]>
Commit | Line | Data |
---|---|---|
6de9cd9a | 1 | /* SSA Dominator optimizations for trees |
75b9aa9f | 2 | Copyright (C) 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc. |
6de9cd9a DN |
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 | |
366ccddb KC |
19 | the Free Software Foundation, 51 Franklin Street, Fifth Floor, |
20 | Boston, MA 02110-1301, USA. */ | |
6de9cd9a DN |
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" | |
d38ffc55 | 32 | #include "cfgloop.h" |
6de9cd9a | 33 | #include "output.h" |
6de9cd9a DN |
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 | 44 | #include "langhooks.h" |
43f31be5 | 45 | #include "params.h" |
6de9cd9a DN |
46 | |
47 | /* This file implements optimizations on the dominator tree. */ | |
48 | ||
efea75f9 JL |
49 | |
50 | /* Structure for recording edge equivalences as well as any pending | |
51 | edge redirections during the dominator optimizer. | |
52 | ||
53 | Computing and storing the edge equivalences instead of creating | |
54 | them on-demand can save significant amounts of time, particularly | |
55 | for pathological cases involving switch statements. | |
56 | ||
57 | These structures live for a single iteration of the dominator | |
58 | optimizer in the edge's AUX field. At the end of an iteration we | |
59 | free each of these structures and update the AUX field to point | |
60 | to any requested redirection target (the code for updating the | |
61 | CFG and SSA graph for edge redirection expects redirection edge | |
62 | targets to be in the AUX field for each edge. */ | |
63 | ||
64 | struct edge_info | |
65 | { | |
66 | /* If this edge creates a simple equivalence, the LHS and RHS of | |
67 | the equivalence will be stored here. */ | |
68 | tree lhs; | |
69 | tree rhs; | |
70 | ||
71 | /* Traversing an edge may also indicate one or more particular conditions | |
72 | are true or false. The number of recorded conditions can vary, but | |
73 | can be determined by the condition's code. So we have an array | |
74 | and its maximum index rather than use a varray. */ | |
75 | tree *cond_equivalences; | |
76 | unsigned int max_cond_equivalences; | |
77 | ||
78 | /* If we can thread this edge this field records the new target. */ | |
79 | edge redirection_target; | |
80 | }; | |
81 | ||
82 | ||
6de9cd9a DN |
83 | /* Hash table with expressions made available during the renaming process. |
84 | When an assignment of the form X_i = EXPR is found, the statement is | |
85 | stored in this table. If the same expression EXPR is later found on the | |
86 | RHS of another statement, it is replaced with X_i (thus performing | |
87 | global redundancy elimination). Similarly as we pass through conditionals | |
88 | we record the conditional itself as having either a true or false value | |
89 | in this table. */ | |
90 | static htab_t avail_exprs; | |
91 | ||
48732f23 JL |
92 | /* Stack of available expressions in AVAIL_EXPRs. Each block pushes any |
93 | expressions it enters into the hash table along with a marker entry | |
b3a27618 | 94 | (null). When we finish processing the block, we pop off entries and |
48732f23 JL |
95 | remove the expressions from the global hash table until we hit the |
96 | marker. */ | |
d4e6fecb | 97 | static VEC(tree,heap) *avail_exprs_stack; |
48732f23 | 98 | |
a6e1aa26 JL |
99 | /* Stack of statements we need to rescan during finalization for newly |
100 | exposed variables. | |
101 | ||
102 | Statement rescanning must occur after the current block's available | |
103 | expressions are removed from AVAIL_EXPRS. Else we may change the | |
104 | hash code for an expression and be unable to find/remove it from | |
105 | AVAIL_EXPRS. */ | |
d4e6fecb | 106 | static VEC(tree,heap) *stmts_to_rescan; |
a6e1aa26 | 107 | |
6de9cd9a DN |
108 | /* Structure for entries in the expression hash table. |
109 | ||
110 | This requires more memory for the hash table entries, but allows us | |
111 | to avoid creating silly tree nodes and annotations for conditionals, | |
112 | eliminates 2 global hash tables and two block local varrays. | |
113 | ||
114 | It also allows us to reduce the number of hash table lookups we | |
115 | have to perform in lookup_avail_expr and finally it allows us to | |
116 | significantly reduce the number of calls into the hashing routine | |
117 | itself. */ | |
56b043c8 | 118 | |
6de9cd9a DN |
119 | struct expr_hash_elt |
120 | { | |
121 | /* The value (lhs) of this expression. */ | |
122 | tree lhs; | |
123 | ||
124 | /* The expression (rhs) we want to record. */ | |
125 | tree rhs; | |
126 | ||
f47c96aa AM |
127 | /* The stmt pointer if this element corresponds to a statement. */ |
128 | tree stmt; | |
6de9cd9a DN |
129 | |
130 | /* The hash value for RHS/ann. */ | |
131 | hashval_t hash; | |
132 | }; | |
133 | ||
b5fefcf6 JL |
134 | /* Stack of dest,src pairs that need to be restored during finalization. |
135 | ||
136 | A NULL entry is used to mark the end of pairs which need to be | |
137 | restored during finalization of this block. */ | |
d4e6fecb | 138 | static VEC(tree,heap) *const_and_copies_stack; |
b5fefcf6 | 139 | |
6de9cd9a DN |
140 | /* Bitmap of SSA_NAMEs known to have a nonzero value, even if we do not |
141 | know their exact value. */ | |
142 | static bitmap nonzero_vars; | |
143 | ||
4aab792d KH |
144 | /* Bitmap of blocks that are scheduled to be threaded through. This |
145 | is used to communicate with thread_through_blocks. */ | |
146 | static bitmap threaded_blocks; | |
147 | ||
fdabe5c2 JL |
148 | /* Stack of SSA_NAMEs which need their NONZERO_VARS property cleared |
149 | when the current block is finalized. | |
150 | ||
151 | A NULL entry is used to mark the end of names needing their | |
152 | entry in NONZERO_VARS cleared during finalization of this block. */ | |
d4e6fecb | 153 | static VEC(tree,heap) *nonzero_vars_stack; |
fdabe5c2 | 154 | |
6de9cd9a DN |
155 | /* Track whether or not we have changed the control flow graph. */ |
156 | static bool cfg_altered; | |
157 | ||
1eaba2f2 | 158 | /* Bitmap of blocks that have had EH statements cleaned. We should |
f6fe65dc | 159 | remove their dead edges eventually. */ |
1eaba2f2 RH |
160 | static bitmap need_eh_cleanup; |
161 | ||
6de9cd9a DN |
162 | /* Statistics for dominator optimizations. */ |
163 | struct opt_stats_d | |
164 | { | |
165 | long num_stmts; | |
166 | long num_exprs_considered; | |
167 | long num_re; | |
0bca51f0 DN |
168 | long num_const_prop; |
169 | long num_copy_prop; | |
6a2cf9dc | 170 | long num_iterations; |
6de9cd9a DN |
171 | }; |
172 | ||
23530866 JL |
173 | static struct opt_stats_d opt_stats; |
174 | ||
6de9cd9a DN |
175 | /* Value range propagation record. Each time we encounter a conditional |
176 | of the form SSA_NAME COND CONST we create a new vrp_element to record | |
177 | how the condition affects the possible values SSA_NAME may have. | |
178 | ||
3f117656 | 179 | Each record contains the condition tested (COND), and the range of |
6de9cd9a DN |
180 | values the variable may legitimately have if COND is true. Note the |
181 | range of values may be a smaller range than COND specifies if we have | |
182 | recorded other ranges for this variable. Each record also contains the | |
183 | block in which the range was recorded for invalidation purposes. | |
184 | ||
185 | Note that the current known range is computed lazily. This allows us | |
186 | to avoid the overhead of computing ranges which are never queried. | |
187 | ||
188 | When we encounter a conditional, we look for records which constrain | |
189 | the SSA_NAME used in the condition. In some cases those records allow | |
190 | us to determine the condition's result at compile time. In other cases | |
191 | they may allow us to simplify the condition. | |
192 | ||
193 | We also use value ranges to do things like transform signed div/mod | |
194 | operations into unsigned div/mod or to simplify ABS_EXPRs. | |
195 | ||
196 | Simple experiments have shown these optimizations to not be all that | |
197 | useful on switch statements (much to my surprise). So switch statement | |
198 | optimizations are not performed. | |
199 | ||
200 | Note carefully we do not propagate information through each statement | |
454ff5cb | 201 | in the block. i.e., if we know variable X has a value defined of |
6de9cd9a DN |
202 | [0, 25] and we encounter Y = X + 1, we do not track a value range |
203 | for Y (which would be [1, 26] if we cared). Similarly we do not | |
204 | constrain values as we encounter narrowing typecasts, etc. */ | |
205 | ||
206 | struct vrp_element | |
207 | { | |
208 | /* The highest and lowest values the variable in COND may contain when | |
209 | COND is true. Note this may not necessarily be the same values | |
210 | tested by COND if the same variable was used in earlier conditionals. | |
211 | ||
212 | Note this is computed lazily and thus can be NULL indicating that | |
213 | the values have not been computed yet. */ | |
214 | tree low; | |
215 | tree high; | |
216 | ||
217 | /* The actual conditional we recorded. This is needed since we compute | |
218 | ranges lazily. */ | |
219 | tree cond; | |
220 | ||
221 | /* The basic block where this record was created. We use this to determine | |
222 | when to remove records. */ | |
223 | basic_block bb; | |
224 | }; | |
225 | ||
23530866 JL |
226 | /* A hash table holding value range records (VRP_ELEMENTs) for a given |
227 | SSA_NAME. We used to use a varray indexed by SSA_NAME_VERSION, but | |
228 | that gets awful wasteful, particularly since the density objects | |
229 | with useful information is very low. */ | |
230 | static htab_t vrp_data; | |
231 | ||
8184759d KH |
232 | typedef struct vrp_element *vrp_element_p; |
233 | ||
234 | DEF_VEC_P(vrp_element_p); | |
235 | DEF_VEC_ALLOC_P(vrp_element_p,heap); | |
236 | ||
23530866 | 237 | /* An entry in the VRP_DATA hash table. We record the variable and a |
471854f8 | 238 | varray of VRP_ELEMENT records associated with that variable. */ |
23530866 JL |
239 | struct vrp_hash_elt |
240 | { | |
241 | tree var; | |
8184759d | 242 | VEC(vrp_element_p,heap) *records; |
23530866 | 243 | }; |
6de9cd9a | 244 | |
fdabe5c2 JL |
245 | /* Array of variables which have their values constrained by operations |
246 | in this basic block. We use this during finalization to know | |
247 | which variables need their VRP data updated. */ | |
6de9cd9a | 248 | |
35fd3193 | 249 | /* Stack of SSA_NAMEs which had their values constrained by operations |
fdabe5c2 JL |
250 | in this basic block. During finalization of this block we use this |
251 | list to determine which variables need their VRP data updated. | |
252 | ||
253 | A NULL entry marks the end of the SSA_NAMEs associated with this block. */ | |
d4e6fecb | 254 | static VEC(tree,heap) *vrp_variables_stack; |
6de9cd9a DN |
255 | |
256 | struct eq_expr_value | |
257 | { | |
258 | tree src; | |
259 | tree dst; | |
260 | }; | |
261 | ||
262 | /* Local functions. */ | |
263 | static void optimize_stmt (struct dom_walk_data *, | |
264 | basic_block bb, | |
265 | block_stmt_iterator); | |
48732f23 | 266 | static tree lookup_avail_expr (tree, bool); |
23530866 JL |
267 | static hashval_t vrp_hash (const void *); |
268 | static int vrp_eq (const void *, const void *); | |
6de9cd9a | 269 | static hashval_t avail_expr_hash (const void *); |
940db2c8 | 270 | static hashval_t real_avail_expr_hash (const void *); |
6de9cd9a DN |
271 | static int avail_expr_eq (const void *, const void *); |
272 | static void htab_statistics (FILE *, htab_t); | |
48732f23 | 273 | static void record_cond (tree, tree); |
b5fefcf6 JL |
274 | static void record_const_or_copy (tree, tree); |
275 | static void record_equality (tree, tree); | |
48732f23 | 276 | static tree simplify_cond_and_lookup_avail_expr (tree, stmt_ann_t, int); |
6de9cd9a | 277 | static tree find_equivalent_equality_comparison (tree); |
fdabe5c2 | 278 | static void record_range (tree, basic_block); |
6de9cd9a | 279 | static bool extract_range_from_cond (tree, tree *, tree *, int *); |
efea75f9 JL |
280 | static void record_equivalences_from_phis (basic_block); |
281 | static void record_equivalences_from_incoming_edge (basic_block); | |
a513fe88 | 282 | static bool eliminate_redundant_computations (tree, stmt_ann_t); |
fdabe5c2 | 283 | static void record_equivalences_from_stmt (tree, int, stmt_ann_t); |
6de9cd9a DN |
284 | static void thread_across_edge (struct dom_walk_data *, edge); |
285 | static void dom_opt_finalize_block (struct dom_walk_data *, basic_block); | |
6de9cd9a | 286 | static void dom_opt_initialize_block (struct dom_walk_data *, basic_block); |
efea75f9 | 287 | static void propagate_to_outgoing_edges (struct dom_walk_data *, basic_block); |
48732f23 | 288 | static void remove_local_expressions_from_table (void); |
b5fefcf6 | 289 | static void restore_vars_to_original_value (void); |
28c008bb | 290 | static edge single_incoming_edge_ignoring_loop_edges (basic_block); |
fdabe5c2 | 291 | static void restore_nonzero_vars_to_original_value (void); |
01d8c00b | 292 | static inline bool unsafe_associative_fp_binop (tree); |
6de9cd9a | 293 | |
0bca51f0 | 294 | |
6de9cd9a DN |
295 | /* Local version of fold that doesn't introduce cruft. */ |
296 | ||
297 | static tree | |
298 | local_fold (tree t) | |
299 | { | |
300 | t = fold (t); | |
301 | ||
302 | /* Strip away useless type conversions. Both the NON_LVALUE_EXPR that | |
303 | may have been added by fold, and "useless" type conversions that might | |
304 | now be apparent due to propagation. */ | |
6de9cd9a DN |
305 | STRIP_USELESS_TYPE_CONVERSION (t); |
306 | ||
307 | return t; | |
308 | } | |
309 | ||
efea75f9 JL |
310 | /* Allocate an EDGE_INFO for edge E and attach it to E. |
311 | Return the new EDGE_INFO structure. */ | |
312 | ||
313 | static struct edge_info * | |
314 | allocate_edge_info (edge e) | |
315 | { | |
316 | struct edge_info *edge_info; | |
317 | ||
e1111e8e | 318 | edge_info = XCNEW (struct edge_info); |
efea75f9 JL |
319 | |
320 | e->aux = edge_info; | |
321 | return edge_info; | |
322 | } | |
323 | ||
324 | /* Free all EDGE_INFO structures associated with edges in the CFG. | |
cbb1cada | 325 | If a particular edge can be threaded, copy the redirection |
efea75f9 JL |
326 | target from the EDGE_INFO structure into the edge's AUX field |
327 | as required by code to update the CFG and SSA graph for | |
328 | jump threading. */ | |
329 | ||
330 | static void | |
331 | free_all_edge_infos (void) | |
332 | { | |
333 | basic_block bb; | |
334 | edge_iterator ei; | |
335 | edge e; | |
336 | ||
337 | FOR_EACH_BB (bb) | |
338 | { | |
339 | FOR_EACH_EDGE (e, ei, bb->preds) | |
340 | { | |
e1111e8e | 341 | struct edge_info *edge_info = (struct edge_info *) e->aux; |
efea75f9 JL |
342 | |
343 | if (edge_info) | |
344 | { | |
345 | e->aux = edge_info->redirection_target; | |
346 | if (edge_info->cond_equivalences) | |
347 | free (edge_info->cond_equivalences); | |
348 | free (edge_info); | |
349 | } | |
350 | } | |
351 | } | |
352 | } | |
353 | ||
8184759d KH |
354 | /* Free an instance of vrp_hash_elt. */ |
355 | ||
356 | static void | |
357 | vrp_free (void *data) | |
358 | { | |
e1111e8e | 359 | struct vrp_hash_elt *elt = (struct vrp_hash_elt *) data; |
8184759d KH |
360 | struct VEC(vrp_element_p,heap) **vrp_elt = &elt->records; |
361 | ||
362 | VEC_free (vrp_element_p, heap, *vrp_elt); | |
363 | free (elt); | |
364 | } | |
365 | ||
6de9cd9a DN |
366 | /* Jump threading, redundancy elimination and const/copy propagation. |
367 | ||
6de9cd9a DN |
368 | This pass may expose new symbols that need to be renamed into SSA. For |
369 | every new symbol exposed, its corresponding bit will be set in | |
ff2ad0f7 | 370 | VARS_TO_RENAME. */ |
6de9cd9a DN |
371 | |
372 | static void | |
373 | tree_ssa_dominator_optimize (void) | |
374 | { | |
6de9cd9a DN |
375 | struct dom_walk_data walk_data; |
376 | unsigned int i; | |
d38ffc55 | 377 | struct loops loops_info; |
6de9cd9a | 378 | |
fded8de7 DN |
379 | memset (&opt_stats, 0, sizeof (opt_stats)); |
380 | ||
6de9cd9a | 381 | /* Create our hash tables. */ |
940db2c8 | 382 | avail_exprs = htab_create (1024, real_avail_expr_hash, avail_expr_eq, free); |
8184759d KH |
383 | vrp_data = htab_create (ceil_log2 (num_ssa_names), vrp_hash, vrp_eq, |
384 | vrp_free); | |
d4e6fecb | 385 | avail_exprs_stack = VEC_alloc (tree, heap, 20); |
d4e6fecb NS |
386 | const_and_copies_stack = VEC_alloc (tree, heap, 20); |
387 | nonzero_vars_stack = VEC_alloc (tree, heap, 20); | |
388 | vrp_variables_stack = VEC_alloc (tree, heap, 20); | |
389 | stmts_to_rescan = VEC_alloc (tree, heap, 20); | |
8bdbfff5 | 390 | nonzero_vars = BITMAP_ALLOC (NULL); |
4aab792d | 391 | threaded_blocks = BITMAP_ALLOC (NULL); |
8bdbfff5 | 392 | need_eh_cleanup = BITMAP_ALLOC (NULL); |
6de9cd9a DN |
393 | |
394 | /* Setup callbacks for the generic dominator tree walker. */ | |
395 | walk_data.walk_stmts_backward = false; | |
396 | walk_data.dom_direction = CDI_DOMINATORS; | |
fdabe5c2 | 397 | walk_data.initialize_block_local_data = NULL; |
6de9cd9a DN |
398 | walk_data.before_dom_children_before_stmts = dom_opt_initialize_block; |
399 | walk_data.before_dom_children_walk_stmts = optimize_stmt; | |
efea75f9 | 400 | walk_data.before_dom_children_after_stmts = propagate_to_outgoing_edges; |
6de9cd9a DN |
401 | walk_data.after_dom_children_before_stmts = NULL; |
402 | walk_data.after_dom_children_walk_stmts = NULL; | |
403 | walk_data.after_dom_children_after_stmts = dom_opt_finalize_block; | |
404 | /* Right now we only attach a dummy COND_EXPR to the global data pointer. | |
405 | When we attach more stuff we'll need to fill this out with a real | |
406 | structure. */ | |
407 | walk_data.global_data = NULL; | |
fdabe5c2 | 408 | walk_data.block_local_data_size = 0; |
0bca51f0 | 409 | walk_data.interesting_blocks = NULL; |
6de9cd9a DN |
410 | |
411 | /* Now initialize the dominator walker. */ | |
412 | init_walk_dominator_tree (&walk_data); | |
413 | ||
6de9cd9a DN |
414 | calculate_dominance_info (CDI_DOMINATORS); |
415 | ||
d38ffc55 JL |
416 | /* We need to know which edges exit loops so that we can |
417 | aggressively thread through loop headers to an exit | |
418 | edge. */ | |
419 | flow_loops_find (&loops_info); | |
420 | mark_loop_exit_edges (&loops_info); | |
421 | flow_loops_free (&loops_info); | |
422 | ||
423 | /* Clean up the CFG so that any forwarder blocks created by loop | |
424 | canonicalization are removed. */ | |
425 | cleanup_tree_cfg (); | |
38965eb2 | 426 | calculate_dominance_info (CDI_DOMINATORS); |
d38ffc55 | 427 | |
6de9cd9a DN |
428 | /* If we prove certain blocks are unreachable, then we want to |
429 | repeat the dominator optimization process as PHI nodes may | |
430 | have turned into copies which allows better propagation of | |
431 | values. So we repeat until we do not identify any new unreachable | |
432 | blocks. */ | |
433 | do | |
434 | { | |
435 | /* Optimize the dominator tree. */ | |
436 | cfg_altered = false; | |
437 | ||
d38ffc55 JL |
438 | /* We need accurate information regarding back edges in the CFG |
439 | for jump threading. */ | |
440 | mark_dfs_back_edges (); | |
441 | ||
6de9cd9a DN |
442 | /* Recursively walk the dominator tree optimizing statements. */ |
443 | walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR); | |
444 | ||
f430bae8 | 445 | { |
a3b609df KH |
446 | block_stmt_iterator bsi; |
447 | basic_block bb; | |
448 | FOR_EACH_BB (bb) | |
f430bae8 | 449 | { |
a3b609df KH |
450 | for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) |
451 | { | |
452 | update_stmt_if_modified (bsi_stmt (bsi)); | |
453 | } | |
f430bae8 AM |
454 | } |
455 | } | |
a3b609df | 456 | |
6e3b9e27 KH |
457 | /* If we exposed any new variables, go ahead and put them into |
458 | SSA form now, before we handle jump threading. This simplifies | |
459 | interactions between rewriting of _DECL nodes into SSA form | |
460 | and rewriting SSA_NAME nodes into SSA form after block | |
461 | duplication and CFG manipulation. */ | |
462 | update_ssa (TODO_update_ssa); | |
463 | ||
464 | free_all_edge_infos (); | |
465 | ||
56b043c8 | 466 | /* Thread jumps, creating duplicate blocks as needed. */ |
4aab792d | 467 | cfg_altered |= thread_through_all_blocks (threaded_blocks); |
6de9cd9a | 468 | |
56b043c8 JL |
469 | /* Removal of statements may make some EH edges dead. Purge |
470 | such edges from the CFG as needed. */ | |
eb59b8de | 471 | if (!bitmap_empty_p (need_eh_cleanup)) |
1eaba2f2 | 472 | { |
56b043c8 | 473 | cfg_altered |= tree_purge_all_dead_eh_edges (need_eh_cleanup); |
1eaba2f2 RH |
474 | bitmap_zero (need_eh_cleanup); |
475 | } | |
476 | ||
13396b14 | 477 | if (cfg_altered) |
d38ffc55 JL |
478 | free_dominance_info (CDI_DOMINATORS); |
479 | ||
608af77d JL |
480 | /* Only iterate if we threaded jumps AND the CFG cleanup did |
481 | something interesting. Other cases generate far fewer | |
482 | optimization opportunities and thus are not worth another | |
483 | full DOM iteration. */ | |
484 | cfg_altered &= cleanup_tree_cfg (); | |
d38ffc55 JL |
485 | |
486 | if (rediscover_loops_after_threading) | |
487 | { | |
488 | /* Rerun basic loop analysis to discover any newly | |
489 | created loops and update the set of exit edges. */ | |
490 | rediscover_loops_after_threading = false; | |
491 | flow_loops_find (&loops_info); | |
492 | mark_loop_exit_edges (&loops_info); | |
493 | flow_loops_free (&loops_info); | |
494 | ||
495 | /* Remove any forwarder blocks inserted by loop | |
496 | header canonicalization. */ | |
497 | cleanup_tree_cfg (); | |
498 | } | |
499 | ||
56b043c8 | 500 | calculate_dominance_info (CDI_DOMINATORS); |
6de9cd9a | 501 | |
84d65814 | 502 | update_ssa (TODO_update_ssa); |
6de9cd9a | 503 | |
6de9cd9a DN |
504 | /* Reinitialize the various tables. */ |
505 | bitmap_clear (nonzero_vars); | |
4aab792d | 506 | bitmap_clear (threaded_blocks); |
6de9cd9a | 507 | htab_empty (avail_exprs); |
23530866 | 508 | htab_empty (vrp_data); |
6de9cd9a | 509 | |
0bd65483 JL |
510 | /* Finally, remove everything except invariants in SSA_NAME_VALUE. |
511 | ||
512 | This must be done before we iterate as we might have a | |
513 | reference to an SSA_NAME which was removed by the call to | |
31117a5c | 514 | update_ssa. |
0bd65483 JL |
515 | |
516 | Long term we will be able to let everything in SSA_NAME_VALUE | |
517 | persist. However, for now, we know this is the safe thing to do. */ | |
518 | for (i = 0; i < num_ssa_names; i++) | |
519 | { | |
520 | tree name = ssa_name (i); | |
521 | tree value; | |
522 | ||
523 | if (!name) | |
524 | continue; | |
525 | ||
526 | value = SSA_NAME_VALUE (name); | |
527 | if (value && !is_gimple_min_invariant (value)) | |
528 | SSA_NAME_VALUE (name) = NULL; | |
529 | } | |
6a2cf9dc DN |
530 | |
531 | opt_stats.num_iterations++; | |
6de9cd9a | 532 | } |
116cb604 | 533 | while (optimize > 1 && cfg_altered); |
6de9cd9a | 534 | |
6de9cd9a DN |
535 | /* Debugging dumps. */ |
536 | if (dump_file && (dump_flags & TDF_STATS)) | |
537 | dump_dominator_optimization_stats (dump_file); | |
538 | ||
61ada8ae | 539 | /* We emptied the hash table earlier, now delete it completely. */ |
6de9cd9a | 540 | htab_delete (avail_exprs); |
23530866 | 541 | htab_delete (vrp_data); |
6de9cd9a | 542 | |
1ea7e6ad | 543 | /* It is not necessary to clear CURRDEFS, REDIRECTION_EDGES, VRP_DATA, |
6de9cd9a DN |
544 | CONST_AND_COPIES, and NONZERO_VARS as they all get cleared at the bottom |
545 | of the do-while loop above. */ | |
546 | ||
547 | /* And finalize the dominator walker. */ | |
548 | fini_walk_dominator_tree (&walk_data); | |
cfa4cb00 | 549 | |
471854f8 | 550 | /* Free nonzero_vars. */ |
8bdbfff5 | 551 | BITMAP_FREE (nonzero_vars); |
4aab792d | 552 | BITMAP_FREE (threaded_blocks); |
8bdbfff5 | 553 | BITMAP_FREE (need_eh_cleanup); |
3a2e4b46 | 554 | |
d4e6fecb NS |
555 | VEC_free (tree, heap, avail_exprs_stack); |
556 | VEC_free (tree, heap, const_and_copies_stack); | |
557 | VEC_free (tree, heap, nonzero_vars_stack); | |
558 | VEC_free (tree, heap, vrp_variables_stack); | |
559 | VEC_free (tree, heap, stmts_to_rescan); | |
6de9cd9a DN |
560 | } |
561 | ||
562 | static bool | |
563 | gate_dominator (void) | |
564 | { | |
565 | return flag_tree_dom != 0; | |
566 | } | |
567 | ||
568 | struct tree_opt_pass pass_dominator = | |
569 | { | |
570 | "dom", /* name */ | |
571 | gate_dominator, /* gate */ | |
572 | tree_ssa_dominator_optimize, /* execute */ | |
573 | NULL, /* sub */ | |
574 | NULL, /* next */ | |
575 | 0, /* static_pass_number */ | |
576 | TV_TREE_SSA_DOMINATOR_OPTS, /* tv_id */ | |
c1b763fa | 577 | PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */ |
6de9cd9a DN |
578 | 0, /* properties_provided */ |
579 | 0, /* properties_destroyed */ | |
580 | 0, /* todo_flags_start */ | |
0bca51f0 DN |
581 | TODO_dump_func |
582 | | TODO_update_ssa | |
9f8628ba PB |
583 | | TODO_verify_ssa, /* todo_flags_finish */ |
584 | 0 /* letter */ | |
6de9cd9a DN |
585 | }; |
586 | ||
587 | ||
0e0ed594 JL |
588 | /* Given a stmt CONDSTMT containing a COND_EXPR, canonicalize the |
589 | COND_EXPR into a canonical form. */ | |
590 | ||
591 | static void | |
592 | canonicalize_comparison (tree condstmt) | |
593 | { | |
594 | tree cond = COND_EXPR_COND (condstmt); | |
595 | tree op0; | |
596 | tree op1; | |
597 | enum tree_code code = TREE_CODE (cond); | |
598 | ||
599 | if (!COMPARISON_CLASS_P (cond)) | |
600 | return; | |
601 | ||
602 | op0 = TREE_OPERAND (cond, 0); | |
603 | op1 = TREE_OPERAND (cond, 1); | |
604 | ||
605 | /* If it would be profitable to swap the operands, then do so to | |
606 | canonicalize the statement, enabling better optimization. | |
607 | ||
608 | By placing canonicalization of such expressions here we | |
609 | transparently keep statements in canonical form, even | |
610 | when the statement is modified. */ | |
611 | if (tree_swap_operands_p (op0, op1, false)) | |
612 | { | |
613 | /* For relationals we need to swap the operands | |
614 | and change the code. */ | |
615 | if (code == LT_EXPR | |
616 | || code == GT_EXPR | |
617 | || code == LE_EXPR | |
618 | || code == GE_EXPR) | |
619 | { | |
620 | TREE_SET_CODE (cond, swap_tree_comparison (code)); | |
621 | swap_tree_operands (condstmt, | |
622 | &TREE_OPERAND (cond, 0), | |
623 | &TREE_OPERAND (cond, 1)); | |
624 | /* If one operand was in the operand cache, but the other is | |
625 | not, because it is a constant, this is a case that the | |
626 | internal updating code of swap_tree_operands can't handle | |
627 | properly. */ | |
628 | if (TREE_CODE_CLASS (TREE_CODE (op0)) | |
629 | != TREE_CODE_CLASS (TREE_CODE (op1))) | |
630 | update_stmt (condstmt); | |
631 | } | |
632 | } | |
633 | } | |
74d2efc7 JL |
634 | /* We are exiting E->src, see if E->dest ends with a conditional |
635 | jump which has a known value when reached via E. | |
636 | ||
637 | Special care is necessary if E is a back edge in the CFG as we | |
638 | will have already recorded equivalences for E->dest into our | |
639 | various tables, including the result of the conditional at | |
640 | the end of E->dest. Threading opportunities are severely | |
641 | limited in that case to avoid short-circuiting the loop | |
642 | incorrectly. | |
643 | ||
644 | Note it is quite common for the first block inside a loop to | |
645 | end with a conditional which is either always true or always | |
646 | false when reached via the loop backedge. Thus we do not want | |
647 | to blindly disable threading across a loop backedge. */ | |
6de9cd9a DN |
648 | |
649 | static void | |
650 | thread_across_edge (struct dom_walk_data *walk_data, edge e) | |
651 | { | |
6de9cd9a DN |
652 | block_stmt_iterator bsi; |
653 | tree stmt = NULL; | |
654 | tree phi; | |
43f31be5 JL |
655 | int stmt_count = 0; |
656 | int max_stmt_count; | |
657 | ||
6de9cd9a | 658 | |
74d2efc7 JL |
659 | /* If E->dest does not end with a conditional, then there is |
660 | nothing to do. */ | |
661 | bsi = bsi_last (e->dest); | |
662 | if (bsi_end_p (bsi) | |
663 | || ! bsi_stmt (bsi) | |
664 | || (TREE_CODE (bsi_stmt (bsi)) != COND_EXPR | |
665 | && TREE_CODE (bsi_stmt (bsi)) != GOTO_EXPR | |
666 | && TREE_CODE (bsi_stmt (bsi)) != SWITCH_EXPR)) | |
667 | return; | |
668 | ||
669 | /* The basic idea here is to use whatever knowledge we have | |
670 | from our dominator walk to simplify statements in E->dest, | |
671 | with the ultimate goal being to simplify the conditional | |
672 | at the end of E->dest. | |
673 | ||
674 | Note that we must undo any changes we make to the underlying | |
675 | statements as the simplifications we are making are control | |
676 | flow sensitive (ie, the simplifications are valid when we | |
677 | traverse E, but may not be valid on other paths to E->dest. */ | |
678 | ||
679 | /* Each PHI creates a temporary equivalence, record them. Again | |
680 | these are context sensitive equivalences and will be removed | |
681 | by our caller. */ | |
17192884 | 682 | for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi)) |
6de9cd9a | 683 | { |
d00ad49b | 684 | tree src = PHI_ARG_DEF_FROM_EDGE (phi, e); |
6de9cd9a | 685 | tree dst = PHI_RESULT (phi); |
fe872835 | 686 | |
43f31be5 JL |
687 | /* Do not include virtual PHIs in our statement count as |
688 | they never generate code. */ | |
689 | if (is_gimple_reg (dst)) | |
690 | stmt_count++; | |
691 | ||
fe872835 | 692 | /* If the desired argument is not the same as this PHI's result |
74d2efc7 JL |
693 | and it is set by a PHI in E->dest, then we can not thread |
694 | through E->dest. */ | |
fe872835 JL |
695 | if (src != dst |
696 | && TREE_CODE (src) == SSA_NAME | |
697 | && TREE_CODE (SSA_NAME_DEF_STMT (src)) == PHI_NODE | |
698 | && bb_for_stmt (SSA_NAME_DEF_STMT (src)) == e->dest) | |
699 | return; | |
700 | ||
b5fefcf6 | 701 | record_const_or_copy (dst, src); |
6de9cd9a DN |
702 | } |
703 | ||
74d2efc7 JL |
704 | /* Try to simplify each statement in E->dest, ultimately leading to |
705 | a simplification of the COND_EXPR at the end of E->dest. | |
706 | ||
707 | We might consider marking just those statements which ultimately | |
708 | feed the COND_EXPR. It's not clear if the overhead of bookkeeping | |
709 | would be recovered by trying to simplify fewer statements. | |
710 | ||
711 | If we are able to simplify a statement into the form | |
712 | SSA_NAME = (SSA_NAME | gimple invariant), then we can record | |
713 | a context sensitive equivalency which may help us simplify | |
714 | later statements in E->dest. | |
715 | ||
716 | Failure to simplify into the form above merely means that the | |
717 | statement provides no equivalences to help simplify later | |
718 | statements. This does not prevent threading through E->dest. */ | |
43f31be5 | 719 | max_stmt_count = PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS); |
6de9cd9a DN |
720 | for (bsi = bsi_start (e->dest); ! bsi_end_p (bsi); bsi_next (&bsi)) |
721 | { | |
46fbb29c | 722 | tree cached_lhs = NULL; |
6de9cd9a DN |
723 | |
724 | stmt = bsi_stmt (bsi); | |
725 | ||
726 | /* Ignore empty statements and labels. */ | |
727 | if (IS_EMPTY_STMT (stmt) || TREE_CODE (stmt) == LABEL_EXPR) | |
728 | continue; | |
729 | ||
43f31be5 JL |
730 | /* If duplicating this block is going to cause too much code |
731 | expansion, then do not thread through this block. */ | |
732 | stmt_count++; | |
733 | if (stmt_count > max_stmt_count) | |
734 | return; | |
735 | ||
74d2efc7 JL |
736 | /* Safely handle threading across loop backedges. This is |
737 | over conservative, but still allows us to capture the | |
738 | majority of the cases where we can thread across a loop | |
739 | backedge. */ | |
740 | if ((e->flags & EDGE_DFS_BACK) != 0 | |
741 | && TREE_CODE (stmt) != COND_EXPR | |
742 | && TREE_CODE (stmt) != SWITCH_EXPR) | |
743 | return; | |
744 | ||
745 | /* If the statement has volatile operands, then we assume we | |
746 | can not thread through this block. This is overly | |
747 | conservative in some ways. */ | |
748 | if (TREE_CODE (stmt) == ASM_EXPR && ASM_VOLATILE_P (stmt)) | |
749 | return; | |
750 | ||
6de9cd9a | 751 | /* If this is not a MODIFY_EXPR which sets an SSA_NAME to a new |
74d2efc7 JL |
752 | value, then do not try to simplify this statement as it will |
753 | not simplify in any way that is helpful for jump threading. */ | |
6de9cd9a DN |
754 | if (TREE_CODE (stmt) != MODIFY_EXPR |
755 | || TREE_CODE (TREE_OPERAND (stmt, 0)) != SSA_NAME) | |
74d2efc7 | 756 | continue; |
6de9cd9a DN |
757 | |
758 | /* At this point we have a statement which assigns an RHS to an | |
74d2efc7 JL |
759 | SSA_VAR on the LHS. We want to try and simplify this statement |
760 | to expose more context sensitive equivalences which in turn may | |
761 | allow us to simplify the condition at the end of the loop. */ | |
6de9cd9a DN |
762 | if (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME) |
763 | cached_lhs = TREE_OPERAND (stmt, 1); | |
764 | else | |
6de9cd9a DN |
765 | { |
766 | /* Copy the operands. */ | |
46fbb29c | 767 | tree *copy, pre_fold_expr; |
f47c96aa AM |
768 | ssa_op_iter iter; |
769 | use_operand_p use_p; | |
770 | unsigned int num, i = 0; | |
6de9cd9a | 771 | |
f47c96aa | 772 | num = NUM_SSA_OPERANDS (stmt, (SSA_OP_USE | SSA_OP_VUSE)); |
e1111e8e | 773 | copy = XCNEWVEC (tree, num); |
6de9cd9a | 774 | |
f47c96aa AM |
775 | /* Make a copy of the uses & vuses into USES_COPY, then cprop into |
776 | the operands. */ | |
777 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE) | |
6de9cd9a DN |
778 | { |
779 | tree tmp = NULL; | |
f47c96aa | 780 | tree use = USE_FROM_PTR (use_p); |
6de9cd9a | 781 | |
f47c96aa AM |
782 | copy[i++] = use; |
783 | if (TREE_CODE (use) == SSA_NAME) | |
784 | tmp = SSA_NAME_VALUE (use); | |
3aecd08b | 785 | if (tmp && TREE_CODE (tmp) != VALUE_HANDLE) |
f47c96aa | 786 | SET_USE (use_p, tmp); |
6de9cd9a DN |
787 | } |
788 | ||
74d2efc7 JL |
789 | /* Try to fold/lookup the new expression. Inserting the |
790 | expression into the hash table is unlikely to help | |
46fbb29c SB |
791 | Sadly, we have to handle conditional assignments specially |
792 | here, because fold expects all the operands of an expression | |
793 | to be folded before the expression itself is folded, but we | |
794 | can't just substitute the folded condition here. */ | |
795 | if (TREE_CODE (TREE_OPERAND (stmt, 1)) == COND_EXPR) | |
796 | { | |
797 | tree cond = COND_EXPR_COND (TREE_OPERAND (stmt, 1)); | |
798 | cond = fold (cond); | |
799 | if (cond == boolean_true_node) | |
800 | pre_fold_expr = COND_EXPR_THEN (TREE_OPERAND (stmt, 1)); | |
801 | else if (cond == boolean_false_node) | |
802 | pre_fold_expr = COND_EXPR_ELSE (TREE_OPERAND (stmt, 1)); | |
803 | else | |
804 | pre_fold_expr = TREE_OPERAND (stmt, 1); | |
805 | } | |
806 | else | |
807 | pre_fold_expr = TREE_OPERAND (stmt, 1); | |
6de9cd9a | 808 | |
46fbb29c SB |
809 | if (pre_fold_expr) |
810 | { | |
811 | cached_lhs = fold (pre_fold_expr); | |
812 | if (TREE_CODE (cached_lhs) != SSA_NAME | |
813 | && !is_gimple_min_invariant (cached_lhs)) | |
814 | cached_lhs = lookup_avail_expr (stmt, false); | |
815 | } | |
6de9cd9a | 816 | |
f47c96aa AM |
817 | /* Restore the statement's original uses/defs. */ |
818 | i = 0; | |
819 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE | SSA_OP_VUSE) | |
820 | SET_USE (use_p, copy[i++]); | |
6de9cd9a | 821 | |
f47c96aa | 822 | free (copy); |
6de9cd9a DN |
823 | } |
824 | ||
74d2efc7 JL |
825 | /* Record the context sensitive equivalence if we were able |
826 | to simplify this statement. */ | |
827 | if (cached_lhs | |
828 | && (TREE_CODE (cached_lhs) == SSA_NAME | |
829 | || is_gimple_min_invariant (cached_lhs))) | |
830 | record_const_or_copy (TREE_OPERAND (stmt, 0), cached_lhs); | |
6de9cd9a DN |
831 | } |
832 | ||
74d2efc7 JL |
833 | /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm |
834 | will be taken. */ | |
6de9cd9a DN |
835 | if (stmt |
836 | && (TREE_CODE (stmt) == COND_EXPR | |
74d2efc7 JL |
837 | || TREE_CODE (stmt) == GOTO_EXPR |
838 | || TREE_CODE (stmt) == SWITCH_EXPR)) | |
6de9cd9a DN |
839 | { |
840 | tree cond, cached_lhs; | |
6de9cd9a DN |
841 | |
842 | /* Now temporarily cprop the operands and try to find the resulting | |
843 | expression in the hash tables. */ | |
844 | if (TREE_CODE (stmt) == COND_EXPR) | |
0e0ed594 JL |
845 | { |
846 | canonicalize_comparison (stmt); | |
847 | cond = COND_EXPR_COND (stmt); | |
848 | } | |
be477406 JL |
849 | else if (TREE_CODE (stmt) == GOTO_EXPR) |
850 | cond = GOTO_DESTINATION (stmt); | |
6de9cd9a DN |
851 | else |
852 | cond = SWITCH_COND (stmt); | |
853 | ||
6615c446 | 854 | if (COMPARISON_CLASS_P (cond)) |
6de9cd9a DN |
855 | { |
856 | tree dummy_cond, op0, op1; | |
857 | enum tree_code cond_code; | |
858 | ||
859 | op0 = TREE_OPERAND (cond, 0); | |
860 | op1 = TREE_OPERAND (cond, 1); | |
861 | cond_code = TREE_CODE (cond); | |
862 | ||
863 | /* Get the current value of both operands. */ | |
864 | if (TREE_CODE (op0) == SSA_NAME) | |
865 | { | |
3aecd08b JL |
866 | tree tmp = SSA_NAME_VALUE (op0); |
867 | if (tmp && TREE_CODE (tmp) != VALUE_HANDLE) | |
6de9cd9a DN |
868 | op0 = tmp; |
869 | } | |
870 | ||
871 | if (TREE_CODE (op1) == SSA_NAME) | |
872 | { | |
3aecd08b JL |
873 | tree tmp = SSA_NAME_VALUE (op1); |
874 | if (tmp && TREE_CODE (tmp) != VALUE_HANDLE) | |
6de9cd9a DN |
875 | op1 = tmp; |
876 | } | |
877 | ||
878 | /* Stuff the operator and operands into our dummy conditional | |
879 | expression, creating the dummy conditional if necessary. */ | |
e1111e8e | 880 | dummy_cond = (tree) walk_data->global_data; |
6de9cd9a DN |
881 | if (! dummy_cond) |
882 | { | |
b4257cfc RG |
883 | dummy_cond = build2 (cond_code, boolean_type_node, op0, op1); |
884 | dummy_cond = build3 (COND_EXPR, void_type_node, | |
885 | dummy_cond, NULL_TREE, NULL_TREE); | |
6de9cd9a DN |
886 | walk_data->global_data = dummy_cond; |
887 | } | |
888 | else | |
889 | { | |
a6234684 KH |
890 | TREE_SET_CODE (COND_EXPR_COND (dummy_cond), cond_code); |
891 | TREE_OPERAND (COND_EXPR_COND (dummy_cond), 0) = op0; | |
892 | TREE_OPERAND (COND_EXPR_COND (dummy_cond), 1) = op1; | |
6de9cd9a DN |
893 | } |
894 | ||
895 | /* If the conditional folds to an invariant, then we are done, | |
896 | otherwise look it up in the hash tables. */ | |
897 | cached_lhs = local_fold (COND_EXPR_COND (dummy_cond)); | |
898 | if (! is_gimple_min_invariant (cached_lhs)) | |
6de9cd9a | 899 | { |
26673217 KH |
900 | cached_lhs = lookup_avail_expr (dummy_cond, false); |
901 | if (!cached_lhs || ! is_gimple_min_invariant (cached_lhs)) | |
902 | cached_lhs = simplify_cond_and_lookup_avail_expr (dummy_cond, | |
903 | NULL, | |
904 | false); | |
6de9cd9a DN |
905 | } |
906 | } | |
907 | /* We can have conditionals which just test the state of a | |
908 | variable rather than use a relational operator. These are | |
909 | simpler to handle. */ | |
910 | else if (TREE_CODE (cond) == SSA_NAME) | |
911 | { | |
912 | cached_lhs = cond; | |
3aecd08b | 913 | cached_lhs = SSA_NAME_VALUE (cached_lhs); |
6de9cd9a | 914 | if (cached_lhs && ! is_gimple_min_invariant (cached_lhs)) |
74d2efc7 | 915 | cached_lhs = NULL; |
6de9cd9a DN |
916 | } |
917 | else | |
48732f23 | 918 | cached_lhs = lookup_avail_expr (stmt, false); |
6de9cd9a DN |
919 | |
920 | if (cached_lhs) | |
921 | { | |
922 | edge taken_edge = find_taken_edge (e->dest, cached_lhs); | |
923 | basic_block dest = (taken_edge ? taken_edge->dest : NULL); | |
924 | ||
8a78744f | 925 | if (dest == e->dest) |
6de9cd9a DN |
926 | return; |
927 | ||
928 | /* If we have a known destination for the conditional, then | |
929 | we can perform this optimization, which saves at least one | |
930 | conditional jump each time it applies since we get to | |
471854f8 | 931 | bypass the conditional at our original destination. */ |
6de9cd9a DN |
932 | if (dest) |
933 | { | |
efea75f9 JL |
934 | struct edge_info *edge_info; |
935 | ||
efea75f9 | 936 | if (e->aux) |
e1111e8e | 937 | edge_info = (struct edge_info *) e->aux; |
efea75f9 JL |
938 | else |
939 | edge_info = allocate_edge_info (e); | |
940 | edge_info->redirection_target = taken_edge; | |
4aab792d | 941 | bitmap_set_bit (threaded_blocks, e->dest->index); |
6de9cd9a DN |
942 | } |
943 | } | |
944 | } | |
945 | } | |
946 | ||
947 | ||
6de9cd9a DN |
948 | /* Initialize local stacks for this optimizer and record equivalences |
949 | upon entry to BB. Equivalences can come from the edge traversed to | |
950 | reach BB or they may come from PHI nodes at the start of BB. */ | |
951 | ||
952 | static void | |
efea75f9 JL |
953 | dom_opt_initialize_block (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED, |
954 | basic_block bb) | |
6de9cd9a DN |
955 | { |
956 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
957 | fprintf (dump_file, "\n\nOptimizing block #%d\n\n", bb->index); | |
958 | ||
9fae925b JL |
959 | /* Push a marker on the stacks of local information so that we know how |
960 | far to unwind when we finalize this block. */ | |
d4e6fecb | 961 | VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE); |
d4e6fecb NS |
962 | VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE); |
963 | VEC_safe_push (tree, heap, nonzero_vars_stack, NULL_TREE); | |
964 | VEC_safe_push (tree, heap, vrp_variables_stack, NULL_TREE); | |
48732f23 | 965 | |
efea75f9 | 966 | record_equivalences_from_incoming_edge (bb); |
6de9cd9a DN |
967 | |
968 | /* PHI nodes can create equivalences too. */ | |
efea75f9 | 969 | record_equivalences_from_phis (bb); |
6de9cd9a DN |
970 | } |
971 | ||
972 | /* Given an expression EXPR (a relational expression or a statement), | |
206048bd | 973 | initialize the hash table element pointed to by ELEMENT. */ |
6de9cd9a DN |
974 | |
975 | static void | |
976 | initialize_hash_element (tree expr, tree lhs, struct expr_hash_elt *element) | |
977 | { | |
978 | /* Hash table elements may be based on conditional expressions or statements. | |
979 | ||
980 | For the former case, we have no annotation and we want to hash the | |
981 | conditional expression. In the latter case we have an annotation and | |
982 | we want to record the expression the statement evaluates. */ | |
6615c446 | 983 | if (COMPARISON_CLASS_P (expr) || TREE_CODE (expr) == TRUTH_NOT_EXPR) |
6de9cd9a | 984 | { |
f47c96aa | 985 | element->stmt = NULL; |
6de9cd9a DN |
986 | element->rhs = expr; |
987 | } | |
988 | else if (TREE_CODE (expr) == COND_EXPR) | |
989 | { | |
f47c96aa | 990 | element->stmt = expr; |
6de9cd9a DN |
991 | element->rhs = COND_EXPR_COND (expr); |
992 | } | |
993 | else if (TREE_CODE (expr) == SWITCH_EXPR) | |
994 | { | |
f47c96aa | 995 | element->stmt = expr; |
6de9cd9a DN |
996 | element->rhs = SWITCH_COND (expr); |
997 | } | |
998 | else if (TREE_CODE (expr) == RETURN_EXPR && TREE_OPERAND (expr, 0)) | |
999 | { | |
f47c96aa | 1000 | element->stmt = expr; |
6de9cd9a DN |
1001 | element->rhs = TREE_OPERAND (TREE_OPERAND (expr, 0), 1); |
1002 | } | |
74d2efc7 JL |
1003 | else if (TREE_CODE (expr) == GOTO_EXPR) |
1004 | { | |
f47c96aa | 1005 | element->stmt = expr; |
74d2efc7 JL |
1006 | element->rhs = GOTO_DESTINATION (expr); |
1007 | } | |
6de9cd9a DN |
1008 | else |
1009 | { | |
f47c96aa | 1010 | element->stmt = expr; |
6de9cd9a DN |
1011 | element->rhs = TREE_OPERAND (expr, 1); |
1012 | } | |
1013 | ||
1014 | element->lhs = lhs; | |
1015 | element->hash = avail_expr_hash (element); | |
1016 | } | |
1017 | ||
1018 | /* Remove all the expressions in LOCALS from TABLE, stopping when there are | |
1019 | LIMIT entries left in LOCALs. */ | |
1020 | ||
1021 | static void | |
48732f23 | 1022 | remove_local_expressions_from_table (void) |
6de9cd9a | 1023 | { |
6de9cd9a | 1024 | /* Remove all the expressions made available in this block. */ |
d4e6fecb | 1025 | while (VEC_length (tree, avail_exprs_stack) > 0) |
6de9cd9a DN |
1026 | { |
1027 | struct expr_hash_elt element; | |
d4e6fecb | 1028 | tree expr = VEC_pop (tree, avail_exprs_stack); |
48732f23 JL |
1029 | |
1030 | if (expr == NULL_TREE) | |
1031 | break; | |
6de9cd9a DN |
1032 | |
1033 | initialize_hash_element (expr, NULL, &element); | |
48732f23 | 1034 | htab_remove_elt_with_hash (avail_exprs, &element, element.hash); |
6de9cd9a DN |
1035 | } |
1036 | } | |
1037 | ||
1038 | /* Use the SSA_NAMES in LOCALS to restore TABLE to its original | |
1ea7e6ad | 1039 | state, stopping when there are LIMIT entries left in LOCALs. */ |
6de9cd9a DN |
1040 | |
1041 | static void | |
76fd4fd7 | 1042 | restore_nonzero_vars_to_original_value (void) |
6de9cd9a | 1043 | { |
d4e6fecb | 1044 | while (VEC_length (tree, nonzero_vars_stack) > 0) |
6de9cd9a | 1045 | { |
d4e6fecb | 1046 | tree name = VEC_pop (tree, nonzero_vars_stack); |
fdabe5c2 JL |
1047 | |
1048 | if (name == NULL) | |
1049 | break; | |
1050 | ||
1051 | bitmap_clear_bit (nonzero_vars, SSA_NAME_VERSION (name)); | |
6de9cd9a DN |
1052 | } |
1053 | } | |
1054 | ||
b5fefcf6 JL |
1055 | /* Use the source/dest pairs in CONST_AND_COPIES_STACK to restore |
1056 | CONST_AND_COPIES to its original state, stopping when we hit a | |
1057 | NULL marker. */ | |
6de9cd9a DN |
1058 | |
1059 | static void | |
b5fefcf6 | 1060 | restore_vars_to_original_value (void) |
6de9cd9a | 1061 | { |
d4e6fecb | 1062 | while (VEC_length (tree, const_and_copies_stack) > 0) |
6de9cd9a DN |
1063 | { |
1064 | tree prev_value, dest; | |
1065 | ||
d4e6fecb | 1066 | dest = VEC_pop (tree, const_and_copies_stack); |
6de9cd9a | 1067 | |
b5fefcf6 JL |
1068 | if (dest == NULL) |
1069 | break; | |
1070 | ||
d4e6fecb | 1071 | prev_value = VEC_pop (tree, const_and_copies_stack); |
3aecd08b | 1072 | SSA_NAME_VALUE (dest) = prev_value; |
6de9cd9a DN |
1073 | } |
1074 | } | |
1075 | ||
6de9cd9a DN |
1076 | /* We have finished processing the dominator children of BB, perform |
1077 | any finalization actions in preparation for leaving this node in | |
1078 | the dominator tree. */ | |
1079 | ||
1080 | static void | |
1081 | dom_opt_finalize_block (struct dom_walk_data *walk_data, basic_block bb) | |
1082 | { | |
6de9cd9a DN |
1083 | tree last; |
1084 | ||
3e352c00 JL |
1085 | /* If we have an outgoing edge to a block with multiple incoming and |
1086 | outgoing edges, then we may be able to thread the edge. ie, we | |
1087 | may be able to statically determine which of the outgoing edges | |
1088 | will be traversed when the incoming edge from BB is traversed. */ | |
c5cbcccf ZD |
1089 | if (single_succ_p (bb) |
1090 | && (single_succ_edge (bb)->flags & EDGE_ABNORMAL) == 0 | |
3e352c00 JL |
1091 | && !single_pred_p (single_succ (bb)) |
1092 | && !single_succ_p (single_succ (bb))) | |
6de9cd9a DN |
1093 | |
1094 | { | |
c5cbcccf | 1095 | thread_across_edge (walk_data, single_succ_edge (bb)); |
6de9cd9a DN |
1096 | } |
1097 | else if ((last = last_stmt (bb)) | |
1098 | && TREE_CODE (last) == COND_EXPR | |
6615c446 | 1099 | && (COMPARISON_CLASS_P (COND_EXPR_COND (last)) |
6de9cd9a | 1100 | || TREE_CODE (COND_EXPR_COND (last)) == SSA_NAME) |
628f6a4e BE |
1101 | && EDGE_COUNT (bb->succs) == 2 |
1102 | && (EDGE_SUCC (bb, 0)->flags & EDGE_ABNORMAL) == 0 | |
1103 | && (EDGE_SUCC (bb, 1)->flags & EDGE_ABNORMAL) == 0) | |
6de9cd9a DN |
1104 | { |
1105 | edge true_edge, false_edge; | |
6de9cd9a DN |
1106 | |
1107 | extract_true_false_edges_from_block (bb, &true_edge, &false_edge); | |
1108 | ||
3e352c00 JL |
1109 | /* Only try to thread the edge if it reaches a target block with |
1110 | more than one predecessor and more than one successor. */ | |
1111 | if (!single_pred_p (true_edge->dest) && !single_succ_p (true_edge->dest)) | |
6de9cd9a | 1112 | { |
efea75f9 JL |
1113 | struct edge_info *edge_info; |
1114 | unsigned int i; | |
1115 | ||
48732f23 JL |
1116 | /* Push a marker onto the available expression stack so that we |
1117 | unwind any expressions related to the TRUE arm before processing | |
1118 | the false arm below. */ | |
d4e6fecb | 1119 | VEC_safe_push (tree, heap, avail_exprs_stack, NULL_TREE); |
d4e6fecb | 1120 | VEC_safe_push (tree, heap, const_and_copies_stack, NULL_TREE); |
48732f23 | 1121 | |
e1111e8e | 1122 | edge_info = (struct edge_info *) true_edge->aux; |
efea75f9 JL |
1123 | |
1124 | /* If we have info associated with this edge, record it into | |
1125 | our equivalency tables. */ | |
1126 | if (edge_info) | |
6de9cd9a | 1127 | { |
efea75f9 JL |
1128 | tree *cond_equivalences = edge_info->cond_equivalences; |
1129 | tree lhs = edge_info->lhs; | |
1130 | tree rhs = edge_info->rhs; | |
1131 | ||
74d2efc7 JL |
1132 | /* If we have a simple NAME = VALUE equivalency record it. */ |
1133 | if (lhs && TREE_CODE (lhs) == SSA_NAME) | |
efea75f9 JL |
1134 | record_const_or_copy (lhs, rhs); |
1135 | ||
1136 | /* If we have 0 = COND or 1 = COND equivalences, record them | |
1137 | into our expression hash tables. */ | |
1138 | if (cond_equivalences) | |
1139 | for (i = 0; i < edge_info->max_cond_equivalences; i += 2) | |
1140 | { | |
1141 | tree expr = cond_equivalences[i]; | |
1142 | tree value = cond_equivalences[i + 1]; | |
1143 | ||
1144 | record_cond (expr, value); | |
1145 | } | |
6de9cd9a | 1146 | } |
6de9cd9a DN |
1147 | |
1148 | /* Now thread the edge. */ | |
1149 | thread_across_edge (walk_data, true_edge); | |
1150 | ||
1151 | /* And restore the various tables to their state before | |
1152 | we threaded this edge. */ | |
48732f23 | 1153 | remove_local_expressions_from_table (); |
b5fefcf6 | 1154 | restore_vars_to_original_value (); |
6de9cd9a DN |
1155 | } |
1156 | ||
1157 | /* Similarly for the ELSE arm. */ | |
3e352c00 | 1158 | if (!single_pred_p (false_edge->dest) && !single_succ_p (false_edge->dest)) |
6de9cd9a | 1159 | { |
efea75f9 JL |
1160 | struct edge_info *edge_info; |
1161 | unsigned int i; | |
1162 | ||
e1111e8e | 1163 | edge_info = (struct edge_info *) false_edge->aux; |
efea75f9 JL |
1164 | |
1165 | /* If we have info associated with this edge, record it into | |
1166 | our equivalency tables. */ | |
1167 | if (edge_info) | |
6de9cd9a | 1168 | { |
efea75f9 JL |
1169 | tree *cond_equivalences = edge_info->cond_equivalences; |
1170 | tree lhs = edge_info->lhs; | |
1171 | tree rhs = edge_info->rhs; | |
1172 | ||
74d2efc7 JL |
1173 | /* If we have a simple NAME = VALUE equivalency record it. */ |
1174 | if (lhs && TREE_CODE (lhs) == SSA_NAME) | |
efea75f9 JL |
1175 | record_const_or_copy (lhs, rhs); |
1176 | ||
1177 | /* If we have 0 = COND or 1 = COND equivalences, record them | |
1178 | into our expression hash tables. */ | |
1179 | if (cond_equivalences) | |
1180 | for (i = 0; i < edge_info->max_cond_equivalences; i += 2) | |
1181 | { | |
1182 | tree expr = cond_equivalences[i]; | |
1183 | tree value = cond_equivalences[i + 1]; | |
1184 | ||
1185 | record_cond (expr, value); | |
1186 | } | |
6de9cd9a | 1187 | } |
6de9cd9a DN |
1188 | |
1189 | thread_across_edge (walk_data, false_edge); | |
1190 | ||
1191 | /* No need to remove local expressions from our tables | |
1192 | or restore vars to their original value as that will | |
1193 | be done immediately below. */ | |
1194 | } | |
1195 | } | |
1196 | ||
48732f23 | 1197 | remove_local_expressions_from_table (); |
fdabe5c2 | 1198 | restore_nonzero_vars_to_original_value (); |
b5fefcf6 | 1199 | restore_vars_to_original_value (); |
6de9cd9a DN |
1200 | |
1201 | /* Remove VRP records associated with this basic block. They are no | |
1202 | longer valid. | |
1203 | ||
1204 | To be efficient, we note which variables have had their values | |
1205 | constrained in this block. So walk over each variable in the | |
1206 | VRP_VARIABLEs array. */ | |
d4e6fecb | 1207 | while (VEC_length (tree, vrp_variables_stack) > 0) |
6de9cd9a | 1208 | { |
d4e6fecb | 1209 | tree var = VEC_pop (tree, vrp_variables_stack); |
b8545fbf | 1210 | struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p; |
23530866 | 1211 | void **slot; |
6de9cd9a DN |
1212 | |
1213 | /* Each variable has a stack of value range records. We want to | |
1214 | invalidate those associated with our basic block. So we walk | |
1215 | the array backwards popping off records associated with our | |
1216 | block. Once we hit a record not associated with our block | |
1217 | we are done. */ | |
8184759d | 1218 | VEC(vrp_element_p,heap) **var_vrp_records; |
fdabe5c2 | 1219 | |
fdabe5c2 JL |
1220 | if (var == NULL) |
1221 | break; | |
6de9cd9a | 1222 | |
23530866 JL |
1223 | vrp_hash_elt.var = var; |
1224 | vrp_hash_elt.records = NULL; | |
1225 | ||
1226 | slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT); | |
1227 | ||
b8545fbf | 1228 | vrp_hash_elt_p = (struct vrp_hash_elt *) *slot; |
8184759d | 1229 | var_vrp_records = &vrp_hash_elt_p->records; |
b8545fbf | 1230 | |
8184759d | 1231 | while (VEC_length (vrp_element_p, *var_vrp_records) > 0) |
6de9cd9a DN |
1232 | { |
1233 | struct vrp_element *element | |
8184759d | 1234 | = VEC_last (vrp_element_p, *var_vrp_records); |
6de9cd9a DN |
1235 | |
1236 | if (element->bb != bb) | |
1237 | break; | |
1238 | ||
8184759d | 1239 | VEC_pop (vrp_element_p, *var_vrp_records); |
6de9cd9a | 1240 | } |
6de9cd9a DN |
1241 | } |
1242 | ||
a6e1aa26 JL |
1243 | /* If we queued any statements to rescan in this block, then |
1244 | go ahead and rescan them now. */ | |
d4e6fecb | 1245 | while (VEC_length (tree, stmts_to_rescan) > 0) |
6de9cd9a | 1246 | { |
d4e6fecb | 1247 | tree stmt = VEC_last (tree, stmts_to_rescan); |
a6e1aa26 JL |
1248 | basic_block stmt_bb = bb_for_stmt (stmt); |
1249 | ||
1250 | if (stmt_bb != bb) | |
1251 | break; | |
1252 | ||
d4e6fecb | 1253 | VEC_pop (tree, stmts_to_rescan); |
0bca51f0 | 1254 | mark_new_vars_to_rename (stmt); |
6de9cd9a DN |
1255 | } |
1256 | } | |
1257 | ||
1258 | /* PHI nodes can create equivalences too. | |
1259 | ||
1260 | Ignoring any alternatives which are the same as the result, if | |
1261 | all the alternatives are equal, then the PHI node creates an | |
dd747311 JL |
1262 | equivalence. |
1263 | ||
1264 | Additionally, if all the PHI alternatives are known to have a nonzero | |
1265 | value, then the result of this PHI is known to have a nonzero value, | |
1266 | even if we do not know its exact value. */ | |
1267 | ||
6de9cd9a | 1268 | static void |
efea75f9 | 1269 | record_equivalences_from_phis (basic_block bb) |
6de9cd9a | 1270 | { |
6de9cd9a DN |
1271 | tree phi; |
1272 | ||
17192884 | 1273 | for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi)) |
6de9cd9a DN |
1274 | { |
1275 | tree lhs = PHI_RESULT (phi); | |
1276 | tree rhs = NULL; | |
1277 | int i; | |
1278 | ||
1279 | for (i = 0; i < PHI_NUM_ARGS (phi); i++) | |
1280 | { | |
1281 | tree t = PHI_ARG_DEF (phi, i); | |
1282 | ||
6e38fea3 KH |
1283 | /* Ignore alternatives which are the same as our LHS. Since |
1284 | LHS is a PHI_RESULT, it is known to be a SSA_NAME, so we | |
1285 | can simply compare pointers. */ | |
073b8140 | 1286 | if (lhs == t) |
a18428f3 KH |
1287 | continue; |
1288 | ||
1289 | /* If we have not processed an alternative yet, then set | |
1290 | RHS to this alternative. */ | |
1291 | if (rhs == NULL) | |
1292 | rhs = t; | |
1293 | /* If we have processed an alternative (stored in RHS), then | |
1294 | see if it is equal to this one. If it isn't, then stop | |
1295 | the search. */ | |
1296 | else if (! operand_equal_for_phi_arg_p (rhs, t)) | |
6de9cd9a DN |
1297 | break; |
1298 | } | |
1299 | ||
1300 | /* If we had no interesting alternatives, then all the RHS alternatives | |
1301 | must have been the same as LHS. */ | |
1302 | if (!rhs) | |
1303 | rhs = lhs; | |
1304 | ||
1305 | /* If we managed to iterate through each PHI alternative without | |
1306 | breaking out of the loop, then we have a PHI which may create | |
1307 | a useful equivalence. We do not need to record unwind data for | |
1308 | this, since this is a true assignment and not an equivalence | |
1ea7e6ad | 1309 | inferred from a comparison. All uses of this ssa name are dominated |
6de9cd9a DN |
1310 | by this assignment, so unwinding just costs time and space. */ |
1311 | if (i == PHI_NUM_ARGS (phi) | |
1312 | && may_propagate_copy (lhs, rhs)) | |
3aecd08b | 1313 | SSA_NAME_VALUE (lhs) = rhs; |
6de9cd9a | 1314 | |
dd747311 JL |
1315 | /* Now see if we know anything about the nonzero property for the |
1316 | result of this PHI. */ | |
1317 | for (i = 0; i < PHI_NUM_ARGS (phi); i++) | |
1318 | { | |
1319 | if (!PHI_ARG_NONZERO (phi, i)) | |
1320 | break; | |
1321 | } | |
1322 | ||
1323 | if (i == PHI_NUM_ARGS (phi)) | |
1324 | bitmap_set_bit (nonzero_vars, SSA_NAME_VERSION (PHI_RESULT (phi))); | |
6de9cd9a DN |
1325 | } |
1326 | } | |
1327 | ||
28c008bb JL |
1328 | /* Ignoring loop backedges, if BB has precisely one incoming edge then |
1329 | return that edge. Otherwise return NULL. */ | |
1330 | static edge | |
1331 | single_incoming_edge_ignoring_loop_edges (basic_block bb) | |
1332 | { | |
1333 | edge retval = NULL; | |
1334 | edge e; | |
628f6a4e | 1335 | edge_iterator ei; |
28c008bb | 1336 | |
628f6a4e | 1337 | FOR_EACH_EDGE (e, ei, bb->preds) |
28c008bb JL |
1338 | { |
1339 | /* A loop back edge can be identified by the destination of | |
1340 | the edge dominating the source of the edge. */ | |
1341 | if (dominated_by_p (CDI_DOMINATORS, e->src, e->dest)) | |
1342 | continue; | |
1343 | ||
1344 | /* If we have already seen a non-loop edge, then we must have | |
1345 | multiple incoming non-loop edges and thus we return NULL. */ | |
1346 | if (retval) | |
1347 | return NULL; | |
1348 | ||
1349 | /* This is the first non-loop incoming edge we have found. Record | |
1350 | it. */ | |
1351 | retval = e; | |
1352 | } | |
1353 | ||
1354 | return retval; | |
1355 | } | |
1356 | ||
6de9cd9a DN |
1357 | /* Record any equivalences created by the incoming edge to BB. If BB |
1358 | has more than one incoming edge, then no equivalence is created. */ | |
1359 | ||
1360 | static void | |
efea75f9 | 1361 | record_equivalences_from_incoming_edge (basic_block bb) |
6de9cd9a | 1362 | { |
efea75f9 | 1363 | edge e; |
6de9cd9a | 1364 | basic_block parent; |
efea75f9 | 1365 | struct edge_info *edge_info; |
6de9cd9a | 1366 | |
35fd3193 | 1367 | /* If our parent block ended with a control statement, then we may be |
6de9cd9a DN |
1368 | able to record some equivalences based on which outgoing edge from |
1369 | the parent was followed. */ | |
1370 | parent = get_immediate_dominator (CDI_DOMINATORS, bb); | |
6de9cd9a | 1371 | |
efea75f9 | 1372 | e = single_incoming_edge_ignoring_loop_edges (bb); |
6de9cd9a | 1373 | |
efea75f9 JL |
1374 | /* If we had a single incoming edge from our parent block, then enter |
1375 | any data associated with the edge into our tables. */ | |
1376 | if (e && e->src == parent) | |
6de9cd9a | 1377 | { |
efea75f9 | 1378 | unsigned int i; |
6de9cd9a | 1379 | |
e1111e8e | 1380 | edge_info = (struct edge_info *) e->aux; |
6de9cd9a | 1381 | |
efea75f9 | 1382 | if (edge_info) |
6de9cd9a | 1383 | { |
efea75f9 JL |
1384 | tree lhs = edge_info->lhs; |
1385 | tree rhs = edge_info->rhs; | |
1386 | tree *cond_equivalences = edge_info->cond_equivalences; | |
1387 | ||
1388 | if (lhs) | |
1389 | record_equality (lhs, rhs); | |
1390 | ||
1391 | if (cond_equivalences) | |
6de9cd9a | 1392 | { |
efea75f9 JL |
1393 | bool recorded_range = false; |
1394 | for (i = 0; i < edge_info->max_cond_equivalences; i += 2) | |
6de9cd9a | 1395 | { |
efea75f9 JL |
1396 | tree expr = cond_equivalences[i]; |
1397 | tree value = cond_equivalences[i + 1]; | |
1398 | ||
1399 | record_cond (expr, value); | |
1400 | ||
1401 | /* For the first true equivalence, record range | |
1402 | information. We only do this for the first | |
1403 | true equivalence as it should dominate any | |
1404 | later true equivalences. */ | |
1405 | if (! recorded_range | |
1406 | && COMPARISON_CLASS_P (expr) | |
1407 | && value == boolean_true_node | |
1408 | && TREE_CONSTANT (TREE_OPERAND (expr, 1))) | |
1409 | { | |
1410 | record_range (expr, bb); | |
1411 | recorded_range = true; | |
1412 | } | |
6de9cd9a DN |
1413 | } |
1414 | } | |
6de9cd9a DN |
1415 | } |
1416 | } | |
6de9cd9a DN |
1417 | } |
1418 | ||
1419 | /* Dump SSA statistics on FILE. */ | |
1420 | ||
1421 | void | |
1422 | dump_dominator_optimization_stats (FILE *file) | |
1423 | { | |
1424 | long n_exprs; | |
1425 | ||
1426 | fprintf (file, "Total number of statements: %6ld\n\n", | |
1427 | opt_stats.num_stmts); | |
1428 | fprintf (file, "Exprs considered for dominator optimizations: %6ld\n", | |
1429 | opt_stats.num_exprs_considered); | |
1430 | ||
1431 | n_exprs = opt_stats.num_exprs_considered; | |
1432 | if (n_exprs == 0) | |
1433 | n_exprs = 1; | |
1434 | ||
1435 | fprintf (file, " Redundant expressions eliminated: %6ld (%.0f%%)\n", | |
1436 | opt_stats.num_re, PERCENT (opt_stats.num_re, | |
1437 | n_exprs)); | |
0bca51f0 DN |
1438 | fprintf (file, " Constants propagated: %6ld\n", |
1439 | opt_stats.num_const_prop); | |
1440 | fprintf (file, " Copies propagated: %6ld\n", | |
1441 | opt_stats.num_copy_prop); | |
6de9cd9a | 1442 | |
6a2cf9dc DN |
1443 | fprintf (file, "\nTotal number of DOM iterations: %6ld\n", |
1444 | opt_stats.num_iterations); | |
1445 | ||
6de9cd9a DN |
1446 | fprintf (file, "\nHash table statistics:\n"); |
1447 | ||
1448 | fprintf (file, " avail_exprs: "); | |
1449 | htab_statistics (file, avail_exprs); | |
1450 | } | |
1451 | ||
1452 | ||
1453 | /* Dump SSA statistics on stderr. */ | |
1454 | ||
1455 | void | |
1456 | debug_dominator_optimization_stats (void) | |
1457 | { | |
1458 | dump_dominator_optimization_stats (stderr); | |
1459 | } | |
1460 | ||
1461 | ||
1462 | /* Dump statistics for the hash table HTAB. */ | |
1463 | ||
1464 | static void | |
1465 | htab_statistics (FILE *file, htab_t htab) | |
1466 | { | |
1467 | fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n", | |
1468 | (long) htab_size (htab), | |
1469 | (long) htab_elements (htab), | |
1470 | htab_collisions (htab)); | |
1471 | } | |
1472 | ||
1473 | /* Record the fact that VAR has a nonzero value, though we may not know | |
1474 | its exact value. Note that if VAR is already known to have a nonzero | |
1475 | value, then we do nothing. */ | |
1476 | ||
1477 | static void | |
fdabe5c2 | 1478 | record_var_is_nonzero (tree var) |
6de9cd9a DN |
1479 | { |
1480 | int indx = SSA_NAME_VERSION (var); | |
1481 | ||
1482 | if (bitmap_bit_p (nonzero_vars, indx)) | |
1483 | return; | |
1484 | ||
1485 | /* Mark it in the global table. */ | |
1486 | bitmap_set_bit (nonzero_vars, indx); | |
1487 | ||
1488 | /* Record this SSA_NAME so that we can reset the global table | |
1489 | when we leave this block. */ | |
d4e6fecb | 1490 | VEC_safe_push (tree, heap, nonzero_vars_stack, var); |
6de9cd9a DN |
1491 | } |
1492 | ||
1493 | /* Enter a statement into the true/false expression hash table indicating | |
1494 | that the condition COND has the value VALUE. */ | |
1495 | ||
1496 | static void | |
48732f23 | 1497 | record_cond (tree cond, tree value) |
6de9cd9a | 1498 | { |
e1111e8e | 1499 | struct expr_hash_elt *element = XCNEW (struct expr_hash_elt); |
6de9cd9a DN |
1500 | void **slot; |
1501 | ||
1502 | initialize_hash_element (cond, value, element); | |
1503 | ||
1504 | slot = htab_find_slot_with_hash (avail_exprs, (void *)element, | |
5746637c | 1505 | element->hash, INSERT); |
6de9cd9a DN |
1506 | if (*slot == NULL) |
1507 | { | |
1508 | *slot = (void *) element; | |
d4e6fecb | 1509 | VEC_safe_push (tree, heap, avail_exprs_stack, cond); |
6de9cd9a DN |
1510 | } |
1511 | else | |
1512 | free (element); | |
1513 | } | |
1514 | ||
efea75f9 JL |
1515 | /* Build a new conditional using NEW_CODE, OP0 and OP1 and store |
1516 | the new conditional into *p, then store a boolean_true_node | |
3f117656 | 1517 | into *(p + 1). */ |
efea75f9 JL |
1518 | |
1519 | static void | |
1520 | build_and_record_new_cond (enum tree_code new_code, tree op0, tree op1, tree *p) | |
1521 | { | |
1522 | *p = build2 (new_code, boolean_type_node, op0, op1); | |
1523 | p++; | |
1524 | *p = boolean_true_node; | |
1525 | } | |
1526 | ||
1527 | /* Record that COND is true and INVERTED is false into the edge information | |
1528 | structure. Also record that any conditions dominated by COND are true | |
1529 | as well. | |
d2d8936f JL |
1530 | |
1531 | For example, if a < b is true, then a <= b must also be true. */ | |
1532 | ||
1533 | static void | |
efea75f9 | 1534 | record_conditions (struct edge_info *edge_info, tree cond, tree inverted) |
d2d8936f | 1535 | { |
efea75f9 JL |
1536 | tree op0, op1; |
1537 | ||
1538 | if (!COMPARISON_CLASS_P (cond)) | |
1539 | return; | |
1540 | ||
1541 | op0 = TREE_OPERAND (cond, 0); | |
1542 | op1 = TREE_OPERAND (cond, 1); | |
1543 | ||
d2d8936f JL |
1544 | switch (TREE_CODE (cond)) |
1545 | { | |
1546 | case LT_EXPR: | |
d2d8936f | 1547 | case GT_EXPR: |
efea75f9 | 1548 | edge_info->max_cond_equivalences = 12; |
e1111e8e | 1549 | edge_info->cond_equivalences = XNEWVEC (tree, 12); |
efea75f9 JL |
1550 | build_and_record_new_cond ((TREE_CODE (cond) == LT_EXPR |
1551 | ? LE_EXPR : GE_EXPR), | |
1552 | op0, op1, &edge_info->cond_equivalences[4]); | |
1553 | build_and_record_new_cond (ORDERED_EXPR, op0, op1, | |
1554 | &edge_info->cond_equivalences[6]); | |
1555 | build_and_record_new_cond (NE_EXPR, op0, op1, | |
1556 | &edge_info->cond_equivalences[8]); | |
1557 | build_and_record_new_cond (LTGT_EXPR, op0, op1, | |
1558 | &edge_info->cond_equivalences[10]); | |
d2d8936f JL |
1559 | break; |
1560 | ||
1561 | case GE_EXPR: | |
1562 | case LE_EXPR: | |
efea75f9 | 1563 | edge_info->max_cond_equivalences = 6; |
e1111e8e | 1564 | edge_info->cond_equivalences = XNEWVEC (tree, 6); |
efea75f9 JL |
1565 | build_and_record_new_cond (ORDERED_EXPR, op0, op1, |
1566 | &edge_info->cond_equivalences[4]); | |
d2d8936f JL |
1567 | break; |
1568 | ||
1569 | case EQ_EXPR: | |
efea75f9 | 1570 | edge_info->max_cond_equivalences = 10; |
e1111e8e | 1571 | edge_info->cond_equivalences = XNEWVEC (tree, 10); |
efea75f9 JL |
1572 | build_and_record_new_cond (ORDERED_EXPR, op0, op1, |
1573 | &edge_info->cond_equivalences[4]); | |
1574 | build_and_record_new_cond (LE_EXPR, op0, op1, | |
1575 | &edge_info->cond_equivalences[6]); | |
1576 | build_and_record_new_cond (GE_EXPR, op0, op1, | |
1577 | &edge_info->cond_equivalences[8]); | |
d2d8936f JL |
1578 | break; |
1579 | ||
1580 | case UNORDERED_EXPR: | |
efea75f9 | 1581 | edge_info->max_cond_equivalences = 16; |
e1111e8e | 1582 | edge_info->cond_equivalences = XNEWVEC (tree, 16); |
efea75f9 JL |
1583 | build_and_record_new_cond (NE_EXPR, op0, op1, |
1584 | &edge_info->cond_equivalences[4]); | |
1585 | build_and_record_new_cond (UNLE_EXPR, op0, op1, | |
1586 | &edge_info->cond_equivalences[6]); | |
1587 | build_and_record_new_cond (UNGE_EXPR, op0, op1, | |
1588 | &edge_info->cond_equivalences[8]); | |
1589 | build_and_record_new_cond (UNEQ_EXPR, op0, op1, | |
1590 | &edge_info->cond_equivalences[10]); | |
1591 | build_and_record_new_cond (UNLT_EXPR, op0, op1, | |
1592 | &edge_info->cond_equivalences[12]); | |
1593 | build_and_record_new_cond (UNGT_EXPR, op0, op1, | |
1594 | &edge_info->cond_equivalences[14]); | |
d2d8936f JL |
1595 | break; |
1596 | ||
1597 | case UNLT_EXPR: | |
d2d8936f | 1598 | case UNGT_EXPR: |
efea75f9 | 1599 | edge_info->max_cond_equivalences = 8; |
e1111e8e | 1600 | edge_info->cond_equivalences = XNEWVEC (tree, 8); |
efea75f9 JL |
1601 | build_and_record_new_cond ((TREE_CODE (cond) == UNLT_EXPR |
1602 | ? UNLE_EXPR : UNGE_EXPR), | |
1603 | op0, op1, &edge_info->cond_equivalences[4]); | |
1604 | build_and_record_new_cond (NE_EXPR, op0, op1, | |
1605 | &edge_info->cond_equivalences[6]); | |
d2d8936f JL |
1606 | break; |
1607 | ||
1608 | case UNEQ_EXPR: | |
efea75f9 | 1609 | edge_info->max_cond_equivalences = 8; |
e1111e8e | 1610 | edge_info->cond_equivalences = XNEWVEC (tree, 8); |
efea75f9 JL |
1611 | build_and_record_new_cond (UNLE_EXPR, op0, op1, |
1612 | &edge_info->cond_equivalences[4]); | |
1613 | build_and_record_new_cond (UNGE_EXPR, op0, op1, | |
1614 | &edge_info->cond_equivalences[6]); | |
d2d8936f JL |
1615 | break; |
1616 | ||
1617 | case LTGT_EXPR: | |
efea75f9 | 1618 | edge_info->max_cond_equivalences = 8; |
e1111e8e | 1619 | edge_info->cond_equivalences = XNEWVEC (tree, 8); |
efea75f9 JL |
1620 | build_and_record_new_cond (NE_EXPR, op0, op1, |
1621 | &edge_info->cond_equivalences[4]); | |
1622 | build_and_record_new_cond (ORDERED_EXPR, op0, op1, | |
1623 | &edge_info->cond_equivalences[6]); | |
1624 | break; | |
d2d8936f JL |
1625 | |
1626 | default: | |
efea75f9 | 1627 | edge_info->max_cond_equivalences = 4; |
e1111e8e | 1628 | edge_info->cond_equivalences = XNEWVEC (tree, 4); |
d2d8936f JL |
1629 | break; |
1630 | } | |
efea75f9 JL |
1631 | |
1632 | /* Now store the original true and false conditions into the first | |
1633 | two slots. */ | |
1634 | edge_info->cond_equivalences[0] = cond; | |
1635 | edge_info->cond_equivalences[1] = boolean_true_node; | |
1636 | edge_info->cond_equivalences[2] = inverted; | |
1637 | edge_info->cond_equivalences[3] = boolean_false_node; | |
d2d8936f JL |
1638 | } |
1639 | ||
6de9cd9a DN |
1640 | /* A helper function for record_const_or_copy and record_equality. |
1641 | Do the work of recording the value and undo info. */ | |
1642 | ||
1643 | static void | |
b5fefcf6 | 1644 | record_const_or_copy_1 (tree x, tree y, tree prev_x) |
6de9cd9a | 1645 | { |
3aecd08b | 1646 | SSA_NAME_VALUE (x) = y; |
6de9cd9a | 1647 | |
d4e6fecb NS |
1648 | VEC_reserve (tree, heap, const_and_copies_stack, 2); |
1649 | VEC_quick_push (tree, const_and_copies_stack, prev_x); | |
1650 | VEC_quick_push (tree, const_and_copies_stack, x); | |
6de9cd9a DN |
1651 | } |
1652 | ||
84dd478f DB |
1653 | |
1654 | /* Return the loop depth of the basic block of the defining statement of X. | |
1655 | This number should not be treated as absolutely correct because the loop | |
1656 | information may not be completely up-to-date when dom runs. However, it | |
1657 | will be relatively correct, and as more passes are taught to keep loop info | |
1658 | up to date, the result will become more and more accurate. */ | |
1659 | ||
0bca51f0 | 1660 | int |
84dd478f DB |
1661 | loop_depth_of_name (tree x) |
1662 | { | |
1663 | tree defstmt; | |
1664 | basic_block defbb; | |
1665 | ||
1666 | /* If it's not an SSA_NAME, we have no clue where the definition is. */ | |
1667 | if (TREE_CODE (x) != SSA_NAME) | |
1668 | return 0; | |
1669 | ||
1670 | /* Otherwise return the loop depth of the defining statement's bb. | |
1671 | Note that there may not actually be a bb for this statement, if the | |
1672 | ssa_name is live on entry. */ | |
1673 | defstmt = SSA_NAME_DEF_STMT (x); | |
1674 | defbb = bb_for_stmt (defstmt); | |
1675 | if (!defbb) | |
1676 | return 0; | |
1677 | ||
1678 | return defbb->loop_depth; | |
1679 | } | |
1680 | ||
1681 | ||
6de9cd9a | 1682 | /* Record that X is equal to Y in const_and_copies. Record undo |
ceb7eb8f | 1683 | information in the block-local vector. */ |
6de9cd9a DN |
1684 | |
1685 | static void | |
b5fefcf6 | 1686 | record_const_or_copy (tree x, tree y) |
6de9cd9a | 1687 | { |
3aecd08b | 1688 | tree prev_x = SSA_NAME_VALUE (x); |
6de9cd9a DN |
1689 | |
1690 | if (TREE_CODE (y) == SSA_NAME) | |
1691 | { | |
3aecd08b | 1692 | tree tmp = SSA_NAME_VALUE (y); |
6de9cd9a DN |
1693 | if (tmp) |
1694 | y = tmp; | |
1695 | } | |
1696 | ||
b5fefcf6 | 1697 | record_const_or_copy_1 (x, y, prev_x); |
6de9cd9a DN |
1698 | } |
1699 | ||
1700 | /* Similarly, but assume that X and Y are the two operands of an EQ_EXPR. | |
1701 | This constrains the cases in which we may treat this as assignment. */ | |
1702 | ||
1703 | static void | |
b5fefcf6 | 1704 | record_equality (tree x, tree y) |
6de9cd9a DN |
1705 | { |
1706 | tree prev_x = NULL, prev_y = NULL; | |
1707 | ||
1708 | if (TREE_CODE (x) == SSA_NAME) | |
3aecd08b | 1709 | prev_x = SSA_NAME_VALUE (x); |
6de9cd9a | 1710 | if (TREE_CODE (y) == SSA_NAME) |
3aecd08b | 1711 | prev_y = SSA_NAME_VALUE (y); |
6de9cd9a | 1712 | |
84dd478f DB |
1713 | /* If one of the previous values is invariant, or invariant in more loops |
1714 | (by depth), then use that. | |
6de9cd9a DN |
1715 | Otherwise it doesn't matter which value we choose, just so |
1716 | long as we canonicalize on one value. */ | |
1717 | if (TREE_INVARIANT (y)) | |
1718 | ; | |
84dd478f | 1719 | else if (TREE_INVARIANT (x) || (loop_depth_of_name (x) <= loop_depth_of_name (y))) |
6de9cd9a DN |
1720 | prev_x = x, x = y, y = prev_x, prev_x = prev_y; |
1721 | else if (prev_x && TREE_INVARIANT (prev_x)) | |
1722 | x = y, y = prev_x, prev_x = prev_y; | |
3aecd08b | 1723 | else if (prev_y && TREE_CODE (prev_y) != VALUE_HANDLE) |
6de9cd9a DN |
1724 | y = prev_y; |
1725 | ||
1726 | /* After the swapping, we must have one SSA_NAME. */ | |
1727 | if (TREE_CODE (x) != SSA_NAME) | |
1728 | return; | |
1729 | ||
1730 | /* For IEEE, -0.0 == 0.0, so we don't necessarily know the sign of a | |
1731 | variable compared against zero. If we're honoring signed zeros, | |
1732 | then we cannot record this value unless we know that the value is | |
1ea7e6ad | 1733 | nonzero. */ |
6de9cd9a DN |
1734 | if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (x))) |
1735 | && (TREE_CODE (y) != REAL_CST | |
1736 | || REAL_VALUES_EQUAL (dconst0, TREE_REAL_CST (y)))) | |
1737 | return; | |
1738 | ||
b5fefcf6 | 1739 | record_const_or_copy_1 (x, y, prev_x); |
6de9cd9a DN |
1740 | } |
1741 | ||
01d8c00b FJ |
1742 | /* Return true, if it is ok to do folding of an associative expression. |
1743 | EXP is the tree for the associative expression. */ | |
1744 | ||
1745 | static inline bool | |
1746 | unsafe_associative_fp_binop (tree exp) | |
1747 | { | |
1748 | enum tree_code code = TREE_CODE (exp); | |
1749 | return !(!flag_unsafe_math_optimizations | |
d05f9c39 FJ |
1750 | && (code == MULT_EXPR || code == PLUS_EXPR |
1751 | || code == MINUS_EXPR) | |
01d8c00b FJ |
1752 | && FLOAT_TYPE_P (TREE_TYPE (exp))); |
1753 | } | |
1754 | ||
f67e783f ZD |
1755 | /* Returns true when STMT is a simple iv increment. It detects the |
1756 | following situation: | |
1757 | ||
1758 | i_1 = phi (..., i_2) | |
1759 | i_2 = i_1 +/- ... */ | |
1760 | ||
1761 | static bool | |
1762 | simple_iv_increment_p (tree stmt) | |
1763 | { | |
1764 | tree lhs, rhs, preinc, phi; | |
1765 | unsigned i; | |
1766 | ||
1767 | if (TREE_CODE (stmt) != MODIFY_EXPR) | |
1768 | return false; | |
1769 | ||
1770 | lhs = TREE_OPERAND (stmt, 0); | |
1771 | if (TREE_CODE (lhs) != SSA_NAME) | |
1772 | return false; | |
1773 | ||
1774 | rhs = TREE_OPERAND (stmt, 1); | |
1775 | ||
1776 | if (TREE_CODE (rhs) != PLUS_EXPR | |
1777 | && TREE_CODE (rhs) != MINUS_EXPR) | |
1778 | return false; | |
1779 | ||
1780 | preinc = TREE_OPERAND (rhs, 0); | |
1781 | if (TREE_CODE (preinc) != SSA_NAME) | |
1782 | return false; | |
1783 | ||
1784 | phi = SSA_NAME_DEF_STMT (preinc); | |
1785 | if (TREE_CODE (phi) != PHI_NODE) | |
1786 | return false; | |
1787 | ||
1788 | for (i = 0; i < (unsigned) PHI_NUM_ARGS (phi); i++) | |
1789 | if (PHI_ARG_DEF (phi, i) == lhs) | |
1790 | return true; | |
1791 | ||
1792 | return false; | |
1793 | } | |
1794 | ||
6de9cd9a DN |
1795 | /* COND is a condition of the form: |
1796 | ||
1797 | x == const or x != const | |
1798 | ||
1799 | Look back to x's defining statement and see if x is defined as | |
1800 | ||
1801 | x = (type) y; | |
1802 | ||
1803 | If const is unchanged if we convert it to type, then we can build | |
1804 | the equivalent expression: | |
1805 | ||
1806 | ||
1807 | y == const or y != const | |
1808 | ||
1809 | Which may allow further optimizations. | |
1810 | ||
1811 | Return the equivalent comparison or NULL if no such equivalent comparison | |
1812 | was found. */ | |
1813 | ||
1814 | static tree | |
1815 | find_equivalent_equality_comparison (tree cond) | |
1816 | { | |
1817 | tree op0 = TREE_OPERAND (cond, 0); | |
1818 | tree op1 = TREE_OPERAND (cond, 1); | |
1819 | tree def_stmt = SSA_NAME_DEF_STMT (op0); | |
1820 | ||
1821 | /* OP0 might have been a parameter, so first make sure it | |
1822 | was defined by a MODIFY_EXPR. */ | |
1823 | if (def_stmt && TREE_CODE (def_stmt) == MODIFY_EXPR) | |
1824 | { | |
1825 | tree def_rhs = TREE_OPERAND (def_stmt, 1); | |
1826 | ||
2d0dab7f JL |
1827 | |
1828 | /* If either operand to the comparison is a pointer to | |
1829 | a function, then we can not apply this optimization | |
1830 | as some targets require function pointers to be | |
1831 | canonicalized and in this case this optimization would | |
1832 | eliminate a necessary canonicalization. */ | |
1833 | if ((POINTER_TYPE_P (TREE_TYPE (op0)) | |
1834 | && TREE_CODE (TREE_TYPE (TREE_TYPE (op0))) == FUNCTION_TYPE) | |
1835 | || (POINTER_TYPE_P (TREE_TYPE (op1)) | |
1836 | && TREE_CODE (TREE_TYPE (TREE_TYPE (op1))) == FUNCTION_TYPE)) | |
1837 | return NULL; | |
1838 | ||
6de9cd9a DN |
1839 | /* Now make sure the RHS of the MODIFY_EXPR is a typecast. */ |
1840 | if ((TREE_CODE (def_rhs) == NOP_EXPR | |
1841 | || TREE_CODE (def_rhs) == CONVERT_EXPR) | |
1842 | && TREE_CODE (TREE_OPERAND (def_rhs, 0)) == SSA_NAME) | |
1843 | { | |
1844 | tree def_rhs_inner = TREE_OPERAND (def_rhs, 0); | |
1845 | tree def_rhs_inner_type = TREE_TYPE (def_rhs_inner); | |
1846 | tree new; | |
1847 | ||
1848 | if (TYPE_PRECISION (def_rhs_inner_type) | |
1849 | > TYPE_PRECISION (TREE_TYPE (def_rhs))) | |
1850 | return NULL; | |
1851 | ||
2d0dab7f JL |
1852 | /* If the inner type of the conversion is a pointer to |
1853 | a function, then we can not apply this optimization | |
1854 | as some targets require function pointers to be | |
1855 | canonicalized. This optimization would result in | |
1856 | canonicalization of the pointer when it was not originally | |
1857 | needed/intended. */ | |
1858 | if (POINTER_TYPE_P (def_rhs_inner_type) | |
1859 | && TREE_CODE (TREE_TYPE (def_rhs_inner_type)) == FUNCTION_TYPE) | |
1860 | return NULL; | |
1861 | ||
6de9cd9a DN |
1862 | /* What we want to prove is that if we convert OP1 to |
1863 | the type of the object inside the NOP_EXPR that the | |
1864 | result is still equivalent to SRC. | |
1865 | ||
1866 | If that is true, the build and return new equivalent | |
1867 | condition which uses the source of the typecast and the | |
1868 | new constant (which has only changed its type). */ | |
1869 | new = build1 (TREE_CODE (def_rhs), def_rhs_inner_type, op1); | |
1870 | new = local_fold (new); | |
1871 | if (is_gimple_val (new) && tree_int_cst_equal (new, op1)) | |
b4257cfc RG |
1872 | return build2 (TREE_CODE (cond), TREE_TYPE (cond), |
1873 | def_rhs_inner, new); | |
6de9cd9a DN |
1874 | } |
1875 | } | |
1876 | return NULL; | |
1877 | } | |
1878 | ||
1879 | /* STMT is a COND_EXPR for which we could not trivially determine its | |
1880 | result. This routine attempts to find equivalent forms of the | |
1881 | condition which we may be able to optimize better. It also | |
1882 | uses simple value range propagation to optimize conditionals. */ | |
1883 | ||
1884 | static tree | |
1885 | simplify_cond_and_lookup_avail_expr (tree stmt, | |
6de9cd9a DN |
1886 | stmt_ann_t ann, |
1887 | int insert) | |
1888 | { | |
1889 | tree cond = COND_EXPR_COND (stmt); | |
1890 | ||
6615c446 | 1891 | if (COMPARISON_CLASS_P (cond)) |
6de9cd9a DN |
1892 | { |
1893 | tree op0 = TREE_OPERAND (cond, 0); | |
1894 | tree op1 = TREE_OPERAND (cond, 1); | |
1895 | ||
1896 | if (TREE_CODE (op0) == SSA_NAME && is_gimple_min_invariant (op1)) | |
1897 | { | |
1898 | int limit; | |
1899 | tree low, high, cond_low, cond_high; | |
1900 | int lowequal, highequal, swapped, no_overlap, subset, cond_inverted; | |
8184759d | 1901 | VEC(vrp_element_p,heap) **vrp_records; |
6de9cd9a | 1902 | struct vrp_element *element; |
b8545fbf | 1903 | struct vrp_hash_elt vrp_hash_elt, *vrp_hash_elt_p; |
23530866 | 1904 | void **slot; |
6de9cd9a DN |
1905 | |
1906 | /* First see if we have test of an SSA_NAME against a constant | |
1907 | where the SSA_NAME is defined by an earlier typecast which | |
1908 | is irrelevant when performing tests against the given | |
1909 | constant. */ | |
1910 | if (TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR) | |
1911 | { | |
1912 | tree new_cond = find_equivalent_equality_comparison (cond); | |
1913 | ||
1914 | if (new_cond) | |
1915 | { | |
1916 | /* Update the statement to use the new equivalent | |
1917 | condition. */ | |
1918 | COND_EXPR_COND (stmt) = new_cond; | |
68b9f53b AM |
1919 | |
1920 | /* If this is not a real stmt, ann will be NULL and we | |
1921 | avoid processing the operands. */ | |
1922 | if (ann) | |
f430bae8 | 1923 | mark_stmt_modified (stmt); |
6de9cd9a DN |
1924 | |
1925 | /* Lookup the condition and return its known value if it | |
1926 | exists. */ | |
48732f23 | 1927 | new_cond = lookup_avail_expr (stmt, insert); |
6de9cd9a DN |
1928 | if (new_cond) |
1929 | return new_cond; | |
1930 | ||
1931 | /* The operands have changed, so update op0 and op1. */ | |
1932 | op0 = TREE_OPERAND (cond, 0); | |
1933 | op1 = TREE_OPERAND (cond, 1); | |
1934 | } | |
1935 | } | |
1936 | ||
1937 | /* Consult the value range records for this variable (if they exist) | |
1938 | to see if we can eliminate or simplify this conditional. | |
1939 | ||
1940 | Note two tests are necessary to determine no records exist. | |
1941 | First we have to see if the virtual array exists, if it | |
1942 | exists, then we have to check its active size. | |
1943 | ||
1944 | Also note the vast majority of conditionals are not testing | |
1945 | a variable which has had its range constrained by an earlier | |
1946 | conditional. So this filter avoids a lot of unnecessary work. */ | |
23530866 JL |
1947 | vrp_hash_elt.var = op0; |
1948 | vrp_hash_elt.records = NULL; | |
1949 | slot = htab_find_slot (vrp_data, &vrp_hash_elt, NO_INSERT); | |
1950 | if (slot == NULL) | |
1951 | return NULL; | |
1952 | ||
b8545fbf | 1953 | vrp_hash_elt_p = (struct vrp_hash_elt *) *slot; |
8184759d | 1954 | vrp_records = &vrp_hash_elt_p->records; |
6de9cd9a | 1955 | |
8184759d | 1956 | limit = VEC_length (vrp_element_p, *vrp_records); |
6de9cd9a DN |
1957 | |
1958 | /* If we have no value range records for this variable, or we are | |
1959 | unable to extract a range for this condition, then there is | |
1960 | nothing to do. */ | |
1961 | if (limit == 0 | |
1962 | || ! extract_range_from_cond (cond, &cond_high, | |
1963 | &cond_low, &cond_inverted)) | |
1964 | return NULL; | |
1965 | ||
1966 | /* We really want to avoid unnecessary computations of range | |
1967 | info. So all ranges are computed lazily; this avoids a | |
454ff5cb | 1968 | lot of unnecessary work. i.e., we record the conditional, |
6de9cd9a DN |
1969 | but do not process how it constrains the variable's |
1970 | potential values until we know that processing the condition | |
1971 | could be helpful. | |
1972 | ||
1973 | However, we do not want to have to walk a potentially long | |
1974 | list of ranges, nor do we want to compute a variable's | |
1975 | range more than once for a given path. | |
1976 | ||
1977 | Luckily, each time we encounter a conditional that can not | |
1978 | be otherwise optimized we will end up here and we will | |
1979 | compute the necessary range information for the variable | |
1980 | used in this condition. | |
1981 | ||
1982 | Thus you can conclude that there will never be more than one | |
1983 | conditional associated with a variable which has not been | |
1984 | processed. So we never need to merge more than one new | |
1985 | conditional into the current range. | |
1986 | ||
1987 | These properties also help us avoid unnecessary work. */ | |
8184759d | 1988 | element = VEC_last (vrp_element_p, *vrp_records); |
6de9cd9a DN |
1989 | |
1990 | if (element->high && element->low) | |
1991 | { | |
1992 | /* The last element has been processed, so there is no range | |
1993 | merging to do, we can simply use the high/low values | |
1994 | recorded in the last element. */ | |
1995 | low = element->low; | |
1996 | high = element->high; | |
1997 | } | |
1998 | else | |
1999 | { | |
2000 | tree tmp_high, tmp_low; | |
2001 | int dummy; | |
2002 | ||
75b9aa9f JJ |
2003 | /* The last element has not been processed. Process it now. |
2004 | record_range should ensure for cond inverted is not set. | |
2005 | This call can only fail if cond is x < min or x > max, | |
2006 | which fold should have optimized into false. | |
2007 | If that doesn't happen, just pretend all values are | |
2008 | in the range. */ | |
2009 | if (! extract_range_from_cond (element->cond, &tmp_high, | |
2010 | &tmp_low, &dummy)) | |
2011 | gcc_unreachable (); | |
2012 | else | |
2013 | gcc_assert (dummy == 0); | |
2014 | ||
6de9cd9a DN |
2015 | /* If this is the only element, then no merging is necessary, |
2016 | the high/low values from extract_range_from_cond are all | |
2017 | we need. */ | |
2018 | if (limit == 1) | |
2019 | { | |
2020 | low = tmp_low; | |
2021 | high = tmp_high; | |
2022 | } | |
2023 | else | |
2024 | { | |
2025 | /* Get the high/low value from the previous element. */ | |
2026 | struct vrp_element *prev | |
8184759d | 2027 | = VEC_index (vrp_element_p, *vrp_records, limit - 2); |
6de9cd9a DN |
2028 | low = prev->low; |
2029 | high = prev->high; | |
2030 | ||
2031 | /* Merge in this element's range with the range from the | |
2032 | previous element. | |
2033 | ||
2034 | The low value for the merged range is the maximum of | |
2035 | the previous low value and the low value of this record. | |
2036 | ||
2037 | Similarly the high value for the merged range is the | |
2038 | minimum of the previous high value and the high value of | |
2039 | this record. */ | |
0bca51f0 | 2040 | low = (low && tree_int_cst_compare (low, tmp_low) == 1 |
6de9cd9a | 2041 | ? low : tmp_low); |
0bca51f0 | 2042 | high = (high && tree_int_cst_compare (high, tmp_high) == -1 |
6de9cd9a DN |
2043 | ? high : tmp_high); |
2044 | } | |
2045 | ||
2046 | /* And record the computed range. */ | |
2047 | element->low = low; | |
2048 | element->high = high; | |
2049 | ||
2050 | } | |
2051 | ||
2052 | /* After we have constrained this variable's potential values, | |
2053 | we try to determine the result of the given conditional. | |
2054 | ||
2055 | To simplify later tests, first determine if the current | |
2056 | low value is the same low value as the conditional. | |
2057 | Similarly for the current high value and the high value | |
2058 | for the conditional. */ | |
2059 | lowequal = tree_int_cst_equal (low, cond_low); | |
2060 | highequal = tree_int_cst_equal (high, cond_high); | |
2061 | ||
2062 | if (lowequal && highequal) | |
2063 | return (cond_inverted ? boolean_false_node : boolean_true_node); | |
2064 | ||
2065 | /* To simplify the overlap/subset tests below we may want | |
2066 | to swap the two ranges so that the larger of the two | |
2067 | ranges occurs "first". */ | |
2068 | swapped = 0; | |
2069 | if (tree_int_cst_compare (low, cond_low) == 1 | |
2070 | || (lowequal | |
2071 | && tree_int_cst_compare (cond_high, high) == 1)) | |
2072 | { | |
2073 | tree temp; | |
2074 | ||
2075 | swapped = 1; | |
2076 | temp = low; | |
2077 | low = cond_low; | |
2078 | cond_low = temp; | |
2079 | temp = high; | |
2080 | high = cond_high; | |
2081 | cond_high = temp; | |
2082 | } | |
2083 | ||
2084 | /* Now determine if there is no overlap in the ranges | |
2085 | or if the second range is a subset of the first range. */ | |
2086 | no_overlap = tree_int_cst_lt (high, cond_low); | |
2087 | subset = tree_int_cst_compare (cond_high, high) != 1; | |
2088 | ||
2089 | /* If there was no overlap in the ranges, then this conditional | |
2090 | always has a false value (unless we had to invert this | |
2091 | conditional, in which case it always has a true value). */ | |
2092 | if (no_overlap) | |
2093 | return (cond_inverted ? boolean_true_node : boolean_false_node); | |
2094 | ||
2095 | /* If the current range is a subset of the condition's range, | |
2096 | then this conditional always has a true value (unless we | |
2097 | had to invert this conditional, in which case it always | |
2098 | has a true value). */ | |
2099 | if (subset && swapped) | |
2100 | return (cond_inverted ? boolean_false_node : boolean_true_node); | |
2101 | ||
2102 | /* We were unable to determine the result of the conditional. | |
2103 | However, we may be able to simplify the conditional. First | |
2104 | merge the ranges in the same manner as range merging above. */ | |
2105 | low = tree_int_cst_compare (low, cond_low) == 1 ? low : cond_low; | |
2106 | high = tree_int_cst_compare (high, cond_high) == -1 ? high : cond_high; | |
2107 | ||
2108 | /* If the range has converged to a single point, then turn this | |
2109 | into an equality comparison. */ | |
2110 | if (TREE_CODE (cond) != EQ_EXPR | |
2111 | && TREE_CODE (cond) != NE_EXPR | |
2112 | && tree_int_cst_equal (low, high)) | |
2113 | { | |
2114 | TREE_SET_CODE (cond, EQ_EXPR); | |
2115 | TREE_OPERAND (cond, 1) = high; | |
2116 | } | |
2117 | } | |
2118 | } | |
2119 | return 0; | |
2120 | } | |
2121 | ||
ff2ad0f7 DN |
2122 | /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current |
2123 | known value for that SSA_NAME (or NULL if no value is known). | |
2124 | ||
2125 | NONZERO_VARS is the set SSA_NAMES known to have a nonzero value, | |
2126 | even if we don't know their precise value. | |
2127 | ||
2128 | Propagate values from CONST_AND_COPIES and NONZERO_VARS into the PHI | |
2129 | nodes of the successors of BB. */ | |
2130 | ||
2131 | static void | |
6f2aec07 | 2132 | cprop_into_successor_phis (basic_block bb, bitmap nonzero_vars) |
ff2ad0f7 DN |
2133 | { |
2134 | edge e; | |
628f6a4e | 2135 | edge_iterator ei; |
ff2ad0f7 | 2136 | |
628f6a4e | 2137 | FOR_EACH_EDGE (e, ei, bb->succs) |
ff2ad0f7 DN |
2138 | { |
2139 | tree phi; | |
0492baf2 | 2140 | int indx; |
ff2ad0f7 DN |
2141 | |
2142 | /* If this is an abnormal edge, then we do not want to copy propagate | |
2143 | into the PHI alternative associated with this edge. */ | |
2144 | if (e->flags & EDGE_ABNORMAL) | |
2145 | continue; | |
2146 | ||
2147 | phi = phi_nodes (e->dest); | |
2148 | if (! phi) | |
2149 | continue; | |
2150 | ||
0492baf2 | 2151 | indx = e->dest_idx; |
ff2ad0f7 DN |
2152 | for ( ; phi; phi = PHI_CHAIN (phi)) |
2153 | { | |
ff2ad0f7 DN |
2154 | tree new; |
2155 | use_operand_p orig_p; | |
2156 | tree orig; | |
2157 | ||
ff2ad0f7 DN |
2158 | /* The alternative may be associated with a constant, so verify |
2159 | it is an SSA_NAME before doing anything with it. */ | |
0492baf2 | 2160 | orig_p = PHI_ARG_DEF_PTR (phi, indx); |
ff2ad0f7 DN |
2161 | orig = USE_FROM_PTR (orig_p); |
2162 | if (TREE_CODE (orig) != SSA_NAME) | |
2163 | continue; | |
2164 | ||
2165 | /* If the alternative is known to have a nonzero value, record | |
2166 | that fact in the PHI node itself for future use. */ | |
2167 | if (bitmap_bit_p (nonzero_vars, SSA_NAME_VERSION (orig))) | |
0492baf2 | 2168 | PHI_ARG_NONZERO (phi, indx) = true; |
ff2ad0f7 DN |
2169 | |
2170 | /* If we have *ORIG_P in our constant/copy table, then replace | |
2171 | ORIG_P with its value in our constant/copy table. */ | |
3aecd08b | 2172 | new = SSA_NAME_VALUE (orig); |
ff2ad0f7 | 2173 | if (new |
0bca51f0 | 2174 | && new != orig |
ff2ad0f7 DN |
2175 | && (TREE_CODE (new) == SSA_NAME |
2176 | || is_gimple_min_invariant (new)) | |
2177 | && may_propagate_copy (orig, new)) | |
0bca51f0 | 2178 | propagate_value (orig_p, new); |
ff2ad0f7 DN |
2179 | } |
2180 | } | |
2181 | } | |
2182 | ||
efea75f9 JL |
2183 | /* We have finished optimizing BB, record any information implied by |
2184 | taking a specific outgoing edge from BB. */ | |
2185 | ||
2186 | static void | |
2187 | record_edge_info (basic_block bb) | |
2188 | { | |
2189 | block_stmt_iterator bsi = bsi_last (bb); | |
2190 | struct edge_info *edge_info; | |
2191 | ||
2192 | if (! bsi_end_p (bsi)) | |
2193 | { | |
2194 | tree stmt = bsi_stmt (bsi); | |
2195 | ||
2196 | if (stmt && TREE_CODE (stmt) == SWITCH_EXPR) | |
2197 | { | |
2198 | tree cond = SWITCH_COND (stmt); | |
2199 | ||
2200 | if (TREE_CODE (cond) == SSA_NAME) | |
2201 | { | |
2202 | tree labels = SWITCH_LABELS (stmt); | |
2203 | int i, n_labels = TREE_VEC_LENGTH (labels); | |
e1111e8e | 2204 | tree *info = XCNEWVEC (tree, last_basic_block); |
efea75f9 JL |
2205 | edge e; |
2206 | edge_iterator ei; | |
2207 | ||
2208 | for (i = 0; i < n_labels; i++) | |
2209 | { | |
2210 | tree label = TREE_VEC_ELT (labels, i); | |
2211 | basic_block target_bb = label_to_block (CASE_LABEL (label)); | |
2212 | ||
2213 | if (CASE_HIGH (label) | |
2214 | || !CASE_LOW (label) | |
2215 | || info[target_bb->index]) | |
2216 | info[target_bb->index] = error_mark_node; | |
2217 | else | |
2218 | info[target_bb->index] = label; | |
2219 | } | |
2220 | ||
2221 | FOR_EACH_EDGE (e, ei, bb->succs) | |
2222 | { | |
2223 | basic_block target_bb = e->dest; | |
2224 | tree node = info[target_bb->index]; | |
ff2ad0f7 | 2225 | |
efea75f9 JL |
2226 | if (node != NULL && node != error_mark_node) |
2227 | { | |
2228 | tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node)); | |
2229 | edge_info = allocate_edge_info (e); | |
2230 | edge_info->lhs = cond; | |
2231 | edge_info->rhs = x; | |
2232 | } | |
2233 | } | |
2234 | free (info); | |
2235 | } | |
2236 | } | |
2237 | ||
2238 | /* A COND_EXPR may create equivalences too. */ | |
2239 | if (stmt && TREE_CODE (stmt) == COND_EXPR) | |
2240 | { | |
2241 | tree cond = COND_EXPR_COND (stmt); | |
2242 | edge true_edge; | |
2243 | edge false_edge; | |
2244 | ||
2245 | extract_true_false_edges_from_block (bb, &true_edge, &false_edge); | |
2246 | ||
cbb1cada | 2247 | /* If the conditional is a single variable 'X', record 'X = 1' |
efea75f9 JL |
2248 | for the true edge and 'X = 0' on the false edge. */ |
2249 | if (SSA_VAR_P (cond)) | |
2250 | { | |
2251 | struct edge_info *edge_info; | |
2252 | ||
2253 | edge_info = allocate_edge_info (true_edge); | |
2254 | edge_info->lhs = cond; | |
2255 | edge_info->rhs = constant_boolean_node (1, TREE_TYPE (cond)); | |
2256 | ||
2257 | edge_info = allocate_edge_info (false_edge); | |
2258 | edge_info->lhs = cond; | |
2259 | edge_info->rhs = constant_boolean_node (0, TREE_TYPE (cond)); | |
2260 | } | |
2261 | /* Equality tests may create one or two equivalences. */ | |
2262 | else if (COMPARISON_CLASS_P (cond)) | |
2263 | { | |
2264 | tree op0 = TREE_OPERAND (cond, 0); | |
2265 | tree op1 = TREE_OPERAND (cond, 1); | |
2266 | ||
2267 | /* Special case comparing booleans against a constant as we | |
2268 | know the value of OP0 on both arms of the branch. i.e., we | |
2269 | can record an equivalence for OP0 rather than COND. */ | |
2270 | if ((TREE_CODE (cond) == EQ_EXPR || TREE_CODE (cond) == NE_EXPR) | |
2271 | && TREE_CODE (op0) == SSA_NAME | |
2272 | && TREE_CODE (TREE_TYPE (op0)) == BOOLEAN_TYPE | |
2273 | && is_gimple_min_invariant (op1)) | |
2274 | { | |
2275 | if (TREE_CODE (cond) == EQ_EXPR) | |
2276 | { | |
2277 | edge_info = allocate_edge_info (true_edge); | |
2278 | edge_info->lhs = op0; | |
2279 | edge_info->rhs = (integer_zerop (op1) | |
2280 | ? boolean_false_node | |
2281 | : boolean_true_node); | |
2282 | ||
2283 | edge_info = allocate_edge_info (false_edge); | |
2284 | edge_info->lhs = op0; | |
2285 | edge_info->rhs = (integer_zerop (op1) | |
2286 | ? boolean_true_node | |
2287 | : boolean_false_node); | |
2288 | } | |
2289 | else | |
2290 | { | |
2291 | edge_info = allocate_edge_info (true_edge); | |
2292 | edge_info->lhs = op0; | |
2293 | edge_info->rhs = (integer_zerop (op1) | |
2294 | ? boolean_true_node | |
2295 | : boolean_false_node); | |
2296 | ||
2297 | edge_info = allocate_edge_info (false_edge); | |
2298 | edge_info->lhs = op0; | |
2299 | edge_info->rhs = (integer_zerop (op1) | |
2300 | ? boolean_false_node | |
2301 | : boolean_true_node); | |
2302 | } | |
2303 | } | |
2304 | ||
3bed147c KH |
2305 | else if (is_gimple_min_invariant (op0) |
2306 | && (TREE_CODE (op1) == SSA_NAME | |
2307 | || is_gimple_min_invariant (op1))) | |
efea75f9 JL |
2308 | { |
2309 | tree inverted = invert_truthvalue (cond); | |
2310 | struct edge_info *edge_info; | |
2311 | ||
2312 | edge_info = allocate_edge_info (true_edge); | |
2313 | record_conditions (edge_info, cond, inverted); | |
2314 | ||
2315 | if (TREE_CODE (cond) == EQ_EXPR) | |
2316 | { | |
2317 | edge_info->lhs = op1; | |
2318 | edge_info->rhs = op0; | |
2319 | } | |
2320 | ||
2321 | edge_info = allocate_edge_info (false_edge); | |
2322 | record_conditions (edge_info, inverted, cond); | |
2323 | ||
2324 | if (TREE_CODE (cond) == NE_EXPR) | |
2325 | { | |
2326 | edge_info->lhs = op1; | |
2327 | edge_info->rhs = op0; | |
2328 | } | |
2329 | } | |
2330 | ||
3bed147c KH |
2331 | else if (TREE_CODE (op0) == SSA_NAME |
2332 | && (is_gimple_min_invariant (op1) | |
2333 | || TREE_CODE (op1) == SSA_NAME)) | |
efea75f9 JL |
2334 | { |
2335 | tree inverted = invert_truthvalue (cond); | |
2336 | struct edge_info *edge_info; | |
2337 | ||
2338 | edge_info = allocate_edge_info (true_edge); | |
2339 | record_conditions (edge_info, cond, inverted); | |
2340 | ||
2341 | if (TREE_CODE (cond) == EQ_EXPR) | |
2342 | { | |
2343 | edge_info->lhs = op0; | |
2344 | edge_info->rhs = op1; | |
2345 | } | |
2346 | ||
2347 | edge_info = allocate_edge_info (false_edge); | |
2348 | record_conditions (edge_info, inverted, cond); | |
2349 | ||
2350 | if (TREE_CODE (cond) == NE_EXPR) | |
2351 | { | |
2352 | edge_info->lhs = op0; | |
2353 | edge_info->rhs = op1; | |
2354 | } | |
2355 | } | |
2356 | } | |
2357 | ||
2358 | /* ??? TRUTH_NOT_EXPR can create an equivalence too. */ | |
2359 | } | |
2360 | } | |
2361 | } | |
2362 | ||
2363 | /* Propagate information from BB to its outgoing edges. | |
2364 | ||
2365 | This can include equivalency information implied by control statements | |
2366 | at the end of BB and const/copy propagation into PHIs in BB's | |
2367 | successor blocks. */ | |
6de9cd9a DN |
2368 | |
2369 | static void | |
efea75f9 JL |
2370 | propagate_to_outgoing_edges (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED, |
2371 | basic_block bb) | |
6de9cd9a | 2372 | { |
efea75f9 | 2373 | record_edge_info (bb); |
6f2aec07 | 2374 | cprop_into_successor_phis (bb, nonzero_vars); |
6de9cd9a DN |
2375 | } |
2376 | ||
2377 | /* Search for redundant computations in STMT. If any are found, then | |
2378 | replace them with the variable holding the result of the computation. | |
2379 | ||
2380 | If safe, record this expression into the available expression hash | |
2381 | table. */ | |
2382 | ||
2383 | static bool | |
a513fe88 | 2384 | eliminate_redundant_computations (tree stmt, stmt_ann_t ann) |
6de9cd9a | 2385 | { |
6de9cd9a DN |
2386 | tree *expr_p, def = NULL_TREE; |
2387 | bool insert = true; | |
2388 | tree cached_lhs; | |
2389 | bool retval = false; | |
019b02f1 | 2390 | bool modify_expr_p = false; |
6de9cd9a DN |
2391 | |
2392 | if (TREE_CODE (stmt) == MODIFY_EXPR) | |
2393 | def = TREE_OPERAND (stmt, 0); | |
2394 | ||
2395 | /* Certain expressions on the RHS can be optimized away, but can not | |
471854f8 | 2396 | themselves be entered into the hash tables. */ |
6de9cd9a DN |
2397 | if (ann->makes_aliased_stores |
2398 | || ! def | |
2399 | || TREE_CODE (def) != SSA_NAME | |
2400 | || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def) | |
f47c96aa | 2401 | || !ZERO_SSA_OPERANDS (stmt, SSA_OP_VMAYDEF) |
f67e783f ZD |
2402 | /* Do not record equivalences for increments of ivs. This would create |
2403 | overlapping live ranges for a very questionable gain. */ | |
2404 | || simple_iv_increment_p (stmt)) | |
6de9cd9a DN |
2405 | insert = false; |
2406 | ||
2407 | /* Check if the expression has been computed before. */ | |
48732f23 | 2408 | cached_lhs = lookup_avail_expr (stmt, insert); |
6de9cd9a | 2409 | |
471eeb83 JL |
2410 | /* If this is a COND_EXPR and we did not find its expression in |
2411 | the hash table, simplify the condition and try again. */ | |
2412 | if (! cached_lhs && TREE_CODE (stmt) == COND_EXPR) | |
48732f23 | 2413 | cached_lhs = simplify_cond_and_lookup_avail_expr (stmt, ann, insert); |
6de9cd9a DN |
2414 | |
2415 | opt_stats.num_exprs_considered++; | |
2416 | ||
2417 | /* Get a pointer to the expression we are trying to optimize. */ | |
2418 | if (TREE_CODE (stmt) == COND_EXPR) | |
2419 | expr_p = &COND_EXPR_COND (stmt); | |
2420 | else if (TREE_CODE (stmt) == SWITCH_EXPR) | |
2421 | expr_p = &SWITCH_COND (stmt); | |
2422 | else if (TREE_CODE (stmt) == RETURN_EXPR && TREE_OPERAND (stmt, 0)) | |
019b02f1 AP |
2423 | { |
2424 | expr_p = &TREE_OPERAND (TREE_OPERAND (stmt, 0), 1); | |
2425 | modify_expr_p = true; | |
2426 | } | |
6de9cd9a | 2427 | else |
019b02f1 AP |
2428 | { |
2429 | expr_p = &TREE_OPERAND (stmt, 1); | |
2430 | modify_expr_p = true; | |
2431 | } | |
6de9cd9a DN |
2432 | |
2433 | /* It is safe to ignore types here since we have already done | |
2434 | type checking in the hashing and equality routines. In fact | |
2435 | type checking here merely gets in the way of constant | |
2436 | propagation. Also, make sure that it is safe to propagate | |
2437 | CACHED_LHS into *EXPR_P. */ | |
2438 | if (cached_lhs | |
019b02f1 AP |
2439 | && ((TREE_CODE (cached_lhs) != SSA_NAME |
2440 | && (modify_expr_p | |
2441 | || tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p), | |
2442 | TREE_TYPE (cached_lhs)))) | |
ff2ad0f7 | 2443 | || may_propagate_copy (*expr_p, cached_lhs))) |
6de9cd9a DN |
2444 | { |
2445 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2446 | { | |
2447 | fprintf (dump_file, " Replaced redundant expr '"); | |
2448 | print_generic_expr (dump_file, *expr_p, dump_flags); | |
2449 | fprintf (dump_file, "' with '"); | |
2450 | print_generic_expr (dump_file, cached_lhs, dump_flags); | |
2451 | fprintf (dump_file, "'\n"); | |
2452 | } | |
2453 | ||
2454 | opt_stats.num_re++; | |
2455 | ||
2456 | #if defined ENABLE_CHECKING | |
1e128c5f GB |
2457 | gcc_assert (TREE_CODE (cached_lhs) == SSA_NAME |
2458 | || is_gimple_min_invariant (cached_lhs)); | |
6de9cd9a DN |
2459 | #endif |
2460 | ||
2461 | if (TREE_CODE (cached_lhs) == ADDR_EXPR | |
2462 | || (POINTER_TYPE_P (TREE_TYPE (*expr_p)) | |
2463 | && is_gimple_min_invariant (cached_lhs))) | |
2464 | retval = true; | |
019b02f1 AP |
2465 | |
2466 | if (modify_expr_p | |
2467 | && !tree_ssa_useless_type_conversion_1 (TREE_TYPE (*expr_p), | |
2468 | TREE_TYPE (cached_lhs))) | |
2469 | cached_lhs = fold_convert (TREE_TYPE (*expr_p), cached_lhs); | |
6de9cd9a | 2470 | |
d00ad49b | 2471 | propagate_tree_value (expr_p, cached_lhs); |
f430bae8 | 2472 | mark_stmt_modified (stmt); |
6de9cd9a DN |
2473 | } |
2474 | return retval; | |
2475 | } | |
2476 | ||
2477 | /* STMT, a MODIFY_EXPR, may create certain equivalences, in either | |
2478 | the available expressions table or the const_and_copies table. | |
2479 | Detect and record those equivalences. */ | |
2480 | ||
2481 | static void | |
2482 | record_equivalences_from_stmt (tree stmt, | |
6de9cd9a DN |
2483 | int may_optimize_p, |
2484 | stmt_ann_t ann) | |
2485 | { | |
2486 | tree lhs = TREE_OPERAND (stmt, 0); | |
2487 | enum tree_code lhs_code = TREE_CODE (lhs); | |
2488 | int i; | |
2489 | ||
2490 | if (lhs_code == SSA_NAME) | |
2491 | { | |
2492 | tree rhs = TREE_OPERAND (stmt, 1); | |
2493 | ||
2494 | /* Strip away any useless type conversions. */ | |
2495 | STRIP_USELESS_TYPE_CONVERSION (rhs); | |
2496 | ||
2497 | /* If the RHS of the assignment is a constant or another variable that | |
2498 | may be propagated, register it in the CONST_AND_COPIES table. We | |
2499 | do not need to record unwind data for this, since this is a true | |
1ea7e6ad | 2500 | assignment and not an equivalence inferred from a comparison. All |
6de9cd9a DN |
2501 | uses of this ssa name are dominated by this assignment, so unwinding |
2502 | just costs time and space. */ | |
2503 | if (may_optimize_p | |
2504 | && (TREE_CODE (rhs) == SSA_NAME | |
2505 | || is_gimple_min_invariant (rhs))) | |
3aecd08b | 2506 | SSA_NAME_VALUE (lhs) = rhs; |
6de9cd9a | 2507 | |
4db8040c | 2508 | if (tree_expr_nonzero_p (rhs)) |
fdabe5c2 | 2509 | record_var_is_nonzero (lhs); |
6de9cd9a DN |
2510 | } |
2511 | ||
2512 | /* Look at both sides for pointer dereferences. If we find one, then | |
2513 | the pointer must be nonnull and we can enter that equivalence into | |
2514 | the hash tables. */ | |
dd747311 JL |
2515 | if (flag_delete_null_pointer_checks) |
2516 | for (i = 0; i < 2; i++) | |
2517 | { | |
2518 | tree t = TREE_OPERAND (stmt, i); | |
2519 | ||
2520 | /* Strip away any COMPONENT_REFs. */ | |
2521 | while (TREE_CODE (t) == COMPONENT_REF) | |
2522 | t = TREE_OPERAND (t, 0); | |
2523 | ||
2524 | /* Now see if this is a pointer dereference. */ | |
1b096a0a | 2525 | if (INDIRECT_REF_P (t)) |
dd747311 JL |
2526 | { |
2527 | tree op = TREE_OPERAND (t, 0); | |
2528 | ||
2529 | /* If the pointer is a SSA variable, then enter new | |
2530 | equivalences into the hash table. */ | |
2531 | while (TREE_CODE (op) == SSA_NAME) | |
2532 | { | |
2533 | tree def = SSA_NAME_DEF_STMT (op); | |
2534 | ||
fdabe5c2 | 2535 | record_var_is_nonzero (op); |
dd747311 JL |
2536 | |
2537 | /* And walk up the USE-DEF chains noting other SSA_NAMEs | |
2538 | which are known to have a nonzero value. */ | |
2539 | if (def | |
2540 | && TREE_CODE (def) == MODIFY_EXPR | |
2541 | && TREE_CODE (TREE_OPERAND (def, 1)) == NOP_EXPR) | |
2542 | op = TREE_OPERAND (TREE_OPERAND (def, 1), 0); | |
2543 | else | |
2544 | break; | |
2545 | } | |
2546 | } | |
2547 | } | |
6de9cd9a DN |
2548 | |
2549 | /* A memory store, even an aliased store, creates a useful | |
2550 | equivalence. By exchanging the LHS and RHS, creating suitable | |
2551 | vops and recording the result in the available expression table, | |
2552 | we may be able to expose more redundant loads. */ | |
2553 | if (!ann->has_volatile_ops | |
2554 | && (TREE_CODE (TREE_OPERAND (stmt, 1)) == SSA_NAME | |
2555 | || is_gimple_min_invariant (TREE_OPERAND (stmt, 1))) | |
2556 | && !is_gimple_reg (lhs)) | |
2557 | { | |
2558 | tree rhs = TREE_OPERAND (stmt, 1); | |
2559 | tree new; | |
6de9cd9a DN |
2560 | |
2561 | /* FIXME: If the LHS of the assignment is a bitfield and the RHS | |
2562 | is a constant, we need to adjust the constant to fit into the | |
2563 | type of the LHS. If the LHS is a bitfield and the RHS is not | |
2564 | a constant, then we can not record any equivalences for this | |
2565 | statement since we would need to represent the widening or | |
2566 | narrowing of RHS. This fixes gcc.c-torture/execute/921016-1.c | |
2567 | and should not be necessary if GCC represented bitfields | |
2568 | properly. */ | |
2569 | if (lhs_code == COMPONENT_REF | |
2570 | && DECL_BIT_FIELD (TREE_OPERAND (lhs, 1))) | |
2571 | { | |
2572 | if (TREE_CONSTANT (rhs)) | |
2573 | rhs = widen_bitfield (rhs, TREE_OPERAND (lhs, 1), lhs); | |
2574 | else | |
2575 | rhs = NULL; | |
2576 | ||
2577 | /* If the value overflowed, then we can not use this equivalence. */ | |
2578 | if (rhs && ! is_gimple_min_invariant (rhs)) | |
2579 | rhs = NULL; | |
2580 | } | |
2581 | ||
2582 | if (rhs) | |
2583 | { | |
6de9cd9a | 2584 | /* Build a new statement with the RHS and LHS exchanged. */ |
b4257cfc | 2585 | new = build2 (MODIFY_EXPR, TREE_TYPE (stmt), rhs, lhs); |
6de9cd9a | 2586 | |
f47c96aa | 2587 | create_ssa_artficial_load_stmt (new, stmt); |
6de9cd9a DN |
2588 | |
2589 | /* Finally enter the statement into the available expression | |
2590 | table. */ | |
48732f23 | 2591 | lookup_avail_expr (new, true); |
6de9cd9a DN |
2592 | } |
2593 | } | |
2594 | } | |
2595 | ||
ff2ad0f7 DN |
2596 | /* Replace *OP_P in STMT with any known equivalent value for *OP_P from |
2597 | CONST_AND_COPIES. */ | |
2598 | ||
2599 | static bool | |
6f2aec07 | 2600 | cprop_operand (tree stmt, use_operand_p op_p) |
ff2ad0f7 DN |
2601 | { |
2602 | bool may_have_exposed_new_symbols = false; | |
2603 | tree val; | |
2604 | tree op = USE_FROM_PTR (op_p); | |
2605 | ||
2606 | /* If the operand has a known constant value or it is known to be a | |
2607 | copy of some other variable, use the value or copy stored in | |
2608 | CONST_AND_COPIES. */ | |
3aecd08b | 2609 | val = SSA_NAME_VALUE (op); |
0bca51f0 | 2610 | if (val && val != op && TREE_CODE (val) != VALUE_HANDLE) |
ff2ad0f7 DN |
2611 | { |
2612 | tree op_type, val_type; | |
2613 | ||
2614 | /* Do not change the base variable in the virtual operand | |
2615 | tables. That would make it impossible to reconstruct | |
2616 | the renamed virtual operand if we later modify this | |
2617 | statement. Also only allow the new value to be an SSA_NAME | |
2618 | for propagation into virtual operands. */ | |
2619 | if (!is_gimple_reg (op) | |
0bca51f0 DN |
2620 | && (TREE_CODE (val) != SSA_NAME |
2621 | || is_gimple_reg (val) | |
2622 | || get_virtual_var (val) != get_virtual_var (op))) | |
ff2ad0f7 DN |
2623 | return false; |
2624 | ||
aa24864c RH |
2625 | /* Do not replace hard register operands in asm statements. */ |
2626 | if (TREE_CODE (stmt) == ASM_EXPR | |
2627 | && !may_propagate_copy_into_asm (op)) | |
2628 | return false; | |
2629 | ||
ff2ad0f7 DN |
2630 | /* Get the toplevel type of each operand. */ |
2631 | op_type = TREE_TYPE (op); | |
2632 | val_type = TREE_TYPE (val); | |
2633 | ||
2634 | /* While both types are pointers, get the type of the object | |
2635 | pointed to. */ | |
2636 | while (POINTER_TYPE_P (op_type) && POINTER_TYPE_P (val_type)) | |
2637 | { | |
2638 | op_type = TREE_TYPE (op_type); | |
2639 | val_type = TREE_TYPE (val_type); | |
2640 | } | |
2641 | ||
63b88252 RH |
2642 | /* Make sure underlying types match before propagating a constant by |
2643 | converting the constant to the proper type. Note that convert may | |
2644 | return a non-gimple expression, in which case we ignore this | |
2645 | propagation opportunity. */ | |
2646 | if (TREE_CODE (val) != SSA_NAME) | |
ff2ad0f7 | 2647 | { |
63b88252 RH |
2648 | if (!lang_hooks.types_compatible_p (op_type, val_type)) |
2649 | { | |
2650 | val = fold_convert (TREE_TYPE (op), val); | |
2651 | if (!is_gimple_min_invariant (val)) | |
2652 | return false; | |
2653 | } | |
ff2ad0f7 DN |
2654 | } |
2655 | ||
2656 | /* Certain operands are not allowed to be copy propagated due | |
2657 | to their interaction with exception handling and some GCC | |
2658 | extensions. */ | |
63b88252 | 2659 | else if (!may_propagate_copy (op, val)) |
ff2ad0f7 | 2660 | return false; |
111e0c9f DB |
2661 | |
2662 | /* Do not propagate copies if the propagated value is at a deeper loop | |
2663 | depth than the propagatee. Otherwise, this may move loop variant | |
2664 | variables outside of their loops and prevent coalescing | |
2665 | opportunities. If the value was loop invariant, it will be hoisted | |
2666 | by LICM and exposed for copy propagation. */ | |
2667 | if (loop_depth_of_name (val) > loop_depth_of_name (op)) | |
2668 | return false; | |
ff2ad0f7 DN |
2669 | |
2670 | /* Dump details. */ | |
2671 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2672 | { | |
2673 | fprintf (dump_file, " Replaced '"); | |
2674 | print_generic_expr (dump_file, op, dump_flags); | |
2675 | fprintf (dump_file, "' with %s '", | |
2676 | (TREE_CODE (val) != SSA_NAME ? "constant" : "variable")); | |
2677 | print_generic_expr (dump_file, val, dump_flags); | |
2678 | fprintf (dump_file, "'\n"); | |
2679 | } | |
2680 | ||
2681 | /* If VAL is an ADDR_EXPR or a constant of pointer type, note | |
2682 | that we may have exposed a new symbol for SSA renaming. */ | |
2683 | if (TREE_CODE (val) == ADDR_EXPR | |
2684 | || (POINTER_TYPE_P (TREE_TYPE (op)) | |
2685 | && is_gimple_min_invariant (val))) | |
2686 | may_have_exposed_new_symbols = true; | |
2687 | ||
0bca51f0 DN |
2688 | if (TREE_CODE (val) != SSA_NAME) |
2689 | opt_stats.num_const_prop++; | |
2690 | else | |
2691 | opt_stats.num_copy_prop++; | |
2692 | ||
ff2ad0f7 DN |
2693 | propagate_value (op_p, val); |
2694 | ||
2695 | /* And note that we modified this statement. This is now | |
2696 | safe, even if we changed virtual operands since we will | |
2697 | rescan the statement and rewrite its operands again. */ | |
f430bae8 | 2698 | mark_stmt_modified (stmt); |
ff2ad0f7 DN |
2699 | } |
2700 | return may_have_exposed_new_symbols; | |
2701 | } | |
2702 | ||
2703 | /* CONST_AND_COPIES is a table which maps an SSA_NAME to the current | |
2704 | known value for that SSA_NAME (or NULL if no value is known). | |
2705 | ||
2706 | Propagate values from CONST_AND_COPIES into the uses, vuses and | |
2707 | v_may_def_ops of STMT. */ | |
2708 | ||
2709 | static bool | |
6f2aec07 | 2710 | cprop_into_stmt (tree stmt) |
ff2ad0f7 DN |
2711 | { |
2712 | bool may_have_exposed_new_symbols = false; | |
4c124b4c AM |
2713 | use_operand_p op_p; |
2714 | ssa_op_iter iter; | |
ff2ad0f7 | 2715 | |
4c124b4c | 2716 | FOR_EACH_SSA_USE_OPERAND (op_p, stmt, iter, SSA_OP_ALL_USES) |
ff2ad0f7 | 2717 | { |
ff2ad0f7 | 2718 | if (TREE_CODE (USE_FROM_PTR (op_p)) == SSA_NAME) |
6f2aec07 | 2719 | may_have_exposed_new_symbols |= cprop_operand (stmt, op_p); |
ff2ad0f7 DN |
2720 | } |
2721 | ||
ff2ad0f7 DN |
2722 | return may_have_exposed_new_symbols; |
2723 | } | |
2724 | ||
2725 | ||
206048bd | 2726 | /* Optimize the statement pointed to by iterator SI. |
6de9cd9a DN |
2727 | |
2728 | We try to perform some simplistic global redundancy elimination and | |
2729 | constant propagation: | |
2730 | ||
2731 | 1- To detect global redundancy, we keep track of expressions that have | |
2732 | been computed in this block and its dominators. If we find that the | |
2733 | same expression is computed more than once, we eliminate repeated | |
2734 | computations by using the target of the first one. | |
2735 | ||
2736 | 2- Constant values and copy assignments. This is used to do very | |
2737 | simplistic constant and copy propagation. When a constant or copy | |
2738 | assignment is found, we map the value on the RHS of the assignment to | |
2739 | the variable in the LHS in the CONST_AND_COPIES table. */ | |
2740 | ||
2741 | static void | |
a513fe88 JL |
2742 | optimize_stmt (struct dom_walk_data *walk_data ATTRIBUTE_UNUSED, |
2743 | basic_block bb, block_stmt_iterator si) | |
6de9cd9a DN |
2744 | { |
2745 | stmt_ann_t ann; | |
af47810a | 2746 | tree stmt, old_stmt; |
6de9cd9a DN |
2747 | bool may_optimize_p; |
2748 | bool may_have_exposed_new_symbols = false; | |
6de9cd9a | 2749 | |
af47810a | 2750 | old_stmt = stmt = bsi_stmt (si); |
0e0ed594 JL |
2751 | |
2752 | if (TREE_CODE (stmt) == COND_EXPR) | |
2753 | canonicalize_comparison (stmt); | |
2754 | ||
f430bae8 | 2755 | update_stmt_if_modified (stmt); |
6de9cd9a | 2756 | ann = stmt_ann (stmt); |
6de9cd9a DN |
2757 | opt_stats.num_stmts++; |
2758 | may_have_exposed_new_symbols = false; | |
2759 | ||
2760 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2761 | { | |
2762 | fprintf (dump_file, "Optimizing statement "); | |
2763 | print_generic_stmt (dump_file, stmt, TDF_SLIM); | |
2764 | } | |
2765 | ||
a32b97a2 | 2766 | /* Const/copy propagate into USES, VUSES and the RHS of V_MAY_DEFs. */ |
6f2aec07 | 2767 | may_have_exposed_new_symbols = cprop_into_stmt (stmt); |
6de9cd9a DN |
2768 | |
2769 | /* If the statement has been modified with constant replacements, | |
2770 | fold its RHS before checking for redundant computations. */ | |
2771 | if (ann->modified) | |
2772 | { | |
6cedb4ac JL |
2773 | tree rhs; |
2774 | ||
6de9cd9a DN |
2775 | /* Try to fold the statement making sure that STMT is kept |
2776 | up to date. */ | |
2777 | if (fold_stmt (bsi_stmt_ptr (si))) | |
2778 | { | |
2779 | stmt = bsi_stmt (si); | |
2780 | ann = stmt_ann (stmt); | |
2781 | ||
2782 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2783 | { | |
2784 | fprintf (dump_file, " Folded to: "); | |
2785 | print_generic_stmt (dump_file, stmt, TDF_SLIM); | |
2786 | } | |
2787 | } | |
2788 | ||
6cedb4ac JL |
2789 | rhs = get_rhs (stmt); |
2790 | if (rhs && TREE_CODE (rhs) == ADDR_EXPR) | |
127203ac | 2791 | recompute_tree_invariant_for_addr_expr (rhs); |
6cedb4ac | 2792 | |
6de9cd9a DN |
2793 | /* Constant/copy propagation above may change the set of |
2794 | virtual operands associated with this statement. Folding | |
2795 | may remove the need for some virtual operands. | |
2796 | ||
2797 | Indicate we will need to rescan and rewrite the statement. */ | |
2798 | may_have_exposed_new_symbols = true; | |
2799 | } | |
2800 | ||
2801 | /* Check for redundant computations. Do this optimization only | |
2802 | for assignments that have no volatile ops and conditionals. */ | |
2803 | may_optimize_p = (!ann->has_volatile_ops | |
2804 | && ((TREE_CODE (stmt) == RETURN_EXPR | |
2805 | && TREE_OPERAND (stmt, 0) | |
2806 | && TREE_CODE (TREE_OPERAND (stmt, 0)) == MODIFY_EXPR | |
2807 | && ! (TREE_SIDE_EFFECTS | |
2808 | (TREE_OPERAND (TREE_OPERAND (stmt, 0), 1)))) | |
2809 | || (TREE_CODE (stmt) == MODIFY_EXPR | |
2810 | && ! TREE_SIDE_EFFECTS (TREE_OPERAND (stmt, 1))) | |
2811 | || TREE_CODE (stmt) == COND_EXPR | |
2812 | || TREE_CODE (stmt) == SWITCH_EXPR)); | |
2813 | ||
2814 | if (may_optimize_p) | |
2815 | may_have_exposed_new_symbols | |
a513fe88 | 2816 | |= eliminate_redundant_computations (stmt, ann); |
6de9cd9a DN |
2817 | |
2818 | /* Record any additional equivalences created by this statement. */ | |
2819 | if (TREE_CODE (stmt) == MODIFY_EXPR) | |
2820 | record_equivalences_from_stmt (stmt, | |
6de9cd9a DN |
2821 | may_optimize_p, |
2822 | ann); | |
2823 | ||
6de9cd9a DN |
2824 | /* If STMT is a COND_EXPR and it was modified, then we may know |
2825 | where it goes. If that is the case, then mark the CFG as altered. | |
2826 | ||
2827 | This will cause us to later call remove_unreachable_blocks and | |
2828 | cleanup_tree_cfg when it is safe to do so. It is not safe to | |
2829 | clean things up here since removal of edges and such can trigger | |
2830 | the removal of PHI nodes, which in turn can release SSA_NAMEs to | |
2831 | the manager. | |
2832 | ||
2833 | That's all fine and good, except that once SSA_NAMEs are released | |
2834 | to the manager, we must not call create_ssa_name until all references | |
2835 | to released SSA_NAMEs have been eliminated. | |
2836 | ||
2837 | All references to the deleted SSA_NAMEs can not be eliminated until | |
2838 | we remove unreachable blocks. | |
2839 | ||
2840 | We can not remove unreachable blocks until after we have completed | |
2841 | any queued jump threading. | |
2842 | ||
2843 | We can not complete any queued jump threads until we have taken | |
2844 | appropriate variables out of SSA form. Taking variables out of | |
2845 | SSA form can call create_ssa_name and thus we lose. | |
2846 | ||
2847 | Ultimately I suspect we're going to need to change the interface | |
2848 | into the SSA_NAME manager. */ | |
2849 | ||
2850 | if (ann->modified) | |
2851 | { | |
2852 | tree val = NULL; | |
2853 | ||
2854 | if (TREE_CODE (stmt) == COND_EXPR) | |
2855 | val = COND_EXPR_COND (stmt); | |
2856 | else if (TREE_CODE (stmt) == SWITCH_EXPR) | |
2857 | val = SWITCH_COND (stmt); | |
2858 | ||
1eaba2f2 | 2859 | if (val && TREE_CODE (val) == INTEGER_CST && find_taken_edge (bb, val)) |
6de9cd9a | 2860 | cfg_altered = true; |
1eaba2f2 RH |
2861 | |
2862 | /* If we simplified a statement in such a way as to be shown that it | |
2863 | cannot trap, update the eh information and the cfg to match. */ | |
af47810a | 2864 | if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt)) |
1eaba2f2 RH |
2865 | { |
2866 | bitmap_set_bit (need_eh_cleanup, bb->index); | |
2867 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2868 | fprintf (dump_file, " Flagged to clear EH edges.\n"); | |
2869 | } | |
6de9cd9a | 2870 | } |
1eaba2f2 | 2871 | |
6de9cd9a | 2872 | if (may_have_exposed_new_symbols) |
d4e6fecb | 2873 | VEC_safe_push (tree, heap, stmts_to_rescan, bsi_stmt (si)); |
6de9cd9a DN |
2874 | } |
2875 | ||
6de9cd9a DN |
2876 | /* Search for an existing instance of STMT in the AVAIL_EXPRS table. If |
2877 | found, return its LHS. Otherwise insert STMT in the table and return | |
2878 | NULL_TREE. | |
2879 | ||
2880 | Also, when an expression is first inserted in the AVAIL_EXPRS table, it | |
206048bd | 2881 | is also added to the stack pointed to by BLOCK_AVAIL_EXPRS_P, so that they |
6de9cd9a DN |
2882 | can be removed when we finish processing this block and its children. |
2883 | ||
2884 | NOTE: This function assumes that STMT is a MODIFY_EXPR node that | |
2885 | contains no CALL_EXPR on its RHS and makes no volatile nor | |
2886 | aliased references. */ | |
2887 | ||
2888 | static tree | |
48732f23 | 2889 | lookup_avail_expr (tree stmt, bool insert) |
6de9cd9a DN |
2890 | { |
2891 | void **slot; | |
2892 | tree lhs; | |
2893 | tree temp; | |
e1111e8e | 2894 | struct expr_hash_elt *element = XNEW (struct expr_hash_elt); |
6de9cd9a DN |
2895 | |
2896 | lhs = TREE_CODE (stmt) == MODIFY_EXPR ? TREE_OPERAND (stmt, 0) : NULL; | |
2897 | ||
2898 | initialize_hash_element (stmt, lhs, element); | |
2899 | ||
2900 | /* Don't bother remembering constant assignments and copy operations. | |
2901 | Constants and copy operations are handled by the constant/copy propagator | |
2902 | in optimize_stmt. */ | |
2903 | if (TREE_CODE (element->rhs) == SSA_NAME | |
2904 | || is_gimple_min_invariant (element->rhs)) | |
2905 | { | |
2906 | free (element); | |
2907 | return NULL_TREE; | |
2908 | } | |
2909 | ||
2910 | /* If this is an equality test against zero, see if we have recorded a | |
2911 | nonzero value for the variable in question. */ | |
2912 | if ((TREE_CODE (element->rhs) == EQ_EXPR | |
2913 | || TREE_CODE (element->rhs) == NE_EXPR) | |
2914 | && TREE_CODE (TREE_OPERAND (element->rhs, 0)) == SSA_NAME | |
2915 | && integer_zerop (TREE_OPERAND (element->rhs, 1))) | |
2916 | { | |
2917 | int indx = SSA_NAME_VERSION (TREE_OPERAND (element->rhs, 0)); | |
2918 | ||
2919 | if (bitmap_bit_p (nonzero_vars, indx)) | |
2920 | { | |
2921 | tree t = element->rhs; | |
2922 | free (element); | |
019b02f1 | 2923 | return constant_boolean_node (TREE_CODE (t) != EQ_EXPR, |
bbcbc3e0 | 2924 | TREE_TYPE (t)); |
6de9cd9a DN |
2925 | } |
2926 | } | |
2927 | ||
2928 | /* Finally try to find the expression in the main expression hash table. */ | |
2929 | slot = htab_find_slot_with_hash (avail_exprs, element, element->hash, | |
2930 | (insert ? INSERT : NO_INSERT)); | |
2931 | if (slot == NULL) | |
2932 | { | |
2933 | free (element); | |
2934 | return NULL_TREE; | |
2935 | } | |
2936 | ||
2937 | if (*slot == NULL) | |
2938 | { | |
2939 | *slot = (void *) element; | |
d4e6fecb | 2940 | VEC_safe_push (tree, heap, avail_exprs_stack, |
ceb7eb8f | 2941 | stmt ? stmt : element->rhs); |
6de9cd9a DN |
2942 | return NULL_TREE; |
2943 | } | |
2944 | ||
2945 | /* Extract the LHS of the assignment so that it can be used as the current | |
2946 | definition of another variable. */ | |
2947 | lhs = ((struct expr_hash_elt *)*slot)->lhs; | |
2948 | ||
2949 | /* See if the LHS appears in the CONST_AND_COPIES table. If it does, then | |
2950 | use the value from the const_and_copies table. */ | |
2951 | if (TREE_CODE (lhs) == SSA_NAME) | |
2952 | { | |
3aecd08b JL |
2953 | temp = SSA_NAME_VALUE (lhs); |
2954 | if (temp && TREE_CODE (temp) != VALUE_HANDLE) | |
6de9cd9a DN |
2955 | lhs = temp; |
2956 | } | |
2957 | ||
2958 | free (element); | |
2959 | return lhs; | |
2960 | } | |
2961 | ||
2962 | /* Given a condition COND, record into HI_P, LO_P and INVERTED_P the | |
2963 | range of values that result in the conditional having a true value. | |
2964 | ||
2965 | Return true if we are successful in extracting a range from COND and | |
2966 | false if we are unsuccessful. */ | |
2967 | ||
2968 | static bool | |
2969 | extract_range_from_cond (tree cond, tree *hi_p, tree *lo_p, int *inverted_p) | |
2970 | { | |
2971 | tree op1 = TREE_OPERAND (cond, 1); | |
2972 | tree high, low, type; | |
2973 | int inverted; | |
ee167207 EB |
2974 | |
2975 | type = TREE_TYPE (op1); | |
2976 | ||
6de9cd9a DN |
2977 | /* Experiments have shown that it's rarely, if ever useful to |
2978 | record ranges for enumerations. Presumably this is due to | |
2979 | the fact that they're rarely used directly. They are typically | |
2980 | cast into an integer type and used that way. */ | |
4d14c1f4 | 2981 | if (TREE_CODE (type) != INTEGER_TYPE) |
6de9cd9a DN |
2982 | return 0; |
2983 | ||
6de9cd9a DN |
2984 | switch (TREE_CODE (cond)) |
2985 | { | |
2986 | case EQ_EXPR: | |
2987 | high = low = op1; | |
2988 | inverted = 0; | |
2989 | break; | |
2990 | ||
2991 | case NE_EXPR: | |
2992 | high = low = op1; | |
2993 | inverted = 1; | |
2994 | break; | |
2995 | ||
2996 | case GE_EXPR: | |
2997 | low = op1; | |
4d14c1f4 RK |
2998 | |
2999 | /* Get the highest value of the type. If not a constant, use that | |
3000 | of its base type, if it has one. */ | |
6de9cd9a | 3001 | high = TYPE_MAX_VALUE (type); |
4d14c1f4 RK |
3002 | if (TREE_CODE (high) != INTEGER_CST && TREE_TYPE (type)) |
3003 | high = TYPE_MAX_VALUE (TREE_TYPE (type)); | |
6de9cd9a DN |
3004 | inverted = 0; |
3005 | break; | |
3006 | ||
3007 | case GT_EXPR: | |
6de9cd9a | 3008 | high = TYPE_MAX_VALUE (type); |
4d14c1f4 RK |
3009 | if (TREE_CODE (high) != INTEGER_CST && TREE_TYPE (type)) |
3010 | high = TYPE_MAX_VALUE (TREE_TYPE (type)); | |
75b9aa9f JJ |
3011 | if (!tree_int_cst_lt (op1, high)) |
3012 | return 0; | |
3013 | low = int_const_binop (PLUS_EXPR, op1, integer_one_node, 1); | |
6de9cd9a DN |
3014 | inverted = 0; |
3015 | break; | |
3016 | ||
3017 | case LE_EXPR: | |
3018 | high = op1; | |
3019 | low = TYPE_MIN_VALUE (type); | |
4d14c1f4 RK |
3020 | if (TREE_CODE (low) != INTEGER_CST && TREE_TYPE (type)) |
3021 | low = TYPE_MIN_VALUE (TREE_TYPE (type)); | |
6de9cd9a DN |
3022 | inverted = 0; |
3023 | break; | |
3024 | ||
3025 | case LT_EXPR: | |
6de9cd9a | 3026 | low = TYPE_MIN_VALUE (type); |
4d14c1f4 RK |
3027 | if (TREE_CODE (low) != INTEGER_CST && TREE_TYPE (type)) |
3028 | low = TYPE_MIN_VALUE (TREE_TYPE (type)); | |
19e1af6e | 3029 | if (!tree_int_cst_lt (low, op1)) |
75b9aa9f JJ |
3030 | return 0; |
3031 | high = int_const_binop (MINUS_EXPR, op1, integer_one_node, 1); | |
6de9cd9a DN |
3032 | inverted = 0; |
3033 | break; | |
3034 | ||
3035 | default: | |
3036 | return 0; | |
3037 | } | |
3038 | ||
3039 | *hi_p = high; | |
3040 | *lo_p = low; | |
3041 | *inverted_p = inverted; | |
3042 | return 1; | |
3043 | } | |
3044 | ||
3045 | /* Record a range created by COND for basic block BB. */ | |
3046 | ||
3047 | static void | |
fdabe5c2 | 3048 | record_range (tree cond, basic_block bb) |
6de9cd9a | 3049 | { |
efea75f9 JL |
3050 | enum tree_code code = TREE_CODE (cond); |
3051 | ||
3052 | /* We explicitly ignore NE_EXPRs and all the unordered comparisons. | |
3053 | They rarely allow for meaningful range optimizations and significantly | |
3054 | complicate the implementation. */ | |
3055 | if ((code == LT_EXPR || code == LE_EXPR || code == GT_EXPR | |
3056 | || code == GE_EXPR || code == EQ_EXPR) | |
6de9cd9a DN |
3057 | && TREE_CODE (TREE_TYPE (TREE_OPERAND (cond, 1))) == INTEGER_TYPE) |
3058 | { | |
23530866 JL |
3059 | struct vrp_hash_elt *vrp_hash_elt; |
3060 | struct vrp_element *element; | |
8184759d | 3061 | VEC(vrp_element_p,heap) **vrp_records_p; |
23530866 JL |
3062 | void **slot; |
3063 | ||
6de9cd9a | 3064 | |
e1111e8e | 3065 | vrp_hash_elt = XNEW (struct vrp_hash_elt); |
23530866 JL |
3066 | vrp_hash_elt->var = TREE_OPERAND (cond, 0); |
3067 | vrp_hash_elt->records = NULL; | |
3068 | slot = htab_find_slot (vrp_data, vrp_hash_elt, INSERT); | |
6de9cd9a | 3069 | |
23530866 | 3070 | if (*slot == NULL) |
4a198dea | 3071 | *slot = (void *) vrp_hash_elt; |
163075a0 | 3072 | else |
8184759d | 3073 | vrp_free (vrp_hash_elt); |
23530866 | 3074 | |
4a198dea | 3075 | vrp_hash_elt = (struct vrp_hash_elt *) *slot; |
23530866 JL |
3076 | vrp_records_p = &vrp_hash_elt->records; |
3077 | ||
e1111e8e | 3078 | element = GGC_NEW (struct vrp_element); |
6de9cd9a DN |
3079 | element->low = NULL; |
3080 | element->high = NULL; | |
3081 | element->cond = cond; | |
3082 | element->bb = bb; | |
3083 | ||
8184759d | 3084 | VEC_safe_push (vrp_element_p, heap, *vrp_records_p, element); |
d4e6fecb | 3085 | VEC_safe_push (tree, heap, vrp_variables_stack, TREE_OPERAND (cond, 0)); |
6de9cd9a DN |
3086 | } |
3087 | } | |
3088 | ||
23530866 JL |
3089 | /* Hashing and equality functions for VRP_DATA. |
3090 | ||
3091 | Since this hash table is addressed by SSA_NAMEs, we can hash on | |
3092 | their version number and equality can be determined with a | |
3093 | pointer comparison. */ | |
3094 | ||
3095 | static hashval_t | |
3096 | vrp_hash (const void *p) | |
3097 | { | |
3098 | tree var = ((struct vrp_hash_elt *)p)->var; | |
3099 | ||
3100 | return SSA_NAME_VERSION (var); | |
3101 | } | |
3102 | ||
3103 | static int | |
3104 | vrp_eq (const void *p1, const void *p2) | |
3105 | { | |
3106 | tree var1 = ((struct vrp_hash_elt *)p1)->var; | |
3107 | tree var2 = ((struct vrp_hash_elt *)p2)->var; | |
3108 | ||
3109 | return var1 == var2; | |
3110 | } | |
3111 | ||
6de9cd9a DN |
3112 | /* Hashing and equality functions for AVAIL_EXPRS. The table stores |
3113 | MODIFY_EXPR statements. We compute a value number for expressions using | |
3114 | the code of the expression and the SSA numbers of its operands. */ | |
3115 | ||
3116 | static hashval_t | |
3117 | avail_expr_hash (const void *p) | |
3118 | { | |
f47c96aa | 3119 | tree stmt = ((struct expr_hash_elt *)p)->stmt; |
6de9cd9a | 3120 | tree rhs = ((struct expr_hash_elt *)p)->rhs; |
f47c96aa AM |
3121 | tree vuse; |
3122 | ssa_op_iter iter; | |
6de9cd9a | 3123 | hashval_t val = 0; |
6de9cd9a DN |
3124 | |
3125 | /* iterative_hash_expr knows how to deal with any expression and | |
3126 | deals with commutative operators as well, so just use it instead | |
3127 | of duplicating such complexities here. */ | |
3128 | val = iterative_hash_expr (rhs, val); | |
3129 | ||
3130 | /* If the hash table entry is not associated with a statement, then we | |
3131 | can just hash the expression and not worry about virtual operands | |
3132 | and such. */ | |
f47c96aa | 3133 | if (!stmt || !stmt_ann (stmt)) |
6de9cd9a DN |
3134 | return val; |
3135 | ||
3136 | /* Add the SSA version numbers of every vuse operand. This is important | |
3137 | because compound variables like arrays are not renamed in the | |
3138 | operands. Rather, the rename is done on the virtual variable | |
3139 | representing all the elements of the array. */ | |
f47c96aa AM |
3140 | FOR_EACH_SSA_TREE_OPERAND (vuse, stmt, iter, SSA_OP_VUSE) |
3141 | val = iterative_hash_expr (vuse, val); | |
6de9cd9a DN |
3142 | |
3143 | return val; | |
3144 | } | |
3145 | ||
940db2c8 RH |
3146 | static hashval_t |
3147 | real_avail_expr_hash (const void *p) | |
3148 | { | |
3149 | return ((const struct expr_hash_elt *)p)->hash; | |
3150 | } | |
6de9cd9a DN |
3151 | |
3152 | static int | |
3153 | avail_expr_eq (const void *p1, const void *p2) | |
3154 | { | |
f47c96aa | 3155 | tree stmt1 = ((struct expr_hash_elt *)p1)->stmt; |
6de9cd9a | 3156 | tree rhs1 = ((struct expr_hash_elt *)p1)->rhs; |
f47c96aa | 3157 | tree stmt2 = ((struct expr_hash_elt *)p2)->stmt; |
6de9cd9a DN |
3158 | tree rhs2 = ((struct expr_hash_elt *)p2)->rhs; |
3159 | ||
3160 | /* If they are the same physical expression, return true. */ | |
f47c96aa | 3161 | if (rhs1 == rhs2 && stmt1 == stmt2) |
6de9cd9a DN |
3162 | return true; |
3163 | ||
3164 | /* If their codes are not equal, then quit now. */ | |
3165 | if (TREE_CODE (rhs1) != TREE_CODE (rhs2)) | |
3166 | return false; | |
3167 | ||
3168 | /* In case of a collision, both RHS have to be identical and have the | |
3169 | same VUSE operands. */ | |
3170 | if ((TREE_TYPE (rhs1) == TREE_TYPE (rhs2) | |
3171 | || lang_hooks.types_compatible_p (TREE_TYPE (rhs1), TREE_TYPE (rhs2))) | |
3172 | && operand_equal_p (rhs1, rhs2, OEP_PURE_SAME)) | |
3173 | { | |
f47c96aa AM |
3174 | bool ret = compare_ssa_operands_equal (stmt1, stmt2, SSA_OP_VUSE); |
3175 | gcc_assert (!ret || ((struct expr_hash_elt *)p1)->hash | |
1e128c5f | 3176 | == ((struct expr_hash_elt *)p2)->hash); |
f47c96aa | 3177 | return ret; |
6de9cd9a DN |
3178 | } |
3179 | ||
3180 | return false; | |
3181 | } |