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