1 | /* SCC value numbering for trees |
2 | Copyright (C) 2006-2024 Free Software Foundation, Inc. |
3 | Contributed by Daniel Berlin <dan@dberlin.org> |
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 3, 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 COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ |
20 | |
21 | #include "config.h" |
22 | #include "system.h" |
23 | #include "coretypes.h" |
24 | #include "splay-tree.h" |
25 | #include "backend.h" |
26 | #include "rtl.h" |
27 | #include "tree.h" |
28 | #include "gimple.h" |
29 | #include "ssa.h" |
30 | #include "expmed.h" |
31 | #include "insn-config.h" |
32 | #include "memmodel.h" |
33 | #include "emit-rtl.h" |
34 | #include "cgraph.h" |
35 | #include "gimple-pretty-print.h" |
36 | #include "alias.h" |
37 | #include "fold-const.h" |
38 | #include "stor-layout.h" |
39 | #include "cfganal.h" |
40 | #include "tree-inline.h" |
41 | #include "internal-fn.h" |
42 | #include "gimple-iterator.h" |
43 | #include "gimple-fold.h" |
44 | #include "tree-eh.h" |
45 | #include "gimplify.h" |
46 | #include "flags.h" |
47 | #include "dojump.h" |
48 | #include "explow.h" |
49 | #include "calls.h" |
50 | #include "varasm.h" |
51 | #include "stmt.h" |
52 | #include "expr.h" |
53 | #include "tree-dfa.h" |
54 | #include "tree-ssa.h" |
55 | #include "dumpfile.h" |
56 | #include "cfgloop.h" |
57 | #include "tree-ssa-propagate.h" |
58 | #include "tree-cfg.h" |
59 | #include "domwalk.h" |
60 | #include "gimple-match.h" |
61 | #include "stringpool.h" |
62 | #include "attribs.h" |
63 | #include "tree-pass.h" |
64 | #include "statistics.h" |
65 | #include "langhooks.h" |
66 | #include "ipa-utils.h" |
67 | #include "dbgcnt.h" |
68 | #include "tree-cfgcleanup.h" |
69 | #include "tree-ssa-loop.h" |
70 | #include "tree-scalar-evolution.h" |
71 | #include "tree-ssa-loop-niter.h" |
72 | #include "builtins.h" |
73 | #include "fold-const-call.h" |
74 | #include "ipa-modref-tree.h" |
75 | #include "ipa-modref.h" |
76 | #include "tree-ssa-sccvn.h" |
77 | #include "alloc-pool.h" |
78 | #include "symbol-summary.h" |
79 | #include "sreal.h" |
80 | #include "ipa-cp.h" |
81 | #include "ipa-prop.h" |
82 | #include "target.h" |
83 | |
84 | /* This algorithm is based on the SCC algorithm presented by Keith |
85 | Cooper and L. Taylor Simpson in "SCC-Based Value numbering" |
86 | (http://citeseer.ist.psu.edu/41805.html). In |
87 | straight line code, it is equivalent to a regular hash based value |
88 | numbering that is performed in reverse postorder. |
89 | |
90 | For code with cycles, there are two alternatives, both of which |
91 | require keeping the hashtables separate from the actual list of |
92 | value numbers for SSA names. |
93 | |
94 | 1. Iterate value numbering in an RPO walk of the blocks, removing |
95 | all the entries from the hashtable after each iteration (but |
96 | keeping the SSA name->value number mapping between iterations). |
97 | Iterate until it does not change. |
98 | |
99 | 2. Perform value numbering as part of an SCC walk on the SSA graph, |
100 | iterating only the cycles in the SSA graph until they do not change |
101 | (using a separate, optimistic hashtable for value numbering the SCC |
102 | operands). |
103 | |
104 | The second is not just faster in practice (because most SSA graph |
105 | cycles do not involve all the variables in the graph), it also has |
106 | some nice properties. |
107 | |
108 | One of these nice properties is that when we pop an SCC off the |
109 | stack, we are guaranteed to have processed all the operands coming from |
110 | *outside of that SCC*, so we do not need to do anything special to |
111 | ensure they have value numbers. |
112 | |
113 | Another nice property is that the SCC walk is done as part of a DFS |
114 | of the SSA graph, which makes it easy to perform combining and |
115 | simplifying operations at the same time. |
116 | |
117 | The code below is deliberately written in a way that makes it easy |
118 | to separate the SCC walk from the other work it does. |
119 | |
120 | In order to propagate constants through the code, we track which |
121 | expressions contain constants, and use those while folding. In |
122 | theory, we could also track expressions whose value numbers are |
123 | replaced, in case we end up folding based on expression |
124 | identities. |
125 | |
126 | In order to value number memory, we assign value numbers to vuses. |
127 | This enables us to note that, for example, stores to the same |
128 | address of the same value from the same starting memory states are |
129 | equivalent. |
130 | TODO: |
131 | |
132 | 1. We can iterate only the changing portions of the SCC's, but |
133 | I have not seen an SCC big enough for this to be a win. |
134 | 2. If you differentiate between phi nodes for loops and phi nodes |
135 | for if-then-else, you can properly consider phi nodes in different |
136 | blocks for equivalence. |
137 | 3. We could value number vuses in more cases, particularly, whole |
138 | structure copies. |
139 | */ |
140 | |
141 | /* There's no BB_EXECUTABLE but we can use BB_VISITED. */ |
142 | #define BB_EXECUTABLE BB_VISITED |
143 | |
144 | static vn_lookup_kind default_vn_walk_kind; |
145 | |
146 | /* vn_nary_op hashtable helpers. */ |
147 | |
148 | struct vn_nary_op_hasher : nofree_ptr_hash <vn_nary_op_s> |
149 | { |
150 | typedef vn_nary_op_s *compare_type; |
151 | static inline hashval_t hash (const vn_nary_op_s *); |
152 | static inline bool equal (const vn_nary_op_s *, const vn_nary_op_s *); |
153 | }; |
154 | |
155 | /* Return the computed hashcode for nary operation P1. */ |
156 | |
157 | inline hashval_t |
158 | vn_nary_op_hasher::hash (const vn_nary_op_s *vno1) |
159 | { |
160 | return vno1->hashcode; |
161 | } |
162 | |
163 | /* Compare nary operations P1 and P2 and return true if they are |
164 | equivalent. */ |
165 | |
166 | inline bool |
167 | vn_nary_op_hasher::equal (const vn_nary_op_s *vno1, const vn_nary_op_s *vno2) |
168 | { |
169 | return vno1 == vno2 || vn_nary_op_eq (vno1, vno2); |
170 | } |
171 | |
172 | typedef hash_table<vn_nary_op_hasher> vn_nary_op_table_type; |
173 | typedef vn_nary_op_table_type::iterator vn_nary_op_iterator_type; |
174 | |
175 | |
176 | /* vn_phi hashtable helpers. */ |
177 | |
178 | static int |
179 | vn_phi_eq (const_vn_phi_t const vp1, const_vn_phi_t const vp2); |
180 | |
181 | struct vn_phi_hasher : nofree_ptr_hash <vn_phi_s> |
182 | { |
183 | static inline hashval_t hash (const vn_phi_s *); |
184 | static inline bool equal (const vn_phi_s *, const vn_phi_s *); |
185 | }; |
186 | |
187 | /* Return the computed hashcode for phi operation P1. */ |
188 | |
189 | inline hashval_t |
190 | vn_phi_hasher::hash (const vn_phi_s *vp1) |
191 | { |
192 | return vp1->hashcode; |
193 | } |
194 | |
195 | /* Compare two phi entries for equality, ignoring VN_TOP arguments. */ |
196 | |
197 | inline bool |
198 | vn_phi_hasher::equal (const vn_phi_s *vp1, const vn_phi_s *vp2) |
199 | { |
200 | return vp1 == vp2 || vn_phi_eq (vp1, vp2); |
201 | } |
202 | |
203 | typedef hash_table<vn_phi_hasher> vn_phi_table_type; |
204 | typedef vn_phi_table_type::iterator vn_phi_iterator_type; |
205 | |
206 | |
207 | /* Compare two reference operands P1 and P2 for equality. Return true if |
208 | they are equal, and false otherwise. */ |
209 | |
210 | static int |
211 | vn_reference_op_eq (const void *p1, const void *p2) |
212 | { |
213 | const_vn_reference_op_t const vro1 = (const_vn_reference_op_t) p1; |
214 | const_vn_reference_op_t const vro2 = (const_vn_reference_op_t) p2; |
215 | |
216 | return (vro1->opcode == vro2->opcode |
217 | /* We do not care for differences in type qualification. */ |
218 | && (vro1->type == vro2->type |
219 | || (vro1->type && vro2->type |
220 | && types_compatible_p (TYPE_MAIN_VARIANT (vro1->type), |
221 | TYPE_MAIN_VARIANT (vro2->type)))) |
222 | && expressions_equal_p (vro1->op0, vro2->op0) |
223 | && expressions_equal_p (vro1->op1, vro2->op1) |
224 | && expressions_equal_p (vro1->op2, vro2->op2) |
225 | && (vro1->opcode != CALL_EXPR || vro1->clique == vro2->clique)); |
226 | } |
227 | |
228 | /* Free a reference operation structure VP. */ |
229 | |
230 | static inline void |
231 | free_reference (vn_reference_s *vr) |
232 | { |
233 | vr->operands.release (); |
234 | } |
235 | |
236 | |
237 | /* vn_reference hashtable helpers. */ |
238 | |
239 | struct vn_reference_hasher : nofree_ptr_hash <vn_reference_s> |
240 | { |
241 | static inline hashval_t hash (const vn_reference_s *); |
242 | static inline bool equal (const vn_reference_s *, const vn_reference_s *); |
243 | }; |
244 | |
245 | /* Return the hashcode for a given reference operation P1. */ |
246 | |
247 | inline hashval_t |
248 | vn_reference_hasher::hash (const vn_reference_s *vr1) |
249 | { |
250 | return vr1->hashcode; |
251 | } |
252 | |
253 | inline bool |
254 | vn_reference_hasher::equal (const vn_reference_s *v, const vn_reference_s *c) |
255 | { |
256 | return v == c || vn_reference_eq (v, c); |
257 | } |
258 | |
259 | typedef hash_table<vn_reference_hasher> vn_reference_table_type; |
260 | typedef vn_reference_table_type::iterator vn_reference_iterator_type; |
261 | |
262 | /* Pretty-print OPS to OUTFILE. */ |
263 | |
264 | void |
265 | print_vn_reference_ops (FILE *outfile, const vec<vn_reference_op_s> ops) |
266 | { |
267 | vn_reference_op_t vro; |
268 | unsigned int i; |
269 | fprintf (stream: outfile, format: "{" ); |
270 | for (i = 0; ops.iterate (ix: i, ptr: &vro); i++) |
271 | { |
272 | bool closebrace = false; |
273 | if (vro->opcode != SSA_NAME |
274 | && TREE_CODE_CLASS (vro->opcode) != tcc_declaration) |
275 | { |
276 | fprintf (stream: outfile, format: "%s" , get_tree_code_name (vro->opcode)); |
277 | if (vro->op0 || vro->opcode == CALL_EXPR) |
278 | { |
279 | fprintf (stream: outfile, format: "<" ); |
280 | closebrace = true; |
281 | } |
282 | } |
283 | if (vro->op0 || vro->opcode == CALL_EXPR) |
284 | { |
285 | if (!vro->op0) |
286 | fprintf (stream: outfile, format: internal_fn_name (fn: (internal_fn)vro->clique)); |
287 | else |
288 | print_generic_expr (outfile, vro->op0); |
289 | if (vro->op1) |
290 | { |
291 | fprintf (stream: outfile, format: "," ); |
292 | print_generic_expr (outfile, vro->op1); |
293 | } |
294 | if (vro->op2) |
295 | { |
296 | fprintf (stream: outfile, format: "," ); |
297 | print_generic_expr (outfile, vro->op2); |
298 | } |
299 | } |
300 | if (closebrace) |
301 | fprintf (stream: outfile, format: ">" ); |
302 | if (i != ops.length () - 1) |
303 | fprintf (stream: outfile, format: "," ); |
304 | } |
305 | fprintf (stream: outfile, format: "}" ); |
306 | } |
307 | |
308 | DEBUG_FUNCTION void |
309 | debug_vn_reference_ops (const vec<vn_reference_op_s> ops) |
310 | { |
311 | print_vn_reference_ops (stderr, ops); |
312 | fputc (c: '\n', stderr); |
313 | } |
314 | |
315 | /* The set of VN hashtables. */ |
316 | |
317 | typedef struct vn_tables_s |
318 | { |
319 | vn_nary_op_table_type *nary; |
320 | vn_phi_table_type *phis; |
321 | vn_reference_table_type *references; |
322 | } *vn_tables_t; |
323 | |
324 | |
325 | /* vn_constant hashtable helpers. */ |
326 | |
327 | struct vn_constant_hasher : free_ptr_hash <vn_constant_s> |
328 | { |
329 | static inline hashval_t hash (const vn_constant_s *); |
330 | static inline bool equal (const vn_constant_s *, const vn_constant_s *); |
331 | }; |
332 | |
333 | /* Hash table hash function for vn_constant_t. */ |
334 | |
335 | inline hashval_t |
336 | vn_constant_hasher::hash (const vn_constant_s *vc1) |
337 | { |
338 | return vc1->hashcode; |
339 | } |
340 | |
341 | /* Hash table equality function for vn_constant_t. */ |
342 | |
343 | inline bool |
344 | vn_constant_hasher::equal (const vn_constant_s *vc1, const vn_constant_s *vc2) |
345 | { |
346 | if (vc1->hashcode != vc2->hashcode) |
347 | return false; |
348 | |
349 | return vn_constant_eq_with_type (c1: vc1->constant, c2: vc2->constant); |
350 | } |
351 | |
352 | static hash_table<vn_constant_hasher> *constant_to_value_id; |
353 | |
354 | |
355 | /* Obstack we allocate the vn-tables elements from. */ |
356 | static obstack vn_tables_obstack; |
357 | /* Special obstack we never unwind. */ |
358 | static obstack vn_tables_insert_obstack; |
359 | |
360 | static vn_reference_t last_inserted_ref; |
361 | static vn_phi_t last_inserted_phi; |
362 | static vn_nary_op_t last_inserted_nary; |
363 | static vn_ssa_aux_t last_pushed_avail; |
364 | |
365 | /* Valid hashtables storing information we have proven to be |
366 | correct. */ |
367 | static vn_tables_t valid_info; |
368 | |
369 | |
370 | /* Valueization hook for simplify_replace_tree. Valueize NAME if it is |
371 | an SSA name, otherwise just return it. */ |
372 | tree (*vn_valueize) (tree); |
373 | static tree |
374 | vn_valueize_for_srt (tree t, void* context ATTRIBUTE_UNUSED) |
375 | { |
376 | basic_block saved_vn_context_bb = vn_context_bb; |
377 | /* Look for sth available at the definition block of the argument. |
378 | This avoids inconsistencies between availability there which |
379 | decides if the stmt can be removed and availability at the |
380 | use site. The SSA property ensures that things available |
381 | at the definition are also available at uses. */ |
382 | if (!SSA_NAME_IS_DEFAULT_DEF (t)) |
383 | vn_context_bb = gimple_bb (SSA_NAME_DEF_STMT (t)); |
384 | tree res = vn_valueize (t); |
385 | vn_context_bb = saved_vn_context_bb; |
386 | return res; |
387 | } |
388 | |
389 | |
390 | /* This represents the top of the VN lattice, which is the universal |
391 | value. */ |
392 | |
393 | tree VN_TOP; |
394 | |
395 | /* Unique counter for our value ids. */ |
396 | |
397 | static unsigned int next_value_id; |
398 | static int next_constant_value_id; |
399 | |
400 | |
401 | /* Table of vn_ssa_aux_t's, one per ssa_name. The vn_ssa_aux_t objects |
402 | are allocated on an obstack for locality reasons, and to free them |
403 | without looping over the vec. */ |
404 | |
405 | struct vn_ssa_aux_hasher : typed_noop_remove <vn_ssa_aux_t> |
406 | { |
407 | typedef vn_ssa_aux_t value_type; |
408 | typedef tree compare_type; |
409 | static inline hashval_t hash (const value_type &); |
410 | static inline bool equal (const value_type &, const compare_type &); |
411 | static inline void mark_deleted (value_type &) {} |
412 | static const bool empty_zero_p = true; |
413 | static inline void mark_empty (value_type &e) { e = NULL; } |
414 | static inline bool is_deleted (value_type &) { return false; } |
415 | static inline bool is_empty (value_type &e) { return e == NULL; } |
416 | }; |
417 | |
418 | hashval_t |
419 | vn_ssa_aux_hasher::hash (const value_type &entry) |
420 | { |
421 | return SSA_NAME_VERSION (entry->name); |
422 | } |
423 | |
424 | bool |
425 | vn_ssa_aux_hasher::equal (const value_type &entry, const compare_type &name) |
426 | { |
427 | return name == entry->name; |
428 | } |
429 | |
430 | static hash_table<vn_ssa_aux_hasher> *vn_ssa_aux_hash; |
431 | typedef hash_table<vn_ssa_aux_hasher>::iterator vn_ssa_aux_iterator_type; |
432 | static struct obstack vn_ssa_aux_obstack; |
433 | |
434 | static vn_nary_op_t vn_nary_op_insert_stmt (gimple *, tree); |
435 | static vn_nary_op_t vn_nary_op_insert_into (vn_nary_op_t, |
436 | vn_nary_op_table_type *); |
437 | static void init_vn_nary_op_from_pieces (vn_nary_op_t, unsigned int, |
438 | enum tree_code, tree, tree *); |
439 | static tree vn_lookup_simplify_result (gimple_match_op *); |
440 | static vn_reference_t vn_reference_lookup_or_insert_for_pieces |
441 | (tree, alias_set_type, alias_set_type, poly_int64, poly_int64, tree, |
442 | vec<vn_reference_op_s, va_heap>, tree); |
443 | |
444 | /* Return whether there is value numbering information for a given SSA name. */ |
445 | |
446 | bool |
447 | has_VN_INFO (tree name) |
448 | { |
449 | return vn_ssa_aux_hash->find_with_hash (comparable: name, SSA_NAME_VERSION (name)); |
450 | } |
451 | |
452 | vn_ssa_aux_t |
453 | VN_INFO (tree name) |
454 | { |
455 | vn_ssa_aux_t *res |
456 | = vn_ssa_aux_hash->find_slot_with_hash (comparable: name, SSA_NAME_VERSION (name), |
457 | insert: INSERT); |
458 | if (*res != NULL) |
459 | return *res; |
460 | |
461 | vn_ssa_aux_t newinfo = *res = XOBNEW (&vn_ssa_aux_obstack, struct vn_ssa_aux); |
462 | memset (s: newinfo, c: 0, n: sizeof (struct vn_ssa_aux)); |
463 | newinfo->name = name; |
464 | newinfo->valnum = VN_TOP; |
465 | /* We are using the visited flag to handle uses with defs not within the |
466 | region being value-numbered. */ |
467 | newinfo->visited = false; |
468 | |
469 | /* Given we create the VN_INFOs on-demand now we have to do initialization |
470 | different than VN_TOP here. */ |
471 | if (SSA_NAME_IS_DEFAULT_DEF (name)) |
472 | switch (TREE_CODE (SSA_NAME_VAR (name))) |
473 | { |
474 | case VAR_DECL: |
475 | /* All undefined vars are VARYING. */ |
476 | newinfo->valnum = name; |
477 | newinfo->visited = true; |
478 | break; |
479 | |
480 | case PARM_DECL: |
481 | /* Parameters are VARYING but we can record a condition |
482 | if we know it is a non-NULL pointer. */ |
483 | newinfo->visited = true; |
484 | newinfo->valnum = name; |
485 | if (POINTER_TYPE_P (TREE_TYPE (name)) |
486 | && nonnull_arg_p (SSA_NAME_VAR (name))) |
487 | { |
488 | tree ops[2]; |
489 | ops[0] = name; |
490 | ops[1] = build_int_cst (TREE_TYPE (name), 0); |
491 | vn_nary_op_t nary; |
492 | /* Allocate from non-unwinding stack. */ |
493 | nary = alloc_vn_nary_op_noinit (2, &vn_tables_insert_obstack); |
494 | init_vn_nary_op_from_pieces (nary, 2, NE_EXPR, |
495 | boolean_type_node, ops); |
496 | nary->predicated_values = 0; |
497 | nary->u.result = boolean_true_node; |
498 | vn_nary_op_insert_into (nary, valid_info->nary); |
499 | gcc_assert (nary->unwind_to == NULL); |
500 | /* Also do not link it into the undo chain. */ |
501 | last_inserted_nary = nary->next; |
502 | nary->next = (vn_nary_op_t)(void *)-1; |
503 | nary = alloc_vn_nary_op_noinit (2, &vn_tables_insert_obstack); |
504 | init_vn_nary_op_from_pieces (nary, 2, EQ_EXPR, |
505 | boolean_type_node, ops); |
506 | nary->predicated_values = 0; |
507 | nary->u.result = boolean_false_node; |
508 | vn_nary_op_insert_into (nary, valid_info->nary); |
509 | gcc_assert (nary->unwind_to == NULL); |
510 | last_inserted_nary = nary->next; |
511 | nary->next = (vn_nary_op_t)(void *)-1; |
512 | if (dump_file && (dump_flags & TDF_DETAILS)) |
513 | { |
514 | fprintf (stream: dump_file, format: "Recording " ); |
515 | print_generic_expr (dump_file, name, TDF_SLIM); |
516 | fprintf (stream: dump_file, format: " != 0\n" ); |
517 | } |
518 | } |
519 | break; |
520 | |
521 | case RESULT_DECL: |
522 | /* If the result is passed by invisible reference the default |
523 | def is initialized, otherwise it's uninitialized. Still |
524 | undefined is varying. */ |
525 | newinfo->visited = true; |
526 | newinfo->valnum = name; |
527 | break; |
528 | |
529 | default: |
530 | gcc_unreachable (); |
531 | } |
532 | return newinfo; |
533 | } |
534 | |
535 | /* Return the SSA value of X. */ |
536 | |
537 | inline tree |
538 | SSA_VAL (tree x, bool *visited = NULL) |
539 | { |
540 | vn_ssa_aux_t tem = vn_ssa_aux_hash->find_with_hash (comparable: x, SSA_NAME_VERSION (x)); |
541 | if (visited) |
542 | *visited = tem && tem->visited; |
543 | return tem && tem->visited ? tem->valnum : x; |
544 | } |
545 | |
546 | /* Return the SSA value of the VUSE x, supporting released VDEFs |
547 | during elimination which will value-number the VDEF to the |
548 | associated VUSE (but not substitute in the whole lattice). */ |
549 | |
550 | static inline tree |
551 | vuse_ssa_val (tree x) |
552 | { |
553 | if (!x) |
554 | return NULL_TREE; |
555 | |
556 | do |
557 | { |
558 | x = SSA_VAL (x); |
559 | gcc_assert (x != VN_TOP); |
560 | } |
561 | while (SSA_NAME_IN_FREE_LIST (x)); |
562 | |
563 | return x; |
564 | } |
565 | |
566 | /* Similar to the above but used as callback for walk_non_aliased_vuses |
567 | and thus should stop at unvisited VUSE to not walk across region |
568 | boundaries. */ |
569 | |
570 | static tree |
571 | vuse_valueize (tree vuse) |
572 | { |
573 | do |
574 | { |
575 | bool visited; |
576 | vuse = SSA_VAL (x: vuse, visited: &visited); |
577 | if (!visited) |
578 | return NULL_TREE; |
579 | gcc_assert (vuse != VN_TOP); |
580 | } |
581 | while (SSA_NAME_IN_FREE_LIST (vuse)); |
582 | return vuse; |
583 | } |
584 | |
585 | |
586 | /* Return the vn_kind the expression computed by the stmt should be |
587 | associated with. */ |
588 | |
589 | enum vn_kind |
590 | vn_get_stmt_kind (gimple *stmt) |
591 | { |
592 | switch (gimple_code (g: stmt)) |
593 | { |
594 | case GIMPLE_CALL: |
595 | return VN_REFERENCE; |
596 | case GIMPLE_PHI: |
597 | return VN_PHI; |
598 | case GIMPLE_ASSIGN: |
599 | { |
600 | enum tree_code code = gimple_assign_rhs_code (gs: stmt); |
601 | tree rhs1 = gimple_assign_rhs1 (gs: stmt); |
602 | switch (get_gimple_rhs_class (code)) |
603 | { |
604 | case GIMPLE_UNARY_RHS: |
605 | case GIMPLE_BINARY_RHS: |
606 | case GIMPLE_TERNARY_RHS: |
607 | return VN_NARY; |
608 | case GIMPLE_SINGLE_RHS: |
609 | switch (TREE_CODE_CLASS (code)) |
610 | { |
611 | case tcc_reference: |
612 | /* VOP-less references can go through unary case. */ |
613 | if ((code == REALPART_EXPR |
614 | || code == IMAGPART_EXPR |
615 | || code == VIEW_CONVERT_EXPR |
616 | || code == BIT_FIELD_REF) |
617 | && (TREE_CODE (TREE_OPERAND (rhs1, 0)) == SSA_NAME |
618 | || is_gimple_min_invariant (TREE_OPERAND (rhs1, 0)))) |
619 | return VN_NARY; |
620 | |
621 | /* Fallthrough. */ |
622 | case tcc_declaration: |
623 | return VN_REFERENCE; |
624 | |
625 | case tcc_constant: |
626 | return VN_CONSTANT; |
627 | |
628 | default: |
629 | if (code == ADDR_EXPR) |
630 | return (is_gimple_min_invariant (rhs1) |
631 | ? VN_CONSTANT : VN_REFERENCE); |
632 | else if (code == CONSTRUCTOR) |
633 | return VN_NARY; |
634 | return VN_NONE; |
635 | } |
636 | default: |
637 | return VN_NONE; |
638 | } |
639 | } |
640 | default: |
641 | return VN_NONE; |
642 | } |
643 | } |
644 | |
645 | /* Lookup a value id for CONSTANT and return it. If it does not |
646 | exist returns 0. */ |
647 | |
648 | unsigned int |
649 | get_constant_value_id (tree constant) |
650 | { |
651 | vn_constant_s **slot; |
652 | struct vn_constant_s vc; |
653 | |
654 | vc.hashcode = vn_hash_constant_with_type (constant); |
655 | vc.constant = constant; |
656 | slot = constant_to_value_id->find_slot (value: &vc, insert: NO_INSERT); |
657 | if (slot) |
658 | return (*slot)->value_id; |
659 | return 0; |
660 | } |
661 | |
662 | /* Lookup a value id for CONSTANT, and if it does not exist, create a |
663 | new one and return it. If it does exist, return it. */ |
664 | |
665 | unsigned int |
666 | get_or_alloc_constant_value_id (tree constant) |
667 | { |
668 | vn_constant_s **slot; |
669 | struct vn_constant_s vc; |
670 | vn_constant_t vcp; |
671 | |
672 | /* If the hashtable isn't initialized we're not running from PRE and thus |
673 | do not need value-ids. */ |
674 | if (!constant_to_value_id) |
675 | return 0; |
676 | |
677 | vc.hashcode = vn_hash_constant_with_type (constant); |
678 | vc.constant = constant; |
679 | slot = constant_to_value_id->find_slot (value: &vc, insert: INSERT); |
680 | if (*slot) |
681 | return (*slot)->value_id; |
682 | |
683 | vcp = XNEW (struct vn_constant_s); |
684 | vcp->hashcode = vc.hashcode; |
685 | vcp->constant = constant; |
686 | vcp->value_id = get_next_constant_value_id (); |
687 | *slot = vcp; |
688 | return vcp->value_id; |
689 | } |
690 | |
691 | /* Compute the hash for a reference operand VRO1. */ |
692 | |
693 | static void |
694 | vn_reference_op_compute_hash (const vn_reference_op_t vro1, inchash::hash &hstate) |
695 | { |
696 | hstate.add_int (v: vro1->opcode); |
697 | if (vro1->opcode == CALL_EXPR && !vro1->op0) |
698 | hstate.add_int (v: vro1->clique); |
699 | if (vro1->op0) |
700 | inchash::add_expr (vro1->op0, hstate); |
701 | if (vro1->op1) |
702 | inchash::add_expr (vro1->op1, hstate); |
703 | if (vro1->op2) |
704 | inchash::add_expr (vro1->op2, hstate); |
705 | } |
706 | |
707 | /* Compute a hash for the reference operation VR1 and return it. */ |
708 | |
709 | static hashval_t |
710 | vn_reference_compute_hash (const vn_reference_t vr1) |
711 | { |
712 | inchash::hash hstate; |
713 | hashval_t result; |
714 | int i; |
715 | vn_reference_op_t vro; |
716 | poly_int64 off = -1; |
717 | bool deref = false; |
718 | |
719 | FOR_EACH_VEC_ELT (vr1->operands, i, vro) |
720 | { |
721 | if (vro->opcode == MEM_REF) |
722 | deref = true; |
723 | else if (vro->opcode != ADDR_EXPR) |
724 | deref = false; |
725 | if (maybe_ne (a: vro->off, b: -1)) |
726 | { |
727 | if (known_eq (off, -1)) |
728 | off = 0; |
729 | off += vro->off; |
730 | } |
731 | else |
732 | { |
733 | if (maybe_ne (a: off, b: -1) |
734 | && maybe_ne (a: off, b: 0)) |
735 | hstate.add_poly_int (v: off); |
736 | off = -1; |
737 | if (deref |
738 | && vro->opcode == ADDR_EXPR) |
739 | { |
740 | if (vro->op0) |
741 | { |
742 | tree op = TREE_OPERAND (vro->op0, 0); |
743 | hstate.add_int (TREE_CODE (op)); |
744 | inchash::add_expr (op, hstate); |
745 | } |
746 | } |
747 | else |
748 | vn_reference_op_compute_hash (vro1: vro, hstate); |
749 | } |
750 | } |
751 | /* Do not hash vr1->offset or vr1->max_size, we want to get collisions |
752 | to be able to identify compatible results. */ |
753 | result = hstate.end (); |
754 | /* ??? We would ICE later if we hash instead of adding that in. */ |
755 | if (vr1->vuse) |
756 | result += SSA_NAME_VERSION (vr1->vuse); |
757 | |
758 | return result; |
759 | } |
760 | |
761 | /* Return true if reference operations VR1 and VR2 are equivalent. This |
762 | means they have the same set of operands and vuses. */ |
763 | |
764 | bool |
765 | vn_reference_eq (const_vn_reference_t const vr1, const_vn_reference_t const vr2) |
766 | { |
767 | unsigned i, j; |
768 | |
769 | /* Early out if this is not a hash collision. */ |
770 | if (vr1->hashcode != vr2->hashcode) |
771 | return false; |
772 | |
773 | /* The VOP needs to be the same. */ |
774 | if (vr1->vuse != vr2->vuse) |
775 | return false; |
776 | |
777 | /* The offset/max_size used for the ao_ref during lookup has to be |
778 | the same. */ |
779 | if (maybe_ne (a: vr1->offset, b: vr2->offset) |
780 | || maybe_ne (a: vr1->max_size, b: vr2->max_size)) |
781 | { |
782 | /* But nothing known in the prevailing entry is OK to be used. */ |
783 | if (maybe_ne (a: vr1->offset, b: 0) || known_size_p (a: vr1->max_size)) |
784 | return false; |
785 | } |
786 | |
787 | /* If the operands are the same we are done. */ |
788 | if (vr1->operands == vr2->operands) |
789 | return true; |
790 | |
791 | if (!vr1->type || !vr2->type) |
792 | { |
793 | if (vr1->type != vr2->type) |
794 | return false; |
795 | } |
796 | else if (vr1->type == vr2->type) |
797 | ; |
798 | else if (COMPLETE_TYPE_P (vr1->type) != COMPLETE_TYPE_P (vr2->type) |
799 | || (COMPLETE_TYPE_P (vr1->type) |
800 | && !expressions_equal_p (TYPE_SIZE (vr1->type), |
801 | TYPE_SIZE (vr2->type)))) |
802 | return false; |
803 | else if (vr1->operands[0].opcode == CALL_EXPR |
804 | && !types_compatible_p (type1: vr1->type, type2: vr2->type)) |
805 | return false; |
806 | else if (INTEGRAL_TYPE_P (vr1->type) |
807 | && INTEGRAL_TYPE_P (vr2->type)) |
808 | { |
809 | if (TYPE_PRECISION (vr1->type) != TYPE_PRECISION (vr2->type)) |
810 | return false; |
811 | } |
812 | else if (INTEGRAL_TYPE_P (vr1->type) |
813 | && (TYPE_PRECISION (vr1->type) |
814 | != TREE_INT_CST_LOW (TYPE_SIZE (vr1->type)))) |
815 | return false; |
816 | else if (INTEGRAL_TYPE_P (vr2->type) |
817 | && (TYPE_PRECISION (vr2->type) |
818 | != TREE_INT_CST_LOW (TYPE_SIZE (vr2->type)))) |
819 | return false; |
820 | else if (VECTOR_BOOLEAN_TYPE_P (vr1->type) |
821 | && VECTOR_BOOLEAN_TYPE_P (vr2->type)) |
822 | { |
823 | /* Vector boolean types can have padding, verify we are dealing with |
824 | the same number of elements, aka the precision of the types. |
825 | For example, In most architecture the precision_size of vbool*_t |
826 | types are caculated like below: |
827 | precision_size = type_size * 8 |
828 | |
829 | Unfortunately, the RISC-V will adjust the precision_size for the |
830 | vbool*_t in order to align the ISA as below: |
831 | type_size = [1, 1, 1, 1, 2, 4, 8] |
832 | precision_size = [1, 2, 4, 8, 16, 32, 64] |
833 | |
834 | Then the precision_size of RISC-V vbool*_t will not be the multiple |
835 | of the type_size. We take care of this case consolidated here. */ |
836 | if (maybe_ne (a: TYPE_VECTOR_SUBPARTS (node: vr1->type), |
837 | b: TYPE_VECTOR_SUBPARTS (node: vr2->type))) |
838 | return false; |
839 | } |
840 | |
841 | i = 0; |
842 | j = 0; |
843 | do |
844 | { |
845 | poly_int64 off1 = 0, off2 = 0; |
846 | vn_reference_op_t vro1, vro2; |
847 | vn_reference_op_s tem1, tem2; |
848 | bool deref1 = false, deref2 = false; |
849 | bool reverse1 = false, reverse2 = false; |
850 | for (; vr1->operands.iterate (ix: i, ptr: &vro1); i++) |
851 | { |
852 | if (vro1->opcode == MEM_REF) |
853 | deref1 = true; |
854 | /* Do not look through a storage order barrier. */ |
855 | else if (vro1->opcode == VIEW_CONVERT_EXPR && vro1->reverse) |
856 | return false; |
857 | reverse1 |= vro1->reverse; |
858 | if (known_eq (vro1->off, -1)) |
859 | break; |
860 | off1 += vro1->off; |
861 | } |
862 | for (; vr2->operands.iterate (ix: j, ptr: &vro2); j++) |
863 | { |
864 | if (vro2->opcode == MEM_REF) |
865 | deref2 = true; |
866 | /* Do not look through a storage order barrier. */ |
867 | else if (vro2->opcode == VIEW_CONVERT_EXPR && vro2->reverse) |
868 | return false; |
869 | reverse2 |= vro2->reverse; |
870 | if (known_eq (vro2->off, -1)) |
871 | break; |
872 | off2 += vro2->off; |
873 | } |
874 | if (maybe_ne (a: off1, b: off2) || reverse1 != reverse2) |
875 | return false; |
876 | if (deref1 && vro1->opcode == ADDR_EXPR) |
877 | { |
878 | memset (s: &tem1, c: 0, n: sizeof (tem1)); |
879 | tem1.op0 = TREE_OPERAND (vro1->op0, 0); |
880 | tem1.type = TREE_TYPE (tem1.op0); |
881 | tem1.opcode = TREE_CODE (tem1.op0); |
882 | vro1 = &tem1; |
883 | deref1 = false; |
884 | } |
885 | if (deref2 && vro2->opcode == ADDR_EXPR) |
886 | { |
887 | memset (s: &tem2, c: 0, n: sizeof (tem2)); |
888 | tem2.op0 = TREE_OPERAND (vro2->op0, 0); |
889 | tem2.type = TREE_TYPE (tem2.op0); |
890 | tem2.opcode = TREE_CODE (tem2.op0); |
891 | vro2 = &tem2; |
892 | deref2 = false; |
893 | } |
894 | if (deref1 != deref2) |
895 | return false; |
896 | if (!vn_reference_op_eq (p1: vro1, p2: vro2)) |
897 | return false; |
898 | ++j; |
899 | ++i; |
900 | } |
901 | while (vr1->operands.length () != i |
902 | || vr2->operands.length () != j); |
903 | |
904 | return true; |
905 | } |
906 | |
907 | /* Copy the operations present in load/store REF into RESULT, a vector of |
908 | vn_reference_op_s's. */ |
909 | |
910 | static void |
911 | copy_reference_ops_from_ref (tree ref, vec<vn_reference_op_s> *result) |
912 | { |
913 | /* For non-calls, store the information that makes up the address. */ |
914 | tree orig = ref; |
915 | while (ref) |
916 | { |
917 | vn_reference_op_s temp; |
918 | |
919 | memset (s: &temp, c: 0, n: sizeof (temp)); |
920 | temp.type = TREE_TYPE (ref); |
921 | temp.opcode = TREE_CODE (ref); |
922 | temp.off = -1; |
923 | |
924 | switch (temp.opcode) |
925 | { |
926 | case MODIFY_EXPR: |
927 | temp.op0 = TREE_OPERAND (ref, 1); |
928 | break; |
929 | case WITH_SIZE_EXPR: |
930 | temp.op0 = TREE_OPERAND (ref, 1); |
931 | temp.off = 0; |
932 | break; |
933 | case MEM_REF: |
934 | /* The base address gets its own vn_reference_op_s structure. */ |
935 | temp.op0 = TREE_OPERAND (ref, 1); |
936 | if (!mem_ref_offset (ref).to_shwi (r: &temp.off)) |
937 | temp.off = -1; |
938 | temp.clique = MR_DEPENDENCE_CLIQUE (ref); |
939 | temp.base = MR_DEPENDENCE_BASE (ref); |
940 | temp.reverse = REF_REVERSE_STORAGE_ORDER (ref); |
941 | break; |
942 | case TARGET_MEM_REF: |
943 | /* The base address gets its own vn_reference_op_s structure. */ |
944 | temp.op0 = TMR_INDEX (ref); |
945 | temp.op1 = TMR_STEP (ref); |
946 | temp.op2 = TMR_OFFSET (ref); |
947 | temp.clique = MR_DEPENDENCE_CLIQUE (ref); |
948 | temp.base = MR_DEPENDENCE_BASE (ref); |
949 | result->safe_push (obj: temp); |
950 | memset (s: &temp, c: 0, n: sizeof (temp)); |
951 | temp.type = NULL_TREE; |
952 | temp.opcode = ERROR_MARK; |
953 | temp.op0 = TMR_INDEX2 (ref); |
954 | temp.off = -1; |
955 | break; |
956 | case BIT_FIELD_REF: |
957 | /* Record bits, position and storage order. */ |
958 | temp.op0 = TREE_OPERAND (ref, 1); |
959 | temp.op1 = TREE_OPERAND (ref, 2); |
960 | if (!multiple_p (a: bit_field_offset (t: ref), BITS_PER_UNIT, multiple: &temp.off)) |
961 | temp.off = -1; |
962 | temp.reverse = REF_REVERSE_STORAGE_ORDER (ref); |
963 | break; |
964 | case COMPONENT_REF: |
965 | /* The field decl is enough to unambiguously specify the field, |
966 | so use its type here. */ |
967 | temp.type = TREE_TYPE (TREE_OPERAND (ref, 1)); |
968 | temp.op0 = TREE_OPERAND (ref, 1); |
969 | temp.op1 = TREE_OPERAND (ref, 2); |
970 | temp.reverse = (AGGREGATE_TYPE_P (TREE_TYPE (TREE_OPERAND (ref, 0))) |
971 | && TYPE_REVERSE_STORAGE_ORDER |
972 | (TREE_TYPE (TREE_OPERAND (ref, 0)))); |
973 | { |
974 | tree this_offset = component_ref_field_offset (ref); |
975 | if (this_offset |
976 | && poly_int_tree_p (t: this_offset)) |
977 | { |
978 | tree bit_offset = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1)); |
979 | if (TREE_INT_CST_LOW (bit_offset) % BITS_PER_UNIT == 0) |
980 | { |
981 | poly_offset_int off |
982 | = (wi::to_poly_offset (t: this_offset) |
983 | + (wi::to_offset (t: bit_offset) >> LOG2_BITS_PER_UNIT)); |
984 | /* Probibit value-numbering zero offset components |
985 | of addresses the same before the pass folding |
986 | __builtin_object_size had a chance to run. */ |
987 | if (TREE_CODE (orig) != ADDR_EXPR |
988 | || maybe_ne (a: off, b: 0) |
989 | || (cfun->curr_properties & PROP_objsz)) |
990 | off.to_shwi (r: &temp.off); |
991 | } |
992 | } |
993 | } |
994 | break; |
995 | case ARRAY_RANGE_REF: |
996 | case ARRAY_REF: |
997 | { |
998 | tree eltype = TREE_TYPE (TREE_TYPE (TREE_OPERAND (ref, 0))); |
999 | /* Record index as operand. */ |
1000 | temp.op0 = TREE_OPERAND (ref, 1); |
1001 | /* Always record lower bounds and element size. */ |
1002 | temp.op1 = array_ref_low_bound (ref); |
1003 | /* But record element size in units of the type alignment. */ |
1004 | temp.op2 = TREE_OPERAND (ref, 3); |
1005 | temp.align = eltype->type_common.align; |
1006 | if (! temp.op2) |
1007 | temp.op2 = size_binop (EXACT_DIV_EXPR, TYPE_SIZE_UNIT (eltype), |
1008 | size_int (TYPE_ALIGN_UNIT (eltype))); |
1009 | if (poly_int_tree_p (t: temp.op0) |
1010 | && poly_int_tree_p (t: temp.op1) |
1011 | && TREE_CODE (temp.op2) == INTEGER_CST) |
1012 | { |
1013 | poly_offset_int off = ((wi::to_poly_offset (t: temp.op0) |
1014 | - wi::to_poly_offset (t: temp.op1)) |
1015 | * wi::to_offset (t: temp.op2) |
1016 | * vn_ref_op_align_unit (op: &temp)); |
1017 | off.to_shwi (r: &temp.off); |
1018 | } |
1019 | temp.reverse = (AGGREGATE_TYPE_P (TREE_TYPE (TREE_OPERAND (ref, 0))) |
1020 | && TYPE_REVERSE_STORAGE_ORDER |
1021 | (TREE_TYPE (TREE_OPERAND (ref, 0)))); |
1022 | } |
1023 | break; |
1024 | case VAR_DECL: |
1025 | if (DECL_HARD_REGISTER (ref)) |
1026 | { |
1027 | temp.op0 = ref; |
1028 | break; |
1029 | } |
1030 | /* Fallthru. */ |
1031 | case PARM_DECL: |
1032 | case CONST_DECL: |
1033 | case RESULT_DECL: |
1034 | /* Canonicalize decls to MEM[&decl] which is what we end up with |
1035 | when valueizing MEM[ptr] with ptr = &decl. */ |
1036 | temp.opcode = MEM_REF; |
1037 | temp.op0 = build_int_cst (build_pointer_type (TREE_TYPE (ref)), 0); |
1038 | temp.off = 0; |
1039 | result->safe_push (obj: temp); |
1040 | temp.opcode = ADDR_EXPR; |
1041 | temp.op0 = build1 (ADDR_EXPR, TREE_TYPE (temp.op0), ref); |
1042 | temp.type = TREE_TYPE (temp.op0); |
1043 | temp.off = -1; |
1044 | break; |
1045 | case STRING_CST: |
1046 | case INTEGER_CST: |
1047 | case POLY_INT_CST: |
1048 | case COMPLEX_CST: |
1049 | case VECTOR_CST: |
1050 | case REAL_CST: |
1051 | case FIXED_CST: |
1052 | case CONSTRUCTOR: |
1053 | case SSA_NAME: |
1054 | temp.op0 = ref; |
1055 | break; |
1056 | case ADDR_EXPR: |
1057 | if (is_gimple_min_invariant (ref)) |
1058 | { |
1059 | temp.op0 = ref; |
1060 | break; |
1061 | } |
1062 | break; |
1063 | /* These are only interesting for their operands, their |
1064 | existence, and their type. They will never be the last |
1065 | ref in the chain of references (IE they require an |
1066 | operand), so we don't have to put anything |
1067 | for op* as it will be handled by the iteration */ |
1068 | case REALPART_EXPR: |
1069 | temp.off = 0; |
1070 | break; |
1071 | case VIEW_CONVERT_EXPR: |
1072 | temp.off = 0; |
1073 | temp.reverse = storage_order_barrier_p (t: ref); |
1074 | break; |
1075 | case IMAGPART_EXPR: |
1076 | /* This is only interesting for its constant offset. */ |
1077 | temp.off = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (ref))); |
1078 | break; |
1079 | default: |
1080 | gcc_unreachable (); |
1081 | } |
1082 | result->safe_push (obj: temp); |
1083 | |
1084 | if (REFERENCE_CLASS_P (ref) |
1085 | || TREE_CODE (ref) == MODIFY_EXPR |
1086 | || TREE_CODE (ref) == WITH_SIZE_EXPR |
1087 | || (TREE_CODE (ref) == ADDR_EXPR |
1088 | && !is_gimple_min_invariant (ref))) |
1089 | ref = TREE_OPERAND (ref, 0); |
1090 | else |
1091 | ref = NULL_TREE; |
1092 | } |
1093 | } |
1094 | |
1095 | /* Build a alias-oracle reference abstraction in *REF from the vn_reference |
1096 | operands in *OPS, the reference alias set SET and the reference type TYPE. |
1097 | Return true if something useful was produced. */ |
1098 | |
1099 | bool |
1100 | ao_ref_init_from_vn_reference (ao_ref *ref, |
1101 | alias_set_type set, alias_set_type base_set, |
1102 | tree type, const vec<vn_reference_op_s> &ops) |
1103 | { |
1104 | unsigned i; |
1105 | tree base = NULL_TREE; |
1106 | tree *op0_p = &base; |
1107 | poly_offset_int offset = 0; |
1108 | poly_offset_int max_size; |
1109 | poly_offset_int size = -1; |
1110 | tree size_tree = NULL_TREE; |
1111 | |
1112 | /* We don't handle calls. */ |
1113 | if (!type) |
1114 | return false; |
1115 | |
1116 | machine_mode mode = TYPE_MODE (type); |
1117 | if (mode == BLKmode) |
1118 | size_tree = TYPE_SIZE (type); |
1119 | else |
1120 | size = GET_MODE_BITSIZE (mode); |
1121 | if (size_tree != NULL_TREE |
1122 | && poly_int_tree_p (t: size_tree)) |
1123 | size = wi::to_poly_offset (t: size_tree); |
1124 | |
1125 | /* Lower the final access size from the outermost expression. */ |
1126 | const_vn_reference_op_t cst_op = &ops[0]; |
1127 | /* Cast away constness for the sake of the const-unsafe |
1128 | FOR_EACH_VEC_ELT(). */ |
1129 | vn_reference_op_t op = const_cast<vn_reference_op_t>(cst_op); |
1130 | size_tree = NULL_TREE; |
1131 | if (op->opcode == COMPONENT_REF) |
1132 | size_tree = DECL_SIZE (op->op0); |
1133 | else if (op->opcode == BIT_FIELD_REF) |
1134 | size_tree = op->op0; |
1135 | if (size_tree != NULL_TREE |
1136 | && poly_int_tree_p (t: size_tree) |
1137 | && (!known_size_p (a: size) |
1138 | || known_lt (wi::to_poly_offset (size_tree), size))) |
1139 | size = wi::to_poly_offset (t: size_tree); |
1140 | |
1141 | /* Initially, maxsize is the same as the accessed element size. |
1142 | In the following it will only grow (or become -1). */ |
1143 | max_size = size; |
1144 | |
1145 | /* Compute cumulative bit-offset for nested component-refs and array-refs, |
1146 | and find the ultimate containing object. */ |
1147 | FOR_EACH_VEC_ELT (ops, i, op) |
1148 | { |
1149 | switch (op->opcode) |
1150 | { |
1151 | /* These may be in the reference ops, but we cannot do anything |
1152 | sensible with them here. */ |
1153 | case ADDR_EXPR: |
1154 | /* Apart from ADDR_EXPR arguments to MEM_REF. */ |
1155 | if (base != NULL_TREE |
1156 | && TREE_CODE (base) == MEM_REF |
1157 | && op->op0 |
1158 | && DECL_P (TREE_OPERAND (op->op0, 0))) |
1159 | { |
1160 | const_vn_reference_op_t pop = &ops[i-1]; |
1161 | base = TREE_OPERAND (op->op0, 0); |
1162 | if (known_eq (pop->off, -1)) |
1163 | { |
1164 | max_size = -1; |
1165 | offset = 0; |
1166 | } |
1167 | else |
1168 | offset += pop->off * BITS_PER_UNIT; |
1169 | op0_p = NULL; |
1170 | break; |
1171 | } |
1172 | /* Fallthru. */ |
1173 | case CALL_EXPR: |
1174 | return false; |
1175 | |
1176 | /* Record the base objects. */ |
1177 | case MEM_REF: |
1178 | *op0_p = build2 (MEM_REF, op->type, |
1179 | NULL_TREE, op->op0); |
1180 | MR_DEPENDENCE_CLIQUE (*op0_p) = op->clique; |
1181 | MR_DEPENDENCE_BASE (*op0_p) = op->base; |
1182 | op0_p = &TREE_OPERAND (*op0_p, 0); |
1183 | break; |
1184 | |
1185 | case VAR_DECL: |
1186 | case PARM_DECL: |
1187 | case RESULT_DECL: |
1188 | case SSA_NAME: |
1189 | *op0_p = op->op0; |
1190 | op0_p = NULL; |
1191 | break; |
1192 | |
1193 | /* And now the usual component-reference style ops. */ |
1194 | case BIT_FIELD_REF: |
1195 | offset += wi::to_poly_offset (t: op->op1); |
1196 | break; |
1197 | |
1198 | case COMPONENT_REF: |
1199 | { |
1200 | tree field = op->op0; |
1201 | /* We do not have a complete COMPONENT_REF tree here so we |
1202 | cannot use component_ref_field_offset. Do the interesting |
1203 | parts manually. */ |
1204 | tree this_offset = DECL_FIELD_OFFSET (field); |
1205 | |
1206 | if (op->op1 || !poly_int_tree_p (t: this_offset)) |
1207 | max_size = -1; |
1208 | else |
1209 | { |
1210 | poly_offset_int woffset = (wi::to_poly_offset (t: this_offset) |
1211 | << LOG2_BITS_PER_UNIT); |
1212 | woffset += wi::to_offset (DECL_FIELD_BIT_OFFSET (field)); |
1213 | offset += woffset; |
1214 | } |
1215 | break; |
1216 | } |
1217 | |
1218 | case ARRAY_RANGE_REF: |
1219 | case ARRAY_REF: |
1220 | /* Use the recorded constant offset. */ |
1221 | if (maybe_eq (a: op->off, b: -1)) |
1222 | max_size = -1; |
1223 | else |
1224 | offset += op->off * BITS_PER_UNIT; |
1225 | break; |
1226 | |
1227 | case REALPART_EXPR: |
1228 | break; |
1229 | |
1230 | case IMAGPART_EXPR: |
1231 | offset += size; |
1232 | break; |
1233 | |
1234 | case VIEW_CONVERT_EXPR: |
1235 | break; |
1236 | |
1237 | case STRING_CST: |
1238 | case INTEGER_CST: |
1239 | case COMPLEX_CST: |
1240 | case VECTOR_CST: |
1241 | case REAL_CST: |
1242 | case CONSTRUCTOR: |
1243 | case CONST_DECL: |
1244 | return false; |
1245 | |
1246 | default: |
1247 | return false; |
1248 | } |
1249 | } |
1250 | |
1251 | if (base == NULL_TREE) |
1252 | return false; |
1253 | |
1254 | ref->ref = NULL_TREE; |
1255 | ref->base = base; |
1256 | ref->ref_alias_set = set; |
1257 | ref->base_alias_set = base_set; |
1258 | /* We discount volatiles from value-numbering elsewhere. */ |
1259 | ref->volatile_p = false; |
1260 | |
1261 | if (!size.to_shwi (r: &ref->size) || maybe_lt (a: ref->size, b: 0)) |
1262 | { |
1263 | ref->offset = 0; |
1264 | ref->size = -1; |
1265 | ref->max_size = -1; |
1266 | return true; |
1267 | } |
1268 | |
1269 | if (!offset.to_shwi (r: &ref->offset)) |
1270 | { |
1271 | ref->offset = 0; |
1272 | ref->max_size = -1; |
1273 | return true; |
1274 | } |
1275 | |
1276 | if (!max_size.to_shwi (r: &ref->max_size) || maybe_lt (a: ref->max_size, b: 0)) |
1277 | ref->max_size = -1; |
1278 | |
1279 | return true; |
1280 | } |
1281 | |
1282 | /* Copy the operations present in load/store/call REF into RESULT, a vector of |
1283 | vn_reference_op_s's. */ |
1284 | |
1285 | static void |
1286 | copy_reference_ops_from_call (gcall *call, |
1287 | vec<vn_reference_op_s> *result) |
1288 | { |
1289 | vn_reference_op_s temp; |
1290 | unsigned i; |
1291 | tree lhs = gimple_call_lhs (gs: call); |
1292 | int lr; |
1293 | |
1294 | /* If 2 calls have a different non-ssa lhs, vdef value numbers should be |
1295 | different. By adding the lhs here in the vector, we ensure that the |
1296 | hashcode is different, guaranteeing a different value number. */ |
1297 | if (lhs && TREE_CODE (lhs) != SSA_NAME) |
1298 | { |
1299 | memset (s: &temp, c: 0, n: sizeof (temp)); |
1300 | temp.opcode = MODIFY_EXPR; |
1301 | temp.type = TREE_TYPE (lhs); |
1302 | temp.op0 = lhs; |
1303 | temp.off = -1; |
1304 | result->safe_push (obj: temp); |
1305 | } |
1306 | |
1307 | /* Copy the type, opcode, function, static chain and EH region, if any. */ |
1308 | memset (s: &temp, c: 0, n: sizeof (temp)); |
1309 | temp.type = gimple_call_fntype (gs: call); |
1310 | temp.opcode = CALL_EXPR; |
1311 | temp.op0 = gimple_call_fn (gs: call); |
1312 | if (gimple_call_internal_p (gs: call)) |
1313 | temp.clique = gimple_call_internal_fn (gs: call); |
1314 | temp.op1 = gimple_call_chain (gs: call); |
1315 | if (stmt_could_throw_p (cfun, call) && (lr = lookup_stmt_eh_lp (call)) > 0) |
1316 | temp.op2 = size_int (lr); |
1317 | temp.off = -1; |
1318 | result->safe_push (obj: temp); |
1319 | |
1320 | /* Copy the call arguments. As they can be references as well, |
1321 | just chain them together. */ |
1322 | for (i = 0; i < gimple_call_num_args (gs: call); ++i) |
1323 | { |
1324 | tree callarg = gimple_call_arg (gs: call, index: i); |
1325 | copy_reference_ops_from_ref (ref: callarg, result); |
1326 | } |
1327 | } |
1328 | |
1329 | /* Fold *& at position *I_P in a vn_reference_op_s vector *OPS. Updates |
1330 | *I_P to point to the last element of the replacement. */ |
1331 | static bool |
1332 | vn_reference_fold_indirect (vec<vn_reference_op_s> *ops, |
1333 | unsigned int *i_p) |
1334 | { |
1335 | unsigned int i = *i_p; |
1336 | vn_reference_op_t op = &(*ops)[i]; |
1337 | vn_reference_op_t mem_op = &(*ops)[i - 1]; |
1338 | tree addr_base; |
1339 | poly_int64 addr_offset = 0; |
1340 | |
1341 | /* The only thing we have to do is from &OBJ.foo.bar add the offset |
1342 | from .foo.bar to the preceding MEM_REF offset and replace the |
1343 | address with &OBJ. */ |
1344 | addr_base = get_addr_base_and_unit_offset_1 (TREE_OPERAND (op->op0, 0), |
1345 | &addr_offset, vn_valueize); |
1346 | gcc_checking_assert (addr_base && TREE_CODE (addr_base) != MEM_REF); |
1347 | if (addr_base != TREE_OPERAND (op->op0, 0)) |
1348 | { |
1349 | poly_offset_int off |
1350 | = (poly_offset_int::from (a: wi::to_poly_wide (t: mem_op->op0), |
1351 | sgn: SIGNED) |
1352 | + addr_offset); |
1353 | mem_op->op0 = wide_int_to_tree (TREE_TYPE (mem_op->op0), cst: off); |
1354 | op->op0 = build_fold_addr_expr (addr_base); |
1355 | if (tree_fits_shwi_p (mem_op->op0)) |
1356 | mem_op->off = tree_to_shwi (mem_op->op0); |
1357 | else |
1358 | mem_op->off = -1; |
1359 | return true; |
1360 | } |
1361 | return false; |
1362 | } |
1363 | |
1364 | /* Fold *& at position *I_P in a vn_reference_op_s vector *OPS. Updates |
1365 | *I_P to point to the last element of the replacement. */ |
1366 | static bool |
1367 | vn_reference_maybe_forwprop_address (vec<vn_reference_op_s> *ops, |
1368 | unsigned int *i_p) |
1369 | { |
1370 | bool changed = false; |
1371 | vn_reference_op_t op; |
1372 | |
1373 | do |
1374 | { |
1375 | unsigned int i = *i_p; |
1376 | op = &(*ops)[i]; |
1377 | vn_reference_op_t mem_op = &(*ops)[i - 1]; |
1378 | gimple *def_stmt; |
1379 | enum tree_code code; |
1380 | poly_offset_int off; |
1381 | |
1382 | def_stmt = SSA_NAME_DEF_STMT (op->op0); |
1383 | if (!is_gimple_assign (gs: def_stmt)) |
1384 | return changed; |
1385 | |
1386 | code = gimple_assign_rhs_code (gs: def_stmt); |
1387 | if (code != ADDR_EXPR |
1388 | && code != POINTER_PLUS_EXPR) |
1389 | return changed; |
1390 | |
1391 | off = poly_offset_int::from (a: wi::to_poly_wide (t: mem_op->op0), sgn: SIGNED); |
1392 | |
1393 | /* The only thing we have to do is from &OBJ.foo.bar add the offset |
1394 | from .foo.bar to the preceding MEM_REF offset and replace the |
1395 | address with &OBJ. */ |
1396 | if (code == ADDR_EXPR) |
1397 | { |
1398 | tree addr, addr_base; |
1399 | poly_int64 addr_offset; |
1400 | |
1401 | addr = gimple_assign_rhs1 (gs: def_stmt); |
1402 | addr_base = get_addr_base_and_unit_offset_1 (TREE_OPERAND (addr, 0), |
1403 | &addr_offset, |
1404 | vn_valueize); |
1405 | /* If that didn't work because the address isn't invariant propagate |
1406 | the reference tree from the address operation in case the current |
1407 | dereference isn't offsetted. */ |
1408 | if (!addr_base |
1409 | && *i_p == ops->length () - 1 |
1410 | && known_eq (off, 0) |
1411 | /* This makes us disable this transform for PRE where the |
1412 | reference ops might be also used for code insertion which |
1413 | is invalid. */ |
1414 | && default_vn_walk_kind == VN_WALKREWRITE) |
1415 | { |
1416 | auto_vec<vn_reference_op_s, 32> tem; |
1417 | copy_reference_ops_from_ref (TREE_OPERAND (addr, 0), result: &tem); |
1418 | /* Make sure to preserve TBAA info. The only objects not |
1419 | wrapped in MEM_REFs that can have their address taken are |
1420 | STRING_CSTs. */ |
1421 | if (tem.length () >= 2 |
1422 | && tem[tem.length () - 2].opcode == MEM_REF) |
1423 | { |
1424 | vn_reference_op_t new_mem_op = &tem[tem.length () - 2]; |
1425 | new_mem_op->op0 |
1426 | = wide_int_to_tree (TREE_TYPE (mem_op->op0), |
1427 | cst: wi::to_poly_wide (t: new_mem_op->op0)); |
1428 | } |
1429 | else |
1430 | gcc_assert (tem.last ().opcode == STRING_CST); |
1431 | ops->pop (); |
1432 | ops->pop (); |
1433 | ops->safe_splice (src: tem); |
1434 | --*i_p; |
1435 | return true; |
1436 | } |
1437 | if (!addr_base |
1438 | || TREE_CODE (addr_base) != MEM_REF |
1439 | || (TREE_CODE (TREE_OPERAND (addr_base, 0)) == SSA_NAME |
1440 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (TREE_OPERAND (addr_base, |
1441 | 0)))) |
1442 | return changed; |
1443 | |
1444 | off += addr_offset; |
1445 | off += mem_ref_offset (addr_base); |
1446 | op->op0 = TREE_OPERAND (addr_base, 0); |
1447 | } |
1448 | else |
1449 | { |
1450 | tree ptr, ptroff; |
1451 | ptr = gimple_assign_rhs1 (gs: def_stmt); |
1452 | ptroff = gimple_assign_rhs2 (gs: def_stmt); |
1453 | if (TREE_CODE (ptr) != SSA_NAME |
1454 | || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (ptr) |
1455 | /* Make sure to not endlessly recurse. |
1456 | See gcc.dg/tree-ssa/20040408-1.c for an example. Can easily |
1457 | happen when we value-number a PHI to its backedge value. */ |
1458 | || SSA_VAL (x: ptr) == op->op0 |
1459 | || !poly_int_tree_p (t: ptroff)) |
1460 | return changed; |
1461 | |
1462 | off += wi::to_poly_offset (t: ptroff); |
1463 | op->op0 = ptr; |
1464 | } |
1465 | |
1466 | mem_op->op0 = wide_int_to_tree (TREE_TYPE (mem_op->op0), cst: off); |
1467 | if (tree_fits_shwi_p (mem_op->op0)) |
1468 | mem_op->off = tree_to_shwi (mem_op->op0); |
1469 | else |
1470 | mem_op->off = -1; |
1471 | /* ??? Can end up with endless recursion here!? |
1472 | gcc.c-torture/execute/strcmp-1.c */ |
1473 | if (TREE_CODE (op->op0) == SSA_NAME) |
1474 | op->op0 = SSA_VAL (x: op->op0); |
1475 | if (TREE_CODE (op->op0) != SSA_NAME) |
1476 | op->opcode = TREE_CODE (op->op0); |
1477 | |
1478 | changed = true; |
1479 | } |
1480 | /* Tail-recurse. */ |
1481 | while (TREE_CODE (op->op0) == SSA_NAME); |
1482 | |
1483 | /* Fold a remaining *&. */ |
1484 | if (TREE_CODE (op->op0) == ADDR_EXPR) |
1485 | vn_reference_fold_indirect (ops, i_p); |
1486 | |
1487 | return changed; |
1488 | } |
1489 | |
1490 | /* Optimize the reference REF to a constant if possible or return |
1491 | NULL_TREE if not. */ |
1492 | |
1493 | tree |
1494 | fully_constant_vn_reference_p (vn_reference_t ref) |
1495 | { |
1496 | vec<vn_reference_op_s> operands = ref->operands; |
1497 | vn_reference_op_t op; |
1498 | |
1499 | /* Try to simplify the translated expression if it is |
1500 | a call to a builtin function with at most two arguments. */ |
1501 | op = &operands[0]; |
1502 | if (op->opcode == CALL_EXPR |
1503 | && (!op->op0 |
1504 | || (TREE_CODE (op->op0) == ADDR_EXPR |
1505 | && TREE_CODE (TREE_OPERAND (op->op0, 0)) == FUNCTION_DECL |
1506 | && fndecl_built_in_p (TREE_OPERAND (op->op0, 0), |
1507 | klass: BUILT_IN_NORMAL))) |
1508 | && operands.length () >= 2 |
1509 | && operands.length () <= 3) |
1510 | { |
1511 | vn_reference_op_t arg0, arg1 = NULL; |
1512 | bool anyconst = false; |
1513 | arg0 = &operands[1]; |
1514 | if (operands.length () > 2) |
1515 | arg1 = &operands[2]; |
1516 | if (TREE_CODE_CLASS (arg0->opcode) == tcc_constant |
1517 | || (arg0->opcode == ADDR_EXPR |
1518 | && is_gimple_min_invariant (arg0->op0))) |
1519 | anyconst = true; |
1520 | if (arg1 |
1521 | && (TREE_CODE_CLASS (arg1->opcode) == tcc_constant |
1522 | || (arg1->opcode == ADDR_EXPR |
1523 | && is_gimple_min_invariant (arg1->op0)))) |
1524 | anyconst = true; |
1525 | if (anyconst) |
1526 | { |
1527 | combined_fn fn; |
1528 | if (op->op0) |
1529 | fn = as_combined_fn (fn: DECL_FUNCTION_CODE |
1530 | (TREE_OPERAND (op->op0, 0))); |
1531 | else |
1532 | fn = as_combined_fn (fn: (internal_fn) op->clique); |
1533 | tree folded; |
1534 | if (arg1) |
1535 | folded = fold_const_call (fn, ref->type, arg0->op0, arg1->op0); |
1536 | else |
1537 | folded = fold_const_call (fn, ref->type, arg0->op0); |
1538 | if (folded |
1539 | && is_gimple_min_invariant (folded)) |
1540 | return folded; |
1541 | } |
1542 | } |
1543 | |
1544 | /* Simplify reads from constants or constant initializers. */ |
1545 | else if (BITS_PER_UNIT == 8 |
1546 | && ref->type |
1547 | && COMPLETE_TYPE_P (ref->type) |
1548 | && is_gimple_reg_type (type: ref->type)) |
1549 | { |
1550 | poly_int64 off = 0; |
1551 | HOST_WIDE_INT size; |
1552 | if (INTEGRAL_TYPE_P (ref->type)) |
1553 | size = TYPE_PRECISION (ref->type); |
1554 | else if (tree_fits_shwi_p (TYPE_SIZE (ref->type))) |
1555 | size = tree_to_shwi (TYPE_SIZE (ref->type)); |
1556 | else |
1557 | return NULL_TREE; |
1558 | if (size % BITS_PER_UNIT != 0 |
1559 | || size > MAX_BITSIZE_MODE_ANY_MODE) |
1560 | return NULL_TREE; |
1561 | size /= BITS_PER_UNIT; |
1562 | unsigned i; |
1563 | for (i = 0; i < operands.length (); ++i) |
1564 | { |
1565 | if (TREE_CODE_CLASS (operands[i].opcode) == tcc_constant) |
1566 | { |
1567 | ++i; |
1568 | break; |
1569 | } |
1570 | if (known_eq (operands[i].off, -1)) |
1571 | return NULL_TREE; |
1572 | off += operands[i].off; |
1573 | if (operands[i].opcode == MEM_REF) |
1574 | { |
1575 | ++i; |
1576 | break; |
1577 | } |
1578 | } |
1579 | vn_reference_op_t base = &operands[--i]; |
1580 | tree ctor = error_mark_node; |
1581 | tree decl = NULL_TREE; |
1582 | if (TREE_CODE_CLASS (base->opcode) == tcc_constant) |
1583 | ctor = base->op0; |
1584 | else if (base->opcode == MEM_REF |
1585 | && base[1].opcode == ADDR_EXPR |
1586 | && (VAR_P (TREE_OPERAND (base[1].op0, 0)) |
1587 | || TREE_CODE (TREE_OPERAND (base[1].op0, 0)) == CONST_DECL |
1588 | || TREE_CODE (TREE_OPERAND (base[1].op0, 0)) == STRING_CST)) |
1589 | { |
1590 | decl = TREE_OPERAND (base[1].op0, 0); |
1591 | if (TREE_CODE (decl) == STRING_CST) |
1592 | ctor = decl; |
1593 | else |
1594 | ctor = ctor_for_folding (decl); |
1595 | } |
1596 | if (ctor == NULL_TREE) |
1597 | return build_zero_cst (ref->type); |
1598 | else if (ctor != error_mark_node) |
1599 | { |
1600 | HOST_WIDE_INT const_off; |
1601 | if (decl) |
1602 | { |
1603 | tree res = fold_ctor_reference (ref->type, ctor, |
1604 | off * BITS_PER_UNIT, |
1605 | size * BITS_PER_UNIT, decl); |
1606 | if (res) |
1607 | { |
1608 | STRIP_USELESS_TYPE_CONVERSION (res); |
1609 | if (is_gimple_min_invariant (res)) |
1610 | return res; |
1611 | } |
1612 | } |
1613 | else if (off.is_constant (const_value: &const_off)) |
1614 | { |
1615 | unsigned char buf[MAX_BITSIZE_MODE_ANY_MODE / BITS_PER_UNIT]; |
1616 | int len = native_encode_expr (ctor, buf, size, off: const_off); |
1617 | if (len > 0) |
1618 | return native_interpret_expr (ref->type, buf, len); |
1619 | } |
1620 | } |
1621 | } |
1622 | |
1623 | return NULL_TREE; |
1624 | } |
1625 | |
1626 | /* Return true if OPS contain a storage order barrier. */ |
1627 | |
1628 | static bool |
1629 | contains_storage_order_barrier_p (vec<vn_reference_op_s> ops) |
1630 | { |
1631 | vn_reference_op_t op; |
1632 | unsigned i; |
1633 | |
1634 | FOR_EACH_VEC_ELT (ops, i, op) |
1635 | if (op->opcode == VIEW_CONVERT_EXPR && op->reverse) |
1636 | return true; |
1637 | |
1638 | return false; |
1639 | } |
1640 | |
1641 | /* Return true if OPS represent an access with reverse storage order. */ |
1642 | |
1643 | static bool |
1644 | reverse_storage_order_for_component_p (vec<vn_reference_op_s> ops) |
1645 | { |
1646 | unsigned i = 0; |
1647 | if (ops[i].opcode == REALPART_EXPR || ops[i].opcode == IMAGPART_EXPR) |
1648 | ++i; |
1649 | switch (ops[i].opcode) |
1650 | { |
1651 | case ARRAY_REF: |
1652 | case COMPONENT_REF: |
1653 | case BIT_FIELD_REF: |
1654 | case MEM_REF: |
1655 | return ops[i].reverse; |
1656 | default: |
1657 | return false; |
1658 | } |
1659 | } |
1660 | |
1661 | /* Transform any SSA_NAME's in a vector of vn_reference_op_s |
1662 | structures into their value numbers. This is done in-place, and |
1663 | the vector passed in is returned. *VALUEIZED_ANYTHING will specify |
1664 | whether any operands were valueized. */ |
1665 | |
1666 | static void |
1667 | valueize_refs_1 (vec<vn_reference_op_s> *orig, bool *valueized_anything, |
1668 | bool with_avail = false) |
1669 | { |
1670 | *valueized_anything = false; |
1671 | |
1672 | for (unsigned i = 0; i < orig->length (); ++i) |
1673 | { |
1674 | re_valueize: |
1675 | vn_reference_op_t vro = &(*orig)[i]; |
1676 | if (vro->opcode == SSA_NAME |
1677 | || (vro->op0 && TREE_CODE (vro->op0) == SSA_NAME)) |
1678 | { |
1679 | tree tem = with_avail ? vn_valueize (vro->op0) : SSA_VAL (x: vro->op0); |
1680 | if (tem != vro->op0) |
1681 | { |
1682 | *valueized_anything = true; |
1683 | vro->op0 = tem; |
1684 | } |
1685 | /* If it transforms from an SSA_NAME to a constant, update |
1686 | the opcode. */ |
1687 | if (TREE_CODE (vro->op0) != SSA_NAME && vro->opcode == SSA_NAME) |
1688 | vro->opcode = TREE_CODE (vro->op0); |
1689 | } |
1690 | if (vro->op1 && TREE_CODE (vro->op1) == SSA_NAME) |
1691 | { |
1692 | tree tem = with_avail ? vn_valueize (vro->op1) : SSA_VAL (x: vro->op1); |
1693 | if (tem != vro->op1) |
1694 | { |
1695 | *valueized_anything = true; |
1696 | vro->op1 = tem; |
1697 | } |
1698 | } |
1699 | if (vro->op2 && TREE_CODE (vro->op2) == SSA_NAME) |
1700 | { |
1701 | tree tem = with_avail ? vn_valueize (vro->op2) : SSA_VAL (x: vro->op2); |
1702 | if (tem != vro->op2) |
1703 | { |
1704 | *valueized_anything = true; |
1705 | vro->op2 = tem; |
1706 | } |
1707 | } |
1708 | /* If it transforms from an SSA_NAME to an address, fold with |
1709 | a preceding indirect reference. */ |
1710 | if (i > 0 |
1711 | && vro->op0 |
1712 | && TREE_CODE (vro->op0) == ADDR_EXPR |
1713 | && (*orig)[i - 1].opcode == MEM_REF) |
1714 | { |
1715 | if (vn_reference_fold_indirect (ops: orig, i_p: &i)) |
1716 | *valueized_anything = true; |
1717 | } |
1718 | else if (i > 0 |
1719 | && vro->opcode == SSA_NAME |
1720 | && (*orig)[i - 1].opcode == MEM_REF) |
1721 | { |
1722 | if (vn_reference_maybe_forwprop_address (ops: orig, i_p: &i)) |
1723 | { |
1724 | *valueized_anything = true; |
1725 | /* Re-valueize the current operand. */ |
1726 | goto re_valueize; |
1727 | } |
1728 | } |
1729 | /* If it transforms a non-constant ARRAY_REF into a constant |
1730 | one, adjust the constant offset. */ |
1731 | else if ((vro->opcode == ARRAY_REF |
1732 | || vro->opcode == ARRAY_RANGE_REF) |
1733 | && known_eq (vro->off, -1) |
1734 | && poly_int_tree_p (t: vro->op0) |
1735 | && poly_int_tree_p (t: vro->op1) |
1736 | && TREE_CODE (vro->op2) == INTEGER_CST) |
1737 | { |
1738 | poly_offset_int off = ((wi::to_poly_offset (t: vro->op0) |
1739 | - wi::to_poly_offset (t: vro->op1)) |
1740 | * wi::to_offset (t: vro->op2) |
1741 | * vn_ref_op_align_unit (op: vro)); |
1742 | off.to_shwi (r: &vro->off); |
1743 | } |
1744 | } |
1745 | } |
1746 | |
1747 | static void |
1748 | valueize_refs (vec<vn_reference_op_s> *orig) |
1749 | { |
1750 | bool tem; |
1751 | valueize_refs_1 (orig, valueized_anything: &tem); |
1752 | } |
1753 | |
1754 | static vec<vn_reference_op_s> shared_lookup_references; |
1755 | |
1756 | /* Create a vector of vn_reference_op_s structures from REF, a |
1757 | REFERENCE_CLASS_P tree. The vector is shared among all callers of |
1758 | this function. *VALUEIZED_ANYTHING will specify whether any |
1759 | operands were valueized. */ |
1760 | |
1761 | static vec<vn_reference_op_s> |
1762 | valueize_shared_reference_ops_from_ref (tree ref, bool *valueized_anything) |
1763 | { |
1764 | if (!ref) |
1765 | return vNULL; |
1766 | shared_lookup_references.truncate (size: 0); |
1767 | copy_reference_ops_from_ref (ref, result: &shared_lookup_references); |
1768 | valueize_refs_1 (orig: &shared_lookup_references, valueized_anything); |
1769 | return shared_lookup_references; |
1770 | } |
1771 | |
1772 | /* Create a vector of vn_reference_op_s structures from CALL, a |
1773 | call statement. The vector is shared among all callers of |
1774 | this function. */ |
1775 | |
1776 | static vec<vn_reference_op_s> |
1777 | valueize_shared_reference_ops_from_call (gcall *call) |
1778 | { |
1779 | if (!call) |
1780 | return vNULL; |
1781 | shared_lookup_references.truncate (size: 0); |
1782 | copy_reference_ops_from_call (call, result: &shared_lookup_references); |
1783 | valueize_refs (orig: &shared_lookup_references); |
1784 | return shared_lookup_references; |
1785 | } |
1786 | |
1787 | /* Lookup a SCCVN reference operation VR in the current hash table. |
1788 | Returns the resulting value number if it exists in the hash table, |
1789 | NULL_TREE otherwise. VNRESULT will be filled in with the actual |
1790 | vn_reference_t stored in the hashtable if something is found. */ |
1791 | |
1792 | static tree |
1793 | vn_reference_lookup_1 (vn_reference_t vr, vn_reference_t *vnresult) |
1794 | { |
1795 | vn_reference_s **slot; |
1796 | hashval_t hash; |
1797 | |
1798 | hash = vr->hashcode; |
1799 | slot = valid_info->references->find_slot_with_hash (comparable: vr, hash, insert: NO_INSERT); |
1800 | if (slot) |
1801 | { |
1802 | if (vnresult) |
1803 | *vnresult = (vn_reference_t)*slot; |
1804 | return ((vn_reference_t)*slot)->result; |
1805 | } |
1806 | |
1807 | return NULL_TREE; |
1808 | } |
1809 | |
1810 | |
1811 | /* Partial definition tracking support. */ |
1812 | |
1813 | struct pd_range |
1814 | { |
1815 | HOST_WIDE_INT offset; |
1816 | HOST_WIDE_INT size; |
1817 | }; |
1818 | |
1819 | struct pd_data |
1820 | { |
1821 | tree rhs; |
1822 | HOST_WIDE_INT rhs_off; |
1823 | HOST_WIDE_INT offset; |
1824 | HOST_WIDE_INT size; |
1825 | }; |
1826 | |
1827 | /* Context for alias walking. */ |
1828 | |
1829 | struct vn_walk_cb_data |
1830 | { |
1831 | vn_walk_cb_data (vn_reference_t vr_, tree orig_ref_, tree *last_vuse_ptr_, |
1832 | vn_lookup_kind vn_walk_kind_, bool tbaa_p_, tree mask_, |
1833 | bool redundant_store_removal_p_) |
1834 | : vr (vr_), last_vuse_ptr (last_vuse_ptr_), last_vuse (NULL_TREE), |
1835 | mask (mask_), masked_result (NULL_TREE), same_val (NULL_TREE), |
1836 | vn_walk_kind (vn_walk_kind_), |
1837 | tbaa_p (tbaa_p_), redundant_store_removal_p (redundant_store_removal_p_), |
1838 | saved_operands (vNULL), first_set (-2), first_base_set (-2), |
1839 | known_ranges (NULL) |
1840 | { |
1841 | if (!last_vuse_ptr) |
1842 | last_vuse_ptr = &last_vuse; |
1843 | ao_ref_init (&orig_ref, orig_ref_); |
1844 | if (mask) |
1845 | { |
1846 | wide_int w = wi::to_wide (t: mask); |
1847 | unsigned int pos = 0, prec = w.get_precision (); |
1848 | pd_data pd; |
1849 | pd.rhs = build_constructor (NULL_TREE, NULL); |
1850 | pd.rhs_off = 0; |
1851 | /* When bitwise and with a constant is done on a memory load, |
1852 | we don't really need all the bits to be defined or defined |
1853 | to constants, we don't really care what is in the position |
1854 | corresponding to 0 bits in the mask. |
1855 | So, push the ranges of those 0 bits in the mask as artificial |
1856 | zero stores and let the partial def handling code do the |
1857 | rest. */ |
1858 | while (pos < prec) |
1859 | { |
1860 | int tz = wi::ctz (w); |
1861 | if (pos + tz > prec) |
1862 | tz = prec - pos; |
1863 | if (tz) |
1864 | { |
1865 | if (BYTES_BIG_ENDIAN) |
1866 | pd.offset = prec - pos - tz; |
1867 | else |
1868 | pd.offset = pos; |
1869 | pd.size = tz; |
1870 | void *r = push_partial_def (pd, 0, 0, 0, prec); |
1871 | gcc_assert (r == NULL_TREE); |
1872 | } |
1873 | pos += tz; |
1874 | if (pos == prec) |
1875 | break; |
1876 | w = wi::lrshift (x: w, y: tz); |
1877 | tz = wi::ctz (wi::bit_not (x: w)); |
1878 | if (pos + tz > prec) |
1879 | tz = prec - pos; |
1880 | pos += tz; |
1881 | w = wi::lrshift (x: w, y: tz); |
1882 | } |
1883 | } |
1884 | } |
1885 | ~vn_walk_cb_data (); |
1886 | void *finish (alias_set_type, alias_set_type, tree); |
1887 | void *push_partial_def (pd_data pd, |
1888 | alias_set_type, alias_set_type, HOST_WIDE_INT, |
1889 | HOST_WIDE_INT); |
1890 | |
1891 | vn_reference_t vr; |
1892 | ao_ref orig_ref; |
1893 | tree *last_vuse_ptr; |
1894 | tree last_vuse; |
1895 | tree mask; |
1896 | tree masked_result; |
1897 | tree same_val; |
1898 | vn_lookup_kind vn_walk_kind; |
1899 | bool tbaa_p; |
1900 | bool redundant_store_removal_p; |
1901 | vec<vn_reference_op_s> saved_operands; |
1902 | |
1903 | /* The VDEFs of partial defs we come along. */ |
1904 | auto_vec<pd_data, 2> partial_defs; |
1905 | /* The first defs range to avoid splay tree setup in most cases. */ |
1906 | pd_range first_range; |
1907 | alias_set_type first_set; |
1908 | alias_set_type first_base_set; |
1909 | splay_tree known_ranges; |
1910 | obstack ranges_obstack; |
1911 | static constexpr HOST_WIDE_INT bufsize = 64; |
1912 | }; |
1913 | |
1914 | vn_walk_cb_data::~vn_walk_cb_data () |
1915 | { |
1916 | if (known_ranges) |
1917 | { |
1918 | splay_tree_delete (known_ranges); |
1919 | obstack_free (&ranges_obstack, NULL); |
1920 | } |
1921 | saved_operands.release (); |
1922 | } |
1923 | |
1924 | void * |
1925 | vn_walk_cb_data::finish (alias_set_type set, alias_set_type base_set, tree val) |
1926 | { |
1927 | if (first_set != -2) |
1928 | { |
1929 | set = first_set; |
1930 | base_set = first_base_set; |
1931 | } |
1932 | if (mask) |
1933 | { |
1934 | masked_result = val; |
1935 | return (void *) -1; |
1936 | } |
1937 | if (same_val && !operand_equal_p (val, same_val)) |
1938 | return (void *) -1; |
1939 | vec<vn_reference_op_s> &operands |
1940 | = saved_operands.exists () ? saved_operands : vr->operands; |
1941 | return vn_reference_lookup_or_insert_for_pieces (last_vuse, set, base_set, |
1942 | vr->offset, vr->max_size, |
1943 | vr->type, operands, val); |
1944 | } |
1945 | |
1946 | /* pd_range splay-tree helpers. */ |
1947 | |
1948 | static int |
1949 | pd_range_compare (splay_tree_key offset1p, splay_tree_key offset2p) |
1950 | { |
1951 | HOST_WIDE_INT offset1 = *(HOST_WIDE_INT *)offset1p; |
1952 | HOST_WIDE_INT offset2 = *(HOST_WIDE_INT *)offset2p; |
1953 | if (offset1 < offset2) |
1954 | return -1; |
1955 | else if (offset1 > offset2) |
1956 | return 1; |
1957 | return 0; |
1958 | } |
1959 | |
1960 | static void * |
1961 | pd_tree_alloc (int size, void *data_) |
1962 | { |
1963 | vn_walk_cb_data *data = (vn_walk_cb_data *)data_; |
1964 | return obstack_alloc (&data->ranges_obstack, size); |
1965 | } |
1966 | |
1967 | static void |
1968 | pd_tree_dealloc (void *, void *) |
1969 | { |
1970 | } |
1971 | |
1972 | /* Push PD to the vector of partial definitions returning a |
1973 | value when we are ready to combine things with VUSE, SET and MAXSIZEI, |
1974 | NULL when we want to continue looking for partial defs or -1 |
1975 | on failure. */ |
1976 | |
1977 | void * |
1978 | vn_walk_cb_data::push_partial_def (pd_data pd, |
1979 | alias_set_type set, alias_set_type base_set, |
1980 | HOST_WIDE_INT offseti, |
1981 | HOST_WIDE_INT maxsizei) |
1982 | { |
1983 | /* We're using a fixed buffer for encoding so fail early if the object |
1984 | we want to interpret is bigger. */ |
1985 | if (maxsizei > bufsize * BITS_PER_UNIT |
1986 | || CHAR_BIT != 8 |
1987 | || BITS_PER_UNIT != 8 |
1988 | /* Not prepared to handle PDP endian. */ |
1989 | || BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN) |
1990 | return (void *)-1; |
1991 | |
1992 | /* Turn too large constant stores into non-constant stores. */ |
1993 | if (CONSTANT_CLASS_P (pd.rhs) && pd.size > bufsize * BITS_PER_UNIT) |
1994 | pd.rhs = error_mark_node; |
1995 | |
1996 | /* And for non-constant or CONSTRUCTOR stores shrink them to only keep at |
1997 | most a partial byte before and/or after the region. */ |
1998 | if (!CONSTANT_CLASS_P (pd.rhs)) |
1999 | { |
2000 | if (pd.offset < offseti) |
2001 | { |
2002 | HOST_WIDE_INT o = ROUND_DOWN (offseti - pd.offset, BITS_PER_UNIT); |
2003 | gcc_assert (pd.size > o); |
2004 | pd.size -= o; |
2005 | pd.offset += o; |
2006 | } |
2007 | if (pd.size > maxsizei) |
2008 | pd.size = maxsizei + ((pd.size - maxsizei) % BITS_PER_UNIT); |
2009 | } |
2010 | |
2011 | pd.offset -= offseti; |
2012 | |
2013 | bool pd_constant_p = (TREE_CODE (pd.rhs) == CONSTRUCTOR |
2014 | || CONSTANT_CLASS_P (pd.rhs)); |
2015 | pd_range *r; |
2016 | if (partial_defs.is_empty ()) |
2017 | { |
2018 | /* If we get a clobber upfront, fail. */ |
2019 | if (TREE_CLOBBER_P (pd.rhs)) |
2020 | return (void *)-1; |
2021 | if (!pd_constant_p) |
2022 | return (void *)-1; |
2023 | partial_defs.safe_push (obj: pd); |
2024 | first_range.offset = pd.offset; |
2025 | first_range.size = pd.size; |
2026 | first_set = set; |
2027 | first_base_set = base_set; |
2028 | last_vuse_ptr = NULL; |
2029 | r = &first_range; |
2030 | /* Go check if the first partial definition was a full one in case |
2031 | the caller didn't optimize for this. */ |
2032 | } |
2033 | else |
2034 | { |
2035 | if (!known_ranges) |
2036 | { |
2037 | /* ??? Optimize the case where the 2nd partial def completes |
2038 | things. */ |
2039 | gcc_obstack_init (&ranges_obstack); |
2040 | known_ranges = splay_tree_new_with_allocator (pd_range_compare, 0, 0, |
2041 | pd_tree_alloc, |
2042 | pd_tree_dealloc, this); |
2043 | splay_tree_insert (known_ranges, |
2044 | (splay_tree_key)&first_range.offset, |
2045 | (splay_tree_value)&first_range); |
2046 | } |
2047 | |
2048 | pd_range newr = { .offset: pd.offset, .size: pd.size }; |
2049 | splay_tree_node n; |
2050 | /* Lookup the predecessor of offset + 1 and see if we need to merge. */ |
2051 | HOST_WIDE_INT loffset = newr.offset + 1; |
2052 | if ((n = splay_tree_predecessor (known_ranges, (splay_tree_key)&loffset)) |
2053 | && ((r = (pd_range *)n->value), true) |
2054 | && ranges_known_overlap_p (pos1: r->offset, size1: r->size + 1, |
2055 | pos2: newr.offset, size2: newr.size)) |
2056 | { |
2057 | /* Ignore partial defs already covered. Here we also drop shadowed |
2058 | clobbers arriving here at the floor. */ |
2059 | if (known_subrange_p (pos1: newr.offset, size1: newr.size, pos2: r->offset, size2: r->size)) |
2060 | return NULL; |
2061 | r->size |
2062 | = MAX (r->offset + r->size, newr.offset + newr.size) - r->offset; |
2063 | } |
2064 | else |
2065 | { |
2066 | /* newr.offset wasn't covered yet, insert the range. */ |
2067 | r = XOBNEW (&ranges_obstack, pd_range); |
2068 | *r = newr; |
2069 | splay_tree_insert (known_ranges, (splay_tree_key)&r->offset, |
2070 | (splay_tree_value)r); |
2071 | } |
2072 | /* Merge r which now contains newr and is a member of the splay tree with |
2073 | adjacent overlapping ranges. */ |
2074 | pd_range *rafter; |
2075 | while ((n = splay_tree_successor (known_ranges, |
2076 | (splay_tree_key)&r->offset)) |
2077 | && ((rafter = (pd_range *)n->value), true) |
2078 | && ranges_known_overlap_p (pos1: r->offset, size1: r->size + 1, |
2079 | pos2: rafter->offset, size2: rafter->size)) |
2080 | { |
2081 | r->size = MAX (r->offset + r->size, |
2082 | rafter->offset + rafter->size) - r->offset; |
2083 | splay_tree_remove (known_ranges, (splay_tree_key)&rafter->offset); |
2084 | } |
2085 | /* If we get a clobber, fail. */ |
2086 | if (TREE_CLOBBER_P (pd.rhs)) |
2087 | return (void *)-1; |
2088 | /* Non-constants are OK as long as they are shadowed by a constant. */ |
2089 | if (!pd_constant_p) |
2090 | return (void *)-1; |
2091 | partial_defs.safe_push (obj: pd); |
2092 | } |
2093 | |
2094 | /* Now we have merged newr into the range tree. When we have covered |
2095 | [offseti, sizei] then the tree will contain exactly one node which has |
2096 | the desired properties and it will be 'r'. */ |
2097 | if (!known_subrange_p (pos1: 0, size1: maxsizei, pos2: r->offset, size2: r->size)) |
2098 | /* Continue looking for partial defs. */ |
2099 | return NULL; |
2100 | |
2101 | /* Now simply native encode all partial defs in reverse order. */ |
2102 | unsigned ndefs = partial_defs.length (); |
2103 | /* We support up to 512-bit values (for V8DFmode). */ |
2104 | unsigned char buffer[bufsize + 1]; |
2105 | unsigned char this_buffer[bufsize + 1]; |
2106 | int len; |
2107 | |
2108 | memset (s: buffer, c: 0, n: bufsize + 1); |
2109 | unsigned needed_len = ROUND_UP (maxsizei, BITS_PER_UNIT) / BITS_PER_UNIT; |
2110 | while (!partial_defs.is_empty ()) |
2111 | { |
2112 | pd_data pd = partial_defs.pop (); |
2113 | unsigned int amnt; |
2114 | if (TREE_CODE (pd.rhs) == CONSTRUCTOR) |
2115 | { |
2116 | /* Empty CONSTRUCTOR. */ |
2117 | if (pd.size >= needed_len * BITS_PER_UNIT) |
2118 | len = needed_len; |
2119 | else |
2120 | len = ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT; |
2121 | memset (s: this_buffer, c: 0, n: len); |
2122 | } |
2123 | else if (pd.rhs_off >= 0) |
2124 | { |
2125 | len = native_encode_expr (pd.rhs, this_buffer, bufsize, |
2126 | off: (MAX (0, -pd.offset) |
2127 | + pd.rhs_off) / BITS_PER_UNIT); |
2128 | if (len <= 0 |
2129 | || len < (ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT |
2130 | - MAX (0, -pd.offset) / BITS_PER_UNIT)) |
2131 | { |
2132 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2133 | fprintf (stream: dump_file, format: "Failed to encode %u " |
2134 | "partial definitions\n" , ndefs); |
2135 | return (void *)-1; |
2136 | } |
2137 | } |
2138 | else /* negative pd.rhs_off indicates we want to chop off first bits */ |
2139 | { |
2140 | if (-pd.rhs_off >= bufsize) |
2141 | return (void *)-1; |
2142 | len = native_encode_expr (pd.rhs, |
2143 | this_buffer + -pd.rhs_off / BITS_PER_UNIT, |
2144 | bufsize - -pd.rhs_off / BITS_PER_UNIT, |
2145 | MAX (0, -pd.offset) / BITS_PER_UNIT); |
2146 | if (len <= 0 |
2147 | || len < (ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT |
2148 | - MAX (0, -pd.offset) / BITS_PER_UNIT)) |
2149 | { |
2150 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2151 | fprintf (stream: dump_file, format: "Failed to encode %u " |
2152 | "partial definitions\n" , ndefs); |
2153 | return (void *)-1; |
2154 | } |
2155 | } |
2156 | |
2157 | unsigned char *p = buffer; |
2158 | HOST_WIDE_INT size = pd.size; |
2159 | if (pd.offset < 0) |
2160 | size -= ROUND_DOWN (-pd.offset, BITS_PER_UNIT); |
2161 | this_buffer[len] = 0; |
2162 | if (BYTES_BIG_ENDIAN) |
2163 | { |
2164 | /* LSB of this_buffer[len - 1] byte should be at |
2165 | pd.offset + pd.size - 1 bits in buffer. */ |
2166 | amnt = ((unsigned HOST_WIDE_INT) pd.offset |
2167 | + pd.size) % BITS_PER_UNIT; |
2168 | if (amnt) |
2169 | shift_bytes_in_array_right (this_buffer, len + 1, amnt); |
2170 | unsigned char *q = this_buffer; |
2171 | unsigned int off = 0; |
2172 | if (pd.offset >= 0) |
2173 | { |
2174 | unsigned int msk; |
2175 | off = pd.offset / BITS_PER_UNIT; |
2176 | gcc_assert (off < needed_len); |
2177 | p = buffer + off; |
2178 | if (size <= amnt) |
2179 | { |
2180 | msk = ((1 << size) - 1) << (BITS_PER_UNIT - amnt); |
2181 | *p = (*p & ~msk) | (this_buffer[len] & msk); |
2182 | size = 0; |
2183 | } |
2184 | else |
2185 | { |
2186 | if (TREE_CODE (pd.rhs) != CONSTRUCTOR) |
2187 | q = (this_buffer + len |
2188 | - (ROUND_UP (size - amnt, BITS_PER_UNIT) |
2189 | / BITS_PER_UNIT)); |
2190 | if (pd.offset % BITS_PER_UNIT) |
2191 | { |
2192 | msk = -1U << (BITS_PER_UNIT |
2193 | - (pd.offset % BITS_PER_UNIT)); |
2194 | *p = (*p & msk) | (*q & ~msk); |
2195 | p++; |
2196 | q++; |
2197 | off++; |
2198 | size -= BITS_PER_UNIT - (pd.offset % BITS_PER_UNIT); |
2199 | gcc_assert (size >= 0); |
2200 | } |
2201 | } |
2202 | } |
2203 | else if (TREE_CODE (pd.rhs) != CONSTRUCTOR) |
2204 | { |
2205 | q = (this_buffer + len |
2206 | - (ROUND_UP (size - amnt, BITS_PER_UNIT) |
2207 | / BITS_PER_UNIT)); |
2208 | if (pd.offset % BITS_PER_UNIT) |
2209 | { |
2210 | q++; |
2211 | size -= BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) pd.offset |
2212 | % BITS_PER_UNIT); |
2213 | gcc_assert (size >= 0); |
2214 | } |
2215 | } |
2216 | if ((unsigned HOST_WIDE_INT) size / BITS_PER_UNIT + off |
2217 | > needed_len) |
2218 | size = (needed_len - off) * BITS_PER_UNIT; |
2219 | memcpy (dest: p, src: q, n: size / BITS_PER_UNIT); |
2220 | if (size % BITS_PER_UNIT) |
2221 | { |
2222 | unsigned int msk |
2223 | = -1U << (BITS_PER_UNIT - (size % BITS_PER_UNIT)); |
2224 | p += size / BITS_PER_UNIT; |
2225 | q += size / BITS_PER_UNIT; |
2226 | *p = (*q & msk) | (*p & ~msk); |
2227 | } |
2228 | } |
2229 | else |
2230 | { |
2231 | if (pd.offset >= 0) |
2232 | { |
2233 | /* LSB of this_buffer[0] byte should be at pd.offset bits |
2234 | in buffer. */ |
2235 | unsigned int msk; |
2236 | size = MIN (size, (HOST_WIDE_INT) needed_len * BITS_PER_UNIT); |
2237 | amnt = pd.offset % BITS_PER_UNIT; |
2238 | if (amnt) |
2239 | shift_bytes_in_array_left (this_buffer, len + 1, amnt); |
2240 | unsigned int off = pd.offset / BITS_PER_UNIT; |
2241 | gcc_assert (off < needed_len); |
2242 | size = MIN (size, |
2243 | (HOST_WIDE_INT) (needed_len - off) * BITS_PER_UNIT); |
2244 | p = buffer + off; |
2245 | if (amnt + size < BITS_PER_UNIT) |
2246 | { |
2247 | /* Low amnt bits come from *p, then size bits |
2248 | from this_buffer[0] and the remaining again from |
2249 | *p. */ |
2250 | msk = ((1 << size) - 1) << amnt; |
2251 | *p = (*p & ~msk) | (this_buffer[0] & msk); |
2252 | size = 0; |
2253 | } |
2254 | else if (amnt) |
2255 | { |
2256 | msk = -1U << amnt; |
2257 | *p = (*p & ~msk) | (this_buffer[0] & msk); |
2258 | p++; |
2259 | size -= (BITS_PER_UNIT - amnt); |
2260 | } |
2261 | } |
2262 | else |
2263 | { |
2264 | amnt = (unsigned HOST_WIDE_INT) pd.offset % BITS_PER_UNIT; |
2265 | if (amnt) |
2266 | size -= BITS_PER_UNIT - amnt; |
2267 | size = MIN (size, (HOST_WIDE_INT) needed_len * BITS_PER_UNIT); |
2268 | if (amnt) |
2269 | shift_bytes_in_array_left (this_buffer, len + 1, amnt); |
2270 | } |
2271 | memcpy (dest: p, src: this_buffer + (amnt != 0), n: size / BITS_PER_UNIT); |
2272 | p += size / BITS_PER_UNIT; |
2273 | if (size % BITS_PER_UNIT) |
2274 | { |
2275 | unsigned int msk = -1U << (size % BITS_PER_UNIT); |
2276 | *p = (this_buffer[(amnt != 0) + size / BITS_PER_UNIT] |
2277 | & ~msk) | (*p & msk); |
2278 | } |
2279 | } |
2280 | } |
2281 | |
2282 | tree type = vr->type; |
2283 | /* Make sure to interpret in a type that has a range covering the whole |
2284 | access size. */ |
2285 | if (INTEGRAL_TYPE_P (vr->type) && maxsizei != TYPE_PRECISION (vr->type)) |
2286 | type = build_nonstandard_integer_type (maxsizei, TYPE_UNSIGNED (type)); |
2287 | tree val; |
2288 | if (BYTES_BIG_ENDIAN) |
2289 | { |
2290 | unsigned sz = needed_len; |
2291 | if (maxsizei % BITS_PER_UNIT) |
2292 | shift_bytes_in_array_right (buffer, needed_len, |
2293 | BITS_PER_UNIT |
2294 | - (maxsizei % BITS_PER_UNIT)); |
2295 | if (INTEGRAL_TYPE_P (type)) |
2296 | { |
2297 | if (TYPE_MODE (type) != BLKmode) |
2298 | sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type)); |
2299 | else |
2300 | sz = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (type)); |
2301 | } |
2302 | if (sz > needed_len) |
2303 | { |
2304 | memcpy (dest: this_buffer + (sz - needed_len), src: buffer, n: needed_len); |
2305 | val = native_interpret_expr (type, this_buffer, sz); |
2306 | } |
2307 | else |
2308 | val = native_interpret_expr (type, buffer, needed_len); |
2309 | } |
2310 | else |
2311 | val = native_interpret_expr (type, buffer, bufsize); |
2312 | /* If we chop off bits because the types precision doesn't match the memory |
2313 | access size this is ok when optimizing reads but not when called from |
2314 | the DSE code during elimination. */ |
2315 | if (val && type != vr->type) |
2316 | { |
2317 | if (! int_fits_type_p (val, vr->type)) |
2318 | val = NULL_TREE; |
2319 | else |
2320 | val = fold_convert (vr->type, val); |
2321 | } |
2322 | |
2323 | if (val) |
2324 | { |
2325 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2326 | fprintf (stream: dump_file, |
2327 | format: "Successfully combined %u partial definitions\n" , ndefs); |
2328 | /* We are using the alias-set of the first store we encounter which |
2329 | should be appropriate here. */ |
2330 | return finish (set: first_set, base_set: first_base_set, val); |
2331 | } |
2332 | else |
2333 | { |
2334 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2335 | fprintf (stream: dump_file, |
2336 | format: "Failed to interpret %u encoded partial definitions\n" , ndefs); |
2337 | return (void *)-1; |
2338 | } |
2339 | } |
2340 | |
2341 | /* Callback for walk_non_aliased_vuses. Adjusts the vn_reference_t VR_ |
2342 | with the current VUSE and performs the expression lookup. */ |
2343 | |
2344 | static void * |
2345 | vn_reference_lookup_2 (ao_ref *op, tree vuse, void *data_) |
2346 | { |
2347 | vn_walk_cb_data *data = (vn_walk_cb_data *)data_; |
2348 | vn_reference_t vr = data->vr; |
2349 | vn_reference_s **slot; |
2350 | hashval_t hash; |
2351 | |
2352 | /* If we have partial definitions recorded we have to go through |
2353 | vn_reference_lookup_3. */ |
2354 | if (!data->partial_defs.is_empty ()) |
2355 | return NULL; |
2356 | |
2357 | if (data->last_vuse_ptr) |
2358 | { |
2359 | *data->last_vuse_ptr = vuse; |
2360 | data->last_vuse = vuse; |
2361 | } |
2362 | |
2363 | /* Fixup vuse and hash. */ |
2364 | if (vr->vuse) |
2365 | vr->hashcode = vr->hashcode - SSA_NAME_VERSION (vr->vuse); |
2366 | vr->vuse = vuse_ssa_val (x: vuse); |
2367 | if (vr->vuse) |
2368 | vr->hashcode = vr->hashcode + SSA_NAME_VERSION (vr->vuse); |
2369 | |
2370 | hash = vr->hashcode; |
2371 | slot = valid_info->references->find_slot_with_hash (comparable: vr, hash, insert: NO_INSERT); |
2372 | if (slot) |
2373 | { |
2374 | if ((*slot)->result && data->saved_operands.exists ()) |
2375 | return data->finish (set: vr->set, base_set: vr->base_set, val: (*slot)->result); |
2376 | return *slot; |
2377 | } |
2378 | |
2379 | if (SSA_NAME_IS_DEFAULT_DEF (vuse)) |
2380 | { |
2381 | HOST_WIDE_INT op_offset, op_size; |
2382 | tree v = NULL_TREE; |
2383 | tree base = ao_ref_base (op); |
2384 | |
2385 | if (base |
2386 | && op->offset.is_constant (const_value: &op_offset) |
2387 | && op->size.is_constant (const_value: &op_size) |
2388 | && op->max_size_known_p () |
2389 | && known_eq (op->size, op->max_size)) |
2390 | { |
2391 | if (TREE_CODE (base) == PARM_DECL) |
2392 | v = ipcp_get_aggregate_const (cfun, parm: base, by_ref: false, bit_offset: op_offset, |
2393 | bit_size: op_size); |
2394 | else if (TREE_CODE (base) == MEM_REF |
2395 | && integer_zerop (TREE_OPERAND (base, 1)) |
2396 | && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME |
2397 | && SSA_NAME_IS_DEFAULT_DEF (TREE_OPERAND (base, 0)) |
2398 | && (TREE_CODE (SSA_NAME_VAR (TREE_OPERAND (base, 0))) |
2399 | == PARM_DECL)) |
2400 | v = ipcp_get_aggregate_const (cfun, |
2401 | SSA_NAME_VAR (TREE_OPERAND (base, 0)), |
2402 | by_ref: true, bit_offset: op_offset, bit_size: op_size); |
2403 | } |
2404 | if (v) |
2405 | return data->finish (set: vr->set, base_set: vr->base_set, val: v); |
2406 | } |
2407 | |
2408 | return NULL; |
2409 | } |
2410 | |
2411 | /* Lookup an existing or insert a new vn_reference entry into the |
2412 | value table for the VUSE, SET, TYPE, OPERANDS reference which |
2413 | has the value VALUE which is either a constant or an SSA name. */ |
2414 | |
2415 | static vn_reference_t |
2416 | vn_reference_lookup_or_insert_for_pieces (tree vuse, |
2417 | alias_set_type set, |
2418 | alias_set_type base_set, |
2419 | poly_int64 offset, |
2420 | poly_int64 max_size, |
2421 | tree type, |
2422 | vec<vn_reference_op_s, |
2423 | va_heap> operands, |
2424 | tree value) |
2425 | { |
2426 | vn_reference_s vr1; |
2427 | vn_reference_t result; |
2428 | unsigned value_id; |
2429 | vr1.vuse = vuse ? SSA_VAL (x: vuse) : NULL_TREE; |
2430 | vr1.operands = operands; |
2431 | vr1.type = type; |
2432 | vr1.set = set; |
2433 | vr1.base_set = base_set; |
2434 | vr1.offset = offset; |
2435 | vr1.max_size = max_size; |
2436 | vr1.hashcode = vn_reference_compute_hash (vr1: &vr1); |
2437 | if (vn_reference_lookup_1 (vr: &vr1, vnresult: &result)) |
2438 | return result; |
2439 | |
2440 | if (TREE_CODE (value) == SSA_NAME) |
2441 | value_id = VN_INFO (name: value)->value_id; |
2442 | else |
2443 | value_id = get_or_alloc_constant_value_id (constant: value); |
2444 | return vn_reference_insert_pieces (vuse, set, base_set, offset, max_size, |
2445 | type, operands.copy (), value, value_id); |
2446 | } |
2447 | |
2448 | /* Return a value-number for RCODE OPS... either by looking up an existing |
2449 | value-number for the possibly simplified result or by inserting the |
2450 | operation if INSERT is true. If SIMPLIFY is false, return a value |
2451 | number for the unsimplified expression. */ |
2452 | |
2453 | static tree |
2454 | vn_nary_build_or_lookup_1 (gimple_match_op *res_op, bool insert, |
2455 | bool simplify) |
2456 | { |
2457 | tree result = NULL_TREE; |
2458 | /* We will be creating a value number for |
2459 | RCODE (OPS...). |
2460 | So first simplify and lookup this expression to see if it |
2461 | is already available. */ |
2462 | /* For simplification valueize. */ |
2463 | unsigned i = 0; |
2464 | if (simplify) |
2465 | for (i = 0; i < res_op->num_ops; ++i) |
2466 | if (TREE_CODE (res_op->ops[i]) == SSA_NAME) |
2467 | { |
2468 | tree tem = vn_valueize (res_op->ops[i]); |
2469 | if (!tem) |
2470 | break; |
2471 | res_op->ops[i] = tem; |
2472 | } |
2473 | /* If valueization of an operand fails (it is not available), skip |
2474 | simplification. */ |
2475 | bool res = false; |
2476 | if (i == res_op->num_ops) |
2477 | { |
2478 | mprts_hook = vn_lookup_simplify_result; |
2479 | res = res_op->resimplify (NULL, vn_valueize); |
2480 | mprts_hook = NULL; |
2481 | } |
2482 | gimple *new_stmt = NULL; |
2483 | if (res |
2484 | && gimple_simplified_result_is_gimple_val (op: res_op)) |
2485 | { |
2486 | /* The expression is already available. */ |
2487 | result = res_op->ops[0]; |
2488 | /* Valueize it, simplification returns sth in AVAIL only. */ |
2489 | if (TREE_CODE (result) == SSA_NAME) |
2490 | result = SSA_VAL (x: result); |
2491 | } |
2492 | else |
2493 | { |
2494 | tree val = vn_lookup_simplify_result (res_op); |
2495 | if (!val && insert) |
2496 | { |
2497 | gimple_seq stmts = NULL; |
2498 | result = maybe_push_res_to_seq (res_op, &stmts); |
2499 | if (result) |
2500 | { |
2501 | gcc_assert (gimple_seq_singleton_p (stmts)); |
2502 | new_stmt = gimple_seq_first_stmt (s: stmts); |
2503 | } |
2504 | } |
2505 | else |
2506 | /* The expression is already available. */ |
2507 | result = val; |
2508 | } |
2509 | if (new_stmt) |
2510 | { |
2511 | /* The expression is not yet available, value-number lhs to |
2512 | the new SSA_NAME we created. */ |
2513 | /* Initialize value-number information properly. */ |
2514 | vn_ssa_aux_t result_info = VN_INFO (name: result); |
2515 | result_info->valnum = result; |
2516 | result_info->value_id = get_next_value_id (); |
2517 | result_info->visited = 1; |
2518 | gimple_seq_add_stmt_without_update (&VN_INFO (name: result)->expr, |
2519 | new_stmt); |
2520 | result_info->needs_insertion = true; |
2521 | /* ??? PRE phi-translation inserts NARYs without corresponding |
2522 | SSA name result. Re-use those but set their result according |
2523 | to the stmt we just built. */ |
2524 | vn_nary_op_t nary = NULL; |
2525 | vn_nary_op_lookup_stmt (new_stmt, &nary); |
2526 | if (nary) |
2527 | { |
2528 | gcc_assert (! nary->predicated_values && nary->u.result == NULL_TREE); |
2529 | nary->u.result = gimple_assign_lhs (gs: new_stmt); |
2530 | } |
2531 | /* As all "inserted" statements are singleton SCCs, insert |
2532 | to the valid table. This is strictly needed to |
2533 | avoid re-generating new value SSA_NAMEs for the same |
2534 | expression during SCC iteration over and over (the |
2535 | optimistic table gets cleared after each iteration). |
2536 | We do not need to insert into the optimistic table, as |
2537 | lookups there will fall back to the valid table. */ |
2538 | else |
2539 | { |
2540 | unsigned int length = vn_nary_length_from_stmt (new_stmt); |
2541 | vn_nary_op_t vno1 |
2542 | = alloc_vn_nary_op_noinit (length, &vn_tables_insert_obstack); |
2543 | vno1->value_id = result_info->value_id; |
2544 | vno1->length = length; |
2545 | vno1->predicated_values = 0; |
2546 | vno1->u.result = result; |
2547 | init_vn_nary_op_from_stmt (vno1, as_a <gassign *> (p: new_stmt)); |
2548 | vn_nary_op_insert_into (vno1, valid_info->nary); |
2549 | /* Also do not link it into the undo chain. */ |
2550 | last_inserted_nary = vno1->next; |
2551 | vno1->next = (vn_nary_op_t)(void *)-1; |
2552 | } |
2553 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2554 | { |
2555 | fprintf (stream: dump_file, format: "Inserting name " ); |
2556 | print_generic_expr (dump_file, result); |
2557 | fprintf (stream: dump_file, format: " for expression " ); |
2558 | print_gimple_expr (dump_file, new_stmt, 0, TDF_SLIM); |
2559 | fprintf (stream: dump_file, format: "\n" ); |
2560 | } |
2561 | } |
2562 | return result; |
2563 | } |
2564 | |
2565 | /* Return a value-number for RCODE OPS... either by looking up an existing |
2566 | value-number for the simplified result or by inserting the operation. */ |
2567 | |
2568 | static tree |
2569 | vn_nary_build_or_lookup (gimple_match_op *res_op) |
2570 | { |
2571 | return vn_nary_build_or_lookup_1 (res_op, insert: true, simplify: true); |
2572 | } |
2573 | |
2574 | /* Try to simplify the expression RCODE OPS... of type TYPE and return |
2575 | its value if present. */ |
2576 | |
2577 | tree |
2578 | vn_nary_simplify (vn_nary_op_t nary) |
2579 | { |
2580 | if (nary->length > gimple_match_op::MAX_NUM_OPS) |
2581 | return NULL_TREE; |
2582 | gimple_match_op op (gimple_match_cond::UNCOND, nary->opcode, |
2583 | nary->type, nary->length); |
2584 | memcpy (dest: op.ops, src: nary->op, n: sizeof (tree) * nary->length); |
2585 | return vn_nary_build_or_lookup_1 (res_op: &op, insert: false, simplify: true); |
2586 | } |
2587 | |
2588 | /* Elimination engine. */ |
2589 | |
2590 | class eliminate_dom_walker : public dom_walker |
2591 | { |
2592 | public: |
2593 | eliminate_dom_walker (cdi_direction, bitmap); |
2594 | ~eliminate_dom_walker (); |
2595 | |
2596 | edge before_dom_children (basic_block) final override; |
2597 | void after_dom_children (basic_block) final override; |
2598 | |
2599 | virtual tree eliminate_avail (basic_block, tree op); |
2600 | virtual void eliminate_push_avail (basic_block, tree op); |
2601 | tree eliminate_insert (basic_block, gimple_stmt_iterator *gsi, tree val); |
2602 | |
2603 | void eliminate_stmt (basic_block, gimple_stmt_iterator *); |
2604 | |
2605 | unsigned eliminate_cleanup (bool region_p = false); |
2606 | |
2607 | bool do_pre; |
2608 | unsigned int el_todo; |
2609 | unsigned int eliminations; |
2610 | unsigned int insertions; |
2611 | |
2612 | /* SSA names that had their defs inserted by PRE if do_pre. */ |
2613 | bitmap inserted_exprs; |
2614 | |
2615 | /* Blocks with statements that have had their EH properties changed. */ |
2616 | bitmap need_eh_cleanup; |
2617 | |
2618 | /* Blocks with statements that have had their AB properties changed. */ |
2619 | bitmap need_ab_cleanup; |
2620 | |
2621 | /* Local state for the eliminate domwalk. */ |
2622 | auto_vec<gimple *> to_remove; |
2623 | auto_vec<gimple *> to_fixup; |
2624 | auto_vec<tree> avail; |
2625 | auto_vec<tree> avail_stack; |
2626 | }; |
2627 | |
2628 | /* Adaptor to the elimination engine using RPO availability. */ |
2629 | |
2630 | class rpo_elim : public eliminate_dom_walker |
2631 | { |
2632 | public: |
2633 | rpo_elim(basic_block entry_) |
2634 | : eliminate_dom_walker (CDI_DOMINATORS, NULL), entry (entry_), |
2635 | m_avail_freelist (NULL) {} |
2636 | |
2637 | tree eliminate_avail (basic_block, tree op) final override; |
2638 | |
2639 | void eliminate_push_avail (basic_block, tree) final override; |
2640 | |
2641 | basic_block entry; |
2642 | /* Freelist of avail entries which are allocated from the vn_ssa_aux |
2643 | obstack. */ |
2644 | vn_avail *m_avail_freelist; |
2645 | }; |
2646 | |
2647 | /* Global RPO state for access from hooks. */ |
2648 | static eliminate_dom_walker *rpo_avail; |
2649 | basic_block vn_context_bb; |
2650 | |
2651 | /* Return true if BASE1 and BASE2 can be adjusted so they have the |
2652 | same address and adjust *OFFSET1 and *OFFSET2 accordingly. |
2653 | Otherwise return false. */ |
2654 | |
2655 | static bool |
2656 | adjust_offsets_for_equal_base_address (tree base1, poly_int64 *offset1, |
2657 | tree base2, poly_int64 *offset2) |
2658 | { |
2659 | poly_int64 soff; |
2660 | if (TREE_CODE (base1) == MEM_REF |
2661 | && TREE_CODE (base2) == MEM_REF) |
2662 | { |
2663 | if (mem_ref_offset (base1).to_shwi (r: &soff)) |
2664 | { |
2665 | base1 = TREE_OPERAND (base1, 0); |
2666 | *offset1 += soff * BITS_PER_UNIT; |
2667 | } |
2668 | if (mem_ref_offset (base2).to_shwi (r: &soff)) |
2669 | { |
2670 | base2 = TREE_OPERAND (base2, 0); |
2671 | *offset2 += soff * BITS_PER_UNIT; |
2672 | } |
2673 | return operand_equal_p (base1, base2, flags: 0); |
2674 | } |
2675 | return operand_equal_p (base1, base2, flags: OEP_ADDRESS_OF); |
2676 | } |
2677 | |
2678 | /* Callback for walk_non_aliased_vuses. Tries to perform a lookup |
2679 | from the statement defining VUSE and if not successful tries to |
2680 | translate *REFP and VR_ through an aggregate copy at the definition |
2681 | of VUSE. If *DISAMBIGUATE_ONLY is true then do not perform translation |
2682 | of *REF and *VR. If only disambiguation was performed then |
2683 | *DISAMBIGUATE_ONLY is set to true. */ |
2684 | |
2685 | static void * |
2686 | vn_reference_lookup_3 (ao_ref *ref, tree vuse, void *data_, |
2687 | translate_flags *disambiguate_only) |
2688 | { |
2689 | vn_walk_cb_data *data = (vn_walk_cb_data *)data_; |
2690 | vn_reference_t vr = data->vr; |
2691 | gimple *def_stmt = SSA_NAME_DEF_STMT (vuse); |
2692 | tree base = ao_ref_base (ref); |
2693 | HOST_WIDE_INT offseti = 0, maxsizei, sizei = 0; |
2694 | static vec<vn_reference_op_s> lhs_ops; |
2695 | ao_ref lhs_ref; |
2696 | bool lhs_ref_ok = false; |
2697 | poly_int64 copy_size; |
2698 | |
2699 | /* First try to disambiguate after value-replacing in the definitions LHS. */ |
2700 | if (is_gimple_assign (gs: def_stmt)) |
2701 | { |
2702 | tree lhs = gimple_assign_lhs (gs: def_stmt); |
2703 | bool valueized_anything = false; |
2704 | /* Avoid re-allocation overhead. */ |
2705 | lhs_ops.truncate (size: 0); |
2706 | basic_block saved_rpo_bb = vn_context_bb; |
2707 | vn_context_bb = gimple_bb (g: def_stmt); |
2708 | if (*disambiguate_only <= TR_VALUEIZE_AND_DISAMBIGUATE) |
2709 | { |
2710 | copy_reference_ops_from_ref (ref: lhs, result: &lhs_ops); |
2711 | valueize_refs_1 (orig: &lhs_ops, valueized_anything: &valueized_anything, with_avail: true); |
2712 | } |
2713 | vn_context_bb = saved_rpo_bb; |
2714 | ao_ref_init (&lhs_ref, lhs); |
2715 | lhs_ref_ok = true; |
2716 | if (valueized_anything |
2717 | && ao_ref_init_from_vn_reference |
2718 | (ref: &lhs_ref, set: ao_ref_alias_set (&lhs_ref), |
2719 | base_set: ao_ref_base_alias_set (&lhs_ref), TREE_TYPE (lhs), ops: lhs_ops) |
2720 | && !refs_may_alias_p_1 (ref, &lhs_ref, data->tbaa_p)) |
2721 | { |
2722 | *disambiguate_only = TR_VALUEIZE_AND_DISAMBIGUATE; |
2723 | return NULL; |
2724 | } |
2725 | |
2726 | /* When the def is a CLOBBER we can optimistically disambiguate |
2727 | against it since any overlap it would be undefined behavior. |
2728 | Avoid this for obvious must aliases to save compile-time though. |
2729 | We also may not do this when the query is used for redundant |
2730 | store removal. */ |
2731 | if (!data->redundant_store_removal_p |
2732 | && gimple_clobber_p (s: def_stmt) |
2733 | && !operand_equal_p (ao_ref_base (&lhs_ref), base, flags: OEP_ADDRESS_OF)) |
2734 | { |
2735 | *disambiguate_only = TR_DISAMBIGUATE; |
2736 | return NULL; |
2737 | } |
2738 | |
2739 | /* Besides valueizing the LHS we can also use access-path based |
2740 | disambiguation on the original non-valueized ref. */ |
2741 | if (!ref->ref |
2742 | && lhs_ref_ok |
2743 | && data->orig_ref.ref) |
2744 | { |
2745 | /* We want to use the non-valueized LHS for this, but avoid redundant |
2746 | work. */ |
2747 | ao_ref *lref = &lhs_ref; |
2748 | ao_ref lref_alt; |
2749 | if (valueized_anything) |
2750 | { |
2751 | ao_ref_init (&lref_alt, lhs); |
2752 | lref = &lref_alt; |
2753 | } |
2754 | if (!refs_may_alias_p_1 (&data->orig_ref, lref, data->tbaa_p)) |
2755 | { |
2756 | *disambiguate_only = (valueized_anything |
2757 | ? TR_VALUEIZE_AND_DISAMBIGUATE |
2758 | : TR_DISAMBIGUATE); |
2759 | return NULL; |
2760 | } |
2761 | } |
2762 | |
2763 | /* If we reach a clobbering statement try to skip it and see if |
2764 | we find a VN result with exactly the same value as the |
2765 | possible clobber. In this case we can ignore the clobber |
2766 | and return the found value. */ |
2767 | if (is_gimple_reg_type (TREE_TYPE (lhs)) |
2768 | && types_compatible_p (TREE_TYPE (lhs), type2: vr->type) |
2769 | && (ref->ref || data->orig_ref.ref) |
2770 | && !data->mask |
2771 | && data->partial_defs.is_empty () |
2772 | && multiple_p (a: get_object_alignment |
2773 | (ref->ref ? ref->ref : data->orig_ref.ref), |
2774 | b: ref->size) |
2775 | && multiple_p (a: get_object_alignment (lhs), b: ref->size)) |
2776 | { |
2777 | tree rhs = gimple_assign_rhs1 (gs: def_stmt); |
2778 | /* ??? We may not compare to ahead values which might be from |
2779 | a different loop iteration but only to loop invariants. Use |
2780 | CONSTANT_CLASS_P (unvalueized!) as conservative approximation. |
2781 | The one-hop lookup below doesn't have this issue since there's |
2782 | a virtual PHI before we ever reach a backedge to cross. |
2783 | We can skip multiple defs as long as they are from the same |
2784 | value though. */ |
2785 | if (data->same_val |
2786 | && !operand_equal_p (data->same_val, rhs)) |
2787 | ; |
2788 | else if (CONSTANT_CLASS_P (rhs)) |
2789 | { |
2790 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2791 | { |
2792 | fprintf (stream: dump_file, |
2793 | format: "Skipping possible redundant definition " ); |
2794 | print_gimple_stmt (dump_file, def_stmt, 0); |
2795 | } |
2796 | /* Delay the actual compare of the values to the end of the walk |
2797 | but do not update last_vuse from here. */ |
2798 | data->last_vuse_ptr = NULL; |
2799 | data->same_val = rhs; |
2800 | return NULL; |
2801 | } |
2802 | else |
2803 | { |
2804 | tree saved_vuse = vr->vuse; |
2805 | hashval_t saved_hashcode = vr->hashcode; |
2806 | if (vr->vuse) |
2807 | vr->hashcode = vr->hashcode - SSA_NAME_VERSION (vr->vuse); |
2808 | vr->vuse = vuse_ssa_val (x: gimple_vuse (g: def_stmt)); |
2809 | if (vr->vuse) |
2810 | vr->hashcode = vr->hashcode + SSA_NAME_VERSION (vr->vuse); |
2811 | vn_reference_t vnresult = NULL; |
2812 | /* Do not use vn_reference_lookup_2 since that might perform |
2813 | expression hashtable insertion but this lookup crosses |
2814 | a possible may-alias making such insertion conditionally |
2815 | invalid. */ |
2816 | vn_reference_lookup_1 (vr, vnresult: &vnresult); |
2817 | /* Need to restore vr->vuse and vr->hashcode. */ |
2818 | vr->vuse = saved_vuse; |
2819 | vr->hashcode = saved_hashcode; |
2820 | if (vnresult) |
2821 | { |
2822 | if (TREE_CODE (rhs) == SSA_NAME) |
2823 | rhs = SSA_VAL (x: rhs); |
2824 | if (vnresult->result |
2825 | && operand_equal_p (vnresult->result, rhs, flags: 0)) |
2826 | return vnresult; |
2827 | } |
2828 | } |
2829 | } |
2830 | } |
2831 | else if (*disambiguate_only <= TR_VALUEIZE_AND_DISAMBIGUATE |
2832 | && gimple_call_builtin_p (def_stmt, BUILT_IN_NORMAL) |
2833 | && gimple_call_num_args (gs: def_stmt) <= 4) |
2834 | { |
2835 | /* For builtin calls valueize its arguments and call the |
2836 | alias oracle again. Valueization may improve points-to |
2837 | info of pointers and constify size and position arguments. |
2838 | Originally this was motivated by PR61034 which has |
2839 | conditional calls to free falsely clobbering ref because |
2840 | of imprecise points-to info of the argument. */ |
2841 | tree oldargs[4]; |
2842 | bool valueized_anything = false; |
2843 | for (unsigned i = 0; i < gimple_call_num_args (gs: def_stmt); ++i) |
2844 | { |
2845 | oldargs[i] = gimple_call_arg (gs: def_stmt, index: i); |
2846 | tree val = vn_valueize (oldargs[i]); |
2847 | if (val != oldargs[i]) |
2848 | { |
2849 | gimple_call_set_arg (gs: def_stmt, index: i, arg: val); |
2850 | valueized_anything = true; |
2851 | } |
2852 | } |
2853 | if (valueized_anything) |
2854 | { |
2855 | bool res = call_may_clobber_ref_p_1 (as_a <gcall *> (p: def_stmt), |
2856 | ref, data->tbaa_p); |
2857 | for (unsigned i = 0; i < gimple_call_num_args (gs: def_stmt); ++i) |
2858 | gimple_call_set_arg (gs: def_stmt, index: i, arg: oldargs[i]); |
2859 | if (!res) |
2860 | { |
2861 | *disambiguate_only = TR_VALUEIZE_AND_DISAMBIGUATE; |
2862 | return NULL; |
2863 | } |
2864 | } |
2865 | } |
2866 | |
2867 | if (*disambiguate_only > TR_TRANSLATE) |
2868 | return (void *)-1; |
2869 | |
2870 | /* If we cannot constrain the size of the reference we cannot |
2871 | test if anything kills it. */ |
2872 | if (!ref->max_size_known_p ()) |
2873 | return (void *)-1; |
2874 | |
2875 | poly_int64 offset = ref->offset; |
2876 | poly_int64 maxsize = ref->max_size; |
2877 | |
2878 | /* def_stmt may-defs *ref. See if we can derive a value for *ref |
2879 | from that definition. |
2880 | 1) Memset. */ |
2881 | if (is_gimple_reg_type (type: vr->type) |
2882 | && (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET) |
2883 | || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET_CHK)) |
2884 | && (integer_zerop (gimple_call_arg (gs: def_stmt, index: 1)) |
2885 | || ((TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST |
2886 | || (INTEGRAL_TYPE_P (vr->type) && known_eq (ref->size, 8))) |
2887 | && CHAR_BIT == 8 |
2888 | && BITS_PER_UNIT == 8 |
2889 | && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN |
2890 | && offset.is_constant (const_value: &offseti) |
2891 | && ref->size.is_constant (const_value: &sizei) |
2892 | && (offseti % BITS_PER_UNIT == 0 |
2893 | || TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST))) |
2894 | && (poly_int_tree_p (t: gimple_call_arg (gs: def_stmt, index: 2)) |
2895 | || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME |
2896 | && poly_int_tree_p (t: SSA_VAL (x: gimple_call_arg (gs: def_stmt, index: 2))))) |
2897 | && (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR |
2898 | || TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME)) |
2899 | { |
2900 | tree base2; |
2901 | poly_int64 offset2, size2, maxsize2; |
2902 | bool reverse; |
2903 | tree ref2 = gimple_call_arg (gs: def_stmt, index: 0); |
2904 | if (TREE_CODE (ref2) == SSA_NAME) |
2905 | { |
2906 | ref2 = SSA_VAL (x: ref2); |
2907 | if (TREE_CODE (ref2) == SSA_NAME |
2908 | && (TREE_CODE (base) != MEM_REF |
2909 | || TREE_OPERAND (base, 0) != ref2)) |
2910 | { |
2911 | gimple *def_stmt = SSA_NAME_DEF_STMT (ref2); |
2912 | if (gimple_assign_single_p (gs: def_stmt) |
2913 | && gimple_assign_rhs_code (gs: def_stmt) == ADDR_EXPR) |
2914 | ref2 = gimple_assign_rhs1 (gs: def_stmt); |
2915 | } |
2916 | } |
2917 | if (TREE_CODE (ref2) == ADDR_EXPR) |
2918 | { |
2919 | ref2 = TREE_OPERAND (ref2, 0); |
2920 | base2 = get_ref_base_and_extent (ref2, &offset2, &size2, &maxsize2, |
2921 | &reverse); |
2922 | if (!known_size_p (a: maxsize2) |
2923 | || !known_eq (maxsize2, size2) |
2924 | || !operand_equal_p (base, base2, flags: OEP_ADDRESS_OF)) |
2925 | return (void *)-1; |
2926 | } |
2927 | else if (TREE_CODE (ref2) == SSA_NAME) |
2928 | { |
2929 | poly_int64 soff; |
2930 | if (TREE_CODE (base) != MEM_REF |
2931 | || !(mem_ref_offset (base) |
2932 | << LOG2_BITS_PER_UNIT).to_shwi (r: &soff)) |
2933 | return (void *)-1; |
2934 | offset += soff; |
2935 | offset2 = 0; |
2936 | if (TREE_OPERAND (base, 0) != ref2) |
2937 | { |
2938 | gimple *def = SSA_NAME_DEF_STMT (ref2); |
2939 | if (is_gimple_assign (gs: def) |
2940 | && gimple_assign_rhs_code (gs: def) == POINTER_PLUS_EXPR |
2941 | && gimple_assign_rhs1 (gs: def) == TREE_OPERAND (base, 0) |
2942 | && poly_int_tree_p (t: gimple_assign_rhs2 (gs: def))) |
2943 | { |
2944 | tree rhs2 = gimple_assign_rhs2 (gs: def); |
2945 | if (!(poly_offset_int::from (a: wi::to_poly_wide (t: rhs2), |
2946 | sgn: SIGNED) |
2947 | << LOG2_BITS_PER_UNIT).to_shwi (r: &offset2)) |
2948 | return (void *)-1; |
2949 | ref2 = gimple_assign_rhs1 (gs: def); |
2950 | if (TREE_CODE (ref2) == SSA_NAME) |
2951 | ref2 = SSA_VAL (x: ref2); |
2952 | } |
2953 | else |
2954 | return (void *)-1; |
2955 | } |
2956 | } |
2957 | else |
2958 | return (void *)-1; |
2959 | tree len = gimple_call_arg (gs: def_stmt, index: 2); |
2960 | HOST_WIDE_INT leni, offset2i; |
2961 | if (TREE_CODE (len) == SSA_NAME) |
2962 | len = SSA_VAL (x: len); |
2963 | /* Sometimes the above trickery is smarter than alias analysis. Take |
2964 | advantage of that. */ |
2965 | if (!ranges_maybe_overlap_p (pos1: offset, size1: maxsize, pos2: offset2, |
2966 | size2: (wi::to_poly_offset (t: len) |
2967 | << LOG2_BITS_PER_UNIT))) |
2968 | return NULL; |
2969 | if (data->partial_defs.is_empty () |
2970 | && known_subrange_p (pos1: offset, size1: maxsize, pos2: offset2, |
2971 | size2: wi::to_poly_offset (t: len) << LOG2_BITS_PER_UNIT)) |
2972 | { |
2973 | tree val; |
2974 | if (integer_zerop (gimple_call_arg (gs: def_stmt, index: 1))) |
2975 | val = build_zero_cst (vr->type); |
2976 | else if (INTEGRAL_TYPE_P (vr->type) |
2977 | && known_eq (ref->size, 8) |
2978 | && offseti % BITS_PER_UNIT == 0) |
2979 | { |
2980 | gimple_match_op res_op (gimple_match_cond::UNCOND, NOP_EXPR, |
2981 | vr->type, gimple_call_arg (gs: def_stmt, index: 1)); |
2982 | val = vn_nary_build_or_lookup (res_op: &res_op); |
2983 | if (!val |
2984 | || (TREE_CODE (val) == SSA_NAME |
2985 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val))) |
2986 | return (void *)-1; |
2987 | } |
2988 | else |
2989 | { |
2990 | unsigned buflen |
2991 | = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (vr->type)) + 1; |
2992 | if (INTEGRAL_TYPE_P (vr->type) |
2993 | && TYPE_MODE (vr->type) != BLKmode) |
2994 | buflen = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (vr->type)) + 1; |
2995 | unsigned char *buf = XALLOCAVEC (unsigned char, buflen); |
2996 | memset (s: buf, TREE_INT_CST_LOW (gimple_call_arg (def_stmt, 1)), |
2997 | n: buflen); |
2998 | if (BYTES_BIG_ENDIAN) |
2999 | { |
3000 | unsigned int amnt |
3001 | = (((unsigned HOST_WIDE_INT) offseti + sizei) |
3002 | % BITS_PER_UNIT); |
3003 | if (amnt) |
3004 | { |
3005 | shift_bytes_in_array_right (buf, buflen, |
3006 | BITS_PER_UNIT - amnt); |
3007 | buf++; |
3008 | buflen--; |
3009 | } |
3010 | } |
3011 | else if (offseti % BITS_PER_UNIT != 0) |
3012 | { |
3013 | unsigned int amnt |
3014 | = BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) offseti |
3015 | % BITS_PER_UNIT); |
3016 | shift_bytes_in_array_left (buf, buflen, amnt); |
3017 | buf++; |
3018 | buflen--; |
3019 | } |
3020 | val = native_interpret_expr (vr->type, buf, buflen); |
3021 | if (!val) |
3022 | return (void *)-1; |
3023 | } |
3024 | return data->finish (set: 0, base_set: 0, val); |
3025 | } |
3026 | /* For now handle clearing memory with partial defs. */ |
3027 | else if (known_eq (ref->size, maxsize) |
3028 | && integer_zerop (gimple_call_arg (gs: def_stmt, index: 1)) |
3029 | && tree_fits_poly_int64_p (len) |
3030 | && tree_to_poly_int64 (len).is_constant (const_value: &leni) |
3031 | && leni <= INTTYPE_MAXIMUM (HOST_WIDE_INT) / BITS_PER_UNIT |
3032 | && offset.is_constant (const_value: &offseti) |
3033 | && offset2.is_constant (const_value: &offset2i) |
3034 | && maxsize.is_constant (const_value: &maxsizei) |
3035 | && ranges_known_overlap_p (pos1: offseti, size1: maxsizei, pos2: offset2i, |
3036 | size2: leni << LOG2_BITS_PER_UNIT)) |
3037 | { |
3038 | pd_data pd; |
3039 | pd.rhs = build_constructor (NULL_TREE, NULL); |
3040 | pd.rhs_off = 0; |
3041 | pd.offset = offset2i; |
3042 | pd.size = leni << LOG2_BITS_PER_UNIT; |
3043 | return data->push_partial_def (pd, set: 0, base_set: 0, offseti, maxsizei); |
3044 | } |
3045 | } |
3046 | |
3047 | /* 2) Assignment from an empty CONSTRUCTOR. */ |
3048 | else if (is_gimple_reg_type (type: vr->type) |
3049 | && gimple_assign_single_p (gs: def_stmt) |
3050 | && gimple_assign_rhs_code (gs: def_stmt) == CONSTRUCTOR |
3051 | && CONSTRUCTOR_NELTS (gimple_assign_rhs1 (def_stmt)) == 0) |
3052 | { |
3053 | tree base2; |
3054 | poly_int64 offset2, size2, maxsize2; |
3055 | HOST_WIDE_INT offset2i, size2i; |
3056 | gcc_assert (lhs_ref_ok); |
3057 | base2 = ao_ref_base (&lhs_ref); |
3058 | offset2 = lhs_ref.offset; |
3059 | size2 = lhs_ref.size; |
3060 | maxsize2 = lhs_ref.max_size; |
3061 | if (known_size_p (a: maxsize2) |
3062 | && known_eq (maxsize2, size2) |
3063 | && adjust_offsets_for_equal_base_address (base1: base, offset1: &offset, |
3064 | base2, offset2: &offset2)) |
3065 | { |
3066 | if (data->partial_defs.is_empty () |
3067 | && known_subrange_p (pos1: offset, size1: maxsize, pos2: offset2, size2)) |
3068 | { |
3069 | /* While technically undefined behavior do not optimize |
3070 | a full read from a clobber. */ |
3071 | if (gimple_clobber_p (s: def_stmt)) |
3072 | return (void *)-1; |
3073 | tree val = build_zero_cst (vr->type); |
3074 | return data->finish (set: ao_ref_alias_set (&lhs_ref), |
3075 | base_set: ao_ref_base_alias_set (&lhs_ref), val); |
3076 | } |
3077 | else if (known_eq (ref->size, maxsize) |
3078 | && maxsize.is_constant (const_value: &maxsizei) |
3079 | && offset.is_constant (const_value: &offseti) |
3080 | && offset2.is_constant (const_value: &offset2i) |
3081 | && size2.is_constant (const_value: &size2i) |
3082 | && ranges_known_overlap_p (pos1: offseti, size1: maxsizei, |
3083 | pos2: offset2i, size2: size2i)) |
3084 | { |
3085 | /* Let clobbers be consumed by the partial-def tracker |
3086 | which can choose to ignore them if they are shadowed |
3087 | by a later def. */ |
3088 | pd_data pd; |
3089 | pd.rhs = gimple_assign_rhs1 (gs: def_stmt); |
3090 | pd.rhs_off = 0; |
3091 | pd.offset = offset2i; |
3092 | pd.size = size2i; |
3093 | return data->push_partial_def (pd, set: ao_ref_alias_set (&lhs_ref), |
3094 | base_set: ao_ref_base_alias_set (&lhs_ref), |
3095 | offseti, maxsizei); |
3096 | } |
3097 | } |
3098 | } |
3099 | |
3100 | /* 3) Assignment from a constant. We can use folds native encode/interpret |
3101 | routines to extract the assigned bits. */ |
3102 | else if (known_eq (ref->size, maxsize) |
3103 | && is_gimple_reg_type (type: vr->type) |
3104 | && !reverse_storage_order_for_component_p (ops: vr->operands) |
3105 | && !contains_storage_order_barrier_p (ops: vr->operands) |
3106 | && gimple_assign_single_p (gs: def_stmt) |
3107 | && CHAR_BIT == 8 |
3108 | && BITS_PER_UNIT == 8 |
3109 | && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN |
3110 | /* native_encode and native_decode operate on arrays of bytes |
3111 | and so fundamentally need a compile-time size and offset. */ |
3112 | && maxsize.is_constant (const_value: &maxsizei) |
3113 | && offset.is_constant (const_value: &offseti) |
3114 | && (is_gimple_min_invariant (gimple_assign_rhs1 (gs: def_stmt)) |
3115 | || (TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME |
3116 | && is_gimple_min_invariant (SSA_VAL (x: gimple_assign_rhs1 (gs: def_stmt)))))) |
3117 | { |
3118 | tree lhs = gimple_assign_lhs (gs: def_stmt); |
3119 | tree base2; |
3120 | poly_int64 offset2, size2, maxsize2; |
3121 | HOST_WIDE_INT offset2i, size2i; |
3122 | bool reverse; |
3123 | gcc_assert (lhs_ref_ok); |
3124 | base2 = ao_ref_base (&lhs_ref); |
3125 | offset2 = lhs_ref.offset; |
3126 | size2 = lhs_ref.size; |
3127 | maxsize2 = lhs_ref.max_size; |
3128 | reverse = reverse_storage_order_for_component_p (t: lhs); |
3129 | if (base2 |
3130 | && !reverse |
3131 | && !storage_order_barrier_p (t: lhs) |
3132 | && known_eq (maxsize2, size2) |
3133 | && adjust_offsets_for_equal_base_address (base1: base, offset1: &offset, |
3134 | base2, offset2: &offset2) |
3135 | && offset.is_constant (const_value: &offseti) |
3136 | && offset2.is_constant (const_value: &offset2i) |
3137 | && size2.is_constant (const_value: &size2i)) |
3138 | { |
3139 | if (data->partial_defs.is_empty () |
3140 | && known_subrange_p (pos1: offseti, size1: maxsizei, pos2: offset2, size2)) |
3141 | { |
3142 | /* We support up to 512-bit values (for V8DFmode). */ |
3143 | unsigned char buffer[65]; |
3144 | int len; |
3145 | |
3146 | tree rhs = gimple_assign_rhs1 (gs: def_stmt); |
3147 | if (TREE_CODE (rhs) == SSA_NAME) |
3148 | rhs = SSA_VAL (x: rhs); |
3149 | len = native_encode_expr (rhs, |
3150 | buffer, sizeof (buffer) - 1, |
3151 | off: (offseti - offset2i) / BITS_PER_UNIT); |
3152 | if (len > 0 && len * BITS_PER_UNIT >= maxsizei) |
3153 | { |
3154 | tree type = vr->type; |
3155 | unsigned char *buf = buffer; |
3156 | unsigned int amnt = 0; |
3157 | /* Make sure to interpret in a type that has a range |
3158 | covering the whole access size. */ |
3159 | if (INTEGRAL_TYPE_P (vr->type) |
3160 | && maxsizei != TYPE_PRECISION (vr->type)) |
3161 | type = build_nonstandard_integer_type (maxsizei, |
3162 | TYPE_UNSIGNED (type)); |
3163 | if (BYTES_BIG_ENDIAN) |
3164 | { |
3165 | /* For big-endian native_encode_expr stored the rhs |
3166 | such that the LSB of it is the LSB of buffer[len - 1]. |
3167 | That bit is stored into memory at position |
3168 | offset2 + size2 - 1, i.e. in byte |
3169 | base + (offset2 + size2 - 1) / BITS_PER_UNIT. |
3170 | E.g. for offset2 1 and size2 14, rhs -1 and memory |
3171 | previously cleared that is: |
3172 | 0 1 |
3173 | 01111111|11111110 |
3174 | Now, if we want to extract offset 2 and size 12 from |
3175 | it using native_interpret_expr (which actually works |
3176 | for integral bitfield types in terms of byte size of |
3177 | the mode), the native_encode_expr stored the value |
3178 | into buffer as |
3179 | XX111111|11111111 |
3180 | and returned len 2 (the X bits are outside of |
3181 | precision). |
3182 | Let sz be maxsize / BITS_PER_UNIT if not extracting |
3183 | a bitfield, and GET_MODE_SIZE otherwise. |
3184 | We need to align the LSB of the value we want to |
3185 | extract as the LSB of buf[sz - 1]. |
3186 | The LSB from memory we need to read is at position |
3187 | offset + maxsize - 1. */ |
3188 | HOST_WIDE_INT sz = maxsizei / BITS_PER_UNIT; |
3189 | if (INTEGRAL_TYPE_P (type)) |
3190 | { |
3191 | if (TYPE_MODE (type) != BLKmode) |
3192 | sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type)); |
3193 | else |
3194 | sz = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (type)); |
3195 | } |
3196 | amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i |
3197 | - offseti - maxsizei) % BITS_PER_UNIT; |
3198 | if (amnt) |
3199 | shift_bytes_in_array_right (buffer, len, amnt); |
3200 | amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i |
3201 | - offseti - maxsizei - amnt) / BITS_PER_UNIT; |
3202 | if ((unsigned HOST_WIDE_INT) sz + amnt > (unsigned) len) |
3203 | len = 0; |
3204 | else |
3205 | { |
3206 | buf = buffer + len - sz - amnt; |
3207 | len -= (buf - buffer); |
3208 | } |
3209 | } |
3210 | else |
3211 | { |
3212 | amnt = ((unsigned HOST_WIDE_INT) offset2i |
3213 | - offseti) % BITS_PER_UNIT; |
3214 | if (amnt) |
3215 | { |
3216 | buffer[len] = 0; |
3217 | shift_bytes_in_array_left (buffer, len + 1, amnt); |
3218 | buf = buffer + 1; |
3219 | } |
3220 | } |
3221 | tree val = native_interpret_expr (type, buf, len); |
3222 | /* If we chop off bits because the types precision doesn't |
3223 | match the memory access size this is ok when optimizing |
3224 | reads but not when called from the DSE code during |
3225 | elimination. */ |
3226 | if (val |
3227 | && type != vr->type) |
3228 | { |
3229 | if (! int_fits_type_p (val, vr->type)) |
3230 | val = NULL_TREE; |
3231 | else |
3232 | val = fold_convert (vr->type, val); |
3233 | } |
3234 | |
3235 | if (val) |
3236 | return data->finish (set: ao_ref_alias_set (&lhs_ref), |
3237 | base_set: ao_ref_base_alias_set (&lhs_ref), val); |
3238 | } |
3239 | } |
3240 | else if (ranges_known_overlap_p (pos1: offseti, size1: maxsizei, pos2: offset2i, |
3241 | size2: size2i)) |
3242 | { |
3243 | pd_data pd; |
3244 | tree rhs = gimple_assign_rhs1 (gs: def_stmt); |
3245 | if (TREE_CODE (rhs) == SSA_NAME) |
3246 | rhs = SSA_VAL (x: rhs); |
3247 | pd.rhs = rhs; |
3248 | pd.rhs_off = 0; |
3249 | pd.offset = offset2i; |
3250 | pd.size = size2i; |
3251 | return data->push_partial_def (pd, set: ao_ref_alias_set (&lhs_ref), |
3252 | base_set: ao_ref_base_alias_set (&lhs_ref), |
3253 | offseti, maxsizei); |
3254 | } |
3255 | } |
3256 | } |
3257 | |
3258 | /* 4) Assignment from an SSA name which definition we may be able |
3259 | to access pieces from or we can combine to a larger entity. */ |
3260 | else if (known_eq (ref->size, maxsize) |
3261 | && is_gimple_reg_type (type: vr->type) |
3262 | && !reverse_storage_order_for_component_p (ops: vr->operands) |
3263 | && !contains_storage_order_barrier_p (ops: vr->operands) |
3264 | && gimple_assign_single_p (gs: def_stmt) |
3265 | && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME) |
3266 | { |
3267 | tree lhs = gimple_assign_lhs (gs: def_stmt); |
3268 | tree base2; |
3269 | poly_int64 offset2, size2, maxsize2; |
3270 | HOST_WIDE_INT offset2i, size2i, offseti; |
3271 | bool reverse; |
3272 | gcc_assert (lhs_ref_ok); |
3273 | base2 = ao_ref_base (&lhs_ref); |
3274 | offset2 = lhs_ref.offset; |
3275 | size2 = lhs_ref.size; |
3276 | maxsize2 = lhs_ref.max_size; |
3277 | reverse = reverse_storage_order_for_component_p (t: lhs); |
3278 | tree def_rhs = gimple_assign_rhs1 (gs: def_stmt); |
3279 | if (!reverse |
3280 | && !storage_order_barrier_p (t: lhs) |
3281 | && known_size_p (a: maxsize2) |
3282 | && known_eq (maxsize2, size2) |
3283 | && adjust_offsets_for_equal_base_address (base1: base, offset1: &offset, |
3284 | base2, offset2: &offset2)) |
3285 | { |
3286 | if (data->partial_defs.is_empty () |
3287 | && known_subrange_p (pos1: offset, size1: maxsize, pos2: offset2, size2) |
3288 | /* ??? We can't handle bitfield precision extracts without |
3289 | either using an alternate type for the BIT_FIELD_REF and |
3290 | then doing a conversion or possibly adjusting the offset |
3291 | according to endianness. */ |
3292 | && (! INTEGRAL_TYPE_P (vr->type) |
3293 | || known_eq (ref->size, TYPE_PRECISION (vr->type))) |
3294 | && multiple_p (a: ref->size, BITS_PER_UNIT)) |
3295 | { |
3296 | tree val = NULL_TREE; |
3297 | if (! INTEGRAL_TYPE_P (TREE_TYPE (def_rhs)) |
3298 | || type_has_mode_precision_p (TREE_TYPE (def_rhs))) |
3299 | { |
3300 | gimple_match_op op (gimple_match_cond::UNCOND, |
3301 | BIT_FIELD_REF, vr->type, |
3302 | SSA_VAL (x: def_rhs), |
3303 | bitsize_int (ref->size), |
3304 | bitsize_int (offset - offset2)); |
3305 | val = vn_nary_build_or_lookup (res_op: &op); |
3306 | } |
3307 | else if (known_eq (ref->size, size2)) |
3308 | { |
3309 | gimple_match_op op (gimple_match_cond::UNCOND, |
3310 | VIEW_CONVERT_EXPR, vr->type, |
3311 | SSA_VAL (x: def_rhs)); |
3312 | val = vn_nary_build_or_lookup (res_op: &op); |
3313 | } |
3314 | if (val |
3315 | && (TREE_CODE (val) != SSA_NAME |
3316 | || ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val))) |
3317 | return data->finish (set: ao_ref_alias_set (&lhs_ref), |
3318 | base_set: ao_ref_base_alias_set (&lhs_ref), val); |
3319 | } |
3320 | else if (maxsize.is_constant (const_value: &maxsizei) |
3321 | && offset.is_constant (const_value: &offseti) |
3322 | && offset2.is_constant (const_value: &offset2i) |
3323 | && size2.is_constant (const_value: &size2i) |
3324 | && ranges_known_overlap_p (pos1: offset, size1: maxsize, pos2: offset2, size2)) |
3325 | { |
3326 | pd_data pd; |
3327 | pd.rhs = SSA_VAL (x: def_rhs); |
3328 | pd.rhs_off = 0; |
3329 | pd.offset = offset2i; |
3330 | pd.size = size2i; |
3331 | return data->push_partial_def (pd, set: ao_ref_alias_set (&lhs_ref), |
3332 | base_set: ao_ref_base_alias_set (&lhs_ref), |
3333 | offseti, maxsizei); |
3334 | } |
3335 | } |
3336 | } |
3337 | |
3338 | /* 4b) Assignment done via one of the vectorizer internal store |
3339 | functions where we may be able to access pieces from or we can |
3340 | combine to a larger entity. */ |
3341 | else if (known_eq (ref->size, maxsize) |
3342 | && is_gimple_reg_type (type: vr->type) |
3343 | && !reverse_storage_order_for_component_p (ops: vr->operands) |
3344 | && !contains_storage_order_barrier_p (ops: vr->operands) |
3345 | && is_gimple_call (gs: def_stmt) |
3346 | && gimple_call_internal_p (gs: def_stmt) |
3347 | && internal_store_fn_p (gimple_call_internal_fn (gs: def_stmt))) |
3348 | { |
3349 | gcall *call = as_a <gcall *> (p: def_stmt); |
3350 | internal_fn fn = gimple_call_internal_fn (gs: call); |
3351 | |
3352 | tree mask = NULL_TREE, len = NULL_TREE, bias = NULL_TREE; |
3353 | switch (fn) |
3354 | { |
3355 | case IFN_MASK_STORE: |
3356 | mask = gimple_call_arg (gs: call, index: internal_fn_mask_index (fn)); |
3357 | mask = vn_valueize (mask); |
3358 | if (TREE_CODE (mask) != VECTOR_CST) |
3359 | return (void *)-1; |
3360 | break; |
3361 | case IFN_LEN_STORE: |
3362 | { |
3363 | int len_index = internal_fn_len_index (fn); |
3364 | len = gimple_call_arg (gs: call, index: len_index); |
3365 | bias = gimple_call_arg (gs: call, index: len_index + 1); |
3366 | if (!tree_fits_uhwi_p (len) || !tree_fits_shwi_p (bias)) |
3367 | return (void *) -1; |
3368 | break; |
3369 | } |
3370 | default: |
3371 | return (void *)-1; |
3372 | } |
3373 | tree def_rhs = gimple_call_arg (gs: call, |
3374 | index: internal_fn_stored_value_index (fn)); |
3375 | def_rhs = vn_valueize (def_rhs); |
3376 | if (TREE_CODE (def_rhs) != VECTOR_CST) |
3377 | return (void *)-1; |
3378 | |
3379 | ao_ref_init_from_ptr_and_size (&lhs_ref, |
3380 | vn_valueize (gimple_call_arg (gs: call, index: 0)), |
3381 | TYPE_SIZE_UNIT (TREE_TYPE (def_rhs))); |
3382 | tree base2; |
3383 | poly_int64 offset2, size2, maxsize2; |
3384 | HOST_WIDE_INT offset2i, size2i, offseti; |
3385 | base2 = ao_ref_base (&lhs_ref); |
3386 | offset2 = lhs_ref.offset; |
3387 | size2 = lhs_ref.size; |
3388 | maxsize2 = lhs_ref.max_size; |
3389 | if (known_size_p (a: maxsize2) |
3390 | && known_eq (maxsize2, size2) |
3391 | && adjust_offsets_for_equal_base_address (base1: base, offset1: &offset, |
3392 | base2, offset2: &offset2) |
3393 | && maxsize.is_constant (const_value: &maxsizei) |
3394 | && offset.is_constant (const_value: &offseti) |
3395 | && offset2.is_constant (const_value: &offset2i) |
3396 | && size2.is_constant (const_value: &size2i)) |
3397 | { |
3398 | if (!ranges_maybe_overlap_p (pos1: offset, size1: maxsize, pos2: offset2, size2)) |
3399 | /* Poor-mans disambiguation. */ |
3400 | return NULL; |
3401 | else if (ranges_known_overlap_p (pos1: offset, size1: maxsize, pos2: offset2, size2)) |
3402 | { |
3403 | pd_data pd; |
3404 | pd.rhs = def_rhs; |
3405 | tree aa = gimple_call_arg (gs: call, index: 1); |
3406 | alias_set_type set = get_deref_alias_set (TREE_TYPE (aa)); |
3407 | tree vectype = TREE_TYPE (def_rhs); |
3408 | unsigned HOST_WIDE_INT elsz |
3409 | = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (vectype))); |
3410 | if (mask) |
3411 | { |
3412 | HOST_WIDE_INT start = 0, length = 0; |
3413 | unsigned mask_idx = 0; |
3414 | do |
3415 | { |
3416 | if (integer_zerop (VECTOR_CST_ELT (mask, mask_idx))) |
3417 | { |
3418 | if (length != 0) |
3419 | { |
3420 | pd.rhs_off = start; |
3421 | pd.offset = offset2i + start; |
3422 | pd.size = length; |
3423 | if (ranges_known_overlap_p |
3424 | (pos1: offset, size1: maxsize, pos2: pd.offset, size2: pd.size)) |
3425 | { |
3426 | void *res = data->push_partial_def |
3427 | (pd, set, base_set: set, offseti, maxsizei); |
3428 | if (res != NULL) |
3429 | return res; |
3430 | } |
3431 | } |
3432 | start = (mask_idx + 1) * elsz; |
3433 | length = 0; |
3434 | } |
3435 | else |
3436 | length += elsz; |
3437 | mask_idx++; |
3438 | } |
3439 | while (known_lt (mask_idx, TYPE_VECTOR_SUBPARTS (vectype))); |
3440 | if (length != 0) |
3441 | { |
3442 | pd.rhs_off = start; |
3443 | pd.offset = offset2i + start; |
3444 | pd.size = length; |
3445 | if (ranges_known_overlap_p (pos1: offset, size1: maxsize, |
3446 | pos2: pd.offset, size2: pd.size)) |
3447 | return data->push_partial_def (pd, set, base_set: set, |
3448 | offseti, maxsizei); |
3449 | } |
3450 | } |
3451 | else if (fn == IFN_LEN_STORE) |
3452 | { |
3453 | pd.offset = offset2i; |
3454 | pd.size = (tree_to_uhwi (len) |
3455 | + -tree_to_shwi (bias)) * BITS_PER_UNIT; |
3456 | if (BYTES_BIG_ENDIAN) |
3457 | pd.rhs_off = pd.size - tree_to_uhwi (TYPE_SIZE (vectype)); |
3458 | else |
3459 | pd.rhs_off = 0; |
3460 | if (ranges_known_overlap_p (pos1: offset, size1: maxsize, |
3461 | pos2: pd.offset, size2: pd.size)) |
3462 | return data->push_partial_def (pd, set, base_set: set, |
3463 | offseti, maxsizei); |
3464 | } |
3465 | else |
3466 | gcc_unreachable (); |
3467 | return NULL; |
3468 | } |
3469 | } |
3470 | } |
3471 | |
3472 | /* 5) For aggregate copies translate the reference through them if |
3473 | the copy kills ref. */ |
3474 | else if (data->vn_walk_kind == VN_WALKREWRITE |
3475 | && gimple_assign_single_p (gs: def_stmt) |
3476 | && (DECL_P (gimple_assign_rhs1 (def_stmt)) |
3477 | || TREE_CODE (gimple_assign_rhs1 (def_stmt)) == MEM_REF |
3478 | || handled_component_p (t: gimple_assign_rhs1 (gs: def_stmt)))) |
3479 | { |
3480 | tree base2; |
3481 | int i, j, k; |
3482 | auto_vec<vn_reference_op_s> rhs; |
3483 | vn_reference_op_t vro; |
3484 | ao_ref r; |
3485 | |
3486 | gcc_assert (lhs_ref_ok); |
3487 | |
3488 | /* See if the assignment kills REF. */ |
3489 | base2 = ao_ref_base (&lhs_ref); |
3490 | if (!lhs_ref.max_size_known_p () |
3491 | || (base != base2 |
3492 | && (TREE_CODE (base) != MEM_REF |
3493 | || TREE_CODE (base2) != MEM_REF |
3494 | || TREE_OPERAND (base, 0) != TREE_OPERAND (base2, 0) |
3495 | || !tree_int_cst_equal (TREE_OPERAND (base, 1), |
3496 | TREE_OPERAND (base2, 1)))) |
3497 | || !stmt_kills_ref_p (def_stmt, ref)) |
3498 | return (void *)-1; |
3499 | |
3500 | /* Find the common base of ref and the lhs. lhs_ops already |
3501 | contains valueized operands for the lhs. */ |
3502 | i = vr->operands.length () - 1; |
3503 | j = lhs_ops.length () - 1; |
3504 | while (j >= 0 && i >= 0 |
3505 | && vn_reference_op_eq (p1: &vr->operands[i], p2: &lhs_ops[j])) |
3506 | { |
3507 | i--; |
3508 | j--; |
3509 | } |
3510 | |
3511 | /* ??? The innermost op should always be a MEM_REF and we already |
3512 | checked that the assignment to the lhs kills vr. Thus for |
3513 | aggregate copies using char[] types the vn_reference_op_eq |
3514 | may fail when comparing types for compatibility. But we really |
3515 | don't care here - further lookups with the rewritten operands |
3516 | will simply fail if we messed up types too badly. */ |
3517 | poly_int64 = 0; |
3518 | if (j == 0 && i >= 0 |
3519 | && lhs_ops[0].opcode == MEM_REF |
3520 | && maybe_ne (a: lhs_ops[0].off, b: -1)) |
3521 | { |
3522 | if (known_eq (lhs_ops[0].off, vr->operands[i].off)) |
3523 | i--, j--; |
3524 | else if (vr->operands[i].opcode == MEM_REF |
3525 | && maybe_ne (a: vr->operands[i].off, b: -1)) |
3526 | { |
3527 | extra_off = vr->operands[i].off - lhs_ops[0].off; |
3528 | i--, j--; |
3529 | } |
3530 | } |
3531 | |
3532 | /* i now points to the first additional op. |
3533 | ??? LHS may not be completely contained in VR, one or more |
3534 | VIEW_CONVERT_EXPRs could be in its way. We could at least |
3535 | try handling outermost VIEW_CONVERT_EXPRs. */ |
3536 | if (j != -1) |
3537 | return (void *)-1; |
3538 | |
3539 | /* Punt if the additional ops contain a storage order barrier. */ |
3540 | for (k = i; k >= 0; k--) |
3541 | { |
3542 | vro = &vr->operands[k]; |
3543 | if (vro->opcode == VIEW_CONVERT_EXPR && vro->reverse) |
3544 | return (void *)-1; |
3545 | } |
3546 | |
3547 | /* Now re-write REF to be based on the rhs of the assignment. */ |
3548 | tree rhs1 = gimple_assign_rhs1 (gs: def_stmt); |
3549 | copy_reference_ops_from_ref (ref: rhs1, result: &rhs); |
3550 | |
3551 | /* Apply an extra offset to the inner MEM_REF of the RHS. */ |
3552 | bool force_no_tbaa = false; |
3553 | if (maybe_ne (a: extra_off, b: 0)) |
3554 | { |
3555 | if (rhs.length () < 2) |
3556 | return (void *)-1; |
3557 | int ix = rhs.length () - 2; |
3558 | if (rhs[ix].opcode != MEM_REF |
3559 | || known_eq (rhs[ix].off, -1)) |
3560 | return (void *)-1; |
3561 | rhs[ix].off += extra_off; |
3562 | rhs[ix].op0 = int_const_binop (PLUS_EXPR, rhs[ix].op0, |
3563 | build_int_cst (TREE_TYPE (rhs[ix].op0), |
3564 | extra_off)); |
3565 | /* When we have offsetted the RHS, reading only parts of it, |
3566 | we can no longer use the original TBAA type, force alias-set |
3567 | zero. */ |
3568 | force_no_tbaa = true; |
3569 | } |
3570 | |
3571 | /* Save the operands since we need to use the original ones for |
3572 | the hash entry we use. */ |
3573 | if (!data->saved_operands.exists ()) |
3574 | data->saved_operands = vr->operands.copy (); |
3575 | |
3576 | /* We need to pre-pend vr->operands[0..i] to rhs. */ |
3577 | vec<vn_reference_op_s> old = vr->operands; |
3578 | if (i + 1 + rhs.length () > vr->operands.length ()) |
3579 | vr->operands.safe_grow (len: i + 1 + rhs.length (), exact: true); |
3580 | else |
3581 | vr->operands.truncate (size: i + 1 + rhs.length ()); |
3582 | FOR_EACH_VEC_ELT (rhs, j, vro) |
3583 | vr->operands[i + 1 + j] = *vro; |
3584 | valueize_refs (orig: &vr->operands); |
3585 | if (old == shared_lookup_references) |
3586 | shared_lookup_references = vr->operands; |
3587 | vr->hashcode = vn_reference_compute_hash (vr1: vr); |
3588 | |
3589 | /* Try folding the new reference to a constant. */ |
3590 | tree val = fully_constant_vn_reference_p (ref: vr); |
3591 | if (val) |
3592 | { |
3593 | if (data->partial_defs.is_empty ()) |
3594 | return data->finish (set: ao_ref_alias_set (&lhs_ref), |
3595 | base_set: ao_ref_base_alias_set (&lhs_ref), val); |
3596 | /* This is the only interesting case for partial-def handling |
3597 | coming from targets that like to gimplify init-ctors as |
3598 | aggregate copies from constant data like aarch64 for |
3599 | PR83518. */ |
3600 | if (maxsize.is_constant (const_value: &maxsizei) && known_eq (ref->size, maxsize)) |
3601 | { |
3602 | pd_data pd; |
3603 | pd.rhs = val; |
3604 | pd.rhs_off = 0; |
3605 | pd.offset = 0; |
3606 | pd.size = maxsizei; |
3607 | return data->push_partial_def (pd, set: ao_ref_alias_set (&lhs_ref), |
3608 | base_set: ao_ref_base_alias_set (&lhs_ref), |
3609 | offseti: 0, maxsizei); |
3610 | } |
3611 | } |
3612 | |
3613 | /* Continuing with partial defs isn't easily possible here, we |
3614 | have to find a full def from further lookups from here. Probably |
3615 | not worth the special-casing everywhere. */ |
3616 | if (!data->partial_defs.is_empty ()) |
3617 | return (void *)-1; |
3618 | |
3619 | /* Adjust *ref from the new operands. */ |
3620 | ao_ref rhs1_ref; |
3621 | ao_ref_init (&rhs1_ref, rhs1); |
3622 | if (!ao_ref_init_from_vn_reference (ref: &r, |
3623 | set: force_no_tbaa ? 0 |
3624 | : ao_ref_alias_set (&rhs1_ref), |
3625 | base_set: force_no_tbaa ? 0 |
3626 | : ao_ref_base_alias_set (&rhs1_ref), |
3627 | type: vr->type, ops: vr->operands)) |
3628 | return (void *)-1; |
3629 | /* This can happen with bitfields. */ |
3630 | if (maybe_ne (a: ref->size, b: r.size)) |
3631 | { |
3632 | /* If the access lacks some subsetting simply apply that by |
3633 | shortening it. That in the end can only be successful |
3634 | if we can pun the lookup result which in turn requires |
3635 | exact offsets. */ |
3636 | if (known_eq (r.size, r.max_size) |
3637 | && known_lt (ref->size, r.size)) |
3638 | r.size = r.max_size = ref->size; |
3639 | else |
3640 | return (void *)-1; |
3641 | } |
3642 | *ref = r; |
3643 | vr->offset = r.offset; |
3644 | vr->max_size = r.max_size; |
3645 | |
3646 | /* Do not update last seen VUSE after translating. */ |
3647 | data->last_vuse_ptr = NULL; |
3648 | /* Invalidate the original access path since it now contains |
3649 | the wrong base. */ |
3650 | data->orig_ref.ref = NULL_TREE; |
3651 | /* Use the alias-set of this LHS for recording an eventual result. */ |
3652 | if (data->first_set == -2) |
3653 | { |
3654 | data->first_set = ao_ref_alias_set (&lhs_ref); |
3655 | data->first_base_set = ao_ref_base_alias_set (&lhs_ref); |
3656 | } |
3657 | |
3658 | /* Keep looking for the adjusted *REF / VR pair. */ |
3659 | return NULL; |
3660 | } |
3661 | |
3662 | /* 6) For memcpy copies translate the reference through them if the copy |
3663 | kills ref. But we cannot (easily) do this translation if the memcpy is |
3664 | a storage order barrier, i.e. is equivalent to a VIEW_CONVERT_EXPR that |
3665 | can modify the storage order of objects (see storage_order_barrier_p). */ |
3666 | else if (data->vn_walk_kind == VN_WALKREWRITE |
3667 | && is_gimple_reg_type (type: vr->type) |
3668 | /* ??? Handle BCOPY as well. */ |
3669 | && (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY) |
3670 | || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY_CHK) |
3671 | || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY) |
3672 | || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY_CHK) |
3673 | || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE) |
3674 | || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE_CHK)) |
3675 | && (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR |
3676 | || TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME) |
3677 | && (TREE_CODE (gimple_call_arg (def_stmt, 1)) == ADDR_EXPR |
3678 | || TREE_CODE (gimple_call_arg (def_stmt, 1)) == SSA_NAME) |
3679 | && (poly_int_tree_p (t: gimple_call_arg (gs: def_stmt, index: 2), value: ©_size) |
3680 | || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME |
3681 | && poly_int_tree_p (t: SSA_VAL (x: gimple_call_arg (gs: def_stmt, index: 2)), |
3682 | value: ©_size))) |
3683 | /* Handling this is more complicated, give up for now. */ |
3684 | && data->partial_defs.is_empty ()) |
3685 | { |
3686 | tree lhs, rhs; |
3687 | ao_ref r; |
3688 | poly_int64 rhs_offset, lhs_offset; |
3689 | vn_reference_op_s op; |
3690 | poly_uint64 mem_offset; |
3691 | poly_int64 at, byte_maxsize; |
3692 | |
3693 | /* Only handle non-variable, addressable refs. */ |
3694 | if (maybe_ne (a: ref->size, b: maxsize) |
3695 | || !multiple_p (a: offset, BITS_PER_UNIT, multiple: &at) |
3696 | || !multiple_p (a: maxsize, BITS_PER_UNIT, multiple: &byte_maxsize)) |
3697 | return (void *)-1; |
3698 | |
3699 | /* Extract a pointer base and an offset for the destination. */ |
3700 | lhs = gimple_call_arg (gs: def_stmt, index: 0); |
3701 | lhs_offset = 0; |
3702 | if (TREE_CODE (lhs) == SSA_NAME) |
3703 | { |
3704 | lhs = vn_valueize (lhs); |
3705 | if (TREE_CODE (lhs) == SSA_NAME) |
3706 | { |
3707 | gimple *def_stmt = SSA_NAME_DEF_STMT (lhs); |
3708 | if (gimple_assign_single_p (gs: def_stmt) |
3709 | && gimple_assign_rhs_code (gs: def_stmt) == ADDR_EXPR) |
3710 | lhs = gimple_assign_rhs1 (gs: def_stmt); |
3711 | } |
3712 | } |
3713 | if (TREE_CODE (lhs) == ADDR_EXPR) |
3714 | { |
3715 | if (AGGREGATE_TYPE_P (TREE_TYPE (TREE_TYPE (lhs))) |
3716 | && TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_TYPE (lhs)))) |
3717 | return (void *)-1; |
3718 | tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (lhs, 0), |
3719 | &lhs_offset); |
3720 | if (!tem) |
3721 | return (void *)-1; |
3722 | if (TREE_CODE (tem) == MEM_REF |
3723 | && poly_int_tree_p (TREE_OPERAND (tem, 1), value: &mem_offset)) |
3724 | { |
3725 | lhs = TREE_OPERAND (tem, 0); |
3726 | if (TREE_CODE (lhs) == SSA_NAME) |
3727 | lhs = vn_valueize (lhs); |
3728 | lhs_offset += mem_offset; |
3729 | } |
3730 | else if (DECL_P (tem)) |
3731 | lhs = build_fold_addr_expr (tem); |
3732 | else |
3733 | return (void *)-1; |
3734 | } |
3735 | if (TREE_CODE (lhs) != SSA_NAME |
3736 | && TREE_CODE (lhs) != ADDR_EXPR) |
3737 | return (void *)-1; |
3738 | |
3739 | /* Extract a pointer base and an offset for the source. */ |
3740 | rhs = gimple_call_arg (gs: def_stmt, index: 1); |
3741 | rhs_offset = 0; |
3742 | if (TREE_CODE (rhs) == SSA_NAME) |
3743 | rhs = vn_valueize (rhs); |
3744 | if (TREE_CODE (rhs) == ADDR_EXPR) |
3745 | { |
3746 | if (AGGREGATE_TYPE_P (TREE_TYPE (TREE_TYPE (rhs))) |
3747 | && TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_TYPE (rhs)))) |
3748 | return (void *)-1; |
3749 | tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs, 0), |
3750 | &rhs_offset); |
3751 | if (!tem) |
3752 | return (void *)-1; |
3753 | if (TREE_CODE (tem) == MEM_REF |
3754 | && poly_int_tree_p (TREE_OPERAND (tem, 1), value: &mem_offset)) |
3755 | { |
3756 | rhs = TREE_OPERAND (tem, 0); |
3757 | rhs_offset += mem_offset; |
3758 | } |
3759 | else if (DECL_P (tem) |
3760 | || TREE_CODE (tem) == STRING_CST) |
3761 | rhs = build_fold_addr_expr (tem); |
3762 | else |
3763 | return (void *)-1; |
3764 | } |
3765 | if (TREE_CODE (rhs) == SSA_NAME) |
3766 | rhs = SSA_VAL (x: rhs); |
3767 | else if (TREE_CODE (rhs) != ADDR_EXPR) |
3768 | return (void *)-1; |
3769 | |
3770 | /* The bases of the destination and the references have to agree. */ |
3771 | if (TREE_CODE (base) == MEM_REF) |
3772 | { |
3773 | if (TREE_OPERAND (base, 0) != lhs |
3774 | || !poly_int_tree_p (TREE_OPERAND (base, 1), value: &mem_offset)) |
3775 | return (void *) -1; |
3776 | at += mem_offset; |
3777 | } |
3778 | else if (!DECL_P (base) |
3779 | || TREE_CODE (lhs) != ADDR_EXPR |
3780 | || TREE_OPERAND (lhs, 0) != base) |
3781 | return (void *)-1; |
3782 | |
3783 | /* If the access is completely outside of the memcpy destination |
3784 | area there is no aliasing. */ |
3785 | if (!ranges_maybe_overlap_p (pos1: lhs_offset, size1: copy_size, pos2: at, size2: byte_maxsize)) |
3786 | return NULL; |
3787 | /* And the access has to be contained within the memcpy destination. */ |
3788 | if (!known_subrange_p (pos1: at, size1: byte_maxsize, pos2: lhs_offset, size2: copy_size)) |
3789 | return (void *)-1; |
3790 | |
3791 | /* Save the operands since we need to use the original ones for |
3792 | the hash entry we use. */ |
3793 | if (!data->saved_operands.exists ()) |
3794 | data->saved_operands = vr->operands.copy (); |
3795 | |
3796 | /* Make room for 2 operands in the new reference. */ |
3797 | if (vr->operands.length () < 2) |
3798 | { |
3799 | vec<vn_reference_op_s> old = vr->operands; |
3800 | vr->operands.safe_grow_cleared (len: 2, exact: true); |
3801 | if (old == shared_lookup_references) |
3802 | shared_lookup_references = vr->operands; |
3803 | } |
3804 | else |
3805 | vr->operands.truncate (size: 2); |
3806 | |
3807 | /* The looked-through reference is a simple MEM_REF. */ |
3808 | memset (s: &op, c: 0, n: sizeof (op)); |
3809 | op.type = vr->type; |
3810 | op.opcode = MEM_REF; |
3811 | op.op0 = build_int_cst (ptr_type_node, at - lhs_offset + rhs_offset); |
3812 | op.off = at - lhs_offset + rhs_offset; |
3813 | vr->operands[0] = op; |
3814 | op.type = TREE_TYPE (rhs); |
3815 | op.opcode = TREE_CODE (rhs); |
3816 | op.op0 = rhs; |
3817 | op.off = -1; |
3818 | vr->operands[1] = op; |
3819 | vr->hashcode = vn_reference_compute_hash (vr1: vr); |
3820 | |
3821 | /* Try folding the new reference to a constant. */ |
3822 | tree val = fully_constant_vn_reference_p (ref: vr); |
3823 | if (val) |
3824 | return data->finish (set: 0, base_set: 0, val); |
3825 | |
3826 | /* Adjust *ref from the new operands. */ |
3827 | if (!ao_ref_init_from_vn_reference (ref: &r, set: 0, base_set: 0, type: vr->type, ops: vr->operands)) |
3828 | return (void *)-1; |
3829 | /* This can happen with bitfields. */ |
3830 | if (maybe_ne (a: ref->size, b: r.size)) |
3831 | return (void *)-1; |
3832 | *ref = r; |
3833 | vr->offset = r.offset; |
3834 | vr->max_size = r.max_size; |
3835 | |
3836 | /* Do not update last seen VUSE after translating. */ |
3837 | data->last_vuse_ptr = NULL; |
3838 | /* Invalidate the original access path since it now contains |
3839 | the wrong base. */ |
3840 | data->orig_ref.ref = NULL_TREE; |
3841 | /* Use the alias-set of this stmt for recording an eventual result. */ |
3842 | if (data->first_set == -2) |
3843 | { |
3844 | data->first_set = 0; |
3845 | data->first_base_set = 0; |
3846 | } |
3847 | |
3848 | /* Keep looking for the adjusted *REF / VR pair. */ |
3849 | return NULL; |
3850 | } |
3851 | |
3852 | /* Bail out and stop walking. */ |
3853 | return (void *)-1; |
3854 | } |
3855 | |
3856 | /* Return a reference op vector from OP that can be used for |
3857 | vn_reference_lookup_pieces. The caller is responsible for releasing |
3858 | the vector. */ |
3859 | |
3860 | vec<vn_reference_op_s> |
3861 | vn_reference_operands_for_lookup (tree op) |
3862 | { |
3863 | bool valueized; |
3864 | return valueize_shared_reference_ops_from_ref (ref: op, valueized_anything: &valueized).copy (); |
3865 | } |
3866 | |
3867 | /* Lookup a reference operation by it's parts, in the current hash table. |
3868 | Returns the resulting value number if it exists in the hash table, |
3869 | NULL_TREE otherwise. VNRESULT will be filled in with the actual |
3870 | vn_reference_t stored in the hashtable if something is found. */ |
3871 | |
3872 | tree |
3873 | vn_reference_lookup_pieces (tree vuse, alias_set_type set, |
3874 | alias_set_type base_set, tree type, |
3875 | vec<vn_reference_op_s> operands, |
3876 | vn_reference_t *vnresult, vn_lookup_kind kind) |
3877 | { |
3878 | struct vn_reference_s vr1; |
3879 | vn_reference_t tmp; |
3880 | tree cst; |
3881 | |
3882 | if (!vnresult) |
3883 | vnresult = &tmp; |
3884 | *vnresult = NULL; |
3885 | |
3886 | vr1.vuse = vuse_ssa_val (x: vuse); |
3887 | shared_lookup_references.truncate (size: 0); |
3888 | shared_lookup_references.safe_grow (len: operands.length (), exact: true); |
3889 | memcpy (dest: shared_lookup_references.address (), |
3890 | src: operands.address (), |
3891 | n: sizeof (vn_reference_op_s) |
3892 | * operands.length ()); |
3893 | bool valueized_p; |
3894 | valueize_refs_1 (orig: &shared_lookup_references, valueized_anything: &valueized_p); |
3895 | vr1.operands = shared_lookup_references; |
3896 | vr1.type = type; |
3897 | vr1.set = set; |
3898 | vr1.base_set = base_set; |
3899 | /* We can pretend there's no extra info fed in since the ao_refs offset |
3900 | and max_size are computed only from the VN reference ops. */ |
3901 | vr1.offset = 0; |
3902 | vr1.max_size = -1; |
3903 | vr1.hashcode = vn_reference_compute_hash (vr1: &vr1); |
3904 | if ((cst = fully_constant_vn_reference_p (ref: &vr1))) |
3905 | return cst; |
3906 | |
3907 | vn_reference_lookup_1 (vr: &vr1, vnresult); |
3908 | if (!*vnresult |
3909 | && kind != VN_NOWALK |
3910 | && vr1.vuse) |
3911 | { |
3912 | ao_ref r; |
3913 | unsigned limit = param_sccvn_max_alias_queries_per_access; |
3914 | vn_walk_cb_data data (&vr1, NULL_TREE, NULL, kind, true, NULL_TREE, |
3915 | false); |
3916 | vec<vn_reference_op_s> ops_for_ref; |
3917 | if (!valueized_p) |
3918 | ops_for_ref = vr1.operands; |
3919 | else |
3920 | { |
3921 | /* For ao_ref_from_mem we have to ensure only available SSA names |
3922 | end up in base and the only convenient way to make this work |
3923 | for PRE is to re-valueize with that in mind. */ |
3924 | ops_for_ref.create (nelems: operands.length ()); |
3925 | ops_for_ref.quick_grow (len: operands.length ()); |
3926 | memcpy (dest: ops_for_ref.address (), |
3927 | src: operands.address (), |
3928 | n: sizeof (vn_reference_op_s) |
3929 | * operands.length ()); |
3930 | valueize_refs_1 (orig: &ops_for_ref, valueized_anything: &valueized_p, with_avail: true); |
3931 | } |
3932 | if (ao_ref_init_from_vn_reference (ref: &r, set, base_set, type, |
3933 | ops: ops_for_ref)) |
3934 | *vnresult |
3935 | = ((vn_reference_t) |
3936 | walk_non_aliased_vuses (&r, vr1.vuse, true, vn_reference_lookup_2, |
3937 | vn_reference_lookup_3, vuse_valueize, |
3938 | limit, &data)); |
3939 | if (ops_for_ref != shared_lookup_references) |
3940 | ops_for_ref.release (); |
3941 | gcc_checking_assert (vr1.operands == shared_lookup_references); |
3942 | if (*vnresult |
3943 | && data.same_val |
3944 | && (!(*vnresult)->result |
3945 | || !operand_equal_p ((*vnresult)->result, data.same_val))) |
3946 | { |
3947 | *vnresult = NULL; |
3948 | return NULL_TREE; |
3949 | } |
3950 | } |
3951 | |
3952 | if (*vnresult) |
3953 | return (*vnresult)->result; |
3954 | |
3955 | return NULL_TREE; |
3956 | } |
3957 | |
3958 | /* Lookup OP in the current hash table, and return the resulting value |
3959 | number if it exists in the hash table. Return NULL_TREE if it does |
3960 | not exist in the hash table or if the result field of the structure |
3961 | was NULL.. VNRESULT will be filled in with the vn_reference_t |
3962 | stored in the hashtable if one exists. When TBAA_P is false assume |
3963 | we are looking up a store and treat it as having alias-set zero. |
3964 | *LAST_VUSE_PTR will be updated with the VUSE the value lookup succeeded. |
3965 | MASK is either NULL_TREE, or can be an INTEGER_CST if the result of the |
3966 | load is bitwise anded with MASK and so we are only interested in a subset |
3967 | of the bits and can ignore if the other bits are uninitialized or |
3968 | not initialized with constants. When doing redundant store removal |
3969 | the caller has to set REDUNDANT_STORE_REMOVAL_P. */ |
3970 | |
3971 | tree |
3972 | vn_reference_lookup (tree op, tree vuse, vn_lookup_kind kind, |
3973 | vn_reference_t *vnresult, bool tbaa_p, |
3974 | tree *last_vuse_ptr, tree mask, |
3975 | bool redundant_store_removal_p) |
3976 | { |
3977 | vec<vn_reference_op_s> operands; |
3978 | struct vn_reference_s vr1; |
3979 | bool valueized_anything; |
3980 | |
3981 | if (vnresult) |
3982 | *vnresult = NULL; |
3983 | |
3984 | vr1.vuse = vuse_ssa_val (x: vuse); |
3985 | vr1.operands = operands |
3986 | = valueize_shared_reference_ops_from_ref (ref: op, valueized_anything: &valueized_anything); |
3987 | |
3988 | /* Handle &MEM[ptr + 5].b[1].c as POINTER_PLUS_EXPR. Avoid doing |
3989 | this before the pass folding __builtin_object_size had a chance to run. */ |
3990 | if ((cfun->curr_properties & PROP_objsz) |
3991 | && operands[0].opcode == ADDR_EXPR |
3992 | && operands.last ().opcode == SSA_NAME) |
3993 | { |
3994 | poly_int64 off = 0; |
3995 | vn_reference_op_t vro; |
3996 | unsigned i; |
3997 | for (i = 1; operands.iterate (ix: i, ptr: &vro); ++i) |
3998 | { |
3999 | if (vro->opcode == SSA_NAME) |
4000 | break; |
4001 | else if (known_eq (vro->off, -1)) |
4002 | break; |
4003 | off += vro->off; |
4004 | } |
4005 | if (i == operands.length () - 1 |
4006 | /* Make sure we the offset we accumulated in a 64bit int |
4007 | fits the address computation carried out in target |
4008 | offset precision. */ |
4009 | && (off.coeffs[0] |
4010 | == sext_hwi (src: off.coeffs[0], TYPE_PRECISION (sizetype)))) |
4011 | { |
4012 | gcc_assert (operands[i-1].opcode == MEM_REF); |
4013 | tree ops[2]; |
4014 | ops[0] = operands[i].op0; |
4015 | ops[1] = wide_int_to_tree (sizetype, cst: off); |
4016 | tree res = vn_nary_op_lookup_pieces (2, POINTER_PLUS_EXPR, |
4017 | TREE_TYPE (op), ops, NULL); |
4018 | if (res) |
4019 | return res; |
4020 | return NULL_TREE; |
4021 | } |
4022 | } |
4023 | |
4024 | vr1.type = TREE_TYPE (op); |
4025 | ao_ref op_ref; |
4026 | ao_ref_init (&op_ref, op); |
4027 | vr1.set = ao_ref_alias_set (&op_ref); |
4028 | vr1.base_set = ao_ref_base_alias_set (&op_ref); |
4029 | vr1.offset = 0; |
4030 | vr1.max_size = -1; |
4031 | vr1.hashcode = vn_reference_compute_hash (vr1: &vr1); |
4032 | if (mask == NULL_TREE) |
4033 | if (tree cst = fully_constant_vn_reference_p (ref: &vr1)) |
4034 | return cst; |
4035 | |
4036 | if (kind != VN_NOWALK && vr1.vuse) |
4037 | { |
4038 | vn_reference_t wvnresult; |
4039 | ao_ref r; |
4040 | unsigned limit = param_sccvn_max_alias_queries_per_access; |
4041 | auto_vec<vn_reference_op_s> ops_for_ref; |
4042 | if (valueized_anything) |
4043 | { |
4044 | copy_reference_ops_from_ref (ref: op, result: &ops_for_ref); |
4045 | bool tem; |
4046 | valueize_refs_1 (orig: &ops_for_ref, valueized_anything: &tem, with_avail: true); |
4047 | } |
4048 | /* Make sure to use a valueized reference if we valueized anything. |
4049 | Otherwise preserve the full reference for advanced TBAA. */ |
4050 | if (!valueized_anything |
4051 | || !ao_ref_init_from_vn_reference (ref: &r, set: vr1.set, base_set: vr1.base_set, |
4052 | type: vr1.type, ops: ops_for_ref)) |
4053 | { |
4054 | ao_ref_init (&r, op); |
4055 | /* Record the extra info we're getting from the full ref. */ |
4056 | ao_ref_base (&r); |
4057 | vr1.offset = r.offset; |
4058 | vr1.max_size = r.max_size; |
4059 | } |
4060 | vn_walk_cb_data data (&vr1, r.ref ? NULL_TREE : op, |
4061 | last_vuse_ptr, kind, tbaa_p, mask, |
4062 | redundant_store_removal_p); |
4063 | |
4064 | wvnresult |
4065 | = ((vn_reference_t) |
4066 | walk_non_aliased_vuses (&r, vr1.vuse, tbaa_p, vn_reference_lookup_2, |
4067 | vn_reference_lookup_3, vuse_valueize, limit, |
4068 | &data)); |
4069 | gcc_checking_assert (vr1.operands == shared_lookup_references); |
4070 | if (wvnresult) |
4071 | { |
4072 | gcc_assert (mask == NULL_TREE); |
4073 | if (data.same_val |
4074 | && (!wvnresult->result |
4075 | || !operand_equal_p (wvnresult->result, data.same_val))) |
4076 | return NULL_TREE; |
4077 | if (vnresult) |
4078 | *vnresult = wvnresult; |
4079 | return wvnresult->result; |
4080 | } |
4081 | else if (mask) |
4082 | return data.masked_result; |
4083 | |
4084 | return NULL_TREE; |
4085 | } |
4086 | |
4087 | if (last_vuse_ptr) |
4088 | *last_vuse_ptr = vr1.vuse; |
4089 | if (mask) |
4090 | return NULL_TREE; |
4091 | return vn_reference_lookup_1 (vr: &vr1, vnresult); |
4092 | } |
4093 | |
4094 | /* Lookup CALL in the current hash table and return the entry in |
4095 | *VNRESULT if found. Populates *VR for the hashtable lookup. */ |
4096 | |
4097 | void |
4098 | vn_reference_lookup_call (gcall *call, vn_reference_t *vnresult, |
4099 | vn_reference_t vr) |
4100 | { |
4101 | if (vnresult) |
4102 | *vnresult = NULL; |
4103 | |
4104 | tree vuse = gimple_vuse (g: call); |
4105 | |
4106 | vr->vuse = vuse ? SSA_VAL (x: vuse) : NULL_TREE; |
4107 | vr->operands = valueize_shared_reference_ops_from_call (call); |
4108 | tree lhs = gimple_call_lhs (gs: call); |
4109 | /* For non-SSA return values the referece ops contain the LHS. */ |
4110 | vr->type = ((lhs && TREE_CODE (lhs) == SSA_NAME) |
4111 | ? TREE_TYPE (lhs) : NULL_TREE); |
4112 | vr->punned = false; |
4113 | vr->set = 0; |
4114 | vr->base_set = 0; |
4115 | vr->offset = 0; |
4116 | vr->max_size = -1; |
4117 | vr->hashcode = vn_reference_compute_hash (vr1: vr); |
4118 | vn_reference_lookup_1 (vr, vnresult); |
4119 | } |
4120 | |
4121 | /* Insert OP into the current hash table with a value number of RESULT. */ |
4122 | |
4123 | static void |
4124 | vn_reference_insert (tree op, tree result, tree vuse, tree vdef) |
4125 | { |
4126 | vn_reference_s **slot; |
4127 | vn_reference_t vr1; |
4128 | bool tem; |
4129 | |
4130 | vec<vn_reference_op_s> operands |
4131 | = valueize_shared_reference_ops_from_ref (ref: op, valueized_anything: &tem); |
4132 | /* Handle &MEM[ptr + 5].b[1].c as POINTER_PLUS_EXPR. Avoid doing this |
4133 | before the pass folding __builtin_object_size had a chance to run. */ |
4134 | if ((cfun->curr_properties & PROP_objsz) |
4135 | && operands[0].opcode == ADDR_EXPR |
4136 | && operands.last ().opcode == SSA_NAME) |
4137 | { |
4138 | poly_int64 off = 0; |
4139 | vn_reference_op_t vro; |
4140 | unsigned i; |
4141 | for (i = 1; operands.iterate (ix: i, ptr: &vro); ++i) |
4142 | { |
4143 | if (vro->opcode == SSA_NAME) |
4144 | break; |
4145 | else if (known_eq (vro->off, -1)) |
4146 | break; |
4147 | off += vro->off; |
4148 | } |
4149 | if (i == operands.length () - 1 |
4150 | /* Make sure we the offset we accumulated in a 64bit int |
4151 | fits the address computation carried out in target |
4152 | offset precision. */ |
4153 | && (off.coeffs[0] |
4154 | == sext_hwi (src: off.coeffs[0], TYPE_PRECISION (sizetype)))) |
4155 | { |
4156 | gcc_assert (operands[i-1].opcode == MEM_REF); |
4157 | tree ops[2]; |
4158 | ops[0] = operands[i].op0; |
4159 | ops[1] = wide_int_to_tree (sizetype, cst: off); |
4160 | vn_nary_op_insert_pieces (2, POINTER_PLUS_EXPR, |
4161 | TREE_TYPE (op), ops, result, |
4162 | VN_INFO (name: result)->value_id); |
4163 | return; |
4164 | } |
4165 | } |
4166 | |
4167 | vr1 = XOBNEW (&vn_tables_obstack, vn_reference_s); |
4168 | if (TREE_CODE (result) == SSA_NAME) |
4169 | vr1->value_id = VN_INFO (name: result)->value_id; |
4170 | else |
4171 | vr1->value_id = get_or_alloc_constant_value_id (constant: result); |
4172 | vr1->vuse = vuse_ssa_val (x: vuse); |
4173 | vr1->operands = operands.copy (); |
4174 | vr1->type = TREE_TYPE (op); |
4175 | vr1->punned = false; |
4176 | ao_ref op_ref; |
4177 | ao_ref_init (&op_ref, op); |
4178 | vr1->set = ao_ref_alias_set (&op_ref); |
4179 | vr1->base_set = ao_ref_base_alias_set (&op_ref); |
4180 | /* Specifically use an unknown extent here, we're not doing any lookup |
4181 | and assume the caller didn't either (or it went VARYING). */ |
4182 | vr1->offset = 0; |
4183 | vr1->max_size = -1; |
4184 | vr1->hashcode = vn_reference_compute_hash (vr1); |
4185 | vr1->result = TREE_CODE (result) == SSA_NAME ? SSA_VAL (x: result) : result; |
4186 | vr1->result_vdef = vdef; |
4187 | |
4188 | slot = valid_info->references->find_slot_with_hash (comparable: vr1, hash: vr1->hashcode, |
4189 | insert: INSERT); |
4190 | |
4191 | /* Because IL walking on reference lookup can end up visiting |
4192 | a def that is only to be visited later in iteration order |
4193 | when we are about to make an irreducible region reducible |
4194 | the def can be effectively processed and its ref being inserted |
4195 | by vn_reference_lookup_3 already. So we cannot assert (!*slot) |
4196 | but save a lookup if we deal with already inserted refs here. */ |
4197 | if (*slot) |
4198 | { |
4199 | /* We cannot assert that we have the same value either because |
4200 | when disentangling an irreducible region we may end up visiting |
4201 | a use before the corresponding def. That's a missed optimization |
4202 | only though. See gcc.dg/tree-ssa/pr87126.c for example. */ |
4203 | if (dump_file && (dump_flags & TDF_DETAILS) |
4204 | && !operand_equal_p ((*slot)->result, vr1->result, flags: 0)) |
4205 | { |
4206 | fprintf (stream: dump_file, format: "Keeping old value " ); |
4207 | print_generic_expr (dump_file, (*slot)->result); |
4208 | fprintf (stream: dump_file, format: " because of collision\n" ); |
4209 | } |
4210 | free_reference (vr: vr1); |
4211 | obstack_free (&vn_tables_obstack, vr1); |
4212 | return; |
4213 | } |
4214 | |
4215 | *slot = vr1; |
4216 | vr1->next = last_inserted_ref; |
4217 | last_inserted_ref = vr1; |
4218 | } |
4219 | |
4220 | /* Insert a reference by it's pieces into the current hash table with |
4221 | a value number of RESULT. Return the resulting reference |
4222 | structure we created. */ |
4223 | |
4224 | vn_reference_t |
4225 | vn_reference_insert_pieces (tree vuse, alias_set_type set, |
4226 | alias_set_type base_set, |
4227 | poly_int64 offset, poly_int64 max_size, tree type, |
4228 | vec<vn_reference_op_s> operands, |
4229 | tree result, unsigned int value_id) |
4230 | |
4231 | { |
4232 | vn_reference_s **slot; |
4233 | vn_reference_t vr1; |
4234 | |
4235 | vr1 = XOBNEW (&vn_tables_obstack, vn_reference_s); |
4236 | vr1->value_id = value_id; |
4237 | vr1->vuse = vuse_ssa_val (x: vuse); |
4238 | vr1->operands = operands; |
4239 | valueize_refs (orig: &vr1->operands); |
4240 | vr1->type = type; |
4241 | vr1->punned = false; |
4242 | vr1->set = set; |
4243 | vr1->base_set = base_set; |
4244 | vr1->offset = offset; |
4245 | vr1->max_size = max_size; |
4246 | vr1->hashcode = vn_reference_compute_hash (vr1); |
4247 | if (result && TREE_CODE (result) == SSA_NAME) |
4248 | result = SSA_VAL (x: result); |
4249 | vr1->result = result; |
4250 | vr1->result_vdef = NULL_TREE; |
4251 | |
4252 | slot = valid_info->references->find_slot_with_hash (comparable: vr1, hash: vr1->hashcode, |
4253 | insert: INSERT); |
4254 | |
4255 | /* At this point we should have all the things inserted that we have |
4256 | seen before, and we should never try inserting something that |
4257 | already exists. */ |
4258 | gcc_assert (!*slot); |
4259 | |
4260 | *slot = vr1; |
4261 | vr1->next = last_inserted_ref; |
4262 | last_inserted_ref = vr1; |
4263 | return vr1; |
4264 | } |
4265 | |
4266 | /* Compute and return the hash value for nary operation VBO1. */ |
4267 | |
4268 | hashval_t |
4269 | vn_nary_op_compute_hash (const vn_nary_op_t vno1) |
4270 | { |
4271 | inchash::hash hstate; |
4272 | unsigned i; |
4273 | |
4274 | if (((vno1->length == 2 |
4275 | && commutative_tree_code (vno1->opcode)) |
4276 | || (vno1->length == 3 |
4277 | && commutative_ternary_tree_code (vno1->opcode))) |
4278 | && tree_swap_operands_p (vno1->op[0], vno1->op[1])) |
4279 | std::swap (a&: vno1->op[0], b&: vno1->op[1]); |
4280 | else if (TREE_CODE_CLASS (vno1->opcode) == tcc_comparison |
4281 | && tree_swap_operands_p (vno1->op[0], vno1->op[1])) |
4282 | { |
4283 | std::swap (a&: vno1->op[0], b&: vno1->op[1]); |
4284 | vno1->opcode = swap_tree_comparison (vno1->opcode); |
4285 | } |
4286 | |
4287 | hstate.add_int (v: vno1->opcode); |
4288 | for (i = 0; i < vno1->length; ++i) |
4289 | inchash::add_expr (vno1->op[i], hstate); |
4290 | |
4291 | return hstate.end (); |
4292 | } |
4293 | |
4294 | /* Compare nary operations VNO1 and VNO2 and return true if they are |
4295 | equivalent. */ |
4296 | |
4297 | bool |
4298 | vn_nary_op_eq (const_vn_nary_op_t const vno1, const_vn_nary_op_t const vno2) |
4299 | { |
4300 | unsigned i; |
4301 | |
4302 | if (vno1->hashcode != vno2->hashcode) |
4303 | return false; |
4304 | |
4305 | if (vno1->length != vno2->length) |
4306 | return false; |
4307 | |
4308 | if (vno1->opcode != vno2->opcode |
4309 | || !types_compatible_p (type1: vno1->type, type2: vno2->type)) |
4310 | return false; |
4311 | |
4312 | for (i = 0; i < vno1->length; ++i) |
4313 | if (!expressions_equal_p (vno1->op[i], vno2->op[i])) |
4314 | return false; |
4315 | |
4316 | /* BIT_INSERT_EXPR has an implict operand as the type precision |
4317 | of op1. Need to check to make sure they are the same. */ |
4318 | if (vno1->opcode == BIT_INSERT_EXPR |
4319 | && TREE_CODE (vno1->op[1]) == INTEGER_CST |
4320 | && TYPE_PRECISION (TREE_TYPE (vno1->op[1])) |
4321 | != TYPE_PRECISION (TREE_TYPE (vno2->op[1]))) |
4322 | return false; |
4323 | |
4324 | return true; |
4325 | } |
4326 | |
4327 | /* Initialize VNO from the pieces provided. */ |
4328 | |
4329 | static void |
4330 | init_vn_nary_op_from_pieces (vn_nary_op_t vno, unsigned int length, |
4331 | enum tree_code code, tree type, tree *ops) |
4332 | { |
4333 | vno->opcode = code; |
4334 | vno->length = length; |
4335 | vno->type = type; |
4336 | memcpy (dest: &vno->op[0], src: ops, n: sizeof (tree) * length); |
4337 | } |
4338 | |
4339 | /* Return the number of operands for a vn_nary ops structure from STMT. */ |
4340 | |
4341 | unsigned int |
4342 | vn_nary_length_from_stmt (gimple *stmt) |
4343 | { |
4344 | switch (gimple_assign_rhs_code (gs: stmt)) |
4345 | { |
4346 | case REALPART_EXPR: |
4347 | case IMAGPART_EXPR: |
4348 | case VIEW_CONVERT_EXPR: |
4349 | return 1; |
4350 | |
4351 | case BIT_FIELD_REF: |
4352 | return 3; |
4353 | |
4354 | case CONSTRUCTOR: |
4355 | return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt)); |
4356 | |
4357 | default: |
4358 | return gimple_num_ops (gs: stmt) - 1; |
4359 | } |
4360 | } |
4361 | |
4362 | /* Initialize VNO from STMT. */ |
4363 | |
4364 | void |
4365 | init_vn_nary_op_from_stmt (vn_nary_op_t vno, gassign *stmt) |
4366 | { |
4367 | unsigned i; |
4368 | |
4369 | vno->opcode = gimple_assign_rhs_code (gs: stmt); |
4370 | vno->type = TREE_TYPE (gimple_assign_lhs (stmt)); |
4371 | switch (vno->opcode) |
4372 | { |
4373 | case REALPART_EXPR: |
4374 | case IMAGPART_EXPR: |
4375 | case VIEW_CONVERT_EXPR: |
4376 | vno->length = 1; |
4377 | vno->op[0] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0); |
4378 | break; |
4379 | |
4380 | case BIT_FIELD_REF: |
4381 | vno->length = 3; |
4382 | vno->op[0] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0); |
4383 | vno->op[1] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 1); |
4384 | vno->op[2] = TREE_OPERAND (gimple_assign_rhs1 (stmt), 2); |
4385 | break; |
4386 | |
4387 | case CONSTRUCTOR: |
4388 | vno->length = CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt)); |
4389 | for (i = 0; i < vno->length; ++i) |
4390 | vno->op[i] = CONSTRUCTOR_ELT (gimple_assign_rhs1 (stmt), i)->value; |
4391 | break; |
4392 | |
4393 | default: |
4394 | gcc_checking_assert (!gimple_assign_single_p (stmt)); |
4395 | vno->length = gimple_num_ops (gs: stmt) - 1; |
4396 | for (i = 0; i < vno->length; ++i) |
4397 | vno->op[i] = gimple_op (gs: stmt, i: i + 1); |
4398 | } |
4399 | } |
4400 | |
4401 | /* Compute the hashcode for VNO and look for it in the hash table; |
4402 | return the resulting value number if it exists in the hash table. |
4403 | Return NULL_TREE if it does not exist in the hash table or if the |
4404 | result field of the operation is NULL. VNRESULT will contain the |
4405 | vn_nary_op_t from the hashtable if it exists. */ |
4406 | |
4407 | static tree |
4408 | vn_nary_op_lookup_1 (vn_nary_op_t vno, vn_nary_op_t *vnresult) |
4409 | { |
4410 | vn_nary_op_s **slot; |
4411 | |
4412 | if (vnresult) |
4413 | *vnresult = NULL; |
4414 | |
4415 | for (unsigned i = 0; i < vno->length; ++i) |
4416 | if (TREE_CODE (vno->op[i]) == SSA_NAME) |
4417 | vno->op[i] = SSA_VAL (x: vno->op[i]); |
4418 | |
4419 | vno->hashcode = vn_nary_op_compute_hash (vno1: vno); |
4420 | slot = valid_info->nary->find_slot_with_hash (comparable: vno, hash: vno->hashcode, insert: NO_INSERT); |
4421 | if (!slot) |
4422 | return NULL_TREE; |
4423 | if (vnresult) |
4424 | *vnresult = *slot; |
4425 | return (*slot)->predicated_values ? NULL_TREE : (*slot)->u.result; |
4426 | } |
4427 | |
4428 | /* Lookup a n-ary operation by its pieces and return the resulting value |
4429 | number if it exists in the hash table. Return NULL_TREE if it does |
4430 | not exist in the hash table or if the result field of the operation |
4431 | is NULL. VNRESULT will contain the vn_nary_op_t from the hashtable |
4432 | if it exists. */ |
4433 | |
4434 | tree |
4435 | vn_nary_op_lookup_pieces (unsigned int length, enum tree_code code, |
4436 | tree type, tree *ops, vn_nary_op_t *vnresult) |
4437 | { |
4438 | vn_nary_op_t vno1 = XALLOCAVAR (struct vn_nary_op_s, |
4439 | sizeof_vn_nary_op (length)); |
4440 | init_vn_nary_op_from_pieces (vno: vno1, length, code, type, ops); |
4441 | return vn_nary_op_lookup_1 (vno: vno1, vnresult); |
4442 | } |
4443 | |
4444 | /* Lookup the rhs of STMT in the current hash table, and return the resulting |
4445 | value number if it exists in the hash table. Return NULL_TREE if |
4446 | it does not exist in the hash table. VNRESULT will contain the |
4447 | vn_nary_op_t from the hashtable if it exists. */ |
4448 | |
4449 | tree |
4450 | vn_nary_op_lookup_stmt (gimple *stmt, vn_nary_op_t *vnresult) |
4451 | { |
4452 | vn_nary_op_t vno1 |
4453 | = XALLOCAVAR (struct vn_nary_op_s, |
4454 | sizeof_vn_nary_op (vn_nary_length_from_stmt (stmt))); |
4455 | init_vn_nary_op_from_stmt (vno: vno1, stmt: as_a <gassign *> (p: stmt)); |
4456 | return vn_nary_op_lookup_1 (vno: vno1, vnresult); |
4457 | } |
4458 | |
4459 | /* Allocate a vn_nary_op_t with LENGTH operands on STACK. */ |
4460 | |
4461 | vn_nary_op_t |
4462 | alloc_vn_nary_op_noinit (unsigned int length, struct obstack *stack) |
4463 | { |
4464 | return (vn_nary_op_t) obstack_alloc (stack, sizeof_vn_nary_op (length)); |
4465 | } |
4466 | |
4467 | /* Allocate and initialize a vn_nary_op_t on CURRENT_INFO's |
4468 | obstack. */ |
4469 | |
4470 | static vn_nary_op_t |
4471 | alloc_vn_nary_op (unsigned int length, tree result, unsigned int value_id) |
4472 | { |
4473 | vn_nary_op_t vno1 = alloc_vn_nary_op_noinit (length, stack: &vn_tables_obstack); |
4474 | |
4475 | vno1->value_id = value_id; |
4476 | vno1->length = length; |
4477 | vno1->predicated_values = 0; |
4478 | vno1->u.result = result; |
4479 | |
4480 | return vno1; |
4481 | } |
4482 | |
4483 | /* Insert VNO into TABLE. */ |
4484 | |
4485 | static vn_nary_op_t |
4486 | vn_nary_op_insert_into (vn_nary_op_t vno, vn_nary_op_table_type *table) |
4487 | { |
4488 | vn_nary_op_s **slot; |
4489 | |
4490 | gcc_assert (! vno->predicated_values |
4491 | || (! vno->u.values->next |
4492 | && vno->u.values->n == 1)); |
4493 | |
4494 | for (unsigned i = 0; i < vno->length; ++i) |
4495 | if (TREE_CODE (vno->op[i]) == SSA_NAME) |
4496 | vno->op[i] = SSA_VAL (x: vno->op[i]); |
4497 | |
4498 | vno->hashcode = vn_nary_op_compute_hash (vno1: vno); |
4499 | slot = table->find_slot_with_hash (comparable: vno, hash: vno->hashcode, insert: INSERT); |
4500 | vno->unwind_to = *slot; |
4501 | if (*slot) |
4502 | { |
4503 | /* Prefer non-predicated values. |
4504 | ??? Only if those are constant, otherwise, with constant predicated |
4505 | value, turn them into predicated values with entry-block validity |
4506 | (??? but we always find the first valid result currently). */ |
4507 | if ((*slot)->predicated_values |
4508 | && ! vno->predicated_values) |
4509 | { |
4510 | /* ??? We cannot remove *slot from the unwind stack list. |
4511 | For the moment we deal with this by skipping not found |
4512 | entries but this isn't ideal ... */ |
4513 | *slot = vno; |
4514 | /* ??? Maintain a stack of states we can unwind in |
4515 | vn_nary_op_s? But how far do we unwind? In reality |
4516 | we need to push change records somewhere... Or not |
4517 | unwind vn_nary_op_s and linking them but instead |
4518 | unwind the results "list", linking that, which also |
4519 | doesn't move on hashtable resize. */ |
4520 | /* We can also have a ->unwind_to recording *slot there. |
4521 | That way we can make u.values a fixed size array with |
4522 | recording the number of entries but of course we then |
4523 | have always N copies for each unwind_to-state. Or we |
4524 | make sure to only ever append and each unwinding will |
4525 | pop off one entry (but how to deal with predicated |
4526 | replaced with non-predicated here?) */ |
4527 | vno->next = last_inserted_nary; |
4528 | last_inserted_nary = vno; |
4529 | return vno; |
4530 | } |
4531 | else if (vno->predicated_values |
4532 | && ! (*slot)->predicated_values) |
4533 | return *slot; |
4534 | else if (vno->predicated_values |
4535 | && (*slot)->predicated_values) |
4536 | { |
4537 | /* ??? Factor this all into a insert_single_predicated_value |
4538 | routine. */ |
4539 | gcc_assert (!vno->u.values->next && vno->u.values->n == 1); |
4540 | basic_block vno_bb |
4541 | = BASIC_BLOCK_FOR_FN (cfun, vno->u.values->valid_dominated_by_p[0]); |
4542 | vn_pval *nval = vno->u.values; |
4543 | vn_pval **next = &vno->u.values; |
4544 | bool found = false; |
4545 | for (vn_pval *val = (*slot)->u.values; val; val = val->next) |
4546 | { |
4547 | if (expressions_equal_p (val->result, nval->result)) |
4548 | { |
4549 | found = true; |
4550 | for (unsigned i = 0; i < val->n; ++i) |
4551 | { |
4552 | basic_block val_bb |
4553 | = BASIC_BLOCK_FOR_FN (cfun, |
4554 | val->valid_dominated_by_p[i]); |
4555 | if (dominated_by_p (CDI_DOMINATORS, vno_bb, val_bb)) |
4556 | /* Value registered with more generic predicate. */ |
4557 | return *slot; |
4558 | else if (flag_checking) |
4559 | /* Shouldn't happen, we insert in RPO order. */ |
4560 | gcc_assert (!dominated_by_p (CDI_DOMINATORS, |
4561 | val_bb, vno_bb)); |
4562 | } |
4563 | /* Append value. */ |
4564 | *next = (vn_pval *) obstack_alloc (&vn_tables_obstack, |
4565 | sizeof (vn_pval) |
4566 | + val->n * sizeof (int)); |
4567 | (*next)->next = NULL; |
4568 | (*next)->result = val->result; |
4569 | (*next)->n = val->n + 1; |
4570 | memcpy (dest: (*next)->valid_dominated_by_p, |
4571 | src: val->valid_dominated_by_p, |
4572 | n: val->n * sizeof (int)); |
4573 | (*next)->valid_dominated_by_p[val->n] = vno_bb->index; |
4574 | next = &(*next)->next; |
4575 | if (dump_file && (dump_flags & TDF_DETAILS)) |
4576 | fprintf (stream: dump_file, format: "Appending predicate to value.\n" ); |
4577 | continue; |
4578 | } |
4579 | /* Copy other predicated values. */ |
4580 | *next = (vn_pval *) obstack_alloc (&vn_tables_obstack, |
4581 | sizeof (vn_pval) |
4582 | + (val->n-1) * sizeof (int)); |
4583 | memcpy (dest: *next, src: val, n: sizeof (vn_pval) + (val->n-1) * sizeof (int)); |
4584 | (*next)->next = NULL; |
4585 | next = &(*next)->next; |
4586 | } |
4587 | if (!found) |
4588 | *next = nval; |
4589 | |
4590 | *slot = vno; |
4591 | vno->next = last_inserted_nary; |
4592 | last_inserted_nary = vno; |
4593 | return vno; |
4594 | } |
4595 | |
4596 | /* While we do not want to insert things twice it's awkward to |
4597 | avoid it in the case where visit_nary_op pattern-matches stuff |
4598 | and ends up simplifying the replacement to itself. We then |
4599 | get two inserts, one from visit_nary_op and one from |
4600 | vn_nary_build_or_lookup. |
4601 | So allow inserts with the same value number. */ |
4602 | if ((*slot)->u.result == vno->u.result) |
4603 | return *slot; |
4604 | } |
4605 | |
4606 | /* ??? There's also optimistic vs. previous commited state merging |
4607 | that is problematic for the case of unwinding. */ |
4608 | |
4609 | /* ??? We should return NULL if we do not use 'vno' and have the |
4610 | caller release it. */ |
4611 | gcc_assert (!*slot); |
4612 | |
4613 | *slot = vno; |
4614 | vno->next = last_inserted_nary; |
4615 | last_inserted_nary = vno; |
4616 | return vno; |
4617 | } |
4618 | |
4619 | /* Insert a n-ary operation into the current hash table using it's |
4620 | pieces. Return the vn_nary_op_t structure we created and put in |
4621 | the hashtable. */ |
4622 | |
4623 | vn_nary_op_t |
4624 | vn_nary_op_insert_pieces (unsigned int length, enum tree_code code, |
4625 | tree type, tree *ops, |
4626 | tree result, unsigned int value_id) |
4627 | { |
4628 | vn_nary_op_t vno1 = alloc_vn_nary_op (length, result, value_id); |
4629 | init_vn_nary_op_from_pieces (vno: vno1, length, code, type, ops); |
4630 | return vn_nary_op_insert_into (vno: vno1, table: valid_info->nary); |
4631 | } |
4632 | |
4633 | /* Return whether we can track a predicate valid when PRED_E is executed. */ |
4634 | |
4635 | static bool |
4636 | can_track_predicate_on_edge (edge pred_e) |
4637 | { |
4638 | /* ??? As we are currently recording the destination basic-block index in |
4639 | vn_pval.valid_dominated_by_p and using dominance for the |
4640 | validity check we cannot track predicates on all edges. */ |
4641 | if (single_pred_p (bb: pred_e->dest)) |
4642 | return true; |
4643 | /* Never record for backedges. */ |
4644 | if (pred_e->flags & EDGE_DFS_BACK) |
4645 | return false; |
4646 | /* When there's more than one predecessor we cannot track |
4647 | predicate validity based on the destination block. The |
4648 | exception is when all other incoming edges sources are |
4649 | dominated by the destination block. */ |
4650 | edge_iterator ei; |
4651 | edge e; |
4652 | FOR_EACH_EDGE (e, ei, pred_e->dest->preds) |
4653 | if (e != pred_e && ! dominated_by_p (CDI_DOMINATORS, e->src, e->dest)) |
4654 | return false; |
4655 | return true; |
4656 | } |
4657 | |
4658 | static vn_nary_op_t |
4659 | vn_nary_op_insert_pieces_predicated (unsigned int length, enum tree_code code, |
4660 | tree type, tree *ops, |
4661 | tree result, unsigned int value_id, |
4662 | edge pred_e) |
4663 | { |
4664 | gcc_assert (can_track_predicate_on_edge (pred_e)); |
4665 | |
4666 | if (dump_file && (dump_flags & TDF_DETAILS) |
4667 | /* ??? Fix dumping, but currently we only get comparisons. */ |
4668 | && TREE_CODE_CLASS (code) == tcc_comparison) |
4669 | { |
4670 | fprintf (stream: dump_file, format: "Recording on edge %d->%d " , pred_e->src->index, |
4671 | pred_e->dest->index); |
4672 | print_generic_expr (dump_file, ops[0], TDF_SLIM); |
4673 | fprintf (stream: dump_file, format: " %s " , get_tree_code_name (code)); |
4674 | print_generic_expr (dump_file, ops[1], TDF_SLIM); |
4675 | fprintf (stream: dump_file, format: " == %s\n" , |
4676 | integer_zerop (result) ? "false" : "true" ); |
4677 | } |
4678 | vn_nary_op_t vno1 = alloc_vn_nary_op (length, NULL_TREE, value_id); |
4679 | init_vn_nary_op_from_pieces (vno: vno1, length, code, type, ops); |
4680 | vno1->predicated_values = 1; |
4681 | vno1->u.values = (vn_pval *) obstack_alloc (&vn_tables_obstack, |
4682 | sizeof (vn_pval)); |
4683 | vno1->u.values->next = NULL; |
4684 | vno1->u.values->result = result; |
4685 | vno1->u.values->n = 1; |
4686 | vno1->u.values->valid_dominated_by_p[0] = pred_e->dest->index; |
4687 | return vn_nary_op_insert_into (vno: vno1, table: valid_info->nary); |
4688 | } |
4689 | |
4690 | static bool |
4691 | dominated_by_p_w_unex (basic_block bb1, basic_block bb2, bool); |
4692 | |
4693 | static tree |
4694 | vn_nary_op_get_predicated_value (vn_nary_op_t vno, basic_block bb, |
4695 | edge e = NULL) |
4696 | { |
4697 | if (! vno->predicated_values) |
4698 | return vno->u.result; |
4699 | for (vn_pval *val = vno->u.values; val; val = val->next) |
4700 | for (unsigned i = 0; i < val->n; ++i) |
4701 | { |
4702 | basic_block cand |
4703 | = BASIC_BLOCK_FOR_FN (cfun, val->valid_dominated_by_p[i]); |
4704 | /* Do not handle backedge executability optimistically since |
4705 | when figuring out whether to iterate we do not consider |
4706 | changed predication. |
4707 | When asking for predicated values on an edge avoid looking |
4708 | at edge executability for edges forward in our iteration |
4709 | as well. */ |
4710 | if (e && (e->flags & EDGE_DFS_BACK)) |
4711 | { |
4712 | if (dominated_by_p (CDI_DOMINATORS, bb, cand)) |
4713 | return val->result; |
4714 | } |
4715 | else if (dominated_by_p_w_unex (bb1: bb, bb2: cand, false)) |
4716 | return val->result; |
4717 | } |
4718 | return NULL_TREE; |
4719 | } |
4720 | |
4721 | static tree |
4722 | vn_nary_op_get_predicated_value (vn_nary_op_t vno, edge e) |
4723 | { |
4724 | return vn_nary_op_get_predicated_value (vno, bb: e->src, e); |
4725 | } |
4726 | |
4727 | /* Insert the rhs of STMT into the current hash table with a value number of |
4728 | RESULT. */ |
4729 | |
4730 | static vn_nary_op_t |
4731 | vn_nary_op_insert_stmt (gimple *stmt, tree result) |
4732 | { |
4733 | vn_nary_op_t vno1 |
4734 | = alloc_vn_nary_op (length: vn_nary_length_from_stmt (stmt), |
4735 | result, value_id: VN_INFO (name: result)->value_id); |
4736 | init_vn_nary_op_from_stmt (vno: vno1, stmt: as_a <gassign *> (p: stmt)); |
4737 | return vn_nary_op_insert_into (vno: vno1, table: valid_info->nary); |
4738 | } |
4739 | |
4740 | /* Compute a hashcode for PHI operation VP1 and return it. */ |
4741 | |
4742 | static inline hashval_t |
4743 | vn_phi_compute_hash (vn_phi_t vp1) |
4744 | { |
4745 | inchash::hash hstate; |
4746 | tree phi1op; |
4747 | tree type; |
4748 | edge e; |
4749 | edge_iterator ei; |
4750 | |
4751 | hstate.add_int (EDGE_COUNT (vp1->block->preds)); |
4752 | switch (EDGE_COUNT (vp1->block->preds)) |
4753 | { |
4754 | case 1: |
4755 | break; |
4756 | case 2: |
4757 | /* When this is a PHI node subject to CSE for different blocks |
4758 | avoid hashing the block index. */ |
4759 | if (vp1->cclhs) |
4760 | break; |
4761 | /* Fallthru. */ |
4762 | default: |
4763 | hstate.add_int (v: vp1->block->index); |
4764 | } |
4765 | |
4766 | /* If all PHI arguments are constants we need to distinguish |
4767 | the PHI node via its type. */ |
4768 | type = vp1->type; |
4769 | hstate.merge_hash (other: vn_hash_type (type)); |
4770 | |
4771 | FOR_EACH_EDGE (e, ei, vp1->block->preds) |
4772 | { |
4773 | /* Don't hash backedge values they need to be handled as VN_TOP |
4774 | for optimistic value-numbering. */ |
4775 | if (e->flags & EDGE_DFS_BACK) |
4776 | continue; |
4777 | |
4778 | phi1op = vp1->phiargs[e->dest_idx]; |
4779 | if (phi1op == VN_TOP) |
4780 | continue; |
4781 | inchash::add_expr (phi1op, hstate); |
4782 | } |
4783 | |
4784 | return hstate.end (); |
4785 | } |
4786 | |
4787 | |
4788 | /* Return true if COND1 and COND2 represent the same condition, set |
4789 | *INVERTED_P if one needs to be inverted to make it the same as |
4790 | the other. */ |
4791 | |
4792 | static bool |
4793 | cond_stmts_equal_p (gcond *cond1, tree lhs1, tree rhs1, |
4794 | gcond *cond2, tree lhs2, tree rhs2, bool *inverted_p) |
4795 | { |
4796 | enum tree_code code1 = gimple_cond_code (gs: cond1); |
4797 | enum tree_code code2 = gimple_cond_code (gs: cond2); |
4798 | |
4799 | *inverted_p = false; |
4800 | if (code1 == code2) |
4801 | ; |
4802 | else if (code1 == swap_tree_comparison (code2)) |
4803 | std::swap (a&: lhs2, b&: rhs2); |
4804 | else if (code1 == invert_tree_comparison (code2, HONOR_NANS (lhs2))) |
4805 | *inverted_p = true; |
4806 | else if (code1 == invert_tree_comparison |
4807 | (swap_tree_comparison (code2), HONOR_NANS (lhs2))) |
4808 | { |
4809 | std::swap (a&: lhs2, b&: rhs2); |
4810 | *inverted_p = true; |
4811 | } |
4812 | else |
4813 | return false; |
4814 | |
4815 | return ((expressions_equal_p (lhs1, lhs2) |
4816 | && expressions_equal_p (rhs1, rhs2)) |
4817 | || (commutative_tree_code (code1) |
4818 | && expressions_equal_p (lhs1, rhs2) |
4819 | && expressions_equal_p (rhs1, lhs2))); |
4820 | } |
4821 | |
4822 | /* Compare two phi entries for equality, ignoring VN_TOP arguments. */ |
4823 | |
4824 | static int |
4825 | vn_phi_eq (const_vn_phi_t const vp1, const_vn_phi_t const vp2) |
4826 | { |
4827 | if (vp1->hashcode != vp2->hashcode) |
4828 | return false; |
4829 | |
4830 | if (vp1->block != vp2->block) |
4831 | { |
4832 | if (EDGE_COUNT (vp1->block->preds) != EDGE_COUNT (vp2->block->preds)) |
4833 | return false; |
4834 | |
4835 | switch (EDGE_COUNT (vp1->block->preds)) |
4836 | { |
4837 | case 1: |
4838 | /* Single-arg PHIs are just copies. */ |
4839 | break; |
4840 | |
4841 | case 2: |
4842 | { |
4843 | /* Make sure both PHIs are classified as CSEable. */ |
4844 | if (! vp1->cclhs || ! vp2->cclhs) |
4845 | return false; |
4846 | |
4847 | /* Rule out backedges into the PHI. */ |
4848 | gcc_checking_assert |
4849 | (vp1->block->loop_father->header != vp1->block |
4850 | && vp2->block->loop_father->header != vp2->block); |
4851 | |
4852 | /* If the PHI nodes do not have compatible types |
4853 | they are not the same. */ |
4854 | if (!types_compatible_p (type1: vp1->type, type2: vp2->type)) |
4855 | return false; |
4856 | |
4857 | /* If the immediate dominator end in switch stmts multiple |
4858 | values may end up in the same PHI arg via intermediate |
4859 | CFG merges. */ |
4860 | basic_block idom1 |
4861 | = get_immediate_dominator (CDI_DOMINATORS, vp1->block); |
4862 | basic_block idom2 |
4863 | = get_immediate_dominator (CDI_DOMINATORS, vp2->block); |
4864 | gcc_checking_assert (EDGE_COUNT (idom1->succs) == 2 |
4865 | && EDGE_COUNT (idom2->succs) == 2); |
4866 | |
4867 | /* Verify the controlling stmt is the same. */ |
4868 | gcond *last1 = as_a <gcond *> (p: *gsi_last_bb (bb: idom1)); |
4869 | gcond *last2 = as_a <gcond *> (p: *gsi_last_bb (bb: idom2)); |
4870 | bool inverted_p; |
4871 | if (! cond_stmts_equal_p (cond1: last1, lhs1: vp1->cclhs, rhs1: vp1->ccrhs, |
4872 | cond2: last2, lhs2: vp2->cclhs, rhs2: vp2->ccrhs, |
4873 | inverted_p: &inverted_p)) |
4874 | return false; |
4875 | |
4876 | /* Get at true/false controlled edges into the PHI. */ |
4877 | edge te1, te2, fe1, fe2; |
4878 | if (! extract_true_false_controlled_edges (idom1, vp1->block, |
4879 | &te1, &fe1) |
4880 | || ! extract_true_false_controlled_edges (idom2, vp2->block, |
4881 | &te2, &fe2)) |
4882 | return false; |
4883 | |
4884 | /* Swap edges if the second condition is the inverted of the |
4885 | first. */ |
4886 | if (inverted_p) |
4887 | std::swap (a&: te2, b&: fe2); |
4888 | |
4889 | /* Since we do not know which edge will be executed we have |
4890 | to be careful when matching VN_TOP. Be conservative and |
4891 | only match VN_TOP == VN_TOP for now, we could allow |
4892 | VN_TOP on the not prevailing PHI though. See for example |
4893 | PR102920. */ |
4894 | if (! expressions_equal_p (vp1->phiargs[te1->dest_idx], |
4895 | vp2->phiargs[te2->dest_idx], false) |
4896 | || ! expressions_equal_p (vp1->phiargs[fe1->dest_idx], |
4897 | vp2->phiargs[fe2->dest_idx], false)) |
4898 | return false; |
4899 | |
4900 | return true; |
4901 | } |
4902 | |
4903 | default: |
4904 | return false; |
4905 | } |
4906 | } |
4907 | |
4908 | /* If the PHI nodes do not have compatible types |
4909 | they are not the same. */ |
4910 | if (!types_compatible_p (type1: vp1->type, type2: vp2->type)) |
4911 | return false; |
4912 | |
4913 | /* Any phi in the same block will have it's arguments in the |
4914 | same edge order, because of how we store phi nodes. */ |
4915 | unsigned nargs = EDGE_COUNT (vp1->block->preds); |
4916 | for (unsigned i = 0; i < nargs; ++i) |
4917 | { |
4918 | tree phi1op = vp1->phiargs[i]; |
4919 | tree phi2op = vp2->phiargs[i]; |
4920 | if (phi1op == phi2op) |
4921 | continue; |
4922 | if (!expressions_equal_p (phi1op, phi2op, false)) |
4923 | return false; |
4924 | } |
4925 | |
4926 | return true; |
4927 | } |
4928 | |
4929 | /* Lookup PHI in the current hash table, and return the resulting |
4930 | value number if it exists in the hash table. Return NULL_TREE if |
4931 | it does not exist in the hash table. */ |
4932 | |
4933 | static tree |
4934 | vn_phi_lookup (gimple *phi, bool backedges_varying_p) |
4935 | { |
4936 | vn_phi_s **slot; |
4937 | struct vn_phi_s *vp1; |
4938 | edge e; |
4939 | edge_iterator ei; |
4940 | |
4941 | vp1 = XALLOCAVAR (struct vn_phi_s, |
4942 | sizeof (struct vn_phi_s) |
4943 | + (gimple_phi_num_args (phi) - 1) * sizeof (tree)); |
4944 | |
4945 | /* Canonicalize the SSA_NAME's to their value number. */ |
4946 | FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds) |
4947 | { |
4948 | tree def = PHI_ARG_DEF_FROM_EDGE (phi, e); |
4949 | if (TREE_CODE (def) == SSA_NAME |
4950 | && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK))) |
4951 | { |
4952 | if (!virtual_operand_p (op: def) |
4953 | && ssa_undefined_value_p (def, false)) |
4954 | def = VN_TOP; |
4955 | else |
4956 | def = SSA_VAL (x: def); |
4957 | } |
4958 | vp1->phiargs[e->dest_idx] = def; |
4959 | } |
4960 | vp1->type = TREE_TYPE (gimple_phi_result (phi)); |
4961 | vp1->block = gimple_bb (g: phi); |
4962 | /* Extract values of the controlling condition. */ |
4963 | vp1->cclhs = NULL_TREE; |
4964 | vp1->ccrhs = NULL_TREE; |
4965 | if (EDGE_COUNT (vp1->block->preds) == 2 |
4966 | && vp1->block->loop_father->header != vp1->block) |
4967 | { |
4968 | basic_block idom1 = get_immediate_dominator (CDI_DOMINATORS, vp1->block); |
4969 | if (EDGE_COUNT (idom1->succs) == 2) |
4970 | if (gcond *last1 = safe_dyn_cast <gcond *> (p: *gsi_last_bb (bb: idom1))) |
4971 | { |
4972 | /* ??? We want to use SSA_VAL here. But possibly not |
4973 | allow VN_TOP. */ |
4974 | vp1->cclhs = vn_valueize (gimple_cond_lhs (gs: last1)); |
4975 | vp1->ccrhs = vn_valueize (gimple_cond_rhs (gs: last1)); |
4976 | } |
4977 | } |
4978 | vp1->hashcode = vn_phi_compute_hash (vp1); |
4979 | slot = valid_info->phis->find_slot_with_hash (comparable: vp1, hash: vp1->hashcode, insert: NO_INSERT); |
4980 | if (!slot) |
4981 | return NULL_TREE; |
4982 | return (*slot)->result; |
4983 | } |
4984 | |
4985 | /* Insert PHI into the current hash table with a value number of |
4986 | RESULT. */ |
4987 | |
4988 | static vn_phi_t |
4989 | vn_phi_insert (gimple *phi, tree result, bool backedges_varying_p) |
4990 | { |
4991 | vn_phi_s **slot; |
4992 | vn_phi_t vp1 = (vn_phi_t) obstack_alloc (&vn_tables_obstack, |
4993 | sizeof (vn_phi_s) |
4994 | + ((gimple_phi_num_args (phi) - 1) |
4995 | * sizeof (tree))); |
4996 | edge e; |
4997 | edge_iterator ei; |
4998 | |
4999 | /* Canonicalize the SSA_NAME's to their value number. */ |
5000 | FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds) |
5001 | { |
5002 | tree def = PHI_ARG_DEF_FROM_EDGE (phi, e); |
5003 | if (TREE_CODE (def) == SSA_NAME |
5004 | && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK))) |
5005 | { |
5006 | if (!virtual_operand_p (op: def) |
5007 | && ssa_undefined_value_p (def, false)) |
5008 | def = VN_TOP; |
5009 | else |
5010 | def = SSA_VAL (x: def); |
5011 | } |
5012 | vp1->phiargs[e->dest_idx] = def; |
5013 | } |
5014 | vp1->value_id = VN_INFO (name: result)->value_id; |
5015 | vp1->type = TREE_TYPE (gimple_phi_result (phi)); |
5016 | vp1->block = gimple_bb (g: phi); |
5017 | /* Extract values of the controlling condition. */ |
5018 | vp1->cclhs = NULL_TREE; |
5019 | vp1->ccrhs = NULL_TREE; |
5020 | if (EDGE_COUNT (vp1->block->preds) == 2 |
5021 | && vp1->block->loop_father->header != vp1->block) |
5022 | { |
5023 | basic_block idom1 = get_immediate_dominator (CDI_DOMINATORS, vp1->block); |
5024 | if (EDGE_COUNT (idom1->succs) == 2) |
5025 | if (gcond *last1 = safe_dyn_cast <gcond *> (p: *gsi_last_bb (bb: idom1))) |
5026 | { |
5027 | /* ??? We want to use SSA_VAL here. But possibly not |
5028 | allow VN_TOP. */ |
5029 | vp1->cclhs = vn_valueize (gimple_cond_lhs (gs: last1)); |
5030 | vp1->ccrhs = vn_valueize (gimple_cond_rhs (gs: last1)); |
5031 | } |
5032 | } |
5033 | vp1->result = result; |
5034 | vp1->hashcode = vn_phi_compute_hash (vp1); |
5035 | |
5036 | slot = valid_info->phis->find_slot_with_hash (comparable: vp1, hash: vp1->hashcode, insert: INSERT); |
5037 | gcc_assert (!*slot); |
5038 | |
5039 | *slot = vp1; |
5040 | vp1->next = last_inserted_phi; |
5041 | last_inserted_phi = vp1; |
5042 | return vp1; |
5043 | } |
5044 | |
5045 | |
5046 | /* Return true if BB1 is dominated by BB2 taking into account edges |
5047 | that are not executable. When ALLOW_BACK is false consider not |
5048 | executable backedges as executable. */ |
5049 | |
5050 | static bool |
5051 | dominated_by_p_w_unex (basic_block bb1, basic_block bb2, bool allow_back) |
5052 | { |
5053 | edge_iterator ei; |
5054 | edge e; |
5055 | |
5056 | if (dominated_by_p (CDI_DOMINATORS, bb1, bb2)) |
5057 | return true; |
5058 | |
5059 | /* Before iterating we'd like to know if there exists a |
5060 | (executable) path from bb2 to bb1 at all, if not we can |
5061 | directly return false. For now simply iterate once. */ |
5062 | |
5063 | /* Iterate to the single executable bb1 predecessor. */ |
5064 | if (EDGE_COUNT (bb1->preds) > 1) |
5065 | { |
5066 | edge prede = NULL; |
5067 | FOR_EACH_EDGE (e, ei, bb1->preds) |
5068 | if ((e->flags & EDGE_EXECUTABLE) |
5069 | || (!allow_back && (e->flags & EDGE_DFS_BACK))) |
5070 | { |
5071 | if (prede) |
5072 | { |
5073 | prede = NULL; |
5074 | break; |
5075 | } |
5076 | prede = e; |
5077 | } |
5078 | if (prede) |
5079 | { |
5080 | bb1 = prede->src; |
5081 | |
5082 | /* Re-do the dominance check with changed bb1. */ |
5083 | if (dominated_by_p (CDI_DOMINATORS, bb1, bb2)) |
5084 | return true; |
5085 | } |
5086 | } |
5087 | |
5088 | /* Iterate to the single executable bb2 successor. */ |
5089 | if (EDGE_COUNT (bb2->succs) > 1) |
5090 | { |
5091 | edge succe = NULL; |
5092 | FOR_EACH_EDGE (e, ei, bb2->succs) |
5093 | if ((e->flags & EDGE_EXECUTABLE) |
5094 | || (!allow_back && (e->flags & EDGE_DFS_BACK))) |
5095 | { |
5096 | if (succe) |
5097 | { |
5098 | succe = NULL; |
5099 | break; |
5100 | } |
5101 | succe = e; |
5102 | } |
5103 | if (succe) |
5104 | { |
5105 | /* Verify the reached block is only reached through succe. |
5106 | If there is only one edge we can spare us the dominator |
5107 | check and iterate directly. */ |
5108 | if (EDGE_COUNT (succe->dest->preds) > 1) |
5109 | { |
5110 | FOR_EACH_EDGE (e, ei, succe->dest->preds) |
5111 | if (e != succe |
5112 | && ((e->flags & EDGE_EXECUTABLE) |
5113 | || (!allow_back && (e->flags & EDGE_DFS_BACK)))) |
5114 | { |
5115 | succe = NULL; |
5116 | break; |
5117 | } |
5118 | } |
5119 | if (succe) |
5120 | { |
5121 | bb2 = succe->dest; |
5122 | |
5123 | /* Re-do the dominance check with changed bb2. */ |
5124 | if (dominated_by_p (CDI_DOMINATORS, bb1, bb2)) |
5125 | return true; |
5126 | } |
5127 | } |
5128 | } |
5129 | |
5130 | /* We could now iterate updating bb1 / bb2. */ |
5131 | return false; |
5132 | } |
5133 | |
5134 | /* Set the value number of FROM to TO, return true if it has changed |
5135 | as a result. */ |
5136 | |
5137 | static inline bool |
5138 | set_ssa_val_to (tree from, tree to) |
5139 | { |
5140 | vn_ssa_aux_t from_info = VN_INFO (name: from); |
5141 | tree currval = from_info->valnum; // SSA_VAL (from) |
5142 | poly_int64 toff, coff; |
5143 | bool curr_undefined = false; |
5144 | bool curr_invariant = false; |
5145 | |
5146 | /* The only thing we allow as value numbers are ssa_names |
5147 | and invariants. So assert that here. We don't allow VN_TOP |
5148 | as visiting a stmt should produce a value-number other than |
5149 | that. |
5150 | ??? Still VN_TOP can happen for unreachable code, so force |
5151 | it to varying in that case. Not all code is prepared to |
5152 | get VN_TOP on valueization. */ |
5153 | if (to == VN_TOP) |
5154 | { |
5155 | /* ??? When iterating and visiting PHI <undef, backedge-value> |
5156 | for the first time we rightfully get VN_TOP and we need to |
5157 | preserve that to optimize for example gcc.dg/tree-ssa/ssa-sccvn-2.c. |
5158 | With SCCVN we were simply lucky we iterated the other PHI |
5159 | cycles first and thus visited the backedge-value DEF. */ |
5160 | if (currval == VN_TOP) |
5161 | goto set_and_exit; |
5162 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5163 | fprintf (stream: dump_file, format: "Forcing value number to varying on " |
5164 | "receiving VN_TOP\n" ); |
5165 | to = from; |
5166 | } |
5167 | |
5168 | gcc_checking_assert (to != NULL_TREE |
5169 | && ((TREE_CODE (to) == SSA_NAME |
5170 | && (to == from || SSA_VAL (to) == to)) |
5171 | || is_gimple_min_invariant (to))); |
5172 | |
5173 | if (from != to) |
5174 | { |
5175 | if (currval == from) |
5176 | { |
5177 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5178 | { |
5179 | fprintf (stream: dump_file, format: "Not changing value number of " ); |
5180 | print_generic_expr (dump_file, from); |
5181 | fprintf (stream: dump_file, format: " from VARYING to " ); |
5182 | print_generic_expr (dump_file, to); |
5183 | fprintf (stream: dump_file, format: "\n" ); |
5184 | } |
5185 | return false; |
5186 | } |
5187 | curr_invariant = is_gimple_min_invariant (currval); |
5188 | curr_undefined = (TREE_CODE (currval) == SSA_NAME |
5189 | && !virtual_operand_p (op: currval) |
5190 | && ssa_undefined_value_p (currval, false)); |
5191 | if (currval != VN_TOP |
5192 | && !curr_invariant |
5193 | && !curr_undefined |
5194 | && is_gimple_min_invariant (to)) |
5195 | { |
5196 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5197 | { |
5198 | fprintf (stream: dump_file, format: "Forcing VARYING instead of changing " |
5199 | "value number of " ); |
5200 | print_generic_expr (dump_file, from); |
5201 | fprintf (stream: dump_file, format: " from " ); |
5202 | print_generic_expr (dump_file, currval); |
5203 | fprintf (stream: dump_file, format: " (non-constant) to " ); |
5204 | print_generic_expr (dump_file, to); |
5205 | fprintf (stream: dump_file, format: " (constant)\n" ); |
5206 | } |
5207 | to = from; |
5208 | } |
5209 | else if (currval != VN_TOP |
5210 | && !curr_undefined |
5211 | && TREE_CODE (to) == SSA_NAME |
5212 | && !virtual_operand_p (op: to) |
5213 | && ssa_undefined_value_p (to, false)) |
5214 | { |
5215 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5216 | { |
5217 | fprintf (stream: dump_file, format: "Forcing VARYING instead of changing " |
5218 | "value number of " ); |
5219 | print_generic_expr (dump_file, from); |
5220 | fprintf (stream: dump_file, format: " from " ); |
5221 | print_generic_expr (dump_file, currval); |
5222 | fprintf (stream: dump_file, format: " (non-undefined) to " ); |
5223 | print_generic_expr (dump_file, to); |
5224 | fprintf (stream: dump_file, format: " (undefined)\n" ); |
5225 | } |
5226 | to = from; |
5227 | } |
5228 | else if (TREE_CODE (to) == SSA_NAME |
5229 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (to)) |
5230 | to = from; |
5231 | } |
5232 | |
5233 | set_and_exit: |
5234 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5235 | { |
5236 | fprintf (stream: dump_file, format: "Setting value number of " ); |
5237 | print_generic_expr (dump_file, from); |
5238 | fprintf (stream: dump_file, format: " to " ); |
5239 | print_generic_expr (dump_file, to); |
5240 | } |
5241 | |
5242 | if (currval != to |
5243 | && !operand_equal_p (currval, to, flags: 0) |
5244 | /* Different undefined SSA names are not actually different. See |
5245 | PR82320 for a testcase were we'd otherwise not terminate iteration. */ |
5246 | && !(curr_undefined |
5247 | && TREE_CODE (to) == SSA_NAME |
5248 | && !virtual_operand_p (op: to) |
5249 | && ssa_undefined_value_p (to, false)) |
5250 | /* ??? For addresses involving volatile objects or types operand_equal_p |
5251 | does not reliably detect ADDR_EXPRs as equal. We know we are only |
5252 | getting invariant gimple addresses here, so can use |
5253 | get_addr_base_and_unit_offset to do this comparison. */ |
5254 | && !(TREE_CODE (currval) == ADDR_EXPR |
5255 | && TREE_CODE (to) == ADDR_EXPR |
5256 | && (get_addr_base_and_unit_offset (TREE_OPERAND (currval, 0), &coff) |
5257 | == get_addr_base_and_unit_offset (TREE_OPERAND (to, 0), &toff)) |
5258 | && known_eq (coff, toff))) |
5259 | { |
5260 | if (to != from |
5261 | && currval != VN_TOP |
5262 | && !curr_undefined |
5263 | /* We do not want to allow lattice transitions from one value |
5264 | to another since that may lead to not terminating iteration |
5265 | (see PR95049). Since there's no convenient way to check |
5266 | for the allowed transition of VAL -> PHI (loop entry value, |
5267 | same on two PHIs, to same PHI result) we restrict the check |
5268 | to invariants. */ |
5269 | && curr_invariant |
5270 | && is_gimple_min_invariant (to)) |
5271 | { |
5272 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5273 | fprintf (stream: dump_file, format: " forced VARYING" ); |
5274 | to = from; |
5275 | } |
5276 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5277 | fprintf (stream: dump_file, format: " (changed)\n" ); |
5278 | from_info->valnum = to; |
5279 | return true; |
5280 | } |
5281 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5282 | fprintf (stream: dump_file, format: "\n" ); |
5283 | return false; |
5284 | } |
5285 | |
5286 | /* Set all definitions in STMT to value number to themselves. |
5287 | Return true if a value number changed. */ |
5288 | |
5289 | static bool |
5290 | defs_to_varying (gimple *stmt) |
5291 | { |
5292 | bool changed = false; |
5293 | ssa_op_iter iter; |
5294 | def_operand_p defp; |
5295 | |
5296 | FOR_EACH_SSA_DEF_OPERAND (defp, stmt, iter, SSA_OP_ALL_DEFS) |
5297 | { |
5298 | tree def = DEF_FROM_PTR (defp); |
5299 | changed |= set_ssa_val_to (from: def, to: def); |
5300 | } |
5301 | return changed; |
5302 | } |
5303 | |
5304 | /* Visit a copy between LHS and RHS, return true if the value number |
5305 | changed. */ |
5306 | |
5307 | static bool |
5308 | visit_copy (tree lhs, tree rhs) |
5309 | { |
5310 | /* Valueize. */ |
5311 | rhs = SSA_VAL (x: rhs); |
5312 | |
5313 | return set_ssa_val_to (from: lhs, to: rhs); |
5314 | } |
5315 | |
5316 | /* Lookup a value for OP in type WIDE_TYPE where the value in type of OP |
5317 | is the same. */ |
5318 | |
5319 | static tree |
5320 | valueized_wider_op (tree wide_type, tree op, bool allow_truncate) |
5321 | { |
5322 | if (TREE_CODE (op) == SSA_NAME) |
5323 | op = vn_valueize (op); |
5324 | |
5325 | /* Either we have the op widened available. */ |
5326 | tree ops[3] = {}; |
5327 | ops[0] = op; |
5328 | tree tem = vn_nary_op_lookup_pieces (length: 1, code: NOP_EXPR, |
5329 | type: wide_type, ops, NULL); |
5330 | if (tem) |
5331 | return tem; |
5332 | |
5333 | /* Or the op is truncated from some existing value. */ |
5334 | if (allow_truncate && TREE_CODE (op) == SSA_NAME) |
5335 | { |
5336 | gimple *def = SSA_NAME_DEF_STMT (op); |
5337 | if (is_gimple_assign (gs: def) |
5338 | && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def))) |
5339 | { |
5340 | tem = gimple_assign_rhs1 (gs: def); |
5341 | if (useless_type_conversion_p (wide_type, TREE_TYPE (tem))) |
5342 | { |
5343 | if (TREE_CODE (tem) == SSA_NAME) |
5344 | tem = vn_valueize (tem); |
5345 | return tem; |
5346 | } |
5347 | } |
5348 | } |
5349 | |
5350 | /* For constants simply extend it. */ |
5351 | if (TREE_CODE (op) == INTEGER_CST) |
5352 | return wide_int_to_tree (type: wide_type, cst: wi::to_widest (t: op)); |
5353 | |
5354 | return NULL_TREE; |
5355 | } |
5356 | |
5357 | /* Visit a nary operator RHS, value number it, and return true if the |
5358 | value number of LHS has changed as a result. */ |
5359 | |
5360 | static bool |
5361 | visit_nary_op (tree lhs, gassign *stmt) |
5362 | { |
5363 | vn_nary_op_t vnresult; |
5364 | tree result = vn_nary_op_lookup_stmt (stmt, vnresult: &vnresult); |
5365 | if (! result && vnresult) |
5366 | result = vn_nary_op_get_predicated_value (vno: vnresult, bb: gimple_bb (g: stmt)); |
5367 | if (result) |
5368 | return set_ssa_val_to (from: lhs, to: result); |
5369 | |
5370 | /* Do some special pattern matching for redundancies of operations |
5371 | in different types. */ |
5372 | enum tree_code code = gimple_assign_rhs_code (gs: stmt); |
5373 | tree type = TREE_TYPE (lhs); |
5374 | tree rhs1 = gimple_assign_rhs1 (gs: stmt); |
5375 | switch (code) |
5376 | { |
5377 | CASE_CONVERT: |
5378 | /* Match arithmetic done in a different type where we can easily |
5379 | substitute the result from some earlier sign-changed or widened |
5380 | operation. */ |
5381 | if (INTEGRAL_TYPE_P (type) |
5382 | && TREE_CODE (rhs1) == SSA_NAME |
5383 | /* We only handle sign-changes, zero-extension -> & mask or |
5384 | sign-extension if we know the inner operation doesn't |
5385 | overflow. */ |
5386 | && (((TYPE_UNSIGNED (TREE_TYPE (rhs1)) |
5387 | || (INTEGRAL_TYPE_P (TREE_TYPE (rhs1)) |
5388 | && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (rhs1)))) |
5389 | && TYPE_PRECISION (type) > TYPE_PRECISION (TREE_TYPE (rhs1))) |
5390 | || TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (rhs1)))) |
5391 | { |
5392 | gassign *def = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (rhs1)); |
5393 | if (def |
5394 | && (gimple_assign_rhs_code (gs: def) == PLUS_EXPR |
5395 | || gimple_assign_rhs_code (gs: def) == MINUS_EXPR |
5396 | || gimple_assign_rhs_code (gs: def) == MULT_EXPR)) |
5397 | { |
5398 | tree ops[3] = {}; |
5399 | /* When requiring a sign-extension we cannot model a |
5400 | previous truncation with a single op so don't bother. */ |
5401 | bool allow_truncate = TYPE_UNSIGNED (TREE_TYPE (rhs1)); |
5402 | /* Either we have the op widened available. */ |
5403 | ops[0] = valueized_wider_op (wide_type: type, op: gimple_assign_rhs1 (gs: def), |
5404 | allow_truncate); |
5405 | if (ops[0]) |
5406 | ops[1] = valueized_wider_op (wide_type: type, op: gimple_assign_rhs2 (gs: def), |
5407 | allow_truncate); |
5408 | if (ops[0] && ops[1]) |
5409 | { |
5410 | ops[0] = vn_nary_op_lookup_pieces |
5411 | (length: 2, code: gimple_assign_rhs_code (gs: def), type, ops, NULL); |
5412 | /* We have wider operation available. */ |
5413 | if (ops[0] |
5414 | /* If the leader is a wrapping operation we can |
5415 | insert it for code hoisting w/o introducing |
5416 | undefined overflow. If it is not it has to |
5417 | be available. See PR86554. */ |
5418 | && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (ops[0])) |
5419 | || (rpo_avail && vn_context_bb |
5420 | && rpo_avail->eliminate_avail (vn_context_bb, |
5421 | op: ops[0])))) |
5422 | { |
5423 | unsigned lhs_prec = TYPE_PRECISION (type); |
5424 | unsigned rhs_prec = TYPE_PRECISION (TREE_TYPE (rhs1)); |
5425 | if (lhs_prec == rhs_prec |
5426 | || (INTEGRAL_TYPE_P (TREE_TYPE (rhs1)) |
5427 | && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (rhs1)))) |
5428 | { |
5429 | gimple_match_op match_op (gimple_match_cond::UNCOND, |
5430 | NOP_EXPR, type, ops[0]); |
5431 | result = vn_nary_build_or_lookup (res_op: &match_op); |
5432 | if (result) |
5433 | { |
5434 | bool changed = set_ssa_val_to (from: lhs, to: result); |
5435 | vn_nary_op_insert_stmt (stmt, result); |
5436 | return changed; |
5437 | } |
5438 | } |
5439 | else |
5440 | { |
5441 | tree mask = wide_int_to_tree |
5442 | (type, cst: wi::mask (width: rhs_prec, negate_p: false, precision: lhs_prec)); |
5443 | gimple_match_op match_op (gimple_match_cond::UNCOND, |
5444 | BIT_AND_EXPR, |
5445 | TREE_TYPE (lhs), |
5446 | ops[0], mask); |
5447 | result = vn_nary_build_or_lookup (res_op: &match_op); |
5448 | if (result) |
5449 | { |
5450 | bool changed = set_ssa_val_to (from: lhs, to: result); |
5451 | vn_nary_op_insert_stmt (stmt, result); |
5452 | return changed; |
5453 | } |
5454 | } |
5455 | } |
5456 | } |
5457 | } |
5458 | } |
5459 | break; |
5460 | case BIT_AND_EXPR: |
5461 | if (INTEGRAL_TYPE_P (type) |
5462 | && TREE_CODE (rhs1) == SSA_NAME |
5463 | && TREE_CODE (gimple_assign_rhs2 (stmt)) == INTEGER_CST |
5464 | && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1) |
5465 | && default_vn_walk_kind != VN_NOWALK |
5466 | && CHAR_BIT == 8 |
5467 | && BITS_PER_UNIT == 8 |
5468 | && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN |
5469 | && TYPE_PRECISION (type) <= vn_walk_cb_data::bufsize * BITS_PER_UNIT |
5470 | && !integer_all_onesp (gimple_assign_rhs2 (gs: stmt)) |
5471 | && !integer_zerop (gimple_assign_rhs2 (gs: stmt))) |
5472 | { |
5473 | gassign *ass = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (rhs1)); |
5474 | if (ass |
5475 | && !gimple_has_volatile_ops (stmt: ass) |
5476 | && vn_get_stmt_kind (stmt: ass) == VN_REFERENCE) |
5477 | { |
5478 | tree last_vuse = gimple_vuse (g: ass); |
5479 | tree op = gimple_assign_rhs1 (gs: ass); |
5480 | tree result = vn_reference_lookup (op, vuse: gimple_vuse (g: ass), |
5481 | kind: default_vn_walk_kind, |
5482 | NULL, tbaa_p: true, last_vuse_ptr: &last_vuse, |
5483 | mask: gimple_assign_rhs2 (gs: stmt)); |
5484 | if (result |
5485 | && useless_type_conversion_p (TREE_TYPE (result), |
5486 | TREE_TYPE (op))) |
5487 | return set_ssa_val_to (from: lhs, to: result); |
5488 | } |
5489 | } |
5490 | break; |
5491 | case TRUNC_DIV_EXPR: |
5492 | if (TYPE_UNSIGNED (type)) |
5493 | break; |
5494 | /* Fallthru. */ |
5495 | case RDIV_EXPR: |
5496 | case MULT_EXPR: |
5497 | /* Match up ([-]a){/,*}([-])b with v=a{/,*}b, replacing it with -v. */ |
5498 | if (! HONOR_SIGN_DEPENDENT_ROUNDING (type)) |
5499 | { |
5500 | tree rhs[2]; |
5501 | rhs[0] = rhs1; |
5502 | rhs[1] = gimple_assign_rhs2 (gs: stmt); |
5503 | for (unsigned i = 0; i <= 1; ++i) |
5504 | { |
5505 | unsigned j = i == 0 ? 1 : 0; |
5506 | tree ops[2]; |
5507 | gimple_match_op match_op (gimple_match_cond::UNCOND, |
5508 | NEGATE_EXPR, type, rhs[i]); |
5509 | ops[i] = vn_nary_build_or_lookup_1 (res_op: &match_op, insert: false, simplify: true); |
5510 | ops[j] = rhs[j]; |
5511 | if (ops[i] |
5512 | && (ops[0] = vn_nary_op_lookup_pieces (length: 2, code, |
5513 | type, ops, NULL))) |
5514 | { |
5515 | gimple_match_op match_op (gimple_match_cond::UNCOND, |
5516 | NEGATE_EXPR, type, ops[0]); |
5517 | result = vn_nary_build_or_lookup_1 (res_op: &match_op, insert: true, simplify: false); |
5518 | if (result) |
5519 | { |
5520 | bool changed = set_ssa_val_to (from: lhs, to: result); |
5521 | vn_nary_op_insert_stmt (stmt, result); |
5522 | return changed; |
5523 | } |
5524 | } |
5525 | } |
5526 | } |
5527 | break; |
5528 | case LSHIFT_EXPR: |
5529 | /* For X << C, use the value number of X * (1 << C). */ |
5530 | if (INTEGRAL_TYPE_P (type) |
5531 | && TYPE_OVERFLOW_WRAPS (type) |
5532 | && !TYPE_SATURATING (type)) |
5533 | { |
5534 | tree rhs2 = gimple_assign_rhs2 (gs: stmt); |
5535 | if (TREE_CODE (rhs2) == INTEGER_CST |
5536 | && tree_fits_uhwi_p (rhs2) |
5537 | && tree_to_uhwi (rhs2) < TYPE_PRECISION (type)) |
5538 | { |
5539 | wide_int w = wi::set_bit_in_zero (bit: tree_to_uhwi (rhs2), |
5540 | TYPE_PRECISION (type)); |
5541 | gimple_match_op match_op (gimple_match_cond::UNCOND, |
5542 | MULT_EXPR, type, rhs1, |
5543 | wide_int_to_tree (type, cst: w)); |
5544 | result = vn_nary_build_or_lookup (res_op: &match_op); |
5545 | if (result) |
5546 | { |
5547 | bool changed = set_ssa_val_to (from: lhs, to: result); |
5548 | if (TREE_CODE (result) == SSA_NAME) |
5549 | vn_nary_op_insert_stmt (stmt, result); |
5550 | return changed; |
5551 | } |
5552 | } |
5553 | } |
5554 | break; |
5555 | default: |
5556 | break; |
5557 | } |
5558 | |
5559 | bool changed = set_ssa_val_to (from: lhs, to: lhs); |
5560 | vn_nary_op_insert_stmt (stmt, result: lhs); |
5561 | return changed; |
5562 | } |
5563 | |
5564 | /* Visit a call STMT storing into LHS. Return true if the value number |
5565 | of the LHS has changed as a result. */ |
5566 | |
5567 | static bool |
5568 | visit_reference_op_call (tree lhs, gcall *stmt) |
5569 | { |
5570 | bool changed = false; |
5571 | struct vn_reference_s vr1; |
5572 | vn_reference_t vnresult = NULL; |
5573 | tree vdef = gimple_vdef (g: stmt); |
5574 | modref_summary *summary; |
5575 | |
5576 | /* Non-ssa lhs is handled in copy_reference_ops_from_call. */ |
5577 | if (lhs && TREE_CODE (lhs) != SSA_NAME) |
5578 | lhs = NULL_TREE; |
5579 | |
5580 | vn_reference_lookup_call (call: stmt, vnresult: &vnresult, vr: &vr1); |
5581 | |
5582 | /* If the lookup did not succeed for pure functions try to use |
5583 | modref info to find a candidate to CSE to. */ |
5584 | const unsigned accesses_limit = 8; |
5585 | if (!vnresult |
5586 | && !vdef |
5587 | && lhs |
5588 | && gimple_vuse (g: stmt) |
5589 | && (((summary = get_modref_function_summary (call: stmt, NULL)) |
5590 | && !summary->global_memory_read |
5591 | && summary->load_accesses < accesses_limit) |
5592 | || gimple_call_flags (stmt) & ECF_CONST)) |
5593 | { |
5594 | /* First search if we can do someting useful and build a |
5595 | vector of all loads we have to check. */ |
5596 | bool unknown_memory_access = false; |
5597 | auto_vec<ao_ref, accesses_limit> accesses; |
5598 | unsigned load_accesses = summary ? summary->load_accesses : 0; |
5599 | if (!unknown_memory_access) |
5600 | /* Add loads done as part of setting up the call arguments. |
5601 | That's also necessary for CONST functions which will |
5602 | not have a modref summary. */ |
5603 | for (unsigned i = 0; i < gimple_call_num_args (gs: stmt); ++i) |
5604 | { |
5605 | tree arg = gimple_call_arg (gs: stmt, index: i); |
5606 | if (TREE_CODE (arg) != SSA_NAME |
5607 | && !is_gimple_min_invariant (arg)) |
5608 | { |
5609 | if (accesses.length () >= accesses_limit - load_accesses) |
5610 | { |
5611 | unknown_memory_access = true; |
5612 | break; |
5613 | } |
5614 | accesses.quick_grow (len: accesses.length () + 1); |
5615 | ao_ref_init (&accesses.last (), arg); |
5616 | } |
5617 | } |
5618 | if (summary && !unknown_memory_access) |
5619 | { |
5620 | /* Add loads as analyzed by IPA modref. */ |
5621 | for (auto base_node : summary->loads->bases) |
5622 | if (unknown_memory_access) |
5623 | break; |
5624 | else for (auto ref_node : base_node->refs) |
5625 | if (unknown_memory_access) |
5626 | break; |
5627 | else for (auto access_node : ref_node->accesses) |
5628 | { |
5629 | accesses.quick_grow (len: accesses.length () + 1); |
5630 | ao_ref *r = &accesses.last (); |
5631 | if (!access_node.get_ao_ref (stmt, ref: r)) |
5632 | { |
5633 | /* Initialize a ref based on the argument and |
5634 | unknown offset if possible. */ |
5635 | tree arg = access_node.get_call_arg (stmt); |
5636 | if (arg && TREE_CODE (arg) == SSA_NAME) |
5637 | arg = SSA_VAL (x: arg); |
5638 | if (arg |
5639 | && TREE_CODE (arg) == ADDR_EXPR |
5640 | && (arg = get_base_address (t: arg)) |
5641 | && DECL_P (arg)) |
5642 | { |
5643 | ao_ref_init (r, arg); |
5644 | r->ref = NULL_TREE; |
5645 | r->base = arg; |
5646 | } |
5647 | else |
5648 | { |
5649 | unknown_memory_access = true; |
5650 | break; |
5651 | } |
5652 | } |
5653 | r->base_alias_set = base_node->base; |
5654 | r->ref_alias_set = ref_node->ref; |
5655 | } |
5656 | } |
5657 | |
5658 | /* Walk the VUSE->VDEF chain optimistically trying to find an entry |
5659 | for the call in the hashtable. */ |
5660 | unsigned limit = (unknown_memory_access |
5661 | ? 0 |
5662 | : (param_sccvn_max_alias_queries_per_access |
5663 | / (accesses.length () + 1))); |
5664 | tree saved_vuse = vr1.vuse; |
5665 | hashval_t saved_hashcode = vr1.hashcode; |
5666 | while (limit > 0 && !vnresult && !SSA_NAME_IS_DEFAULT_DEF (vr1.vuse)) |
5667 | { |
5668 | vr1.hashcode = vr1.hashcode - SSA_NAME_VERSION (vr1.vuse); |
5669 | gimple *def = SSA_NAME_DEF_STMT (vr1.vuse); |
5670 | /* ??? We could use fancy stuff like in walk_non_aliased_vuses, but |
5671 | do not bother for now. */ |
5672 | if (is_a <gphi *> (p: def)) |
5673 | break; |
5674 | vr1.vuse = vuse_ssa_val (x: gimple_vuse (g: def)); |
5675 | vr1.hashcode = vr1.hashcode + SSA_NAME_VERSION (vr1.vuse); |
5676 | vn_reference_lookup_1 (vr: &vr1, vnresult: &vnresult); |
5677 | limit--; |
5678 | } |
5679 | |
5680 | /* If we found a candidate to CSE to verify it is valid. */ |
5681 | if (vnresult && !accesses.is_empty ()) |
5682 | { |
5683 | tree vuse = vuse_ssa_val (x: gimple_vuse (g: stmt)); |
5684 | while (vnresult && vuse != vr1.vuse) |
5685 | { |
5686 | gimple *def = SSA_NAME_DEF_STMT (vuse); |
5687 | for (auto &ref : accesses) |
5688 | { |
5689 | /* ??? stmt_may_clobber_ref_p_1 does per stmt constant |
5690 | analysis overhead that we might be able to cache. */ |
5691 | if (stmt_may_clobber_ref_p_1 (def, &ref, true)) |
5692 | { |
5693 | vnresult = NULL; |
5694 | break; |
5695 | } |
5696 | } |
5697 | vuse = vuse_ssa_val (x: gimple_vuse (g: def)); |
5698 | } |
5699 | } |
5700 | vr1.vuse = saved_vuse; |
5701 | vr1.hashcode = saved_hashcode; |
5702 | } |
5703 | |
5704 | if (vnresult) |
5705 | { |
5706 | if (vdef) |
5707 | { |
5708 | if (vnresult->result_vdef) |
5709 | changed |= set_ssa_val_to (from: vdef, to: vnresult->result_vdef); |
5710 | else if (!lhs && gimple_call_lhs (gs: stmt)) |
5711 | /* If stmt has non-SSA_NAME lhs, value number the vdef to itself, |
5712 | as the call still acts as a lhs store. */ |
5713 | changed |= set_ssa_val_to (from: vdef, to: vdef); |
5714 | else |
5715 | /* If the call was discovered to be pure or const reflect |
5716 | that as far as possible. */ |
5717 | changed |= set_ssa_val_to (from: vdef, |
5718 | to: vuse_ssa_val (x: gimple_vuse (g: stmt))); |
5719 | } |
5720 | |
5721 | if (!vnresult->result && lhs) |
5722 | vnresult->result = lhs; |
5723 | |
5724 | if (vnresult->result && lhs) |
5725 | changed |= set_ssa_val_to (from: lhs, to: vnresult->result); |
5726 | } |
5727 | else |
5728 | { |
5729 | vn_reference_t vr2; |
5730 | vn_reference_s **slot; |
5731 | tree vdef_val = vdef; |
5732 | if (vdef) |
5733 | { |
5734 | /* If we value numbered an indirect functions function to |
5735 | one not clobbering memory value number its VDEF to its |
5736 | VUSE. */ |
5737 | tree fn = gimple_call_fn (gs: stmt); |
5738 | if (fn && TREE_CODE (fn) == SSA_NAME) |
5739 | { |
5740 | fn = SSA_VAL (x: fn); |
5741 | if (TREE_CODE (fn) == ADDR_EXPR |
5742 | && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL |
5743 | && (flags_from_decl_or_type (TREE_OPERAND (fn, 0)) |
5744 | & (ECF_CONST | ECF_PURE)) |
5745 | /* If stmt has non-SSA_NAME lhs, value number the |
5746 | vdef to itself, as the call still acts as a lhs |
5747 | store. */ |
5748 | && (lhs || gimple_call_lhs (gs: stmt) == NULL_TREE)) |
5749 | vdef_val = vuse_ssa_val (x: gimple_vuse (g: stmt)); |
5750 | } |
5751 | changed |= set_ssa_val_to (from: vdef, to: vdef_val); |
5752 | } |
5753 | if (lhs) |
5754 | changed |= set_ssa_val_to (from: lhs, to: lhs); |
5755 | vr2 = XOBNEW (&vn_tables_obstack, vn_reference_s); |
5756 | vr2->vuse = vr1.vuse; |
5757 | /* As we are not walking the virtual operand chain we know the |
5758 | shared_lookup_references are still original so we can re-use |
5759 | them here. */ |
5760 | vr2->operands = vr1.operands.copy (); |
5761 | vr2->type = vr1.type; |
5762 | vr2->punned = vr1.punned; |
5763 | vr2->set = vr1.set; |
5764 | vr2->offset = vr1.offset; |
5765 | vr2->max_size = vr1.max_size; |
5766 | vr2->base_set = vr1.base_set; |
5767 | vr2->hashcode = vr1.hashcode; |
5768 | vr2->result = lhs; |
5769 | vr2->result_vdef = vdef_val; |
5770 | vr2->value_id = 0; |
5771 | slot = valid_info->references->find_slot_with_hash (comparable: vr2, hash: vr2->hashcode, |
5772 | insert: INSERT); |
5773 | gcc_assert (!*slot); |
5774 | *slot = vr2; |
5775 | vr2->next = last_inserted_ref; |
5776 | last_inserted_ref = vr2; |
5777 | } |
5778 | |
5779 | return changed; |
5780 | } |
5781 | |
5782 | /* Visit a load from a reference operator RHS, part of STMT, value number it, |
5783 | and return true if the value number of the LHS has changed as a result. */ |
5784 | |
5785 | static bool |
5786 | visit_reference_op_load (tree lhs, tree op, gimple *stmt) |
5787 | { |
5788 | bool changed = false; |
5789 | tree result; |
5790 | vn_reference_t res; |
5791 | |
5792 | tree vuse = gimple_vuse (g: stmt); |
5793 | tree last_vuse = vuse; |
5794 | result = vn_reference_lookup (op, vuse, kind: default_vn_walk_kind, vnresult: &res, tbaa_p: true, last_vuse_ptr: &last_vuse); |
5795 | |
5796 | /* We handle type-punning through unions by value-numbering based |
5797 | on offset and size of the access. Be prepared to handle a |
5798 | type-mismatch here via creating a VIEW_CONVERT_EXPR. */ |
5799 | if (result |
5800 | && !useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (op))) |
5801 | { |
5802 | /* Avoid the type punning in case the result mode has padding where |
5803 | the op we lookup has not. */ |
5804 | if (TYPE_MODE (TREE_TYPE (result)) != BLKmode |
5805 | && maybe_lt (a: GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (result))), |
5806 | b: GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (op))))) |
5807 | result = NULL_TREE; |
5808 | else if (CONSTANT_CLASS_P (result)) |
5809 | result = const_unop (VIEW_CONVERT_EXPR, TREE_TYPE (op), result); |
5810 | else |
5811 | { |
5812 | /* We will be setting the value number of lhs to the value number |
5813 | of VIEW_CONVERT_EXPR <TREE_TYPE (result)> (result). |
5814 | So first simplify and lookup this expression to see if it |
5815 | is already available. */ |
5816 | gimple_match_op res_op (gimple_match_cond::UNCOND, |
5817 | VIEW_CONVERT_EXPR, TREE_TYPE (op), result); |
5818 | result = vn_nary_build_or_lookup (res_op: &res_op); |
5819 | if (result |
5820 | && TREE_CODE (result) == SSA_NAME |
5821 | && VN_INFO (name: result)->needs_insertion) |
5822 | /* Track whether this is the canonical expression for different |
5823 | typed loads. We use that as a stopgap measure for code |
5824 | hoisting when dealing with floating point loads. */ |
5825 | res->punned = true; |
5826 | } |
5827 | |
5828 | /* When building the conversion fails avoid inserting the reference |
5829 | again. */ |
5830 | if (!result) |
5831 | return set_ssa_val_to (from: lhs, to: lhs); |
5832 | } |
5833 | |
5834 | if (result) |
5835 | changed = set_ssa_val_to (from: lhs, to: result); |
5836 | else |
5837 | { |
5838 | changed = set_ssa_val_to (from: lhs, to: lhs); |
5839 | vn_reference_insert (op, result: lhs, vuse: last_vuse, NULL_TREE); |
5840 | if (vuse && SSA_VAL (x: last_vuse) != SSA_VAL (x: vuse)) |
5841 | { |
5842 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5843 | { |
5844 | fprintf (stream: dump_file, format: "Using extra use virtual operand " ); |
5845 | print_generic_expr (dump_file, last_vuse); |
5846 | fprintf (stream: dump_file, format: "\n" ); |
5847 | } |
5848 | vn_reference_insert (op, result: lhs, vuse, NULL_TREE); |
5849 | } |
5850 | } |
5851 | |
5852 | return changed; |
5853 | } |
5854 | |
5855 | |
5856 | /* Visit a store to a reference operator LHS, part of STMT, value number it, |
5857 | and return true if the value number of the LHS has changed as a result. */ |
5858 | |
5859 | static bool |
5860 | visit_reference_op_store (tree lhs, tree op, gimple *stmt) |
5861 | { |
5862 | bool changed = false; |
5863 | vn_reference_t vnresult = NULL; |
5864 | tree assign; |
5865 | bool resultsame = false; |
5866 | tree vuse = gimple_vuse (g: stmt); |
5867 | tree vdef = gimple_vdef (g: stmt); |
5868 | |
5869 | if (TREE_CODE (op) == SSA_NAME) |
5870 | op = SSA_VAL (x: op); |
5871 | |
5872 | /* First we want to lookup using the *vuses* from the store and see |
5873 | if there the last store to this location with the same address |
5874 | had the same value. |
5875 | |
5876 | The vuses represent the memory state before the store. If the |
5877 | memory state, address, and value of the store is the same as the |
5878 | last store to this location, then this store will produce the |
5879 | same memory state as that store. |
5880 | |
5881 | In this case the vdef versions for this store are value numbered to those |
5882 | vuse versions, since they represent the same memory state after |
5883 | this store. |
5884 | |
5885 | Otherwise, the vdefs for the store are used when inserting into |
5886 | the table, since the store generates a new memory state. */ |
5887 | |
5888 | vn_reference_lookup (op: lhs, vuse, kind: VN_NOWALK, vnresult: &vnresult, tbaa_p: false); |
5889 | if (vnresult |
5890 | && vnresult->result) |
5891 | { |
5892 | tree result = vnresult->result; |
5893 | gcc_checking_assert (TREE_CODE (result) != SSA_NAME |
5894 | || result == SSA_VAL (result)); |
5895 | resultsame = expressions_equal_p (result, op); |
5896 | if (resultsame) |
5897 | { |
5898 | /* If the TBAA state isn't compatible for downstream reads |
5899 | we cannot value-number the VDEFs the same. */ |
5900 | ao_ref lhs_ref; |
5901 | ao_ref_init (&lhs_ref, lhs); |
5902 | alias_set_type set = ao_ref_alias_set (&lhs_ref); |
5903 | alias_set_type base_set = ao_ref_base_alias_set (&lhs_ref); |
5904 | if ((vnresult->set != set |
5905 | && ! alias_set_subset_of (set, vnresult->set)) |
5906 | || (vnresult->base_set != base_set |
5907 | && ! alias_set_subset_of (base_set, vnresult->base_set))) |
5908 | resultsame = false; |
5909 | } |
5910 | } |
5911 | |
5912 | if (!resultsame) |
5913 | { |
5914 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5915 | { |
5916 | fprintf (stream: dump_file, format: "No store match\n" ); |
5917 | fprintf (stream: dump_file, format: "Value numbering store " ); |
5918 | print_generic_expr (dump_file, lhs); |
5919 | fprintf (stream: dump_file, format: " to " ); |
5920 | print_generic_expr (dump_file, op); |
5921 | fprintf (stream: dump_file, format: "\n" ); |
5922 | } |
5923 | /* Have to set value numbers before insert, since insert is |
5924 | going to valueize the references in-place. */ |
5925 | if (vdef) |
5926 | changed |= set_ssa_val_to (from: vdef, to: vdef); |
5927 | |
5928 | /* Do not insert structure copies into the tables. */ |
5929 | if (is_gimple_min_invariant (op) |
5930 | || is_gimple_reg (op)) |
5931 | vn_reference_insert (op: lhs, result: op, vuse: vdef, NULL); |
5932 | |
5933 | /* Only perform the following when being called from PRE |
5934 | which embeds tail merging. */ |
5935 | if (default_vn_walk_kind == VN_WALK) |
5936 | { |
5937 | assign = build2 (MODIFY_EXPR, TREE_TYPE (lhs), lhs, op); |
5938 | vn_reference_lookup (op: assign, vuse, kind: VN_NOWALK, vnresult: &vnresult, tbaa_p: false); |
5939 | if (!vnresult) |
5940 | vn_reference_insert (op: assign, result: lhs, vuse, vdef); |
5941 | } |
5942 | } |
5943 | else |
5944 | { |
5945 | /* We had a match, so value number the vdef to have the value |
5946 | number of the vuse it came from. */ |
5947 | |
5948 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5949 | fprintf (stream: dump_file, format: "Store matched earlier value, " |
5950 | "value numbering store vdefs to matching vuses.\n" ); |
5951 | |
5952 | changed |= set_ssa_val_to (from: vdef, to: SSA_VAL (x: vuse)); |
5953 | } |
5954 | |
5955 | return changed; |
5956 | } |
5957 | |
5958 | /* Visit and value number PHI, return true if the value number |
5959 | changed. When BACKEDGES_VARYING_P is true then assume all |
5960 | backedge values are varying. When INSERTED is not NULL then |
5961 | this is just a ahead query for a possible iteration, set INSERTED |
5962 | to true if we'd insert into the hashtable. */ |
5963 | |
5964 | static bool |
5965 | visit_phi (gimple *phi, bool *inserted, bool backedges_varying_p) |
5966 | { |
5967 | tree result, sameval = VN_TOP, seen_undef = NULL_TREE; |
5968 | bool seen_undef_visited = false; |
5969 | tree backedge_val = NULL_TREE; |
5970 | bool seen_non_backedge = false; |
5971 | tree sameval_base = NULL_TREE; |
5972 | poly_int64 soff, doff; |
5973 | unsigned n_executable = 0; |
5974 | edge_iterator ei; |
5975 | edge e, sameval_e = NULL; |
5976 | |
5977 | /* TODO: We could check for this in initialization, and replace this |
5978 | with a gcc_assert. */ |
5979 | if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (phi))) |
5980 | return set_ssa_val_to (PHI_RESULT (phi), PHI_RESULT (phi)); |
5981 | |
5982 | /* We track whether a PHI was CSEd to avoid excessive iterations |
5983 | that would be necessary only because the PHI changed arguments |
5984 | but not value. */ |
5985 | if (!inserted) |
5986 | gimple_set_plf (stmt: phi, plf: GF_PLF_1, val_p: false); |
5987 | |
5988 | /* See if all non-TOP arguments have the same value. TOP is |
5989 | equivalent to everything, so we can ignore it. */ |
5990 | basic_block bb = gimple_bb (g: phi); |
5991 | FOR_EACH_EDGE (e, ei, bb->preds) |
5992 | if (e->flags & EDGE_EXECUTABLE) |
5993 | { |
5994 | tree def = PHI_ARG_DEF_FROM_EDGE (phi, e); |
5995 | |
5996 | if (def == PHI_RESULT (phi)) |
5997 | continue; |
5998 | ++n_executable; |
5999 | bool visited = true; |
6000 | if (TREE_CODE (def) == SSA_NAME) |
6001 | { |
6002 | tree val = SSA_VAL (x: def, visited: &visited); |
6003 | if (SSA_NAME_IS_DEFAULT_DEF (def)) |
6004 | visited = true; |
6005 | if (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK)) |
6006 | def = val; |
6007 | if (e->flags & EDGE_DFS_BACK) |
6008 | backedge_val = def; |
6009 | } |
6010 | if (!(e->flags & EDGE_DFS_BACK)) |
6011 | seen_non_backedge = true; |
6012 | if (def == VN_TOP) |
6013 | ; |
6014 | /* Ignore undefined defs for sameval but record one. */ |
6015 | else if (TREE_CODE (def) == SSA_NAME |
6016 | && ! virtual_operand_p (op: def) |
6017 | && ssa_undefined_value_p (def, false)) |
6018 | { |
6019 | if (!seen_undef |
6020 | /* Avoid having not visited undefined defs if we also have |
6021 | a visited one. */ |
6022 | || (!seen_undef_visited && visited)) |
6023 | { |
6024 | seen_undef = def; |
6025 | seen_undef_visited = visited; |
6026 | } |
6027 | } |
6028 | else if (sameval == VN_TOP) |
6029 | { |
6030 | sameval = def; |
6031 | sameval_e = e; |
6032 | } |
6033 | else if (expressions_equal_p (def, sameval)) |
6034 | sameval_e = NULL; |
6035 | else if (virtual_operand_p (op: def)) |
6036 | { |
6037 | sameval = NULL_TREE; |
6038 | break; |
6039 | } |
6040 | else |
6041 | { |
6042 | /* We know we're arriving only with invariant addresses here, |
6043 | try harder comparing them. We can do some caching here |
6044 | which we cannot do in expressions_equal_p. */ |
6045 | if (TREE_CODE (def) == ADDR_EXPR |
6046 | && TREE_CODE (sameval) == ADDR_EXPR |
6047 | && sameval_base != (void *)-1) |
6048 | { |
6049 | if (!sameval_base) |
6050 | sameval_base = get_addr_base_and_unit_offset |
6051 | (TREE_OPERAND (sameval, 0), &soff); |
6052 | if (!sameval_base) |
6053 | sameval_base = (tree)(void *)-1; |
6054 | else if ((get_addr_base_and_unit_offset |
6055 | (TREE_OPERAND (def, 0), &doff) == sameval_base) |
6056 | && known_eq (soff, doff)) |
6057 | continue; |
6058 | } |
6059 | /* There's also the possibility to use equivalences. */ |
6060 | if (!FLOAT_TYPE_P (TREE_TYPE (def)) |
6061 | /* But only do this if we didn't force any of sameval or |
6062 | val to VARYING because of backedge processing rules. */ |
6063 | && (TREE_CODE (sameval) != SSA_NAME |
6064 | || SSA_VAL (x: sameval) == sameval) |
6065 | && (TREE_CODE (def) != SSA_NAME || SSA_VAL (x: def) == def)) |
6066 | { |
6067 | vn_nary_op_t vnresult; |
6068 | tree ops[2]; |
6069 | ops[0] = def; |
6070 | ops[1] = sameval; |
6071 | tree val = vn_nary_op_lookup_pieces (length: 2, code: EQ_EXPR, |
6072 | boolean_type_node, |
6073 | ops, vnresult: &vnresult); |
6074 | if (! val && vnresult && vnresult->predicated_values) |
6075 | { |
6076 | val = vn_nary_op_get_predicated_value (vno: vnresult, e); |
6077 | if (val && integer_truep (val) |
6078 | && !(sameval_e && (sameval_e->flags & EDGE_DFS_BACK))) |
6079 | { |
6080 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6081 | { |
6082 | fprintf (stream: dump_file, format: "Predication says " ); |
6083 | print_generic_expr (dump_file, def, TDF_NONE); |
6084 | fprintf (stream: dump_file, format: " and " ); |
6085 | print_generic_expr (dump_file, sameval, TDF_NONE); |
6086 | fprintf (stream: dump_file, format: " are equal on edge %d -> %d\n" , |
6087 | e->src->index, e->dest->index); |
6088 | } |
6089 | continue; |
6090 | } |
6091 | /* If on all previous edges the value was equal to def |
6092 | we can change sameval to def. */ |
6093 | if (EDGE_COUNT (bb->preds) == 2 |
6094 | && (val = vn_nary_op_get_predicated_value |
6095 | (vno: vnresult, EDGE_PRED (bb, 0))) |
6096 | && integer_truep (val) |
6097 | && !(e->flags & EDGE_DFS_BACK)) |
6098 | { |
6099 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6100 | { |
6101 | fprintf (stream: dump_file, format: "Predication says " ); |
6102 | print_generic_expr (dump_file, def, TDF_NONE); |
6103 | fprintf (stream: dump_file, format: " and " ); |
6104 | print_generic_expr (dump_file, sameval, TDF_NONE); |
6105 | fprintf (stream: dump_file, format: " are equal on edge %d -> %d\n" , |
6106 | EDGE_PRED (bb, 0)->src->index, |
6107 | EDGE_PRED (bb, 0)->dest->index); |
6108 | } |
6109 | sameval = def; |
6110 | continue; |
6111 | } |
6112 | } |
6113 | } |
6114 | sameval = NULL_TREE; |
6115 | break; |
6116 | } |
6117 | } |
6118 | |
6119 | /* If the value we want to use is flowing over the backedge and we |
6120 | should take it as VARYING but it has a non-VARYING value drop to |
6121 | VARYING. |
6122 | If we value-number a virtual operand never value-number to the |
6123 | value from the backedge as that confuses the alias-walking code. |
6124 | See gcc.dg/torture/pr87176.c. If the value is the same on a |
6125 | non-backedge everything is OK though. */ |
6126 | bool visited_p; |
6127 | if ((backedge_val |
6128 | && !seen_non_backedge |
6129 | && TREE_CODE (backedge_val) == SSA_NAME |
6130 | && sameval == backedge_val |
6131 | && (SSA_NAME_IS_VIRTUAL_OPERAND (backedge_val) |
6132 | || SSA_VAL (x: backedge_val) != backedge_val)) |
6133 | /* Do not value-number a virtual operand to sth not visited though |
6134 | given that allows us to escape a region in alias walking. */ |
6135 | || (sameval |
6136 | && TREE_CODE (sameval) == SSA_NAME |
6137 | && !SSA_NAME_IS_DEFAULT_DEF (sameval) |
6138 | && SSA_NAME_IS_VIRTUAL_OPERAND (sameval) |
6139 | && (SSA_VAL (x: sameval, visited: &visited_p), !visited_p))) |
6140 | /* Note this just drops to VARYING without inserting the PHI into |
6141 | the hashes. */ |
6142 | result = PHI_RESULT (phi); |
6143 | /* If none of the edges was executable keep the value-number at VN_TOP, |
6144 | if only a single edge is exectuable use its value. */ |
6145 | else if (n_executable <= 1) |
6146 | result = seen_undef ? seen_undef : sameval; |
6147 | /* If we saw only undefined values and VN_TOP use one of the |
6148 | undefined values. */ |
6149 | else if (sameval == VN_TOP) |
6150 | result = (seen_undef && seen_undef_visited) ? seen_undef : sameval; |
6151 | /* First see if it is equivalent to a phi node in this block. We prefer |
6152 | this as it allows IV elimination - see PRs 66502 and 67167. */ |
6153 | else if ((result = vn_phi_lookup (phi, backedges_varying_p))) |
6154 | { |
6155 | if (!inserted |
6156 | && TREE_CODE (result) == SSA_NAME |
6157 | && gimple_code (SSA_NAME_DEF_STMT (result)) == GIMPLE_PHI) |
6158 | { |
6159 | gimple_set_plf (SSA_NAME_DEF_STMT (result), plf: GF_PLF_1, val_p: true); |
6160 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6161 | { |
6162 | fprintf (stream: dump_file, format: "Marking CSEd to PHI node " ); |
6163 | print_gimple_expr (dump_file, SSA_NAME_DEF_STMT (result), |
6164 | 0, TDF_SLIM); |
6165 | fprintf (stream: dump_file, format: "\n" ); |
6166 | } |
6167 | } |
6168 | } |
6169 | /* If all values are the same use that, unless we've seen undefined |
6170 | values as well and the value isn't constant. |
6171 | CCP/copyprop have the same restriction to not remove uninit warnings. */ |
6172 | else if (sameval |
6173 | && (! seen_undef || is_gimple_min_invariant (sameval))) |
6174 | result = sameval; |
6175 | else |
6176 | { |
6177 | result = PHI_RESULT (phi); |
6178 | /* Only insert PHIs that are varying, for constant value numbers |
6179 | we mess up equivalences otherwise as we are only comparing |
6180 | the immediate controlling predicates. */ |
6181 | vn_phi_insert (phi, result, backedges_varying_p); |
6182 | if (inserted) |
6183 | *inserted = true; |
6184 | } |
6185 | |
6186 | return set_ssa_val_to (PHI_RESULT (phi), to: result); |
6187 | } |
6188 | |
6189 | /* Try to simplify RHS using equivalences and constant folding. */ |
6190 | |
6191 | static tree |
6192 | try_to_simplify (gassign *stmt) |
6193 | { |
6194 | enum tree_code code = gimple_assign_rhs_code (gs: stmt); |
6195 | tree tem; |
6196 | |
6197 | /* For stores we can end up simplifying a SSA_NAME rhs. Just return |
6198 | in this case, there is no point in doing extra work. */ |
6199 | if (code == SSA_NAME) |
6200 | return NULL_TREE; |
6201 | |
6202 | /* First try constant folding based on our current lattice. */ |
6203 | mprts_hook = vn_lookup_simplify_result; |
6204 | tem = gimple_fold_stmt_to_constant_1 (stmt, vn_valueize, vn_valueize); |
6205 | mprts_hook = NULL; |
6206 | if (tem |
6207 | && (TREE_CODE (tem) == SSA_NAME |
6208 | || is_gimple_min_invariant (tem))) |
6209 | return tem; |
6210 | |
6211 | return NULL_TREE; |
6212 | } |
6213 | |
6214 | /* Visit and value number STMT, return true if the value number |
6215 | changed. */ |
6216 | |
6217 | static bool |
6218 | visit_stmt (gimple *stmt, bool backedges_varying_p = false) |
6219 | { |
6220 | bool changed = false; |
6221 | |
6222 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6223 | { |
6224 | fprintf (stream: dump_file, format: "Value numbering stmt = " ); |
6225 | print_gimple_stmt (dump_file, stmt, 0); |
6226 | } |
6227 | |
6228 | if (gimple_code (g: stmt) == GIMPLE_PHI) |
6229 | changed = visit_phi (phi: stmt, NULL, backedges_varying_p); |
6230 | else if (gimple_has_volatile_ops (stmt)) |
6231 | changed = defs_to_varying (stmt); |
6232 | else if (gassign *ass = dyn_cast <gassign *> (p: stmt)) |
6233 | { |
6234 | enum tree_code code = gimple_assign_rhs_code (gs: ass); |
6235 | tree lhs = gimple_assign_lhs (gs: ass); |
6236 | tree rhs1 = gimple_assign_rhs1 (gs: ass); |
6237 | tree simplified; |
6238 | |
6239 | /* Shortcut for copies. Simplifying copies is pointless, |
6240 | since we copy the expression and value they represent. */ |
6241 | if (code == SSA_NAME |
6242 | && TREE_CODE (lhs) == SSA_NAME) |
6243 | { |
6244 | changed = visit_copy (lhs, rhs: rhs1); |
6245 | goto done; |
6246 | } |
6247 | simplified = try_to_simplify (stmt: ass); |
6248 | if (simplified) |
6249 | { |
6250 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6251 | { |
6252 | fprintf (stream: dump_file, format: "RHS " ); |
6253 | print_gimple_expr (dump_file, ass, 0); |
6254 | fprintf (stream: dump_file, format: " simplified to " ); |
6255 | print_generic_expr (dump_file, simplified); |
6256 | fprintf (stream: dump_file, format: "\n" ); |
6257 | } |
6258 | } |
6259 | /* Setting value numbers to constants will occasionally |
6260 | screw up phi congruence because constants are not |
6261 | uniquely associated with a single ssa name that can be |
6262 | looked up. */ |
6263 | if (simplified |
6264 | && is_gimple_min_invariant (simplified) |
6265 | && TREE_CODE (lhs) == SSA_NAME) |
6266 | { |
6267 | changed = set_ssa_val_to (from: lhs, to: simplified); |
6268 | goto done; |
6269 | } |
6270 | else if (simplified |
6271 | && TREE_CODE (simplified) == SSA_NAME |
6272 | && TREE_CODE (lhs) == SSA_NAME) |
6273 | { |
6274 | changed = visit_copy (lhs, rhs: simplified); |
6275 | goto done; |
6276 | } |
6277 | |
6278 | if ((TREE_CODE (lhs) == SSA_NAME |
6279 | /* We can substitute SSA_NAMEs that are live over |
6280 | abnormal edges with their constant value. */ |
6281 | && !(gimple_assign_copy_p (ass) |
6282 | && is_gimple_min_invariant (rhs1)) |
6283 | && !(simplified |
6284 | && is_gimple_min_invariant (simplified)) |
6285 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)) |
6286 | /* Stores or copies from SSA_NAMEs that are live over |
6287 | abnormal edges are a problem. */ |
6288 | || (code == SSA_NAME |
6289 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1))) |
6290 | changed = defs_to_varying (stmt: ass); |
6291 | else if (REFERENCE_CLASS_P (lhs) |
6292 | || DECL_P (lhs)) |
6293 | changed = visit_reference_op_store (lhs, op: rhs1, stmt: ass); |
6294 | else if (TREE_CODE (lhs) == SSA_NAME) |
6295 | { |
6296 | if ((gimple_assign_copy_p (ass) |
6297 | && is_gimple_min_invariant (rhs1)) |
6298 | || (simplified |
6299 | && is_gimple_min_invariant (simplified))) |
6300 | { |
6301 | if (simplified) |
6302 | changed = set_ssa_val_to (from: lhs, to: simplified); |
6303 | else |
6304 | changed = set_ssa_val_to (from: lhs, to: rhs1); |
6305 | } |
6306 | else |
6307 | { |
6308 | /* Visit the original statement. */ |
6309 | switch (vn_get_stmt_kind (stmt: ass)) |
6310 | { |
6311 | case VN_NARY: |
6312 | changed = visit_nary_op (lhs, stmt: ass); |
6313 | break; |
6314 | case VN_REFERENCE: |
6315 | changed = visit_reference_op_load (lhs, op: rhs1, stmt: ass); |
6316 | break; |
6317 | default: |
6318 | changed = defs_to_varying (stmt: ass); |
6319 | break; |
6320 | } |
6321 | } |
6322 | } |
6323 | else |
6324 | changed = defs_to_varying (stmt: ass); |
6325 | } |
6326 | else if (gcall *call_stmt = dyn_cast <gcall *> (p: stmt)) |
6327 | { |
6328 | tree lhs = gimple_call_lhs (gs: call_stmt); |
6329 | if (lhs && TREE_CODE (lhs) == SSA_NAME) |
6330 | { |
6331 | /* Try constant folding based on our current lattice. */ |
6332 | tree simplified = gimple_fold_stmt_to_constant_1 (call_stmt, |
6333 | vn_valueize); |
6334 | if (simplified) |
6335 | { |
6336 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6337 | { |
6338 | fprintf (stream: dump_file, format: "call " ); |
6339 | print_gimple_expr (dump_file, call_stmt, 0); |
6340 | fprintf (stream: dump_file, format: " simplified to " ); |
6341 | print_generic_expr (dump_file, simplified); |
6342 | fprintf (stream: dump_file, format: "\n" ); |
6343 | } |
6344 | } |
6345 | /* Setting value numbers to constants will occasionally |
6346 | screw up phi congruence because constants are not |
6347 | uniquely associated with a single ssa name that can be |
6348 | looked up. */ |
6349 | if (simplified |
6350 | && is_gimple_min_invariant (simplified)) |
6351 | { |
6352 | changed = set_ssa_val_to (from: lhs, to: simplified); |
6353 | if (gimple_vdef (g: call_stmt)) |
6354 | changed |= set_ssa_val_to (from: gimple_vdef (g: call_stmt), |
6355 | to: SSA_VAL (x: gimple_vuse (g: call_stmt))); |
6356 | goto done; |
6357 | } |
6358 | else if (simplified |
6359 | && TREE_CODE (simplified) == SSA_NAME) |
6360 | { |
6361 | changed = visit_copy (lhs, rhs: simplified); |
6362 | if (gimple_vdef (g: call_stmt)) |
6363 | changed |= set_ssa_val_to (from: gimple_vdef (g: call_stmt), |
6364 | to: SSA_VAL (x: gimple_vuse (g: call_stmt))); |
6365 | goto done; |
6366 | } |
6367 | else if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)) |
6368 | { |
6369 | changed = defs_to_varying (stmt: call_stmt); |
6370 | goto done; |
6371 | } |
6372 | } |
6373 | |
6374 | /* Pick up flags from a devirtualization target. */ |
6375 | tree fn = gimple_call_fn (gs: stmt); |
6376 | int = 0; |
6377 | if (fn && TREE_CODE (fn) == SSA_NAME) |
6378 | { |
6379 | fn = SSA_VAL (x: fn); |
6380 | if (TREE_CODE (fn) == ADDR_EXPR |
6381 | && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL) |
6382 | extra_fnflags = flags_from_decl_or_type (TREE_OPERAND (fn, 0)); |
6383 | } |
6384 | if ((/* Calls to the same function with the same vuse |
6385 | and the same operands do not necessarily return the same |
6386 | value, unless they're pure or const. */ |
6387 | ((gimple_call_flags (call_stmt) | extra_fnflags) |
6388 | & (ECF_PURE | ECF_CONST)) |
6389 | /* If calls have a vdef, subsequent calls won't have |
6390 | the same incoming vuse. So, if 2 calls with vdef have the |
6391 | same vuse, we know they're not subsequent. |
6392 | We can value number 2 calls to the same function with the |
6393 | same vuse and the same operands which are not subsequent |
6394 | the same, because there is no code in the program that can |
6395 | compare the 2 values... */ |
6396 | || (gimple_vdef (g: call_stmt) |
6397 | /* ... unless the call returns a pointer which does |
6398 | not alias with anything else. In which case the |
6399 | information that the values are distinct are encoded |
6400 | in the IL. */ |
6401 | && !(gimple_call_return_flags (call_stmt) & ERF_NOALIAS) |
6402 | /* Only perform the following when being called from PRE |
6403 | which embeds tail merging. */ |
6404 | && default_vn_walk_kind == VN_WALK)) |
6405 | /* Do not process .DEFERRED_INIT since that confuses uninit |
6406 | analysis. */ |
6407 | && !gimple_call_internal_p (gs: call_stmt, fn: IFN_DEFERRED_INIT)) |
6408 | changed = visit_reference_op_call (lhs, stmt: call_stmt); |
6409 | else |
6410 | changed = defs_to_varying (stmt: call_stmt); |
6411 | } |
6412 | else |
6413 | changed = defs_to_varying (stmt); |
6414 | done: |
6415 | return changed; |
6416 | } |
6417 | |
6418 | |
6419 | /* Allocate a value number table. */ |
6420 | |
6421 | static void |
6422 | allocate_vn_table (vn_tables_t table, unsigned size) |
6423 | { |
6424 | table->phis = new vn_phi_table_type (size); |
6425 | table->nary = new vn_nary_op_table_type (size); |
6426 | table->references = new vn_reference_table_type (size); |
6427 | } |
6428 | |
6429 | /* Free a value number table. */ |
6430 | |
6431 | static void |
6432 | free_vn_table (vn_tables_t table) |
6433 | { |
6434 | /* Walk over elements and release vectors. */ |
6435 | vn_reference_iterator_type hir; |
6436 | vn_reference_t vr; |
6437 | FOR_EACH_HASH_TABLE_ELEMENT (*table->references, vr, vn_reference_t, hir) |
6438 | vr->operands.release (); |
6439 | delete table->phis; |
6440 | table->phis = NULL; |
6441 | delete table->nary; |
6442 | table->nary = NULL; |
6443 | delete table->references; |
6444 | table->references = NULL; |
6445 | } |
6446 | |
6447 | /* Set *ID according to RESULT. */ |
6448 | |
6449 | static void |
6450 | set_value_id_for_result (tree result, unsigned int *id) |
6451 | { |
6452 | if (result && TREE_CODE (result) == SSA_NAME) |
6453 | *id = VN_INFO (name: result)->value_id; |
6454 | else if (result && is_gimple_min_invariant (result)) |
6455 | *id = get_or_alloc_constant_value_id (constant: result); |
6456 | else |
6457 | *id = get_next_value_id (); |
6458 | } |
6459 | |
6460 | /* Set the value ids in the valid hash tables. */ |
6461 | |
6462 | static void |
6463 | set_hashtable_value_ids (void) |
6464 | { |
6465 | vn_nary_op_iterator_type hin; |
6466 | vn_phi_iterator_type hip; |
6467 | vn_reference_iterator_type hir; |
6468 | vn_nary_op_t vno; |
6469 | vn_reference_t vr; |
6470 | vn_phi_t vp; |
6471 | |
6472 | /* Now set the value ids of the things we had put in the hash |
6473 | table. */ |
6474 | |
6475 | FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->nary, vno, vn_nary_op_t, hin) |
6476 | if (! vno->predicated_values) |
6477 | set_value_id_for_result (result: vno->u.result, id: &vno->value_id); |
6478 | |
6479 | FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->phis, vp, vn_phi_t, hip) |
6480 | set_value_id_for_result (result: vp->result, id: &vp->value_id); |
6481 | |
6482 | FOR_EACH_HASH_TABLE_ELEMENT (*valid_info->references, vr, vn_reference_t, |
6483 | hir) |
6484 | set_value_id_for_result (result: vr->result, id: &vr->value_id); |
6485 | } |
6486 | |
6487 | /* Return the maximum value id we have ever seen. */ |
6488 | |
6489 | unsigned int |
6490 | get_max_value_id (void) |
6491 | { |
6492 | return next_value_id; |
6493 | } |
6494 | |
6495 | /* Return the maximum constant value id we have ever seen. */ |
6496 | |
6497 | unsigned int |
6498 | get_max_constant_value_id (void) |
6499 | { |
6500 | return -next_constant_value_id; |
6501 | } |
6502 | |
6503 | /* Return the next unique value id. */ |
6504 | |
6505 | unsigned int |
6506 | get_next_value_id (void) |
6507 | { |
6508 | gcc_checking_assert ((int)next_value_id > 0); |
6509 | return next_value_id++; |
6510 | } |
6511 | |
6512 | /* Return the next unique value id for constants. */ |
6513 | |
6514 | unsigned int |
6515 | get_next_constant_value_id (void) |
6516 | { |
6517 | gcc_checking_assert (next_constant_value_id < 0); |
6518 | return next_constant_value_id--; |
6519 | } |
6520 | |
6521 | |
6522 | /* Compare two expressions E1 and E2 and return true if they are equal. |
6523 | If match_vn_top_optimistically is true then VN_TOP is equal to anything, |
6524 | otherwise VN_TOP only matches VN_TOP. */ |
6525 | |
6526 | bool |
6527 | expressions_equal_p (tree e1, tree e2, bool match_vn_top_optimistically) |
6528 | { |
6529 | /* The obvious case. */ |
6530 | if (e1 == e2) |
6531 | return true; |
6532 | |
6533 | /* If either one is VN_TOP consider them equal. */ |
6534 | if (match_vn_top_optimistically |
6535 | && (e1 == VN_TOP || e2 == VN_TOP)) |
6536 | return true; |
6537 | |
6538 | /* If only one of them is null, they cannot be equal. While in general |
6539 | this should not happen for operations like TARGET_MEM_REF some |
6540 | operands are optional and an identity value we could substitute |
6541 | has differing semantics. */ |
6542 | if (!e1 || !e2) |
6543 | return false; |
6544 | |
6545 | /* SSA_NAME compare pointer equal. */ |
6546 | if (TREE_CODE (e1) == SSA_NAME || TREE_CODE (e2) == SSA_NAME) |
6547 | return false; |
6548 | |
6549 | /* Now perform the actual comparison. */ |
6550 | if (TREE_CODE (e1) == TREE_CODE (e2) |
6551 | && operand_equal_p (e1, e2, flags: OEP_PURE_SAME)) |
6552 | return true; |
6553 | |
6554 | return false; |
6555 | } |
6556 | |
6557 | |
6558 | /* Return true if the nary operation NARY may trap. This is a copy |
6559 | of stmt_could_throw_1_p adjusted to the SCCVN IL. */ |
6560 | |
6561 | bool |
6562 | vn_nary_may_trap (vn_nary_op_t nary) |
6563 | { |
6564 | tree type; |
6565 | tree rhs2 = NULL_TREE; |
6566 | bool honor_nans = false; |
6567 | bool honor_snans = false; |
6568 | bool fp_operation = false; |
6569 | bool honor_trapv = false; |
6570 | bool handled, ret; |
6571 | unsigned i; |
6572 | |
6573 | if (TREE_CODE_CLASS (nary->opcode) == tcc_comparison |
6574 | || TREE_CODE_CLASS (nary->opcode) == tcc_unary |
6575 | || TREE_CODE_CLASS (nary->opcode) == tcc_binary) |
6576 | { |
6577 | type = nary->type; |
6578 | fp_operation = FLOAT_TYPE_P (type); |
6579 | if (fp_operation) |
6580 | { |
6581 | honor_nans = flag_trapping_math && !flag_finite_math_only; |
6582 | honor_snans = flag_signaling_nans != 0; |
6583 | } |
6584 | else if (INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type)) |
6585 | honor_trapv = true; |
6586 | } |
6587 | if (nary->length >= 2) |
6588 | rhs2 = nary->op[1]; |
6589 | ret = operation_could_trap_helper_p (nary->opcode, fp_operation, |
6590 | honor_trapv, honor_nans, honor_snans, |
6591 | rhs2, &handled); |
6592 | if (handled && ret) |
6593 | return true; |
6594 | |
6595 | for (i = 0; i < nary->length; ++i) |
6596 | if (tree_could_trap_p (nary->op[i])) |
6597 | return true; |
6598 | |
6599 | return false; |
6600 | } |
6601 | |
6602 | /* Return true if the reference operation REF may trap. */ |
6603 | |
6604 | bool |
6605 | vn_reference_may_trap (vn_reference_t ref) |
6606 | { |
6607 | switch (ref->operands[0].opcode) |
6608 | { |
6609 | case MODIFY_EXPR: |
6610 | case CALL_EXPR: |
6611 | /* We do not handle calls. */ |
6612 | return true; |
6613 | case ADDR_EXPR: |
6614 | /* And toplevel address computations never trap. */ |
6615 | return false; |
6616 | default:; |
6617 | } |
6618 | |
6619 | vn_reference_op_t op; |
6620 | unsigned i; |
6621 | FOR_EACH_VEC_ELT (ref->operands, i, op) |
6622 | { |
6623 | switch (op->opcode) |
6624 | { |
6625 | case WITH_SIZE_EXPR: |
6626 | case TARGET_MEM_REF: |
6627 | /* Always variable. */ |
6628 | return true; |
6629 | case COMPONENT_REF: |
6630 | if (op->op1 && TREE_CODE (op->op1) == SSA_NAME) |
6631 | return true; |
6632 | break; |
6633 | case ARRAY_RANGE_REF: |
6634 | if (TREE_CODE (op->op0) == SSA_NAME) |
6635 | return true; |
6636 | break; |
6637 | case ARRAY_REF: |
6638 | { |
6639 | if (TREE_CODE (op->op0) != INTEGER_CST) |
6640 | return true; |
6641 | |
6642 | /* !in_array_bounds */ |
6643 | tree domain_type = TYPE_DOMAIN (ref->operands[i+1].type); |
6644 | if (!domain_type) |
6645 | return true; |
6646 | |
6647 | tree min = op->op1; |
6648 | tree max = TYPE_MAX_VALUE (domain_type); |
6649 | if (!min |
6650 | || !max |
6651 | || TREE_CODE (min) != INTEGER_CST |
6652 | || TREE_CODE (max) != INTEGER_CST) |
6653 | return true; |
6654 | |
6655 | if (tree_int_cst_lt (t1: op->op0, t2: min) |
6656 | || tree_int_cst_lt (t1: max, t2: op->op0)) |
6657 | return true; |
6658 | |
6659 | break; |
6660 | } |
6661 | case MEM_REF: |
6662 | /* Nothing interesting in itself, the base is separate. */ |
6663 | break; |
6664 | /* The following are the address bases. */ |
6665 | case SSA_NAME: |
6666 | return true; |
6667 | case ADDR_EXPR: |
6668 | if (op->op0) |
6669 | return tree_could_trap_p (TREE_OPERAND (op->op0, 0)); |
6670 | return false; |
6671 | default:; |
6672 | } |
6673 | } |
6674 | return false; |
6675 | } |
6676 | |
6677 | eliminate_dom_walker::eliminate_dom_walker (cdi_direction direction, |
6678 | bitmap inserted_exprs_) |
6679 | : dom_walker (direction), do_pre (inserted_exprs_ != NULL), |
6680 | el_todo (0), eliminations (0), insertions (0), |
6681 | inserted_exprs (inserted_exprs_) |
6682 | { |
6683 | need_eh_cleanup = BITMAP_ALLOC (NULL); |
6684 | need_ab_cleanup = BITMAP_ALLOC (NULL); |
6685 | } |
6686 | |
6687 | eliminate_dom_walker::~eliminate_dom_walker () |
6688 | { |
6689 | BITMAP_FREE (need_eh_cleanup); |
6690 | BITMAP_FREE (need_ab_cleanup); |
6691 | } |
6692 | |
6693 | /* Return a leader for OP that is available at the current point of the |
6694 | eliminate domwalk. */ |
6695 | |
6696 | tree |
6697 | eliminate_dom_walker::eliminate_avail (basic_block, tree op) |
6698 | { |
6699 | tree valnum = VN_INFO (name: op)->valnum; |
6700 | if (TREE_CODE (valnum) == SSA_NAME) |
6701 | { |
6702 | if (SSA_NAME_IS_DEFAULT_DEF (valnum)) |
6703 | return valnum; |
6704 | if (avail.length () > SSA_NAME_VERSION (valnum)) |
6705 | { |
6706 | tree av = avail[SSA_NAME_VERSION (valnum)]; |
6707 | /* When PRE discovers a new redundancy there's no way to unite |
6708 | the value classes so it instead inserts a copy old-val = new-val. |
6709 | Look through such copies here, providing one more level of |
6710 | simplification at elimination time. */ |
6711 | gassign *ass; |
6712 | if (av && (ass = dyn_cast <gassign *> (SSA_NAME_DEF_STMT (av)))) |
6713 | if (gimple_assign_rhs_class (gs: ass) == GIMPLE_SINGLE_RHS) |
6714 | { |
6715 | tree rhs1 = gimple_assign_rhs1 (gs: ass); |
6716 | if (CONSTANT_CLASS_P (rhs1) |
6717 | || (TREE_CODE (rhs1) == SSA_NAME |
6718 | && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1))) |
6719 | av = rhs1; |
6720 | } |
6721 | return av; |
6722 | } |
6723 | } |
6724 | else if (is_gimple_min_invariant (valnum)) |
6725 | return valnum; |
6726 | return NULL_TREE; |
6727 | } |
6728 | |
6729 | /* At the current point of the eliminate domwalk make OP available. */ |
6730 | |
6731 | void |
6732 | eliminate_dom_walker::eliminate_push_avail (basic_block, tree op) |
6733 | { |
6734 | tree valnum = VN_INFO (name: op)->valnum; |
6735 | if (TREE_CODE (valnum) == SSA_NAME) |
6736 | { |
6737 | if (avail.length () <= SSA_NAME_VERSION (valnum)) |
6738 | avail.safe_grow_cleared (SSA_NAME_VERSION (valnum) + 1, exact: true); |
6739 | tree pushop = op; |
6740 | if (avail[SSA_NAME_VERSION (valnum)]) |
6741 | pushop = avail[SSA_NAME_VERSION (valnum)]; |
6742 | avail_stack.safe_push (obj: pushop); |
6743 | avail[SSA_NAME_VERSION (valnum)] = op; |
6744 | } |
6745 | } |
6746 | |
6747 | /* Insert the expression recorded by SCCVN for VAL at *GSI. Returns |
6748 | the leader for the expression if insertion was successful. */ |
6749 | |
6750 | tree |
6751 | eliminate_dom_walker::eliminate_insert (basic_block bb, |
6752 | gimple_stmt_iterator *gsi, tree val) |
6753 | { |
6754 | /* We can insert a sequence with a single assignment only. */ |
6755 | gimple_seq stmts = VN_INFO (name: val)->expr; |
6756 | if (!gimple_seq_singleton_p (seq: stmts)) |
6757 | return NULL_TREE; |
6758 | gassign *stmt = dyn_cast <gassign *> (p: gimple_seq_first_stmt (s: stmts)); |
6759 | if (!stmt |
6760 | || (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt)) |
6761 | && gimple_assign_rhs_code (gs: stmt) != VIEW_CONVERT_EXPR |
6762 | && gimple_assign_rhs_code (gs: stmt) != NEGATE_EXPR |
6763 | && gimple_assign_rhs_code (gs: stmt) != BIT_FIELD_REF |
6764 | && (gimple_assign_rhs_code (gs: stmt) != BIT_AND_EXPR |
6765 | || TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST))) |
6766 | return NULL_TREE; |
6767 | |
6768 | tree op = gimple_assign_rhs1 (gs: stmt); |
6769 | if (gimple_assign_rhs_code (gs: stmt) == VIEW_CONVERT_EXPR |
6770 | || gimple_assign_rhs_code (gs: stmt) == BIT_FIELD_REF) |
6771 | op = TREE_OPERAND (op, 0); |
6772 | tree leader = TREE_CODE (op) == SSA_NAME ? eliminate_avail (bb, op) : op; |
6773 | if (!leader) |
6774 | return NULL_TREE; |
6775 | |
6776 | tree res; |
6777 | stmts = NULL; |
6778 | if (gimple_assign_rhs_code (gs: stmt) == BIT_FIELD_REF) |
6779 | res = gimple_build (seq: &stmts, code: BIT_FIELD_REF, |
6780 | TREE_TYPE (val), ops: leader, |
6781 | TREE_OPERAND (gimple_assign_rhs1 (stmt), 1), |
6782 | TREE_OPERAND (gimple_assign_rhs1 (stmt), 2)); |
6783 | else if (gimple_assign_rhs_code (gs: stmt) == BIT_AND_EXPR) |
6784 | res = gimple_build (seq: &stmts, code: BIT_AND_EXPR, |
6785 | TREE_TYPE (val), ops: leader, ops: gimple_assign_rhs2 (gs: stmt)); |
6786 | else |
6787 | res = gimple_build (seq: &stmts, code: gimple_assign_rhs_code (gs: stmt), |
6788 | TREE_TYPE (val), ops: leader); |
6789 | if (TREE_CODE (res) != SSA_NAME |
6790 | || SSA_NAME_IS_DEFAULT_DEF (res) |
6791 | || gimple_bb (SSA_NAME_DEF_STMT (res))) |
6792 | { |
6793 | gimple_seq_discard (stmts); |
6794 | |
6795 | /* During propagation we have to treat SSA info conservatively |
6796 | and thus we can end up simplifying the inserted expression |
6797 | at elimination time to sth not defined in stmts. */ |
6798 | /* But then this is a redundancy we failed to detect. Which means |
6799 | res now has two values. That doesn't play well with how |
6800 | we track availability here, so give up. */ |
6801 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6802 | { |
6803 | if (TREE_CODE (res) == SSA_NAME) |
6804 | res = eliminate_avail (bb, op: res); |
6805 | if (res) |
6806 | { |
6807 | fprintf (stream: dump_file, format: "Failed to insert expression for value " ); |
6808 | print_generic_expr (dump_file, val); |
6809 | fprintf (stream: dump_file, format: " which is really fully redundant to " ); |
6810 | print_generic_expr (dump_file, res); |
6811 | fprintf (stream: dump_file, format: "\n" ); |
6812 | } |
6813 | } |
6814 | |
6815 | return NULL_TREE; |
6816 | } |
6817 | else |
6818 | { |
6819 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
6820 | vn_ssa_aux_t vn_info = VN_INFO (name: res); |
6821 | vn_info->valnum = val; |
6822 | vn_info->visited = true; |
6823 | } |
6824 | |
6825 | insertions++; |
6826 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6827 | { |
6828 | fprintf (stream: dump_file, format: "Inserted " ); |
6829 | print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (res), 0); |
6830 | } |
6831 | |
6832 | return res; |
6833 | } |
6834 | |
6835 | void |
6836 | eliminate_dom_walker::eliminate_stmt (basic_block b, gimple_stmt_iterator *gsi) |
6837 | { |
6838 | tree sprime = NULL_TREE; |
6839 | gimple *stmt = gsi_stmt (i: *gsi); |
6840 | tree lhs = gimple_get_lhs (stmt); |
6841 | if (lhs && TREE_CODE (lhs) == SSA_NAME |
6842 | && !gimple_has_volatile_ops (stmt) |
6843 | /* See PR43491. Do not replace a global register variable when |
6844 | it is a the RHS of an assignment. Do replace local register |
6845 | variables since gcc does not guarantee a local variable will |
6846 | be allocated in register. |
6847 | ??? The fix isn't effective here. This should instead |
6848 | be ensured by not value-numbering them the same but treating |
6849 | them like volatiles? */ |
6850 | && !(gimple_assign_single_p (gs: stmt) |
6851 | && (TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL |
6852 | && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt)) |
6853 | && is_global_var (t: gimple_assign_rhs1 (gs: stmt))))) |
6854 | { |
6855 | sprime = eliminate_avail (b, op: lhs); |
6856 | if (!sprime) |
6857 | { |
6858 | /* If there is no existing usable leader but SCCVN thinks |
6859 | it has an expression it wants to use as replacement, |
6860 | insert that. */ |
6861 | tree val = VN_INFO (name: lhs)->valnum; |
6862 | vn_ssa_aux_t vn_info; |
6863 | if (val != VN_TOP |
6864 | && TREE_CODE (val) == SSA_NAME |
6865 | && (vn_info = VN_INFO (name: val), true) |
6866 | && vn_info->needs_insertion |
6867 | && vn_info->expr != NULL |
6868 | && (sprime = eliminate_insert (bb: b, gsi, val)) != NULL_TREE) |
6869 | eliminate_push_avail (b, op: sprime); |
6870 | } |
6871 | |
6872 | /* If this now constitutes a copy duplicate points-to |
6873 | and range info appropriately. This is especially |
6874 | important for inserted code. See tree-ssa-copy.cc |
6875 | for similar code. */ |
6876 | if (sprime |
6877 | && TREE_CODE (sprime) == SSA_NAME) |
6878 | { |
6879 | basic_block sprime_b = gimple_bb (SSA_NAME_DEF_STMT (sprime)); |
6880 | if (POINTER_TYPE_P (TREE_TYPE (lhs)) |
6881 | && SSA_NAME_PTR_INFO (lhs) |
6882 | && ! SSA_NAME_PTR_INFO (sprime)) |
6883 | { |
6884 | duplicate_ssa_name_ptr_info (sprime, |
6885 | SSA_NAME_PTR_INFO (lhs)); |
6886 | if (b != sprime_b) |
6887 | reset_flow_sensitive_info (sprime); |
6888 | } |
6889 | else if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)) |
6890 | && SSA_NAME_RANGE_INFO (lhs) |
6891 | && ! SSA_NAME_RANGE_INFO (sprime) |
6892 | && b == sprime_b) |
6893 | duplicate_ssa_name_range_info (dest: sprime, src: lhs); |
6894 | } |
6895 | |
6896 | /* Inhibit the use of an inserted PHI on a loop header when |
6897 | the address of the memory reference is a simple induction |
6898 | variable. In other cases the vectorizer won't do anything |
6899 | anyway (either it's loop invariant or a complicated |
6900 | expression). */ |
6901 | if (sprime |
6902 | && TREE_CODE (sprime) == SSA_NAME |
6903 | && do_pre |
6904 | && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1) |
6905 | && loop_outer (loop: b->loop_father) |
6906 | && has_zero_uses (var: sprime) |
6907 | && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime)) |
6908 | && gimple_assign_load_p (stmt)) |
6909 | { |
6910 | gimple *def_stmt = SSA_NAME_DEF_STMT (sprime); |
6911 | basic_block def_bb = gimple_bb (g: def_stmt); |
6912 | if (gimple_code (g: def_stmt) == GIMPLE_PHI |
6913 | && def_bb->loop_father->header == def_bb) |
6914 | { |
6915 | loop_p loop = def_bb->loop_father; |
6916 | ssa_op_iter iter; |
6917 | tree op; |
6918 | bool found = false; |
6919 | FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE) |
6920 | { |
6921 | affine_iv iv; |
6922 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (op)); |
6923 | if (def_bb |
6924 | && flow_bb_inside_loop_p (loop, def_bb) |
6925 | && simple_iv (loop, loop, op, &iv, true)) |
6926 | { |
6927 | found = true; |
6928 | break; |
6929 | } |
6930 | } |
6931 | if (found) |
6932 | { |
6933 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6934 | { |
6935 | fprintf (stream: dump_file, format: "Not replacing " ); |
6936 | print_gimple_expr (dump_file, stmt, 0); |
6937 | fprintf (stream: dump_file, format: " with " ); |
6938 | print_generic_expr (dump_file, sprime); |
6939 | fprintf (stream: dump_file, format: " which would add a loop" |
6940 | " carried dependence to loop %d\n" , |
6941 | loop->num); |
6942 | } |
6943 | /* Don't keep sprime available. */ |
6944 | sprime = NULL_TREE; |
6945 | } |
6946 | } |
6947 | } |
6948 | |
6949 | if (sprime) |
6950 | { |
6951 | /* If we can propagate the value computed for LHS into |
6952 | all uses don't bother doing anything with this stmt. */ |
6953 | if (may_propagate_copy (lhs, sprime)) |
6954 | { |
6955 | /* Mark it for removal. */ |
6956 | to_remove.safe_push (obj: stmt); |
6957 | |
6958 | /* ??? Don't count copy/constant propagations. */ |
6959 | if (gimple_assign_single_p (gs: stmt) |
6960 | && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME |
6961 | || gimple_assign_rhs1 (gs: stmt) == sprime)) |
6962 | return; |
6963 | |
6964 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6965 | { |
6966 | fprintf (stream: dump_file, format: "Replaced " ); |
6967 | print_gimple_expr (dump_file, stmt, 0); |
6968 | fprintf (stream: dump_file, format: " with " ); |
6969 | print_generic_expr (dump_file, sprime); |
6970 | fprintf (stream: dump_file, format: " in all uses of " ); |
6971 | print_gimple_stmt (dump_file, stmt, 0); |
6972 | } |
6973 | |
6974 | eliminations++; |
6975 | return; |
6976 | } |
6977 | |
6978 | /* If this is an assignment from our leader (which |
6979 | happens in the case the value-number is a constant) |
6980 | then there is nothing to do. Likewise if we run into |
6981 | inserted code that needed a conversion because of |
6982 | our type-agnostic value-numbering of loads. */ |
6983 | if ((gimple_assign_single_p (gs: stmt) |
6984 | || (is_gimple_assign (gs: stmt) |
6985 | && (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt)) |
6986 | || gimple_assign_rhs_code (gs: stmt) == VIEW_CONVERT_EXPR))) |
6987 | && sprime == gimple_assign_rhs1 (gs: stmt)) |
6988 | return; |
6989 | |
6990 | /* Else replace its RHS. */ |
6991 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6992 | { |
6993 | fprintf (stream: dump_file, format: "Replaced " ); |
6994 | print_gimple_expr (dump_file, stmt, 0); |
6995 | fprintf (stream: dump_file, format: " with " ); |
6996 | print_generic_expr (dump_file, sprime); |
6997 | fprintf (stream: dump_file, format: " in " ); |
6998 | print_gimple_stmt (dump_file, stmt, 0); |
6999 | } |
7000 | eliminations++; |
7001 | |
7002 | bool can_make_abnormal_goto = (is_gimple_call (gs: stmt) |
7003 | && stmt_can_make_abnormal_goto (stmt)); |
7004 | gimple *orig_stmt = stmt; |
7005 | if (!useless_type_conversion_p (TREE_TYPE (lhs), |
7006 | TREE_TYPE (sprime))) |
7007 | { |
7008 | /* We preserve conversions to but not from function or method |
7009 | types. This asymmetry makes it necessary to re-instantiate |
7010 | conversions here. */ |
7011 | if (POINTER_TYPE_P (TREE_TYPE (lhs)) |
7012 | && FUNC_OR_METHOD_TYPE_P (TREE_TYPE (TREE_TYPE (lhs)))) |
7013 | sprime = fold_convert (TREE_TYPE (lhs), sprime); |
7014 | else |
7015 | gcc_unreachable (); |
7016 | } |
7017 | tree vdef = gimple_vdef (g: stmt); |
7018 | tree vuse = gimple_vuse (g: stmt); |
7019 | propagate_tree_value_into_stmt (gsi, sprime); |
7020 | stmt = gsi_stmt (i: *gsi); |
7021 | update_stmt (s: stmt); |
7022 | /* In case the VDEF on the original stmt was released, value-number |
7023 | it to the VUSE. This is to make vuse_ssa_val able to skip |
7024 | released virtual operands. */ |
7025 | if (vdef != gimple_vdef (g: stmt)) |
7026 | { |
7027 | gcc_assert (SSA_NAME_IN_FREE_LIST (vdef)); |
7028 | VN_INFO (name: vdef)->valnum = vuse; |
7029 | } |
7030 | |
7031 | /* If we removed EH side-effects from the statement, clean |
7032 | its EH information. */ |
7033 | if (maybe_clean_or_replace_eh_stmt (orig_stmt, stmt)) |
7034 | { |
7035 | bitmap_set_bit (need_eh_cleanup, |
7036 | gimple_bb (g: stmt)->index); |
7037 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7038 | fprintf (stream: dump_file, format: " Removed EH side-effects.\n" ); |
7039 | } |
7040 | |
7041 | /* Likewise for AB side-effects. */ |
7042 | if (can_make_abnormal_goto |
7043 | && !stmt_can_make_abnormal_goto (stmt)) |
7044 | { |
7045 | bitmap_set_bit (need_ab_cleanup, |
7046 | gimple_bb (g: stmt)->index); |
7047 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7048 | fprintf (stream: dump_file, format: " Removed AB side-effects.\n" ); |
7049 | } |
7050 | |
7051 | return; |
7052 | } |
7053 | } |
7054 | |
7055 | /* If the statement is a scalar store, see if the expression |
7056 | has the same value number as its rhs. If so, the store is |
7057 | dead. */ |
7058 | if (gimple_assign_single_p (gs: stmt) |
7059 | && !gimple_has_volatile_ops (stmt) |
7060 | && !is_gimple_reg (gimple_assign_lhs (gs: stmt)) |
7061 | && (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME |
7062 | || is_gimple_min_invariant (gimple_assign_rhs1 (gs: stmt)))) |
7063 | { |
7064 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
7065 | vn_reference_t vnresult; |
7066 | /* ??? gcc.dg/torture/pr91445.c shows that we lookup a boolean |
7067 | typed load of a byte known to be 0x11 as 1 so a store of |
7068 | a boolean 1 is detected as redundant. Because of this we |
7069 | have to make sure to lookup with a ref where its size |
7070 | matches the precision. */ |
7071 | tree lookup_lhs = lhs; |
7072 | if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)) |
7073 | && (TREE_CODE (lhs) != COMPONENT_REF |
7074 | || !DECL_BIT_FIELD_TYPE (TREE_OPERAND (lhs, 1))) |
7075 | && !type_has_mode_precision_p (TREE_TYPE (lhs))) |
7076 | { |
7077 | if (TREE_CODE (TREE_TYPE (lhs)) == BITINT_TYPE |
7078 | && TYPE_PRECISION (TREE_TYPE (lhs)) > MAX_FIXED_MODE_SIZE) |
7079 | lookup_lhs = NULL_TREE; |
7080 | else if (TREE_CODE (lhs) == COMPONENT_REF |
7081 | || TREE_CODE (lhs) == MEM_REF) |
7082 | { |
7083 | tree ltype = build_nonstandard_integer_type |
7084 | (TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (lhs))), |
7085 | TYPE_UNSIGNED (TREE_TYPE (lhs))); |
7086 | if (TREE_CODE (lhs) == COMPONENT_REF) |
7087 | { |
7088 | tree foff = component_ref_field_offset (lhs); |
7089 | tree f = TREE_OPERAND (lhs, 1); |
7090 | if (!poly_int_tree_p (t: foff)) |
7091 | lookup_lhs = NULL_TREE; |
7092 | else |
7093 | lookup_lhs = build3 (BIT_FIELD_REF, ltype, |
7094 | TREE_OPERAND (lhs, 0), |
7095 | TYPE_SIZE (TREE_TYPE (lhs)), |
7096 | bit_from_pos |
7097 | (foff, DECL_FIELD_BIT_OFFSET (f))); |
7098 | } |
7099 | else |
7100 | lookup_lhs = build2 (MEM_REF, ltype, |
7101 | TREE_OPERAND (lhs, 0), |
7102 | TREE_OPERAND (lhs, 1)); |
7103 | } |
7104 | else |
7105 | lookup_lhs = NULL_TREE; |
7106 | } |
7107 | tree val = NULL_TREE; |
7108 | if (lookup_lhs) |
7109 | val = vn_reference_lookup (op: lookup_lhs, vuse: gimple_vuse (g: stmt), |
7110 | kind: VN_WALKREWRITE, vnresult: &vnresult, tbaa_p: false, |
7111 | NULL, NULL_TREE, redundant_store_removal_p: true); |
7112 | if (TREE_CODE (rhs) == SSA_NAME) |
7113 | rhs = VN_INFO (name: rhs)->valnum; |
7114 | if (val |
7115 | && (operand_equal_p (val, rhs, flags: 0) |
7116 | /* Due to the bitfield lookups above we can get bit |
7117 | interpretations of the same RHS as values here. Those |
7118 | are redundant as well. */ |
7119 | || (TREE_CODE (val) == SSA_NAME |
7120 | && gimple_assign_single_p (SSA_NAME_DEF_STMT (val)) |
7121 | && (val = gimple_assign_rhs1 (SSA_NAME_DEF_STMT (val))) |
7122 | && TREE_CODE (val) == VIEW_CONVERT_EXPR |
7123 | && TREE_OPERAND (val, 0) == rhs))) |
7124 | { |
7125 | /* We can only remove the later store if the former aliases |
7126 | at least all accesses the later one does or if the store |
7127 | was to readonly memory storing the same value. */ |
7128 | ao_ref lhs_ref; |
7129 | ao_ref_init (&lhs_ref, lhs); |
7130 | alias_set_type set = ao_ref_alias_set (&lhs_ref); |
7131 | alias_set_type base_set = ao_ref_base_alias_set (&lhs_ref); |
7132 | if (! vnresult |
7133 | || ((vnresult->set == set |
7134 | || alias_set_subset_of (set, vnresult->set)) |
7135 | && (vnresult->base_set == base_set |
7136 | || alias_set_subset_of (base_set, vnresult->base_set)))) |
7137 | { |
7138 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7139 | { |
7140 | fprintf (stream: dump_file, format: "Deleted redundant store " ); |
7141 | print_gimple_stmt (dump_file, stmt, 0); |
7142 | } |
7143 | |
7144 | /* Queue stmt for removal. */ |
7145 | to_remove.safe_push (obj: stmt); |
7146 | return; |
7147 | } |
7148 | } |
7149 | } |
7150 | |
7151 | /* If this is a control statement value numbering left edges |
7152 | unexecuted on force the condition in a way consistent with |
7153 | that. */ |
7154 | if (gcond *cond = dyn_cast <gcond *> (p: stmt)) |
7155 | { |
7156 | if ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) |
7157 | ^ (EDGE_SUCC (b, 1)->flags & EDGE_EXECUTABLE)) |
7158 | { |
7159 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7160 | { |
7161 | fprintf (stream: dump_file, format: "Removing unexecutable edge from " ); |
7162 | print_gimple_stmt (dump_file, stmt, 0); |
7163 | } |
7164 | if (((EDGE_SUCC (b, 0)->flags & EDGE_TRUE_VALUE) != 0) |
7165 | == ((EDGE_SUCC (b, 0)->flags & EDGE_EXECUTABLE) != 0)) |
7166 | gimple_cond_make_true (gs: cond); |
7167 | else |
7168 | gimple_cond_make_false (gs: cond); |
7169 | update_stmt (s: cond); |
7170 | el_todo |= TODO_cleanup_cfg; |
7171 | return; |
7172 | } |
7173 | } |
7174 | |
7175 | bool can_make_abnormal_goto = stmt_can_make_abnormal_goto (stmt); |
7176 | bool was_noreturn = (is_gimple_call (gs: stmt) |
7177 | && gimple_call_noreturn_p (s: stmt)); |
7178 | tree vdef = gimple_vdef (g: stmt); |
7179 | tree vuse = gimple_vuse (g: stmt); |
7180 | |
7181 | /* If we didn't replace the whole stmt (or propagate the result |
7182 | into all uses), replace all uses on this stmt with their |
7183 | leaders. */ |
7184 | bool modified = false; |
7185 | use_operand_p use_p; |
7186 | ssa_op_iter iter; |
7187 | FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) |
7188 | { |
7189 | tree use = USE_FROM_PTR (use_p); |
7190 | /* ??? The call code above leaves stmt operands un-updated. */ |
7191 | if (TREE_CODE (use) != SSA_NAME) |
7192 | continue; |
7193 | tree sprime; |
7194 | if (SSA_NAME_IS_DEFAULT_DEF (use)) |
7195 | /* ??? For default defs BB shouldn't matter, but we have to |
7196 | solve the inconsistency between rpo eliminate and |
7197 | dom eliminate avail valueization first. */ |
7198 | sprime = eliminate_avail (b, op: use); |
7199 | else |
7200 | /* Look for sth available at the definition block of the argument. |
7201 | This avoids inconsistencies between availability there which |
7202 | decides if the stmt can be removed and availability at the |
7203 | use site. The SSA property ensures that things available |
7204 | at the definition are also available at uses. */ |
7205 | sprime = eliminate_avail (gimple_bb (SSA_NAME_DEF_STMT (use)), op: use); |
7206 | if (sprime && sprime != use |
7207 | && may_propagate_copy (use, sprime, true) |
7208 | /* We substitute into debug stmts to avoid excessive |
7209 | debug temporaries created by removed stmts, but we need |
7210 | to avoid doing so for inserted sprimes as we never want |
7211 | to create debug temporaries for them. */ |
7212 | && (!inserted_exprs |
7213 | || TREE_CODE (sprime) != SSA_NAME |
7214 | || !is_gimple_debug (gs: stmt) |
7215 | || !bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime)))) |
7216 | { |
7217 | propagate_value (use_p, sprime); |
7218 | modified = true; |
7219 | } |
7220 | } |
7221 | |
7222 | /* Fold the stmt if modified, this canonicalizes MEM_REFs we propagated |
7223 | into which is a requirement for the IPA devirt machinery. */ |
7224 | gimple *old_stmt = stmt; |
7225 | if (modified) |
7226 | { |
7227 | /* If a formerly non-invariant ADDR_EXPR is turned into an |
7228 | invariant one it was on a separate stmt. */ |
7229 | if (gimple_assign_single_p (gs: stmt) |
7230 | && TREE_CODE (gimple_assign_rhs1 (stmt)) == ADDR_EXPR) |
7231 | recompute_tree_invariant_for_addr_expr (gimple_assign_rhs1 (gs: stmt)); |
7232 | gimple_stmt_iterator prev = *gsi; |
7233 | gsi_prev (i: &prev); |
7234 | if (fold_stmt (gsi, follow_all_ssa_edges)) |
7235 | { |
7236 | /* fold_stmt may have created new stmts inbetween |
7237 | the previous stmt and the folded stmt. Mark |
7238 | all defs created there as varying to not confuse |
7239 | the SCCVN machinery as we're using that even during |
7240 | elimination. */ |
7241 | if (gsi_end_p (i: prev)) |
7242 | prev = gsi_start_bb (bb: b); |
7243 | else |
7244 | gsi_next (i: &prev); |
7245 | if (gsi_stmt (i: prev) != gsi_stmt (i: *gsi)) |
7246 | do |
7247 | { |
7248 | tree def; |
7249 | ssa_op_iter dit; |
7250 | FOR_EACH_SSA_TREE_OPERAND (def, gsi_stmt (prev), |
7251 | dit, SSA_OP_ALL_DEFS) |
7252 | /* As existing DEFs may move between stmts |
7253 | only process new ones. */ |
7254 | if (! has_VN_INFO (name: def)) |
7255 | { |
7256 | vn_ssa_aux_t vn_info = VN_INFO (name: def); |
7257 | vn_info->valnum = def; |
7258 | vn_info->visited = true; |
7259 | } |
7260 | if (gsi_stmt (i: prev) == gsi_stmt (i: *gsi)) |
7261 | break; |
7262 | gsi_next (i: &prev); |
7263 | } |
7264 | while (1); |
7265 | } |
7266 | stmt = gsi_stmt (i: *gsi); |
7267 | /* In case we folded the stmt away schedule the NOP for removal. */ |
7268 | if (gimple_nop_p (g: stmt)) |
7269 | to_remove.safe_push (obj: stmt); |
7270 | } |
7271 | |
7272 | /* Visit indirect calls and turn them into direct calls if |
7273 | possible using the devirtualization machinery. Do this before |
7274 | checking for required EH/abnormal/noreturn cleanup as devird |
7275 | may expose more of those. */ |
7276 | if (gcall *call_stmt = dyn_cast <gcall *> (p: stmt)) |
7277 | { |
7278 | tree fn = gimple_call_fn (gs: call_stmt); |
7279 | if (fn |
7280 | && flag_devirtualize |
7281 | && virtual_method_call_p (fn)) |
7282 | { |
7283 | tree otr_type = obj_type_ref_class (ref: fn); |
7284 | unsigned HOST_WIDE_INT otr_tok |
7285 | = tree_to_uhwi (OBJ_TYPE_REF_TOKEN (fn)); |
7286 | tree instance; |
7287 | ipa_polymorphic_call_context context (current_function_decl, |
7288 | fn, stmt, &instance); |
7289 | context.get_dynamic_type (instance, OBJ_TYPE_REF_OBJECT (fn), |
7290 | otr_type, stmt, NULL); |
7291 | bool final; |
7292 | vec <cgraph_node *> targets |
7293 | = possible_polymorphic_call_targets (obj_type_ref_class (ref: fn), |
7294 | otr_tok, context, copletep: &final); |
7295 | if (dump_file) |
7296 | dump_possible_polymorphic_call_targets (dump_file, |
7297 | obj_type_ref_class (ref: fn), |
7298 | otr_tok, context); |
7299 | if (final && targets.length () <= 1 && dbg_cnt (index: devirt)) |
7300 | { |
7301 | tree fn; |
7302 | if (targets.length () == 1) |
7303 | fn = targets[0]->decl; |
7304 | else |
7305 | fn = builtin_decl_unreachable (); |
7306 | if (dump_enabled_p ()) |
7307 | { |
7308 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, stmt, |
7309 | "converting indirect call to " |
7310 | "function %s\n" , |
7311 | lang_hooks.decl_printable_name (fn, 2)); |
7312 | } |
7313 | gimple_call_set_fndecl (gs: call_stmt, decl: fn); |
7314 | /* If changing the call to __builtin_unreachable |
7315 | or similar noreturn function, adjust gimple_call_fntype |
7316 | too. */ |
7317 | if (gimple_call_noreturn_p (s: call_stmt) |
7318 | && VOID_TYPE_P (TREE_TYPE (TREE_TYPE (fn))) |
7319 | && TYPE_ARG_TYPES (TREE_TYPE (fn)) |
7320 | && (TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fn))) |
7321 | == void_type_node)) |
7322 | gimple_call_set_fntype (call_stmt, TREE_TYPE (fn)); |
7323 | maybe_remove_unused_call_args (cfun, call_stmt); |
7324 | modified = true; |
7325 | } |
7326 | } |
7327 | } |
7328 | |
7329 | if (modified) |
7330 | { |
7331 | /* When changing a call into a noreturn call, cfg cleanup |
7332 | is needed to fix up the noreturn call. */ |
7333 | if (!was_noreturn |
7334 | && is_gimple_call (gs: stmt) && gimple_call_noreturn_p (s: stmt)) |
7335 | to_fixup.safe_push (obj: stmt); |
7336 | /* When changing a condition or switch into one we know what |
7337 | edge will be executed, schedule a cfg cleanup. */ |
7338 | if ((gimple_code (g: stmt) == GIMPLE_COND |
7339 | && (gimple_cond_true_p (gs: as_a <gcond *> (p: stmt)) |
7340 | || gimple_cond_false_p (gs: as_a <gcond *> (p: stmt)))) |
7341 | || (gimple_code (g: stmt) == GIMPLE_SWITCH |
7342 | && TREE_CODE (gimple_switch_index |
7343 | (as_a <gswitch *> (stmt))) == INTEGER_CST)) |
7344 | el_todo |= TODO_cleanup_cfg; |
7345 | /* If we removed EH side-effects from the statement, clean |
7346 | its EH information. */ |
7347 | if (maybe_clean_or_replace_eh_stmt (old_stmt, stmt)) |
7348 | { |
7349 | bitmap_set_bit (need_eh_cleanup, |
7350 | gimple_bb (g: stmt)->index); |
7351 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7352 | fprintf (stream: dump_file, format: " Removed EH side-effects.\n" ); |
7353 | } |
7354 | /* Likewise for AB side-effects. */ |
7355 | if (can_make_abnormal_goto |
7356 | && !stmt_can_make_abnormal_goto (stmt)) |
7357 | { |
7358 | bitmap_set_bit (need_ab_cleanup, |
7359 | gimple_bb (g: stmt)->index); |
7360 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7361 | fprintf (stream: dump_file, format: " Removed AB side-effects.\n" ); |
7362 | } |
7363 | update_stmt (s: stmt); |
7364 | /* In case the VDEF on the original stmt was released, value-number |
7365 | it to the VUSE. This is to make vuse_ssa_val able to skip |
7366 | released virtual operands. */ |
7367 | if (vdef && SSA_NAME_IN_FREE_LIST (vdef)) |
7368 | VN_INFO (name: vdef)->valnum = vuse; |
7369 | } |
7370 | |
7371 | /* Make new values available - for fully redundant LHS we |
7372 | continue with the next stmt above and skip this. |
7373 | But avoid picking up dead defs. */ |
7374 | tree def; |
7375 | FOR_EACH_SSA_TREE_OPERAND (def, stmt, iter, SSA_OP_DEF) |
7376 | if (! has_zero_uses (var: def) |
7377 | || (inserted_exprs |
7378 | && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (def)))) |
7379 | eliminate_push_avail (b, op: def); |
7380 | } |
7381 | |
7382 | /* Perform elimination for the basic-block B during the domwalk. */ |
7383 | |
7384 | edge |
7385 | eliminate_dom_walker::before_dom_children (basic_block b) |
7386 | { |
7387 | /* Mark new bb. */ |
7388 | avail_stack.safe_push (NULL_TREE); |
7389 | |
7390 | /* Skip unreachable blocks marked unreachable during the SCCVN domwalk. */ |
7391 | if (!(b->flags & BB_EXECUTABLE)) |
7392 | return NULL; |
7393 | |
7394 | vn_context_bb = b; |
7395 | |
7396 | for (gphi_iterator gsi = gsi_start_phis (b); !gsi_end_p (i: gsi);) |
7397 | { |
7398 | gphi *phi = gsi.phi (); |
7399 | tree res = PHI_RESULT (phi); |
7400 | |
7401 | if (virtual_operand_p (op: res)) |
7402 | { |
7403 | gsi_next (i: &gsi); |
7404 | continue; |
7405 | } |
7406 | |
7407 | tree sprime = eliminate_avail (b, op: res); |
7408 | if (sprime |
7409 | && sprime != res) |
7410 | { |
7411 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7412 | { |
7413 | fprintf (stream: dump_file, format: "Replaced redundant PHI node defining " ); |
7414 | print_generic_expr (dump_file, res); |
7415 | fprintf (stream: dump_file, format: " with " ); |
7416 | print_generic_expr (dump_file, sprime); |
7417 | fprintf (stream: dump_file, format: "\n" ); |
7418 | } |
7419 | |
7420 | /* If we inserted this PHI node ourself, it's not an elimination. */ |
7421 | if (! inserted_exprs |
7422 | || ! bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (res))) |
7423 | eliminations++; |
7424 | |
7425 | /* If we will propagate into all uses don't bother to do |
7426 | anything. */ |
7427 | if (may_propagate_copy (res, sprime)) |
7428 | { |
7429 | /* Mark the PHI for removal. */ |
7430 | to_remove.safe_push (obj: phi); |
7431 | gsi_next (i: &gsi); |
7432 | continue; |
7433 | } |
7434 | |
7435 | remove_phi_node (&gsi, false); |
7436 | |
7437 | if (!useless_type_conversion_p (TREE_TYPE (res), TREE_TYPE (sprime))) |
7438 | sprime = fold_convert (TREE_TYPE (res), sprime); |
7439 | gimple *stmt = gimple_build_assign (res, sprime); |
7440 | gimple_stmt_iterator gsi2 = gsi_after_labels (bb: b); |
7441 | gsi_insert_before (&gsi2, stmt, GSI_NEW_STMT); |
7442 | continue; |
7443 | } |
7444 | |
7445 | eliminate_push_avail (b, op: res); |
7446 | gsi_next (i: &gsi); |
7447 | } |
7448 | |
7449 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb: b); |
7450 | !gsi_end_p (i: gsi); |
7451 | gsi_next (i: &gsi)) |
7452 | eliminate_stmt (b, gsi: &gsi); |
7453 | |
7454 | /* Replace destination PHI arguments. */ |
7455 | edge_iterator ei; |
7456 | edge e; |
7457 | FOR_EACH_EDGE (e, ei, b->succs) |
7458 | if (e->flags & EDGE_EXECUTABLE) |
7459 | for (gphi_iterator gsi = gsi_start_phis (e->dest); |
7460 | !gsi_end_p (i: gsi); |
7461 | gsi_next (i: &gsi)) |
7462 | { |
7463 | gphi *phi = gsi.phi (); |
7464 | use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e); |
7465 | tree arg = USE_FROM_PTR (use_p); |
7466 | if (TREE_CODE (arg) != SSA_NAME |
7467 | || virtual_operand_p (op: arg)) |
7468 | continue; |
7469 | tree sprime = eliminate_avail (b, op: arg); |
7470 | if (sprime && may_propagate_copy (arg, sprime, |
7471 | !(e->flags & EDGE_ABNORMAL))) |
7472 | propagate_value (use_p, sprime); |
7473 | } |
7474 | |
7475 | vn_context_bb = NULL; |
7476 | |
7477 | return NULL; |
7478 | } |
7479 | |
7480 | /* Make no longer available leaders no longer available. */ |
7481 | |
7482 | void |
7483 | eliminate_dom_walker::after_dom_children (basic_block) |
7484 | { |
7485 | tree entry; |
7486 | while ((entry = avail_stack.pop ()) != NULL_TREE) |
7487 | { |
7488 | tree valnum = VN_INFO (name: entry)->valnum; |
7489 | tree old = avail[SSA_NAME_VERSION (valnum)]; |
7490 | if (old == entry) |
7491 | avail[SSA_NAME_VERSION (valnum)] = NULL_TREE; |
7492 | else |
7493 | avail[SSA_NAME_VERSION (valnum)] = entry; |
7494 | } |
7495 | } |
7496 | |
7497 | /* Remove queued stmts and perform delayed cleanups. */ |
7498 | |
7499 | unsigned |
7500 | eliminate_dom_walker::eliminate_cleanup (bool region_p) |
7501 | { |
7502 | statistics_counter_event (cfun, "Eliminated" , eliminations); |
7503 | statistics_counter_event (cfun, "Insertions" , insertions); |
7504 | |
7505 | /* We cannot remove stmts during BB walk, especially not release SSA |
7506 | names there as this confuses the VN machinery. The stmts ending |
7507 | up in to_remove are either stores or simple copies. |
7508 | Remove stmts in reverse order to make debug stmt creation possible. */ |
7509 | while (!to_remove.is_empty ()) |
7510 | { |
7511 | bool do_release_defs = true; |
7512 | gimple *stmt = to_remove.pop (); |
7513 | |
7514 | /* When we are value-numbering a region we do not require exit PHIs to |
7515 | be present so we have to make sure to deal with uses outside of the |
7516 | region of stmts that we thought are eliminated. |
7517 | ??? Note we may be confused by uses in dead regions we didn't run |
7518 | elimination on. Rather than checking individual uses we accept |
7519 | dead copies to be generated here (gcc.c-torture/execute/20060905-1.c |
7520 | contains such example). */ |
7521 | if (region_p) |
7522 | { |
7523 | if (gphi *phi = dyn_cast <gphi *> (p: stmt)) |
7524 | { |
7525 | tree lhs = gimple_phi_result (gs: phi); |
7526 | if (!has_zero_uses (var: lhs)) |
7527 | { |
7528 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7529 | fprintf (stream: dump_file, format: "Keeping eliminated stmt live " |
7530 | "as copy because of out-of-region uses\n" ); |
7531 | tree sprime = eliminate_avail (gimple_bb (g: stmt), op: lhs); |
7532 | gimple *copy = gimple_build_assign (lhs, sprime); |
7533 | gimple_stmt_iterator gsi |
7534 | = gsi_after_labels (bb: gimple_bb (g: stmt)); |
7535 | gsi_insert_before (&gsi, copy, GSI_SAME_STMT); |
7536 | do_release_defs = false; |
7537 | } |
7538 | } |
7539 | else if (tree lhs = gimple_get_lhs (stmt)) |
7540 | if (TREE_CODE (lhs) == SSA_NAME |
7541 | && !has_zero_uses (var: lhs)) |
7542 | { |
7543 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7544 | fprintf (stream: dump_file, format: "Keeping eliminated stmt live " |
7545 | "as copy because of out-of-region uses\n" ); |
7546 | tree sprime = eliminate_avail (gimple_bb (g: stmt), op: lhs); |
7547 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); |
7548 | if (is_gimple_assign (gs: stmt)) |
7549 | { |
7550 | gimple_assign_set_rhs_from_tree (&gsi, sprime); |
7551 | stmt = gsi_stmt (i: gsi); |
7552 | update_stmt (s: stmt); |
7553 | if (maybe_clean_or_replace_eh_stmt (stmt, stmt)) |
7554 | bitmap_set_bit (need_eh_cleanup, gimple_bb (g: stmt)->index); |
7555 | continue; |
7556 | } |
7557 | else |
7558 | { |
7559 | gimple *copy = gimple_build_assign (lhs, sprime); |
7560 | gsi_insert_before (&gsi, copy, GSI_SAME_STMT); |
7561 | do_release_defs = false; |
7562 | } |
7563 | } |
7564 | } |
7565 | |
7566 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7567 | { |
7568 | fprintf (stream: dump_file, format: "Removing dead stmt " ); |
7569 | print_gimple_stmt (dump_file, stmt, 0, TDF_NONE); |
7570 | } |
7571 | |
7572 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); |
7573 | if (gimple_code (g: stmt) == GIMPLE_PHI) |
7574 | remove_phi_node (&gsi, do_release_defs); |
7575 | else |
7576 | { |
7577 | basic_block bb = gimple_bb (g: stmt); |
7578 | unlink_stmt_vdef (stmt); |
7579 | if (gsi_remove (&gsi, true)) |
7580 | bitmap_set_bit (need_eh_cleanup, bb->index); |
7581 | if (is_gimple_call (gs: stmt) && stmt_can_make_abnormal_goto (stmt)) |
7582 | bitmap_set_bit (need_ab_cleanup, bb->index); |
7583 | if (do_release_defs) |
7584 | release_defs (stmt); |
7585 | } |
7586 | |
7587 | /* Removing a stmt may expose a forwarder block. */ |
7588 | el_todo |= TODO_cleanup_cfg; |
7589 | } |
7590 | |
7591 | /* Fixup stmts that became noreturn calls. This may require splitting |
7592 | blocks and thus isn't possible during the dominator walk. Do this |
7593 | in reverse order so we don't inadvertedly remove a stmt we want to |
7594 | fixup by visiting a dominating now noreturn call first. */ |
7595 | while (!to_fixup.is_empty ()) |
7596 | { |
7597 | gimple *stmt = to_fixup.pop (); |
7598 | |
7599 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7600 | { |
7601 | fprintf (stream: dump_file, format: "Fixing up noreturn call " ); |
7602 | print_gimple_stmt (dump_file, stmt, 0); |
7603 | } |
7604 | |
7605 | if (fixup_noreturn_call (stmt)) |
7606 | el_todo |= TODO_cleanup_cfg; |
7607 | } |
7608 | |
7609 | bool do_eh_cleanup = !bitmap_empty_p (map: need_eh_cleanup); |
7610 | bool do_ab_cleanup = !bitmap_empty_p (map: need_ab_cleanup); |
7611 | |
7612 | if (do_eh_cleanup) |
7613 | gimple_purge_all_dead_eh_edges (need_eh_cleanup); |
7614 | |
7615 | if (do_ab_cleanup) |
7616 | gimple_purge_all_dead_abnormal_call_edges (need_ab_cleanup); |
7617 | |
7618 | if (do_eh_cleanup || do_ab_cleanup) |
7619 | el_todo |= TODO_cleanup_cfg; |
7620 | |
7621 | return el_todo; |
7622 | } |
7623 | |
7624 | /* Eliminate fully redundant computations. */ |
7625 | |
7626 | unsigned |
7627 | eliminate_with_rpo_vn (bitmap inserted_exprs) |
7628 | { |
7629 | eliminate_dom_walker walker (CDI_DOMINATORS, inserted_exprs); |
7630 | |
7631 | eliminate_dom_walker *saved_rpo_avail = rpo_avail; |
7632 | rpo_avail = &walker; |
7633 | walker.walk (cfun->cfg->x_entry_block_ptr); |
7634 | rpo_avail = saved_rpo_avail; |
7635 | |
7636 | return walker.eliminate_cleanup (); |
7637 | } |
7638 | |
7639 | static unsigned |
7640 | do_rpo_vn_1 (function *fn, edge entry, bitmap exit_bbs, |
7641 | bool iterate, bool eliminate, bool skip_entry_phis, |
7642 | vn_lookup_kind kind); |
7643 | |
7644 | void |
7645 | run_rpo_vn (vn_lookup_kind kind) |
7646 | { |
7647 | do_rpo_vn_1 (cfun, NULL, NULL, iterate: true, eliminate: false, skip_entry_phis: false, kind); |
7648 | |
7649 | /* ??? Prune requirement of these. */ |
7650 | constant_to_value_id = new hash_table<vn_constant_hasher> (23); |
7651 | |
7652 | /* Initialize the value ids and prune out remaining VN_TOPs |
7653 | from dead code. */ |
7654 | tree name; |
7655 | unsigned i; |
7656 | FOR_EACH_SSA_NAME (i, name, cfun) |
7657 | { |
7658 | vn_ssa_aux_t info = VN_INFO (name); |
7659 | if (!info->visited |
7660 | || info->valnum == VN_TOP) |
7661 | info->valnum = name; |
7662 | if (info->valnum == name) |
7663 | info->value_id = get_next_value_id (); |
7664 | else if (is_gimple_min_invariant (info->valnum)) |
7665 | info->value_id = get_or_alloc_constant_value_id (constant: info->valnum); |
7666 | } |
7667 | |
7668 | /* Propagate. */ |
7669 | FOR_EACH_SSA_NAME (i, name, cfun) |
7670 | { |
7671 | vn_ssa_aux_t info = VN_INFO (name); |
7672 | if (TREE_CODE (info->valnum) == SSA_NAME |
7673 | && info->valnum != name |
7674 | && info->value_id != VN_INFO (name: info->valnum)->value_id) |
7675 | info->value_id = VN_INFO (name: info->valnum)->value_id; |
7676 | } |
7677 | |
7678 | set_hashtable_value_ids (); |
7679 | |
7680 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7681 | { |
7682 | fprintf (stream: dump_file, format: "Value numbers:\n" ); |
7683 | FOR_EACH_SSA_NAME (i, name, cfun) |
7684 | { |
7685 | if (VN_INFO (name)->visited |
7686 | && SSA_VAL (x: name) != name) |
7687 | { |
7688 | print_generic_expr (dump_file, name); |
7689 | fprintf (stream: dump_file, format: " = " ); |
7690 | print_generic_expr (dump_file, SSA_VAL (x: name)); |
7691 | fprintf (stream: dump_file, format: " (%04d)\n" , VN_INFO (name)->value_id); |
7692 | } |
7693 | } |
7694 | } |
7695 | } |
7696 | |
7697 | /* Free VN associated data structures. */ |
7698 | |
7699 | void |
7700 | free_rpo_vn (void) |
7701 | { |
7702 | free_vn_table (table: valid_info); |
7703 | XDELETE (valid_info); |
7704 | obstack_free (&vn_tables_obstack, NULL); |
7705 | obstack_free (&vn_tables_insert_obstack, NULL); |
7706 | |
7707 | vn_ssa_aux_iterator_type it; |
7708 | vn_ssa_aux_t info; |
7709 | FOR_EACH_HASH_TABLE_ELEMENT (*vn_ssa_aux_hash, info, vn_ssa_aux_t, it) |
7710 | if (info->needs_insertion) |
7711 | release_ssa_name (name: info->name); |
7712 | obstack_free (&vn_ssa_aux_obstack, NULL); |
7713 | delete vn_ssa_aux_hash; |
7714 | |
7715 | delete constant_to_value_id; |
7716 | constant_to_value_id = NULL; |
7717 | } |
7718 | |
7719 | /* Hook for maybe_push_res_to_seq, lookup the expression in the VN tables. */ |
7720 | |
7721 | static tree |
7722 | vn_lookup_simplify_result (gimple_match_op *res_op) |
7723 | { |
7724 | if (!res_op->code.is_tree_code ()) |
7725 | return NULL_TREE; |
7726 | tree *ops = res_op->ops; |
7727 | unsigned int length = res_op->num_ops; |
7728 | if (res_op->code == CONSTRUCTOR |
7729 | /* ??? We're arriving here with SCCVNs view, decomposed CONSTRUCTOR |
7730 | and GIMPLEs / match-and-simplifies, CONSTRUCTOR as GENERIC tree. */ |
7731 | && TREE_CODE (res_op->ops[0]) == CONSTRUCTOR) |
7732 | { |
7733 | length = CONSTRUCTOR_NELTS (res_op->ops[0]); |
7734 | ops = XALLOCAVEC (tree, length); |
7735 | for (unsigned i = 0; i < length; ++i) |
7736 | ops[i] = CONSTRUCTOR_ELT (res_op->ops[0], i)->value; |
7737 | } |
7738 | vn_nary_op_t vnresult = NULL; |
7739 | tree res = vn_nary_op_lookup_pieces (length, code: (tree_code) res_op->code, |
7740 | type: res_op->type, ops, vnresult: &vnresult); |
7741 | /* If this is used from expression simplification make sure to |
7742 | return an available expression. */ |
7743 | if (res && TREE_CODE (res) == SSA_NAME && mprts_hook && rpo_avail) |
7744 | res = rpo_avail->eliminate_avail (vn_context_bb, op: res); |
7745 | return res; |
7746 | } |
7747 | |
7748 | /* Return a leader for OPs value that is valid at BB. */ |
7749 | |
7750 | tree |
7751 | rpo_elim::eliminate_avail (basic_block bb, tree op) |
7752 | { |
7753 | bool visited; |
7754 | tree valnum = SSA_VAL (x: op, visited: &visited); |
7755 | /* If we didn't visit OP then it must be defined outside of the |
7756 | region we process and also dominate it. So it is available. */ |
7757 | if (!visited) |
7758 | return op; |
7759 | if (TREE_CODE (valnum) == SSA_NAME) |
7760 | { |
7761 | if (SSA_NAME_IS_DEFAULT_DEF (valnum)) |
7762 | return valnum; |
7763 | vn_ssa_aux_t valnum_info = VN_INFO (name: valnum); |
7764 | vn_avail *av = valnum_info->avail; |
7765 | if (!av) |
7766 | { |
7767 | /* See above. But when there's availability info prefer |
7768 | what we recorded there for example to preserve LC SSA. */ |
7769 | if (!valnum_info->visited) |
7770 | return valnum; |
7771 | return NULL_TREE; |
7772 | } |
7773 | if (av->location == bb->index) |
7774 | /* On tramp3d 90% of the cases are here. */ |
7775 | return ssa_name (av->leader); |
7776 | do |
7777 | { |
7778 | basic_block abb = BASIC_BLOCK_FOR_FN (cfun, av->location); |
7779 | /* ??? During elimination we have to use availability at the |
7780 | definition site of a use we try to replace. This |
7781 | is required to not run into inconsistencies because |
7782 | of dominated_by_p_w_unex behavior and removing a definition |
7783 | while not replacing all uses. |
7784 | ??? We could try to consistently walk dominators |
7785 | ignoring non-executable regions. The nearest common |
7786 | dominator of bb and abb is where we can stop walking. We |
7787 | may also be able to "pre-compute" (bits of) the next immediate |
7788 | (non-)dominator during the RPO walk when marking edges as |
7789 | executable. */ |
7790 | if (dominated_by_p_w_unex (bb1: bb, bb2: abb, allow_back: true)) |
7791 | { |
7792 | tree leader = ssa_name (av->leader); |
7793 | /* Prevent eliminations that break loop-closed SSA. */ |
7794 | if (loops_state_satisfies_p (flags: LOOP_CLOSED_SSA) |
7795 | && ! SSA_NAME_IS_DEFAULT_DEF (leader) |
7796 | && ! flow_bb_inside_loop_p (gimple_bb (SSA_NAME_DEF_STMT |
7797 | (leader))->loop_father, |
7798 | bb)) |
7799 | return NULL_TREE; |
7800 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7801 | { |
7802 | print_generic_expr (dump_file, leader); |
7803 | fprintf (stream: dump_file, format: " is available for " ); |
7804 | print_generic_expr (dump_file, valnum); |
7805 | fprintf (stream: dump_file, format: "\n" ); |
7806 | } |
7807 | /* On tramp3d 99% of the _remaining_ cases succeed at |
7808 | the first enty. */ |
7809 | return leader; |
7810 | } |
7811 | /* ??? Can we somehow skip to the immediate dominator |
7812 | RPO index (bb_to_rpo)? Again, maybe not worth, on |
7813 | tramp3d the worst number of elements in the vector is 9. */ |
7814 | av = av->next; |
7815 | } |
7816 | while (av); |
7817 | /* While we prefer avail we have to fallback to using the value |
7818 | directly if defined outside of the region when none of the |
7819 | available defs suit. */ |
7820 | if (!valnum_info->visited) |
7821 | return valnum; |
7822 | } |
7823 | else if (valnum != VN_TOP) |
7824 | /* valnum is is_gimple_min_invariant. */ |
7825 | return valnum; |
7826 | return NULL_TREE; |
7827 | } |
7828 | |
7829 | /* Make LEADER a leader for its value at BB. */ |
7830 | |
7831 | void |
7832 | rpo_elim::eliminate_push_avail (basic_block bb, tree leader) |
7833 | { |
7834 | tree valnum = VN_INFO (name: leader)->valnum; |
7835 | if (valnum == VN_TOP |
7836 | || is_gimple_min_invariant (valnum)) |
7837 | return; |
7838 | if (dump_file && (dump_flags & TDF_DETAILS)) |
7839 | { |
7840 | fprintf (stream: dump_file, format: "Making available beyond BB%d " , bb->index); |
7841 | print_generic_expr (dump_file, leader); |
7842 | fprintf (stream: dump_file, format: " for value " ); |
7843 | print_generic_expr (dump_file, valnum); |
7844 | fprintf (stream: dump_file, format: "\n" ); |
7845 | } |
7846 | vn_ssa_aux_t value = VN_INFO (name: valnum); |
7847 | vn_avail *av; |
7848 | if (m_avail_freelist) |
7849 | { |
7850 | av = m_avail_freelist; |
7851 | m_avail_freelist = m_avail_freelist->next; |
7852 | } |
7853 | else |
7854 | av = XOBNEW (&vn_ssa_aux_obstack, vn_avail); |
7855 | av->location = bb->index; |
7856 | av->leader = SSA_NAME_VERSION (leader); |
7857 | av->next = value->avail; |
7858 | av->next_undo = last_pushed_avail; |
7859 | last_pushed_avail = value; |
7860 | value->avail = av; |
7861 | } |
7862 | |
7863 | /* Valueization hook for RPO VN plus required state. */ |
7864 | |
7865 | tree |
7866 | rpo_vn_valueize (tree name) |
7867 | { |
7868 | if (TREE_CODE (name) == SSA_NAME) |
7869 | { |
7870 | vn_ssa_aux_t val = VN_INFO (name); |
7871 | if (val) |
7872 | { |
7873 | tree tem = val->valnum; |
7874 | if (tem != VN_TOP && tem != name) |
7875 | { |
7876 | if (TREE_CODE (tem) != SSA_NAME) |
7877 | return tem; |
7878 | /* For all values we only valueize to an available leader |
7879 | which means we can use SSA name info without restriction. */ |
7880 | tem = rpo_avail->eliminate_avail (vn_context_bb, op: tem); |
7881 | if (tem) |
7882 | return tem; |
7883 | } |
7884 | } |
7885 | } |
7886 | return name; |
7887 | } |
7888 | |
7889 | /* Insert on PRED_E predicates derived from CODE OPS being true besides the |
7890 | inverted condition. */ |
7891 | |
7892 | static void |
7893 | insert_related_predicates_on_edge (enum tree_code code, tree *ops, edge pred_e) |
7894 | { |
7895 | switch (code) |
7896 | { |
7897 | case LT_EXPR: |
7898 | /* a < b -> a {!,<}= b */ |
7899 | vn_nary_op_insert_pieces_predicated (length: 2, code: NE_EXPR, boolean_type_node, |
7900 | ops, boolean_true_node, value_id: 0, pred_e); |
7901 | vn_nary_op_insert_pieces_predicated (length: 2, code: LE_EXPR, boolean_type_node, |
7902 | ops, boolean_true_node, value_id: 0, pred_e); |
7903 | /* a < b -> ! a {>,=} b */ |
7904 | vn_nary_op_insert_pieces_predicated (length: 2, code: GT_EXPR, boolean_type_node, |
7905 | ops, boolean_false_node, value_id: 0, pred_e); |
7906 | vn_nary_op_insert_pieces_predicated (length: 2, code: EQ_EXPR, boolean_type_node, |
7907 | ops, boolean_false_node, value_id: 0, pred_e); |
7908 | break; |
7909 | case GT_EXPR: |
7910 | /* a > b -> a {!,>}= b */ |
7911 | vn_nary_op_insert_pieces_predicated (length: 2, code: NE_EXPR, boolean_type_node, |
7912 | ops, boolean_true_node, value_id: 0, pred_e); |
7913 | vn_nary_op_insert_pieces_predicated (length: 2, code: GE_EXPR, boolean_type_node, |
7914 | ops, boolean_true_node, value_id: 0, pred_e); |
7915 | /* a > b -> ! a {<,=} b */ |
7916 | vn_nary_op_insert_pieces_predicated (length: 2, code: LT_EXPR, boolean_type_node, |
7917 | ops, boolean_false_node, value_id: 0, pred_e); |
7918 | vn_nary_op_insert_pieces_predicated (length: 2, code: EQ_EXPR, boolean_type_node, |
7919 | ops, boolean_false_node, value_id: 0, pred_e); |
7920 | break; |
7921 | case EQ_EXPR: |
7922 | /* a == b -> ! a {<,>} b */ |
7923 | vn_nary_op_insert_pieces_predicated (length: 2, code: LT_EXPR, boolean_type_node, |
7924 | ops, boolean_false_node, value_id: 0, pred_e); |
7925 | vn_nary_op_insert_pieces_predicated (length: 2, code: GT_EXPR, boolean_type_node, |
7926 | ops, boolean_false_node, value_id: 0, pred_e); |
7927 | break; |
7928 | case LE_EXPR: |
7929 | case GE_EXPR: |
7930 | case NE_EXPR: |
7931 | /* Nothing besides inverted condition. */ |
7932 | break; |
7933 | default:; |
7934 | } |
7935 | } |
7936 | |
7937 | /* Main stmt worker for RPO VN, process BB. */ |
7938 | |
7939 | static unsigned |
7940 | process_bb (rpo_elim &avail, basic_block bb, |
7941 | bool bb_visited, bool iterate_phis, bool iterate, bool eliminate, |
7942 | bool do_region, bitmap exit_bbs, bool skip_phis) |
7943 | { |
7944 | unsigned todo = 0; |
7945 | edge_iterator ei; |
7946 | edge e; |
7947 | |
7948 | vn_context_bb = bb; |
7949 | |
7950 | /* If we are in loop-closed SSA preserve this state. This is |
7951 | relevant when called on regions from outside of FRE/PRE. */ |
7952 | bool lc_phi_nodes = false; |
7953 | if (!skip_phis |
7954 | && loops_state_satisfies_p (flags: LOOP_CLOSED_SSA)) |
7955 | FOR_EACH_EDGE (e, ei, bb->preds) |
7956 | if (e->src->loop_father != e->dest->loop_father |
7957 | && flow_loop_nested_p (e->dest->loop_father, |
7958 | e->src->loop_father)) |
7959 | { |
7960 | lc_phi_nodes = true; |
7961 | break; |
7962 | } |
7963 | |
7964 | /* When we visit a loop header substitute into loop info. */ |
7965 | if (!iterate && eliminate && bb->loop_father->header == bb) |
7966 | { |
7967 | /* Keep fields in sync with substitute_in_loop_info. */ |
7968 | if (bb->loop_father->nb_iterations) |
7969 | bb->loop_father->nb_iterations |
7970 | = simplify_replace_tree (bb->loop_father->nb_iterations, |
7971 | NULL_TREE, NULL_TREE, &vn_valueize_for_srt); |
7972 | } |
7973 | |
7974 | /* Value-number all defs in the basic-block. */ |
7975 | if (!skip_phis) |
7976 | for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (i: gsi); |
7977 | gsi_next (i: &gsi)) |
7978 | { |
7979 | gphi *phi = gsi.phi (); |
7980 | tree res = PHI_RESULT (phi); |
7981 | vn_ssa_aux_t res_info = VN_INFO (name: res); |
7982 | if (!bb_visited) |
7983 | { |
7984 | gcc_assert (!res_info->visited); |
7985 | res_info->valnum = VN_TOP; |
7986 | res_info->visited = true; |
7987 | } |
7988 | |
7989 | /* When not iterating force backedge values to varying. */ |
7990 | visit_stmt (stmt: phi, backedges_varying_p: !iterate_phis); |
7991 | if (virtual_operand_p (op: res)) |
7992 | continue; |
7993 | |
7994 | /* Eliminate */ |
7995 | /* The interesting case is gcc.dg/tree-ssa/pr22230.c for correctness |
7996 | how we handle backedges and availability. |
7997 | And gcc.dg/tree-ssa/ssa-sccvn-2.c for optimization. */ |
7998 | tree val = res_info->valnum; |
7999 | if (res != val && !iterate && eliminate) |
8000 | { |
8001 | if (tree leader = avail.eliminate_avail (bb, op: res)) |
8002 | { |
8003 | if (leader != res |
8004 | /* Preserve loop-closed SSA form. */ |
8005 | && (! lc_phi_nodes |
8006 | || is_gimple_min_invariant (leader))) |
8007 | { |
8008 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8009 | { |
8010 | fprintf (stream: dump_file, format: "Replaced redundant PHI node " |
8011 | "defining " ); |
8012 | print_generic_expr (dump_file, res); |
8013 | fprintf (stream: dump_file, format: " with " ); |
8014 | print_generic_expr (dump_file, leader); |
8015 | fprintf (stream: dump_file, format: "\n" ); |
8016 | } |
8017 | avail.eliminations++; |
8018 | |
8019 | if (may_propagate_copy (res, leader)) |
8020 | { |
8021 | /* Schedule for removal. */ |
8022 | avail.to_remove.safe_push (obj: phi); |
8023 | continue; |
8024 | } |
8025 | /* ??? Else generate a copy stmt. */ |
8026 | } |
8027 | } |
8028 | } |
8029 | /* Only make defs available that not already are. But make |
8030 | sure loop-closed SSA PHI node defs are picked up for |
8031 | downstream uses. */ |
8032 | if (lc_phi_nodes |
8033 | || res == val |
8034 | || ! avail.eliminate_avail (bb, op: res)) |
8035 | avail.eliminate_push_avail (bb, leader: res); |
8036 | } |
8037 | |
8038 | /* For empty BBs mark outgoing edges executable. For non-empty BBs |
8039 | we do this when processing the last stmt as we have to do this |
8040 | before elimination which otherwise forces GIMPLE_CONDs to |
8041 | if (1 != 0) style when seeing non-executable edges. */ |
8042 | if (gsi_end_p (i: gsi_start_bb (bb))) |
8043 | { |
8044 | FOR_EACH_EDGE (e, ei, bb->succs) |
8045 | { |
8046 | if (!(e->flags & EDGE_EXECUTABLE)) |
8047 | { |
8048 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8049 | fprintf (stream: dump_file, |
8050 | format: "marking outgoing edge %d -> %d executable\n" , |
8051 | e->src->index, e->dest->index); |
8052 | e->flags |= EDGE_EXECUTABLE; |
8053 | e->dest->flags |= BB_EXECUTABLE; |
8054 | } |
8055 | else if (!(e->dest->flags & BB_EXECUTABLE)) |
8056 | { |
8057 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8058 | fprintf (stream: dump_file, |
8059 | format: "marking destination block %d reachable\n" , |
8060 | e->dest->index); |
8061 | e->dest->flags |= BB_EXECUTABLE; |
8062 | } |
8063 | } |
8064 | } |
8065 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); |
8066 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
8067 | { |
8068 | ssa_op_iter i; |
8069 | tree op; |
8070 | if (!bb_visited) |
8071 | { |
8072 | FOR_EACH_SSA_TREE_OPERAND (op, gsi_stmt (gsi), i, SSA_OP_ALL_DEFS) |
8073 | { |
8074 | vn_ssa_aux_t op_info = VN_INFO (name: op); |
8075 | gcc_assert (!op_info->visited); |
8076 | op_info->valnum = VN_TOP; |
8077 | op_info->visited = true; |
8078 | } |
8079 | |
8080 | /* We somehow have to deal with uses that are not defined |
8081 | in the processed region. Forcing unvisited uses to |
8082 | varying here doesn't play well with def-use following during |
8083 | expression simplification, so we deal with this by checking |
8084 | the visited flag in SSA_VAL. */ |
8085 | } |
8086 | |
8087 | visit_stmt (stmt: gsi_stmt (i: gsi)); |
8088 | |
8089 | gimple *last = gsi_stmt (i: gsi); |
8090 | e = NULL; |
8091 | switch (gimple_code (g: last)) |
8092 | { |
8093 | case GIMPLE_SWITCH: |
8094 | e = find_taken_edge (bb, vn_valueize (gimple_switch_index |
8095 | (gs: as_a <gswitch *> (p: last)))); |
8096 | break; |
8097 | case GIMPLE_COND: |
8098 | { |
8099 | tree lhs = vn_valueize (gimple_cond_lhs (gs: last)); |
8100 | tree rhs = vn_valueize (gimple_cond_rhs (gs: last)); |
8101 | tree val = gimple_simplify (gimple_cond_code (gs: last), |
8102 | boolean_type_node, lhs, rhs, |
8103 | NULL, vn_valueize); |
8104 | /* If the condition didn't simplfy see if we have recorded |
8105 | an expression from sofar taken edges. */ |
8106 | if (! val || TREE_CODE (val) != INTEGER_CST) |
8107 | { |
8108 | vn_nary_op_t vnresult; |
8109 | tree ops[2]; |
8110 | ops[0] = lhs; |
8111 | ops[1] = rhs; |
8112 | val = vn_nary_op_lookup_pieces (length: 2, code: gimple_cond_code (gs: last), |
8113 | boolean_type_node, ops, |
8114 | vnresult: &vnresult); |
8115 | /* Did we get a predicated value? */ |
8116 | if (! val && vnresult && vnresult->predicated_values) |
8117 | { |
8118 | val = vn_nary_op_get_predicated_value (vno: vnresult, bb); |
8119 | if (val && dump_file && (dump_flags & TDF_DETAILS)) |
8120 | { |
8121 | fprintf (stream: dump_file, format: "Got predicated value " ); |
8122 | print_generic_expr (dump_file, val, TDF_NONE); |
8123 | fprintf (stream: dump_file, format: " for " ); |
8124 | print_gimple_stmt (dump_file, last, TDF_SLIM); |
8125 | } |
8126 | } |
8127 | } |
8128 | if (val) |
8129 | e = find_taken_edge (bb, val); |
8130 | if (! e) |
8131 | { |
8132 | /* If we didn't manage to compute the taken edge then |
8133 | push predicated expressions for the condition itself |
8134 | and related conditions to the hashtables. This allows |
8135 | simplification of redundant conditions which is |
8136 | important as early cleanup. */ |
8137 | edge true_e, false_e; |
8138 | extract_true_false_edges_from_block (bb, &true_e, &false_e); |
8139 | enum tree_code code = gimple_cond_code (gs: last); |
8140 | enum tree_code icode |
8141 | = invert_tree_comparison (code, HONOR_NANS (lhs)); |
8142 | tree ops[2]; |
8143 | ops[0] = lhs; |
8144 | ops[1] = rhs; |
8145 | if ((do_region && bitmap_bit_p (exit_bbs, true_e->dest->index)) |
8146 | || !can_track_predicate_on_edge (pred_e: true_e)) |
8147 | true_e = NULL; |
8148 | if ((do_region && bitmap_bit_p (exit_bbs, false_e->dest->index)) |
8149 | || !can_track_predicate_on_edge (pred_e: false_e)) |
8150 | false_e = NULL; |
8151 | if (true_e) |
8152 | vn_nary_op_insert_pieces_predicated |
8153 | (length: 2, code, boolean_type_node, ops, |
8154 | boolean_true_node, value_id: 0, pred_e: true_e); |
8155 | if (false_e) |
8156 | vn_nary_op_insert_pieces_predicated |
8157 | (length: 2, code, boolean_type_node, ops, |
8158 | boolean_false_node, value_id: 0, pred_e: false_e); |
8159 | if (icode != ERROR_MARK) |
8160 | { |
8161 | if (true_e) |
8162 | vn_nary_op_insert_pieces_predicated |
8163 | (length: 2, code: icode, boolean_type_node, ops, |
8164 | boolean_false_node, value_id: 0, pred_e: true_e); |
8165 | if (false_e) |
8166 | vn_nary_op_insert_pieces_predicated |
8167 | (length: 2, code: icode, boolean_type_node, ops, |
8168 | boolean_true_node, value_id: 0, pred_e: false_e); |
8169 | } |
8170 | /* Relax for non-integers, inverted condition handled |
8171 | above. */ |
8172 | if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))) |
8173 | { |
8174 | if (true_e) |
8175 | insert_related_predicates_on_edge (code, ops, pred_e: true_e); |
8176 | if (false_e) |
8177 | insert_related_predicates_on_edge (code: icode, ops, pred_e: false_e); |
8178 | } |
8179 | } |
8180 | break; |
8181 | } |
8182 | case GIMPLE_GOTO: |
8183 | e = find_taken_edge (bb, vn_valueize (gimple_goto_dest (gs: last))); |
8184 | break; |
8185 | default: |
8186 | e = NULL; |
8187 | } |
8188 | if (e) |
8189 | { |
8190 | todo = TODO_cleanup_cfg; |
8191 | if (!(e->flags & EDGE_EXECUTABLE)) |
8192 | { |
8193 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8194 | fprintf (stream: dump_file, |
8195 | format: "marking known outgoing %sedge %d -> %d executable\n" , |
8196 | e->flags & EDGE_DFS_BACK ? "back-" : "" , |
8197 | e->src->index, e->dest->index); |
8198 | e->flags |= EDGE_EXECUTABLE; |
8199 | e->dest->flags |= BB_EXECUTABLE; |
8200 | } |
8201 | else if (!(e->dest->flags & BB_EXECUTABLE)) |
8202 | { |
8203 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8204 | fprintf (stream: dump_file, |
8205 | format: "marking destination block %d reachable\n" , |
8206 | e->dest->index); |
8207 | e->dest->flags |= BB_EXECUTABLE; |
8208 | } |
8209 | } |
8210 | else if (gsi_one_before_end_p (i: gsi)) |
8211 | { |
8212 | FOR_EACH_EDGE (e, ei, bb->succs) |
8213 | { |
8214 | if (!(e->flags & EDGE_EXECUTABLE)) |
8215 | { |
8216 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8217 | fprintf (stream: dump_file, |
8218 | format: "marking outgoing edge %d -> %d executable\n" , |
8219 | e->src->index, e->dest->index); |
8220 | e->flags |= EDGE_EXECUTABLE; |
8221 | e->dest->flags |= BB_EXECUTABLE; |
8222 | } |
8223 | else if (!(e->dest->flags & BB_EXECUTABLE)) |
8224 | { |
8225 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8226 | fprintf (stream: dump_file, |
8227 | format: "marking destination block %d reachable\n" , |
8228 | e->dest->index); |
8229 | e->dest->flags |= BB_EXECUTABLE; |
8230 | } |
8231 | } |
8232 | } |
8233 | |
8234 | /* Eliminate. That also pushes to avail. */ |
8235 | if (eliminate && ! iterate) |
8236 | avail.eliminate_stmt (b: bb, gsi: &gsi); |
8237 | else |
8238 | /* If not eliminating, make all not already available defs |
8239 | available. But avoid picking up dead defs. */ |
8240 | FOR_EACH_SSA_TREE_OPERAND (op, gsi_stmt (gsi), i, SSA_OP_DEF) |
8241 | if (! has_zero_uses (var: op) |
8242 | && ! avail.eliminate_avail (bb, op)) |
8243 | avail.eliminate_push_avail (bb, leader: op); |
8244 | } |
8245 | |
8246 | /* Eliminate in destination PHI arguments. Always substitute in dest |
8247 | PHIs, even for non-executable edges. This handles region |
8248 | exits PHIs. */ |
8249 | if (!iterate && eliminate) |
8250 | FOR_EACH_EDGE (e, ei, bb->succs) |
8251 | for (gphi_iterator gsi = gsi_start_phis (e->dest); |
8252 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
8253 | { |
8254 | gphi *phi = gsi.phi (); |
8255 | use_operand_p use_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, e); |
8256 | tree arg = USE_FROM_PTR (use_p); |
8257 | if (TREE_CODE (arg) != SSA_NAME |
8258 | || virtual_operand_p (op: arg)) |
8259 | continue; |
8260 | tree sprime; |
8261 | if (SSA_NAME_IS_DEFAULT_DEF (arg)) |
8262 | { |
8263 | sprime = SSA_VAL (x: arg); |
8264 | gcc_assert (TREE_CODE (sprime) != SSA_NAME |
8265 | || SSA_NAME_IS_DEFAULT_DEF (sprime)); |
8266 | } |
8267 | else |
8268 | /* Look for sth available at the definition block of the argument. |
8269 | This avoids inconsistencies between availability there which |
8270 | decides if the stmt can be removed and availability at the |
8271 | use site. The SSA property ensures that things available |
8272 | at the definition are also available at uses. */ |
8273 | sprime = avail.eliminate_avail (bb: gimple_bb (SSA_NAME_DEF_STMT (arg)), |
8274 | op: arg); |
8275 | if (sprime |
8276 | && sprime != arg |
8277 | && may_propagate_copy (arg, sprime, !(e->flags & EDGE_ABNORMAL))) |
8278 | propagate_value (use_p, sprime); |
8279 | } |
8280 | |
8281 | vn_context_bb = NULL; |
8282 | return todo; |
8283 | } |
8284 | |
8285 | /* Unwind state per basic-block. */ |
8286 | |
8287 | struct unwind_state |
8288 | { |
8289 | /* Times this block has been visited. */ |
8290 | unsigned visited; |
8291 | /* Whether to handle this as iteration point or whether to treat |
8292 | incoming backedge PHI values as varying. */ |
8293 | bool iterate; |
8294 | /* Maximum RPO index this block is reachable from. */ |
8295 | int max_rpo; |
8296 | /* Unwind state. */ |
8297 | void *ob_top; |
8298 | vn_reference_t ref_top; |
8299 | vn_phi_t phi_top; |
8300 | vn_nary_op_t nary_top; |
8301 | vn_avail *avail_top; |
8302 | }; |
8303 | |
8304 | /* Unwind the RPO VN state for iteration. */ |
8305 | |
8306 | static void |
8307 | do_unwind (unwind_state *to, rpo_elim &avail) |
8308 | { |
8309 | gcc_assert (to->iterate); |
8310 | for (; last_inserted_nary != to->nary_top; |
8311 | last_inserted_nary = last_inserted_nary->next) |
8312 | { |
8313 | vn_nary_op_t *slot; |
8314 | slot = valid_info->nary->find_slot_with_hash |
8315 | (comparable: last_inserted_nary, hash: last_inserted_nary->hashcode, insert: NO_INSERT); |
8316 | /* Predication causes the need to restore previous state. */ |
8317 | if ((*slot)->unwind_to) |
8318 | *slot = (*slot)->unwind_to; |
8319 | else |
8320 | valid_info->nary->clear_slot (slot); |
8321 | } |
8322 | for (; last_inserted_phi != to->phi_top; |
8323 | last_inserted_phi = last_inserted_phi->next) |
8324 | { |
8325 | vn_phi_t *slot; |
8326 | slot = valid_info->phis->find_slot_with_hash |
8327 | (comparable: last_inserted_phi, hash: last_inserted_phi->hashcode, insert: NO_INSERT); |
8328 | valid_info->phis->clear_slot (slot); |
8329 | } |
8330 | for (; last_inserted_ref != to->ref_top; |
8331 | last_inserted_ref = last_inserted_ref->next) |
8332 | { |
8333 | vn_reference_t *slot; |
8334 | slot = valid_info->references->find_slot_with_hash |
8335 | (comparable: last_inserted_ref, hash: last_inserted_ref->hashcode, insert: NO_INSERT); |
8336 | (*slot)->operands.release (); |
8337 | valid_info->references->clear_slot (slot); |
8338 | } |
8339 | obstack_free (&vn_tables_obstack, to->ob_top); |
8340 | |
8341 | /* Prune [rpo_idx, ] from avail. */ |
8342 | for (; last_pushed_avail && last_pushed_avail->avail != to->avail_top;) |
8343 | { |
8344 | vn_ssa_aux_t val = last_pushed_avail; |
8345 | vn_avail *av = val->avail; |
8346 | val->avail = av->next; |
8347 | last_pushed_avail = av->next_undo; |
8348 | av->next = avail.m_avail_freelist; |
8349 | avail.m_avail_freelist = av; |
8350 | } |
8351 | } |
8352 | |
8353 | /* Do VN on a SEME region specified by ENTRY and EXIT_BBS in FN. |
8354 | If ITERATE is true then treat backedges optimistically as not |
8355 | executed and iterate. If ELIMINATE is true then perform |
8356 | elimination, otherwise leave that to the caller. If SKIP_ENTRY_PHIS |
8357 | is true then force PHI nodes in ENTRY->dest to VARYING. */ |
8358 | |
8359 | static unsigned |
8360 | do_rpo_vn_1 (function *fn, edge entry, bitmap exit_bbs, |
8361 | bool iterate, bool eliminate, bool skip_entry_phis, |
8362 | vn_lookup_kind kind) |
8363 | { |
8364 | unsigned todo = 0; |
8365 | default_vn_walk_kind = kind; |
8366 | |
8367 | /* We currently do not support region-based iteration when |
8368 | elimination is requested. */ |
8369 | gcc_assert (!entry || !iterate || !eliminate); |
8370 | /* When iterating we need loop info up-to-date. */ |
8371 | gcc_assert (!iterate || !loops_state_satisfies_p (LOOPS_NEED_FIXUP)); |
8372 | |
8373 | bool do_region = entry != NULL; |
8374 | if (!do_region) |
8375 | { |
8376 | entry = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (fn)); |
8377 | exit_bbs = BITMAP_ALLOC (NULL); |
8378 | bitmap_set_bit (exit_bbs, EXIT_BLOCK); |
8379 | } |
8380 | |
8381 | /* Clear EDGE_DFS_BACK on "all" entry edges, RPO order compute will |
8382 | re-mark those that are contained in the region. */ |
8383 | edge_iterator ei; |
8384 | edge e; |
8385 | FOR_EACH_EDGE (e, ei, entry->dest->preds) |
8386 | e->flags &= ~EDGE_DFS_BACK; |
8387 | |
8388 | int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (fn) - NUM_FIXED_BLOCKS); |
8389 | auto_vec<std::pair<int, int> > toplevel_scc_extents; |
8390 | int n = rev_post_order_and_mark_dfs_back_seme |
8391 | (fn, entry, exit_bbs, true, rpo, !iterate ? &toplevel_scc_extents : NULL); |
8392 | |
8393 | if (!do_region) |
8394 | BITMAP_FREE (exit_bbs); |
8395 | |
8396 | /* If there are any non-DFS_BACK edges into entry->dest skip |
8397 | processing PHI nodes for that block. This supports |
8398 | value-numbering loop bodies w/o the actual loop. */ |
8399 | FOR_EACH_EDGE (e, ei, entry->dest->preds) |
8400 | if (e != entry |
8401 | && !(e->flags & EDGE_DFS_BACK)) |
8402 | break; |
8403 | if (e != NULL && dump_file && (dump_flags & TDF_DETAILS)) |
8404 | fprintf (stream: dump_file, format: "Region does not contain all edges into " |
8405 | "the entry block, skipping its PHIs.\n" ); |
8406 | skip_entry_phis |= e != NULL; |
8407 | |
8408 | int *bb_to_rpo = XNEWVEC (int, last_basic_block_for_fn (fn)); |
8409 | for (int i = 0; i < n; ++i) |
8410 | bb_to_rpo[rpo[i]] = i; |
8411 | |
8412 | unwind_state *rpo_state = XNEWVEC (unwind_state, n); |
8413 | |
8414 | rpo_elim avail (entry->dest); |
8415 | rpo_avail = &avail; |
8416 | |
8417 | /* Verify we have no extra entries into the region. */ |
8418 | if (flag_checking && do_region) |
8419 | { |
8420 | auto_bb_flag bb_in_region (fn); |
8421 | for (int i = 0; i < n; ++i) |
8422 | { |
8423 | basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]); |
8424 | bb->flags |= bb_in_region; |
8425 | } |
8426 | /* We can't merge the first two loops because we cannot rely |
8427 | on EDGE_DFS_BACK for edges not within the region. But if |
8428 | we decide to always have the bb_in_region flag we can |
8429 | do the checking during the RPO walk itself (but then it's |
8430 | also easy to handle MEME conservatively). */ |
8431 | for (int i = 0; i < n; ++i) |
8432 | { |
8433 | basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]); |
8434 | edge e; |
8435 | edge_iterator ei; |
8436 | FOR_EACH_EDGE (e, ei, bb->preds) |
8437 | gcc_assert (e == entry |
8438 | || (skip_entry_phis && bb == entry->dest) |
8439 | || (e->src->flags & bb_in_region)); |
8440 | } |
8441 | for (int i = 0; i < n; ++i) |
8442 | { |
8443 | basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]); |
8444 | bb->flags &= ~bb_in_region; |
8445 | } |
8446 | } |
8447 | |
8448 | /* Create the VN state. For the initial size of the various hashtables |
8449 | use a heuristic based on region size and number of SSA names. */ |
8450 | unsigned region_size = (((unsigned HOST_WIDE_INT)n * num_ssa_names) |
8451 | / (n_basic_blocks_for_fn (fn) - NUM_FIXED_BLOCKS)); |
8452 | VN_TOP = create_tmp_var_raw (void_type_node, "vn_top" ); |
8453 | next_value_id = 1; |
8454 | next_constant_value_id = -1; |
8455 | |
8456 | vn_ssa_aux_hash = new hash_table <vn_ssa_aux_hasher> (region_size * 2); |
8457 | gcc_obstack_init (&vn_ssa_aux_obstack); |
8458 | |
8459 | gcc_obstack_init (&vn_tables_obstack); |
8460 | gcc_obstack_init (&vn_tables_insert_obstack); |
8461 | valid_info = XCNEW (struct vn_tables_s); |
8462 | allocate_vn_table (table: valid_info, size: region_size); |
8463 | last_inserted_ref = NULL; |
8464 | last_inserted_phi = NULL; |
8465 | last_inserted_nary = NULL; |
8466 | last_pushed_avail = NULL; |
8467 | |
8468 | vn_valueize = rpo_vn_valueize; |
8469 | |
8470 | /* Initialize the unwind state and edge/BB executable state. */ |
8471 | unsigned curr_scc = 0; |
8472 | for (int i = 0; i < n; ++i) |
8473 | { |
8474 | basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]); |
8475 | rpo_state[i].visited = 0; |
8476 | rpo_state[i].max_rpo = i; |
8477 | if (!iterate && curr_scc < toplevel_scc_extents.length ()) |
8478 | { |
8479 | if (i >= toplevel_scc_extents[curr_scc].first |
8480 | && i <= toplevel_scc_extents[curr_scc].second) |
8481 | rpo_state[i].max_rpo = toplevel_scc_extents[curr_scc].second; |
8482 | if (i == toplevel_scc_extents[curr_scc].second) |
8483 | curr_scc++; |
8484 | } |
8485 | bb->flags &= ~BB_EXECUTABLE; |
8486 | bool has_backedges = false; |
8487 | edge e; |
8488 | edge_iterator ei; |
8489 | FOR_EACH_EDGE (e, ei, bb->preds) |
8490 | { |
8491 | if (e->flags & EDGE_DFS_BACK) |
8492 | has_backedges = true; |
8493 | e->flags &= ~EDGE_EXECUTABLE; |
8494 | if (iterate || e == entry || (skip_entry_phis && bb == entry->dest)) |
8495 | continue; |
8496 | } |
8497 | rpo_state[i].iterate = iterate && has_backedges; |
8498 | } |
8499 | entry->flags |= EDGE_EXECUTABLE; |
8500 | entry->dest->flags |= BB_EXECUTABLE; |
8501 | |
8502 | /* As heuristic to improve compile-time we handle only the N innermost |
8503 | loops and the outermost one optimistically. */ |
8504 | if (iterate) |
8505 | { |
8506 | unsigned max_depth = param_rpo_vn_max_loop_depth; |
8507 | for (auto loop : loops_list (cfun, LI_ONLY_INNERMOST)) |
8508 | if (loop_depth (loop) > max_depth) |
8509 | for (unsigned i = 2; |
8510 | i < loop_depth (loop) - max_depth; ++i) |
8511 | { |
8512 | basic_block = superloop_at_depth (loop, i)->header; |
8513 | bool non_latch_backedge = false; |
8514 | edge e; |
8515 | edge_iterator ei; |
8516 | FOR_EACH_EDGE (e, ei, header->preds) |
8517 | if (e->flags & EDGE_DFS_BACK) |
8518 | { |
8519 | /* There can be a non-latch backedge into the header |
8520 | which is part of an outer irreducible region. We |
8521 | cannot avoid iterating this block then. */ |
8522 | if (!dominated_by_p (CDI_DOMINATORS, |
8523 | e->src, e->dest)) |
8524 | { |
8525 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8526 | fprintf (stream: dump_file, format: "non-latch backedge %d -> %d " |
8527 | "forces iteration of loop %d\n" , |
8528 | e->src->index, e->dest->index, loop->num); |
8529 | non_latch_backedge = true; |
8530 | } |
8531 | else |
8532 | e->flags |= EDGE_EXECUTABLE; |
8533 | } |
8534 | rpo_state[bb_to_rpo[header->index]].iterate = non_latch_backedge; |
8535 | } |
8536 | } |
8537 | |
8538 | uint64_t nblk = 0; |
8539 | int idx = 0; |
8540 | if (iterate) |
8541 | /* Go and process all blocks, iterating as necessary. */ |
8542 | do |
8543 | { |
8544 | basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[idx]); |
8545 | |
8546 | /* If the block has incoming backedges remember unwind state. This |
8547 | is required even for non-executable blocks since in irreducible |
8548 | regions we might reach them via the backedge and re-start iterating |
8549 | from there. |
8550 | Note we can individually mark blocks with incoming backedges to |
8551 | not iterate where we then handle PHIs conservatively. We do that |
8552 | heuristically to reduce compile-time for degenerate cases. */ |
8553 | if (rpo_state[idx].iterate) |
8554 | { |
8555 | rpo_state[idx].ob_top = obstack_alloc (&vn_tables_obstack, 0); |
8556 | rpo_state[idx].ref_top = last_inserted_ref; |
8557 | rpo_state[idx].phi_top = last_inserted_phi; |
8558 | rpo_state[idx].nary_top = last_inserted_nary; |
8559 | rpo_state[idx].avail_top |
8560 | = last_pushed_avail ? last_pushed_avail->avail : NULL; |
8561 | } |
8562 | |
8563 | if (!(bb->flags & BB_EXECUTABLE)) |
8564 | { |
8565 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8566 | fprintf (stream: dump_file, format: "Block %d: BB%d found not executable\n" , |
8567 | idx, bb->index); |
8568 | idx++; |
8569 | continue; |
8570 | } |
8571 | |
8572 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8573 | fprintf (stream: dump_file, format: "Processing block %d: BB%d\n" , idx, bb->index); |
8574 | nblk++; |
8575 | todo |= process_bb (avail, bb, |
8576 | bb_visited: rpo_state[idx].visited != 0, |
8577 | iterate_phis: rpo_state[idx].iterate, |
8578 | iterate, eliminate, do_region, exit_bbs, skip_phis: false); |
8579 | rpo_state[idx].visited++; |
8580 | |
8581 | /* Verify if changed values flow over executable outgoing backedges |
8582 | and those change destination PHI values (that's the thing we |
8583 | can easily verify). Reduce over all such edges to the farthest |
8584 | away PHI. */ |
8585 | int iterate_to = -1; |
8586 | edge_iterator ei; |
8587 | edge e; |
8588 | FOR_EACH_EDGE (e, ei, bb->succs) |
8589 | if ((e->flags & (EDGE_DFS_BACK|EDGE_EXECUTABLE)) |
8590 | == (EDGE_DFS_BACK|EDGE_EXECUTABLE) |
8591 | && rpo_state[bb_to_rpo[e->dest->index]].iterate) |
8592 | { |
8593 | int destidx = bb_to_rpo[e->dest->index]; |
8594 | if (!rpo_state[destidx].visited) |
8595 | { |
8596 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8597 | fprintf (stream: dump_file, format: "Unvisited destination %d\n" , |
8598 | e->dest->index); |
8599 | if (iterate_to == -1 || destidx < iterate_to) |
8600 | iterate_to = destidx; |
8601 | continue; |
8602 | } |
8603 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8604 | fprintf (stream: dump_file, format: "Looking for changed values of backedge" |
8605 | " %d->%d destination PHIs\n" , |
8606 | e->src->index, e->dest->index); |
8607 | vn_context_bb = e->dest; |
8608 | gphi_iterator gsi; |
8609 | for (gsi = gsi_start_phis (e->dest); |
8610 | !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
8611 | { |
8612 | bool inserted = false; |
8613 | /* While we'd ideally just iterate on value changes |
8614 | we CSE PHIs and do that even across basic-block |
8615 | boundaries. So even hashtable state changes can |
8616 | be important (which is roughly equivalent to |
8617 | PHI argument value changes). To not excessively |
8618 | iterate because of that we track whether a PHI |
8619 | was CSEd to with GF_PLF_1. */ |
8620 | bool phival_changed; |
8621 | if ((phival_changed = visit_phi (phi: gsi.phi (), |
8622 | inserted: &inserted, backedges_varying_p: false)) |
8623 | || (inserted && gimple_plf (stmt: gsi.phi (), plf: GF_PLF_1))) |
8624 | { |
8625 | if (!phival_changed |
8626 | && dump_file && (dump_flags & TDF_DETAILS)) |
8627 | fprintf (stream: dump_file, format: "PHI was CSEd and hashtable " |
8628 | "state (changed)\n" ); |
8629 | if (iterate_to == -1 || destidx < iterate_to) |
8630 | iterate_to = destidx; |
8631 | break; |
8632 | } |
8633 | } |
8634 | vn_context_bb = NULL; |
8635 | } |
8636 | if (iterate_to != -1) |
8637 | { |
8638 | do_unwind (to: &rpo_state[iterate_to], avail); |
8639 | idx = iterate_to; |
8640 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8641 | fprintf (stream: dump_file, format: "Iterating to %d BB%d\n" , |
8642 | iterate_to, rpo[iterate_to]); |
8643 | continue; |
8644 | } |
8645 | |
8646 | idx++; |
8647 | } |
8648 | while (idx < n); |
8649 | |
8650 | else /* !iterate */ |
8651 | { |
8652 | /* Process all blocks greedily with a worklist that enforces RPO |
8653 | processing of reachable blocks. */ |
8654 | auto_bitmap worklist; |
8655 | bitmap_set_bit (worklist, 0); |
8656 | while (!bitmap_empty_p (map: worklist)) |
8657 | { |
8658 | int idx = bitmap_clear_first_set_bit (worklist); |
8659 | basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[idx]); |
8660 | gcc_assert ((bb->flags & BB_EXECUTABLE) |
8661 | && !rpo_state[idx].visited); |
8662 | |
8663 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8664 | fprintf (stream: dump_file, format: "Processing block %d: BB%d\n" , idx, bb->index); |
8665 | |
8666 | /* When we run into predecessor edges where we cannot trust its |
8667 | executable state mark them executable so PHI processing will |
8668 | be conservative. |
8669 | ??? Do we need to force arguments flowing over that edge |
8670 | to be varying or will they even always be? */ |
8671 | edge_iterator ei; |
8672 | edge e; |
8673 | FOR_EACH_EDGE (e, ei, bb->preds) |
8674 | if (!(e->flags & EDGE_EXECUTABLE) |
8675 | && (bb == entry->dest |
8676 | || (!rpo_state[bb_to_rpo[e->src->index]].visited |
8677 | && (rpo_state[bb_to_rpo[e->src->index]].max_rpo |
8678 | >= (int)idx)))) |
8679 | { |
8680 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8681 | fprintf (stream: dump_file, format: "Cannot trust state of predecessor " |
8682 | "edge %d -> %d, marking executable\n" , |
8683 | e->src->index, e->dest->index); |
8684 | e->flags |= EDGE_EXECUTABLE; |
8685 | } |
8686 | |
8687 | nblk++; |
8688 | todo |= process_bb (avail, bb, bb_visited: false, iterate_phis: false, iterate: false, eliminate, |
8689 | do_region, exit_bbs, |
8690 | skip_phis: skip_entry_phis && bb == entry->dest); |
8691 | rpo_state[idx].visited++; |
8692 | |
8693 | FOR_EACH_EDGE (e, ei, bb->succs) |
8694 | if ((e->flags & EDGE_EXECUTABLE) |
8695 | && e->dest->index != EXIT_BLOCK |
8696 | && (!do_region || !bitmap_bit_p (exit_bbs, e->dest->index)) |
8697 | && !rpo_state[bb_to_rpo[e->dest->index]].visited) |
8698 | bitmap_set_bit (worklist, bb_to_rpo[e->dest->index]); |
8699 | } |
8700 | } |
8701 | |
8702 | /* If statistics or dump file active. */ |
8703 | int nex = 0; |
8704 | unsigned max_visited = 1; |
8705 | for (int i = 0; i < n; ++i) |
8706 | { |
8707 | basic_block bb = BASIC_BLOCK_FOR_FN (fn, rpo[i]); |
8708 | if (bb->flags & BB_EXECUTABLE) |
8709 | nex++; |
8710 | statistics_histogram_event (cfun, "RPO block visited times" , |
8711 | rpo_state[i].visited); |
8712 | if (rpo_state[i].visited > max_visited) |
8713 | max_visited = rpo_state[i].visited; |
8714 | } |
8715 | unsigned nvalues = 0, navail = 0; |
8716 | for (hash_table<vn_ssa_aux_hasher>::iterator i = vn_ssa_aux_hash->begin (); |
8717 | i != vn_ssa_aux_hash->end (); ++i) |
8718 | { |
8719 | nvalues++; |
8720 | vn_avail *av = (*i)->avail; |
8721 | while (av) |
8722 | { |
8723 | navail++; |
8724 | av = av->next; |
8725 | } |
8726 | } |
8727 | statistics_counter_event (cfun, "RPO blocks" , n); |
8728 | statistics_counter_event (cfun, "RPO blocks visited" , nblk); |
8729 | statistics_counter_event (cfun, "RPO blocks executable" , nex); |
8730 | statistics_histogram_event (cfun, "RPO iterations" , 10*nblk / nex); |
8731 | statistics_histogram_event (cfun, "RPO num values" , nvalues); |
8732 | statistics_histogram_event (cfun, "RPO num avail" , navail); |
8733 | statistics_histogram_event (cfun, "RPO num lattice" , |
8734 | vn_ssa_aux_hash->elements ()); |
8735 | if (dump_file && (dump_flags & (TDF_DETAILS|TDF_STATS))) |
8736 | { |
8737 | fprintf (stream: dump_file, format: "RPO iteration over %d blocks visited %" PRIu64 |
8738 | " blocks in total discovering %d executable blocks iterating " |
8739 | "%d.%d times, a block was visited max. %u times\n" , |
8740 | n, nblk, nex, |
8741 | (int)((10*nblk / nex)/10), (int)((10*nblk / nex)%10), |
8742 | max_visited); |
8743 | fprintf (stream: dump_file, format: "RPO tracked %d values available at %d locations " |
8744 | "and %" PRIu64 " lattice elements\n" , |
8745 | nvalues, navail, (uint64_t) vn_ssa_aux_hash->elements ()); |
8746 | } |
8747 | |
8748 | if (eliminate) |
8749 | { |
8750 | /* When !iterate we already performed elimination during the RPO |
8751 | walk. */ |
8752 | if (iterate) |
8753 | { |
8754 | /* Elimination for region-based VN needs to be done within the |
8755 | RPO walk. */ |
8756 | gcc_assert (! do_region); |
8757 | /* Note we can't use avail.walk here because that gets confused |
8758 | by the existing availability and it will be less efficient |
8759 | as well. */ |
8760 | todo |= eliminate_with_rpo_vn (NULL); |
8761 | } |
8762 | else |
8763 | todo |= avail.eliminate_cleanup (region_p: do_region); |
8764 | } |
8765 | |
8766 | vn_valueize = NULL; |
8767 | rpo_avail = NULL; |
8768 | |
8769 | XDELETEVEC (bb_to_rpo); |
8770 | XDELETEVEC (rpo); |
8771 | XDELETEVEC (rpo_state); |
8772 | |
8773 | return todo; |
8774 | } |
8775 | |
8776 | /* Region-based entry for RPO VN. Performs value-numbering and elimination |
8777 | on the SEME region specified by ENTRY and EXIT_BBS. If ENTRY is not |
8778 | the only edge into the region at ENTRY->dest PHI nodes in ENTRY->dest |
8779 | are not considered. |
8780 | If ITERATE is true then treat backedges optimistically as not |
8781 | executed and iterate. If ELIMINATE is true then perform |
8782 | elimination, otherwise leave that to the caller. |
8783 | If SKIP_ENTRY_PHIS is true then force PHI nodes in ENTRY->dest to VARYING. |
8784 | KIND specifies the amount of work done for handling memory operations. */ |
8785 | |
8786 | unsigned |
8787 | do_rpo_vn (function *fn, edge entry, bitmap exit_bbs, |
8788 | bool iterate, bool eliminate, bool skip_entry_phis, |
8789 | vn_lookup_kind kind) |
8790 | { |
8791 | auto_timevar tv (TV_TREE_RPO_VN); |
8792 | unsigned todo = do_rpo_vn_1 (fn, entry, exit_bbs, iterate, eliminate, |
8793 | skip_entry_phis, kind); |
8794 | free_rpo_vn (); |
8795 | return todo; |
8796 | } |
8797 | |
8798 | |
8799 | namespace { |
8800 | |
8801 | const pass_data pass_data_fre = |
8802 | { |
8803 | .type: GIMPLE_PASS, /* type */ |
8804 | .name: "fre" , /* name */ |
8805 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
8806 | .tv_id: TV_TREE_FRE, /* tv_id */ |
8807 | .properties_required: ( PROP_cfg | PROP_ssa ), /* properties_required */ |
8808 | .properties_provided: 0, /* properties_provided */ |
8809 | .properties_destroyed: 0, /* properties_destroyed */ |
8810 | .todo_flags_start: 0, /* todo_flags_start */ |
8811 | .todo_flags_finish: 0, /* todo_flags_finish */ |
8812 | }; |
8813 | |
8814 | class pass_fre : public gimple_opt_pass |
8815 | { |
8816 | public: |
8817 | pass_fre (gcc::context *ctxt) |
8818 | : gimple_opt_pass (pass_data_fre, ctxt), may_iterate (true) |
8819 | {} |
8820 | |
8821 | /* opt_pass methods: */ |
8822 | opt_pass * clone () final override { return new pass_fre (m_ctxt); } |
8823 | void set_pass_param (unsigned int n, bool param) final override |
8824 | { |
8825 | gcc_assert (n == 0); |
8826 | may_iterate = param; |
8827 | } |
8828 | bool gate (function *) final override |
8829 | { |
8830 | return flag_tree_fre != 0 && (may_iterate || optimize > 1); |
8831 | } |
8832 | unsigned int execute (function *) final override; |
8833 | |
8834 | private: |
8835 | bool may_iterate; |
8836 | }; // class pass_fre |
8837 | |
8838 | unsigned int |
8839 | pass_fre::execute (function *fun) |
8840 | { |
8841 | unsigned todo = 0; |
8842 | |
8843 | /* At -O[1g] use the cheap non-iterating mode. */ |
8844 | bool iterate_p = may_iterate && (optimize > 1); |
8845 | calculate_dominance_info (CDI_DOMINATORS); |
8846 | if (iterate_p) |
8847 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS); |
8848 | |
8849 | todo = do_rpo_vn_1 (fn: fun, NULL, NULL, iterate: iterate_p, eliminate: true, skip_entry_phis: false, kind: VN_WALKREWRITE); |
8850 | free_rpo_vn (); |
8851 | |
8852 | if (iterate_p) |
8853 | loop_optimizer_finalize (); |
8854 | |
8855 | if (scev_initialized_p ()) |
8856 | scev_reset_htab (); |
8857 | |
8858 | /* For late FRE after IVOPTs and unrolling, see if we can |
8859 | remove some TREE_ADDRESSABLE and rewrite stuff into SSA. */ |
8860 | if (!may_iterate) |
8861 | todo |= TODO_update_address_taken; |
8862 | |
8863 | return todo; |
8864 | } |
8865 | |
8866 | } // anon namespace |
8867 | |
8868 | gimple_opt_pass * |
8869 | make_pass_fre (gcc::context *ctxt) |
8870 | { |
8871 | return new pass_fre (ctxt); |
8872 | } |
8873 | |
8874 | #undef BB_EXECUTABLE |
8875 | |