]>
Commit | Line | Data |
---|---|---|
6de9cd9a DN |
1 | /* SSA Dominator optimizations for trees |
2 | Copyright (C) 2001, 2002, 2003, 2004 Free Software Foundation, Inc. | |
3 | Contributed by Diego Novillo <dnovillo@redhat.com> | |
4 | ||
5 | This file is part of GCC. | |
6 | ||
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. | |
11 | ||
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. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING. If not, write to | |
19 | the Free Software Foundation, 59 Temple Place - Suite 330, | |
20 | Boston, MA 02111-1307, USA. */ | |
21 | ||
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 "output.h" | |
33 | #include "errors.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" | |
c7f90219 | 43 | #include "tree-ssa-propagate.h" |
6de9cd9a DN |
44 | #include "langhooks.h" |
45 | ||
46 | /* This file implements optimizations on the dominator tree. */ | |
47 | ||
48 | /* Hash table with expressions made available during the renaming process. | |
49 | When an assignment of the form X_i = EXPR is found, the statement is | |
50 | stored in this table. If the same expression EXPR is later found on the | |
51 | RHS of another statement, it is replaced with X_i (thus performing | |
52 | global redundancy elimination). Similarly as we pass through conditionals | |
53 | we record the conditional itself as having either a true or false value | |
54 | in this table. */ | |
55 | static htab_t avail_exprs; | |
56 | ||
48732f23 JL |
57 | /* Stack of available expressions in AVAIL_EXPRs. Each block pushes any |
58 | expressions it enters into the hash table along with a marker entry | |
b3a27618 | 59 | (null). When we finish processing the block, we pop off entries and |
48732f23 JL |
60 | remove the expressions from the global hash table until we hit the |
61 | marker. */ | |
62 | static varray_type avail_exprs_stack; | |
63 | ||
9fae925b JL |
64 | /* Stack of trees used to restore the global currdefs to its original |
65 | state after completing optimization of a block and its dominator children. | |
66 | ||
67 | An SSA_NAME indicates that the current definition of the underlying | |
68 | variable should be set to the given SSA_NAME. | |
69 | ||
70 | A _DECL node indicates that the underlying variable has no current | |
71 | definition. | |
72 | ||
73 | A NULL node is used to mark the last node associated with the | |
74 | current block. */ | |
75 | varray_type block_defs_stack; | |
76 | ||
a6e1aa26 JL |
77 | /* Stack of statements we need to rescan during finalization for newly |
78 | exposed variables. | |
79 | ||
80 | Statement rescanning must occur after the current block's available | |
81 | expressions are removed from AVAIL_EXPRS. Else we may change the | |
82 | hash code for an expression and be unable to find/remove it from | |
83 | AVAIL_EXPRS. */ | |
84 | varray_type stmts_to_rescan; | |
85 | ||
6de9cd9a DN |
86 | /* Structure for entries in the expression hash table. |
87 | ||
88 | This requires more memory for the hash table entries, but allows us | |
89 | to avoid creating silly tree nodes and annotations for conditionals, | |
90 | eliminates 2 global hash tables and two block local varrays. | |
91 | ||
92 | It also allows us to reduce the number of hash table lookups we | |
93 | have to perform in lookup_avail_expr and finally it allows us to | |
94 | significantly reduce the number of calls into the hashing routine | |
95 | itself. */ | |
56b043c8 | 96 | |
6de9cd9a DN |
97 | struct expr_hash_elt |
98 | { | |
99 | /* The value (lhs) of this expression. */ | |
100 | tree lhs; | |
101 | ||
102 | /* The expression (rhs) we want to record. */ | |
103 | tree rhs; | |
104 | ||
105 | /* The annotation if this element corresponds to a statement. */ | |
106 | stmt_ann_t ann; | |
107 | ||
108 | /* The hash value for RHS/ann. */ | |
109 | hashval_t hash; | |
110 | }; | |
111 | ||
112 | /* Table of constant values and copies indexed by SSA name. When the | |
113 | renaming pass finds an assignment of a constant (X_i = C) or a copy | |
114 | assignment from another SSA variable (X_i = Y_j), it creates a mapping | |
115 | between X_i and the RHS in this table. This mapping is used later on, | |
116 | when renaming uses of X_i. If an assignment to X_i is found in this | |
117 | table, instead of using X_i, we use the RHS of the statement stored in | |
118 | this table (thus performing very simplistic copy and constant | |
119 | propagation). */ | |
120 | static varray_type const_and_copies; | |
121 | ||
b5fefcf6 JL |
122 | /* Stack of dest,src pairs that need to be restored during finalization. |
123 | ||
124 | A NULL entry is used to mark the end of pairs which need to be | |
125 | restored during finalization of this block. */ | |
126 | static varray_type const_and_copies_stack; | |
127 | ||
6de9cd9a DN |
128 | /* Bitmap of SSA_NAMEs known to have a nonzero value, even if we do not |
129 | know their exact value. */ | |
130 | static bitmap nonzero_vars; | |
131 | ||
132 | /* Track whether or not we have changed the control flow graph. */ | |
133 | static bool cfg_altered; | |
134 | ||
1eaba2f2 | 135 | /* Bitmap of blocks that have had EH statements cleaned. We should |
f6fe65dc | 136 | remove their dead edges eventually. */ |
1eaba2f2 RH |
137 | static bitmap need_eh_cleanup; |
138 | ||
6de9cd9a DN |
139 | /* Statistics for dominator optimizations. */ |
140 | struct opt_stats_d | |
141 | { | |
142 | long num_stmts; | |
143 | long num_exprs_considered; | |
144 | long num_re; | |
145 | }; | |
146 | ||
147 | /* Value range propagation record. Each time we encounter a conditional | |
148 | of the form SSA_NAME COND CONST we create a new vrp_element to record | |
149 | how the condition affects the possible values SSA_NAME may have. | |
150 | ||
151 | Each record contains the condition tested (COND), and the the range of | |
152 | values the variable may legitimately have if COND is true. Note the | |
153 | range of values may be a smaller range than COND specifies if we have | |
154 | recorded other ranges for this variable. Each record also contains the | |
155 | block in which the range was recorded for invalidation purposes. | |
156 | ||
157 | Note that the current known range is computed lazily. This allows us | |
158 | to avoid the overhead of computing ranges which are never queried. | |
159 | ||
160 | When we encounter a conditional, we look for records which constrain | |
161 | the SSA_NAME used in the condition. In some cases those records allow | |
162 | us to determine the condition's result at compile time. In other cases | |
163 | they may allow us to simplify the condition. | |
164 | ||
165 | We also use value ranges to do things like transform signed div/mod | |
166 | operations into unsigned div/mod or to simplify ABS_EXPRs. | |
167 | ||
168 | Simple experiments have shown these optimizations to not be all that | |
169 | useful on switch statements (much to my surprise). So switch statement | |
170 | optimizations are not performed. | |
171 | ||
172 | Note carefully we do not propagate information through each statement | |
454ff5cb | 173 | in the block. i.e., if we know variable X has a value defined of |
6de9cd9a DN |
174 | [0, 25] and we encounter Y = X + 1, we do not track a value range |
175 | for Y (which would be [1, 26] if we cared). Similarly we do not | |
176 | constrain values as we encounter narrowing typecasts, etc. */ | |
177 | ||
178 | struct vrp_element | |
179 | { | |
180 | /* The highest and lowest values the variable in COND may contain when | |
181 | COND is true. Note this may not necessarily be the same values | |
182 | tested by COND if the same variable was used in earlier conditionals. | |
183 | ||
184 | Note this is computed lazily and thus can be NULL indicating that | |
185 | the values have not been computed yet. */ | |
186 | tree low; | |
187 | tree high; | |
188 | ||
189 | /* The actual conditional we recorded. This is needed since we compute | |
190 | ranges lazily. */ | |
191 | tree cond; | |
192 | ||
193 | /* The basic block where this record was created. We use this to determine | |
194 | when to remove records. */ | |
195 | basic_block bb; | |
196 | }; | |
197 | ||
198 | static struct opt_stats_d opt_stats; | |
199 | ||
6de9cd9a DN |
200 | /* A virtual array holding value range records for the variable identified |
201 | by the index, SSA_VERSION. */ | |
202 | static varray_type vrp_data; | |
203 | ||
204 | /* Datastructure for block local data used during the dominator walk. | |
205 | We maintain a stack of these as we recursively walk down the | |
206 | dominator tree. */ | |
207 | ||
208 | struct dom_walk_block_data | |
209 | { | |
6de9cd9a DN |
210 | /* Similarly for the nonzero state of variables that needs to be |
211 | restored during finalization. */ | |
212 | varray_type nonzero_vars; | |
213 | ||
6de9cd9a DN |
214 | /* Array of variables which have their values constrained by operations |
215 | in this basic block. We use this during finalization to know | |
216 | which variables need their VRP data updated. */ | |
217 | varray_type vrp_variables; | |
6de9cd9a DN |
218 | }; |
219 | ||
220 | struct eq_expr_value | |
221 | { | |
222 | tree src; | |
223 | tree dst; | |
224 | }; | |
225 | ||
226 | /* Local functions. */ | |
227 | static void optimize_stmt (struct dom_walk_data *, | |
228 | basic_block bb, | |
229 | block_stmt_iterator); | |
230 | static inline tree get_value_for (tree, varray_type table); | |
231 | static inline void set_value_for (tree, tree, varray_type table); | |
48732f23 JL |
232 | static tree lookup_avail_expr (tree, bool); |
233 | static struct eq_expr_value get_eq_expr_value (tree, int, | |
6de9cd9a DN |
234 | basic_block, varray_type *); |
235 | static hashval_t avail_expr_hash (const void *); | |
940db2c8 | 236 | static hashval_t real_avail_expr_hash (const void *); |
6de9cd9a DN |
237 | static int avail_expr_eq (const void *, const void *); |
238 | static void htab_statistics (FILE *, htab_t); | |
48732f23 JL |
239 | static void record_cond (tree, tree); |
240 | static void record_dominating_conditions (tree); | |
b5fefcf6 JL |
241 | static void record_const_or_copy (tree, tree); |
242 | static void record_equality (tree, tree); | |
48732f23 | 243 | static tree update_rhs_and_lookup_avail_expr (tree, tree, bool); |
6de9cd9a | 244 | static tree simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *, |
68b9f53b | 245 | tree, int); |
48732f23 JL |
246 | static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int); |
247 | static tree simplify_switch_and_lookup_avail_expr (tree, int); | |
6de9cd9a DN |
248 | static tree find_equivalent_equality_comparison (tree); |
249 | static void record_range (tree, basic_block, varray_type *); | |
250 | static bool extract_range_from_cond (tree, tree *, tree *, int *); | |
251 | static void record_equivalences_from_phis (struct dom_walk_data *, basic_block); | |
252 | static void record_equivalences_from_incoming_edge (struct dom_walk_data *, | |
253 | basic_block); | |
254 | static bool eliminate_redundant_computations (struct dom_walk_data *, | |
255 | tree, stmt_ann_t); | |
48732f23 | 256 | static void record_equivalences_from_stmt (tree, varray_type *, |
6de9cd9a DN |
257 | int, stmt_ann_t); |
258 | static void thread_across_edge (struct dom_walk_data *, edge); | |
259 | static void dom_opt_finalize_block (struct dom_walk_data *, basic_block); | |
260 | static void dom_opt_initialize_block_local_data (struct dom_walk_data *, | |
261 | basic_block, bool); | |
262 | static void dom_opt_initialize_block (struct dom_walk_data *, basic_block); | |
263 | static void cprop_into_phis (struct dom_walk_data *, basic_block); | |
48732f23 | 264 | static void remove_local_expressions_from_table (void); |
b5fefcf6 | 265 | static void restore_vars_to_original_value (void); |
9fae925b JL |
266 | static void restore_currdefs_to_original_value (void); |
267 | static void register_definitions_for_stmt (tree); | |
28c008bb | 268 | static edge single_incoming_edge_ignoring_loop_edges (basic_block); |
6de9cd9a DN |
269 | |
270 | /* Local version of fold that doesn't introduce cruft. */ | |
271 | ||
272 | static tree | |
273 | local_fold (tree t) | |
274 | { | |
275 | t = fold (t); | |
276 | ||
277 | /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that | |
278 | may have been added by fold, and "useless" type conversions that might | |
279 | now be apparent due to propagation. */ | |
6de9cd9a DN |
280 | STRIP_USELESS_TYPE_CONVERSION (t); |
281 | ||
282 | return t; | |
283 | } | |
284 | ||
285 | /* Return the value associated with variable VAR in TABLE. */ | |
286 | ||
287 | static inline tree | |
288 | get_value_for (tree var, varray_type table) | |
289 | { | |
290 | return VARRAY_TREE (table, SSA_NAME_VERSION (var)); | |
291 | } | |
292 | ||
293 | /* Associate VALUE to variable VAR in TABLE. */ | |
294 | ||
295 | static inline void | |
296 | set_value_for (tree var, tree value, varray_type table) | |
297 | { | |
298 | VARRAY_TREE (table, SSA_NAME_VERSION (var)) = value; | |
299 | } | |
300 | ||
6de9cd9a DN |
301 | /* Jump threading, redundancy elimination and const/copy propagation. |
302 | ||
6de9cd9a DN |
303 | This pass may expose new symbols that need to be renamed into SSA. For |
304 | every new symbol exposed, its corresponding bit will be set in | |
ff2ad0f7 | 305 | VARS_TO_RENAME. */ |
6de9cd9a DN |
306 | |
307 | static void | |
308 | tree_ssa_dominator_optimize (void) | |
309 | { | |
6de9cd9a DN |
310 | struct dom_walk_data walk_data; |
311 | unsigned int i; | |
312 | ||
313 | for (i = 0; i < num_referenced_vars; i++) | |
314 | var_ann (referenced_var (i))->current_def = NULL; | |
315 | ||
316 | /* Mark loop edges so we avoid threading across loop boundaries. | |
317 | This may result in transforming natural loop into irreducible | |
318 | region. */ | |
319 | mark_dfs_back_edges (); | |
320 | ||
321 | /* Create our hash tables. */ | |
940db2c8 | 322 | avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free); |
48732f23 | 323 | VARRAY_TREE_INIT (avail_exprs_stack, 20, "Available expression stack"); |
9fae925b | 324 | VARRAY_TREE_INIT (block_defs_stack, 20, "Block DEFS stack"); |
95a3742c | 325 | VARRAY_TREE_INIT (const_and_copies, num_ssa_names, "const_and_copies"); |
b5fefcf6 | 326 | VARRAY_TREE_INIT (const_and_copies_stack, 20, "Block const_and_copies stack"); |
6de9cd9a | 327 | nonzero_vars = BITMAP_XMALLOC (); |
95a3742c | 328 | VARRAY_GENERIC_PTR_INIT (vrp_data, num_ssa_names, "vrp_data"); |
1eaba2f2 | 329 | need_eh_cleanup = BITMAP_XMALLOC (); |
a6e1aa26 | 330 | VARRAY_TREE_INIT (stmts_to_rescan, 20, "Statements to rescan"); |
6de9cd9a DN |
331 | |
332 | /* Setup callbacks for the generic dominator tree walker. */ | |
333 | walk_data.walk_stmts_backward = false; | |
334 | walk_data.dom_direction = CDI_DOMINATORS; | |
335 | walk_data.initialize_block_local_data = dom_opt_initialize_block_local_data; | |
336 | walk_data.before_dom_children_before_stmts = dom_opt_initialize_block; | |
337 | walk_data.before_dom_children_walk_stmts = optimize_stmt; | |
338 | walk_data.before_dom_children_after_stmts = cprop_into_phis; | |
339 | walk_data.after_dom_children_before_stmts = NULL; | |
340 | walk_data.after_dom_children_walk_stmts = NULL; | |
341 | walk_data.after_dom_children_after_stmts = dom_opt_finalize_block; | |
342 | /* Right now we only attach a dummy COND_EXPR to the global data pointer. | |
343 | When we attach more stuff we'll need to fill this out with a real | |
344 | structure. */ | |
345 | walk_data.global_data = NULL; | |
346 | walk_data.block_local_data_size = sizeof (struct dom_walk_block_data); | |
347 | ||
348 | /* Now initialize the dominator walker. */ | |
349 | init_walk_dominator_tree (&walk_data); | |
350 | ||
6de9cd9a DN |
351 | calculate_dominance_info (CDI_DOMINATORS); |
352 | ||
353 | /* If we prove certain blocks are unreachable, then we want to | |
354 | repeat the dominator optimization process as PHI nodes may | |
355 | have turned into copies which allows better propagation of | |
356 | values. So we repeat until we do not identify any new unreachable | |
357 | blocks. */ | |
358 | do | |
359 | { | |
360 | /* Optimize the dominator tree. */ | |
361 | cfg_altered = false; | |
362 | ||
363 | /* Recursively walk the dominator tree optimizing statements. */ | |
364 | walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR); | |
365 | ||
56b043c8 JL |
366 | /* If we exposed any new variables, go ahead and put them into |
367 | SSA form now, before we handle jump threading. This simplifies | |
368 | interactions between rewriting of _DECL nodes into SSA form | |
369 | and rewriting SSA_NAME nodes into SSA form after block | |
370 | duplication and CFG manipulation. */ | |
371 | if (bitmap_first_set_bit (vars_to_rename) >= 0) | |
372 | { | |
373 | rewrite_into_ssa (false); | |
374 | bitmap_clear (vars_to_rename); | |
375 | } | |
6de9cd9a | 376 | |
56b043c8 JL |
377 | /* Thread jumps, creating duplicate blocks as needed. */ |
378 | cfg_altered = thread_through_all_blocks (); | |
6de9cd9a | 379 | |
56b043c8 JL |
380 | /* Removal of statements may make some EH edges dead. Purge |
381 | such edges from the CFG as needed. */ | |
1eaba2f2 RH |
382 | if (bitmap_first_set_bit (need_eh_cleanup) >= 0) |
383 | { | |
56b043c8 | 384 | cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup); |
1eaba2f2 RH |
385 | bitmap_zero (need_eh_cleanup); |
386 | } | |
387 | ||
56b043c8 JL |
388 | free_dominance_info (CDI_DOMINATORS); |
389 | cfg_altered = cleanup_tree_cfg (); | |
390 | calculate_dominance_info (CDI_DOMINATORS); | |
6de9cd9a | 391 | |
56b043c8 | 392 | rewrite_ssa_into_ssa (); |
6de9cd9a | 393 | |
56b043c8 | 394 | if (VARRAY_ACTIVE_SIZE (const_and_copies) <= num_ssa_names) |
6de9cd9a | 395 | { |
95a3742c DN |
396 | VARRAY_GROW (const_and_copies, num_ssa_names); |
397 | VARRAY_GROW (vrp_data, num_ssa_names); | |
6de9cd9a DN |
398 | } |
399 | ||
400 | /* Reinitialize the various tables. */ | |
401 | bitmap_clear (nonzero_vars); | |
402 | htab_empty (avail_exprs); | |
403 | VARRAY_CLEAR (const_and_copies); | |
404 | VARRAY_CLEAR (vrp_data); | |
405 | ||
406 | for (i = 0; i < num_referenced_vars; i++) | |
407 | var_ann (referenced_var (i))->current_def = NULL; | |
408 | } | |
409 | while (cfg_altered); | |
410 | ||
6de9cd9a DN |
411 | /* Debugging dumps. */ |
412 | if (dump_file && (dump_flags & TDF_STATS)) | |
413 | dump_dominator_optimization_stats (dump_file); | |
414 | ||
61ada8ae | 415 | /* We emptied the hash table earlier, now delete it completely. */ |
6de9cd9a DN |
416 | htab_delete (avail_exprs); |
417 | ||
1ea7e6ad | 418 | /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA, |
6de9cd9a DN |
419 | CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom |
420 | of the do-while loop above. */ | |
421 | ||
422 | /* And finalize the dominator walker. */ | |
423 | fini_walk_dominator_tree (&walk_data); | |
cfa4cb00 AP |
424 | |
425 | /* Free nonzero_vars. */ | |
426 | BITMAP_XFREE (nonzero_vars); | |
1eaba2f2 | 427 | BITMAP_XFREE (need_eh_cleanup); |
6de9cd9a DN |
428 | } |
429 | ||
430 | static bool | |
431 | gate_dominator (void) | |
432 | { | |
433 | return flag_tree_dom != 0; | |
434 | } | |
435 | ||
436 | struct tree_opt_pass pass_dominator = | |
437 | { | |
438 | "dom", /* name */ | |
439 | gate_dominator, /* gate */ | |
440 | tree_ssa_dominator_optimize, /* execute */ | |
441 | NULL, /* sub */ | |
442 | NULL, /* next */ | |
443 | 0, /* static_pass_number */ | |
444 | TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */ | |
c1b763fa | 445 | PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */ |
6de9cd9a DN |
446 | 0, /* properties_provided */ |
447 | 0, /* properties_destroyed */ | |
448 | 0, /* todo_flags_start */ | |
449 | TODO_dump_func | TODO_rename_vars | |
9f8628ba PB |
450 | | TODO_verify_ssa, /* todo_flags_finish */ |
451 | 0 /* letter */ | |
6de9cd9a DN |
452 | }; |
453 | ||
454 | ||
455 | /* We are exiting BB, see if the target block begins with a conditional | |
456 | jump which has a known value when reached via BB. */ | |
457 | ||
458 | static void | |
459 | thread_across_edge (struct dom_walk_data *walk_data, edge e) | |
460 | { | |
6de9cd9a DN |
461 | block_stmt_iterator bsi; |
462 | tree stmt = NULL; | |
463 | tree phi; | |
464 | ||
465 | /* Each PHI creates a temporary equivalence, record them. */ | |
17192884 | 466 | for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi)) |
6de9cd9a | 467 | { |
d00ad49b | 468 | tree src = PHI_ARG_DEF_FROM_EDGE (phi, e); |
6de9cd9a | 469 | tree dst = PHI_RESULT (phi); |
b5fefcf6 | 470 | record_const_or_copy (dst, src); |
9fae925b | 471 | register_new_def (dst, &block_defs_stack); |
6de9cd9a DN |
472 | } |
473 | ||
474 | for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi)) | |
475 | { | |
476 | tree lhs, cached_lhs; | |
477 | ||
478 | stmt = bsi_stmt (bsi); | |
479 | ||
480 | /* Ignore empty statements and labels. */ | |
481 | if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR) | |
482 | continue; | |
483 | ||
484 | /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new | |
485 | value, then stop our search here. Ideally when we stop a | |
486 | search we stop on a COND_EXPR or SWITCH_EXPR. */ | |
487 | if (TREE_CODE (stmt) != MODIFY_EXPR | |
488 | || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME) | |
489 | break; | |
490 | ||
491 | /* At this point we have a statement which assigns an RHS to an | |
492 | SSA_VAR on the LHS. We want to prove that the RHS is already | |
493 | available and that its value is held in the current definition | |
494 | of the LHS -- meaning that this assignment is a NOP when | |
495 | reached via edge E. */ | |
496 | if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME) | |
497 | cached_lhs = TREE_OPERAND (stmt, 1); | |
498 | else | |
48732f23 | 499 | cached_lhs = lookup_avail_expr (stmt, false); |
6de9cd9a DN |
500 | |
501 | lhs = TREE_OPERAND (stmt, 0); | |
502 | ||
503 | /* This can happen if we thread around to the start of a loop. */ | |
504 | if (lhs == cached_lhs) | |
505 | break; | |
506 | ||
507 | /* If we did not find RHS in the hash table, then try again after | |
508 | temporarily const/copy propagating the operands. */ | |
509 | if (!cached_lhs) | |
510 | { | |
511 | /* Copy the operands. */ | |
512 | stmt_ann_t ann = stmt_ann (stmt); | |
513 | use_optype uses = USE_OPS (ann); | |
514 | vuse_optype vuses = VUSE_OPS (ann); | |
515 | tree *uses_copy = xcalloc (NUM_USES (uses), sizeof (tree)); | |
516 | tree *vuses_copy = xcalloc (NUM_VUSES (vuses), sizeof (tree)); | |
517 | unsigned int i; | |
518 | ||
519 | /* Make a copy of the uses into USES_COPY, then cprop into | |
520 | the use operands. */ | |
521 | for (i = 0; i < NUM_USES (uses); i++) | |
522 | { | |
523 | tree tmp = NULL; | |
524 | ||
525 | uses_copy[i] = USE_OP (uses, i); | |
526 | if (TREE_CODE (USE_OP (uses, i)) == SSA_NAME) | |
527 | tmp = get_value_for (USE_OP (uses, i), const_and_copies); | |
528 | if (tmp) | |
d00ad49b | 529 | SET_USE_OP (uses, i, tmp); |
6de9cd9a DN |
530 | } |
531 | ||
532 | /* Similarly for virtual uses. */ | |
533 | for (i = 0; i < NUM_VUSES (vuses); i++) | |
534 | { | |
535 | tree tmp = NULL; | |
536 | ||
537 | vuses_copy[i] = VUSE_OP (vuses, i); | |
538 | if (TREE_CODE (VUSE_OP (vuses, i)) == SSA_NAME) | |
539 | tmp = get_value_for (VUSE_OP (vuses, i), const_and_copies); | |
540 | if (tmp) | |
d00ad49b | 541 | SET_VUSE_OP (vuses, i, tmp); |
6de9cd9a DN |
542 | } |
543 | ||
544 | /* Try to lookup the new expression. */ | |
48732f23 | 545 | cached_lhs = lookup_avail_expr (stmt, false); |
6de9cd9a DN |
546 | |
547 | /* Restore the statement's original uses/defs. */ | |
548 | for (i = 0; i < NUM_USES (uses); i++) | |
d00ad49b | 549 | SET_USE_OP (uses, i, uses_copy[i]); |
6de9cd9a DN |
550 | |
551 | for (i = 0; i < NUM_VUSES (vuses); i++) | |
d00ad49b | 552 | SET_VUSE_OP (vuses, i, vuses_copy[i]); |
6de9cd9a DN |
553 | |
554 | free (uses_copy); | |
555 | free (vuses_copy); | |
556 | ||
557 | /* If we still did not find the expression in the hash table, | |
558 | then we can not ignore this statement. */ | |
559 | if (! cached_lhs) | |
560 | break; | |
561 | } | |
562 | ||
563 | /* If the expression in the hash table was not assigned to an | |
564 | SSA_NAME, then we can not ignore this statement. */ | |
565 | if (TREE_CODE (cached_lhs) != SSA_NAME) | |
566 | break; | |
567 | ||
568 | /* If we have different underlying variables, then we can not | |
569 | ignore this statement. */ | |
570 | if (SSA_NAME_VAR (cached_lhs) != SSA_NAME_VAR (lhs)) | |
571 | break; | |
572 | ||
573 | /* If CACHED_LHS does not represent the current value of the undering | |
574 | variable in CACHED_LHS/LHS, then we can not ignore this statement. */ | |
575 | if (var_ann (SSA_NAME_VAR (lhs))->current_def != cached_lhs) | |
576 | break; | |
577 | ||
578 | /* If we got here, then we can ignore this statement and continue | |
579 | walking through the statements in the block looking for a threadable | |
580 | COND_EXPR. | |
581 | ||
582 | We want to record an equivalence lhs = cache_lhs so that if | |
583 | the result of this statement is used later we can copy propagate | |
584 | suitably. */ | |
b5fefcf6 | 585 | record_const_or_copy (lhs, cached_lhs); |
9fae925b | 586 | register_new_def (lhs, &block_defs_stack); |
6de9cd9a DN |
587 | } |
588 | ||
589 | /* If we stopped at a COND_EXPR or SWITCH_EXPR, then see if we know which | |
590 | arm will be taken. */ | |
591 | if (stmt | |
592 | && (TREE_CODE (stmt) == COND_EXPR | |
593 | || TREE_CODE (stmt) == SWITCH_EXPR)) | |
594 | { | |
595 | tree cond, cached_lhs; | |
596 | edge e1; | |
597 | ||
598 | /* Do not forward entry edges into the loop. In the case loop | |
599 | has multiple entry edges we may end up in constructing irreducible | |
600 | region. | |
601 | ??? We may consider forwarding the edges in the case all incoming | |
602 | edges forward to the same destination block. */ | |
603 | if (!e->flags & EDGE_DFS_BACK) | |
604 | { | |
605 | for (e1 = e->dest->pred; e; e = e->pred_next) | |
606 | if (e1->flags & EDGE_DFS_BACK) | |
607 | break; | |
608 | if (e1) | |
609 | return; | |
610 | } | |
611 | ||
612 | /* Now temporarily cprop the operands and try to find the resulting | |
613 | expression in the hash tables. */ | |
614 | if (TREE_CODE (stmt) == COND_EXPR) | |
615 | cond = COND_EXPR_COND (stmt); | |
616 | else | |
617 | cond = SWITCH_COND (stmt); | |
618 | ||
6615c446 | 619 | if (COMPARISON_CLASS_P (cond)) |
6de9cd9a DN |
620 | { |
621 | tree dummy_cond, op0, op1; | |
622 | enum tree_code cond_code; | |
623 | ||
624 | op0 = TREE_OPERAND (cond, 0); | |
625 | op1 = TREE_OPERAND (cond, 1); | |
626 | cond_code = TREE_CODE (cond); | |
627 | ||
628 | /* Get the current value of both operands. */ | |
629 | if (TREE_CODE (op0) == SSA_NAME) | |
630 | { | |
631 | tree tmp = get_value_for (op0, const_and_copies); | |
632 | if (tmp) | |
633 | op0 = tmp; | |
634 | } | |
635 | ||
636 | if (TREE_CODE (op1) == SSA_NAME) | |
637 | { | |
638 | tree tmp = get_value_for (op1, const_and_copies); | |
639 | if (tmp) | |
640 | op1 = tmp; | |
641 | } | |
642 | ||
643 | /* Stuff the operator and operands into our dummy conditional | |
644 | expression, creating the dummy conditional if necessary. */ | |
645 | dummy_cond = walk_data->global_data; | |
646 | if (! dummy_cond) | |
647 | { | |
648 | dummy_cond = build (cond_code, boolean_type_node, op0, op1); | |
649 | dummy_cond = build (COND_EXPR, void_type_node, | |
650 | dummy_cond, NULL, NULL); | |
651 | walk_data->global_data = dummy_cond; | |
652 | } | |
653 | else | |
654 | { | |
655 | TREE_SET_CODE (TREE_OPERAND (dummy_cond, 0), cond_code); | |
656 | TREE_OPERAND (TREE_OPERAND (dummy_cond, 0), 0) = op0; | |
657 | TREE_OPERAND (TREE_OPERAND (dummy_cond, 0), 1) = op1; | |
658 | } | |
659 | ||
660 | /* If the conditional folds to an invariant, then we are done, | |
661 | otherwise look it up in the hash tables. */ | |
662 | cached_lhs = local_fold (COND_EXPR_COND (dummy_cond)); | |
663 | if (! is_gimple_min_invariant (cached_lhs)) | |
48732f23 | 664 | cached_lhs = lookup_avail_expr (dummy_cond, false); |
6de9cd9a DN |
665 | if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs)) |
666 | { | |
6de9cd9a | 667 | cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond, |
68b9f53b | 668 | NULL, |
6de9cd9a DN |
669 | false); |
670 | } | |
671 | } | |
672 | /* We can have conditionals which just test the state of a | |
673 | variable rather than use a relational operator. These are | |
674 | simpler to handle. */ | |
675 | else if (TREE_CODE (cond) == SSA_NAME) | |
676 | { | |
677 | cached_lhs = cond; | |
678 | cached_lhs = get_value_for (cached_lhs, const_and_copies); | |
679 | if (cached_lhs && ! is_gimple_min_invariant (cached_lhs)) | |
680 | cached_lhs = 0; | |
681 | } | |
682 | else | |
48732f23 | 683 | cached_lhs = lookup_avail_expr (stmt, false); |
6de9cd9a DN |
684 | |
685 | if (cached_lhs) | |
686 | { | |
687 | edge taken_edge = find_taken_edge (e->dest, cached_lhs); | |
688 | basic_block dest = (taken_edge ? taken_edge->dest : NULL); | |
689 | ||
8a78744f | 690 | if (dest == e->dest) |
6de9cd9a DN |
691 | return; |
692 | ||
693 | /* If we have a known destination for the conditional, then | |
694 | we can perform this optimization, which saves at least one | |
695 | conditional jump each time it applies since we get to | |
56b043c8 | 696 | bypass the conditional at our original destination. */ |
6de9cd9a DN |
697 | if (dest) |
698 | { | |
56b043c8 JL |
699 | e->aux = taken_edge; |
700 | bb_ann (e->dest)->incoming_edge_threaded = true; | |
6de9cd9a DN |
701 | } |
702 | } | |
703 | } | |
704 | } | |
705 | ||
706 | ||
707 | /* Initialize the local stacks. | |
708 | ||
709 | AVAIL_EXPRS stores all the expressions made available in this block. | |
710 | ||
711 | CONST_AND_COPIES stores var/value pairs to restore at the end of this | |
712 | block. | |
713 | ||
714 | NONZERO_VARS stores the vars which have a nonzero value made in this | |
715 | block. | |
716 | ||
717 | STMTS_TO_RESCAN is a list of statements we will rescan for operands. | |
718 | ||
719 | VRP_VARIABLES is the list of variables which have had their values | |
720 | constrained by an operation in this block. | |
721 | ||
722 | These stacks are cleared in the finalization routine run for each | |
723 | block. */ | |
724 | ||
725 | static void | |
726 | dom_opt_initialize_block_local_data (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED, | |
727 | basic_block bb ATTRIBUTE_UNUSED, | |
728 | bool recycled ATTRIBUTE_UNUSED) | |
729 | { | |
6de9cd9a DN |
730 | struct dom_walk_block_data *bd |
731 | = (struct dom_walk_block_data *)VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack); | |
732 | ||
733 | /* We get cleared memory from the allocator, so if the memory is not | |
734 | cleared, then we are re-using a previously allocated entry. In | |
735 | that case, we can also re-use the underlying virtual arrays. Just | |
736 | make sure we clear them before using them! */ | |
737 | if (recycled) | |
738 | { | |
1e128c5f GB |
739 | gcc_assert (!bd->nonzero_vars |
740 | || VARRAY_ACTIVE_SIZE (bd->nonzero_vars) == 0); | |
1e128c5f GB |
741 | gcc_assert (!bd->vrp_variables |
742 | || VARRAY_ACTIVE_SIZE (bd->vrp_variables) == 0); | |
6de9cd9a | 743 | } |
6de9cd9a DN |
744 | } |
745 | ||
746 | /* Initialize local stacks for this optimizer and record equivalences | |
747 | upon entry to BB. Equivalences can come from the edge traversed to | |
748 | reach BB or they may come from PHI nodes at the start of BB. */ | |
749 | ||
750 | static void | |
751 | dom_opt_initialize_block (struct dom_walk_data *walk_data, basic_block bb) | |
752 | { | |
753 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
754 | fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index); | |
755 | ||
9fae925b JL |
756 | /* Push a marker on the stacks of local information so that we know how |
757 | far to unwind when we finalize this block. */ | |
48732f23 | 758 | VARRAY_PUSH_TREE (avail_exprs_stack, NULL_TREE); |
9fae925b | 759 | VARRAY_PUSH_TREE (block_defs_stack, NULL_TREE); |
b5fefcf6 | 760 | VARRAY_PUSH_TREE (const_and_copies_stack, NULL_TREE); |
48732f23 | 761 | |
6de9cd9a DN |
762 | record_equivalences_from_incoming_edge (walk_data, bb); |
763 | ||
764 | /* PHI nodes can create equivalences too. */ | |
765 | record_equivalences_from_phis (walk_data, bb); | |
766 | } | |
767 | ||
768 | /* Given an expression EXPR (a relational expression or a statement), | |
769 | initialize the hash table element pointed by by ELEMENT. */ | |
770 | ||
771 | static void | |
772 | initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element) | |
773 | { | |
774 | /* Hash table elements may be based on conditional expressions or statements. | |
775 | ||
776 | For the former case, we have no annotation and we want to hash the | |
777 | conditional expression. In the latter case we have an annotation and | |
778 | we want to record the expression the statement evaluates. */ | |
6615c446 | 779 | if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR) |
6de9cd9a DN |
780 | { |
781 | element->ann = NULL; | |
782 | element->rhs = expr; | |
783 | } | |
784 | else if (TREE_CODE (expr) == COND_EXPR) | |
785 | { | |
786 | element->ann = stmt_ann (expr); | |
787 | element->rhs = COND_EXPR_COND (expr); | |
788 | } | |
789 | else if (TREE_CODE (expr) == SWITCH_EXPR) | |
790 | { | |
791 | element->ann = stmt_ann (expr); | |
792 | element->rhs = SWITCH_COND (expr); | |
793 | } | |
794 | else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0)) | |
795 | { | |
796 | element->ann = stmt_ann (expr); | |
797 | element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1); | |
798 | } | |
799 | else | |
800 | { | |
801 | element->ann = stmt_ann (expr); | |
802 | element->rhs = TREE_OPERAND (expr, 1); | |
803 | } | |
804 | ||
805 | element->lhs = lhs; | |
806 | element->hash = avail_expr_hash (element); | |
807 | } | |
808 | ||
809 | /* Remove all the expressions in LOCALS from TABLE, stopping when there are | |
810 | LIMIT entries left in LOCALs. */ | |
811 | ||
812 | static void | |
48732f23 | 813 | remove_local_expressions_from_table (void) |
6de9cd9a | 814 | { |
6de9cd9a | 815 | /* Remove all the expressions made available in this block. */ |
48732f23 | 816 | while (VARRAY_ACTIVE_SIZE (avail_exprs_stack) > 0) |
6de9cd9a DN |
817 | { |
818 | struct expr_hash_elt element; | |
48732f23 JL |
819 | tree expr = VARRAY_TOP_TREE (avail_exprs_stack); |
820 | VARRAY_POP (avail_exprs_stack); | |
821 | ||
822 | if (expr == NULL_TREE) | |
823 | break; | |
6de9cd9a DN |
824 | |
825 | initialize_hash_element (expr, NULL, &element); | |
48732f23 | 826 | htab_remove_elt_with_hash (avail_exprs, &element, element.hash); |
6de9cd9a DN |
827 | } |
828 | } | |
829 | ||
830 | /* Use the SSA_NAMES in LOCALS to restore TABLE to its original | |
1ea7e6ad | 831 | state, stopping when there are LIMIT entries left in LOCALs. */ |
6de9cd9a DN |
832 | |
833 | static void | |
834 | restore_nonzero_vars_to_original_value (varray_type locals, | |
835 | unsigned limit, | |
836 | bitmap table) | |
837 | { | |
838 | if (!locals) | |
839 | return; | |
840 | ||
841 | while (VARRAY_ACTIVE_SIZE (locals) > limit) | |
842 | { | |
843 | tree name = VARRAY_TOP_TREE (locals); | |
844 | VARRAY_POP (locals); | |
845 | bitmap_clear_bit (table, SSA_NAME_VERSION (name)); | |
846 | } | |
847 | } | |
848 | ||
b5fefcf6 JL |
849 | /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore |
850 | CONST_AND_COPIES to its original state, stopping when we hit a | |
851 | NULL marker. */ | |
6de9cd9a DN |
852 | |
853 | static void | |
b5fefcf6 | 854 | restore_vars_to_original_value (void) |
6de9cd9a | 855 | { |
b5fefcf6 | 856 | while (VARRAY_ACTIVE_SIZE (const_and_copies_stack) > 0) |
6de9cd9a DN |
857 | { |
858 | tree prev_value, dest; | |
859 | ||
b5fefcf6 JL |
860 | dest = VARRAY_TOP_TREE (const_and_copies_stack); |
861 | VARRAY_POP (const_and_copies_stack); | |
6de9cd9a | 862 | |
b5fefcf6 JL |
863 | if (dest == NULL) |
864 | break; | |
865 | ||
866 | prev_value = VARRAY_TOP_TREE (const_and_copies_stack); | |
867 | VARRAY_POP (const_and_copies_stack); | |
868 | ||
869 | set_value_for (dest, prev_value, const_and_copies); | |
6de9cd9a DN |
870 | } |
871 | } | |
872 | ||
873 | /* Similar to restore_vars_to_original_value, except that it restores | |
874 | CURRDEFS to its original value. */ | |
875 | static void | |
9fae925b | 876 | restore_currdefs_to_original_value (void) |
6de9cd9a | 877 | { |
6de9cd9a | 878 | /* Restore CURRDEFS to its original state. */ |
9fae925b | 879 | while (VARRAY_ACTIVE_SIZE (block_defs_stack) > 0) |
6de9cd9a | 880 | { |
9fae925b | 881 | tree tmp = VARRAY_TOP_TREE (block_defs_stack); |
6de9cd9a DN |
882 | tree saved_def, var; |
883 | ||
9fae925b JL |
884 | VARRAY_POP (block_defs_stack); |
885 | ||
886 | if (tmp == NULL_TREE) | |
887 | break; | |
6de9cd9a DN |
888 | |
889 | /* If we recorded an SSA_NAME, then make the SSA_NAME the current | |
890 | definition of its underlying variable. If we recorded anything | |
891 | else, it must have been an _DECL node and its current reaching | |
892 | definition must have been NULL. */ | |
893 | if (TREE_CODE (tmp) == SSA_NAME) | |
894 | { | |
895 | saved_def = tmp; | |
896 | var = SSA_NAME_VAR (saved_def); | |
897 | } | |
898 | else | |
899 | { | |
900 | saved_def = NULL; | |
901 | var = tmp; | |
902 | } | |
903 | ||
904 | var_ann (var)->current_def = saved_def; | |
905 | } | |
906 | } | |
907 | ||
908 | /* We have finished processing the dominator children of BB, perform | |
909 | any finalization actions in preparation for leaving this node in | |
910 | the dominator tree. */ | |
911 | ||
912 | static void | |
913 | dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb) | |
914 | { | |
915 | struct dom_walk_block_data *bd | |
916 | = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack); | |
917 | tree last; | |
918 | ||
919 | /* If we are at a leaf node in the dominator graph, see if we can thread | |
920 | the edge from BB through its successor. | |
921 | ||
922 | Do this before we remove entries from our equivalence tables. */ | |
923 | if (bb->succ | |
924 | && ! bb->succ->succ_next | |
925 | && (bb->succ->flags & EDGE_ABNORMAL) == 0 | |
926 | && (get_immediate_dominator (CDI_DOMINATORS, bb->succ->dest) != bb | |
927 | || phi_nodes (bb->succ->dest))) | |
928 | ||
929 | { | |
930 | thread_across_edge (walk_data, bb->succ); | |
931 | } | |
932 | else if ((last = last_stmt (bb)) | |
933 | && TREE_CODE (last) == COND_EXPR | |
6615c446 | 934 | && (COMPARISON_CLASS_P (COND_EXPR_COND (last)) |
6de9cd9a DN |
935 | || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME) |
936 | && bb->succ | |
937 | && (bb->succ->flags & EDGE_ABNORMAL) == 0 | |
938 | && bb->succ->succ_next | |
939 | && (bb->succ->succ_next->flags & EDGE_ABNORMAL) == 0 | |
940 | && ! bb->succ->succ_next->succ_next) | |
941 | { | |
942 | edge true_edge, false_edge; | |
943 | tree cond, inverted = NULL; | |
944 | enum tree_code cond_code; | |
945 | ||
946 | extract_true_false_edges_from_block (bb, &true_edge, &false_edge); | |
947 | ||
948 | cond = COND_EXPR_COND (last); | |
949 | cond_code = TREE_CODE (cond); | |
950 | ||
6615c446 | 951 | if (TREE_CODE_CLASS (cond_code) == tcc_comparison) |
6de9cd9a DN |
952 | inverted = invert_truthvalue (cond); |
953 | ||
954 | /* If the THEN arm is the end of a dominator tree or has PHI nodes, | |
955 | then try to thread through its edge. */ | |
956 | if (get_immediate_dominator (CDI_DOMINATORS, true_edge->dest) != bb | |
957 | || phi_nodes (true_edge->dest)) | |
958 | { | |
48732f23 JL |
959 | /* Push a marker onto the available expression stack so that we |
960 | unwind any expressions related to the TRUE arm before processing | |
961 | the false arm below. */ | |
962 | VARRAY_PUSH_TREE (avail_exprs_stack, NULL_TREE); | |
9fae925b | 963 | VARRAY_PUSH_TREE (block_defs_stack, NULL_TREE); |
b5fefcf6 | 964 | VARRAY_PUSH_TREE (const_and_copies_stack, NULL_TREE); |
48732f23 | 965 | |
6de9cd9a | 966 | /* Record any equivalences created by following this edge. */ |
6615c446 | 967 | if (TREE_CODE_CLASS (cond_code) == tcc_comparison) |
6de9cd9a | 968 | { |
48732f23 JL |
969 | record_cond (cond, boolean_true_node); |
970 | record_dominating_conditions (cond); | |
971 | record_cond (inverted, boolean_false_node); | |
6de9cd9a DN |
972 | } |
973 | else if (cond_code == SSA_NAME) | |
b5fefcf6 | 974 | record_const_or_copy (cond, boolean_true_node); |
6de9cd9a DN |
975 | |
976 | /* Now thread the edge. */ | |
977 | thread_across_edge (walk_data, true_edge); | |
978 | ||
979 | /* And restore the various tables to their state before | |
980 | we threaded this edge. */ | |
48732f23 | 981 | remove_local_expressions_from_table (); |
b5fefcf6 | 982 | restore_vars_to_original_value (); |
9fae925b | 983 | restore_currdefs_to_original_value (); |
6de9cd9a DN |
984 | } |
985 | ||
986 | /* Similarly for the ELSE arm. */ | |
987 | if (get_immediate_dominator (CDI_DOMINATORS, false_edge->dest) != bb | |
988 | || phi_nodes (false_edge->dest)) | |
989 | { | |
990 | /* Record any equivalences created by following this edge. */ | |
6615c446 | 991 | if (TREE_CODE_CLASS (cond_code) == tcc_comparison) |
6de9cd9a | 992 | { |
48732f23 JL |
993 | record_cond (cond, boolean_false_node); |
994 | record_cond (inverted, boolean_true_node); | |
995 | record_dominating_conditions (inverted); | |
6de9cd9a DN |
996 | } |
997 | else if (cond_code == SSA_NAME) | |
b5fefcf6 | 998 | record_const_or_copy (cond, boolean_false_node); |
6de9cd9a DN |
999 | |
1000 | thread_across_edge (walk_data, false_edge); | |
1001 | ||
1002 | /* No need to remove local expressions from our tables | |
1003 | or restore vars to their original value as that will | |
1004 | be done immediately below. */ | |
1005 | } | |
1006 | } | |
1007 | ||
48732f23 | 1008 | remove_local_expressions_from_table (); |
6de9cd9a | 1009 | restore_nonzero_vars_to_original_value (bd->nonzero_vars, 0, nonzero_vars); |
b5fefcf6 | 1010 | restore_vars_to_original_value (); |
9fae925b | 1011 | restore_currdefs_to_original_value (); |
6de9cd9a DN |
1012 | |
1013 | /* Remove VRP records associated with this basic block. They are no | |
1014 | longer valid. | |
1015 | ||
1016 | To be efficient, we note which variables have had their values | |
1017 | constrained in this block. So walk over each variable in the | |
1018 | VRP_VARIABLEs array. */ | |
1019 | while (bd->vrp_variables && VARRAY_ACTIVE_SIZE (bd->vrp_variables) > 0) | |
1020 | { | |
1021 | tree var = VARRAY_TOP_TREE (bd->vrp_variables); | |
1022 | ||
1023 | /* Each variable has a stack of value range records. We want to | |
1024 | invalidate those associated with our basic block. So we walk | |
1025 | the array backwards popping off records associated with our | |
1026 | block. Once we hit a record not associated with our block | |
1027 | we are done. */ | |
1028 | varray_type var_vrp_records = VARRAY_GENERIC_PTR (vrp_data, | |
1029 | SSA_NAME_VERSION (var)); | |
1030 | ||
1031 | while (VARRAY_ACTIVE_SIZE (var_vrp_records) > 0) | |
1032 | { | |
1033 | struct vrp_element *element | |
1034 | = (struct vrp_element *)VARRAY_TOP_GENERIC_PTR (var_vrp_records); | |
1035 | ||
1036 | if (element->bb != bb) | |
1037 | break; | |
1038 | ||
1039 | VARRAY_POP (var_vrp_records); | |
1040 | } | |
1041 | ||
1042 | VARRAY_POP (bd->vrp_variables); | |
1043 | } | |
1044 | ||
a6e1aa26 JL |
1045 | /* If we queued any statements to rescan in this block, then |
1046 | go ahead and rescan them now. */ | |
1047 | while (VARRAY_ACTIVE_SIZE (stmts_to_rescan) > 0) | |
6de9cd9a | 1048 | { |
a6e1aa26 JL |
1049 | tree stmt = VARRAY_TOP_TREE (stmts_to_rescan); |
1050 | basic_block stmt_bb = bb_for_stmt (stmt); | |
1051 | ||
1052 | if (stmt_bb != bb) | |
1053 | break; | |
1054 | ||
1055 | VARRAY_POP (stmts_to_rescan); | |
6de9cd9a DN |
1056 | mark_new_vars_to_rename (stmt, vars_to_rename); |
1057 | } | |
1058 | } | |
1059 | ||
1060 | /* PHI nodes can create equivalences too. | |
1061 | ||
1062 | Ignoring any alternatives which are the same as the result, if | |
1063 | all the alternatives are equal, then the PHI node creates an | |
dd747311 JL |
1064 | equivalence. |
1065 | ||
1066 | Additionally, if all the PHI alternatives are known to have a nonzero | |
1067 | value, then the result of this PHI is known to have a nonzero value, | |
1068 | even if we do not know its exact value. */ | |
1069 | ||
6de9cd9a | 1070 | static void |
9fae925b JL |
1071 | record_equivalences_from_phis (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED, |
1072 | basic_block bb) | |
6de9cd9a | 1073 | { |
6de9cd9a DN |
1074 | tree phi; |
1075 | ||
17192884 | 1076 | for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi)) |
6de9cd9a DN |
1077 | { |
1078 | tree lhs = PHI_RESULT (phi); | |
1079 | tree rhs = NULL; | |
1080 | int i; | |
1081 | ||
1082 | for (i = 0; i < PHI_NUM_ARGS (phi); i++) | |
1083 | { | |
1084 | tree t = PHI_ARG_DEF (phi, i); | |
1085 | ||
1086 | if (TREE_CODE (t) == SSA_NAME || is_gimple_min_invariant (t)) | |
1087 | { | |
1088 | /* Ignore alternatives which are the same as our LHS. */ | |
1089 | if (operand_equal_p (lhs, t, 0)) | |
1090 | continue; | |
1091 | ||
1092 | /* If we have not processed an alternative yet, then set | |
1093 | RHS to this alternative. */ | |
1094 | if (rhs == NULL) | |
1095 | rhs = t; | |
1096 | /* If we have processed an alternative (stored in RHS), then | |
1097 | see if it is equal to this one. If it isn't, then stop | |
1098 | the search. */ | |
1099 | else if (! operand_equal_p (rhs, t, 0)) | |
1100 | break; | |
1101 | } | |
1102 | else | |
1103 | break; | |
1104 | } | |
1105 | ||
1106 | /* If we had no interesting alternatives, then all the RHS alternatives | |
1107 | must have been the same as LHS. */ | |
1108 | if (!rhs) | |
1109 | rhs = lhs; | |
1110 | ||
1111 | /* If we managed to iterate through each PHI alternative without | |
1112 | breaking out of the loop, then we have a PHI which may create | |
1113 | a useful equivalence. We do not need to record unwind data for | |
1114 | this, since this is a true assignment and not an equivalence | |
1ea7e6ad | 1115 | inferred from a comparison. All uses of this ssa name are dominated |
6de9cd9a DN |
1116 | by this assignment, so unwinding just costs time and space. */ |
1117 | if (i == PHI_NUM_ARGS (phi) | |
1118 | && may_propagate_copy (lhs, rhs)) | |
1119 | set_value_for (lhs, rhs, const_and_copies); | |
1120 | ||
dd747311 JL |
1121 | /* Now see if we know anything about the nonzero property for the |
1122 | result of this PHI. */ | |
1123 | for (i = 0; i < PHI_NUM_ARGS (phi); i++) | |
1124 | { | |
1125 | if (!PHI_ARG_NONZERO (phi, i)) | |
1126 | break; | |
1127 | } | |
1128 | ||
1129 | if (i == PHI_NUM_ARGS (phi)) | |
1130 | bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi))); | |
1131 | ||
9fae925b | 1132 | register_new_def (lhs, &block_defs_stack); |
6de9cd9a DN |
1133 | } |
1134 | } | |
1135 | ||
28c008bb JL |
1136 | /* Ignoring loop backedges, if BB has precisely one incoming edge then |
1137 | return that edge. Otherwise return NULL. */ | |
1138 | static edge | |
1139 | single_incoming_edge_ignoring_loop_edges (basic_block bb) | |
1140 | { | |
1141 | edge retval = NULL; | |
1142 | edge e; | |
1143 | ||
1144 | for (e = bb->pred; e; e = e->pred_next) | |
1145 | { | |
1146 | /* A loop back edge can be identified by the destination of | |
1147 | the edge dominating the source of the edge. */ | |
1148 | if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest)) | |
1149 | continue; | |
1150 | ||
1151 | /* If we have already seen a non-loop edge, then we must have | |
1152 | multiple incoming non-loop edges and thus we return NULL. */ | |
1153 | if (retval) | |
1154 | return NULL; | |
1155 | ||
1156 | /* This is the first non-loop incoming edge we have found. Record | |
1157 | it. */ | |
1158 | retval = e; | |
1159 | } | |
1160 | ||
1161 | return retval; | |
1162 | } | |
1163 | ||
6de9cd9a DN |
1164 | /* Record any equivalences created by the incoming edge to BB. If BB |
1165 | has more than one incoming edge, then no equivalence is created. */ | |
1166 | ||
1167 | static void | |
1168 | record_equivalences_from_incoming_edge (struct dom_walk_data *walk_data, | |
1169 | basic_block bb) | |
1170 | { | |
1171 | int edge_flags; | |
1172 | basic_block parent; | |
1173 | struct eq_expr_value eq_expr_value; | |
1174 | tree parent_block_last_stmt = NULL; | |
1175 | struct dom_walk_block_data *bd | |
1176 | = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack); | |
1177 | ||
1178 | /* If our parent block ended with a control statment, then we may be | |
1179 | able to record some equivalences based on which outgoing edge from | |
1180 | the parent was followed. */ | |
1181 | parent = get_immediate_dominator (CDI_DOMINATORS, bb); | |
1182 | if (parent) | |
1183 | { | |
1184 | parent_block_last_stmt = last_stmt (parent); | |
1185 | if (parent_block_last_stmt && !is_ctrl_stmt (parent_block_last_stmt)) | |
1186 | parent_block_last_stmt = NULL; | |
1187 | } | |
1188 | ||
1189 | eq_expr_value.src = NULL; | |
1190 | eq_expr_value.dst = NULL; | |
1191 | ||
28c008bb JL |
1192 | /* If we have a single predecessor (ignoring loop backedges), then extract |
1193 | EDGE_FLAGS from the single incoming edge. Otherwise just return as | |
1194 | there is nothing to do. */ | |
6de9cd9a | 1195 | if (bb->pred |
28c008bb | 1196 | && parent_block_last_stmt) |
6de9cd9a | 1197 | { |
28c008bb JL |
1198 | edge e = single_incoming_edge_ignoring_loop_edges (bb); |
1199 | if (e && bb_for_stmt (parent_block_last_stmt) == e->src) | |
1200 | edge_flags = e->flags; | |
1201 | else | |
1202 | return; | |
6de9cd9a DN |
1203 | } |
1204 | else | |
28c008bb | 1205 | return; |
6de9cd9a DN |
1206 | |
1207 | /* If our parent block ended in a COND_EXPR, add any equivalences | |
1208 | created by the COND_EXPR to the hash table and initialize | |
1209 | EQ_EXPR_VALUE appropriately. | |
1210 | ||
1211 | EQ_EXPR_VALUE is an assignment expression created when BB's immediate | |
1212 | dominator ends in a COND_EXPR statement whose predicate is of the form | |
1213 | 'VAR == VALUE', where VALUE may be another variable or a constant. | |
1214 | This is used to propagate VALUE on the THEN_CLAUSE of that | |
1215 | conditional. This assignment is inserted in CONST_AND_COPIES so that | |
1216 | the copy and constant propagator can find more propagation | |
1217 | opportunities. */ | |
28c008bb | 1218 | if (TREE_CODE (parent_block_last_stmt) == COND_EXPR |
6de9cd9a DN |
1219 | && (edge_flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) |
1220 | eq_expr_value = get_eq_expr_value (parent_block_last_stmt, | |
1221 | (edge_flags & EDGE_TRUE_VALUE) != 0, | |
6de9cd9a DN |
1222 | bb, |
1223 | &bd->vrp_variables); | |
1c052514 SB |
1224 | /* Similarly when the parent block ended in a SWITCH_EXPR. |
1225 | We can only know the value of the switch's condition if the dominator | |
1226 | parent is also the only predecessor of this block. */ | |
28c008bb | 1227 | else if (bb->pred->src == parent |
6de9cd9a DN |
1228 | && TREE_CODE (parent_block_last_stmt) == SWITCH_EXPR) |
1229 | { | |
1230 | tree switch_cond = SWITCH_COND (parent_block_last_stmt); | |
1231 | ||
1232 | /* If the switch's condition is an SSA variable, then we may | |
1233 | know its value at each of the case labels. */ | |
1234 | if (TREE_CODE (switch_cond) == SSA_NAME) | |
1235 | { | |
1236 | tree switch_vec = SWITCH_LABELS (parent_block_last_stmt); | |
1237 | size_t i, n = TREE_VEC_LENGTH (switch_vec); | |
1238 | int case_count = 0; | |
1239 | tree match_case = NULL_TREE; | |
1240 | ||
1241 | /* Search the case labels for those whose destination is | |
1242 | the current basic block. */ | |
1243 | for (i = 0; i < n; ++i) | |
1244 | { | |
1245 | tree elt = TREE_VEC_ELT (switch_vec, i); | |
1246 | if (label_to_block (CASE_LABEL (elt)) == bb) | |
1247 | { | |
1c052514 | 1248 | if (++case_count > 1 || CASE_HIGH (elt)) |
6de9cd9a DN |
1249 | break; |
1250 | match_case = elt; | |
1251 | } | |
1252 | } | |
1253 | ||
1254 | /* If we encountered precisely one CASE_LABEL_EXPR and it | |
1255 | was not the default case, or a case range, then we know | |
1256 | the exact value of SWITCH_COND which caused us to get to | |
1257 | this block. Record that equivalence in EQ_EXPR_VALUE. */ | |
1258 | if (case_count == 1 | |
1c052514 | 1259 | && match_case |
6de9cd9a DN |
1260 | && CASE_LOW (match_case) |
1261 | && !CASE_HIGH (match_case)) | |
1262 | { | |
1263 | eq_expr_value.dst = switch_cond; | |
e9ea8bd5 RS |
1264 | eq_expr_value.src = fold_convert (TREE_TYPE (switch_cond), |
1265 | CASE_LOW (match_case)); | |
6de9cd9a DN |
1266 | } |
1267 | } | |
1268 | } | |
1269 | ||
1270 | /* If EQ_EXPR_VALUE (VAR == VALUE) is given, register the VALUE as a | |
1271 | new value for VAR, so that occurrences of VAR can be replaced with | |
1272 | VALUE while re-writing the THEN arm of a COND_EXPR. */ | |
1273 | if (eq_expr_value.src && eq_expr_value.dst) | |
b5fefcf6 | 1274 | record_equality (eq_expr_value.dst, eq_expr_value.src); |
6de9cd9a DN |
1275 | } |
1276 | ||
1277 | /* Dump SSA statistics on FILE. */ | |
1278 | ||
1279 | void | |
1280 | dump_dominator_optimization_stats (FILE *file) | |
1281 | { | |
1282 | long n_exprs; | |
1283 | ||
1284 | fprintf (file, "Total number of statements: %6ld\n\n", | |
1285 | opt_stats.num_stmts); | |
1286 | fprintf (file, "Exprs considered for dominator optimizations: %6ld\n", | |
1287 | opt_stats.num_exprs_considered); | |
1288 | ||
1289 | n_exprs = opt_stats.num_exprs_considered; | |
1290 | if (n_exprs == 0) | |
1291 | n_exprs = 1; | |
1292 | ||
1293 | fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n", | |
1294 | opt_stats.num_re, PERCENT (opt_stats.num_re, | |
1295 | n_exprs)); | |
1296 | ||
1297 | fprintf (file, "\nHash table statistics:\n"); | |
1298 | ||
1299 | fprintf (file, " avail_exprs: "); | |
1300 | htab_statistics (file, avail_exprs); | |
1301 | } | |
1302 | ||
1303 | ||
1304 | /* Dump SSA statistics on stderr. */ | |
1305 | ||
1306 | void | |
1307 | debug_dominator_optimization_stats (void) | |
1308 | { | |
1309 | dump_dominator_optimization_stats (stderr); | |
1310 | } | |
1311 | ||
1312 | ||
1313 | /* Dump statistics for the hash table HTAB. */ | |
1314 | ||
1315 | static void | |
1316 | htab_statistics (FILE *file, htab_t htab) | |
1317 | { | |
1318 | fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n", | |
1319 | (long) htab_size (htab), | |
1320 | (long) htab_elements (htab), | |
1321 | htab_collisions (htab)); | |
1322 | } | |
1323 | ||
1324 | /* Record the fact that VAR has a nonzero value, though we may not know | |
1325 | its exact value. Note that if VAR is already known to have a nonzero | |
1326 | value, then we do nothing. */ | |
1327 | ||
1328 | static void | |
1329 | record_var_is_nonzero (tree var, varray_type *block_nonzero_vars_p) | |
1330 | { | |
1331 | int indx = SSA_NAME_VERSION (var); | |
1332 | ||
1333 | if (bitmap_bit_p (nonzero_vars, indx)) | |
1334 | return; | |
1335 | ||
1336 | /* Mark it in the global table. */ | |
1337 | bitmap_set_bit (nonzero_vars, indx); | |
1338 | ||
1339 | /* Record this SSA_NAME so that we can reset the global table | |
1340 | when we leave this block. */ | |
1341 | if (! *block_nonzero_vars_p) | |
1342 | VARRAY_TREE_INIT (*block_nonzero_vars_p, 2, "block_nonzero_vars"); | |
1343 | VARRAY_PUSH_TREE (*block_nonzero_vars_p, var); | |
1344 | } | |
1345 | ||
1346 | /* Enter a statement into the true/false expression hash table indicating | |
1347 | that the condition COND has the value VALUE. */ | |
1348 | ||
1349 | static void | |
48732f23 | 1350 | record_cond (tree cond, tree value) |
6de9cd9a DN |
1351 | { |
1352 | struct expr_hash_elt *element = xmalloc (sizeof (struct expr_hash_elt)); | |
1353 | void **slot; | |
1354 | ||
1355 | initialize_hash_element (cond, value, element); | |
1356 | ||
1357 | slot = htab_find_slot_with_hash (avail_exprs, (void *)element, | |
1358 | element->hash, true); | |
1359 | if (*slot == NULL) | |
1360 | { | |
1361 | *slot = (void *) element; | |
48732f23 | 1362 | VARRAY_PUSH_TREE (avail_exprs_stack, cond); |
6de9cd9a DN |
1363 | } |
1364 | else | |
1365 | free (element); | |
1366 | } | |
1367 | ||
d2d8936f JL |
1368 | /* COND is a condition which is known to be true. Record variants of |
1369 | COND which must also be true. | |
1370 | ||
1371 | For example, if a < b is true, then a <= b must also be true. */ | |
1372 | ||
1373 | static void | |
48732f23 | 1374 | record_dominating_conditions (tree cond) |
d2d8936f JL |
1375 | { |
1376 | switch (TREE_CODE (cond)) | |
1377 | { | |
1378 | case LT_EXPR: | |
1379 | record_cond (build2 (LE_EXPR, boolean_type_node, | |
1380 | TREE_OPERAND (cond, 0), | |
1381 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1382 | boolean_true_node); |
d2d8936f JL |
1383 | record_cond (build2 (ORDERED_EXPR, boolean_type_node, |
1384 | TREE_OPERAND (cond, 0), | |
1385 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1386 | boolean_true_node); |
d2d8936f JL |
1387 | record_cond (build2 (NE_EXPR, boolean_type_node, |
1388 | TREE_OPERAND (cond, 0), | |
1389 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1390 | boolean_true_node); |
d2d8936f JL |
1391 | record_cond (build2 (LTGT_EXPR, boolean_type_node, |
1392 | TREE_OPERAND (cond, 0), | |
1393 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1394 | boolean_true_node); |
d2d8936f JL |
1395 | break; |
1396 | ||
1397 | case GT_EXPR: | |
1398 | record_cond (build2 (GE_EXPR, boolean_type_node, | |
1399 | TREE_OPERAND (cond, 0), | |
1400 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1401 | boolean_true_node); |
d2d8936f JL |
1402 | record_cond (build2 (ORDERED_EXPR, boolean_type_node, |
1403 | TREE_OPERAND (cond, 0), | |
1404 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1405 | boolean_true_node); |
d2d8936f JL |
1406 | record_cond (build2 (NE_EXPR, boolean_type_node, |
1407 | TREE_OPERAND (cond, 0), | |
1408 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1409 | boolean_true_node); |
d2d8936f JL |
1410 | record_cond (build2 (LTGT_EXPR, boolean_type_node, |
1411 | TREE_OPERAND (cond, 0), | |
1412 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1413 | boolean_true_node); |
d2d8936f JL |
1414 | break; |
1415 | ||
1416 | case GE_EXPR: | |
1417 | case LE_EXPR: | |
1418 | record_cond (build2 (ORDERED_EXPR, boolean_type_node, | |
1419 | TREE_OPERAND (cond, 0), | |
1420 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1421 | boolean_true_node); |
d2d8936f JL |
1422 | break; |
1423 | ||
1424 | case EQ_EXPR: | |
1425 | record_cond (build2 (ORDERED_EXPR, boolean_type_node, | |
1426 | TREE_OPERAND (cond, 0), | |
1427 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1428 | boolean_true_node); |
d2d8936f JL |
1429 | record_cond (build2 (LE_EXPR, boolean_type_node, |
1430 | TREE_OPERAND (cond, 0), | |
1431 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1432 | boolean_true_node); |
d2d8936f JL |
1433 | record_cond (build2 (GE_EXPR, boolean_type_node, |
1434 | TREE_OPERAND (cond, 0), | |
1435 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1436 | boolean_true_node); |
d2d8936f JL |
1437 | break; |
1438 | ||
1439 | case UNORDERED_EXPR: | |
1440 | record_cond (build2 (NE_EXPR, boolean_type_node, | |
1441 | TREE_OPERAND (cond, 0), | |
1442 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1443 | boolean_true_node); |
d2d8936f JL |
1444 | record_cond (build2 (UNLE_EXPR, boolean_type_node, |
1445 | TREE_OPERAND (cond, 0), | |
1446 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1447 | boolean_true_node); |
d2d8936f JL |
1448 | record_cond (build2 (UNGE_EXPR, boolean_type_node, |
1449 | TREE_OPERAND (cond, 0), | |
1450 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1451 | boolean_true_node); |
d2d8936f JL |
1452 | record_cond (build2 (UNEQ_EXPR, boolean_type_node, |
1453 | TREE_OPERAND (cond, 0), | |
1454 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1455 | boolean_true_node); |
d2d8936f JL |
1456 | record_cond (build2 (UNLT_EXPR, boolean_type_node, |
1457 | TREE_OPERAND (cond, 0), | |
1458 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1459 | boolean_true_node); |
d2d8936f JL |
1460 | record_cond (build2 (UNGT_EXPR, boolean_type_node, |
1461 | TREE_OPERAND (cond, 0), | |
1462 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1463 | boolean_true_node); |
d2d8936f JL |
1464 | break; |
1465 | ||
1466 | case UNLT_EXPR: | |
1467 | record_cond (build2 (UNLE_EXPR, boolean_type_node, | |
1468 | TREE_OPERAND (cond, 0), | |
1469 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1470 | boolean_true_node); |
d2d8936f JL |
1471 | record_cond (build2 (NE_EXPR, boolean_type_node, |
1472 | TREE_OPERAND (cond, 0), | |
1473 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1474 | boolean_true_node); |
d2d8936f JL |
1475 | break; |
1476 | ||
1477 | case UNGT_EXPR: | |
1478 | record_cond (build2 (UNGE_EXPR, boolean_type_node, | |
1479 | TREE_OPERAND (cond, 0), | |
1480 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1481 | boolean_true_node); |
d2d8936f JL |
1482 | record_cond (build2 (NE_EXPR, boolean_type_node, |
1483 | TREE_OPERAND (cond, 0), | |
1484 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1485 | boolean_true_node); |
d2d8936f JL |
1486 | break; |
1487 | ||
1488 | case UNEQ_EXPR: | |
1489 | record_cond (build2 (UNLE_EXPR, boolean_type_node, | |
1490 | TREE_OPERAND (cond, 0), | |
1491 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1492 | boolean_true_node); |
d2d8936f JL |
1493 | record_cond (build2 (UNGE_EXPR, boolean_type_node, |
1494 | TREE_OPERAND (cond, 0), | |
1495 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1496 | boolean_true_node); |
d2d8936f JL |
1497 | break; |
1498 | ||
1499 | case LTGT_EXPR: | |
1500 | record_cond (build2 (NE_EXPR, boolean_type_node, | |
1501 | TREE_OPERAND (cond, 0), | |
1502 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1503 | boolean_true_node); |
d2d8936f JL |
1504 | record_cond (build2 (ORDERED_EXPR, boolean_type_node, |
1505 | TREE_OPERAND (cond, 0), | |
1506 | TREE_OPERAND (cond, 1)), | |
48732f23 | 1507 | boolean_true_node); |
d2d8936f JL |
1508 | |
1509 | default: | |
1510 | break; | |
1511 | } | |
1512 | } | |
1513 | ||
6de9cd9a DN |
1514 | /* A helper function for record_const_or_copy and record_equality. |
1515 | Do the work of recording the value and undo info. */ | |
1516 | ||
1517 | static void | |
b5fefcf6 | 1518 | record_const_or_copy_1 (tree x, tree y, tree prev_x) |
6de9cd9a DN |
1519 | { |
1520 | set_value_for (x, y, const_and_copies); | |
1521 | ||
b5fefcf6 JL |
1522 | VARRAY_PUSH_TREE (const_and_copies_stack, prev_x); |
1523 | VARRAY_PUSH_TREE (const_and_copies_stack, x); | |
6de9cd9a DN |
1524 | } |
1525 | ||
1526 | /* Record that X is equal to Y in const_and_copies. Record undo | |
1527 | information in the block-local varray. */ | |
1528 | ||
1529 | static void | |
b5fefcf6 | 1530 | record_const_or_copy (tree x, tree y) |
6de9cd9a DN |
1531 | { |
1532 | tree prev_x = get_value_for (x, const_and_copies); | |
1533 | ||
1534 | if (TREE_CODE (y) == SSA_NAME) | |
1535 | { | |
1536 | tree tmp = get_value_for (y, const_and_copies); | |
1537 | if (tmp) | |
1538 | y = tmp; | |
1539 | } | |
1540 | ||
b5fefcf6 | 1541 | record_const_or_copy_1 (x, y, prev_x); |
6de9cd9a DN |
1542 | } |
1543 | ||
1544 | /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR. | |
1545 | This constrains the cases in which we may treat this as assignment. */ | |
1546 | ||
1547 | static void | |
b5fefcf6 | 1548 | record_equality (tree x, tree y) |
6de9cd9a DN |
1549 | { |
1550 | tree prev_x = NULL, prev_y = NULL; | |
1551 | ||
1552 | if (TREE_CODE (x) == SSA_NAME) | |
1553 | prev_x = get_value_for (x, const_and_copies); | |
1554 | if (TREE_CODE (y) == SSA_NAME) | |
1555 | prev_y = get_value_for (y, const_and_copies); | |
1556 | ||
1557 | /* If one of the previous values is invariant, then use that. | |
1558 | Otherwise it doesn't matter which value we choose, just so | |
1559 | long as we canonicalize on one value. */ | |
1560 | if (TREE_INVARIANT (y)) | |
1561 | ; | |
1562 | else if (TREE_INVARIANT (x)) | |
1563 | prev_x = x, x = y, y = prev_x, prev_x = prev_y; | |
1564 | else if (prev_x && TREE_INVARIANT (prev_x)) | |
1565 | x = y, y = prev_x, prev_x = prev_y; | |
1566 | else if (prev_y) | |
1567 | y = prev_y; | |
1568 | ||
1569 | /* After the swapping, we must have one SSA_NAME. */ | |
1570 | if (TREE_CODE (x) != SSA_NAME) | |
1571 | return; | |
1572 | ||
1573 | /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a | |
1574 | variable compared against zero. If we're honoring signed zeros, | |
1575 | then we cannot record this value unless we know that the value is | |
1ea7e6ad | 1576 | nonzero. */ |
6de9cd9a DN |
1577 | if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x))) |
1578 | && (TREE_CODE (y) != REAL_CST | |
1579 | || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y)))) | |
1580 | return; | |
1581 | ||
b5fefcf6 | 1582 | record_const_or_copy_1 (x, y, prev_x); |
6de9cd9a DN |
1583 | } |
1584 | ||
1585 | /* STMT is a MODIFY_EXPR for which we were unable to find RHS in the | |
1586 | hash tables. Try to simplify the RHS using whatever equivalences | |
1587 | we may have recorded. | |
1588 | ||
1589 | If we are able to simplify the RHS, then lookup the simplified form in | |
1590 | the hash table and return the result. Otherwise return NULL. */ | |
1591 | ||
1592 | static tree | |
1593 | simplify_rhs_and_lookup_avail_expr (struct dom_walk_data *walk_data, | |
68b9f53b | 1594 | tree stmt, int insert) |
6de9cd9a DN |
1595 | { |
1596 | tree rhs = TREE_OPERAND (stmt, 1); | |
1597 | enum tree_code rhs_code = TREE_CODE (rhs); | |
1598 | tree result = NULL; | |
6de9cd9a DN |
1599 | |
1600 | /* If we have lhs = ~x, look and see if we earlier had x = ~y. | |
1601 | In which case we can change this statement to be lhs = y. | |
1602 | Which can then be copy propagated. | |
1603 | ||
1604 | Similarly for negation. */ | |
1605 | if ((rhs_code == BIT_NOT_EXPR || rhs_code == NEGATE_EXPR) | |
1606 | && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME) | |
1607 | { | |
1608 | /* Get the definition statement for our RHS. */ | |
1609 | tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0)); | |
1610 | ||
1611 | /* See if the RHS_DEF_STMT has the same form as our statement. */ | |
1612 | if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR | |
1613 | && TREE_CODE (TREE_OPERAND (rhs_def_stmt, 1)) == rhs_code) | |
1614 | { | |
1615 | tree rhs_def_operand; | |
1616 | ||
1617 | rhs_def_operand = TREE_OPERAND (TREE_OPERAND (rhs_def_stmt, 1), 0); | |
1618 | ||
1619 | /* Verify that RHS_DEF_OPERAND is a suitable SSA variable. */ | |
1620 | if (TREE_CODE (rhs_def_operand) == SSA_NAME | |
1621 | && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs_def_operand)) | |
1622 | result = update_rhs_and_lookup_avail_expr (stmt, | |
1623 | rhs_def_operand, | |
6de9cd9a DN |
1624 | insert); |
1625 | } | |
1626 | } | |
1627 | ||
1628 | /* If we have z = (x OP C1), see if we earlier had x = y OP C2. | |
1629 | If OP is associative, create and fold (y OP C2) OP C1 which | |
1630 | should result in (y OP C3), use that as the RHS for the | |
1631 | assignment. Add minus to this, as we handle it specially below. */ | |
1632 | if ((associative_tree_code (rhs_code) || rhs_code == MINUS_EXPR) | |
1633 | && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME | |
1634 | && is_gimple_min_invariant (TREE_OPERAND (rhs, 1))) | |
1635 | { | |
1636 | tree rhs_def_stmt = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0)); | |
1637 | ||
1638 | /* See if the RHS_DEF_STMT has the same form as our statement. */ | |
1639 | if (TREE_CODE (rhs_def_stmt) == MODIFY_EXPR) | |
1640 | { | |
1641 | tree rhs_def_rhs = TREE_OPERAND (rhs_def_stmt, 1); | |
1642 | enum tree_code rhs_def_code = TREE_CODE (rhs_def_rhs); | |
1643 | ||
1644 | if (rhs_code == rhs_def_code | |
1645 | || (rhs_code == PLUS_EXPR && rhs_def_code == MINUS_EXPR) | |
1646 | || (rhs_code == MINUS_EXPR && rhs_def_code == PLUS_EXPR)) | |
1647 | { | |
1648 | tree def_stmt_op0 = TREE_OPERAND (rhs_def_rhs, 0); | |
1649 | tree def_stmt_op1 = TREE_OPERAND (rhs_def_rhs, 1); | |
1650 | ||
1651 | if (TREE_CODE (def_stmt_op0) == SSA_NAME | |
1652 | && ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def_stmt_op0) | |
1653 | && is_gimple_min_invariant (def_stmt_op1)) | |
1654 | { | |
1655 | tree outer_const = TREE_OPERAND (rhs, 1); | |
1656 | tree type = TREE_TYPE (TREE_OPERAND (stmt, 0)); | |
1657 | tree t; | |
1658 | ||
f05ef422 RH |
1659 | /* If we care about correct floating point results, then |
1660 | don't fold x + c1 - c2. Note that we need to take both | |
1661 | the codes and the signs to figure this out. */ | |
1662 | if (FLOAT_TYPE_P (type) | |
1663 | && !flag_unsafe_math_optimizations | |
1664 | && (rhs_def_code == PLUS_EXPR | |
1665 | || rhs_def_code == MINUS_EXPR)) | |
1666 | { | |
1667 | bool neg = false; | |
1668 | ||
1669 | neg ^= (rhs_code == MINUS_EXPR); | |
1670 | neg ^= (rhs_def_code == MINUS_EXPR); | |
1671 | neg ^= real_isneg (TREE_REAL_CST_PTR (outer_const)); | |
1672 | neg ^= real_isneg (TREE_REAL_CST_PTR (def_stmt_op1)); | |
1673 | ||
1674 | if (neg) | |
1675 | goto dont_fold_assoc; | |
1676 | } | |
1677 | ||
6de9cd9a DN |
1678 | /* Ho hum. So fold will only operate on the outermost |
1679 | thingy that we give it, so we have to build the new | |
1680 | expression in two pieces. This requires that we handle | |
1681 | combinations of plus and minus. */ | |
1682 | if (rhs_def_code != rhs_code) | |
1683 | { | |
1684 | if (rhs_def_code == MINUS_EXPR) | |
1685 | t = build (MINUS_EXPR, type, outer_const, def_stmt_op1); | |
1686 | else | |
1687 | t = build (MINUS_EXPR, type, def_stmt_op1, outer_const); | |
1688 | rhs_code = PLUS_EXPR; | |
1689 | } | |
1690 | else if (rhs_def_code == MINUS_EXPR) | |
1691 | t = build (PLUS_EXPR, type, def_stmt_op1, outer_const); | |
1692 | else | |
1693 | t = build (rhs_def_code, type, def_stmt_op1, outer_const); | |
1694 | t = local_fold (t); | |
1695 | t = build (rhs_code, type, def_stmt_op0, t); | |
1696 | t = local_fold (t); | |
1697 | ||
1698 | /* If the result is a suitable looking gimple expression, | |
1699 | then use it instead of the original for STMT. */ | |
1700 | if (TREE_CODE (t) == SSA_NAME | |
6615c446 | 1701 | || (UNARY_CLASS_P (t) |
6de9cd9a | 1702 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME) |
6615c446 | 1703 | || ((BINARY_CLASS_P (t) || COMPARISON_CLASS_P (t)) |
6de9cd9a DN |
1704 | && TREE_CODE (TREE_OPERAND (t, 0)) == SSA_NAME |
1705 | && is_gimple_val (TREE_OPERAND (t, 1)))) | |
48732f23 | 1706 | result = update_rhs_and_lookup_avail_expr (stmt, t, insert); |
6de9cd9a DN |
1707 | } |
1708 | } | |
1709 | } | |
f05ef422 | 1710 | dont_fold_assoc:; |
6de9cd9a DN |
1711 | } |
1712 | ||
1713 | /* Transform TRUNC_DIV_EXPR and TRUNC_MOD_EXPR into RSHIFT_EXPR | |
1714 | and BIT_AND_EXPR respectively if the first operand is greater | |
1715 | than zero and the second operand is an exact power of two. */ | |
1716 | if ((rhs_code == TRUNC_DIV_EXPR || rhs_code == TRUNC_MOD_EXPR) | |
1717 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0))) | |
1718 | && integer_pow2p (TREE_OPERAND (rhs, 1))) | |
1719 | { | |
1720 | tree val; | |
1721 | tree op = TREE_OPERAND (rhs, 0); | |
1722 | ||
1723 | if (TYPE_UNSIGNED (TREE_TYPE (op))) | |
1724 | { | |
1725 | val = integer_one_node; | |
1726 | } | |
1727 | else | |
1728 | { | |
1729 | tree dummy_cond = walk_data->global_data; | |
1730 | ||
1731 | if (! dummy_cond) | |
1732 | { | |
1733 | dummy_cond = build (GT_EXPR, boolean_type_node, | |
1734 | op, integer_zero_node); | |
1735 | dummy_cond = build (COND_EXPR, void_type_node, | |
1736 | dummy_cond, NULL, NULL); | |
1737 | walk_data->global_data = dummy_cond; | |
1738 | } | |
1739 | else | |
1740 | { | |
1741 | TREE_SET_CODE (TREE_OPERAND (dummy_cond, 0), GT_EXPR); | |
1742 | TREE_OPERAND (TREE_OPERAND (dummy_cond, 0), 0) = op; | |
1743 | TREE_OPERAND (TREE_OPERAND (dummy_cond, 0), 1) | |
1744 | = integer_zero_node; | |
1745 | } | |
48732f23 | 1746 | val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false); |
6de9cd9a DN |
1747 | } |
1748 | ||
1749 | if (val && integer_onep (val)) | |
1750 | { | |
1751 | tree t; | |
1752 | tree op0 = TREE_OPERAND (rhs, 0); | |
1753 | tree op1 = TREE_OPERAND (rhs, 1); | |
1754 | ||
1755 | if (rhs_code == TRUNC_DIV_EXPR) | |
1756 | t = build (RSHIFT_EXPR, TREE_TYPE (op0), op0, | |
7d60be94 | 1757 | build_int_cst (NULL_TREE, tree_log2 (op1))); |
6de9cd9a DN |
1758 | else |
1759 | t = build (BIT_AND_EXPR, TREE_TYPE (op0), op0, | |
1760 | local_fold (build (MINUS_EXPR, TREE_TYPE (op1), | |
1761 | op1, integer_one_node))); | |
1762 | ||
48732f23 | 1763 | result = update_rhs_and_lookup_avail_expr (stmt, t, insert); |
6de9cd9a DN |
1764 | } |
1765 | } | |
1766 | ||
1767 | /* Transform ABS (X) into X or -X as appropriate. */ | |
1768 | if (rhs_code == ABS_EXPR | |
1769 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (rhs, 0)))) | |
1770 | { | |
1771 | tree val; | |
1772 | tree op = TREE_OPERAND (rhs, 0); | |
1773 | tree type = TREE_TYPE (op); | |
1774 | ||
1775 | if (TYPE_UNSIGNED (type)) | |
1776 | { | |
1777 | val = integer_zero_node; | |
1778 | } | |
1779 | else | |
1780 | { | |
1781 | tree dummy_cond = walk_data->global_data; | |
1782 | ||
1783 | if (! dummy_cond) | |
1784 | { | |
14bc8dc2 | 1785 | dummy_cond = build (LE_EXPR, boolean_type_node, |
6de9cd9a DN |
1786 | op, integer_zero_node); |
1787 | dummy_cond = build (COND_EXPR, void_type_node, | |
1788 | dummy_cond, NULL, NULL); | |
1789 | walk_data->global_data = dummy_cond; | |
1790 | } | |
1791 | else | |
1792 | { | |
14bc8dc2 | 1793 | TREE_SET_CODE (TREE_OPERAND (dummy_cond, 0), LE_EXPR); |
6de9cd9a DN |
1794 | TREE_OPERAND (TREE_OPERAND (dummy_cond, 0), 0) = op; |
1795 | TREE_OPERAND (TREE_OPERAND (dummy_cond, 0), 1) | |
5212068f | 1796 | = build_int_cst (type, 0); |
6de9cd9a | 1797 | } |
48732f23 | 1798 | val = simplify_cond_and_lookup_avail_expr (dummy_cond, NULL, false); |
14bc8dc2 JL |
1799 | |
1800 | if (!val) | |
1801 | { | |
1802 | TREE_SET_CODE (TREE_OPERAND (dummy_cond, 0), GE_EXPR); | |
1803 | TREE_OPERAND (TREE_OPERAND (dummy_cond, 0), 0) = op; | |
1804 | TREE_OPERAND (TREE_OPERAND (dummy_cond, 0), 1) | |
5212068f | 1805 | = build_int_cst (type, 0); |
14bc8dc2 JL |
1806 | |
1807 | val = simplify_cond_and_lookup_avail_expr (dummy_cond, | |
14bc8dc2 JL |
1808 | NULL, false); |
1809 | ||
1810 | if (val) | |
1811 | { | |
1812 | if (integer_zerop (val)) | |
1813 | val = integer_one_node; | |
1814 | else if (integer_onep (val)) | |
1815 | val = integer_zero_node; | |
1816 | } | |
1817 | } | |
6de9cd9a DN |
1818 | } |
1819 | ||
1820 | if (val | |
1821 | && (integer_onep (val) || integer_zerop (val))) | |
1822 | { | |
1823 | tree t; | |
1824 | ||
1825 | if (integer_onep (val)) | |
1826 | t = build1 (NEGATE_EXPR, TREE_TYPE (op), op); | |
1827 | else | |
1828 | t = op; | |
1829 | ||
48732f23 | 1830 | result = update_rhs_and_lookup_avail_expr (stmt, t, insert); |
6de9cd9a DN |
1831 | } |
1832 | } | |
1833 | ||
1834 | /* Optimize *"foo" into 'f'. This is done here rather than | |
1835 | in fold to avoid problems with stuff like &*"foo". */ | |
1836 | if (TREE_CODE (rhs) == INDIRECT_REF || TREE_CODE (rhs) == ARRAY_REF) | |
1837 | { | |
1838 | tree t = fold_read_from_constant_string (rhs); | |
1839 | ||
1840 | if (t) | |
48732f23 | 1841 | result = update_rhs_and_lookup_avail_expr (stmt, t, insert); |
6de9cd9a DN |
1842 | } |
1843 | ||
1844 | return result; | |
1845 | } | |
1846 | ||
1847 | /* COND is a condition of the form: | |
1848 | ||
1849 | x == const or x != const | |
1850 | ||
1851 | Look back to x's defining statement and see if x is defined as | |
1852 | ||
1853 | x = (type) y; | |
1854 | ||
1855 | If const is unchanged if we convert it to type, then we can build | |
1856 | the equivalent expression: | |
1857 | ||
1858 | ||
1859 | y == const or y != const | |
1860 | ||
1861 | Which may allow further optimizations. | |
1862 | ||
1863 | Return the equivalent comparison or NULL if no such equivalent comparison | |
1864 | was found. */ | |
1865 | ||
1866 | static tree | |
1867 | find_equivalent_equality_comparison (tree cond) | |
1868 | { | |
1869 | tree op0 = TREE_OPERAND (cond, 0); | |
1870 | tree op1 = TREE_OPERAND (cond, 1); | |
1871 | tree def_stmt = SSA_NAME_DEF_STMT (op0); | |
1872 | ||
1873 | /* OP0 might have been a parameter, so first make sure it | |
1874 | was defined by a MODIFY_EXPR. */ | |
1875 | if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR) | |
1876 | { | |
1877 | tree def_rhs = TREE_OPERAND (def_stmt, 1); | |
1878 | ||
1879 | /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */ | |
1880 | if ((TREE_CODE (def_rhs) == NOP_EXPR | |
1881 | || TREE_CODE (def_rhs) == CONVERT_EXPR) | |
1882 | && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME) | |
1883 | { | |
1884 | tree def_rhs_inner = TREE_OPERAND (def_rhs, 0); | |
1885 | tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner); | |
1886 | tree new; | |
1887 | ||
1888 | if (TYPE_PRECISION (def_rhs_inner_type) | |
1889 | > TYPE_PRECISION (TREE_TYPE (def_rhs))) | |
1890 | return NULL; | |
1891 | ||
1892 | /* What we want to prove is that if we convert OP1 to | |
1893 | the type of the object inside the NOP_EXPR that the | |
1894 | result is still equivalent to SRC. | |
1895 | ||
1896 | If that is true, the build and return new equivalent | |
1897 | condition which uses the source of the typecast and the | |
1898 | new constant (which has only changed its type). */ | |
1899 | new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1); | |
1900 | new = local_fold (new); | |
1901 | if (is_gimple_val (new) && tree_int_cst_equal (new, op1)) | |
1902 | return build (TREE_CODE (cond), TREE_TYPE (cond), | |
1903 | def_rhs_inner, new); | |
1904 | } | |
1905 | } | |
1906 | return NULL; | |
1907 | } | |
1908 | ||
1909 | /* STMT is a COND_EXPR for which we could not trivially determine its | |
1910 | result. This routine attempts to find equivalent forms of the | |
1911 | condition which we may be able to optimize better. It also | |
1912 | uses simple value range propagation to optimize conditionals. */ | |
1913 | ||
1914 | static tree | |
1915 | simplify_cond_and_lookup_avail_expr (tree stmt, | |
6de9cd9a DN |
1916 | stmt_ann_t ann, |
1917 | int insert) | |
1918 | { | |
1919 | tree cond = COND_EXPR_COND (stmt); | |
1920 | ||
6615c446 | 1921 | if (COMPARISON_CLASS_P (cond)) |
6de9cd9a DN |
1922 | { |
1923 | tree op0 = TREE_OPERAND (cond, 0); | |
1924 | tree op1 = TREE_OPERAND (cond, 1); | |
1925 | ||
1926 | if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1)) | |
1927 | { | |
1928 | int limit; | |
1929 | tree low, high, cond_low, cond_high; | |
1930 | int lowequal, highequal, swapped, no_overlap, subset, cond_inverted; | |
1931 | varray_type vrp_records; | |
1932 | struct vrp_element *element; | |
1933 | ||
1934 | /* First see if we have test of an SSA_NAME against a constant | |
1935 | where the SSA_NAME is defined by an earlier typecast which | |
1936 | is irrelevant when performing tests against the given | |
1937 | constant. */ | |
1938 | if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR) | |
1939 | { | |
1940 | tree new_cond = find_equivalent_equality_comparison (cond); | |
1941 | ||
1942 | if (new_cond) | |
1943 | { | |
1944 | /* Update the statement to use the new equivalent | |
1945 | condition. */ | |
1946 | COND_EXPR_COND (stmt) = new_cond; | |
68b9f53b AM |
1947 | |
1948 | /* If this is not a real stmt, ann will be NULL and we | |
1949 | avoid processing the operands. */ | |
1950 | if (ann) | |
1951 | modify_stmt (stmt); | |
6de9cd9a DN |
1952 | |
1953 | /* Lookup the condition and return its known value if it | |
1954 | exists. */ | |
48732f23 | 1955 | new_cond = lookup_avail_expr (stmt, insert); |
6de9cd9a DN |
1956 | if (new_cond) |
1957 | return new_cond; | |
1958 | ||
1959 | /* The operands have changed, so update op0 and op1. */ | |
1960 | op0 = TREE_OPERAND (cond, 0); | |
1961 | op1 = TREE_OPERAND (cond, 1); | |
1962 | } | |
1963 | } | |
1964 | ||
1965 | /* Consult the value range records for this variable (if they exist) | |
1966 | to see if we can eliminate or simplify this conditional. | |
1967 | ||
1968 | Note two tests are necessary to determine no records exist. | |
1969 | First we have to see if the virtual array exists, if it | |
1970 | exists, then we have to check its active size. | |
1971 | ||
1972 | Also note the vast majority of conditionals are not testing | |
1973 | a variable which has had its range constrained by an earlier | |
1974 | conditional. So this filter avoids a lot of unnecessary work. */ | |
1975 | vrp_records = VARRAY_GENERIC_PTR (vrp_data, SSA_NAME_VERSION (op0)); | |
1976 | if (vrp_records == NULL) | |
1977 | return NULL; | |
1978 | ||
1979 | limit = VARRAY_ACTIVE_SIZE (vrp_records); | |
1980 | ||
1981 | /* If we have no value range records for this variable, or we are | |
1982 | unable to extract a range for this condition, then there is | |
1983 | nothing to do. */ | |
1984 | if (limit == 0 | |
1985 | || ! extract_range_from_cond (cond, &cond_high, | |
1986 | &cond_low, &cond_inverted)) | |
1987 | return NULL; | |
1988 | ||
1989 | /* We really want to avoid unnecessary computations of range | |
1990 | info. So all ranges are computed lazily; this avoids a | |
454ff5cb | 1991 | lot of unnecessary work. i.e., we record the conditional, |
6de9cd9a DN |
1992 | but do not process how it constrains the variable's |
1993 | potential values until we know that processing the condition | |
1994 | could be helpful. | |
1995 | ||
1996 | However, we do not want to have to walk a potentially long | |
1997 | list of ranges, nor do we want to compute a variable's | |
1998 | range more than once for a given path. | |
1999 | ||
2000 | Luckily, each time we encounter a conditional that can not | |
2001 | be otherwise optimized we will end up here and we will | |
2002 | compute the necessary range information for the variable | |
2003 | used in this condition. | |
2004 | ||
2005 | Thus you can conclude that there will never be more than one | |
2006 | conditional associated with a variable which has not been | |
2007 | processed. So we never need to merge more than one new | |
2008 | conditional into the current range. | |
2009 | ||
2010 | These properties also help us avoid unnecessary work. */ | |
2011 | element | |
2012 | = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records, limit - 1); | |
2013 | ||
2014 | if (element->high && element->low) | |
2015 | { | |
2016 | /* The last element has been processed, so there is no range | |
2017 | merging to do, we can simply use the high/low values | |
2018 | recorded in the last element. */ | |
2019 | low = element->low; | |
2020 | high = element->high; | |
2021 | } | |
2022 | else | |
2023 | { | |
2024 | tree tmp_high, tmp_low; | |
2025 | int dummy; | |
2026 | ||
2027 | /* The last element has not been processed. Process it now. */ | |
2028 | extract_range_from_cond (element->cond, &tmp_high, | |
2029 | &tmp_low, &dummy); | |
2030 | ||
2031 | /* If this is the only element, then no merging is necessary, | |
2032 | the high/low values from extract_range_from_cond are all | |
2033 | we need. */ | |
2034 | if (limit == 1) | |
2035 | { | |
2036 | low = tmp_low; | |
2037 | high = tmp_high; | |
2038 | } | |
2039 | else | |
2040 | { | |
2041 | /* Get the high/low value from the previous element. */ | |
2042 | struct vrp_element *prev | |
2043 | = (struct vrp_element *)VARRAY_GENERIC_PTR (vrp_records, | |
2044 | limit - 2); | |
2045 | low = prev->low; | |
2046 | high = prev->high; | |
2047 | ||
2048 | /* Merge in this element's range with the range from the | |
2049 | previous element. | |
2050 | ||
2051 | The low value for the merged range is the maximum of | |
2052 | the previous low value and the low value of this record. | |
2053 | ||
2054 | Similarly the high value for the merged range is the | |
2055 | minimum of the previous high value and the high value of | |
2056 | this record. */ | |
2057 | low = (tree_int_cst_compare (low, tmp_low) == 1 | |
2058 | ? low : tmp_low); | |
2059 | high = (tree_int_cst_compare (high, tmp_high) == -1 | |
2060 | ? high : tmp_high); | |
2061 | } | |
2062 | ||
2063 | /* And record the computed range. */ | |
2064 | element->low = low; | |
2065 | element->high = high; | |
2066 | ||
2067 | } | |
2068 | ||
2069 | /* After we have constrained this variable's potential values, | |
2070 | we try to determine the result of the given conditional. | |
2071 | ||
2072 | To simplify later tests, first determine if the current | |
2073 | low value is the same low value as the conditional. | |
2074 | Similarly for the current high value and the high value | |
2075 | for the conditional. */ | |
2076 | lowequal = tree_int_cst_equal (low, cond_low); | |
2077 | highequal = tree_int_cst_equal (high, cond_high); | |
2078 | ||
2079 | if (lowequal && highequal) | |
2080 | return (cond_inverted ? boolean_false_node : boolean_true_node); | |
2081 | ||
2082 | /* To simplify the overlap/subset tests below we may want | |
2083 | to swap the two ranges so that the larger of the two | |
2084 | ranges occurs "first". */ | |
2085 | swapped = 0; | |
2086 | if (tree_int_cst_compare (low, cond_low) == 1 | |
2087 | || (lowequal | |
2088 | && tree_int_cst_compare (cond_high, high) == 1)) | |
2089 | { | |
2090 | tree temp; | |
2091 | ||
2092 | swapped = 1; | |
2093 | temp = low; | |
2094 | low = cond_low; | |
2095 | cond_low = temp; | |
2096 | temp = high; | |
2097 | high = cond_high; | |
2098 | cond_high = temp; | |
2099 | } | |
2100 | ||
2101 | /* Now determine if there is no overlap in the ranges | |
2102 | or if the second range is a subset of the first range. */ | |
2103 | no_overlap = tree_int_cst_lt (high, cond_low); | |
2104 | subset = tree_int_cst_compare (cond_high, high) != 1; | |
2105 | ||
2106 | /* If there was no overlap in the ranges, then this conditional | |
2107 | always has a false value (unless we had to invert this | |
2108 | conditional, in which case it always has a true value). */ | |
2109 | if (no_overlap) | |
2110 | return (cond_inverted ? boolean_true_node : boolean_false_node); | |
2111 | ||
2112 | /* If the current range is a subset of the condition's range, | |
2113 | then this conditional always has a true value (unless we | |
2114 | had to invert this conditional, in which case it always | |
2115 | has a true value). */ | |
2116 | if (subset && swapped) | |
2117 | return (cond_inverted ? boolean_false_node : boolean_true_node); | |
2118 | ||
2119 | /* We were unable to determine the result of the conditional. | |
2120 | However, we may be able to simplify the conditional. First | |
2121 | merge the ranges in the same manner as range merging above. */ | |
2122 | low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low; | |
2123 | high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high; | |
2124 | ||
2125 | /* If the range has converged to a single point, then turn this | |
2126 | into an equality comparison. */ | |
2127 | if (TREE_CODE (cond) != EQ_EXPR | |
2128 | && TREE_CODE (cond) != NE_EXPR | |
2129 | && tree_int_cst_equal (low, high)) | |
2130 | { | |
2131 | TREE_SET_CODE (cond, EQ_EXPR); | |
2132 | TREE_OPERAND (cond, 1) = high; | |
2133 | } | |
2134 | } | |
2135 | } | |
2136 | return 0; | |
2137 | } | |
2138 | ||
2139 | /* STMT is a SWITCH_EXPR for which we could not trivially determine its | |
2140 | result. This routine attempts to find equivalent forms of the | |
2141 | condition which we may be able to optimize better. */ | |
2142 | ||
2143 | static tree | |
48732f23 | 2144 | simplify_switch_and_lookup_avail_expr (tree stmt, int insert) |
6de9cd9a DN |
2145 | { |
2146 | tree cond = SWITCH_COND (stmt); | |
2147 | tree def, to, ti; | |
2148 | ||
2149 | /* The optimization that we really care about is removing unnecessary | |
2150 | casts. That will let us do much better in propagating the inferred | |
2151 | constant at the switch target. */ | |
2152 | if (TREE_CODE (cond) == SSA_NAME) | |
2153 | { | |
2154 | def = SSA_NAME_DEF_STMT (cond); | |
2155 | if (TREE_CODE (def) == MODIFY_EXPR) | |
2156 | { | |
2157 | def = TREE_OPERAND (def, 1); | |
2158 | if (TREE_CODE (def) == NOP_EXPR) | |
2159 | { | |
d969ee71 RH |
2160 | int need_precision; |
2161 | bool fail; | |
2162 | ||
6de9cd9a | 2163 | def = TREE_OPERAND (def, 0); |
d969ee71 RH |
2164 | |
2165 | #ifdef ENABLE_CHECKING | |
2166 | /* ??? Why was Jeff testing this? We are gimple... */ | |
1e128c5f | 2167 | gcc_assert (is_gimple_val (def)); |
d969ee71 RH |
2168 | #endif |
2169 | ||
6de9cd9a DN |
2170 | to = TREE_TYPE (cond); |
2171 | ti = TREE_TYPE (def); | |
2172 | ||
d969ee71 | 2173 | /* If we have an extension that preserves value, then we |
6de9cd9a | 2174 | can copy the source value into the switch. */ |
d969ee71 RH |
2175 | |
2176 | need_precision = TYPE_PRECISION (ti); | |
2177 | fail = false; | |
2178 | if (TYPE_UNSIGNED (to) && !TYPE_UNSIGNED (ti)) | |
2179 | fail = true; | |
2180 | else if (!TYPE_UNSIGNED (to) && TYPE_UNSIGNED (ti)) | |
2181 | need_precision += 1; | |
2182 | if (TYPE_PRECISION (to) < need_precision) | |
2183 | fail = true; | |
2184 | ||
2185 | if (!fail) | |
6de9cd9a DN |
2186 | { |
2187 | SWITCH_COND (stmt) = def; | |
68b9f53b | 2188 | modify_stmt (stmt); |
6de9cd9a | 2189 | |
48732f23 | 2190 | return lookup_avail_expr (stmt, insert); |
6de9cd9a DN |
2191 | } |
2192 | } | |
2193 | } | |
2194 | } | |
2195 | ||
2196 | return 0; | |
2197 | } | |
2198 | ||
ff2ad0f7 DN |
2199 | |
2200 | /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current | |
2201 | known value for that SSA_NAME (or NULL if no value is known). | |
2202 | ||
2203 | NONZERO_VARS is the set SSA_NAMES known to have a nonzero value, | |
2204 | even if we don't know their precise value. | |
2205 | ||
2206 | Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI | |
2207 | nodes of the successors of BB. */ | |
2208 | ||
2209 | static void | |
2210 | cprop_into_successor_phis (basic_block bb, | |
2211 | varray_type const_and_copies, | |
2212 | bitmap nonzero_vars) | |
2213 | { | |
2214 | edge e; | |
2215 | ||
2216 | /* This can get rather expensive if the implementation is naive in | |
2217 | how it finds the phi alternative associated with a particular edge. */ | |
2218 | for (e = bb->succ; e; e = e->succ_next) | |
2219 | { | |
2220 | tree phi; | |
2221 | int phi_num_args; | |
2222 | int hint; | |
2223 | ||
2224 | /* If this is an abnormal edge, then we do not want to copy propagate | |
2225 | into the PHI alternative associated with this edge. */ | |
2226 | if (e->flags & EDGE_ABNORMAL) | |
2227 | continue; | |
2228 | ||
2229 | phi = phi_nodes (e->dest); | |
2230 | if (! phi) | |
2231 | continue; | |
2232 | ||
2233 | /* There is no guarantee that for any two PHI nodes in a block that | |
2234 | the phi alternative associated with a particular edge will be | |
2235 | at the same index in the phi alternative array. | |
2236 | ||
2237 | However, it is very likely they will be the same. So we keep | |
2238 | track of the index of the alternative where we found the edge in | |
2239 | the previous phi node and check that index first in the next | |
2240 | phi node. If that hint fails, then we actually search all | |
2241 | the entries. */ | |
2242 | phi_num_args = PHI_NUM_ARGS (phi); | |
2243 | hint = phi_num_args; | |
2244 | for ( ; phi; phi = PHI_CHAIN (phi)) | |
2245 | { | |
2246 | int i; | |
2247 | tree new; | |
2248 | use_operand_p orig_p; | |
2249 | tree orig; | |
2250 | ||
2251 | /* If the hint is valid (!= phi_num_args), see if it points | |
2252 | us to the desired phi alternative. */ | |
2253 | if (hint != phi_num_args && PHI_ARG_EDGE (phi, hint) == e) | |
2254 | ; | |
2255 | else | |
2256 | { | |
2257 | /* The hint was either invalid or did not point to the | |
2258 | correct phi alternative. Search all the alternatives | |
2259 | for the correct one. Update the hint. */ | |
2260 | for (i = 0; i < phi_num_args; i++) | |
2261 | if (PHI_ARG_EDGE (phi, i) == e) | |
2262 | break; | |
2263 | hint = i; | |
2264 | } | |
2265 | ||
ff2ad0f7 DN |
2266 | /* If we did not find the proper alternative, then something is |
2267 | horribly wrong. */ | |
1e128c5f | 2268 | gcc_assert (hint != phi_num_args); |
ff2ad0f7 DN |
2269 | |
2270 | /* The alternative may be associated with a constant, so verify | |
2271 | it is an SSA_NAME before doing anything with it. */ | |
2272 | orig_p = PHI_ARG_DEF_PTR (phi, hint); | |
2273 | orig = USE_FROM_PTR (orig_p); | |
2274 | if (TREE_CODE (orig) != SSA_NAME) | |
2275 | continue; | |
2276 | ||
2277 | /* If the alternative is known to have a nonzero value, record | |
2278 | that fact in the PHI node itself for future use. */ | |
2279 | if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig))) | |
2280 | PHI_ARG_NONZERO (phi, hint) = true; | |
2281 | ||
2282 | /* If we have *ORIG_P in our constant/copy table, then replace | |
2283 | ORIG_P with its value in our constant/copy table. */ | |
2284 | new = VARRAY_TREE (const_and_copies, SSA_NAME_VERSION (orig)); | |
2285 | if (new | |
2286 | && (TREE_CODE (new) == SSA_NAME | |
2287 | || is_gimple_min_invariant (new)) | |
2288 | && may_propagate_copy (orig, new)) | |
2289 | { | |
2290 | propagate_value (orig_p, new); | |
2291 | } | |
2292 | } | |
2293 | } | |
2294 | } | |
2295 | ||
2296 | ||
6de9cd9a DN |
2297 | /* Propagate known constants/copies into PHI nodes of BB's successor |
2298 | blocks. */ | |
2299 | ||
2300 | static void | |
2301 | cprop_into_phis (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED, | |
2302 | basic_block bb) | |
2303 | { | |
dd747311 | 2304 | cprop_into_successor_phis (bb, const_and_copies, nonzero_vars); |
6de9cd9a DN |
2305 | } |
2306 | ||
2307 | /* Search for redundant computations in STMT. If any are found, then | |
2308 | replace them with the variable holding the result of the computation. | |
2309 | ||
2310 | If safe, record this expression into the available expression hash | |
2311 | table. */ | |
2312 | ||
2313 | static bool | |
2314 | eliminate_redundant_computations (struct dom_walk_data *walk_data, | |
2315 | tree stmt, stmt_ann_t ann) | |
2316 | { | |
a32b97a2 | 2317 | v_may_def_optype v_may_defs = V_MAY_DEF_OPS (ann); |
6de9cd9a DN |
2318 | tree *expr_p, def = NULL_TREE; |
2319 | bool insert = true; | |
2320 | tree cached_lhs; | |
2321 | bool retval = false; | |
6de9cd9a DN |
2322 | |
2323 | if (TREE_CODE (stmt) == MODIFY_EXPR) | |
2324 | def = TREE_OPERAND (stmt, 0); | |
2325 | ||
2326 | /* Certain expressions on the RHS can be optimized away, but can not | |
2327 | themselves be entered into the hash tables. */ | |
2328 | if (ann->makes_aliased_stores | |
2329 | || ! def | |
2330 | || TREE_CODE (def) != SSA_NAME | |
2331 | || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def) | |
a32b97a2 | 2332 | || NUM_V_MAY_DEFS (v_may_defs) != 0) |
6de9cd9a DN |
2333 | insert = false; |
2334 | ||
2335 | /* Check if the expression has been computed before. */ | |
48732f23 | 2336 | cached_lhs = lookup_avail_expr (stmt, insert); |
6de9cd9a DN |
2337 | |
2338 | /* If this is an assignment and the RHS was not in the hash table, | |
2339 | then try to simplify the RHS and lookup the new RHS in the | |
2340 | hash table. */ | |
2341 | if (! cached_lhs && TREE_CODE (stmt) == MODIFY_EXPR) | |
48732f23 | 2342 | cached_lhs = simplify_rhs_and_lookup_avail_expr (walk_data, stmt, insert); |
6de9cd9a DN |
2343 | /* Similarly if this is a COND_EXPR and we did not find its |
2344 | expression in the hash table, simplify the condition and | |
2345 | try again. */ | |
2346 | else if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR) | |
48732f23 | 2347 | cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert); |
6de9cd9a DN |
2348 | /* Similarly for a SWITCH_EXPR. */ |
2349 | else if (!cached_lhs && TREE_CODE (stmt) == SWITCH_EXPR) | |
48732f23 | 2350 | cached_lhs = simplify_switch_and_lookup_avail_expr (stmt, insert); |
6de9cd9a DN |
2351 | |
2352 | opt_stats.num_exprs_considered++; | |
2353 | ||
2354 | /* Get a pointer to the expression we are trying to optimize. */ | |
2355 | if (TREE_CODE (stmt) == COND_EXPR) | |
2356 | expr_p = &COND_EXPR_COND (stmt); | |
2357 | else if (TREE_CODE (stmt) == SWITCH_EXPR) | |
2358 | expr_p = &SWITCH_COND (stmt); | |
2359 | else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0)) | |
2360 | expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1); | |
2361 | else | |
2362 | expr_p = &TREE_OPERAND (stmt, 1); | |
2363 | ||
2364 | /* It is safe to ignore types here since we have already done | |
2365 | type checking in the hashing and equality routines. In fact | |
2366 | type checking here merely gets in the way of constant | |
2367 | propagation. Also, make sure that it is safe to propagate | |
2368 | CACHED_LHS into *EXPR_P. */ | |
2369 | if (cached_lhs | |
2370 | && (TREE_CODE (cached_lhs) != SSA_NAME | |
ff2ad0f7 | 2371 | || may_propagate_copy (*expr_p, cached_lhs))) |
6de9cd9a DN |
2372 | { |
2373 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2374 | { | |
2375 | fprintf (dump_file, " Replaced redundant expr '"); | |
2376 | print_generic_expr (dump_file, *expr_p, dump_flags); | |
2377 | fprintf (dump_file, "' with '"); | |
2378 | print_generic_expr (dump_file, cached_lhs, dump_flags); | |
2379 | fprintf (dump_file, "'\n"); | |
2380 | } | |
2381 | ||
2382 | opt_stats.num_re++; | |
2383 | ||
2384 | #if defined ENABLE_CHECKING | |
1e128c5f GB |
2385 | gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME |
2386 | || is_gimple_min_invariant (cached_lhs)); | |
6de9cd9a DN |
2387 | #endif |
2388 | ||
2389 | if (TREE_CODE (cached_lhs) == ADDR_EXPR | |
2390 | || (POINTER_TYPE_P (TREE_TYPE (*expr_p)) | |
2391 | && is_gimple_min_invariant (cached_lhs))) | |
2392 | retval = true; | |
2393 | ||
d00ad49b | 2394 | propagate_tree_value (expr_p, cached_lhs); |
68b9f53b | 2395 | modify_stmt (stmt); |
6de9cd9a DN |
2396 | } |
2397 | return retval; | |
2398 | } | |
2399 | ||
2400 | /* STMT, a MODIFY_EXPR, may create certain equivalences, in either | |
2401 | the available expressions table or the const_and_copies table. | |
2402 | Detect and record those equivalences. */ | |
2403 | ||
2404 | static void | |
2405 | record_equivalences_from_stmt (tree stmt, | |
6de9cd9a DN |
2406 | varray_type *block_nonzero_vars_p, |
2407 | int may_optimize_p, | |
2408 | stmt_ann_t ann) | |
2409 | { | |
2410 | tree lhs = TREE_OPERAND (stmt, 0); | |
2411 | enum tree_code lhs_code = TREE_CODE (lhs); | |
2412 | int i; | |
2413 | ||
2414 | if (lhs_code == SSA_NAME) | |
2415 | { | |
2416 | tree rhs = TREE_OPERAND (stmt, 1); | |
2417 | ||
2418 | /* Strip away any useless type conversions. */ | |
2419 | STRIP_USELESS_TYPE_CONVERSION (rhs); | |
2420 | ||
2421 | /* If the RHS of the assignment is a constant or another variable that | |
2422 | may be propagated, register it in the CONST_AND_COPIES table. We | |
2423 | do not need to record unwind data for this, since this is a true | |
1ea7e6ad | 2424 | assignment and not an equivalence inferred from a comparison. All |
6de9cd9a DN |
2425 | uses of this ssa name are dominated by this assignment, so unwinding |
2426 | just costs time and space. */ | |
2427 | if (may_optimize_p | |
2428 | && (TREE_CODE (rhs) == SSA_NAME | |
2429 | || is_gimple_min_invariant (rhs))) | |
2430 | set_value_for (lhs, rhs, const_and_copies); | |
2431 | ||
2432 | /* alloca never returns zero and the address of a non-weak symbol | |
2433 | is never zero. NOP_EXPRs and CONVERT_EXPRs can be completely | |
2434 | stripped as they do not affect this equivalence. */ | |
2435 | while (TREE_CODE (rhs) == NOP_EXPR | |
2436 | || TREE_CODE (rhs) == CONVERT_EXPR) | |
2437 | rhs = TREE_OPERAND (rhs, 0); | |
2438 | ||
2439 | if (alloca_call_p (rhs) | |
2440 | || (TREE_CODE (rhs) == ADDR_EXPR | |
2441 | && DECL_P (TREE_OPERAND (rhs, 0)) | |
2442 | && ! DECL_WEAK (TREE_OPERAND (rhs, 0)))) | |
2443 | record_var_is_nonzero (lhs, block_nonzero_vars_p); | |
2444 | ||
2445 | /* IOR of any value with a nonzero value will result in a nonzero | |
2446 | value. Even if we do not know the exact result recording that | |
2447 | the result is nonzero is worth the effort. */ | |
2448 | if (TREE_CODE (rhs) == BIT_IOR_EXPR | |
2449 | && integer_nonzerop (TREE_OPERAND (rhs, 1))) | |
2450 | record_var_is_nonzero (lhs, block_nonzero_vars_p); | |
2451 | } | |
2452 | ||
2453 | /* Look at both sides for pointer dereferences. If we find one, then | |
2454 | the pointer must be nonnull and we can enter that equivalence into | |
2455 | the hash tables. */ | |
dd747311 JL |
2456 | if (flag_delete_null_pointer_checks) |
2457 | for (i = 0; i < 2; i++) | |
2458 | { | |
2459 | tree t = TREE_OPERAND (stmt, i); | |
2460 | ||
2461 | /* Strip away any COMPONENT_REFs. */ | |
2462 | while (TREE_CODE (t) == COMPONENT_REF) | |
2463 | t = TREE_OPERAND (t, 0); | |
2464 | ||
2465 | /* Now see if this is a pointer dereference. */ | |
2466 | if (TREE_CODE (t) == INDIRECT_REF) | |
2467 | { | |
2468 | tree op = TREE_OPERAND (t, 0); | |
2469 | ||
2470 | /* If the pointer is a SSA variable, then enter new | |
2471 | equivalences into the hash table. */ | |
2472 | while (TREE_CODE (op) == SSA_NAME) | |
2473 | { | |
2474 | tree def = SSA_NAME_DEF_STMT (op); | |
2475 | ||
2476 | record_var_is_nonzero (op, block_nonzero_vars_p); | |
2477 | ||
2478 | /* And walk up the USE-DEF chains noting other SSA_NAMEs | |
2479 | which are known to have a nonzero value. */ | |
2480 | if (def | |
2481 | && TREE_CODE (def) == MODIFY_EXPR | |
2482 | && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR) | |
2483 | op = TREE_OPERAND (TREE_OPERAND (def, 1), 0); | |
2484 | else | |
2485 | break; | |
2486 | } | |
2487 | } | |
2488 | } | |
6de9cd9a DN |
2489 | |
2490 | /* A memory store, even an aliased store, creates a useful | |
2491 | equivalence. By exchanging the LHS and RHS, creating suitable | |
2492 | vops and recording the result in the available expression table, | |
2493 | we may be able to expose more redundant loads. */ | |
2494 | if (!ann->has_volatile_ops | |
2495 | && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME | |
2496 | || is_gimple_min_invariant (TREE_OPERAND (stmt, 1))) | |
2497 | && !is_gimple_reg (lhs)) | |
2498 | { | |
2499 | tree rhs = TREE_OPERAND (stmt, 1); | |
2500 | tree new; | |
6de9cd9a DN |
2501 | |
2502 | /* FIXME: If the LHS of the assignment is a bitfield and the RHS | |
2503 | is a constant, we need to adjust the constant to fit into the | |
2504 | type of the LHS. If the LHS is a bitfield and the RHS is not | |
2505 | a constant, then we can not record any equivalences for this | |
2506 | statement since we would need to represent the widening or | |
2507 | narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c | |
2508 | and should not be necessary if GCC represented bitfields | |
2509 | properly. */ | |
2510 | if (lhs_code == COMPONENT_REF | |
2511 | && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1))) | |
2512 | { | |
2513 | if (TREE_CONSTANT (rhs)) | |
2514 | rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs); | |
2515 | else | |
2516 | rhs = NULL; | |
2517 | ||
2518 | /* If the value overflowed, then we can not use this equivalence. */ | |
2519 | if (rhs && ! is_gimple_min_invariant (rhs)) | |
2520 | rhs = NULL; | |
2521 | } | |
2522 | ||
2523 | if (rhs) | |
2524 | { | |
6de9cd9a DN |
2525 | /* Build a new statement with the RHS and LHS exchanged. */ |
2526 | new = build (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs); | |
2527 | ||
1a24f92f | 2528 | create_ssa_artficial_load_stmt (&(ann->operands), new); |
6de9cd9a DN |
2529 | |
2530 | /* Finally enter the statement into the available expression | |
2531 | table. */ | |
48732f23 | 2532 | lookup_avail_expr (new, true); |
6de9cd9a DN |
2533 | } |
2534 | } | |
2535 | } | |
2536 | ||
ff2ad0f7 DN |
2537 | /* Replace *OP_P in STMT with any known equivalent value for *OP_P from |
2538 | CONST_AND_COPIES. */ | |
2539 | ||
2540 | static bool | |
68b9f53b | 2541 | cprop_operand (tree stmt, use_operand_p op_p, varray_type const_and_copies) |
ff2ad0f7 DN |
2542 | { |
2543 | bool may_have_exposed_new_symbols = false; | |
2544 | tree val; | |
2545 | tree op = USE_FROM_PTR (op_p); | |
2546 | ||
2547 | /* If the operand has a known constant value or it is known to be a | |
2548 | copy of some other variable, use the value or copy stored in | |
2549 | CONST_AND_COPIES. */ | |
2550 | val = VARRAY_TREE (const_and_copies, SSA_NAME_VERSION (op)); | |
2551 | if (val) | |
2552 | { | |
2553 | tree op_type, val_type; | |
2554 | ||
2555 | /* Do not change the base variable in the virtual operand | |
2556 | tables. That would make it impossible to reconstruct | |
2557 | the renamed virtual operand if we later modify this | |
2558 | statement. Also only allow the new value to be an SSA_NAME | |
2559 | for propagation into virtual operands. */ | |
2560 | if (!is_gimple_reg (op) | |
2561 | && (get_virtual_var (val) != get_virtual_var (op) | |
2562 | || TREE_CODE (val) != SSA_NAME)) | |
2563 | return false; | |
2564 | ||
2565 | /* Get the toplevel type of each operand. */ | |
2566 | op_type = TREE_TYPE (op); | |
2567 | val_type = TREE_TYPE (val); | |
2568 | ||
2569 | /* While both types are pointers, get the type of the object | |
2570 | pointed to. */ | |
2571 | while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type)) | |
2572 | { | |
2573 | op_type = TREE_TYPE (op_type); | |
2574 | val_type = TREE_TYPE (val_type); | |
2575 | } | |
2576 | ||
63b88252 RH |
2577 | /* Make sure underlying types match before propagating a constant by |
2578 | converting the constant to the proper type. Note that convert may | |
2579 | return a non-gimple expression, in which case we ignore this | |
2580 | propagation opportunity. */ | |
2581 | if (TREE_CODE (val) != SSA_NAME) | |
ff2ad0f7 | 2582 | { |
63b88252 RH |
2583 | if (!lang_hooks.types_compatible_p (op_type, val_type)) |
2584 | { | |
2585 | val = fold_convert (TREE_TYPE (op), val); | |
2586 | if (!is_gimple_min_invariant (val)) | |
2587 | return false; | |
2588 | } | |
ff2ad0f7 DN |
2589 | } |
2590 | ||
2591 | /* Certain operands are not allowed to be copy propagated due | |
2592 | to their interaction with exception handling and some GCC | |
2593 | extensions. */ | |
63b88252 | 2594 | else if (!may_propagate_copy (op, val)) |
ff2ad0f7 DN |
2595 | return false; |
2596 | ||
2597 | /* Dump details. */ | |
2598 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2599 | { | |
2600 | fprintf (dump_file, " Replaced '"); | |
2601 | print_generic_expr (dump_file, op, dump_flags); | |
2602 | fprintf (dump_file, "' with %s '", | |
2603 | (TREE_CODE (val) != SSA_NAME ? "constant" : "variable")); | |
2604 | print_generic_expr (dump_file, val, dump_flags); | |
2605 | fprintf (dump_file, "'\n"); | |
2606 | } | |
2607 | ||
2608 | /* If VAL is an ADDR_EXPR or a constant of pointer type, note | |
2609 | that we may have exposed a new symbol for SSA renaming. */ | |
2610 | if (TREE_CODE (val) == ADDR_EXPR | |
2611 | || (POINTER_TYPE_P (TREE_TYPE (op)) | |
2612 | && is_gimple_min_invariant (val))) | |
2613 | may_have_exposed_new_symbols = true; | |
2614 | ||
2615 | propagate_value (op_p, val); | |
2616 | ||
2617 | /* And note that we modified this statement. This is now | |
2618 | safe, even if we changed virtual operands since we will | |
2619 | rescan the statement and rewrite its operands again. */ | |
68b9f53b | 2620 | modify_stmt (stmt); |
ff2ad0f7 DN |
2621 | } |
2622 | return may_have_exposed_new_symbols; | |
2623 | } | |
2624 | ||
2625 | /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current | |
2626 | known value for that SSA_NAME (or NULL if no value is known). | |
2627 | ||
2628 | Propagate values from CONST_AND_COPIES into the uses, vuses and | |
2629 | v_may_def_ops of STMT. */ | |
2630 | ||
2631 | static bool | |
2632 | cprop_into_stmt (tree stmt, varray_type const_and_copies) | |
2633 | { | |
2634 | bool may_have_exposed_new_symbols = false; | |
4c124b4c AM |
2635 | use_operand_p op_p; |
2636 | ssa_op_iter iter; | |
c7f90219 | 2637 | tree rhs; |
ff2ad0f7 | 2638 | |
4c124b4c | 2639 | FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES) |
ff2ad0f7 | 2640 | { |
ff2ad0f7 DN |
2641 | if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME) |
2642 | may_have_exposed_new_symbols | |
68b9f53b | 2643 | |= cprop_operand (stmt, op_p, const_and_copies); |
ff2ad0f7 DN |
2644 | } |
2645 | ||
c7f90219 SB |
2646 | if (may_have_exposed_new_symbols) |
2647 | { | |
2648 | rhs = get_rhs (stmt); | |
2649 | if (rhs && TREE_CODE (rhs) == ADDR_EXPR) | |
2650 | recompute_tree_invarant_for_addr_expr (rhs); | |
2651 | } | |
2652 | ||
ff2ad0f7 DN |
2653 | return may_have_exposed_new_symbols; |
2654 | } | |
2655 | ||
2656 | ||
6de9cd9a DN |
2657 | /* Optimize the statement pointed by iterator SI. |
2658 | ||
2659 | We try to perform some simplistic global redundancy elimination and | |
2660 | constant propagation: | |
2661 | ||
2662 | 1- To detect global redundancy, we keep track of expressions that have | |
2663 | been computed in this block and its dominators. If we find that the | |
2664 | same expression is computed more than once, we eliminate repeated | |
2665 | computations by using the target of the first one. | |
2666 | ||
2667 | 2- Constant values and copy assignments. This is used to do very | |
2668 | simplistic constant and copy propagation. When a constant or copy | |
2669 | assignment is found, we map the value on the RHS of the assignment to | |
2670 | the variable in the LHS in the CONST_AND_COPIES table. */ | |
2671 | ||
2672 | static void | |
1eaba2f2 | 2673 | optimize_stmt (struct dom_walk_data *walk_data, basic_block bb, |
6de9cd9a DN |
2674 | block_stmt_iterator si) |
2675 | { | |
2676 | stmt_ann_t ann; | |
2677 | tree stmt; | |
6de9cd9a DN |
2678 | bool may_optimize_p; |
2679 | bool may_have_exposed_new_symbols = false; | |
2680 | struct dom_walk_block_data *bd | |
2681 | = VARRAY_TOP_GENERIC_PTR (walk_data->block_data_stack); | |
2682 | ||
2683 | stmt = bsi_stmt (si); | |
2684 | ||
2685 | get_stmt_operands (stmt); | |
2686 | ann = stmt_ann (stmt); | |
6de9cd9a DN |
2687 | opt_stats.num_stmts++; |
2688 | may_have_exposed_new_symbols = false; | |
2689 | ||
2690 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2691 | { | |
2692 | fprintf (dump_file, "Optimizing statement "); | |
2693 | print_generic_stmt (dump_file, stmt, TDF_SLIM); | |
2694 | } | |
2695 | ||
a32b97a2 | 2696 | /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */ |
6de9cd9a DN |
2697 | may_have_exposed_new_symbols = cprop_into_stmt (stmt, const_and_copies); |
2698 | ||
2699 | /* If the statement has been modified with constant replacements, | |
2700 | fold its RHS before checking for redundant computations. */ | |
2701 | if (ann->modified) | |
2702 | { | |
2703 | /* Try to fold the statement making sure that STMT is kept | |
2704 | up to date. */ | |
2705 | if (fold_stmt (bsi_stmt_ptr (si))) | |
2706 | { | |
2707 | stmt = bsi_stmt (si); | |
2708 | ann = stmt_ann (stmt); | |
2709 | ||
2710 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2711 | { | |
2712 | fprintf (dump_file, " Folded to: "); | |
2713 | print_generic_stmt (dump_file, stmt, TDF_SLIM); | |
2714 | } | |
2715 | } | |
2716 | ||
2717 | /* Constant/copy propagation above may change the set of | |
2718 | virtual operands associated with this statement. Folding | |
2719 | may remove the need for some virtual operands. | |
2720 | ||
2721 | Indicate we will need to rescan and rewrite the statement. */ | |
2722 | may_have_exposed_new_symbols = true; | |
2723 | } | |
2724 | ||
2725 | /* Check for redundant computations. Do this optimization only | |
2726 | for assignments that have no volatile ops and conditionals. */ | |
2727 | may_optimize_p = (!ann->has_volatile_ops | |
2728 | && ((TREE_CODE (stmt) == RETURN_EXPR | |
2729 | && TREE_OPERAND (stmt, 0) | |
2730 | && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR | |
2731 | && ! (TREE_SIDE_EFFECTS | |
2732 | (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1)))) | |
2733 | || (TREE_CODE (stmt) == MODIFY_EXPR | |
2734 | && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1))) | |
2735 | || TREE_CODE (stmt) == COND_EXPR | |
2736 | || TREE_CODE (stmt) == SWITCH_EXPR)); | |
2737 | ||
2738 | if (may_optimize_p) | |
2739 | may_have_exposed_new_symbols | |
2740 | |= eliminate_redundant_computations (walk_data, stmt, ann); | |
2741 | ||
2742 | /* Record any additional equivalences created by this statement. */ | |
2743 | if (TREE_CODE (stmt) == MODIFY_EXPR) | |
2744 | record_equivalences_from_stmt (stmt, | |
6de9cd9a DN |
2745 | &bd->nonzero_vars, |
2746 | may_optimize_p, | |
2747 | ann); | |
2748 | ||
9fae925b | 2749 | register_definitions_for_stmt (stmt); |
6de9cd9a DN |
2750 | |
2751 | /* If STMT is a COND_EXPR and it was modified, then we may know | |
2752 | where it goes. If that is the case, then mark the CFG as altered. | |
2753 | ||
2754 | This will cause us to later call remove_unreachable_blocks and | |
2755 | cleanup_tree_cfg when it is safe to do so. It is not safe to | |
2756 | clean things up here since removal of edges and such can trigger | |
2757 | the removal of PHI nodes, which in turn can release SSA_NAMEs to | |
2758 | the manager. | |
2759 | ||
2760 | That's all fine and good, except that once SSA_NAMEs are released | |
2761 | to the manager, we must not call create_ssa_name until all references | |
2762 | to released SSA_NAMEs have been eliminated. | |
2763 | ||
2764 | All references to the deleted SSA_NAMEs can not be eliminated until | |
2765 | we remove unreachable blocks. | |
2766 | ||
2767 | We can not remove unreachable blocks until after we have completed | |
2768 | any queued jump threading. | |
2769 | ||
2770 | We can not complete any queued jump threads until we have taken | |
2771 | appropriate variables out of SSA form. Taking variables out of | |
2772 | SSA form can call create_ssa_name and thus we lose. | |
2773 | ||
2774 | Ultimately I suspect we're going to need to change the interface | |
2775 | into the SSA_NAME manager. */ | |
2776 | ||
2777 | if (ann->modified) | |
2778 | { | |
2779 | tree val = NULL; | |
2780 | ||
2781 | if (TREE_CODE (stmt) == COND_EXPR) | |
2782 | val = COND_EXPR_COND (stmt); | |
2783 | else if (TREE_CODE (stmt) == SWITCH_EXPR) | |
2784 | val = SWITCH_COND (stmt); | |
2785 | ||
1eaba2f2 | 2786 | if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val)) |
6de9cd9a | 2787 | cfg_altered = true; |
1eaba2f2 RH |
2788 | |
2789 | /* If we simplified a statement in such a way as to be shown that it | |
2790 | cannot trap, update the eh information and the cfg to match. */ | |
2791 | if (maybe_clean_eh_stmt (stmt)) | |
2792 | { | |
2793 | bitmap_set_bit (need_eh_cleanup, bb->index); | |
2794 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2795 | fprintf (dump_file, " Flagged to clear EH edges.\n"); | |
2796 | } | |
6de9cd9a | 2797 | } |
1eaba2f2 | 2798 | |
6de9cd9a | 2799 | if (may_have_exposed_new_symbols) |
a6e1aa26 | 2800 | VARRAY_PUSH_TREE (stmts_to_rescan, bsi_stmt (si)); |
6de9cd9a DN |
2801 | } |
2802 | ||
2803 | /* Replace the RHS of STMT with NEW_RHS. If RHS can be found in the | |
2804 | available expression hashtable, then return the LHS from the hash | |
2805 | table. | |
2806 | ||
2807 | If INSERT is true, then we also update the available expression | |
2808 | hash table to account for the changes made to STMT. */ | |
2809 | ||
2810 | static tree | |
48732f23 | 2811 | update_rhs_and_lookup_avail_expr (tree stmt, tree new_rhs, bool insert) |
6de9cd9a DN |
2812 | { |
2813 | tree cached_lhs = NULL; | |
2814 | ||
2815 | /* Remove the old entry from the hash table. */ | |
2816 | if (insert) | |
2817 | { | |
2818 | struct expr_hash_elt element; | |
2819 | ||
2820 | initialize_hash_element (stmt, NULL, &element); | |
2821 | htab_remove_elt_with_hash (avail_exprs, &element, element.hash); | |
2822 | } | |
2823 | ||
2824 | /* Now update the RHS of the assignment. */ | |
2825 | TREE_OPERAND (stmt, 1) = new_rhs; | |
2826 | ||
2827 | /* Now lookup the updated statement in the hash table. */ | |
48732f23 | 2828 | cached_lhs = lookup_avail_expr (stmt, insert); |
6de9cd9a DN |
2829 | |
2830 | /* We have now called lookup_avail_expr twice with two different | |
2831 | versions of this same statement, once in optimize_stmt, once here. | |
2832 | ||
2833 | We know the call in optimize_stmt did not find an existing entry | |
2834 | in the hash table, so a new entry was created. At the same time | |
2835 | this statement was pushed onto the BLOCK_AVAIL_EXPRS varray. | |
2836 | ||
2837 | If this call failed to find an existing entry on the hash table, | |
2838 | then the new version of this statement was entered into the | |
2839 | hash table. And this statement was pushed onto BLOCK_AVAIL_EXPR | |
2840 | for the second time. So there are two copies on BLOCK_AVAIL_EXPRs | |
2841 | ||
2842 | If this call succeeded, we still have one copy of this statement | |
2843 | on the BLOCK_AVAIL_EXPRs varray. | |
2844 | ||
2845 | For both cases, we need to pop the most recent entry off the | |
2846 | BLOCK_AVAIL_EXPRs varray. For the case where we never found this | |
2847 | statement in the hash tables, that will leave precisely one | |
2848 | copy of this statement on BLOCK_AVAIL_EXPRs. For the case where | |
2849 | we found a copy of this statement in the second hash table lookup | |
2850 | we want _no_ copies of this statement in BLOCK_AVAIL_EXPRs. */ | |
2851 | if (insert) | |
48732f23 | 2852 | VARRAY_POP (avail_exprs_stack); |
6de9cd9a DN |
2853 | |
2854 | /* And make sure we record the fact that we modified this | |
2855 | statement. */ | |
68b9f53b | 2856 | modify_stmt (stmt); |
6de9cd9a DN |
2857 | |
2858 | return cached_lhs; | |
2859 | } | |
2860 | ||
2861 | /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If | |
2862 | found, return its LHS. Otherwise insert STMT in the table and return | |
2863 | NULL_TREE. | |
2864 | ||
2865 | Also, when an expression is first inserted in the AVAIL_EXPRS table, it | |
2866 | is also added to the stack pointed by BLOCK_AVAIL_EXPRS_P, so that they | |
2867 | can be removed when we finish processing this block and its children. | |
2868 | ||
2869 | NOTE: This function assumes that STMT is a MODIFY_EXPR node that | |
2870 | contains no CALL_EXPR on its RHS and makes no volatile nor | |
2871 | aliased references. */ | |
2872 | ||
2873 | static tree | |
48732f23 | 2874 | lookup_avail_expr (tree stmt, bool insert) |
6de9cd9a DN |
2875 | { |
2876 | void **slot; | |
2877 | tree lhs; | |
2878 | tree temp; | |
2879 | struct expr_hash_elt *element = xcalloc (sizeof (struct expr_hash_elt), 1); | |
2880 | ||
2881 | lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL; | |
2882 | ||
2883 | initialize_hash_element (stmt, lhs, element); | |
2884 | ||
2885 | /* Don't bother remembering constant assignments and copy operations. | |
2886 | Constants and copy operations are handled by the constant/copy propagator | |
2887 | in optimize_stmt. */ | |
2888 | if (TREE_CODE (element->rhs) == SSA_NAME | |
2889 | || is_gimple_min_invariant (element->rhs)) | |
2890 | { | |
2891 | free (element); | |
2892 | return NULL_TREE; | |
2893 | } | |
2894 | ||
2895 | /* If this is an equality test against zero, see if we have recorded a | |
2896 | nonzero value for the variable in question. */ | |
2897 | if ((TREE_CODE (element->rhs) == EQ_EXPR | |
2898 | || TREE_CODE (element->rhs) == NE_EXPR) | |
2899 | && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME | |
2900 | && integer_zerop (TREE_OPERAND (element->rhs, 1))) | |
2901 | { | |
2902 | int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0)); | |
2903 | ||
2904 | if (bitmap_bit_p (nonzero_vars, indx)) | |
2905 | { | |
2906 | tree t = element->rhs; | |
2907 | free (element); | |
2908 | ||
2909 | if (TREE_CODE (t) == EQ_EXPR) | |
2910 | return boolean_false_node; | |
2911 | else | |
2912 | return boolean_true_node; | |
2913 | } | |
2914 | } | |
2915 | ||
2916 | /* Finally try to find the expression in the main expression hash table. */ | |
2917 | slot = htab_find_slot_with_hash (avail_exprs, element, element->hash, | |
2918 | (insert ? INSERT : NO_INSERT)); | |
2919 | if (slot == NULL) | |
2920 | { | |
2921 | free (element); | |
2922 | return NULL_TREE; | |
2923 | } | |
2924 | ||
2925 | if (*slot == NULL) | |
2926 | { | |
2927 | *slot = (void *) element; | |
48732f23 | 2928 | VARRAY_PUSH_TREE (avail_exprs_stack, stmt ? stmt : element->rhs); |
6de9cd9a DN |
2929 | return NULL_TREE; |
2930 | } | |
2931 | ||
2932 | /* Extract the LHS of the assignment so that it can be used as the current | |
2933 | definition of another variable. */ | |
2934 | lhs = ((struct expr_hash_elt *)*slot)->lhs; | |
2935 | ||
2936 | /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then | |
2937 | use the value from the const_and_copies table. */ | |
2938 | if (TREE_CODE (lhs) == SSA_NAME) | |
2939 | { | |
2940 | temp = get_value_for (lhs, const_and_copies); | |
2941 | if (temp) | |
2942 | lhs = temp; | |
2943 | } | |
2944 | ||
2945 | free (element); | |
2946 | return lhs; | |
2947 | } | |
2948 | ||
2949 | /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the | |
2950 | range of values that result in the conditional having a true value. | |
2951 | ||
2952 | Return true if we are successful in extracting a range from COND and | |
2953 | false if we are unsuccessful. */ | |
2954 | ||
2955 | static bool | |
2956 | extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p) | |
2957 | { | |
2958 | tree op1 = TREE_OPERAND (cond, 1); | |
2959 | tree high, low, type; | |
2960 | int inverted; | |
2961 | ||
2962 | /* Experiments have shown that it's rarely, if ever useful to | |
2963 | record ranges for enumerations. Presumably this is due to | |
2964 | the fact that they're rarely used directly. They are typically | |
2965 | cast into an integer type and used that way. */ | |
2966 | if (TREE_CODE (TREE_TYPE (op1)) != INTEGER_TYPE) | |
2967 | return 0; | |
2968 | ||
2969 | type = TREE_TYPE (op1); | |
2970 | ||
2971 | switch (TREE_CODE (cond)) | |
2972 | { | |
2973 | case EQ_EXPR: | |
2974 | high = low = op1; | |
2975 | inverted = 0; | |
2976 | break; | |
2977 | ||
2978 | case NE_EXPR: | |
2979 | high = low = op1; | |
2980 | inverted = 1; | |
2981 | break; | |
2982 | ||
2983 | case GE_EXPR: | |
2984 | low = op1; | |
2985 | high = TYPE_MAX_VALUE (type); | |
2986 | inverted = 0; | |
2987 | break; | |
2988 | ||
2989 | case GT_EXPR: | |
2990 | low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1); | |
2991 | high = TYPE_MAX_VALUE (type); | |
2992 | inverted = 0; | |
2993 | break; | |
2994 | ||
2995 | case LE_EXPR: | |
2996 | high = op1; | |
2997 | low = TYPE_MIN_VALUE (type); | |
2998 | inverted = 0; | |
2999 | break; | |
3000 | ||
3001 | case LT_EXPR: | |
3002 | high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1); | |
3003 | low = TYPE_MIN_VALUE (type); | |
3004 | inverted = 0; | |
3005 | break; | |
3006 | ||
3007 | default: | |
3008 | return 0; | |
3009 | } | |
3010 | ||
3011 | *hi_p = high; | |
3012 | *lo_p = low; | |
3013 | *inverted_p = inverted; | |
3014 | return 1; | |
3015 | } | |
3016 | ||
3017 | /* Record a range created by COND for basic block BB. */ | |
3018 | ||
3019 | static void | |
3020 | record_range (tree cond, basic_block bb, varray_type *vrp_variables_p) | |
3021 | { | |
3022 | /* We explicitly ignore NE_EXPRs. They rarely allow for meaningful | |
3023 | range optimizations and significantly complicate the implementation. */ | |
6615c446 | 3024 | if (COMPARISON_CLASS_P (cond) |
6de9cd9a DN |
3025 | && TREE_CODE (cond) != NE_EXPR |
3026 | && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE) | |
3027 | { | |
3028 | struct vrp_element *element = ggc_alloc (sizeof (struct vrp_element)); | |
3029 | int ssa_version = SSA_NAME_VERSION (TREE_OPERAND (cond, 0)); | |
3030 | ||
3031 | varray_type *vrp_records_p | |
3032 | = (varray_type *)&VARRAY_GENERIC_PTR (vrp_data, ssa_version); | |
3033 | ||
3034 | element->low = NULL; | |
3035 | element->high = NULL; | |
3036 | element->cond = cond; | |
3037 | element->bb = bb; | |
3038 | ||
3039 | if (*vrp_records_p == NULL) | |
3040 | { | |
3041 | VARRAY_GENERIC_PTR_INIT (*vrp_records_p, 2, "vrp records"); | |
3042 | VARRAY_GENERIC_PTR (vrp_data, ssa_version) = *vrp_records_p; | |
3043 | } | |
3044 | ||
3045 | VARRAY_PUSH_GENERIC_PTR (*vrp_records_p, element); | |
3046 | if (! *vrp_variables_p) | |
3047 | VARRAY_TREE_INIT (*vrp_variables_p, 2, "vrp_variables"); | |
3048 | VARRAY_PUSH_TREE (*vrp_variables_p, TREE_OPERAND (cond, 0)); | |
3049 | } | |
3050 | } | |
3051 | ||
3052 | /* Given a conditional statement IF_STMT, return the assignment 'X = Y' | |
3053 | known to be true depending on which arm of IF_STMT is taken. | |
3054 | ||
3055 | Not all conditional statements will result in a useful assignment. | |
3056 | Return NULL_TREE in that case. | |
3057 | ||
3058 | Also enter into the available expression table statements of | |
3059 | the form: | |
3060 | ||
3061 | TRUE ARM FALSE ARM | |
3062 | 1 = cond 1 = cond' | |
3063 | 0 = cond' 0 = cond | |
3064 | ||
3065 | This allows us to lookup the condition in a dominated block and | |
3066 | get back a constant indicating if the condition is true. */ | |
3067 | ||
3068 | static struct eq_expr_value | |
3069 | get_eq_expr_value (tree if_stmt, | |
3070 | int true_arm, | |
6de9cd9a DN |
3071 | basic_block bb, |
3072 | varray_type *vrp_variables_p) | |
3073 | { | |
3074 | tree cond; | |
3075 | struct eq_expr_value retval; | |
3076 | ||
3077 | cond = COND_EXPR_COND (if_stmt); | |
3078 | retval.src = NULL; | |
3079 | retval.dst = NULL; | |
3080 | ||
3081 | /* If the conditional is a single variable 'X', return 'X = 1' for | |
3082 | the true arm and 'X = 0' on the false arm. */ | |
3083 | if (TREE_CODE (cond) == SSA_NAME) | |
3084 | { | |
3085 | retval.dst = cond; | |
e9ea8bd5 | 3086 | retval.src = constant_boolean_node (true_arm, TREE_TYPE (cond)); |
6de9cd9a DN |
3087 | return retval; |
3088 | } | |
3089 | ||
3090 | /* If we have a comparison expression, then record its result into | |
3091 | the available expression table. */ | |
6615c446 | 3092 | if (COMPARISON_CLASS_P (cond)) |
6de9cd9a DN |
3093 | { |
3094 | tree op0 = TREE_OPERAND (cond, 0); | |
3095 | tree op1 = TREE_OPERAND (cond, 1); | |
3096 | ||
3097 | /* Special case comparing booleans against a constant as we know | |
454ff5cb | 3098 | the value of OP0 on both arms of the branch. i.e., we can record |
6de9cd9a DN |
3099 | an equivalence for OP0 rather than COND. */ |
3100 | if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR) | |
3101 | && TREE_CODE (op0) == SSA_NAME | |
3102 | && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE | |
3103 | && is_gimple_min_invariant (op1)) | |
3104 | { | |
3105 | if ((TREE_CODE (cond) == EQ_EXPR && true_arm) | |
3106 | || (TREE_CODE (cond) == NE_EXPR && ! true_arm)) | |
3107 | { | |
3108 | retval.src = op1; | |
3109 | } | |
3110 | else | |
3111 | { | |
3112 | if (integer_zerop (op1)) | |
3113 | retval.src = boolean_true_node; | |
3114 | else | |
3115 | retval.src = boolean_false_node; | |
3116 | } | |
3117 | retval.dst = op0; | |
3118 | return retval; | |
3119 | } | |
3120 | ||
3121 | if (TREE_CODE (op0) == SSA_NAME | |
3122 | && (is_gimple_min_invariant (op1) || TREE_CODE (op1) == SSA_NAME)) | |
3123 | { | |
3124 | tree inverted = invert_truthvalue (cond); | |
3125 | ||
3126 | /* When we find an available expression in the hash table, we replace | |
3127 | the expression with the LHS of the statement in the hash table. | |
3128 | ||
3129 | So, we want to build statements such as "1 = <condition>" on the | |
3130 | true arm and "0 = <condition>" on the false arm. That way if we | |
3131 | find the expression in the table, we will replace it with its | |
3132 | known constant value. Also insert inversions of the result and | |
3133 | condition into the hash table. */ | |
3134 | if (true_arm) | |
3135 | { | |
48732f23 JL |
3136 | record_cond (cond, boolean_true_node); |
3137 | record_dominating_conditions (cond); | |
3138 | record_cond (inverted, boolean_false_node); | |
6de9cd9a DN |
3139 | |
3140 | if (TREE_CONSTANT (op1)) | |
3141 | record_range (cond, bb, vrp_variables_p); | |
3142 | ||
3143 | /* If the conditional is of the form 'X == Y', return 'X = Y' | |
3144 | for the true arm. */ | |
3145 | if (TREE_CODE (cond) == EQ_EXPR) | |
3146 | { | |
3147 | retval.dst = op0; | |
3148 | retval.src = op1; | |
3149 | return retval; | |
3150 | } | |
3151 | } | |
3152 | else | |
3153 | { | |
3154 | ||
48732f23 JL |
3155 | record_cond (inverted, boolean_true_node); |
3156 | record_dominating_conditions (inverted); | |
3157 | record_cond (cond, boolean_false_node); | |
6de9cd9a DN |
3158 | |
3159 | if (TREE_CONSTANT (op1)) | |
3160 | record_range (inverted, bb, vrp_variables_p); | |
3161 | ||
3162 | /* If the conditional is of the form 'X != Y', return 'X = Y' | |
3163 | for the false arm. */ | |
3164 | if (TREE_CODE (cond) == NE_EXPR) | |
3165 | { | |
3166 | retval.dst = op0; | |
3167 | retval.src = op1; | |
3168 | return retval; | |
3169 | } | |
3170 | } | |
3171 | } | |
3172 | } | |
3173 | ||
3174 | return retval; | |
3175 | } | |
3176 | ||
3177 | /* Hashing and equality functions for AVAIL_EXPRS. The table stores | |
3178 | MODIFY_EXPR statements. We compute a value number for expressions using | |
3179 | the code of the expression and the SSA numbers of its operands. */ | |
3180 | ||
3181 | static hashval_t | |
3182 | avail_expr_hash (const void *p) | |
3183 | { | |
3184 | stmt_ann_t ann = ((struct expr_hash_elt *)p)->ann; | |
3185 | tree rhs = ((struct expr_hash_elt *)p)->rhs; | |
3186 | hashval_t val = 0; | |
3187 | size_t i; | |
3188 | vuse_optype vuses; | |
3189 | ||
3190 | /* iterative_hash_expr knows how to deal with any expression and | |
3191 | deals with commutative operators as well, so just use it instead | |
3192 | of duplicating such complexities here. */ | |
3193 | val = iterative_hash_expr (rhs, val); | |
3194 | ||
3195 | /* If the hash table entry is not associated with a statement, then we | |
3196 | can just hash the expression and not worry about virtual operands | |
3197 | and such. */ | |
3198 | if (!ann) | |
3199 | return val; | |
3200 | ||
3201 | /* Add the SSA version numbers of every vuse operand. This is important | |
3202 | because compound variables like arrays are not renamed in the | |
3203 | operands. Rather, the rename is done on the virtual variable | |
3204 | representing all the elements of the array. */ | |
3205 | vuses = VUSE_OPS (ann); | |
3206 | for (i = 0; i < NUM_VUSES (vuses); i++) | |
3207 | val = iterative_hash_expr (VUSE_OP (vuses, i), val); | |
3208 | ||
3209 | return val; | |
3210 | } | |
3211 | ||
940db2c8 RH |
3212 | static hashval_t |
3213 | real_avail_expr_hash (const void *p) | |
3214 | { | |
3215 | return ((const struct expr_hash_elt *)p)->hash; | |
3216 | } | |
6de9cd9a DN |
3217 | |
3218 | static int | |
3219 | avail_expr_eq (const void *p1, const void *p2) | |
3220 | { | |
3221 | stmt_ann_t ann1 = ((struct expr_hash_elt *)p1)->ann; | |
3222 | tree rhs1 = ((struct expr_hash_elt *)p1)->rhs; | |
3223 | stmt_ann_t ann2 = ((struct expr_hash_elt *)p2)->ann; | |
3224 | tree rhs2 = ((struct expr_hash_elt *)p2)->rhs; | |
3225 | ||
3226 | /* If they are the same physical expression, return true. */ | |
3227 | if (rhs1 == rhs2 && ann1 == ann2) | |
3228 | return true; | |
3229 | ||
3230 | /* If their codes are not equal, then quit now. */ | |
3231 | if (TREE_CODE (rhs1) != TREE_CODE (rhs2)) | |
3232 | return false; | |
3233 | ||
3234 | /* In case of a collision, both RHS have to be identical and have the | |
3235 | same VUSE operands. */ | |
3236 | if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2) | |
3237 | || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2))) | |
3238 | && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME)) | |
3239 | { | |
3240 | vuse_optype ops1 = NULL; | |
3241 | vuse_optype ops2 = NULL; | |
3242 | size_t num_ops1 = 0; | |
3243 | size_t num_ops2 = 0; | |
3244 | size_t i; | |
3245 | ||
3246 | if (ann1) | |
3247 | { | |
3248 | ops1 = VUSE_OPS (ann1); | |
3249 | num_ops1 = NUM_VUSES (ops1); | |
3250 | } | |
3251 | ||
3252 | if (ann2) | |
3253 | { | |
3254 | ops2 = VUSE_OPS (ann2); | |
3255 | num_ops2 = NUM_VUSES (ops2); | |
3256 | } | |
3257 | ||
3258 | /* If the number of virtual uses is different, then we consider | |
3259 | them not equal. */ | |
3260 | if (num_ops1 != num_ops2) | |
3261 | return false; | |
3262 | ||
3263 | for (i = 0; i < num_ops1; i++) | |
3264 | if (VUSE_OP (ops1, i) != VUSE_OP (ops2, i)) | |
3265 | return false; | |
3266 | ||
1e128c5f GB |
3267 | gcc_assert (((struct expr_hash_elt *)p1)->hash |
3268 | == ((struct expr_hash_elt *)p2)->hash); | |
6de9cd9a DN |
3269 | return true; |
3270 | } | |
3271 | ||
3272 | return false; | |
3273 | } | |
3274 | ||
61ada8ae | 3275 | /* Given STMT and a pointer to the block local definitions BLOCK_DEFS_P, |
6de9cd9a DN |
3276 | register register all objects set by this statement into BLOCK_DEFS_P |
3277 | and CURRDEFS. */ | |
3278 | ||
3279 | static void | |
9fae925b | 3280 | register_definitions_for_stmt (tree stmt) |
6de9cd9a | 3281 | { |
4c124b4c AM |
3282 | tree def; |
3283 | ssa_op_iter iter; | |
6de9cd9a | 3284 | |
4c124b4c | 3285 | FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_ALL_DEFS) |
6de9cd9a | 3286 | { |
6de9cd9a DN |
3287 | |
3288 | /* FIXME: We shouldn't be registering new defs if the variable | |
3289 | doesn't need to be renamed. */ | |
9fae925b | 3290 | register_new_def (def, &block_defs_stack); |
6de9cd9a | 3291 | } |
6de9cd9a DN |
3292 | } |
3293 |