* configure.ac: Enable checking assembler dwarf2 support for bfin
[official-gcc/alias-decl.git] / gcc / tree-ssa-dom.c
blobcb5eeb03a2f28cf6b8a80730a1a55bf7c6c41eec
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 update_rhs_and_lookup_avail_expr (tree, tree, bool);
277 static tree simplify_rhs_and_lookup_avail_expr (tree, int);
278 static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int);
279 static tree simplify_switch_and_lookup_avail_expr (tree, int);
280 static tree find_equivalent_equality_comparison (tree);
281 static void record_range (tree, basic_block);
282 static bool extract_range_from_cond (tree, tree *, tree *, int *);
283 static void record_equivalences_from_phis (basic_block);
284 static void record_equivalences_from_incoming_edge (basic_block);
285 static bool eliminate_redundant_computations (tree, stmt_ann_t);
286 static void record_equivalences_from_stmt (tree, int, stmt_ann_t);
287 static void thread_across_edge (struct dom_walk_data *, edge);
288 static void dom_opt_finalize_block (struct dom_walk_data *, basic_block);
289 static void dom_opt_initialize_block (struct dom_walk_data *, basic_block);
290 static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block);
291 static void remove_local_expressions_from_table (void);
292 static void restore_vars_to_original_value (void);
293 static edge single_incoming_edge_ignoring_loop_edges (basic_block);
294 static void restore_nonzero_vars_to_original_value (void);
295 static inline bool unsafe_associative_fp_binop (tree);
298 /* Local version of fold that doesn't introduce cruft. */
300 static tree
301 local_fold (tree t)
303 t = fold (t);
305 /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that
306 may have been added by fold, and "useless" type conversions that might
307 now be apparent due to propagation. */
308 STRIP_USELESS_TYPE_CONVERSION (t);
310 return t;
313 /* Allocate an EDGE_INFO for edge E and attach it to E.
314 Return the new EDGE_INFO structure. */
316 static struct edge_info *
317 allocate_edge_info (edge e)
319 struct edge_info *edge_info;
321 edge_info = xcalloc (1, sizeof (struct edge_info));
323 e->aux = edge_info;
324 return edge_info;
327 /* Free all EDGE_INFO structures associated with edges in the CFG.
328 If a particular edge can be threaded, copy the redirection
329 target from the EDGE_INFO structure into the edge's AUX field
330 as required by code to update the CFG and SSA graph for
331 jump threading. */
333 static void
334 free_all_edge_infos (void)
336 basic_block bb;
337 edge_iterator ei;
338 edge e;
340 FOR_EACH_BB (bb)
342 FOR_EACH_EDGE (e, ei, bb->preds)
344 struct edge_info *edge_info = e->aux;
346 if (edge_info)
348 e->aux = edge_info->redirection_target;
349 if (edge_info->cond_equivalences)
350 free (edge_info->cond_equivalences);
351 free (edge_info);
357 /* Free an instance of vrp_hash_elt. */
359 static void
360 vrp_free (void *data)
362 struct vrp_hash_elt *elt = data;
363 struct VEC(vrp_element_p,heap) **vrp_elt = &elt->records;
365 VEC_free (vrp_element_p, heap, *vrp_elt);
366 free (elt);
369 /* Jump threading, redundancy elimination and const/copy propagation.
371 This pass may expose new symbols that need to be renamed into SSA. For
372 every new symbol exposed, its corresponding bit will be set in
373 VARS_TO_RENAME. */
375 static void
376 tree_ssa_dominator_optimize (void)
378 struct dom_walk_data walk_data;
379 unsigned int i;
380 struct loops loops_info;
382 memset (&opt_stats, 0, sizeof (opt_stats));
384 /* Create our hash tables. */
385 avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free);
386 vrp_data = htab_create (ceil_log2 (num_ssa_names), vrp_hash, vrp_eq,
387 vrp_free);
388 avail_exprs_stack = VEC_alloc (tree, heap, 20);
389 const_and_copies_stack = VEC_alloc (tree, heap, 20);
390 nonzero_vars_stack = VEC_alloc (tree, heap, 20);
391 vrp_variables_stack = VEC_alloc (tree, heap, 20);
392 stmts_to_rescan = VEC_alloc (tree, heap, 20);
393 nonzero_vars = BITMAP_ALLOC (NULL);
394 threaded_blocks = BITMAP_ALLOC (NULL);
395 need_eh_cleanup = BITMAP_ALLOC (NULL);
397 /* Setup callbacks for the generic dominator tree walker. */
398 walk_data.walk_stmts_backward = false;
399 walk_data.dom_direction = CDI_DOMINATORS;
400 walk_data.initialize_block_local_data = NULL;
401 walk_data.before_dom_children_before_stmts = dom_opt_initialize_block;
402 walk_data.before_dom_children_walk_stmts = optimize_stmt;
403 walk_data.before_dom_children_after_stmts = propagate_to_outgoing_edges;
404 walk_data.after_dom_children_before_stmts = NULL;
405 walk_data.after_dom_children_walk_stmts = NULL;
406 walk_data.after_dom_children_after_stmts = dom_opt_finalize_block;
407 /* Right now we only attach a dummy COND_EXPR to the global data pointer.
408 When we attach more stuff we'll need to fill this out with a real
409 structure. */
410 walk_data.global_data = NULL;
411 walk_data.block_local_data_size = 0;
412 walk_data.interesting_blocks = NULL;
414 /* Now initialize the dominator walker. */
415 init_walk_dominator_tree (&walk_data);
417 calculate_dominance_info (CDI_DOMINATORS);
419 /* We need to know which edges exit loops so that we can
420 aggressively thread through loop headers to an exit
421 edge. */
422 flow_loops_find (&loops_info);
423 mark_loop_exit_edges (&loops_info);
424 flow_loops_free (&loops_info);
426 /* Clean up the CFG so that any forwarder blocks created by loop
427 canonicalization are removed. */
428 cleanup_tree_cfg ();
429 calculate_dominance_info (CDI_DOMINATORS);
431 /* If we prove certain blocks are unreachable, then we want to
432 repeat the dominator optimization process as PHI nodes may
433 have turned into copies which allows better propagation of
434 values. So we repeat until we do not identify any new unreachable
435 blocks. */
438 /* Optimize the dominator tree. */
439 cfg_altered = false;
441 /* We need accurate information regarding back edges in the CFG
442 for jump threading. */
443 mark_dfs_back_edges ();
445 /* Recursively walk the dominator tree optimizing statements. */
446 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
449 block_stmt_iterator bsi;
450 basic_block bb;
451 FOR_EACH_BB (bb)
453 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
455 update_stmt_if_modified (bsi_stmt (bsi));
460 /* If we exposed any new variables, go ahead and put them into
461 SSA form now, before we handle jump threading. This simplifies
462 interactions between rewriting of _DECL nodes into SSA form
463 and rewriting SSA_NAME nodes into SSA form after block
464 duplication and CFG manipulation. */
465 update_ssa (TODO_update_ssa);
467 free_all_edge_infos ();
469 /* Thread jumps, creating duplicate blocks as needed. */
470 cfg_altered |= thread_through_all_blocks (threaded_blocks);
472 /* Removal of statements may make some EH edges dead. Purge
473 such edges from the CFG as needed. */
474 if (!bitmap_empty_p (need_eh_cleanup))
476 cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup);
477 bitmap_zero (need_eh_cleanup);
480 if (cfg_altered)
481 free_dominance_info (CDI_DOMINATORS);
483 /* Only iterate if we threaded jumps AND the CFG cleanup did
484 something interesting. Other cases generate far fewer
485 optimization opportunities and thus are not worth another
486 full DOM iteration. */
487 cfg_altered &= cleanup_tree_cfg ();
489 if (rediscover_loops_after_threading)
491 /* Rerun basic loop analysis to discover any newly
492 created loops and update the set of exit edges. */
493 rediscover_loops_after_threading = false;
494 flow_loops_find (&loops_info);
495 mark_loop_exit_edges (&loops_info);
496 flow_loops_free (&loops_info);
498 /* Remove any forwarder blocks inserted by loop
499 header canonicalization. */
500 cleanup_tree_cfg ();
503 calculate_dominance_info (CDI_DOMINATORS);
505 update_ssa (TODO_update_ssa);
507 /* Reinitialize the various tables. */
508 bitmap_clear (nonzero_vars);
509 bitmap_clear (threaded_blocks);
510 htab_empty (avail_exprs);
511 htab_empty (vrp_data);
513 /* Finally, remove everything except invariants in SSA_NAME_VALUE.
515 This must be done before we iterate as we might have a
516 reference to an SSA_NAME which was removed by the call to
517 update_ssa.
519 Long term we will be able to let everything in SSA_NAME_VALUE
520 persist. However, for now, we know this is the safe thing to do. */
521 for (i = 0; i < num_ssa_names; i++)
523 tree name = ssa_name (i);
524 tree value;
526 if (!name)
527 continue;
529 value = SSA_NAME_VALUE (name);
530 if (value && !is_gimple_min_invariant (value))
531 SSA_NAME_VALUE (name) = NULL;
534 opt_stats.num_iterations++;
536 while (optimize > 1 && cfg_altered);
538 /* Debugging dumps. */
539 if (dump_file && (dump_flags & TDF_STATS))
540 dump_dominator_optimization_stats (dump_file);
542 /* We emptied the hash table earlier, now delete it completely. */
543 htab_delete (avail_exprs);
544 htab_delete (vrp_data);
546 /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA,
547 CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom
548 of the do-while loop above. */
550 /* And finalize the dominator walker. */
551 fini_walk_dominator_tree (&walk_data);
553 /* Free nonzero_vars. */
554 BITMAP_FREE (nonzero_vars);
555 BITMAP_FREE (threaded_blocks);
556 BITMAP_FREE (need_eh_cleanup);
558 VEC_free (tree, heap, avail_exprs_stack);
559 VEC_free (tree, heap, const_and_copies_stack);
560 VEC_free (tree, heap, nonzero_vars_stack);
561 VEC_free (tree, heap, vrp_variables_stack);
562 VEC_free (tree, heap, stmts_to_rescan);
565 static bool
566 gate_dominator (void)
568 return flag_tree_dom != 0;
571 struct tree_opt_pass pass_dominator =
573 "dom", /* name */
574 gate_dominator, /* gate */
575 tree_ssa_dominator_optimize, /* execute */
576 NULL, /* sub */
577 NULL, /* next */
578 0, /* static_pass_number */
579 TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */
580 PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
581 0, /* properties_provided */
582 0, /* properties_destroyed */
583 0, /* todo_flags_start */
584 TODO_dump_func
585 | TODO_update_ssa
586 | TODO_verify_ssa, /* todo_flags_finish */
587 0 /* letter */
591 /* We are exiting E->src, see if E->dest ends with a conditional
592 jump which has a known value when reached via E.
594 Special care is necessary if E is a back edge in the CFG as we
595 will have already recorded equivalences for E->dest into our
596 various tables, including the result of the conditional at
597 the end of E->dest. Threading opportunities are severely
598 limited in that case to avoid short-circuiting the loop
599 incorrectly.
601 Note it is quite common for the first block inside a loop to
602 end with a conditional which is either always true or always
603 false when reached via the loop backedge. Thus we do not want
604 to blindly disable threading across a loop backedge. */
606 static void
607 thread_across_edge (struct dom_walk_data *walk_data, edge e)
609 block_stmt_iterator bsi;
610 tree stmt = NULL;
611 tree phi;
612 int stmt_count = 0;
613 int max_stmt_count;
616 /* If E->dest does not end with a conditional, then there is
617 nothing to do. */
618 bsi = bsi_last (e->dest);
619 if (bsi_end_p (bsi)
620 || ! bsi_stmt (bsi)
621 || (TREE_CODE (bsi_stmt (bsi)) != COND_EXPR
622 && TREE_CODE (bsi_stmt (bsi)) != GOTO_EXPR
623 && TREE_CODE (bsi_stmt (bsi)) != SWITCH_EXPR))
624 return;
626 /* The basic idea here is to use whatever knowledge we have
627 from our dominator walk to simplify statements in E->dest,
628 with the ultimate goal being to simplify the conditional
629 at the end of E->dest.
631 Note that we must undo any changes we make to the underlying
632 statements as the simplifications we are making are control
633 flow sensitive (ie, the simplifications are valid when we
634 traverse E, but may not be valid on other paths to E->dest. */
636 /* Each PHI creates a temporary equivalence, record them. Again
637 these are context sensitive equivalences and will be removed
638 by our caller. */
639 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
641 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e);
642 tree dst = PHI_RESULT (phi);
644 /* Do not include virtual PHIs in our statement count as
645 they never generate code. */
646 if (is_gimple_reg (dst))
647 stmt_count++;
649 /* If the desired argument is not the same as this PHI's result
650 and it is set by a PHI in E->dest, then we can not thread
651 through E->dest. */
652 if (src != dst
653 && TREE_CODE (src) == SSA_NAME
654 && TREE_CODE (SSA_NAME_DEF_STMT (src)) == PHI_NODE
655 && bb_for_stmt (SSA_NAME_DEF_STMT (src)) == e->dest)
656 return;
658 record_const_or_copy (dst, src);
661 /* Try to simplify each statement in E->dest, ultimately leading to
662 a simplification of the COND_EXPR at the end of E->dest.
664 We might consider marking just those statements which ultimately
665 feed the COND_EXPR. It's not clear if the overhead of bookkeeping
666 would be recovered by trying to simplify fewer statements.
668 If we are able to simplify a statement into the form
669 SSA_NAME = (SSA_NAME | gimple invariant), then we can record
670 a context sensitive equivalency which may help us simplify
671 later statements in E->dest.
673 Failure to simplify into the form above merely means that the
674 statement provides no equivalences to help simplify later
675 statements. This does not prevent threading through E->dest. */
676 max_stmt_count = PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS);
677 for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi))
679 tree cached_lhs = NULL;
681 stmt = bsi_stmt (bsi);
683 /* Ignore empty statements and labels. */
684 if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR)
685 continue;
687 /* If duplicating this block is going to cause too much code
688 expansion, then do not thread through this block. */
689 stmt_count++;
690 if (stmt_count > max_stmt_count)
691 return;
693 /* Safely handle threading across loop backedges. This is
694 over conservative, but still allows us to capture the
695 majority of the cases where we can thread across a loop
696 backedge. */
697 if ((e->flags & EDGE_DFS_BACK) != 0
698 && TREE_CODE (stmt) != COND_EXPR
699 && TREE_CODE (stmt) != SWITCH_EXPR)
700 return;
702 /* If the statement has volatile operands, then we assume we
703 can not thread through this block. This is overly
704 conservative in some ways. */
705 if (TREE_CODE (stmt) == ASM_EXPR && ASM_VOLATILE_P (stmt))
706 return;
708 /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new
709 value, then do not try to simplify this statement as it will
710 not simplify in any way that is helpful for jump threading. */
711 if (TREE_CODE (stmt) != MODIFY_EXPR
712 || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME)
713 continue;
715 /* At this point we have a statement which assigns an RHS to an
716 SSA_VAR on the LHS. We want to try and simplify this statement
717 to expose more context sensitive equivalences which in turn may
718 allow us to simplify the condition at the end of the loop. */
719 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME)
720 cached_lhs = TREE_OPERAND (stmt, 1);
721 else
723 /* Copy the operands. */
724 tree *copy, pre_fold_expr;
725 ssa_op_iter iter;
726 use_operand_p use_p;
727 unsigned int num, i = 0;
729 num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE));
730 copy = xcalloc (num, sizeof (tree));
732 /* Make a copy of the uses & vuses into USES_COPY, then cprop into
733 the operands. */
734 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
736 tree tmp = NULL;
737 tree use = USE_FROM_PTR (use_p);
739 copy[i++] = use;
740 if (TREE_CODE (use) == SSA_NAME)
741 tmp = SSA_NAME_VALUE (use);
742 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
743 SET_USE (use_p, tmp);
746 /* Try to fold/lookup the new expression. Inserting the
747 expression into the hash table is unlikely to help
748 Sadly, we have to handle conditional assignments specially
749 here, because fold expects all the operands of an expression
750 to be folded before the expression itself is folded, but we
751 can't just substitute the folded condition here. */
752 if (TREE_CODE (TREE_OPERAND (stmt, 1)) == COND_EXPR)
754 tree cond = COND_EXPR_COND (TREE_OPERAND (stmt, 1));
755 cond = fold (cond);
756 if (cond == boolean_true_node)
757 pre_fold_expr = COND_EXPR_THEN (TREE_OPERAND (stmt, 1));
758 else if (cond == boolean_false_node)
759 pre_fold_expr = COND_EXPR_ELSE (TREE_OPERAND (stmt, 1));
760 else
761 pre_fold_expr = TREE_OPERAND (stmt, 1);
763 else
764 pre_fold_expr = TREE_OPERAND (stmt, 1);
766 if (pre_fold_expr)
768 cached_lhs = fold (pre_fold_expr);
769 if (TREE_CODE (cached_lhs) != SSA_NAME
770 && !is_gimple_min_invariant (cached_lhs))
771 cached_lhs = lookup_avail_expr (stmt, false);
774 /* Restore the statement's original uses/defs. */
775 i = 0;
776 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE)
777 SET_USE (use_p, copy[i++]);
779 free (copy);
782 /* Record the context sensitive equivalence if we were able
783 to simplify this statement. */
784 if (cached_lhs
785 && (TREE_CODE (cached_lhs) == SSA_NAME
786 || is_gimple_min_invariant (cached_lhs)))
787 record_const_or_copy (TREE_OPERAND (stmt, 0), cached_lhs);
790 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm
791 will be taken. */
792 if (stmt
793 && (TREE_CODE (stmt) == COND_EXPR
794 || TREE_CODE (stmt) == GOTO_EXPR
795 || TREE_CODE (stmt) == SWITCH_EXPR))
797 tree cond, cached_lhs;
799 /* Now temporarily cprop the operands and try to find the resulting
800 expression in the hash tables. */
801 if (TREE_CODE (stmt) == COND_EXPR)
802 cond = COND_EXPR_COND (stmt);
803 else if (TREE_CODE (stmt) == GOTO_EXPR)
804 cond = GOTO_DESTINATION (stmt);
805 else
806 cond = SWITCH_COND (stmt);
808 if (COMPARISON_CLASS_P (cond))
810 tree dummy_cond, op0, op1;
811 enum tree_code cond_code;
813 op0 = TREE_OPERAND (cond, 0);
814 op1 = TREE_OPERAND (cond, 1);
815 cond_code = TREE_CODE (cond);
817 /* Get the current value of both operands. */
818 if (TREE_CODE (op0) == SSA_NAME)
820 tree tmp = SSA_NAME_VALUE (op0);
821 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
822 op0 = tmp;
825 if (TREE_CODE (op1) == SSA_NAME)
827 tree tmp = SSA_NAME_VALUE (op1);
828 if (tmp && TREE_CODE (tmp) != VALUE_HANDLE)
829 op1 = tmp;
832 /* Stuff the operator and operands into our dummy conditional
833 expression, creating the dummy conditional if necessary. */
834 dummy_cond = walk_data->global_data;
835 if (! dummy_cond)
837 dummy_cond = build (cond_code, boolean_type_node, op0, op1);
838 dummy_cond = build (COND_EXPR, void_type_node,
839 dummy_cond, NULL, NULL);
840 walk_data->global_data = dummy_cond;
842 else
844 TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code);
845 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0;
846 TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1;
849 /* If the conditional folds to an invariant, then we are done,
850 otherwise look it up in the hash tables. */
851 cached_lhs = local_fold (COND_EXPR_COND (dummy_cond));
852 if (! is_gimple_min_invariant (cached_lhs))
854 cached_lhs = lookup_avail_expr (dummy_cond, false);
855 if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs))
856 cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond,
857 NULL,
858 false);
861 /* We can have conditionals which just test the state of a
862 variable rather than use a relational operator. These are
863 simpler to handle. */
864 else if (TREE_CODE (cond) == SSA_NAME)
866 cached_lhs = cond;
867 cached_lhs = SSA_NAME_VALUE (cached_lhs);
868 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs))
869 cached_lhs = NULL;
871 else
872 cached_lhs = lookup_avail_expr (stmt, false);
874 if (cached_lhs)
876 edge taken_edge = find_taken_edge (e->dest, cached_lhs);
877 basic_block dest = (taken_edge ? taken_edge->dest : NULL);
879 if (dest == e->dest)
880 return;
882 /* If we have a known destination for the conditional, then
883 we can perform this optimization, which saves at least one
884 conditional jump each time it applies since we get to
885 bypass the conditional at our original destination. */
886 if (dest)
888 struct edge_info *edge_info;
890 if (e->aux)
891 edge_info = e->aux;
892 else
893 edge_info = allocate_edge_info (e);
894 edge_info->redirection_target = taken_edge;
895 bitmap_set_bit (threaded_blocks, e->dest->index);
902 /* Initialize local stacks for this optimizer and record equivalences
903 upon entry to BB. Equivalences can come from the edge traversed to
904 reach BB or they may come from PHI nodes at the start of BB. */
906 static void
907 dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
908 basic_block bb)
910 if (dump_file && (dump_flags & TDF_DETAILS))
911 fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index);
913 /* Push a marker on the stacks of local information so that we know how
914 far to unwind when we finalize this block. */
915 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
916 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
917 VEC_safe_push (tree, heap, nonzero_vars_stack, NULL_TREE);
918 VEC_safe_push (tree, heap, vrp_variables_stack, NULL_TREE);
920 record_equivalences_from_incoming_edge (bb);
922 /* PHI nodes can create equivalences too. */
923 record_equivalences_from_phis (bb);
926 /* Given an expression EXPR (a relational expression or a statement),
927 initialize the hash table element pointed to by ELEMENT. */
929 static void
930 initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element)
932 /* Hash table elements may be based on conditional expressions or statements.
934 For the former case, we have no annotation and we want to hash the
935 conditional expression. In the latter case we have an annotation and
936 we want to record the expression the statement evaluates. */
937 if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR)
939 element->stmt = NULL;
940 element->rhs = expr;
942 else if (TREE_CODE (expr) == COND_EXPR)
944 element->stmt = expr;
945 element->rhs = COND_EXPR_COND (expr);
947 else if (TREE_CODE (expr) == SWITCH_EXPR)
949 element->stmt = expr;
950 element->rhs = SWITCH_COND (expr);
952 else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0))
954 element->stmt = expr;
955 element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1);
957 else if (TREE_CODE (expr) == GOTO_EXPR)
959 element->stmt = expr;
960 element->rhs = GOTO_DESTINATION (expr);
962 else
964 element->stmt = expr;
965 element->rhs = TREE_OPERAND (expr, 1);
968 element->lhs = lhs;
969 element->hash = avail_expr_hash (element);
972 /* Remove all the expressions in LOCALS from TABLE, stopping when there are
973 LIMIT entries left in LOCALs. */
975 static void
976 remove_local_expressions_from_table (void)
978 /* Remove all the expressions made available in this block. */
979 while (VEC_length (tree, avail_exprs_stack) > 0)
981 struct expr_hash_elt element;
982 tree expr = VEC_pop (tree, avail_exprs_stack);
984 if (expr == NULL_TREE)
985 break;
987 initialize_hash_element (expr, NULL, &element);
988 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
992 /* Use the SSA_NAMES in LOCALS to restore TABLE to its original
993 state, stopping when there are LIMIT entries left in LOCALs. */
995 static void
996 restore_nonzero_vars_to_original_value (void)
998 while (VEC_length (tree, nonzero_vars_stack) > 0)
1000 tree name = VEC_pop (tree, nonzero_vars_stack);
1002 if (name == NULL)
1003 break;
1005 bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name));
1009 /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore
1010 CONST_AND_COPIES to its original state, stopping when we hit a
1011 NULL marker. */
1013 static void
1014 restore_vars_to_original_value (void)
1016 while (VEC_length (tree, const_and_copies_stack) > 0)
1018 tree prev_value, dest;
1020 dest = VEC_pop (tree, const_and_copies_stack);
1022 if (dest == NULL)
1023 break;
1025 prev_value = VEC_pop (tree, const_and_copies_stack);
1026 SSA_NAME_VALUE (dest) = prev_value;
1030 /* We have finished processing the dominator children of BB, perform
1031 any finalization actions in preparation for leaving this node in
1032 the dominator tree. */
1034 static void
1035 dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb)
1037 tree last;
1039 /* If we have an outgoing edge to a block with multiple incoming and
1040 outgoing edges, then we may be able to thread the edge. ie, we
1041 may be able to statically determine which of the outgoing edges
1042 will be traversed when the incoming edge from BB is traversed. */
1043 if (single_succ_p (bb)
1044 && (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0
1045 && !single_pred_p (single_succ (bb))
1046 && !single_succ_p (single_succ (bb)))
1049 thread_across_edge (walk_data, single_succ_edge (bb));
1051 else if ((last = last_stmt (bb))
1052 && TREE_CODE (last) == COND_EXPR
1053 && (COMPARISON_CLASS_P (COND_EXPR_COND (last))
1054 || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME)
1055 && EDGE_COUNT (bb->succs) == 2
1056 && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0
1057 && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0)
1059 edge true_edge, false_edge;
1061 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
1063 /* Only try to thread the edge if it reaches a target block with
1064 more than one predecessor and more than one successor. */
1065 if (!single_pred_p (true_edge->dest) && !single_succ_p (true_edge->dest))
1067 struct edge_info *edge_info;
1068 unsigned int i;
1070 /* Push a marker onto the available expression stack so that we
1071 unwind any expressions related to the TRUE arm before processing
1072 the false arm below. */
1073 VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE);
1074 VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE);
1076 edge_info = true_edge->aux;
1078 /* If we have info associated with this edge, record it into
1079 our equivalency tables. */
1080 if (edge_info)
1082 tree *cond_equivalences = edge_info->cond_equivalences;
1083 tree lhs = edge_info->lhs;
1084 tree rhs = edge_info->rhs;
1086 /* If we have a simple NAME = VALUE equivalency record it. */
1087 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1088 record_const_or_copy (lhs, rhs);
1090 /* If we have 0 = COND or 1 = COND equivalences, record them
1091 into our expression hash tables. */
1092 if (cond_equivalences)
1093 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1095 tree expr = cond_equivalences[i];
1096 tree value = cond_equivalences[i + 1];
1098 record_cond (expr, value);
1102 /* Now thread the edge. */
1103 thread_across_edge (walk_data, true_edge);
1105 /* And restore the various tables to their state before
1106 we threaded this edge. */
1107 remove_local_expressions_from_table ();
1108 restore_vars_to_original_value ();
1111 /* Similarly for the ELSE arm. */
1112 if (!single_pred_p (false_edge->dest) && !single_succ_p (false_edge->dest))
1114 struct edge_info *edge_info;
1115 unsigned int i;
1117 edge_info = false_edge->aux;
1119 /* If we have info associated with this edge, record it into
1120 our equivalency tables. */
1121 if (edge_info)
1123 tree *cond_equivalences = edge_info->cond_equivalences;
1124 tree lhs = edge_info->lhs;
1125 tree rhs = edge_info->rhs;
1127 /* If we have a simple NAME = VALUE equivalency record it. */
1128 if (lhs && TREE_CODE (lhs) == SSA_NAME)
1129 record_const_or_copy (lhs, rhs);
1131 /* If we have 0 = COND or 1 = COND equivalences, record them
1132 into our expression hash tables. */
1133 if (cond_equivalences)
1134 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1136 tree expr = cond_equivalences[i];
1137 tree value = cond_equivalences[i + 1];
1139 record_cond (expr, value);
1143 thread_across_edge (walk_data, false_edge);
1145 /* No need to remove local expressions from our tables
1146 or restore vars to their original value as that will
1147 be done immediately below. */
1151 remove_local_expressions_from_table ();
1152 restore_nonzero_vars_to_original_value ();
1153 restore_vars_to_original_value ();
1155 /* Remove VRP records associated with this basic block. They are no
1156 longer valid.
1158 To be efficient, we note which variables have had their values
1159 constrained in this block. So walk over each variable in the
1160 VRP_VARIABLEs array. */
1161 while (VEC_length (tree, vrp_variables_stack) > 0)
1163 tree var = VEC_pop (tree, vrp_variables_stack);
1164 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
1165 void **slot;
1167 /* Each variable has a stack of value range records. We want to
1168 invalidate those associated with our basic block. So we walk
1169 the array backwards popping off records associated with our
1170 block. Once we hit a record not associated with our block
1171 we are done. */
1172 VEC(vrp_element_p,heap) **var_vrp_records;
1174 if (var == NULL)
1175 break;
1177 vrp_hash_elt.var = var;
1178 vrp_hash_elt.records = NULL;
1180 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
1182 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
1183 var_vrp_records = &vrp_hash_elt_p->records;
1185 while (VEC_length (vrp_element_p, *var_vrp_records) > 0)
1187 struct vrp_element *element
1188 = VEC_last (vrp_element_p, *var_vrp_records);
1190 if (element->bb != bb)
1191 break;
1193 VEC_pop (vrp_element_p, *var_vrp_records);
1197 /* If we queued any statements to rescan in this block, then
1198 go ahead and rescan them now. */
1199 while (VEC_length (tree, stmts_to_rescan) > 0)
1201 tree stmt = VEC_last (tree, stmts_to_rescan);
1202 basic_block stmt_bb = bb_for_stmt (stmt);
1204 if (stmt_bb != bb)
1205 break;
1207 VEC_pop (tree, stmts_to_rescan);
1208 mark_new_vars_to_rename (stmt);
1212 /* PHI nodes can create equivalences too.
1214 Ignoring any alternatives which are the same as the result, if
1215 all the alternatives are equal, then the PHI node creates an
1216 equivalence.
1218 Additionally, if all the PHI alternatives are known to have a nonzero
1219 value, then the result of this PHI is known to have a nonzero value,
1220 even if we do not know its exact value. */
1222 static void
1223 record_equivalences_from_phis (basic_block bb)
1225 tree phi;
1227 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1229 tree lhs = PHI_RESULT (phi);
1230 tree rhs = NULL;
1231 int i;
1233 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1235 tree t = PHI_ARG_DEF (phi, i);
1237 /* Ignore alternatives which are the same as our LHS. Since
1238 LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we
1239 can simply compare pointers. */
1240 if (lhs == t)
1241 continue;
1243 /* If we have not processed an alternative yet, then set
1244 RHS to this alternative. */
1245 if (rhs == NULL)
1246 rhs = t;
1247 /* If we have processed an alternative (stored in RHS), then
1248 see if it is equal to this one. If it isn't, then stop
1249 the search. */
1250 else if (! operand_equal_for_phi_arg_p (rhs, t))
1251 break;
1254 /* If we had no interesting alternatives, then all the RHS alternatives
1255 must have been the same as LHS. */
1256 if (!rhs)
1257 rhs = lhs;
1259 /* If we managed to iterate through each PHI alternative without
1260 breaking out of the loop, then we have a PHI which may create
1261 a useful equivalence. We do not need to record unwind data for
1262 this, since this is a true assignment and not an equivalence
1263 inferred from a comparison. All uses of this ssa name are dominated
1264 by this assignment, so unwinding just costs time and space. */
1265 if (i == PHI_NUM_ARGS (phi)
1266 && may_propagate_copy (lhs, rhs))
1267 SSA_NAME_VALUE (lhs) = rhs;
1269 /* Now see if we know anything about the nonzero property for the
1270 result of this PHI. */
1271 for (i = 0; i < PHI_NUM_ARGS (phi); i++)
1273 if (!PHI_ARG_NONZERO (phi, i))
1274 break;
1277 if (i == PHI_NUM_ARGS (phi))
1278 bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi)));
1282 /* Ignoring loop backedges, if BB has precisely one incoming edge then
1283 return that edge. Otherwise return NULL. */
1284 static edge
1285 single_incoming_edge_ignoring_loop_edges (basic_block bb)
1287 edge retval = NULL;
1288 edge e;
1289 edge_iterator ei;
1291 FOR_EACH_EDGE (e, ei, bb->preds)
1293 /* A loop back edge can be identified by the destination of
1294 the edge dominating the source of the edge. */
1295 if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest))
1296 continue;
1298 /* If we have already seen a non-loop edge, then we must have
1299 multiple incoming non-loop edges and thus we return NULL. */
1300 if (retval)
1301 return NULL;
1303 /* This is the first non-loop incoming edge we have found. Record
1304 it. */
1305 retval = e;
1308 return retval;
1311 /* Record any equivalences created by the incoming edge to BB. If BB
1312 has more than one incoming edge, then no equivalence is created. */
1314 static void
1315 record_equivalences_from_incoming_edge (basic_block bb)
1317 edge e;
1318 basic_block parent;
1319 struct edge_info *edge_info;
1321 /* If our parent block ended with a control statement, then we may be
1322 able to record some equivalences based on which outgoing edge from
1323 the parent was followed. */
1324 parent = get_immediate_dominator (CDI_DOMINATORS, bb);
1326 e = single_incoming_edge_ignoring_loop_edges (bb);
1328 /* If we had a single incoming edge from our parent block, then enter
1329 any data associated with the edge into our tables. */
1330 if (e && e->src == parent)
1332 unsigned int i;
1334 edge_info = e->aux;
1336 if (edge_info)
1338 tree lhs = edge_info->lhs;
1339 tree rhs = edge_info->rhs;
1340 tree *cond_equivalences = edge_info->cond_equivalences;
1342 if (lhs)
1343 record_equality (lhs, rhs);
1345 if (cond_equivalences)
1347 bool recorded_range = false;
1348 for (i = 0; i < edge_info->max_cond_equivalences; i += 2)
1350 tree expr = cond_equivalences[i];
1351 tree value = cond_equivalences[i + 1];
1353 record_cond (expr, value);
1355 /* For the first true equivalence, record range
1356 information. We only do this for the first
1357 true equivalence as it should dominate any
1358 later true equivalences. */
1359 if (! recorded_range
1360 && COMPARISON_CLASS_P (expr)
1361 && value == boolean_true_node
1362 && TREE_CONSTANT (TREE_OPERAND (expr, 1)))
1364 record_range (expr, bb);
1365 recorded_range = true;
1373 /* Dump SSA statistics on FILE. */
1375 void
1376 dump_dominator_optimization_stats (FILE *file)
1378 long n_exprs;
1380 fprintf (file, "Total number of statements: %6ld\n\n",
1381 opt_stats.num_stmts);
1382 fprintf (file, "Exprs considered for dominator optimizations: %6ld\n",
1383 opt_stats.num_exprs_considered);
1385 n_exprs = opt_stats.num_exprs_considered;
1386 if (n_exprs == 0)
1387 n_exprs = 1;
1389 fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n",
1390 opt_stats.num_re, PERCENT (opt_stats.num_re,
1391 n_exprs));
1392 fprintf (file, " Constants propagated: %6ld\n",
1393 opt_stats.num_const_prop);
1394 fprintf (file, " Copies propagated: %6ld\n",
1395 opt_stats.num_copy_prop);
1397 fprintf (file, "\nTotal number of DOM iterations: %6ld\n",
1398 opt_stats.num_iterations);
1400 fprintf (file, "\nHash table statistics:\n");
1402 fprintf (file, " avail_exprs: ");
1403 htab_statistics (file, avail_exprs);
1407 /* Dump SSA statistics on stderr. */
1409 void
1410 debug_dominator_optimization_stats (void)
1412 dump_dominator_optimization_stats (stderr);
1416 /* Dump statistics for the hash table HTAB. */
1418 static void
1419 htab_statistics (FILE *file, htab_t htab)
1421 fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n",
1422 (long) htab_size (htab),
1423 (long) htab_elements (htab),
1424 htab_collisions (htab));
1427 /* Record the fact that VAR has a nonzero value, though we may not know
1428 its exact value. Note that if VAR is already known to have a nonzero
1429 value, then we do nothing. */
1431 static void
1432 record_var_is_nonzero (tree var)
1434 int indx = SSA_NAME_VERSION (var);
1436 if (bitmap_bit_p (nonzero_vars, indx))
1437 return;
1439 /* Mark it in the global table. */
1440 bitmap_set_bit (nonzero_vars, indx);
1442 /* Record this SSA_NAME so that we can reset the global table
1443 when we leave this block. */
1444 VEC_safe_push (tree, heap, nonzero_vars_stack, var);
1447 /* Enter a statement into the true/false expression hash table indicating
1448 that the condition COND has the value VALUE. */
1450 static void
1451 record_cond (tree cond, tree value)
1453 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
1454 void **slot;
1456 initialize_hash_element (cond, value, element);
1458 slot = htab_find_slot_with_hash (avail_exprs, (void *)element,
1459 element->hash, INSERT);
1460 if (*slot == NULL)
1462 *slot = (void *) element;
1463 VEC_safe_push (tree, heap, avail_exprs_stack, cond);
1465 else
1466 free (element);
1469 /* Build a new conditional using NEW_CODE, OP0 and OP1 and store
1470 the new conditional into *p, then store a boolean_true_node
1471 into *(p + 1). */
1473 static void
1474 build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p)
1476 *p = build2 (new_code, boolean_type_node, op0, op1);
1477 p++;
1478 *p = boolean_true_node;
1481 /* Record that COND is true and INVERTED is false into the edge information
1482 structure. Also record that any conditions dominated by COND are true
1483 as well.
1485 For example, if a < b is true, then a <= b must also be true. */
1487 static void
1488 record_conditions (struct edge_info *edge_info, tree cond, tree inverted)
1490 tree op0, op1;
1492 if (!COMPARISON_CLASS_P (cond))
1493 return;
1495 op0 = TREE_OPERAND (cond, 0);
1496 op1 = TREE_OPERAND (cond, 1);
1498 switch (TREE_CODE (cond))
1500 case LT_EXPR:
1501 case GT_EXPR:
1502 edge_info->max_cond_equivalences = 12;
1503 edge_info->cond_equivalences = xmalloc (12 * sizeof (tree));
1504 build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR
1505 ? LE_EXPR : GE_EXPR),
1506 op0, op1, &edge_info->cond_equivalences[4]);
1507 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1508 &edge_info->cond_equivalences[6]);
1509 build_and_record_new_cond (NE_EXPR, op0, op1,
1510 &edge_info->cond_equivalences[8]);
1511 build_and_record_new_cond (LTGT_EXPR, op0, op1,
1512 &edge_info->cond_equivalences[10]);
1513 break;
1515 case GE_EXPR:
1516 case LE_EXPR:
1517 edge_info->max_cond_equivalences = 6;
1518 edge_info->cond_equivalences = xmalloc (6 * sizeof (tree));
1519 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1520 &edge_info->cond_equivalences[4]);
1521 break;
1523 case EQ_EXPR:
1524 edge_info->max_cond_equivalences = 10;
1525 edge_info->cond_equivalences = xmalloc (10 * sizeof (tree));
1526 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1527 &edge_info->cond_equivalences[4]);
1528 build_and_record_new_cond (LE_EXPR, op0, op1,
1529 &edge_info->cond_equivalences[6]);
1530 build_and_record_new_cond (GE_EXPR, op0, op1,
1531 &edge_info->cond_equivalences[8]);
1532 break;
1534 case UNORDERED_EXPR:
1535 edge_info->max_cond_equivalences = 16;
1536 edge_info->cond_equivalences = xmalloc (16 * sizeof (tree));
1537 build_and_record_new_cond (NE_EXPR, op0, op1,
1538 &edge_info->cond_equivalences[4]);
1539 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1540 &edge_info->cond_equivalences[6]);
1541 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1542 &edge_info->cond_equivalences[8]);
1543 build_and_record_new_cond (UNEQ_EXPR, op0, op1,
1544 &edge_info->cond_equivalences[10]);
1545 build_and_record_new_cond (UNLT_EXPR, op0, op1,
1546 &edge_info->cond_equivalences[12]);
1547 build_and_record_new_cond (UNGT_EXPR, op0, op1,
1548 &edge_info->cond_equivalences[14]);
1549 break;
1551 case UNLT_EXPR:
1552 case UNGT_EXPR:
1553 edge_info->max_cond_equivalences = 8;
1554 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1555 build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR
1556 ? UNLE_EXPR : UNGE_EXPR),
1557 op0, op1, &edge_info->cond_equivalences[4]);
1558 build_and_record_new_cond (NE_EXPR, op0, op1,
1559 &edge_info->cond_equivalences[6]);
1560 break;
1562 case UNEQ_EXPR:
1563 edge_info->max_cond_equivalences = 8;
1564 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1565 build_and_record_new_cond (UNLE_EXPR, op0, op1,
1566 &edge_info->cond_equivalences[4]);
1567 build_and_record_new_cond (UNGE_EXPR, op0, op1,
1568 &edge_info->cond_equivalences[6]);
1569 break;
1571 case LTGT_EXPR:
1572 edge_info->max_cond_equivalences = 8;
1573 edge_info->cond_equivalences = xmalloc (8 * sizeof (tree));
1574 build_and_record_new_cond (NE_EXPR, op0, op1,
1575 &edge_info->cond_equivalences[4]);
1576 build_and_record_new_cond (ORDERED_EXPR, op0, op1,
1577 &edge_info->cond_equivalences[6]);
1578 break;
1580 default:
1581 edge_info->max_cond_equivalences = 4;
1582 edge_info->cond_equivalences = xmalloc (4 * sizeof (tree));
1583 break;
1586 /* Now store the original true and false conditions into the first
1587 two slots. */
1588 edge_info->cond_equivalences[0] = cond;
1589 edge_info->cond_equivalences[1] = boolean_true_node;
1590 edge_info->cond_equivalences[2] = inverted;
1591 edge_info->cond_equivalences[3] = boolean_false_node;
1594 /* A helper function for record_const_or_copy and record_equality.
1595 Do the work of recording the value and undo info. */
1597 static void
1598 record_const_or_copy_1 (tree x, tree y, tree prev_x)
1600 SSA_NAME_VALUE (x) = y;
1602 VEC_reserve (tree, heap, const_and_copies_stack, 2);
1603 VEC_quick_push (tree, const_and_copies_stack, prev_x);
1604 VEC_quick_push (tree, const_and_copies_stack, x);
1608 /* Return the loop depth of the basic block of the defining statement of X.
1609 This number should not be treated as absolutely correct because the loop
1610 information may not be completely up-to-date when dom runs. However, it
1611 will be relatively correct, and as more passes are taught to keep loop info
1612 up to date, the result will become more and more accurate. */
1615 loop_depth_of_name (tree x)
1617 tree defstmt;
1618 basic_block defbb;
1620 /* If it's not an SSA_NAME, we have no clue where the definition is. */
1621 if (TREE_CODE (x) != SSA_NAME)
1622 return 0;
1624 /* Otherwise return the loop depth of the defining statement's bb.
1625 Note that there may not actually be a bb for this statement, if the
1626 ssa_name is live on entry. */
1627 defstmt = SSA_NAME_DEF_STMT (x);
1628 defbb = bb_for_stmt (defstmt);
1629 if (!defbb)
1630 return 0;
1632 return defbb->loop_depth;
1636 /* Record that X is equal to Y in const_and_copies. Record undo
1637 information in the block-local vector. */
1639 static void
1640 record_const_or_copy (tree x, tree y)
1642 tree prev_x = SSA_NAME_VALUE (x);
1644 if (TREE_CODE (y) == SSA_NAME)
1646 tree tmp = SSA_NAME_VALUE (y);
1647 if (tmp)
1648 y = tmp;
1651 record_const_or_copy_1 (x, y, prev_x);
1654 /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR.
1655 This constrains the cases in which we may treat this as assignment. */
1657 static void
1658 record_equality (tree x, tree y)
1660 tree prev_x = NULL, prev_y = NULL;
1662 if (TREE_CODE (x) == SSA_NAME)
1663 prev_x = SSA_NAME_VALUE (x);
1664 if (TREE_CODE (y) == SSA_NAME)
1665 prev_y = SSA_NAME_VALUE (y);
1667 /* If one of the previous values is invariant, or invariant in more loops
1668 (by depth), then use that.
1669 Otherwise it doesn't matter which value we choose, just so
1670 long as we canonicalize on one value. */
1671 if (TREE_INVARIANT (y))
1673 else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y)))
1674 prev_x = x, x = y, y = prev_x, prev_x = prev_y;
1675 else if (prev_x && TREE_INVARIANT (prev_x))
1676 x = y, y = prev_x, prev_x = prev_y;
1677 else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE)
1678 y = prev_y;
1680 /* After the swapping, we must have one SSA_NAME. */
1681 if (TREE_CODE (x) != SSA_NAME)
1682 return;
1684 /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a
1685 variable compared against zero. If we're honoring signed zeros,
1686 then we cannot record this value unless we know that the value is
1687 nonzero. */
1688 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x)))
1689 && (TREE_CODE (y) != REAL_CST
1690 || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y))))
1691 return;
1693 record_const_or_copy_1 (x, y, prev_x);
1696 /* Return true, if it is ok to do folding of an associative expression.
1697 EXP is the tree for the associative expression. */
1699 static inline bool
1700 unsafe_associative_fp_binop (tree exp)
1702 enum tree_code code = TREE_CODE (exp);
1703 return !(!flag_unsafe_math_optimizations
1704 && (code == MULT_EXPR || code == PLUS_EXPR
1705 || code == MINUS_EXPR)
1706 && FLOAT_TYPE_P (TREE_TYPE (exp)));
1709 /* Returns true when STMT is a simple iv increment. It detects the
1710 following situation:
1712 i_1 = phi (..., i_2)
1713 i_2 = i_1 +/- ... */
1715 static bool
1716 simple_iv_increment_p (tree stmt)
1718 tree lhs, rhs, preinc, phi;
1719 unsigned i;
1721 if (TREE_CODE (stmt) != MODIFY_EXPR)
1722 return false;
1724 lhs = TREE_OPERAND (stmt, 0);
1725 if (TREE_CODE (lhs) != SSA_NAME)
1726 return false;
1728 rhs = TREE_OPERAND (stmt, 1);
1730 if (TREE_CODE (rhs) != PLUS_EXPR
1731 && TREE_CODE (rhs) != MINUS_EXPR)
1732 return false;
1734 preinc = TREE_OPERAND (rhs, 0);
1735 if (TREE_CODE (preinc) != SSA_NAME)
1736 return false;
1738 phi = SSA_NAME_DEF_STMT (preinc);
1739 if (TREE_CODE (phi) != PHI_NODE)
1740 return false;
1742 for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++)
1743 if (PHI_ARG_DEF (phi, i) == lhs)
1744 return true;
1746 return false;
1749 /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the
1750 hash tables. Try to simplify the RHS using whatever equivalences
1751 we may have recorded.
1753 If we are able to simplify the RHS, then lookup the simplified form in
1754 the hash table and return the result. Otherwise return NULL. */
1756 static tree
1757 simplify_rhs_and_lookup_avail_expr (tree stmt, int insert)
1759 tree rhs = TREE_OPERAND (stmt, 1);
1760 enum tree_code rhs_code = TREE_CODE (rhs);
1761 tree result = NULL;
1763 /* If we have lhs = ~x, look and see if we earlier had x = ~y.
1764 In which case we can change this statement to be lhs = y.
1765 Which can then be copy propagated.
1767 Similarly for negation. */
1768 if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR)
1769 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME)
1771 /* Get the definition statement for our RHS. */
1772 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1774 /* See if the RHS_DEF_STMT has the same form as our statement. */
1775 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR
1776 && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code)
1778 tree rhs_def_operand;
1780 rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0);
1782 /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */
1783 if (TREE_CODE (rhs_def_operand) == SSA_NAME
1784 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand))
1785 result = update_rhs_and_lookup_avail_expr (stmt,
1786 rhs_def_operand,
1787 insert);
1791 /* If we have z = (x OP C1), see if we earlier had x = y OP C2.
1792 If OP is associative, create and fold (y OP C2) OP C1 which
1793 should result in (y OP C3), use that as the RHS for the
1794 assignment. Add minus to this, as we handle it specially below. */
1795 if ((associative_tree_code (rhs_code) || rhs_code == MINUS_EXPR)
1796 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME
1797 && is_gimple_min_invariant (TREE_OPERAND (rhs, 1)))
1799 tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
1801 /* If the statement defines an induction variable, do not propagate
1802 its value, so that we do not create overlapping life ranges. */
1803 if (simple_iv_increment_p (rhs_def_stmt))
1804 goto dont_fold_assoc;
1806 /* See if the RHS_DEF_STMT has the same form as our statement. */
1807 if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR)
1809 tree rhs_def_rhs = TREE_OPERAND (rhs_def_stmt, 1);
1810 enum tree_code rhs_def_code = TREE_CODE (rhs_def_rhs);
1812 if ((rhs_code == rhs_def_code && unsafe_associative_fp_binop (rhs))
1813 || (rhs_code == PLUS_EXPR && rhs_def_code == MINUS_EXPR)
1814 || (rhs_code == MINUS_EXPR && rhs_def_code == PLUS_EXPR))
1816 tree def_stmt_op0 = TREE_OPERAND (rhs_def_rhs, 0);
1817 tree def_stmt_op1 = TREE_OPERAND (rhs_def_rhs, 1);
1819 if (TREE_CODE (def_stmt_op0) == SSA_NAME
1820 && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0)
1821 && is_gimple_min_invariant (def_stmt_op1))
1823 tree outer_const = TREE_OPERAND (rhs, 1);
1824 tree type = TREE_TYPE (TREE_OPERAND (stmt, 0));
1825 tree t;
1827 /* If we care about correct floating point results, then
1828 don't fold x + c1 - c2. Note that we need to take both
1829 the codes and the signs to figure this out. */
1830 if (FLOAT_TYPE_P (type)
1831 && !flag_unsafe_math_optimizations
1832 && (rhs_def_code == PLUS_EXPR
1833 || rhs_def_code == MINUS_EXPR))
1835 bool neg = false;
1837 neg ^= (rhs_code == MINUS_EXPR);
1838 neg ^= (rhs_def_code == MINUS_EXPR);
1839 neg ^= real_isneg (TREE_REAL_CST_PTR (outer_const));
1840 neg ^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1));
1842 if (neg)
1843 goto dont_fold_assoc;
1846 /* Ho hum. So fold will only operate on the outermost
1847 thingy that we give it, so we have to build the new
1848 expression in two pieces. This requires that we handle
1849 combinations of plus and minus. */
1850 if (rhs_def_code != rhs_code)
1852 if (rhs_def_code == MINUS_EXPR)
1853 t = build (MINUS_EXPR, type, outer_const, def_stmt_op1);
1854 else
1855 t = build (MINUS_EXPR, type, def_stmt_op1, outer_const);
1856 rhs_code = PLUS_EXPR;
1858 else if (rhs_def_code == MINUS_EXPR)
1859 t = build (PLUS_EXPR, type, def_stmt_op1, outer_const);
1860 else
1861 t = build (rhs_def_code, type, def_stmt_op1, outer_const);
1862 t = local_fold (t);
1863 t = build (rhs_code, type, def_stmt_op0, t);
1864 t = local_fold (t);
1866 /* If the result is a suitable looking gimple expression,
1867 then use it instead of the original for STMT. */
1868 if (TREE_CODE (t) == SSA_NAME
1869 || (UNARY_CLASS_P (t)
1870 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME)
1871 || ((BINARY_CLASS_P (t) || COMPARISON_CLASS_P (t))
1872 && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME
1873 && is_gimple_val (TREE_OPERAND (t, 1))))
1874 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1878 dont_fold_assoc:;
1881 /* Optimize *"foo" into 'f'. This is done here rather than
1882 in fold to avoid problems with stuff like &*"foo". */
1883 if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF)
1885 tree t = fold_read_from_constant_string (rhs);
1887 if (t)
1888 result = update_rhs_and_lookup_avail_expr (stmt, t, insert);
1891 return result;
1894 /* COND is a condition of the form:
1896 x == const or x != const
1898 Look back to x's defining statement and see if x is defined as
1900 x = (type) y;
1902 If const is unchanged if we convert it to type, then we can build
1903 the equivalent expression:
1906 y == const or y != const
1908 Which may allow further optimizations.
1910 Return the equivalent comparison or NULL if no such equivalent comparison
1911 was found. */
1913 static tree
1914 find_equivalent_equality_comparison (tree cond)
1916 tree op0 = TREE_OPERAND (cond, 0);
1917 tree op1 = TREE_OPERAND (cond, 1);
1918 tree def_stmt = SSA_NAME_DEF_STMT (op0);
1920 /* OP0 might have been a parameter, so first make sure it
1921 was defined by a MODIFY_EXPR. */
1922 if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR)
1924 tree def_rhs = TREE_OPERAND (def_stmt, 1);
1927 /* If either operand to the comparison is a pointer to
1928 a function, then we can not apply this optimization
1929 as some targets require function pointers to be
1930 canonicalized and in this case this optimization would
1931 eliminate a necessary canonicalization. */
1932 if ((POINTER_TYPE_P (TREE_TYPE (op0))
1933 && TREE_CODE (TREE_TYPE (TREE_TYPE (op0))) == FUNCTION_TYPE)
1934 || (POINTER_TYPE_P (TREE_TYPE (op1))
1935 && TREE_CODE (TREE_TYPE (TREE_TYPE (op1))) == FUNCTION_TYPE))
1936 return NULL;
1938 /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */
1939 if ((TREE_CODE (def_rhs) == NOP_EXPR
1940 || TREE_CODE (def_rhs) == CONVERT_EXPR)
1941 && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME)
1943 tree def_rhs_inner = TREE_OPERAND (def_rhs, 0);
1944 tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner);
1945 tree new;
1947 if (TYPE_PRECISION (def_rhs_inner_type)
1948 > TYPE_PRECISION (TREE_TYPE (def_rhs)))
1949 return NULL;
1951 /* If the inner type of the conversion is a pointer to
1952 a function, then we can not apply this optimization
1953 as some targets require function pointers to be
1954 canonicalized. This optimization would result in
1955 canonicalization of the pointer when it was not originally
1956 needed/intended. */
1957 if (POINTER_TYPE_P (def_rhs_inner_type)
1958 && TREE_CODE (TREE_TYPE (def_rhs_inner_type)) == FUNCTION_TYPE)
1959 return NULL;
1961 /* What we want to prove is that if we convert OP1 to
1962 the type of the object inside the NOP_EXPR that the
1963 result is still equivalent to SRC.
1965 If that is true, the build and return new equivalent
1966 condition which uses the source of the typecast and the
1967 new constant (which has only changed its type). */
1968 new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1);
1969 new = local_fold (new);
1970 if (is_gimple_val (new) && tree_int_cst_equal (new, op1))
1971 return build (TREE_CODE (cond), TREE_TYPE (cond),
1972 def_rhs_inner, new);
1975 return NULL;
1978 /* STMT is a COND_EXPR for which we could not trivially determine its
1979 result. This routine attempts to find equivalent forms of the
1980 condition which we may be able to optimize better. It also
1981 uses simple value range propagation to optimize conditionals. */
1983 static tree
1984 simplify_cond_and_lookup_avail_expr (tree stmt,
1985 stmt_ann_t ann,
1986 int insert)
1988 tree cond = COND_EXPR_COND (stmt);
1990 if (COMPARISON_CLASS_P (cond))
1992 tree op0 = TREE_OPERAND (cond, 0);
1993 tree op1 = TREE_OPERAND (cond, 1);
1995 if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1))
1997 int limit;
1998 tree low, high, cond_low, cond_high;
1999 int lowequal, highequal, swapped, no_overlap, subset, cond_inverted;
2000 VEC(vrp_element_p,heap) **vrp_records;
2001 struct vrp_element *element;
2002 struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p;
2003 void **slot;
2005 /* First see if we have test of an SSA_NAME against a constant
2006 where the SSA_NAME is defined by an earlier typecast which
2007 is irrelevant when performing tests against the given
2008 constant. */
2009 if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2011 tree new_cond = find_equivalent_equality_comparison (cond);
2013 if (new_cond)
2015 /* Update the statement to use the new equivalent
2016 condition. */
2017 COND_EXPR_COND (stmt) = new_cond;
2019 /* If this is not a real stmt, ann will be NULL and we
2020 avoid processing the operands. */
2021 if (ann)
2022 mark_stmt_modified (stmt);
2024 /* Lookup the condition and return its known value if it
2025 exists. */
2026 new_cond = lookup_avail_expr (stmt, insert);
2027 if (new_cond)
2028 return new_cond;
2030 /* The operands have changed, so update op0 and op1. */
2031 op0 = TREE_OPERAND (cond, 0);
2032 op1 = TREE_OPERAND (cond, 1);
2036 /* Consult the value range records for this variable (if they exist)
2037 to see if we can eliminate or simplify this conditional.
2039 Note two tests are necessary to determine no records exist.
2040 First we have to see if the virtual array exists, if it
2041 exists, then we have to check its active size.
2043 Also note the vast majority of conditionals are not testing
2044 a variable which has had its range constrained by an earlier
2045 conditional. So this filter avoids a lot of unnecessary work. */
2046 vrp_hash_elt.var = op0;
2047 vrp_hash_elt.records = NULL;
2048 slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT);
2049 if (slot == NULL)
2050 return NULL;
2052 vrp_hash_elt_p = (struct vrp_hash_elt *) *slot;
2053 vrp_records = &vrp_hash_elt_p->records;
2055 limit = VEC_length (vrp_element_p, *vrp_records);
2057 /* If we have no value range records for this variable, or we are
2058 unable to extract a range for this condition, then there is
2059 nothing to do. */
2060 if (limit == 0
2061 || ! extract_range_from_cond (cond, &cond_high,
2062 &cond_low, &cond_inverted))
2063 return NULL;
2065 /* We really want to avoid unnecessary computations of range
2066 info. So all ranges are computed lazily; this avoids a
2067 lot of unnecessary work. i.e., we record the conditional,
2068 but do not process how it constrains the variable's
2069 potential values until we know that processing the condition
2070 could be helpful.
2072 However, we do not want to have to walk a potentially long
2073 list of ranges, nor do we want to compute a variable's
2074 range more than once for a given path.
2076 Luckily, each time we encounter a conditional that can not
2077 be otherwise optimized we will end up here and we will
2078 compute the necessary range information for the variable
2079 used in this condition.
2081 Thus you can conclude that there will never be more than one
2082 conditional associated with a variable which has not been
2083 processed. So we never need to merge more than one new
2084 conditional into the current range.
2086 These properties also help us avoid unnecessary work. */
2087 element = VEC_last (vrp_element_p, *vrp_records);
2089 if (element->high && element->low)
2091 /* The last element has been processed, so there is no range
2092 merging to do, we can simply use the high/low values
2093 recorded in the last element. */
2094 low = element->low;
2095 high = element->high;
2097 else
2099 tree tmp_high, tmp_low;
2100 int dummy;
2102 /* The last element has not been processed. Process it now.
2103 record_range should ensure for cond inverted is not set.
2104 This call can only fail if cond is x < min or x > max,
2105 which fold should have optimized into false.
2106 If that doesn't happen, just pretend all values are
2107 in the range. */
2108 if (! extract_range_from_cond (element->cond, &tmp_high,
2109 &tmp_low, &dummy))
2110 gcc_unreachable ();
2111 else
2112 gcc_assert (dummy == 0);
2114 /* If this is the only element, then no merging is necessary,
2115 the high/low values from extract_range_from_cond are all
2116 we need. */
2117 if (limit == 1)
2119 low = tmp_low;
2120 high = tmp_high;
2122 else
2124 /* Get the high/low value from the previous element. */
2125 struct vrp_element *prev
2126 = VEC_index (vrp_element_p, *vrp_records, limit - 2);
2127 low = prev->low;
2128 high = prev->high;
2130 /* Merge in this element's range with the range from the
2131 previous element.
2133 The low value for the merged range is the maximum of
2134 the previous low value and the low value of this record.
2136 Similarly the high value for the merged range is the
2137 minimum of the previous high value and the high value of
2138 this record. */
2139 low = (low && tree_int_cst_compare (low, tmp_low) == 1
2140 ? low : tmp_low);
2141 high = (high && tree_int_cst_compare (high, tmp_high) == -1
2142 ? high : tmp_high);
2145 /* And record the computed range. */
2146 element->low = low;
2147 element->high = high;
2151 /* After we have constrained this variable's potential values,
2152 we try to determine the result of the given conditional.
2154 To simplify later tests, first determine if the current
2155 low value is the same low value as the conditional.
2156 Similarly for the current high value and the high value
2157 for the conditional. */
2158 lowequal = tree_int_cst_equal (low, cond_low);
2159 highequal = tree_int_cst_equal (high, cond_high);
2161 if (lowequal && highequal)
2162 return (cond_inverted ? boolean_false_node : boolean_true_node);
2164 /* To simplify the overlap/subset tests below we may want
2165 to swap the two ranges so that the larger of the two
2166 ranges occurs "first". */
2167 swapped = 0;
2168 if (tree_int_cst_compare (low, cond_low) == 1
2169 || (lowequal
2170 && tree_int_cst_compare (cond_high, high) == 1))
2172 tree temp;
2174 swapped = 1;
2175 temp = low;
2176 low = cond_low;
2177 cond_low = temp;
2178 temp = high;
2179 high = cond_high;
2180 cond_high = temp;
2183 /* Now determine if there is no overlap in the ranges
2184 or if the second range is a subset of the first range. */
2185 no_overlap = tree_int_cst_lt (high, cond_low);
2186 subset = tree_int_cst_compare (cond_high, high) != 1;
2188 /* If there was no overlap in the ranges, then this conditional
2189 always has a false value (unless we had to invert this
2190 conditional, in which case it always has a true value). */
2191 if (no_overlap)
2192 return (cond_inverted ? boolean_true_node : boolean_false_node);
2194 /* If the current range is a subset of the condition's range,
2195 then this conditional always has a true value (unless we
2196 had to invert this conditional, in which case it always
2197 has a true value). */
2198 if (subset && swapped)
2199 return (cond_inverted ? boolean_false_node : boolean_true_node);
2201 /* We were unable to determine the result of the conditional.
2202 However, we may be able to simplify the conditional. First
2203 merge the ranges in the same manner as range merging above. */
2204 low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low;
2205 high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high;
2207 /* If the range has converged to a single point, then turn this
2208 into an equality comparison. */
2209 if (TREE_CODE (cond) != EQ_EXPR
2210 && TREE_CODE (cond) != NE_EXPR
2211 && tree_int_cst_equal (low, high))
2213 TREE_SET_CODE (cond, EQ_EXPR);
2214 TREE_OPERAND (cond, 1) = high;
2218 return 0;
2221 /* STMT is a SWITCH_EXPR for which we could not trivially determine its
2222 result. This routine attempts to find equivalent forms of the
2223 condition which we may be able to optimize better. */
2225 static tree
2226 simplify_switch_and_lookup_avail_expr (tree stmt, int insert)
2228 tree cond = SWITCH_COND (stmt);
2229 tree def, to, ti;
2231 /* The optimization that we really care about is removing unnecessary
2232 casts. That will let us do much better in propagating the inferred
2233 constant at the switch target. */
2234 if (TREE_CODE (cond) == SSA_NAME)
2236 def = SSA_NAME_DEF_STMT (cond);
2237 if (TREE_CODE (def) == MODIFY_EXPR)
2239 def = TREE_OPERAND (def, 1);
2240 if (TREE_CODE (def) == NOP_EXPR)
2242 int need_precision;
2243 bool fail;
2245 def = TREE_OPERAND (def, 0);
2247 #ifdef ENABLE_CHECKING
2248 /* ??? Why was Jeff testing this? We are gimple... */
2249 gcc_assert (is_gimple_val (def));
2250 #endif
2252 to = TREE_TYPE (cond);
2253 ti = TREE_TYPE (def);
2255 /* If we have an extension that preserves value, then we
2256 can copy the source value into the switch. */
2258 need_precision = TYPE_PRECISION (ti);
2259 fail = false;
2260 if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti))
2261 fail = true;
2262 else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti))
2263 need_precision += 1;
2264 if (TYPE_PRECISION (to) < need_precision)
2265 fail = true;
2267 if (!fail)
2269 SWITCH_COND (stmt) = def;
2270 mark_stmt_modified (stmt);
2272 return lookup_avail_expr (stmt, insert);
2278 return 0;
2282 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2283 known value for that SSA_NAME (or NULL if no value is known).
2285 NONZERO_VARS is the set SSA_NAMES known to have a nonzero value,
2286 even if we don't know their precise value.
2288 Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI
2289 nodes of the successors of BB. */
2291 static void
2292 cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars)
2294 edge e;
2295 edge_iterator ei;
2297 FOR_EACH_EDGE (e, ei, bb->succs)
2299 tree phi;
2300 int indx;
2302 /* If this is an abnormal edge, then we do not want to copy propagate
2303 into the PHI alternative associated with this edge. */
2304 if (e->flags & EDGE_ABNORMAL)
2305 continue;
2307 phi = phi_nodes (e->dest);
2308 if (! phi)
2309 continue;
2311 indx = e->dest_idx;
2312 for ( ; phi; phi = PHI_CHAIN (phi))
2314 tree new;
2315 use_operand_p orig_p;
2316 tree orig;
2318 /* The alternative may be associated with a constant, so verify
2319 it is an SSA_NAME before doing anything with it. */
2320 orig_p = PHI_ARG_DEF_PTR (phi, indx);
2321 orig = USE_FROM_PTR (orig_p);
2322 if (TREE_CODE (orig) != SSA_NAME)
2323 continue;
2325 /* If the alternative is known to have a nonzero value, record
2326 that fact in the PHI node itself for future use. */
2327 if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig)))
2328 PHI_ARG_NONZERO (phi, indx) = true;
2330 /* If we have *ORIG_P in our constant/copy table, then replace
2331 ORIG_P with its value in our constant/copy table. */
2332 new = SSA_NAME_VALUE (orig);
2333 if (new
2334 && new != orig
2335 && (TREE_CODE (new) == SSA_NAME
2336 || is_gimple_min_invariant (new))
2337 && may_propagate_copy (orig, new))
2338 propagate_value (orig_p, new);
2343 /* We have finished optimizing BB, record any information implied by
2344 taking a specific outgoing edge from BB. */
2346 static void
2347 record_edge_info (basic_block bb)
2349 block_stmt_iterator bsi = bsi_last (bb);
2350 struct edge_info *edge_info;
2352 if (! bsi_end_p (bsi))
2354 tree stmt = bsi_stmt (bsi);
2356 if (stmt && TREE_CODE (stmt) == SWITCH_EXPR)
2358 tree cond = SWITCH_COND (stmt);
2360 if (TREE_CODE (cond) == SSA_NAME)
2362 tree labels = SWITCH_LABELS (stmt);
2363 int i, n_labels = TREE_VEC_LENGTH (labels);
2364 tree *info = xcalloc (last_basic_block, sizeof (tree));
2365 edge e;
2366 edge_iterator ei;
2368 for (i = 0; i < n_labels; i++)
2370 tree label = TREE_VEC_ELT (labels, i);
2371 basic_block target_bb = label_to_block (CASE_LABEL (label));
2373 if (CASE_HIGH (label)
2374 || !CASE_LOW (label)
2375 || info[target_bb->index])
2376 info[target_bb->index] = error_mark_node;
2377 else
2378 info[target_bb->index] = label;
2381 FOR_EACH_EDGE (e, ei, bb->succs)
2383 basic_block target_bb = e->dest;
2384 tree node = info[target_bb->index];
2386 if (node != NULL && node != error_mark_node)
2388 tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
2389 edge_info = allocate_edge_info (e);
2390 edge_info->lhs = cond;
2391 edge_info->rhs = x;
2394 free (info);
2398 /* A COND_EXPR may create equivalences too. */
2399 if (stmt && TREE_CODE (stmt) == COND_EXPR)
2401 tree cond = COND_EXPR_COND (stmt);
2402 edge true_edge;
2403 edge false_edge;
2405 extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
2407 /* If the conditional is a single variable 'X', record 'X = 1'
2408 for the true edge and 'X = 0' on the false edge. */
2409 if (SSA_VAR_P (cond))
2411 struct edge_info *edge_info;
2413 edge_info = allocate_edge_info (true_edge);
2414 edge_info->lhs = cond;
2415 edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond));
2417 edge_info = allocate_edge_info (false_edge);
2418 edge_info->lhs = cond;
2419 edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond));
2421 /* Equality tests may create one or two equivalences. */
2422 else if (COMPARISON_CLASS_P (cond))
2424 tree op0 = TREE_OPERAND (cond, 0);
2425 tree op1 = TREE_OPERAND (cond, 1);
2427 /* Special case comparing booleans against a constant as we
2428 know the value of OP0 on both arms of the branch. i.e., we
2429 can record an equivalence for OP0 rather than COND. */
2430 if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR)
2431 && TREE_CODE (op0) == SSA_NAME
2432 && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE
2433 && is_gimple_min_invariant (op1))
2435 if (TREE_CODE (cond) == EQ_EXPR)
2437 edge_info = allocate_edge_info (true_edge);
2438 edge_info->lhs = op0;
2439 edge_info->rhs = (integer_zerop (op1)
2440 ? boolean_false_node
2441 : boolean_true_node);
2443 edge_info = allocate_edge_info (false_edge);
2444 edge_info->lhs = op0;
2445 edge_info->rhs = (integer_zerop (op1)
2446 ? boolean_true_node
2447 : boolean_false_node);
2449 else
2451 edge_info = allocate_edge_info (true_edge);
2452 edge_info->lhs = op0;
2453 edge_info->rhs = (integer_zerop (op1)
2454 ? boolean_true_node
2455 : boolean_false_node);
2457 edge_info = allocate_edge_info (false_edge);
2458 edge_info->lhs = op0;
2459 edge_info->rhs = (integer_zerop (op1)
2460 ? boolean_false_node
2461 : boolean_true_node);
2465 else if (is_gimple_min_invariant (op0)
2466 && (TREE_CODE (op1) == SSA_NAME
2467 || is_gimple_min_invariant (op1)))
2469 tree inverted = invert_truthvalue (cond);
2470 struct edge_info *edge_info;
2472 edge_info = allocate_edge_info (true_edge);
2473 record_conditions (edge_info, cond, inverted);
2475 if (TREE_CODE (cond) == EQ_EXPR)
2477 edge_info->lhs = op1;
2478 edge_info->rhs = op0;
2481 edge_info = allocate_edge_info (false_edge);
2482 record_conditions (edge_info, inverted, cond);
2484 if (TREE_CODE (cond) == NE_EXPR)
2486 edge_info->lhs = op1;
2487 edge_info->rhs = op0;
2491 else if (TREE_CODE (op0) == SSA_NAME
2492 && (is_gimple_min_invariant (op1)
2493 || TREE_CODE (op1) == SSA_NAME))
2495 tree inverted = invert_truthvalue (cond);
2496 struct edge_info *edge_info;
2498 edge_info = allocate_edge_info (true_edge);
2499 record_conditions (edge_info, cond, inverted);
2501 if (TREE_CODE (cond) == EQ_EXPR)
2503 edge_info->lhs = op0;
2504 edge_info->rhs = op1;
2507 edge_info = allocate_edge_info (false_edge);
2508 record_conditions (edge_info, inverted, cond);
2510 if (TREE_CODE (cond) == NE_EXPR)
2512 edge_info->lhs = op0;
2513 edge_info->rhs = op1;
2518 /* ??? TRUTH_NOT_EXPR can create an equivalence too. */
2523 /* Propagate information from BB to its outgoing edges.
2525 This can include equivalency information implied by control statements
2526 at the end of BB and const/copy propagation into PHIs in BB's
2527 successor blocks. */
2529 static void
2530 propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2531 basic_block bb)
2533 record_edge_info (bb);
2534 cprop_into_successor_phis (bb, nonzero_vars);
2537 /* Search for redundant computations in STMT. If any are found, then
2538 replace them with the variable holding the result of the computation.
2540 If safe, record this expression into the available expression hash
2541 table. */
2543 static bool
2544 eliminate_redundant_computations (tree stmt, stmt_ann_t ann)
2546 tree *expr_p, def = NULL_TREE;
2547 bool insert = true;
2548 tree cached_lhs;
2549 bool retval = false;
2550 bool modify_expr_p = false;
2552 if (TREE_CODE (stmt) == MODIFY_EXPR)
2553 def = TREE_OPERAND (stmt, 0);
2555 /* Certain expressions on the RHS can be optimized away, but can not
2556 themselves be entered into the hash tables. */
2557 if (ann->makes_aliased_stores
2558 || ! def
2559 || TREE_CODE (def) != SSA_NAME
2560 || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def)
2561 || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VMAYDEF)
2562 /* Do not record equivalences for increments of ivs. This would create
2563 overlapping live ranges for a very questionable gain. */
2564 || simple_iv_increment_p (stmt))
2565 insert = false;
2567 /* Check if the expression has been computed before. */
2568 cached_lhs = lookup_avail_expr (stmt, insert);
2570 /* If this is an assignment and the RHS was not in the hash table,
2571 then try to simplify the RHS and lookup the new RHS in the
2572 hash table. */
2573 if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR)
2574 cached_lhs = simplify_rhs_and_lookup_avail_expr (stmt, insert);
2575 /* Similarly if this is a COND_EXPR and we did not find its
2576 expression in the hash table, simplify the condition and
2577 try again. */
2578 else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR)
2579 cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert);
2580 /* Similarly for a SWITCH_EXPR. */
2581 else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR)
2582 cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert);
2584 opt_stats.num_exprs_considered++;
2586 /* Get a pointer to the expression we are trying to optimize. */
2587 if (TREE_CODE (stmt) == COND_EXPR)
2588 expr_p = &COND_EXPR_COND (stmt);
2589 else if (TREE_CODE (stmt) == SWITCH_EXPR)
2590 expr_p = &SWITCH_COND (stmt);
2591 else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0))
2593 expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1);
2594 modify_expr_p = true;
2596 else
2598 expr_p = &TREE_OPERAND (stmt, 1);
2599 modify_expr_p = true;
2602 /* It is safe to ignore types here since we have already done
2603 type checking in the hashing and equality routines. In fact
2604 type checking here merely gets in the way of constant
2605 propagation. Also, make sure that it is safe to propagate
2606 CACHED_LHS into *EXPR_P. */
2607 if (cached_lhs
2608 && ((TREE_CODE (cached_lhs) != SSA_NAME
2609 && (modify_expr_p
2610 || tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
2611 TREE_TYPE (cached_lhs))))
2612 || may_propagate_copy (*expr_p, cached_lhs)))
2614 if (dump_file && (dump_flags & TDF_DETAILS))
2616 fprintf (dump_file, " Replaced redundant expr '");
2617 print_generic_expr (dump_file, *expr_p, dump_flags);
2618 fprintf (dump_file, "' with '");
2619 print_generic_expr (dump_file, cached_lhs, dump_flags);
2620 fprintf (dump_file, "'\n");
2623 opt_stats.num_re++;
2625 #if defined ENABLE_CHECKING
2626 gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME
2627 || is_gimple_min_invariant (cached_lhs));
2628 #endif
2630 if (TREE_CODE (cached_lhs) == ADDR_EXPR
2631 || (POINTER_TYPE_P (TREE_TYPE (*expr_p))
2632 && is_gimple_min_invariant (cached_lhs)))
2633 retval = true;
2635 if (modify_expr_p
2636 && !tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p),
2637 TREE_TYPE (cached_lhs)))
2638 cached_lhs = fold_convert (TREE_TYPE (*expr_p), cached_lhs);
2640 propagate_tree_value (expr_p, cached_lhs);
2641 mark_stmt_modified (stmt);
2643 return retval;
2646 /* STMT, a MODIFY_EXPR, may create certain equivalences, in either
2647 the available expressions table or the const_and_copies table.
2648 Detect and record those equivalences. */
2650 static void
2651 record_equivalences_from_stmt (tree stmt,
2652 int may_optimize_p,
2653 stmt_ann_t ann)
2655 tree lhs = TREE_OPERAND (stmt, 0);
2656 enum tree_code lhs_code = TREE_CODE (lhs);
2657 int i;
2659 if (lhs_code == SSA_NAME)
2661 tree rhs = TREE_OPERAND (stmt, 1);
2663 /* Strip away any useless type conversions. */
2664 STRIP_USELESS_TYPE_CONVERSION (rhs);
2666 /* If the RHS of the assignment is a constant or another variable that
2667 may be propagated, register it in the CONST_AND_COPIES table. We
2668 do not need to record unwind data for this, since this is a true
2669 assignment and not an equivalence inferred from a comparison. All
2670 uses of this ssa name are dominated by this assignment, so unwinding
2671 just costs time and space. */
2672 if (may_optimize_p
2673 && (TREE_CODE (rhs) == SSA_NAME
2674 || is_gimple_min_invariant (rhs)))
2675 SSA_NAME_VALUE (lhs) = rhs;
2677 if (tree_expr_nonzero_p (rhs))
2678 record_var_is_nonzero (lhs);
2681 /* Look at both sides for pointer dereferences. If we find one, then
2682 the pointer must be nonnull and we can enter that equivalence into
2683 the hash tables. */
2684 if (flag_delete_null_pointer_checks)
2685 for (i = 0; i < 2; i++)
2687 tree t = TREE_OPERAND (stmt, i);
2689 /* Strip away any COMPONENT_REFs. */
2690 while (TREE_CODE (t) == COMPONENT_REF)
2691 t = TREE_OPERAND (t, 0);
2693 /* Now see if this is a pointer dereference. */
2694 if (INDIRECT_REF_P (t))
2696 tree op = TREE_OPERAND (t, 0);
2698 /* If the pointer is a SSA variable, then enter new
2699 equivalences into the hash table. */
2700 while (TREE_CODE (op) == SSA_NAME)
2702 tree def = SSA_NAME_DEF_STMT (op);
2704 record_var_is_nonzero (op);
2706 /* And walk up the USE-DEF chains noting other SSA_NAMEs
2707 which are known to have a nonzero value. */
2708 if (def
2709 && TREE_CODE (def) == MODIFY_EXPR
2710 && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR)
2711 op = TREE_OPERAND (TREE_OPERAND (def, 1), 0);
2712 else
2713 break;
2718 /* A memory store, even an aliased store, creates a useful
2719 equivalence. By exchanging the LHS and RHS, creating suitable
2720 vops and recording the result in the available expression table,
2721 we may be able to expose more redundant loads. */
2722 if (!ann->has_volatile_ops
2723 && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME
2724 || is_gimple_min_invariant (TREE_OPERAND (stmt, 1)))
2725 && !is_gimple_reg (lhs))
2727 tree rhs = TREE_OPERAND (stmt, 1);
2728 tree new;
2730 /* FIXME: If the LHS of the assignment is a bitfield and the RHS
2731 is a constant, we need to adjust the constant to fit into the
2732 type of the LHS. If the LHS is a bitfield and the RHS is not
2733 a constant, then we can not record any equivalences for this
2734 statement since we would need to represent the widening or
2735 narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c
2736 and should not be necessary if GCC represented bitfields
2737 properly. */
2738 if (lhs_code == COMPONENT_REF
2739 && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1)))
2741 if (TREE_CONSTANT (rhs))
2742 rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs);
2743 else
2744 rhs = NULL;
2746 /* If the value overflowed, then we can not use this equivalence. */
2747 if (rhs && ! is_gimple_min_invariant (rhs))
2748 rhs = NULL;
2751 if (rhs)
2753 /* Build a new statement with the RHS and LHS exchanged. */
2754 new = build (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs);
2756 create_ssa_artficial_load_stmt (new, stmt);
2758 /* Finally enter the statement into the available expression
2759 table. */
2760 lookup_avail_expr (new, true);
2765 /* Replace *OP_P in STMT with any known equivalent value for *OP_P from
2766 CONST_AND_COPIES. */
2768 static bool
2769 cprop_operand (tree stmt, use_operand_p op_p)
2771 bool may_have_exposed_new_symbols = false;
2772 tree val;
2773 tree op = USE_FROM_PTR (op_p);
2775 /* If the operand has a known constant value or it is known to be a
2776 copy of some other variable, use the value or copy stored in
2777 CONST_AND_COPIES. */
2778 val = SSA_NAME_VALUE (op);
2779 if (val && val != op && TREE_CODE (val) != VALUE_HANDLE)
2781 tree op_type, val_type;
2783 /* Do not change the base variable in the virtual operand
2784 tables. That would make it impossible to reconstruct
2785 the renamed virtual operand if we later modify this
2786 statement. Also only allow the new value to be an SSA_NAME
2787 for propagation into virtual operands. */
2788 if (!is_gimple_reg (op)
2789 && (TREE_CODE (val) != SSA_NAME
2790 || is_gimple_reg (val)
2791 || get_virtual_var (val) != get_virtual_var (op)))
2792 return false;
2794 /* Do not replace hard register operands in asm statements. */
2795 if (TREE_CODE (stmt) == ASM_EXPR
2796 && !may_propagate_copy_into_asm (op))
2797 return false;
2799 /* Get the toplevel type of each operand. */
2800 op_type = TREE_TYPE (op);
2801 val_type = TREE_TYPE (val);
2803 /* While both types are pointers, get the type of the object
2804 pointed to. */
2805 while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type))
2807 op_type = TREE_TYPE (op_type);
2808 val_type = TREE_TYPE (val_type);
2811 /* Make sure underlying types match before propagating a constant by
2812 converting the constant to the proper type. Note that convert may
2813 return a non-gimple expression, in which case we ignore this
2814 propagation opportunity. */
2815 if (TREE_CODE (val) != SSA_NAME)
2817 if (!lang_hooks.types_compatible_p (op_type, val_type))
2819 val = fold_convert (TREE_TYPE (op), val);
2820 if (!is_gimple_min_invariant (val))
2821 return false;
2825 /* Certain operands are not allowed to be copy propagated due
2826 to their interaction with exception handling and some GCC
2827 extensions. */
2828 else if (!may_propagate_copy (op, val))
2829 return false;
2831 /* Do not propagate copies if the propagated value is at a deeper loop
2832 depth than the propagatee. Otherwise, this may move loop variant
2833 variables outside of their loops and prevent coalescing
2834 opportunities. If the value was loop invariant, it will be hoisted
2835 by LICM and exposed for copy propagation. */
2836 if (loop_depth_of_name (val) > loop_depth_of_name (op))
2837 return false;
2839 /* Dump details. */
2840 if (dump_file && (dump_flags & TDF_DETAILS))
2842 fprintf (dump_file, " Replaced '");
2843 print_generic_expr (dump_file, op, dump_flags);
2844 fprintf (dump_file, "' with %s '",
2845 (TREE_CODE (val) != SSA_NAME ? "constant" : "variable"));
2846 print_generic_expr (dump_file, val, dump_flags);
2847 fprintf (dump_file, "'\n");
2850 /* If VAL is an ADDR_EXPR or a constant of pointer type, note
2851 that we may have exposed a new symbol for SSA renaming. */
2852 if (TREE_CODE (val) == ADDR_EXPR
2853 || (POINTER_TYPE_P (TREE_TYPE (op))
2854 && is_gimple_min_invariant (val)))
2855 may_have_exposed_new_symbols = true;
2857 if (TREE_CODE (val) != SSA_NAME)
2858 opt_stats.num_const_prop++;
2859 else
2860 opt_stats.num_copy_prop++;
2862 propagate_value (op_p, val);
2864 /* And note that we modified this statement. This is now
2865 safe, even if we changed virtual operands since we will
2866 rescan the statement and rewrite its operands again. */
2867 mark_stmt_modified (stmt);
2869 return may_have_exposed_new_symbols;
2872 /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current
2873 known value for that SSA_NAME (or NULL if no value is known).
2875 Propagate values from CONST_AND_COPIES into the uses, vuses and
2876 v_may_def_ops of STMT. */
2878 static bool
2879 cprop_into_stmt (tree stmt)
2881 bool may_have_exposed_new_symbols = false;
2882 use_operand_p op_p;
2883 ssa_op_iter iter;
2885 FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES)
2887 if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME)
2888 may_have_exposed_new_symbols |= cprop_operand (stmt, op_p);
2891 return may_have_exposed_new_symbols;
2895 /* Optimize the statement pointed to by iterator SI.
2897 We try to perform some simplistic global redundancy elimination and
2898 constant propagation:
2900 1- To detect global redundancy, we keep track of expressions that have
2901 been computed in this block and its dominators. If we find that the
2902 same expression is computed more than once, we eliminate repeated
2903 computations by using the target of the first one.
2905 2- Constant values and copy assignments. This is used to do very
2906 simplistic constant and copy propagation. When a constant or copy
2907 assignment is found, we map the value on the RHS of the assignment to
2908 the variable in the LHS in the CONST_AND_COPIES table. */
2910 static void
2911 optimize_stmt (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED,
2912 basic_block bb, block_stmt_iterator si)
2914 stmt_ann_t ann;
2915 tree stmt, old_stmt;
2916 bool may_optimize_p;
2917 bool may_have_exposed_new_symbols = false;
2919 old_stmt = stmt = bsi_stmt (si);
2921 update_stmt_if_modified (stmt);
2922 ann = stmt_ann (stmt);
2923 opt_stats.num_stmts++;
2924 may_have_exposed_new_symbols = false;
2926 if (dump_file && (dump_flags & TDF_DETAILS))
2928 fprintf (dump_file, "Optimizing statement ");
2929 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2932 /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */
2933 may_have_exposed_new_symbols = cprop_into_stmt (stmt);
2935 /* If the statement has been modified with constant replacements,
2936 fold its RHS before checking for redundant computations. */
2937 if (ann->modified)
2939 tree rhs;
2941 /* Try to fold the statement making sure that STMT is kept
2942 up to date. */
2943 if (fold_stmt (bsi_stmt_ptr (si)))
2945 stmt = bsi_stmt (si);
2946 ann = stmt_ann (stmt);
2948 if (dump_file && (dump_flags & TDF_DETAILS))
2950 fprintf (dump_file, " Folded to: ");
2951 print_generic_stmt (dump_file, stmt, TDF_SLIM);
2955 rhs = get_rhs (stmt);
2956 if (rhs && TREE_CODE (rhs) == ADDR_EXPR)
2957 recompute_tree_invarant_for_addr_expr (rhs);
2959 /* Constant/copy propagation above may change the set of
2960 virtual operands associated with this statement. Folding
2961 may remove the need for some virtual operands.
2963 Indicate we will need to rescan and rewrite the statement. */
2964 may_have_exposed_new_symbols = true;
2967 /* Check for redundant computations. Do this optimization only
2968 for assignments that have no volatile ops and conditionals. */
2969 may_optimize_p = (!ann->has_volatile_ops
2970 && ((TREE_CODE (stmt) == RETURN_EXPR
2971 && TREE_OPERAND (stmt, 0)
2972 && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR
2973 && ! (TREE_SIDE_EFFECTS
2974 (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1))))
2975 || (TREE_CODE (stmt) == MODIFY_EXPR
2976 && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1)))
2977 || TREE_CODE (stmt) == COND_EXPR
2978 || TREE_CODE (stmt) == SWITCH_EXPR));
2980 if (may_optimize_p)
2981 may_have_exposed_new_symbols
2982 |= eliminate_redundant_computations (stmt, ann);
2984 /* Record any additional equivalences created by this statement. */
2985 if (TREE_CODE (stmt) == MODIFY_EXPR)
2986 record_equivalences_from_stmt (stmt,
2987 may_optimize_p,
2988 ann);
2990 /* If STMT is a COND_EXPR and it was modified, then we may know
2991 where it goes. If that is the case, then mark the CFG as altered.
2993 This will cause us to later call remove_unreachable_blocks and
2994 cleanup_tree_cfg when it is safe to do so. It is not safe to
2995 clean things up here since removal of edges and such can trigger
2996 the removal of PHI nodes, which in turn can release SSA_NAMEs to
2997 the manager.
2999 That's all fine and good, except that once SSA_NAMEs are released
3000 to the manager, we must not call create_ssa_name until all references
3001 to released SSA_NAMEs have been eliminated.
3003 All references to the deleted SSA_NAMEs can not be eliminated until
3004 we remove unreachable blocks.
3006 We can not remove unreachable blocks until after we have completed
3007 any queued jump threading.
3009 We can not complete any queued jump threads until we have taken
3010 appropriate variables out of SSA form. Taking variables out of
3011 SSA form can call create_ssa_name and thus we lose.
3013 Ultimately I suspect we're going to need to change the interface
3014 into the SSA_NAME manager. */
3016 if (ann->modified)
3018 tree val = NULL;
3020 if (TREE_CODE (stmt) == COND_EXPR)
3021 val = COND_EXPR_COND (stmt);
3022 else if (TREE_CODE (stmt) == SWITCH_EXPR)
3023 val = SWITCH_COND (stmt);
3025 if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val))
3026 cfg_altered = true;
3028 /* If we simplified a statement in such a way as to be shown that it
3029 cannot trap, update the eh information and the cfg to match. */
3030 if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt))
3032 bitmap_set_bit (need_eh_cleanup, bb->index);
3033 if (dump_file && (dump_flags & TDF_DETAILS))
3034 fprintf (dump_file, " Flagged to clear EH edges.\n");
3038 if (may_have_exposed_new_symbols)
3039 VEC_safe_push (tree, heap, stmts_to_rescan, bsi_stmt (si));
3042 /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the
3043 available expression hashtable, then return the LHS from the hash
3044 table.
3046 If INSERT is true, then we also update the available expression
3047 hash table to account for the changes made to STMT. */
3049 static tree
3050 update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert)
3052 tree cached_lhs = NULL;
3054 /* Remove the old entry from the hash table. */
3055 if (insert)
3057 struct expr_hash_elt element;
3059 initialize_hash_element (stmt, NULL, &element);
3060 htab_remove_elt_with_hash (avail_exprs, &element, element.hash);
3063 /* Now update the RHS of the assignment. */
3064 TREE_OPERAND (stmt, 1) = new_rhs;
3066 /* Now lookup the updated statement in the hash table. */
3067 cached_lhs = lookup_avail_expr (stmt, insert);
3069 /* We have now called lookup_avail_expr twice with two different
3070 versions of this same statement, once in optimize_stmt, once here.
3072 We know the call in optimize_stmt did not find an existing entry
3073 in the hash table, so a new entry was created. At the same time
3074 this statement was pushed onto the AVAIL_EXPRS_STACK vector.
3076 If this call failed to find an existing entry on the hash table,
3077 then the new version of this statement was entered into the
3078 hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR
3079 for the second time. So there are two copies on BLOCK_AVAIL_EXPRs
3081 If this call succeeded, we still have one copy of this statement
3082 on the BLOCK_AVAIL_EXPRs vector.
3084 For both cases, we need to pop the most recent entry off the
3085 BLOCK_AVAIL_EXPRs vector. For the case where we never found this
3086 statement in the hash tables, that will leave precisely one
3087 copy of this statement on BLOCK_AVAIL_EXPRs. For the case where
3088 we found a copy of this statement in the second hash table lookup
3089 we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */
3090 if (insert)
3091 VEC_pop (tree, avail_exprs_stack);
3093 /* And make sure we record the fact that we modified this
3094 statement. */
3095 mark_stmt_modified (stmt);
3097 return cached_lhs;
3100 /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If
3101 found, return its LHS. Otherwise insert STMT in the table and return
3102 NULL_TREE.
3104 Also, when an expression is first inserted in the AVAIL_EXPRS table, it
3105 is also added to the stack pointed to by BLOCK_AVAIL_EXPRS_P, so that they
3106 can be removed when we finish processing this block and its children.
3108 NOTE: This function assumes that STMT is a MODIFY_EXPR node that
3109 contains no CALL_EXPR on its RHS and makes no volatile nor
3110 aliased references. */
3112 static tree
3113 lookup_avail_expr (tree stmt, bool insert)
3115 void **slot;
3116 tree lhs;
3117 tree temp;
3118 struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt));
3120 lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL;
3122 initialize_hash_element (stmt, lhs, element);
3124 /* Don't bother remembering constant assignments and copy operations.
3125 Constants and copy operations are handled by the constant/copy propagator
3126 in optimize_stmt. */
3127 if (TREE_CODE (element->rhs) == SSA_NAME
3128 || is_gimple_min_invariant (element->rhs))
3130 free (element);
3131 return NULL_TREE;
3134 /* If this is an equality test against zero, see if we have recorded a
3135 nonzero value for the variable in question. */
3136 if ((TREE_CODE (element->rhs) == EQ_EXPR
3137 || TREE_CODE (element->rhs) == NE_EXPR)
3138 && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME
3139 && integer_zerop (TREE_OPERAND (element->rhs, 1)))
3141 int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0));
3143 if (bitmap_bit_p (nonzero_vars, indx))
3145 tree t = element->rhs;
3146 free (element);
3147 return constant_boolean_node (TREE_CODE (t) != EQ_EXPR,
3148 TREE_TYPE (t));
3152 /* Finally try to find the expression in the main expression hash table. */
3153 slot = htab_find_slot_with_hash (avail_exprs, element, element->hash,
3154 (insert ? INSERT : NO_INSERT));
3155 if (slot == NULL)
3157 free (element);
3158 return NULL_TREE;
3161 if (*slot == NULL)
3163 *slot = (void *) element;
3164 VEC_safe_push (tree, heap, avail_exprs_stack,
3165 stmt ? stmt : element->rhs);
3166 return NULL_TREE;
3169 /* Extract the LHS of the assignment so that it can be used as the current
3170 definition of another variable. */
3171 lhs = ((struct expr_hash_elt *)*slot)->lhs;
3173 /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then
3174 use the value from the const_and_copies table. */
3175 if (TREE_CODE (lhs) == SSA_NAME)
3177 temp = SSA_NAME_VALUE (lhs);
3178 if (temp && TREE_CODE (temp) != VALUE_HANDLE)
3179 lhs = temp;
3182 free (element);
3183 return lhs;
3186 /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the
3187 range of values that result in the conditional having a true value.
3189 Return true if we are successful in extracting a range from COND and
3190 false if we are unsuccessful. */
3192 static bool
3193 extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p)
3195 tree op1 = TREE_OPERAND (cond, 1);
3196 tree high, low, type;
3197 int inverted;
3199 type = TREE_TYPE (op1);
3201 /* Experiments have shown that it's rarely, if ever useful to
3202 record ranges for enumerations. Presumably this is due to
3203 the fact that they're rarely used directly. They are typically
3204 cast into an integer type and used that way. */
3205 if (TREE_CODE (type) != INTEGER_TYPE
3206 /* We don't know how to deal with types with variable bounds. */
3207 || TREE_CODE (TYPE_MIN_VALUE (type)) != INTEGER_CST
3208 || TREE_CODE (TYPE_MAX_VALUE (type)) != INTEGER_CST)
3209 return 0;
3211 switch (TREE_CODE (cond))
3213 case EQ_EXPR:
3214 high = low = op1;
3215 inverted = 0;
3216 break;
3218 case NE_EXPR:
3219 high = low = op1;
3220 inverted = 1;
3221 break;
3223 case GE_EXPR:
3224 low = op1;
3225 high = TYPE_MAX_VALUE (type);
3226 inverted = 0;
3227 break;
3229 case GT_EXPR:
3230 high = TYPE_MAX_VALUE (type);
3231 if (!tree_int_cst_lt (op1, high))
3232 return 0;
3233 low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1);
3234 inverted = 0;
3235 break;
3237 case LE_EXPR:
3238 high = op1;
3239 low = TYPE_MIN_VALUE (type);
3240 inverted = 0;
3241 break;
3243 case LT_EXPR:
3244 low = TYPE_MIN_VALUE (type);
3245 if (!tree_int_cst_lt (low, op1))
3246 return 0;
3247 high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1);
3248 inverted = 0;
3249 break;
3251 default:
3252 return 0;
3255 *hi_p = high;
3256 *lo_p = low;
3257 *inverted_p = inverted;
3258 return 1;
3261 /* Record a range created by COND for basic block BB. */
3263 static void
3264 record_range (tree cond, basic_block bb)
3266 enum tree_code code = TREE_CODE (cond);
3268 /* We explicitly ignore NE_EXPRs and all the unordered comparisons.
3269 They rarely allow for meaningful range optimizations and significantly
3270 complicate the implementation. */
3271 if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR
3272 || code == GE_EXPR || code == EQ_EXPR)
3273 && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE)
3275 struct vrp_hash_elt *vrp_hash_elt;
3276 struct vrp_element *element;
3277 VEC(vrp_element_p,heap) **vrp_records_p;
3278 void **slot;
3281 vrp_hash_elt = xmalloc (sizeof (struct vrp_hash_elt));
3282 vrp_hash_elt->var = TREE_OPERAND (cond, 0);
3283 vrp_hash_elt->records = NULL;
3284 slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT);
3286 if (*slot == NULL)
3287 *slot = (void *) vrp_hash_elt;
3288 else
3289 vrp_free (vrp_hash_elt);
3291 vrp_hash_elt = (struct vrp_hash_elt *) *slot;
3292 vrp_records_p = &vrp_hash_elt->records;
3294 element = ggc_alloc (sizeof (struct vrp_element));
3295 element->low = NULL;
3296 element->high = NULL;
3297 element->cond = cond;
3298 element->bb = bb;
3300 VEC_safe_push (vrp_element_p, heap, *vrp_records_p, element);
3301 VEC_safe_push (tree, heap, vrp_variables_stack, TREE_OPERAND (cond, 0));
3305 /* Hashing and equality functions for VRP_DATA.
3307 Since this hash table is addressed by SSA_NAMEs, we can hash on
3308 their version number and equality can be determined with a
3309 pointer comparison. */
3311 static hashval_t
3312 vrp_hash (const void *p)
3314 tree var = ((struct vrp_hash_elt *)p)->var;
3316 return SSA_NAME_VERSION (var);
3319 static int
3320 vrp_eq (const void *p1, const void *p2)
3322 tree var1 = ((struct vrp_hash_elt *)p1)->var;
3323 tree var2 = ((struct vrp_hash_elt *)p2)->var;
3325 return var1 == var2;
3328 /* Hashing and equality functions for AVAIL_EXPRS. The table stores
3329 MODIFY_EXPR statements. We compute a value number for expressions using
3330 the code of the expression and the SSA numbers of its operands. */
3332 static hashval_t
3333 avail_expr_hash (const void *p)
3335 tree stmt = ((struct expr_hash_elt *)p)->stmt;
3336 tree rhs = ((struct expr_hash_elt *)p)->rhs;
3337 tree vuse;
3338 ssa_op_iter iter;
3339 hashval_t val = 0;
3341 /* iterative_hash_expr knows how to deal with any expression and
3342 deals with commutative operators as well, so just use it instead
3343 of duplicating such complexities here. */
3344 val = iterative_hash_expr (rhs, val);
3346 /* If the hash table entry is not associated with a statement, then we
3347 can just hash the expression and not worry about virtual operands
3348 and such. */
3349 if (!stmt || !stmt_ann (stmt))
3350 return val;
3352 /* Add the SSA version numbers of every vuse operand. This is important
3353 because compound variables like arrays are not renamed in the
3354 operands. Rather, the rename is done on the virtual variable
3355 representing all the elements of the array. */
3356 FOR_EACH_SSA_TREE_OPERAND (vuse, stmt, iter, SSA_OP_VUSE)
3357 val = iterative_hash_expr (vuse, val);
3359 return val;
3362 static hashval_t
3363 real_avail_expr_hash (const void *p)
3365 return ((const struct expr_hash_elt *)p)->hash;
3368 static int
3369 avail_expr_eq (const void *p1, const void *p2)
3371 tree stmt1 = ((struct expr_hash_elt *)p1)->stmt;
3372 tree rhs1 = ((struct expr_hash_elt *)p1)->rhs;
3373 tree stmt2 = ((struct expr_hash_elt *)p2)->stmt;
3374 tree rhs2 = ((struct expr_hash_elt *)p2)->rhs;
3376 /* If they are the same physical expression, return true. */
3377 if (rhs1 == rhs2 && stmt1 == stmt2)
3378 return true;
3380 /* If their codes are not equal, then quit now. */
3381 if (TREE_CODE (rhs1) != TREE_CODE (rhs2))
3382 return false;
3384 /* In case of a collision, both RHS have to be identical and have the
3385 same VUSE operands. */
3386 if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2)
3387 || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2)))
3388 && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME))
3390 bool ret = compare_ssa_operands_equal (stmt1, stmt2, SSA_OP_VUSE);
3391 gcc_assert (!ret || ((struct expr_hash_elt *)p1)->hash
3392 == ((struct expr_hash_elt *)p2)->hash);
3393 return ret;
3396 return false;