (stmt);
2781 if (dump_file && (dump_flags & TDF_DETAILS))
2783 fprintf (dump_file, " Folded to: ");
2784 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2788 rhs = get_rhs (stmt);
2789 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
2790 recompute_tree_invariant_for_addr_expr (rhs);
2792 /* Constant/copy propagation above may change the set of
2793 virtual operands associated with this statement. Folding
2794 may remove the need for some virtual operands.
2796 Indicate we will need to rescan and rewrite the statement. */
2797 may_have_exposed_new_symbols = true;
2800 /* Check for redundant computations. Do this optimization only
2801 for assignments that have no volatile ops and conditionals. */
2802 may_optimize_p = (!ann->has_volatile_ops
2803 && ((TREE_CODE (stmt) == RETURN_EXPR
2804 && TREE_OPERAND (stmt, 0)
2805 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
2806 && ! (TREE_SIDE_EFFECTS
2807 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
2808 || (TREE_CODE (stmt) == MODIFY_EXPR
2809 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
2810 || TREE_CODE (stmt) == COND_EXPR
2811 || TREE_CODE (stmt) == SWITCH_EXPR));
2813 if (may_optimize_p)
2814 may_have_exposed_new_symbols
2815 |= eliminate_redundant_computations (stmt, ann);
2817 /* Record any additional equivalences created by this statement. */
2818 if (TREE_CODE (stmt) == MODIFY_EXPR)
2819 record_equivalences_from_stmt (stmt,
2820 may_optimize_p,
2821 ann);
2823 /* If STMT is a COND_EXPR and it was modified, then we may know
2824 where it goes. If that is the case, then mark the CFG as altered.
2826 This will cause us to later call remove_unreachable_blocks and
2827 cleanup_tree_cfg when it is safe to do so. It is not safe to
2828 clean things up here since removal of edges and such can trigger
2829 the removal of PHI nodes, which in turn can release SSA_NAMEs to
2830 the manager.
2832 That's all fine and good, except that once SSA_NAMEs are released
2833 to the manager, we must not call create_ssa_name until all references
2834 to released SSA_NAMEs have been eliminated.
2836 All references to the deleted SSA_NAMEs can not be eliminated until
2837 we remove unreachable blocks.
2839 We can not remove unreachable blocks until after we have completed
2840 any queued jump threading.
2842 We can not complete any queued jump threads until we have taken
2843 appropriate variables out of SSA form. Taking variables out of
2844 SSA form can call create_ssa_name and thus we lose.
2846 Ultimately I suspect we're going to need to change the interface
2847 into the SSA_NAME manager. */
2849 if (ann->modified)
2851 tree val = NULL;
2853 if (TREE_CODE (stmt) == COND_EXPR)
2854 val = COND_EXPR_COND (stmt);
2855 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2856 val = SWITCH_COND (stmt);
2858 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
2859 cfg_altered = true;
2861 /* If we simplified a statement in such a way as to be shown that it
2862 cannot trap, update the eh information and the cfg to match. */
2863 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
2865 bitmap_set_bit (need_eh_cleanup, bb->index);
2866 if (dump_file && (dump_flags & TDF_DETAILS))
2867 fprintf (dump_file, " Flagged to clear EH edges.\n");
2871 if (may_have_exposed_new_symbols)
2872 VEC_safe_push (tree, heap, stmts_to_rescan, bsi_stmt (si));
2875 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
2876 found, return its LHS. Otherwise insert STMT in the table and return
2877 NULL_TREE.
2879 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
2880 is also added to the stack pointed to by BLOCK_AVAIL_EXPRS_P, so that they
2881 can be removed when we finish processing this block and its children.
2883 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
2884 contains no CALL_EXPR on its RHS and makes no volatile nor
2885 aliased references. */
2887 static tree
2888 lookup_avail_expr (tree stmt, bool insert)
2890 void **slot;
2891 tree lhs;
2892 tree temp;
2893 struct expr_hash_elt *element = XNEW (struct expr_hash_elt);
2895 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
2897 initialize_hash_element (stmt, lhs, element);
2899 /* Don't bother remembering constant assignments and copy operations.
2900 Constants and copy operations are handled by the constant/copy propagator
2901 in optimize_stmt. */
2902 if (TREE_CODE (element->rhs) == SSA_NAME
2903 || is_gimple_min_invariant (element->rhs))
2905 free (element);
2906 return NULL_TREE;
2909 /* If this is an equality test against zero, see if we have recorded a
2910 nonzero value for the variable in question. */
2911 if ((TREE_CODE (element->rhs) == EQ_EXPR
2912 || TREE_CODE (element->rhs) == NE_EXPR)
2913 && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
2914 && integer_zerop (TREE_OPERAND (element->rhs, 1)))
2916 int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
2918 if (bitmap_bit_p (nonzero_vars, indx))
2920 tree t = element->rhs;
2921 free (element);
2922 return constant_boolean_node (TREE_CODE (t) != EQ_EXPR,
2923 TREE_TYPE (t));
2927 /* Finally try to find the expression in the main expression hash table. */
2928 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
2929 (insert ? INSERT : NO_INSERT));
2930 if (slot == NULL)
2932 free (element);
2933 return NULL_TREE;
2936 if (*slot == NULL)
2938 *slot = (void *) element;
2939 VEC_safe_push (tree, heap, avail_exprs_stack,
2940 stmt ? stmt : element->rhs);
2941 return NULL_TREE;
2944 /* Extract the LHS of the assignment so that it can be used as the current
2945 definition of another variable. */
2946 lhs = ((struct expr_hash_elt *)*slot)->lhs;
2948 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
2949 use the value from the const_and_copies table. */
2950 if (TREE_CODE (lhs) == SSA_NAME)
2952 temp = SSA_NAME_VALUE (lhs);
2953 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
2954 lhs = temp;
2957 free (element);
2958 return lhs;
2961 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
2962 range of values that result in the conditional having a true value.
2964 Return true if we are successful in extracting a range from COND and
2965 false if we are unsuccessful. */
2967 static bool
2968 extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
2970 tree op1 = TREE_OPERAND (cond, 1);
2971 tree high, low, type;
2972 int inverted;
2974 type = TREE_TYPE (op1);
2976 /* Experiments have shown that it's rarely, if ever useful to
2977 record ranges for enumerations. Presumably this is due to
2978 the fact that they're rarely used directly. They are typically
2979 cast into an integer type and used that way. */
2980 if (TREE_CODE (type) != INTEGER_TYPE)
2981 return 0;
2983 switch (TREE_CODE (cond))
2985 case EQ_EXPR:
2986 high = low = op1;
2987 inverted = 0;
2988 break;
2990 case NE_EXPR:
2991 high = low = op1;
2992 inverted = 1;
2993 break;
2995 case GE_EXPR:
2996 low = op1;
2998 /* Get the highest value of the type. If not a constant, use that
2999 of its base type, if it has one. */
3000 high = TYPE_MAX_VALUE (type);
3001 if (TREE_CODE (high) != INTEGER_CST && TREE_TYPE (type))
3002 high = TYPE_MAX_VALUE (TREE_TYPE (type));
3003 inverted = 0;
3004 break;
3006 case GT_EXPR:
3007 high = TYPE_MAX_VALUE (type);
3008 if (TREE_CODE (high) != INTEGER_CST && TREE_TYPE (type))
3009 high = TYPE_MAX_VALUE (TREE_TYPE (type));
3010 if (!tree_int_cst_lt (op1, high))
3011 return 0;
3012 low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
3013 inverted = 0;
3014 break;
3016 case LE_EXPR:
3017 high = op1;
3018 low = TYPE_MIN_VALUE (type);
3019 if (TREE_CODE (low) != INTEGER_CST && TREE_TYPE (type))
3020 low = TYPE_MIN_VALUE (TREE_TYPE (type));
3021 inverted = 0;
3022 break;
3024 case LT_EXPR:
3025 low = TYPE_MIN_VALUE (type);
3026 if (TREE_CODE (low) != INTEGER_CST && TREE_TYPE (type))
3027 low = TYPE_MIN_VALUE (TREE_TYPE (type));
3028 if (!tree_int_cst_lt (low, op1))
3029 return 0;
3030 high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
3031 inverted = 0;
3032 break;
3034 default:
3035 return 0;
3038 *hi_p = high;
3039 *lo_p = low;
3040 *inverted_p = inverted;
3041 return 1;
3044 /* Record a range created by COND for basic block BB. */
3046 static void
3047 record_range (tree cond, basic_block bb)
3049 enum tree_code code = TREE_CODE (cond);
3051 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3052 They rarely allow for meaningful range optimizations and significantly
3053 complicate the implementation. */
3054 if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
3055 || code == GE_EXPR || code == EQ_EXPR)
3056 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
3058 struct vrp_hash_elt *vrp_hash_elt;
3059 struct vrp_element *element;
3060 VEC(vrp_element_p,heap) **vrp_records_p;
3061 void **slot;
3064 vrp_hash_elt = XNEW (struct vrp_hash_elt);
3065 vrp_hash_elt->var = TREE_OPERAND (cond, 0);
3066 vrp_hash_elt->records = NULL;
3067 slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
3069 if (*slot == NULL)
3070 *slot = (void *) vrp_hash_elt;
3071 else
3072 vrp_free (vrp_hash_elt);
3074 vrp_hash_elt = (struct vrp_hash_elt *) *slot;
3075 vrp_records_p = &vrp_hash_elt->records;
3077 element = GGC_NEW (struct vrp_element);
3078 element->low = NULL;
3079 element->high = NULL;
3080 element->cond = cond;
3081 element->bb = bb;
3083 VEC_safe_push (vrp_element_p, heap, *vrp_records_p, element);
3084 VEC_safe_push (tree, heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
3088 /* Hashing and equality functions for VRP_DATA.
3090 Since this hash table is addressed by SSA_NAMEs, we can hash on
3091 their version number and equality can be determined with a
3092 pointer comparison. */
3094 static hashval_t
3095 vrp_hash (const void *p)
3097 tree var = ((struct vrp_hash_elt *)p)->var;
3099 return SSA_NAME_VERSION (var);
3102 static int
3103 vrp_eq (const void *p1, const void *p2)
3105 tree var1 = ((struct vrp_hash_elt *)p1)->var;
3106 tree var2 = ((struct vrp_hash_elt *)p2)->var;
3108 return var1 == var2;
3111 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3112 MODIFY_EXPR statements. We compute a value number for expressions using
3113 the code of the expression and the SSA numbers of its operands. */
3115 static hashval_t
3116 avail_expr_hash (const void *p)
3118 tree stmt = ((struct expr_hash_elt *)p)->stmt;
3119 tree rhs = ((struct expr_hash_elt *)p)->rhs;
3120 tree vuse;
3121 ssa_op_iter iter;
3122 hashval_t val = 0;
3124 /* iterative_hash_expr knows how to deal with any expression and
3125 deals with commutative operators as well, so just use it instead
3126 of duplicating such complexities here. */
3127 val = iterative_hash_expr (rhs, val);
3129 /* If the hash table entry is not associated with a statement, then we
3130 can just hash the expression and not worry about virtual operands
3131 and such. */
3132 if (!stmt || !stmt_ann (stmt))
3133 return val;
3135 /* Add the SSA version numbers of every vuse operand. This is important
3136 because compound variables like arrays are not renamed in the
3137 operands. Rather, the rename is done on the virtual variable
3138 representing all the elements of the array. */
3139 FOR_EACH_SSA_TREE_OPERAND (vuse, stmt, iter, SSA_OP_VUSE)
3140 val = iterative_hash_expr (vuse, val);
3142 return val;
3145 static hashval_t
3146 real_avail_expr_hash (const void *p)
3148 return ((const struct expr_hash_elt *)p)->hash;
3151 static int
3152 avail_expr_eq (const void *p1, const void *p2)
3154 tree stmt1 = ((struct expr_hash_elt *)p1)->stmt;
3155 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
3156 tree stmt2 = ((struct expr_hash_elt *)p2)->stmt;
3157 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
3159 /* If they are the same physical expression, return true. */
3160 if (rhs1 == rhs2 && stmt1 == stmt2)
3161 return true;
3163 /* If their codes are not equal, then quit now. */
3164 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
3165 return false;
3167 /* In case of a collision, both RHS have to be identical and have the
3168 same VUSE operands. */
3169 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
3170 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
3171 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
3173 bool ret = compare_ssa_operands_equal (stmt1, stmt2, SSA_OP_VUSE);
3174 gcc_assert (!ret || ((struct expr_hash_elt *)p1)->hash
3175 == ((struct expr_hash_elt *)p2)->hash);
3176 return ret;
3179 return false;