Branch data Line data Source code
1 : : /* Full and partial redundancy elimination and code hoisting on SSA GIMPLE.
2 : : Copyright (C) 2001-2025 Free Software Foundation, Inc.
3 : : Contributed by Daniel Berlin <dan@dberlin.org> and Steven Bosscher
4 : : <stevenb@suse.de>
5 : :
6 : : This file is part of GCC.
7 : :
8 : : GCC is free software; you can redistribute it and/or modify
9 : : it under the terms of the GNU General Public License as published by
10 : : the Free Software Foundation; either version 3, or (at your option)
11 : : any later version.
12 : :
13 : : GCC is distributed in the hope that it will be useful,
14 : : but WITHOUT ANY WARRANTY; without even the implied warranty of
15 : : MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 : : GNU General Public License for more details.
17 : :
18 : : You should have received a copy of the GNU General Public License
19 : : along with GCC; see the file COPYING3. If not see
20 : : <http://www.gnu.org/licenses/>. */
21 : :
22 : : #include "config.h"
23 : : #include "system.h"
24 : : #include "coretypes.h"
25 : : #include "backend.h"
26 : : #include "rtl.h"
27 : : #include "tree.h"
28 : : #include "gimple.h"
29 : : #include "predict.h"
30 : : #include "alloc-pool.h"
31 : : #include "tree-pass.h"
32 : : #include "ssa.h"
33 : : #include "cgraph.h"
34 : : #include "gimple-pretty-print.h"
35 : : #include "fold-const.h"
36 : : #include "cfganal.h"
37 : : #include "gimple-iterator.h"
38 : : #include "gimple-fold.h"
39 : : #include "tree-eh.h"
40 : : #include "gimplify.h"
41 : : #include "tree-cfg.h"
42 : : #include "tree-into-ssa.h"
43 : : #include "tree-dfa.h"
44 : : #include "tree-ssa.h"
45 : : #include "cfgloop.h"
46 : : #include "tree-ssa-sccvn.h"
47 : : #include "tree-scalar-evolution.h"
48 : : #include "dbgcnt.h"
49 : : #include "domwalk.h"
50 : : #include "tree-ssa-propagate.h"
51 : : #include "tree-ssa-dce.h"
52 : : #include "tree-cfgcleanup.h"
53 : : #include "alias.h"
54 : : #include "gimple-range.h"
55 : :
56 : : /* Even though this file is called tree-ssa-pre.cc, we actually
57 : : implement a bit more than just PRE here. All of them piggy-back
58 : : on GVN which is implemented in tree-ssa-sccvn.cc.
59 : :
60 : : 1. Full Redundancy Elimination (FRE)
61 : : This is the elimination phase of GVN.
62 : :
63 : : 2. Partial Redundancy Elimination (PRE)
64 : : This is adds computation of AVAIL_OUT and ANTIC_IN and
65 : : doing expression insertion to form GVN-PRE.
66 : :
67 : : 3. Code hoisting
68 : : This optimization uses the ANTIC_IN sets computed for PRE
69 : : to move expressions further up than PRE would do, to make
70 : : multiple computations of the same value fully redundant.
71 : : This pass is explained below (after the explanation of the
72 : : basic algorithm for PRE).
73 : : */
74 : :
75 : : /* TODO:
76 : :
77 : : 1. Avail sets can be shared by making an avail_find_leader that
78 : : walks up the dominator tree and looks in those avail sets.
79 : : This might affect code optimality, it's unclear right now.
80 : : Currently the AVAIL_OUT sets are the remaining quadraticness in
81 : : memory of GVN-PRE.
82 : : 2. Strength reduction can be performed by anticipating expressions
83 : : we can repair later on.
84 : : 3. We can do back-substitution or smarter value numbering to catch
85 : : commutative expressions split up over multiple statements.
86 : : */
87 : :
88 : : /* For ease of terminology, "expression node" in the below refers to
89 : : every expression node but GIMPLE_ASSIGN, because GIMPLE_ASSIGNs
90 : : represent the actual statement containing the expressions we care about,
91 : : and we cache the value number by putting it in the expression. */
92 : :
93 : : /* Basic algorithm for Partial Redundancy Elimination:
94 : :
95 : : First we walk the statements to generate the AVAIL sets, the
96 : : EXP_GEN sets, and the tmp_gen sets. EXP_GEN sets represent the
97 : : generation of values/expressions by a given block. We use them
98 : : when computing the ANTIC sets. The AVAIL sets consist of
99 : : SSA_NAME's that represent values, so we know what values are
100 : : available in what blocks. AVAIL is a forward dataflow problem. In
101 : : SSA, values are never killed, so we don't need a kill set, or a
102 : : fixpoint iteration, in order to calculate the AVAIL sets. In
103 : : traditional parlance, AVAIL sets tell us the downsafety of the
104 : : expressions/values.
105 : :
106 : : Next, we generate the ANTIC sets. These sets represent the
107 : : anticipatable expressions. ANTIC is a backwards dataflow
108 : : problem. An expression is anticipatable in a given block if it could
109 : : be generated in that block. This means that if we had to perform
110 : : an insertion in that block, of the value of that expression, we
111 : : could. Calculating the ANTIC sets requires phi translation of
112 : : expressions, because the flow goes backwards through phis. We must
113 : : iterate to a fixpoint of the ANTIC sets, because we have a kill
114 : : set. Even in SSA form, values are not live over the entire
115 : : function, only from their definition point onwards. So we have to
116 : : remove values from the ANTIC set once we go past the definition
117 : : point of the leaders that make them up.
118 : : compute_antic/compute_antic_aux performs this computation.
119 : :
120 : : Third, we perform insertions to make partially redundant
121 : : expressions fully redundant.
122 : :
123 : : An expression is partially redundant (excluding partial
124 : : anticipation) if:
125 : :
126 : : 1. It is AVAIL in some, but not all, of the predecessors of a
127 : : given block.
128 : : 2. It is ANTIC in all the predecessors.
129 : :
130 : : In order to make it fully redundant, we insert the expression into
131 : : the predecessors where it is not available, but is ANTIC.
132 : :
133 : : When optimizing for size, we only eliminate the partial redundancy
134 : : if we need to insert in only one predecessor. This avoids almost
135 : : completely the code size increase that PRE usually causes.
136 : :
137 : : For the partial anticipation case, we only perform insertion if it
138 : : is partially anticipated in some block, and fully available in all
139 : : of the predecessors.
140 : :
141 : : do_pre_regular_insertion/do_pre_partial_partial_insertion
142 : : performs these steps, driven by insert/insert_aux.
143 : :
144 : : Fourth, we eliminate fully redundant expressions.
145 : : This is a simple statement walk that replaces redundant
146 : : calculations with the now available values. */
147 : :
148 : : /* Basic algorithm for Code Hoisting:
149 : :
150 : : Code hoisting is: Moving value computations up in the control flow
151 : : graph to make multiple copies redundant. Typically this is a size
152 : : optimization, but there are cases where it also is helpful for speed.
153 : :
154 : : A simple code hoisting algorithm is implemented that piggy-backs on
155 : : the PRE infrastructure. For code hoisting, we have to know ANTIC_OUT
156 : : which is effectively ANTIC_IN - AVAIL_OUT. The latter two have to be
157 : : computed for PRE, and we can use them to perform a limited version of
158 : : code hoisting, too.
159 : :
160 : : For the purpose of this implementation, a value is hoistable to a basic
161 : : block B if the following properties are met:
162 : :
163 : : 1. The value is in ANTIC_IN(B) -- the value will be computed on all
164 : : paths from B to function exit and it can be computed in B);
165 : :
166 : : 2. The value is not in AVAIL_OUT(B) -- there would be no need to
167 : : compute the value again and make it available twice;
168 : :
169 : : 3. All successors of B are dominated by B -- makes sure that inserting
170 : : a computation of the value in B will make the remaining
171 : : computations fully redundant;
172 : :
173 : : 4. At least one successor has the value in AVAIL_OUT -- to avoid
174 : : hoisting values up too far;
175 : :
176 : : 5. There are at least two successors of B -- hoisting in straight
177 : : line code is pointless.
178 : :
179 : : The third condition is not strictly necessary, but it would complicate
180 : : the hoisting pass a lot. In fact, I don't know of any code hoisting
181 : : algorithm that does not have this requirement. Fortunately, experiments
182 : : have show that most candidate hoistable values are in regions that meet
183 : : this condition (e.g. diamond-shape regions).
184 : :
185 : : The forth condition is necessary to avoid hoisting things up too far
186 : : away from the uses of the value. Nothing else limits the algorithm
187 : : from hoisting everything up as far as ANTIC_IN allows. Experiments
188 : : with SPEC and CSiBE have shown that hoisting up too far results in more
189 : : spilling, less benefits for code size, and worse benchmark scores.
190 : : Fortunately, in practice most of the interesting hoisting opportunities
191 : : are caught despite this limitation.
192 : :
193 : : For hoistable values that meet all conditions, expressions are inserted
194 : : to make the calculation of the hoistable value fully redundant. We
195 : : perform code hoisting insertions after each round of PRE insertions,
196 : : because code hoisting never exposes new PRE opportunities, but PRE can
197 : : create new code hoisting opportunities.
198 : :
199 : : The code hoisting algorithm is implemented in do_hoist_insert, driven
200 : : by insert/insert_aux. */
201 : :
202 : : /* Representations of value numbers:
203 : :
204 : : Value numbers are represented by a representative SSA_NAME. We
205 : : will create fake SSA_NAME's in situations where we need a
206 : : representative but do not have one (because it is a complex
207 : : expression). In order to facilitate storing the value numbers in
208 : : bitmaps, and keep the number of wasted SSA_NAME's down, we also
209 : : associate a value_id with each value number, and create full blown
210 : : ssa_name's only where we actually need them (IE in operands of
211 : : existing expressions).
212 : :
213 : : Theoretically you could replace all the value_id's with
214 : : SSA_NAME_VERSION, but this would allocate a large number of
215 : : SSA_NAME's (which are each > 30 bytes) just to get a 4 byte number.
216 : : It would also require an additional indirection at each point we
217 : : use the value id. */
218 : :
219 : : /* Representation of expressions on value numbers:
220 : :
221 : : Expressions consisting of value numbers are represented the same
222 : : way as our VN internally represents them, with an additional
223 : : "pre_expr" wrapping around them in order to facilitate storing all
224 : : of the expressions in the same sets. */
225 : :
226 : : /* Representation of sets:
227 : :
228 : : The dataflow sets do not need to be sorted in any particular order
229 : : for the majority of their lifetime, are simply represented as two
230 : : bitmaps, one that keeps track of values present in the set, and one
231 : : that keeps track of expressions present in the set.
232 : :
233 : : When we need them in topological order, we produce it on demand by
234 : : transforming the bitmap into an array and sorting it into topo
235 : : order. */
236 : :
237 : : /* Type of expression, used to know which member of the PRE_EXPR union
238 : : is valid. */
239 : :
240 : : enum pre_expr_kind
241 : : {
242 : : NAME,
243 : : NARY,
244 : : REFERENCE,
245 : : CONSTANT
246 : : };
247 : :
248 : : union pre_expr_union
249 : : {
250 : : tree name;
251 : : tree constant;
252 : : vn_nary_op_t nary;
253 : : vn_reference_t reference;
254 : : };
255 : :
256 : : typedef struct pre_expr_d : nofree_ptr_hash <pre_expr_d>
257 : : {
258 : : enum pre_expr_kind kind;
259 : : unsigned int id;
260 : : unsigned value_id;
261 : : location_t loc;
262 : : pre_expr_union u;
263 : :
264 : : /* hash_table support. */
265 : : static inline hashval_t hash (const pre_expr_d *);
266 : : static inline int equal (const pre_expr_d *, const pre_expr_d *);
267 : : } *pre_expr;
268 : :
269 : : #define PRE_EXPR_NAME(e) (e)->u.name
270 : : #define PRE_EXPR_NARY(e) (e)->u.nary
271 : : #define PRE_EXPR_REFERENCE(e) (e)->u.reference
272 : : #define PRE_EXPR_CONSTANT(e) (e)->u.constant
273 : :
274 : : /* Compare E1 and E1 for equality. */
275 : :
276 : : inline int
277 : 54089947 : pre_expr_d::equal (const pre_expr_d *e1, const pre_expr_d *e2)
278 : : {
279 : 54089947 : if (e1->kind != e2->kind)
280 : : return false;
281 : :
282 : 34205573 : switch (e1->kind)
283 : : {
284 : 4528138 : case CONSTANT:
285 : 4528138 : return vn_constant_eq_with_type (PRE_EXPR_CONSTANT (e1),
286 : 4528138 : PRE_EXPR_CONSTANT (e2));
287 : 150564 : case NAME:
288 : 150564 : return PRE_EXPR_NAME (e1) == PRE_EXPR_NAME (e2);
289 : 21159862 : case NARY:
290 : 21159862 : return vn_nary_op_eq (PRE_EXPR_NARY (e1), PRE_EXPR_NARY (e2));
291 : 8367009 : case REFERENCE:
292 : 8367009 : return vn_reference_eq (PRE_EXPR_REFERENCE (e1),
293 : 8367009 : PRE_EXPR_REFERENCE (e2));
294 : 0 : default:
295 : 0 : gcc_unreachable ();
296 : : }
297 : : }
298 : :
299 : : /* Hash E. */
300 : :
301 : : inline hashval_t
302 : 87667924 : pre_expr_d::hash (const pre_expr_d *e)
303 : : {
304 : 87667924 : switch (e->kind)
305 : : {
306 : 6941654 : case CONSTANT:
307 : 6941654 : return vn_hash_constant_with_type (PRE_EXPR_CONSTANT (e));
308 : 0 : case NAME:
309 : 0 : return SSA_NAME_VERSION (PRE_EXPR_NAME (e));
310 : 53576124 : case NARY:
311 : 53576124 : return PRE_EXPR_NARY (e)->hashcode;
312 : 27150146 : case REFERENCE:
313 : 27150146 : return PRE_EXPR_REFERENCE (e)->hashcode;
314 : 0 : default:
315 : 0 : gcc_unreachable ();
316 : : }
317 : : }
318 : :
319 : : /* Next global expression id number. */
320 : : static unsigned int next_expression_id;
321 : :
322 : : /* Mapping from expression to id number we can use in bitmap sets. */
323 : : static vec<pre_expr> expressions;
324 : : static hash_table<pre_expr_d> *expression_to_id;
325 : : static vec<unsigned> name_to_id;
326 : : static obstack pre_expr_obstack;
327 : :
328 : : /* Allocate an expression id for EXPR. */
329 : :
330 : : static inline unsigned int
331 : 42966335 : alloc_expression_id (pre_expr expr)
332 : : {
333 : 42966335 : struct pre_expr_d **slot;
334 : : /* Make sure we won't overflow. */
335 : 42966335 : gcc_assert (next_expression_id + 1 > next_expression_id);
336 : 42966335 : expr->id = next_expression_id++;
337 : 42966335 : expressions.safe_push (expr);
338 : 42966335 : if (expr->kind == NAME)
339 : : {
340 : 23548365 : unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
341 : : /* vec::safe_grow_cleared allocates no headroom. Avoid frequent
342 : : re-allocations by using vec::reserve upfront. */
343 : 23548365 : unsigned old_len = name_to_id.length ();
344 : 47096730 : name_to_id.reserve (num_ssa_names - old_len);
345 : 47096730 : name_to_id.quick_grow_cleared (num_ssa_names);
346 : 23548365 : gcc_assert (name_to_id[version] == 0);
347 : 23548365 : name_to_id[version] = expr->id;
348 : : }
349 : : else
350 : : {
351 : 19417970 : slot = expression_to_id->find_slot (expr, INSERT);
352 : 19417970 : gcc_assert (!*slot);
353 : 19417970 : *slot = expr;
354 : : }
355 : 42966335 : return next_expression_id - 1;
356 : : }
357 : :
358 : : /* Return the expression id for tree EXPR. */
359 : :
360 : : static inline unsigned int
361 : 247764595 : get_expression_id (const pre_expr expr)
362 : : {
363 : 247764595 : return expr->id;
364 : : }
365 : :
366 : : static inline unsigned int
367 : 77296492 : lookup_expression_id (const pre_expr expr)
368 : : {
369 : 77296492 : struct pre_expr_d **slot;
370 : :
371 : 77296492 : if (expr->kind == NAME)
372 : : {
373 : 52123588 : unsigned version = SSA_NAME_VERSION (PRE_EXPR_NAME (expr));
374 : 71541558 : if (name_to_id.length () <= version)
375 : : return 0;
376 : 49324412 : return name_to_id[version];
377 : : }
378 : : else
379 : : {
380 : 25172904 : slot = expression_to_id->find_slot (expr, NO_INSERT);
381 : 25172904 : if (!slot)
382 : : return 0;
383 : 5754934 : return ((pre_expr)*slot)->id;
384 : : }
385 : : }
386 : :
387 : : /* Return the expression that has expression id ID */
388 : :
389 : : static inline pre_expr
390 : 496351994 : expression_for_id (unsigned int id)
391 : : {
392 : 992703988 : return expressions[id];
393 : : }
394 : :
395 : : static object_allocator<pre_expr_d> pre_expr_pool ("pre_expr nodes");
396 : :
397 : : /* Given an SSA_NAME NAME, get or create a pre_expr to represent it. */
398 : :
399 : : static pre_expr
400 : 52123588 : get_or_alloc_expr_for_name (tree name)
401 : : {
402 : 52123588 : struct pre_expr_d expr;
403 : 52123588 : pre_expr result;
404 : 52123588 : unsigned int result_id;
405 : :
406 : 52123588 : expr.kind = NAME;
407 : 52123588 : expr.id = 0;
408 : 52123588 : PRE_EXPR_NAME (&expr) = name;
409 : 52123588 : result_id = lookup_expression_id (&expr);
410 : 52123588 : if (result_id != 0)
411 : 28575223 : return expression_for_id (result_id);
412 : :
413 : 23548365 : result = pre_expr_pool.allocate ();
414 : 23548365 : result->kind = NAME;
415 : 23548365 : result->loc = UNKNOWN_LOCATION;
416 : 23548365 : result->value_id = VN_INFO (name)->value_id;
417 : 23548365 : PRE_EXPR_NAME (result) = name;
418 : 23548365 : alloc_expression_id (result);
419 : 23548365 : return result;
420 : : }
421 : :
422 : : /* Given an NARY, get or create a pre_expr to represent it. Assign
423 : : VALUE_ID to it or allocate a new value-id if it is zero. Record
424 : : LOC as the original location of the expression. */
425 : :
426 : : static pre_expr
427 : 13302849 : get_or_alloc_expr_for_nary (vn_nary_op_t nary, unsigned value_id,
428 : : location_t loc = UNKNOWN_LOCATION)
429 : : {
430 : 13302849 : struct pre_expr_d expr;
431 : 13302849 : pre_expr result;
432 : 13302849 : unsigned int result_id;
433 : :
434 : 13302849 : gcc_assert (value_id == 0 || !value_id_constant_p (value_id));
435 : :
436 : 13302849 : expr.kind = NARY;
437 : 13302849 : expr.id = 0;
438 : 13302849 : nary->hashcode = vn_nary_op_compute_hash (nary);
439 : 13302849 : PRE_EXPR_NARY (&expr) = nary;
440 : 13302849 : result_id = lookup_expression_id (&expr);
441 : 13302849 : if (result_id != 0)
442 : 955914 : return expression_for_id (result_id);
443 : :
444 : 12346935 : result = pre_expr_pool.allocate ();
445 : 12346935 : result->kind = NARY;
446 : 12346935 : result->loc = loc;
447 : 12346935 : result->value_id = value_id ? value_id : get_next_value_id ();
448 : 12346935 : PRE_EXPR_NARY (result)
449 : 12346935 : = alloc_vn_nary_op_noinit (nary->length, &pre_expr_obstack);
450 : 12346935 : memcpy (PRE_EXPR_NARY (result), nary, sizeof_vn_nary_op (nary->length));
451 : 12346935 : alloc_expression_id (result);
452 : 12346935 : return result;
453 : : }
454 : :
455 : : /* Given an REFERENCE, get or create a pre_expr to represent it. */
456 : :
457 : : static pre_expr
458 : 7068955 : get_or_alloc_expr_for_reference (vn_reference_t reference,
459 : : location_t loc = UNKNOWN_LOCATION)
460 : : {
461 : 7068955 : struct pre_expr_d expr;
462 : 7068955 : pre_expr result;
463 : 7068955 : unsigned int result_id;
464 : :
465 : 7068955 : expr.kind = REFERENCE;
466 : 7068955 : expr.id = 0;
467 : 7068955 : PRE_EXPR_REFERENCE (&expr) = reference;
468 : 7068955 : result_id = lookup_expression_id (&expr);
469 : 7068955 : if (result_id != 0)
470 : 817817 : return expression_for_id (result_id);
471 : :
472 : 6251138 : result = pre_expr_pool.allocate ();
473 : 6251138 : result->kind = REFERENCE;
474 : 6251138 : result->loc = loc;
475 : 6251138 : result->value_id = reference->value_id;
476 : 6251138 : PRE_EXPR_REFERENCE (result) = reference;
477 : 6251138 : alloc_expression_id (result);
478 : 6251138 : return result;
479 : : }
480 : :
481 : :
482 : : /* An unordered bitmap set. One bitmap tracks values, the other,
483 : : expressions. */
484 : 145764104 : typedef class bitmap_set
485 : : {
486 : : public:
487 : : bitmap_head expressions;
488 : : bitmap_head values;
489 : : } *bitmap_set_t;
490 : :
491 : : #define FOR_EACH_EXPR_ID_IN_SET(set, id, bi) \
492 : : EXECUTE_IF_SET_IN_BITMAP (&(set)->expressions, 0, (id), (bi))
493 : :
494 : : #define FOR_EACH_VALUE_ID_IN_SET(set, id, bi) \
495 : : EXECUTE_IF_SET_IN_BITMAP (&(set)->values, 0, (id), (bi))
496 : :
497 : : /* Mapping from value id to expressions with that value_id. */
498 : : static vec<bitmap> value_expressions;
499 : : /* We just record a single expression for each constant value,
500 : : one of kind CONSTANT. */
501 : : static vec<pre_expr> constant_value_expressions;
502 : :
503 : :
504 : : /* This structure is used to keep track of statistics on what
505 : : optimization PRE was able to perform. */
506 : : static struct
507 : : {
508 : : /* The number of new expressions/temporaries generated by PRE. */
509 : : int insertions;
510 : :
511 : : /* The number of inserts found due to partial anticipation */
512 : : int pa_insert;
513 : :
514 : : /* The number of inserts made for code hoisting. */
515 : : int hoist_insert;
516 : :
517 : : /* The number of new PHI nodes added by PRE. */
518 : : int phis;
519 : : } pre_stats;
520 : :
521 : : static bool do_partial_partial;
522 : : static pre_expr bitmap_find_leader (bitmap_set_t, unsigned int);
523 : : static void bitmap_value_insert_into_set (bitmap_set_t, pre_expr);
524 : : static bool bitmap_value_replace_in_set (bitmap_set_t, pre_expr);
525 : : static void bitmap_set_copy (bitmap_set_t, bitmap_set_t);
526 : : static bool bitmap_set_contains_value (bitmap_set_t, unsigned int);
527 : : static void bitmap_insert_into_set (bitmap_set_t, pre_expr);
528 : : static bitmap_set_t bitmap_set_new (void);
529 : : static tree create_expression_by_pieces (basic_block, pre_expr, gimple_seq *,
530 : : tree);
531 : : static tree find_or_generate_expression (basic_block, tree, gimple_seq *);
532 : : static unsigned int get_expr_value_id (pre_expr);
533 : :
534 : : /* We can add and remove elements and entries to and from sets
535 : : and hash tables, so we use alloc pools for them. */
536 : :
537 : : static object_allocator<bitmap_set> bitmap_set_pool ("Bitmap sets");
538 : : static bitmap_obstack grand_bitmap_obstack;
539 : :
540 : : /* A three tuple {e, pred, v} used to cache phi translations in the
541 : : phi_translate_table. */
542 : :
543 : : typedef struct expr_pred_trans_d : public typed_noop_remove <expr_pred_trans_d>
544 : : {
545 : : typedef expr_pred_trans_d value_type;
546 : : typedef expr_pred_trans_d compare_type;
547 : :
548 : : /* The expression ID. */
549 : : unsigned e;
550 : :
551 : : /* The value expression ID that resulted from the translation. */
552 : : unsigned v;
553 : :
554 : : /* hash_table support. */
555 : : static inline void mark_empty (expr_pred_trans_d &);
556 : : static inline bool is_empty (const expr_pred_trans_d &);
557 : : static inline void mark_deleted (expr_pred_trans_d &);
558 : : static inline bool is_deleted (const expr_pred_trans_d &);
559 : : static const bool empty_zero_p = true;
560 : : static inline hashval_t hash (const expr_pred_trans_d &);
561 : : static inline int equal (const expr_pred_trans_d &, const expr_pred_trans_d &);
562 : : } *expr_pred_trans_t;
563 : : typedef const struct expr_pred_trans_d *const_expr_pred_trans_t;
564 : :
565 : : inline bool
566 : 1314571820 : expr_pred_trans_d::is_empty (const expr_pred_trans_d &e)
567 : : {
568 : 1314571820 : return e.e == 0;
569 : : }
570 : :
571 : : inline bool
572 : 256146348 : expr_pred_trans_d::is_deleted (const expr_pred_trans_d &e)
573 : : {
574 : 256146348 : return e.e == -1u;
575 : : }
576 : :
577 : : inline void
578 : 1935572 : expr_pred_trans_d::mark_empty (expr_pred_trans_d &e)
579 : : {
580 : 1935572 : e.e = 0;
581 : : }
582 : :
583 : : inline void
584 : 3405989 : expr_pred_trans_d::mark_deleted (expr_pred_trans_d &e)
585 : : {
586 : 3405989 : e.e = -1u;
587 : : }
588 : :
589 : : inline hashval_t
590 : : expr_pred_trans_d::hash (const expr_pred_trans_d &e)
591 : : {
592 : : return e.e;
593 : : }
594 : :
595 : : inline int
596 : 200380061 : expr_pred_trans_d::equal (const expr_pred_trans_d &ve1,
597 : : const expr_pred_trans_d &ve2)
598 : : {
599 : 200380061 : return ve1.e == ve2.e;
600 : : }
601 : :
602 : : /* Sets that we need to keep track of. */
603 : : typedef struct bb_bitmap_sets
604 : : {
605 : : /* The EXP_GEN set, which represents expressions/values generated in
606 : : a basic block. */
607 : : bitmap_set_t exp_gen;
608 : :
609 : : /* The PHI_GEN set, which represents PHI results generated in a
610 : : basic block. */
611 : : bitmap_set_t phi_gen;
612 : :
613 : : /* The TMP_GEN set, which represents results/temporaries generated
614 : : in a basic block. IE the LHS of an expression. */
615 : : bitmap_set_t tmp_gen;
616 : :
617 : : /* The AVAIL_OUT set, which represents which values are available in
618 : : a given basic block. */
619 : : bitmap_set_t avail_out;
620 : :
621 : : /* The ANTIC_IN set, which represents which values are anticipatable
622 : : in a given basic block. */
623 : : bitmap_set_t antic_in;
624 : :
625 : : /* The PA_IN set, which represents which values are
626 : : partially anticipatable in a given basic block. */
627 : : bitmap_set_t pa_in;
628 : :
629 : : /* The NEW_SETS set, which is used during insertion to augment the
630 : : AVAIL_OUT set of blocks with the new insertions performed during
631 : : the current iteration. */
632 : : bitmap_set_t new_sets;
633 : :
634 : : /* A cache for value_dies_in_block_x. */
635 : : bitmap expr_dies;
636 : :
637 : : /* The live virtual operand on successor edges. */
638 : : tree vop_on_exit;
639 : :
640 : : /* PHI translate cache for the single successor edge. */
641 : : hash_table<expr_pred_trans_d> *phi_translate_table;
642 : :
643 : : /* True if we have visited this block during ANTIC calculation. */
644 : : unsigned int visited : 1;
645 : :
646 : : /* True when the block contains a call that might not return. */
647 : : unsigned int contains_may_not_return_call : 1;
648 : : } *bb_value_sets_t;
649 : :
650 : : #define EXP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->exp_gen
651 : : #define PHI_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->phi_gen
652 : : #define TMP_GEN(BB) ((bb_value_sets_t) ((BB)->aux))->tmp_gen
653 : : #define AVAIL_OUT(BB) ((bb_value_sets_t) ((BB)->aux))->avail_out
654 : : #define ANTIC_IN(BB) ((bb_value_sets_t) ((BB)->aux))->antic_in
655 : : #define PA_IN(BB) ((bb_value_sets_t) ((BB)->aux))->pa_in
656 : : #define NEW_SETS(BB) ((bb_value_sets_t) ((BB)->aux))->new_sets
657 : : #define EXPR_DIES(BB) ((bb_value_sets_t) ((BB)->aux))->expr_dies
658 : : #define PHI_TRANS_TABLE(BB) ((bb_value_sets_t) ((BB)->aux))->phi_translate_table
659 : : #define BB_VISITED(BB) ((bb_value_sets_t) ((BB)->aux))->visited
660 : : #define BB_MAY_NOTRETURN(BB) ((bb_value_sets_t) ((BB)->aux))->contains_may_not_return_call
661 : : #define BB_LIVE_VOP_ON_EXIT(BB) ((bb_value_sets_t) ((BB)->aux))->vop_on_exit
662 : :
663 : :
664 : : /* Add the tuple mapping from {expression E, basic block PRED} to
665 : : the phi translation table and return whether it pre-existed. */
666 : :
667 : : static inline bool
668 : 82067566 : phi_trans_add (expr_pred_trans_t *entry, pre_expr e, basic_block pred)
669 : : {
670 : 82067566 : if (!PHI_TRANS_TABLE (pred))
671 : 182545 : PHI_TRANS_TABLE (pred) = new hash_table<expr_pred_trans_d> (11);
672 : :
673 : 82067566 : expr_pred_trans_t slot;
674 : 82067566 : expr_pred_trans_d tem;
675 : 82067566 : unsigned id = get_expression_id (e);
676 : 82067566 : tem.e = id;
677 : 82067566 : slot = PHI_TRANS_TABLE (pred)->find_slot_with_hash (tem, id, INSERT);
678 : 82067566 : if (slot->e)
679 : : {
680 : 59750349 : *entry = slot;
681 : 59750349 : return true;
682 : : }
683 : :
684 : 22317217 : *entry = slot;
685 : 22317217 : slot->e = id;
686 : 22317217 : return false;
687 : : }
688 : :
689 : :
690 : : /* Add expression E to the expression set of value id V. */
691 : :
692 : : static void
693 : 44740066 : add_to_value (unsigned int v, pre_expr e)
694 : : {
695 : 0 : gcc_checking_assert (get_expr_value_id (e) == v);
696 : :
697 : 44740066 : if (value_id_constant_p (v))
698 : : {
699 : 867928 : if (e->kind != CONSTANT)
700 : : return;
701 : :
702 : 819897 : if (-v >= constant_value_expressions.length ())
703 : 487141 : constant_value_expressions.safe_grow_cleared (-v + 1);
704 : :
705 : 819897 : pre_expr leader = constant_value_expressions[-v];
706 : 819897 : if (!leader)
707 : 819897 : constant_value_expressions[-v] = e;
708 : : }
709 : : else
710 : : {
711 : 43872138 : if (v >= value_expressions.length ())
712 : 6678641 : value_expressions.safe_grow_cleared (v + 1);
713 : :
714 : 43872138 : bitmap set = value_expressions[v];
715 : 43872138 : if (!set)
716 : : {
717 : 24450866 : set = BITMAP_ALLOC (&grand_bitmap_obstack);
718 : 24450866 : value_expressions[v] = set;
719 : : }
720 : 43872138 : bitmap_set_bit (set, get_expression_id (e));
721 : : }
722 : : }
723 : :
724 : : /* Create a new bitmap set and return it. */
725 : :
726 : : static bitmap_set_t
727 : 145764104 : bitmap_set_new (void)
728 : : {
729 : 145764104 : bitmap_set_t ret = bitmap_set_pool.allocate ();
730 : 145764104 : bitmap_initialize (&ret->expressions, &grand_bitmap_obstack);
731 : 145764104 : bitmap_initialize (&ret->values, &grand_bitmap_obstack);
732 : 145764104 : return ret;
733 : : }
734 : :
735 : : /* Return the value id for a PRE expression EXPR. */
736 : :
737 : : static unsigned int
738 : 517091104 : get_expr_value_id (pre_expr expr)
739 : : {
740 : : /* ??? We cannot assert that expr has a value-id (it can be 0), because
741 : : we assign value-ids only to expressions that have a result
742 : : in set_hashtable_value_ids. */
743 : 44740066 : return expr->value_id;
744 : : }
745 : :
746 : : /* Return a VN valnum (SSA name or constant) for the PRE value-id VAL. */
747 : :
748 : : static tree
749 : 1234562 : vn_valnum_from_value_id (unsigned int val)
750 : : {
751 : 1234562 : if (value_id_constant_p (val))
752 : : {
753 : 0 : pre_expr vexpr = constant_value_expressions[-val];
754 : 0 : if (vexpr)
755 : 0 : return PRE_EXPR_CONSTANT (vexpr);
756 : : return NULL_TREE;
757 : : }
758 : :
759 : 1234562 : bitmap exprset = value_expressions[val];
760 : 1234562 : bitmap_iterator bi;
761 : 1234562 : unsigned int i;
762 : 1838111 : EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
763 : : {
764 : 1505725 : pre_expr vexpr = expression_for_id (i);
765 : 1505725 : if (vexpr->kind == NAME)
766 : 902176 : return VN_INFO (PRE_EXPR_NAME (vexpr))->valnum;
767 : : }
768 : : return NULL_TREE;
769 : : }
770 : :
771 : : /* Insert an expression EXPR into a bitmapped set. */
772 : :
773 : : static void
774 : 72309873 : bitmap_insert_into_set (bitmap_set_t set, pre_expr expr)
775 : : {
776 : 72309873 : unsigned int val = get_expr_value_id (expr);
777 : 72309873 : if (! value_id_constant_p (val))
778 : : {
779 : : /* Note this is the only function causing multiple expressions
780 : : for the same value to appear in a set. This is needed for
781 : : TMP_GEN, PHI_GEN and NEW_SETs. */
782 : 69659953 : bitmap_set_bit (&set->values, val);
783 : 69659953 : bitmap_set_bit (&set->expressions, get_expression_id (expr));
784 : : }
785 : 72309873 : }
786 : :
787 : : /* Copy a bitmapped set ORIG, into bitmapped set DEST. */
788 : :
789 : : static void
790 : 27027434 : bitmap_set_copy (bitmap_set_t dest, bitmap_set_t orig)
791 : : {
792 : 27027434 : bitmap_copy (&dest->expressions, &orig->expressions);
793 : 27027434 : bitmap_copy (&dest->values, &orig->values);
794 : 27027434 : }
795 : :
796 : :
797 : : /* Free memory used up by SET. */
798 : : static void
799 : 72591304 : bitmap_set_free (bitmap_set_t set)
800 : : {
801 : 0 : bitmap_clear (&set->expressions);
802 : 19524418 : bitmap_clear (&set->values);
803 : 48939069 : }
804 : :
805 : : static void
806 : : pre_expr_DFS (pre_expr expr, bitmap_set_t set, bitmap val_visited,
807 : : vec<pre_expr> &post);
808 : :
809 : : /* DFS walk leaders of VAL to their operands with leaders in SET, collecting
810 : : expressions in SET in postorder into POST. */
811 : :
812 : : static void
813 : 82923482 : pre_expr_DFS (unsigned val, bitmap_set_t set, bitmap val_visited,
814 : : vec<pre_expr> &post)
815 : : {
816 : 82923482 : unsigned int i;
817 : 82923482 : bitmap_iterator bi;
818 : :
819 : : /* Iterate over all leaders and DFS recurse. Borrowed from
820 : : bitmap_find_leader. */
821 : 82923482 : bitmap exprset = value_expressions[val];
822 : 82923482 : if (!exprset->first->next)
823 : : {
824 : 197653745 : EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
825 : 125298629 : if (bitmap_bit_p (&set->expressions, i))
826 : 72535929 : pre_expr_DFS (expression_for_id (i), set, val_visited, post);
827 : 72355116 : return;
828 : : }
829 : :
830 : 21356546 : EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
831 : 10788180 : pre_expr_DFS (expression_for_id (i), set, val_visited, post);
832 : : }
833 : :
834 : : /* DFS walk EXPR to its operands with leaders in SET, collecting
835 : : expressions in SET in postorder into POST. */
836 : :
837 : : static void
838 : 83324109 : pre_expr_DFS (pre_expr expr, bitmap_set_t set, bitmap val_visited,
839 : : vec<pre_expr> &post)
840 : : {
841 : 83324109 : switch (expr->kind)
842 : : {
843 : 38146515 : case NARY:
844 : 38146515 : {
845 : 38146515 : vn_nary_op_t nary = PRE_EXPR_NARY (expr);
846 : 102566903 : for (unsigned i = 0; i < nary->length; i++)
847 : : {
848 : 64420388 : if (TREE_CODE (nary->op[i]) != SSA_NAME)
849 : 19745401 : continue;
850 : 44674987 : unsigned int op_val_id = VN_INFO (nary->op[i])->value_id;
851 : : /* If we already found a leader for the value we've
852 : : recursed already. Avoid the costly bitmap_find_leader. */
853 : 44674987 : if (bitmap_bit_p (&set->values, op_val_id)
854 : 44674987 : && bitmap_set_bit (val_visited, op_val_id))
855 : 7848337 : pre_expr_DFS (op_val_id, set, val_visited, post);
856 : : }
857 : : break;
858 : : }
859 : 12782020 : case REFERENCE:
860 : 12782020 : {
861 : 12782020 : vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
862 : 12782020 : vec<vn_reference_op_s> operands = ref->operands;
863 : 12782020 : vn_reference_op_t operand;
864 : 50556091 : for (unsigned i = 0; operands.iterate (i, &operand); i++)
865 : : {
866 : 37774071 : tree op[3];
867 : 37774071 : op[0] = operand->op0;
868 : 37774071 : op[1] = operand->op1;
869 : 37774071 : op[2] = operand->op2;
870 : 151096284 : for (unsigned n = 0; n < 3; ++n)
871 : : {
872 : 113322213 : if (!op[n] || TREE_CODE (op[n]) != SSA_NAME)
873 : 104839128 : continue;
874 : 8483085 : unsigned op_val_id = VN_INFO (op[n])->value_id;
875 : 8483085 : if (bitmap_bit_p (&set->values, op_val_id)
876 : 8483085 : && bitmap_set_bit (val_visited, op_val_id))
877 : 1459020 : pre_expr_DFS (op_val_id, set, val_visited, post);
878 : : }
879 : : }
880 : : break;
881 : : }
882 : 83324109 : default:;
883 : : }
884 : 83324109 : post.quick_push (expr);
885 : 83324109 : }
886 : :
887 : : /* Generate an topological-ordered array of bitmap set SET. */
888 : :
889 : : static vec<pre_expr>
890 : 18337233 : sorted_array_from_bitmap_set (bitmap_set_t set)
891 : : {
892 : 18337233 : unsigned int i;
893 : 18337233 : bitmap_iterator bi;
894 : 18337233 : vec<pre_expr> result;
895 : :
896 : : /* Pre-allocate enough space for the array. */
897 : 18337233 : result.create (bitmap_count_bits (&set->expressions));
898 : :
899 : 18337233 : auto_bitmap val_visited (&grand_bitmap_obstack);
900 : 18337233 : bitmap_tree_view (val_visited);
901 : 101260715 : FOR_EACH_VALUE_ID_IN_SET (set, i, bi)
902 : 82923482 : if (bitmap_set_bit (val_visited, i))
903 : 73616125 : pre_expr_DFS (i, set, val_visited, result);
904 : :
905 : 18337233 : return result;
906 : 18337233 : }
907 : :
908 : : /* Subtract all expressions contained in ORIG from DEST. */
909 : :
910 : : static bitmap_set_t
911 : 32246510 : bitmap_set_subtract_expressions (bitmap_set_t dest, bitmap_set_t orig)
912 : : {
913 : 32246510 : bitmap_set_t result = bitmap_set_new ();
914 : 32246510 : bitmap_iterator bi;
915 : 32246510 : unsigned int i;
916 : :
917 : 32246510 : bitmap_and_compl (&result->expressions, &dest->expressions,
918 : 32246510 : &orig->expressions);
919 : :
920 : 109905411 : FOR_EACH_EXPR_ID_IN_SET (result, i, bi)
921 : : {
922 : 77658901 : pre_expr expr = expression_for_id (i);
923 : 77658901 : unsigned int value_id = get_expr_value_id (expr);
924 : 77658901 : bitmap_set_bit (&result->values, value_id);
925 : : }
926 : :
927 : 32246510 : return result;
928 : : }
929 : :
930 : : /* Subtract all values in bitmap set B from bitmap set A. */
931 : :
932 : : static void
933 : 1132698 : bitmap_set_subtract_values (bitmap_set_t a, bitmap_set_t b)
934 : : {
935 : 1132698 : unsigned int i;
936 : 1132698 : bitmap_iterator bi;
937 : 1132698 : unsigned to_remove = -1U;
938 : 1132698 : bitmap_and_compl_into (&a->values, &b->values);
939 : 10721823 : FOR_EACH_EXPR_ID_IN_SET (a, i, bi)
940 : : {
941 : 9589125 : if (to_remove != -1U)
942 : : {
943 : 1347125 : bitmap_clear_bit (&a->expressions, to_remove);
944 : 1347125 : to_remove = -1U;
945 : : }
946 : 9589125 : pre_expr expr = expression_for_id (i);
947 : 9589125 : if (! bitmap_bit_p (&a->values, get_expr_value_id (expr)))
948 : 1398748 : to_remove = i;
949 : : }
950 : 1132698 : if (to_remove != -1U)
951 : 51623 : bitmap_clear_bit (&a->expressions, to_remove);
952 : 1132698 : }
953 : :
954 : :
955 : : /* Return true if bitmapped set SET contains the value VALUE_ID. */
956 : :
957 : : static bool
958 : 190507403 : bitmap_set_contains_value (bitmap_set_t set, unsigned int value_id)
959 : : {
960 : 0 : if (value_id_constant_p (value_id))
961 : : return true;
962 : :
963 : 91888505 : return bitmap_bit_p (&set->values, value_id);
964 : : }
965 : :
966 : : /* Return true if two bitmap sets are equal. */
967 : :
968 : : static bool
969 : 15556906 : bitmap_set_equal (bitmap_set_t a, bitmap_set_t b)
970 : : {
971 : 0 : return bitmap_equal_p (&a->values, &b->values);
972 : : }
973 : :
974 : : /* Replace an instance of EXPR's VALUE with EXPR in SET if it exists,
975 : : and add it otherwise. Return true if any changes were made. */
976 : :
977 : : static bool
978 : 32161321 : bitmap_value_replace_in_set (bitmap_set_t set, pre_expr expr)
979 : : {
980 : 32161321 : unsigned int val = get_expr_value_id (expr);
981 : 32161321 : if (value_id_constant_p (val))
982 : : return false;
983 : :
984 : 32161321 : if (bitmap_set_contains_value (set, val))
985 : : {
986 : : /* The number of expressions having a given value is usually
987 : : significantly less than the total number of expressions in SET.
988 : : Thus, rather than check, for each expression in SET, whether it
989 : : has the value LOOKFOR, we walk the reverse mapping that tells us
990 : : what expressions have a given value, and see if any of those
991 : : expressions are in our set. For large testcases, this is about
992 : : 5-10x faster than walking the bitmap. If this is somehow a
993 : : significant lose for some cases, we can choose which set to walk
994 : : based on the set size. */
995 : 13725941 : unsigned int i;
996 : 13725941 : bitmap_iterator bi;
997 : 13725941 : bitmap exprset = value_expressions[val];
998 : 15605828 : EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
999 : : {
1000 : 15605828 : if (bitmap_clear_bit (&set->expressions, i))
1001 : : {
1002 : 13725941 : bitmap_set_bit (&set->expressions, get_expression_id (expr));
1003 : 13725941 : return i != get_expression_id (expr);
1004 : : }
1005 : : }
1006 : 0 : gcc_unreachable ();
1007 : : }
1008 : :
1009 : 18435380 : bitmap_insert_into_set (set, expr);
1010 : 18435380 : return true;
1011 : : }
1012 : :
1013 : : /* Insert EXPR into SET if EXPR's value is not already present in
1014 : : SET. */
1015 : :
1016 : : static void
1017 : 62074468 : bitmap_value_insert_into_set (bitmap_set_t set, pre_expr expr)
1018 : : {
1019 : 62074468 : unsigned int val = get_expr_value_id (expr);
1020 : :
1021 : 62074468 : gcc_checking_assert (expr->id == get_expression_id (expr));
1022 : :
1023 : : /* Constant values are always considered to be part of the set. */
1024 : 62074468 : if (value_id_constant_p (val))
1025 : : return;
1026 : :
1027 : : /* If the value membership changed, add the expression. */
1028 : 62012565 : if (bitmap_set_bit (&set->values, val))
1029 : 47927867 : bitmap_set_bit (&set->expressions, expr->id);
1030 : : }
1031 : :
1032 : : /* Print out EXPR to outfile. */
1033 : :
1034 : : static void
1035 : 4355 : print_pre_expr (FILE *outfile, const pre_expr expr)
1036 : : {
1037 : 4355 : if (! expr)
1038 : : {
1039 : 0 : fprintf (outfile, "NULL");
1040 : 0 : return;
1041 : : }
1042 : 4355 : switch (expr->kind)
1043 : : {
1044 : 0 : case CONSTANT:
1045 : 0 : print_generic_expr (outfile, PRE_EXPR_CONSTANT (expr));
1046 : 0 : break;
1047 : 3085 : case NAME:
1048 : 3085 : print_generic_expr (outfile, PRE_EXPR_NAME (expr));
1049 : 3085 : break;
1050 : 965 : case NARY:
1051 : 965 : {
1052 : 965 : unsigned int i;
1053 : 965 : vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1054 : 965 : fprintf (outfile, "{%s,", get_tree_code_name (nary->opcode));
1055 : 3688 : for (i = 0; i < nary->length; i++)
1056 : : {
1057 : 1758 : print_generic_expr (outfile, nary->op[i]);
1058 : 1758 : if (i != (unsigned) nary->length - 1)
1059 : 793 : fprintf (outfile, ",");
1060 : : }
1061 : 965 : fprintf (outfile, "}");
1062 : : }
1063 : 965 : break;
1064 : :
1065 : 305 : case REFERENCE:
1066 : 305 : {
1067 : 305 : vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1068 : 305 : print_vn_reference_ops (outfile, ref->operands);
1069 : 305 : if (ref->vuse)
1070 : : {
1071 : 293 : fprintf (outfile, "@");
1072 : 293 : print_generic_expr (outfile, ref->vuse);
1073 : : }
1074 : : }
1075 : : break;
1076 : : }
1077 : : }
1078 : : void debug_pre_expr (pre_expr);
1079 : :
1080 : : /* Like print_pre_expr but always prints to stderr. */
1081 : : DEBUG_FUNCTION void
1082 : 0 : debug_pre_expr (pre_expr e)
1083 : : {
1084 : 0 : print_pre_expr (stderr, e);
1085 : 0 : fprintf (stderr, "\n");
1086 : 0 : }
1087 : :
1088 : : /* Print out SET to OUTFILE. */
1089 : :
1090 : : static void
1091 : 880 : print_bitmap_set (FILE *outfile, bitmap_set_t set,
1092 : : const char *setname, int blockindex)
1093 : : {
1094 : 880 : fprintf (outfile, "%s[%d] := { ", setname, blockindex);
1095 : 880 : if (set)
1096 : : {
1097 : 880 : bool first = true;
1098 : 880 : unsigned i;
1099 : 880 : bitmap_iterator bi;
1100 : :
1101 : 5089 : FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1102 : : {
1103 : 4209 : const pre_expr expr = expression_for_id (i);
1104 : :
1105 : 4209 : if (!first)
1106 : 3606 : fprintf (outfile, ", ");
1107 : 4209 : first = false;
1108 : 4209 : print_pre_expr (outfile, expr);
1109 : :
1110 : 4209 : fprintf (outfile, " (%04d)", get_expr_value_id (expr));
1111 : : }
1112 : : }
1113 : 880 : fprintf (outfile, " }\n");
1114 : 880 : }
1115 : :
1116 : : void debug_bitmap_set (bitmap_set_t);
1117 : :
1118 : : DEBUG_FUNCTION void
1119 : 0 : debug_bitmap_set (bitmap_set_t set)
1120 : : {
1121 : 0 : print_bitmap_set (stderr, set, "debug", 0);
1122 : 0 : }
1123 : :
1124 : : void debug_bitmap_sets_for (basic_block);
1125 : :
1126 : : DEBUG_FUNCTION void
1127 : 0 : debug_bitmap_sets_for (basic_block bb)
1128 : : {
1129 : 0 : print_bitmap_set (stderr, AVAIL_OUT (bb), "avail_out", bb->index);
1130 : 0 : print_bitmap_set (stderr, EXP_GEN (bb), "exp_gen", bb->index);
1131 : 0 : print_bitmap_set (stderr, PHI_GEN (bb), "phi_gen", bb->index);
1132 : 0 : print_bitmap_set (stderr, TMP_GEN (bb), "tmp_gen", bb->index);
1133 : 0 : print_bitmap_set (stderr, ANTIC_IN (bb), "antic_in", bb->index);
1134 : 0 : if (do_partial_partial)
1135 : 0 : print_bitmap_set (stderr, PA_IN (bb), "pa_in", bb->index);
1136 : 0 : print_bitmap_set (stderr, NEW_SETS (bb), "new_sets", bb->index);
1137 : 0 : }
1138 : :
1139 : : /* Print out the expressions that have VAL to OUTFILE. */
1140 : :
1141 : : static void
1142 : 0 : print_value_expressions (FILE *outfile, unsigned int val)
1143 : : {
1144 : 0 : bitmap set = value_expressions[val];
1145 : 0 : if (set)
1146 : : {
1147 : 0 : bitmap_set x;
1148 : 0 : char s[10];
1149 : 0 : sprintf (s, "%04d", val);
1150 : 0 : x.expressions = *set;
1151 : 0 : print_bitmap_set (outfile, &x, s, 0);
1152 : : }
1153 : 0 : }
1154 : :
1155 : :
1156 : : DEBUG_FUNCTION void
1157 : 0 : debug_value_expressions (unsigned int val)
1158 : : {
1159 : 0 : print_value_expressions (stderr, val);
1160 : 0 : }
1161 : :
1162 : : /* Given a CONSTANT, allocate a new CONSTANT type PRE_EXPR to
1163 : : represent it. */
1164 : :
1165 : : static pre_expr
1166 : 4801100 : get_or_alloc_expr_for_constant (tree constant)
1167 : : {
1168 : 4801100 : unsigned int result_id;
1169 : 4801100 : struct pre_expr_d expr;
1170 : 4801100 : pre_expr newexpr;
1171 : :
1172 : 4801100 : expr.kind = CONSTANT;
1173 : 4801100 : PRE_EXPR_CONSTANT (&expr) = constant;
1174 : 4801100 : result_id = lookup_expression_id (&expr);
1175 : 4801100 : if (result_id != 0)
1176 : 3981203 : return expression_for_id (result_id);
1177 : :
1178 : 819897 : newexpr = pre_expr_pool.allocate ();
1179 : 819897 : newexpr->kind = CONSTANT;
1180 : 819897 : newexpr->loc = UNKNOWN_LOCATION;
1181 : 819897 : PRE_EXPR_CONSTANT (newexpr) = constant;
1182 : 819897 : alloc_expression_id (newexpr);
1183 : 819897 : newexpr->value_id = get_or_alloc_constant_value_id (constant);
1184 : 819897 : add_to_value (newexpr->value_id, newexpr);
1185 : 819897 : return newexpr;
1186 : : }
1187 : :
1188 : : /* Translate the VUSE backwards through phi nodes in E->dest, so that
1189 : : it has the value it would have in E->src. Set *SAME_VALID to true
1190 : : in case the new vuse doesn't change the value id of the OPERANDS. */
1191 : :
1192 : : static tree
1193 : 4361816 : translate_vuse_through_block (vec<vn_reference_op_s> operands,
1194 : : alias_set_type set, alias_set_type base_set,
1195 : : tree type, tree vuse, edge e, bool *same_valid)
1196 : : {
1197 : 4361816 : basic_block phiblock = e->dest;
1198 : 4361816 : gimple *phi = SSA_NAME_DEF_STMT (vuse);
1199 : 4361816 : ao_ref ref;
1200 : :
1201 : 4361816 : if (same_valid)
1202 : 3159041 : *same_valid = true;
1203 : :
1204 : : /* If value-numbering provided a memory state for this
1205 : : that dominates PHIBLOCK we can just use that. */
1206 : 4361816 : if (gimple_nop_p (phi)
1207 : 4361816 : || (gimple_bb (phi) != phiblock
1208 : 1115537 : && dominated_by_p (CDI_DOMINATORS, phiblock, gimple_bb (phi))))
1209 : 1803445 : return vuse;
1210 : :
1211 : : /* We have pruned expressions that are killed in PHIBLOCK via
1212 : : prune_clobbered_mems but we have not rewritten the VUSE to the one
1213 : : live at the start of the block. If there is no virtual PHI to translate
1214 : : through return the VUSE live at entry. Otherwise the VUSE to translate
1215 : : is the def of the virtual PHI node. */
1216 : 2558371 : phi = get_virtual_phi (phiblock);
1217 : 2558371 : if (!phi)
1218 : 88674 : return BB_LIVE_VOP_ON_EXIT
1219 : : (get_immediate_dominator (CDI_DOMINATORS, phiblock));
1220 : :
1221 : 2469697 : if (same_valid
1222 : 2469697 : && ao_ref_init_from_vn_reference (&ref, set, base_set, type, operands))
1223 : : {
1224 : 1809256 : bitmap visited = NULL;
1225 : : /* Try to find a vuse that dominates this phi node by skipping
1226 : : non-clobbering statements. */
1227 : 1809256 : unsigned int cnt = param_sccvn_max_alias_queries_per_access;
1228 : 1809256 : vuse = get_continuation_for_phi (phi, &ref, true,
1229 : : cnt, &visited, false, NULL, NULL);
1230 : 1809256 : if (visited)
1231 : 1801870 : BITMAP_FREE (visited);
1232 : : }
1233 : : else
1234 : : vuse = NULL_TREE;
1235 : : /* If we didn't find any, the value ID can't stay the same. */
1236 : 2469697 : if (!vuse && same_valid)
1237 : 1576546 : *same_valid = false;
1238 : :
1239 : : /* ??? We would like to return vuse here as this is the canonical
1240 : : upmost vdef that this reference is associated with. But during
1241 : : insertion of the references into the hash tables we only ever
1242 : : directly insert with their direct gimple_vuse, hence returning
1243 : : something else would make us not find the other expression. */
1244 : 2469697 : return PHI_ARG_DEF (phi, e->dest_idx);
1245 : : }
1246 : :
1247 : : /* Like bitmap_find_leader, but checks for the value existing in SET1 *or*
1248 : : SET2 *or* SET3. This is used to avoid making a set consisting of the union
1249 : : of PA_IN and ANTIC_IN during insert and phi-translation. */
1250 : :
1251 : : static inline pre_expr
1252 : 23142471 : find_leader_in_sets (unsigned int val, bitmap_set_t set1, bitmap_set_t set2,
1253 : : bitmap_set_t set3 = NULL)
1254 : : {
1255 : 23142471 : pre_expr result = NULL;
1256 : :
1257 : 23142471 : if (set1)
1258 : 23012687 : result = bitmap_find_leader (set1, val);
1259 : 23142471 : if (!result && set2)
1260 : 1416055 : result = bitmap_find_leader (set2, val);
1261 : 23142471 : if (!result && set3)
1262 : 0 : result = bitmap_find_leader (set3, val);
1263 : 23142471 : return result;
1264 : : }
1265 : :
1266 : : /* Get the tree type for our PRE expression e. */
1267 : :
1268 : : static tree
1269 : 7117211 : get_expr_type (const pre_expr e)
1270 : : {
1271 : 7117211 : switch (e->kind)
1272 : : {
1273 : 982264 : case NAME:
1274 : 982264 : return TREE_TYPE (PRE_EXPR_NAME (e));
1275 : 181980 : case CONSTANT:
1276 : 181980 : return TREE_TYPE (PRE_EXPR_CONSTANT (e));
1277 : 1318050 : case REFERENCE:
1278 : 1318050 : return PRE_EXPR_REFERENCE (e)->type;
1279 : 4634917 : case NARY:
1280 : 4634917 : return PRE_EXPR_NARY (e)->type;
1281 : : }
1282 : 0 : gcc_unreachable ();
1283 : : }
1284 : :
1285 : : /* Get a representative SSA_NAME for a given expression that is available in B.
1286 : : Since all of our sub-expressions are treated as values, we require
1287 : : them to be SSA_NAME's for simplicity.
1288 : : Prior versions of GVNPRE used to use "value handles" here, so that
1289 : : an expression would be VH.11 + VH.10 instead of d_3 + e_6. In
1290 : : either case, the operands are really values (IE we do not expect
1291 : : them to be usable without finding leaders). */
1292 : :
1293 : : static tree
1294 : 18711853 : get_representative_for (const pre_expr e, basic_block b = NULL)
1295 : : {
1296 : 18711853 : tree name, valnum = NULL_TREE;
1297 : 18711853 : unsigned int value_id = get_expr_value_id (e);
1298 : :
1299 : 18711853 : switch (e->kind)
1300 : : {
1301 : 8540593 : case NAME:
1302 : 8540593 : return PRE_EXPR_NAME (e);
1303 : 1848317 : case CONSTANT:
1304 : 1848317 : return PRE_EXPR_CONSTANT (e);
1305 : 8322943 : case NARY:
1306 : 8322943 : case REFERENCE:
1307 : 8322943 : {
1308 : : /* Go through all of the expressions representing this value
1309 : : and pick out an SSA_NAME. */
1310 : 8322943 : unsigned int i;
1311 : 8322943 : bitmap_iterator bi;
1312 : 8322943 : bitmap exprs = value_expressions[value_id];
1313 : 21425310 : EXECUTE_IF_SET_IN_BITMAP (exprs, 0, i, bi)
1314 : : {
1315 : 17560298 : pre_expr rep = expression_for_id (i);
1316 : 17560298 : if (rep->kind == NAME)
1317 : : {
1318 : 8005233 : tree name = PRE_EXPR_NAME (rep);
1319 : 8005233 : valnum = VN_INFO (name)->valnum;
1320 : 8005233 : gimple *def = SSA_NAME_DEF_STMT (name);
1321 : : /* We have to return either a new representative or one
1322 : : that can be used for expression simplification and thus
1323 : : is available in B. */
1324 : 8005233 : if (! b
1325 : 7738815 : || gimple_nop_p (def)
1326 : 11823636 : || dominated_by_p (CDI_DOMINATORS, b, gimple_bb (def)))
1327 : 4457931 : return name;
1328 : : }
1329 : 9555065 : else if (rep->kind == CONSTANT)
1330 : 0 : return PRE_EXPR_CONSTANT (rep);
1331 : : }
1332 : : }
1333 : 3865012 : break;
1334 : : }
1335 : :
1336 : : /* If we reached here we couldn't find an SSA_NAME. This can
1337 : : happen when we've discovered a value that has never appeared in
1338 : : the program as set to an SSA_NAME, as the result of phi translation.
1339 : : Create one here.
1340 : : ??? We should be able to re-use this when we insert the statement
1341 : : to compute it. */
1342 : 3865012 : name = make_temp_ssa_name (get_expr_type (e), gimple_build_nop (), "pretmp");
1343 : 3865012 : vn_ssa_aux_t vn_info = VN_INFO (name);
1344 : 3865012 : vn_info->value_id = value_id;
1345 : 3865012 : vn_info->valnum = valnum ? valnum : name;
1346 : 3865012 : vn_info->visited = true;
1347 : : /* ??? For now mark this SSA name for release by VN. */
1348 : 3865012 : vn_info->needs_insertion = true;
1349 : 3865012 : add_to_value (value_id, get_or_alloc_expr_for_name (name));
1350 : 3865012 : if (dump_file && (dump_flags & TDF_DETAILS))
1351 : : {
1352 : 47 : fprintf (dump_file, "Created SSA_NAME representative ");
1353 : 47 : print_generic_expr (dump_file, name);
1354 : 47 : fprintf (dump_file, " for expression:");
1355 : 47 : print_pre_expr (dump_file, e);
1356 : 47 : fprintf (dump_file, " (%04d)\n", value_id);
1357 : : }
1358 : :
1359 : : return name;
1360 : : }
1361 : :
1362 : :
1363 : : static pre_expr
1364 : : phi_translate (bitmap_set_t, pre_expr, bitmap_set_t, bitmap_set_t, edge);
1365 : :
1366 : : /* Translate EXPR using phis in PHIBLOCK, so that it has the values of
1367 : : the phis in PRED. Return NULL if we can't find a leader for each part
1368 : : of the translated expression. */
1369 : :
1370 : : static pre_expr
1371 : 46410815 : phi_translate_1 (bitmap_set_t dest,
1372 : : pre_expr expr, bitmap_set_t set1, bitmap_set_t set2, edge e)
1373 : : {
1374 : 46410815 : basic_block pred = e->src;
1375 : 46410815 : basic_block phiblock = e->dest;
1376 : 46410815 : location_t expr_loc = expr->loc;
1377 : 46410815 : switch (expr->kind)
1378 : : {
1379 : 17579823 : case NARY:
1380 : 17579823 : {
1381 : 17579823 : unsigned int i;
1382 : 17579823 : bool changed = false;
1383 : 17579823 : vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1384 : 17579823 : vn_nary_op_t newnary = XALLOCAVAR (struct vn_nary_op_s,
1385 : : sizeof_vn_nary_op (nary->length));
1386 : 17579823 : memcpy (newnary, nary, sizeof_vn_nary_op (nary->length));
1387 : :
1388 : 41893942 : for (i = 0; i < newnary->length; i++)
1389 : : {
1390 : 27439856 : if (TREE_CODE (newnary->op[i]) != SSA_NAME)
1391 : 8563179 : continue;
1392 : : else
1393 : : {
1394 : 18876677 : pre_expr leader, result;
1395 : 18876677 : unsigned int op_val_id = VN_INFO (newnary->op[i])->value_id;
1396 : 18876677 : leader = find_leader_in_sets (op_val_id, set1, set2);
1397 : 18876677 : result = phi_translate (dest, leader, set1, set2, e);
1398 : 18876677 : if (result)
1399 : : /* If op has a leader in the sets we translate make
1400 : : sure to use the value of the translated expression.
1401 : : We might need a new representative for that. */
1402 : 15750940 : newnary->op[i] = get_representative_for (result, pred);
1403 : : else if (!result)
1404 : : return NULL;
1405 : :
1406 : 15750940 : changed |= newnary->op[i] != nary->op[i];
1407 : : }
1408 : : }
1409 : 14454086 : if (changed)
1410 : : {
1411 : 7280366 : unsigned int new_val_id;
1412 : :
1413 : : /* Try to simplify the new NARY. */
1414 : 7280366 : tree res = vn_nary_simplify (newnary);
1415 : 7280366 : if (res)
1416 : : {
1417 : 2322525 : if (is_gimple_min_invariant (res))
1418 : 1212676 : return get_or_alloc_expr_for_constant (res);
1419 : :
1420 : : /* For non-CONSTANTs we have to make sure we can eventually
1421 : : insert the expression. Which means we need to have a
1422 : : leader for it. */
1423 : 1109849 : gcc_assert (TREE_CODE (res) == SSA_NAME);
1424 : :
1425 : : /* Do not allow simplifications to non-constants over
1426 : : backedges as this will likely result in a loop PHI node
1427 : : to be inserted and increased register pressure.
1428 : : See PR77498 - this avoids doing predcoms work in
1429 : : a less efficient way. */
1430 : 1109849 : if (e->flags & EDGE_DFS_BACK)
1431 : : ;
1432 : : else
1433 : : {
1434 : 1025855 : unsigned value_id = VN_INFO (res)->value_id;
1435 : : /* We want a leader in ANTIC_OUT or AVAIL_OUT here.
1436 : : dest has what we computed into ANTIC_OUT sofar
1437 : : so pick from that - since topological sorting
1438 : : by sorted_array_from_bitmap_set isn't perfect
1439 : : we may lose some cases here. */
1440 : 2051710 : pre_expr constant = find_leader_in_sets (value_id, dest,
1441 : 1025855 : AVAIL_OUT (pred));
1442 : 1025855 : if (constant)
1443 : : {
1444 : 309783 : if (dump_file && (dump_flags & TDF_DETAILS))
1445 : : {
1446 : 7 : fprintf (dump_file, "simplifying ");
1447 : 7 : print_pre_expr (dump_file, expr);
1448 : 7 : fprintf (dump_file, " translated %d -> %d to ",
1449 : : phiblock->index, pred->index);
1450 : 7 : PRE_EXPR_NARY (expr) = newnary;
1451 : 7 : print_pre_expr (dump_file, expr);
1452 : 7 : PRE_EXPR_NARY (expr) = nary;
1453 : 7 : fprintf (dump_file, " to ");
1454 : 7 : print_pre_expr (dump_file, constant);
1455 : 7 : fprintf (dump_file, "\n");
1456 : : }
1457 : 309783 : return constant;
1458 : : }
1459 : : }
1460 : : }
1461 : :
1462 : 11515814 : tree result = vn_nary_op_lookup_pieces (newnary->length,
1463 : 5757907 : newnary->opcode,
1464 : : newnary->type,
1465 : : &newnary->op[0],
1466 : : &nary);
1467 : 5757907 : if (result && is_gimple_min_invariant (result))
1468 : 0 : return get_or_alloc_expr_for_constant (result);
1469 : :
1470 : 5757907 : if (!nary || nary->predicated_values)
1471 : : new_val_id = 0;
1472 : : else
1473 : 770267 : new_val_id = nary->value_id;
1474 : 5757907 : expr = get_or_alloc_expr_for_nary (newnary, new_val_id, expr_loc);
1475 : 5757907 : add_to_value (get_expr_value_id (expr), expr);
1476 : : }
1477 : : return expr;
1478 : : }
1479 : 4737394 : break;
1480 : :
1481 : 4737394 : case REFERENCE:
1482 : 4737394 : {
1483 : 4737394 : vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1484 : 4737394 : vec<vn_reference_op_s> operands = ref->operands;
1485 : 4737394 : tree vuse = ref->vuse;
1486 : 4737394 : tree newvuse = vuse;
1487 : 4737394 : vec<vn_reference_op_s> newoperands = vNULL;
1488 : 4737394 : bool changed = false, same_valid = true;
1489 : 4737394 : unsigned int i, n;
1490 : 4737394 : vn_reference_op_t operand;
1491 : 4737394 : vn_reference_t newref;
1492 : :
1493 : 18181190 : for (i = 0; operands.iterate (i, &operand); i++)
1494 : : {
1495 : 13722822 : pre_expr opresult;
1496 : 13722822 : pre_expr leader;
1497 : 13722822 : tree op[3];
1498 : 13722822 : tree type = operand->type;
1499 : 13722822 : vn_reference_op_s newop = *operand;
1500 : 13722822 : op[0] = operand->op0;
1501 : 13722822 : op[1] = operand->op1;
1502 : 13722822 : op[2] = operand->op2;
1503 : 54054270 : for (n = 0; n < 3; ++n)
1504 : : {
1505 : 40610474 : unsigned int op_val_id;
1506 : 40610474 : if (!op[n])
1507 : 24625181 : continue;
1508 : 15985293 : if (TREE_CODE (op[n]) != SSA_NAME)
1509 : : {
1510 : : /* We can't possibly insert these. */
1511 : 12745354 : if (n != 0
1512 : 12745354 : && !is_gimple_min_invariant (op[n]))
1513 : : break;
1514 : 12745354 : continue;
1515 : : }
1516 : 3239939 : op_val_id = VN_INFO (op[n])->value_id;
1517 : 3239939 : leader = find_leader_in_sets (op_val_id, set1, set2);
1518 : 3239939 : opresult = phi_translate (dest, leader, set1, set2, e);
1519 : 3239939 : if (opresult)
1520 : : {
1521 : 2960913 : tree name = get_representative_for (opresult);
1522 : 2960913 : changed |= name != op[n];
1523 : 2960913 : op[n] = name;
1524 : : }
1525 : : else if (!opresult)
1526 : : break;
1527 : : }
1528 : 13722822 : if (n != 3)
1529 : : {
1530 : 279026 : newoperands.release ();
1531 : 279026 : return NULL;
1532 : : }
1533 : : /* When we translate a MEM_REF across a backedge and we have
1534 : : restrict info that's not from our functions parameters
1535 : : we have to remap it since we now may deal with a different
1536 : : instance where the dependence info is no longer valid.
1537 : : See PR102970. Note instead of keeping a remapping table
1538 : : per backedge we simply throw away restrict info. */
1539 : 13443796 : if ((newop.opcode == MEM_REF
1540 : 13443796 : || newop.opcode == TARGET_MEM_REF)
1541 : 4514504 : && newop.clique > 1
1542 : 150939 : && (e->flags & EDGE_DFS_BACK))
1543 : : {
1544 : : newop.clique = 0;
1545 : : newop.base = 0;
1546 : : changed = true;
1547 : : }
1548 : 13425428 : if (!changed)
1549 : 11059689 : continue;
1550 : 2384107 : if (!newoperands.exists ())
1551 : 1228753 : newoperands = operands.copy ();
1552 : : /* We may have changed from an SSA_NAME to a constant */
1553 : 2384107 : if (newop.opcode == SSA_NAME && TREE_CODE (op[0]) != SSA_NAME)
1554 : : newop.opcode = TREE_CODE (op[0]);
1555 : 2384107 : newop.type = type;
1556 : 2384107 : newop.op0 = op[0];
1557 : 2384107 : newop.op1 = op[1];
1558 : 2384107 : newop.op2 = op[2];
1559 : 2384107 : newoperands[i] = newop;
1560 : : }
1561 : 8916736 : gcc_checking_assert (i == operands.length ());
1562 : :
1563 : 4458368 : if (vuse)
1564 : : {
1565 : 10679898 : newvuse = translate_vuse_through_block (newoperands.exists ()
1566 : 4361816 : ? newoperands : operands,
1567 : : ref->set, ref->base_set,
1568 : : ref->type, vuse, e,
1569 : : changed
1570 : : ? NULL : &same_valid);
1571 : 4361816 : if (newvuse == NULL_TREE)
1572 : : {
1573 : 0 : newoperands.release ();
1574 : 0 : return NULL;
1575 : : }
1576 : : }
1577 : :
1578 : 4458368 : if (changed || newvuse != vuse)
1579 : : {
1580 : 3143070 : unsigned int new_val_id;
1581 : :
1582 : 5058318 : tree result = vn_reference_lookup_pieces (newvuse, ref->set,
1583 : : ref->base_set,
1584 : : ref->type,
1585 : 3143070 : newoperands.exists ()
1586 : 3143070 : ? newoperands : operands,
1587 : : &newref, VN_WALK);
1588 : 3143070 : if (result)
1589 : 641240 : newoperands.release ();
1590 : :
1591 : : /* We can always insert constants, so if we have a partial
1592 : : redundant constant load of another type try to translate it
1593 : : to a constant of appropriate type. */
1594 : 641240 : if (result && is_gimple_min_invariant (result))
1595 : : {
1596 : 69732 : tree tem = result;
1597 : 69732 : if (!useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1598 : : {
1599 : 82 : tem = fold_unary (VIEW_CONVERT_EXPR, ref->type, result);
1600 : 82 : if (tem && !is_gimple_min_invariant (tem))
1601 : : tem = NULL_TREE;
1602 : : }
1603 : 69732 : if (tem)
1604 : 69732 : return get_or_alloc_expr_for_constant (tem);
1605 : : }
1606 : :
1607 : : /* If we'd have to convert things we would need to validate
1608 : : if we can insert the translated expression. So fail
1609 : : here for now - we cannot insert an alias with a different
1610 : : type in the VN tables either, as that would assert. */
1611 : 3073338 : if (result
1612 : 3073338 : && !useless_type_conversion_p (ref->type, TREE_TYPE (result)))
1613 : : return NULL;
1614 : 2501830 : else if (!result && newref
1615 : 3258260 : && !useless_type_conversion_p (ref->type, newref->type))
1616 : : {
1617 : 224 : newoperands.release ();
1618 : 224 : return NULL;
1619 : : }
1620 : :
1621 : 3072112 : if (newref)
1622 : 756206 : new_val_id = newref->value_id;
1623 : : else
1624 : : {
1625 : 2315906 : if (changed || !same_valid)
1626 : 2254542 : new_val_id = get_next_value_id ();
1627 : : else
1628 : 61364 : new_val_id = ref->value_id;
1629 : 2315906 : if (!newoperands.exists ())
1630 : 1288019 : newoperands = operands.copy ();
1631 : 2315906 : newref = vn_reference_insert_pieces (newvuse, ref->set,
1632 : : ref->base_set,
1633 : : ref->offset, ref->max_size,
1634 : : ref->type, newoperands,
1635 : : result, new_val_id);
1636 : 2315906 : newoperands = vNULL;
1637 : : }
1638 : 3072112 : expr = get_or_alloc_expr_for_reference (newref, expr_loc);
1639 : 3072112 : add_to_value (new_val_id, expr);
1640 : : }
1641 : 4387410 : newoperands.release ();
1642 : 4387410 : return expr;
1643 : : }
1644 : 24093598 : break;
1645 : :
1646 : 24093598 : case NAME:
1647 : 24093598 : {
1648 : 24093598 : tree name = PRE_EXPR_NAME (expr);
1649 : 24093598 : gimple *def_stmt = SSA_NAME_DEF_STMT (name);
1650 : : /* If the SSA name is defined by a PHI node in this block,
1651 : : translate it. */
1652 : 24093598 : if (gimple_code (def_stmt) == GIMPLE_PHI
1653 : 24093598 : && gimple_bb (def_stmt) == phiblock)
1654 : : {
1655 : 7555084 : tree def = PHI_ARG_DEF (def_stmt, e->dest_idx);
1656 : :
1657 : : /* Handle constant. */
1658 : 7555084 : if (is_gimple_min_invariant (def))
1659 : 2198390 : return get_or_alloc_expr_for_constant (def);
1660 : :
1661 : 5356694 : return get_or_alloc_expr_for_name (def);
1662 : : }
1663 : : /* Otherwise return it unchanged - it will get removed if its
1664 : : value is not available in PREDs AVAIL_OUT set of expressions
1665 : : by the subtraction of TMP_GEN. */
1666 : : return expr;
1667 : : }
1668 : :
1669 : 0 : default:
1670 : 0 : gcc_unreachable ();
1671 : : }
1672 : : }
1673 : :
1674 : : /* Wrapper around phi_translate_1 providing caching functionality. */
1675 : :
1676 : : static pre_expr
1677 : 85561999 : phi_translate (bitmap_set_t dest, pre_expr expr,
1678 : : bitmap_set_t set1, bitmap_set_t set2, edge e)
1679 : : {
1680 : 85561999 : expr_pred_trans_t slot = NULL;
1681 : 85561999 : pre_expr phitrans;
1682 : :
1683 : 85561999 : if (!expr)
1684 : : return NULL;
1685 : :
1686 : : /* Constants contain no values that need translation. */
1687 : 83844043 : if (expr->kind == CONSTANT)
1688 : : return expr;
1689 : :
1690 : 83843947 : if (value_id_constant_p (get_expr_value_id (expr)))
1691 : : return expr;
1692 : :
1693 : : /* Don't add translations of NAMEs as those are cheap to translate. */
1694 : 83843947 : if (expr->kind != NAME)
1695 : : {
1696 : 59750349 : if (phi_trans_add (&slot, expr, e->src))
1697 : 37433132 : return slot->v == 0 ? NULL : expression_for_id (slot->v);
1698 : : /* Store NULL for the value we want to return in the case of
1699 : : recursing. */
1700 : 22317217 : slot->v = 0;
1701 : : }
1702 : :
1703 : : /* Translate. */
1704 : 46410815 : basic_block saved_valueize_bb = vn_context_bb;
1705 : 46410815 : vn_context_bb = e->src;
1706 : 46410815 : phitrans = phi_translate_1 (dest, expr, set1, set2, e);
1707 : 46410815 : vn_context_bb = saved_valueize_bb;
1708 : :
1709 : 46410815 : if (slot)
1710 : : {
1711 : : /* We may have reallocated. */
1712 : 22317217 : phi_trans_add (&slot, expr, e->src);
1713 : 22317217 : if (phitrans)
1714 : 18911228 : slot->v = get_expression_id (phitrans);
1715 : : else
1716 : : /* Remove failed translations again, they cause insert
1717 : : iteration to not pick up new opportunities reliably. */
1718 : 3405989 : PHI_TRANS_TABLE (e->src)->clear_slot (slot);
1719 : : }
1720 : :
1721 : : return phitrans;
1722 : : }
1723 : :
1724 : :
1725 : : /* For each expression in SET, translate the values through phi nodes
1726 : : in PHIBLOCK using edge PHIBLOCK->PRED, and store the resulting
1727 : : expressions in DEST. */
1728 : :
1729 : : static void
1730 : 20322412 : phi_translate_set (bitmap_set_t dest, bitmap_set_t set, edge e)
1731 : : {
1732 : 20322412 : bitmap_iterator bi;
1733 : 20322412 : unsigned int i;
1734 : :
1735 : 20322412 : if (gimple_seq_empty_p (phi_nodes (e->dest)))
1736 : : {
1737 : 13696600 : bitmap_set_copy (dest, set);
1738 : 13696600 : return;
1739 : : }
1740 : :
1741 : : /* Allocate the phi-translation cache where we have an idea about
1742 : : its size. hash-table implementation internals tell us that
1743 : : allocating the table to fit twice the number of elements will
1744 : : make sure we do not usually re-allocate. */
1745 : 6625812 : if (!PHI_TRANS_TABLE (e->src))
1746 : 5944814 : PHI_TRANS_TABLE (e->src) = new hash_table<expr_pred_trans_d>
1747 : 5944814 : (2 * bitmap_count_bits (&set->expressions));
1748 : 42811095 : FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1749 : : {
1750 : 36185283 : pre_expr expr = expression_for_id (i);
1751 : 36185283 : pre_expr translated = phi_translate (dest, expr, set, NULL, e);
1752 : 36185283 : if (!translated)
1753 : 1718334 : continue;
1754 : :
1755 : 34466949 : bitmap_insert_into_set (dest, translated);
1756 : : }
1757 : : }
1758 : :
1759 : : /* Find the leader for a value (i.e., the name representing that
1760 : : value) in a given set, and return it. Return NULL if no leader
1761 : : is found. */
1762 : :
1763 : : static pre_expr
1764 : 54699553 : bitmap_find_leader (bitmap_set_t set, unsigned int val)
1765 : : {
1766 : 54699553 : if (value_id_constant_p (val))
1767 : 1695569 : return constant_value_expressions[-val];
1768 : :
1769 : 53003984 : if (bitmap_set_contains_value (set, val))
1770 : : {
1771 : : /* Rather than walk the entire bitmap of expressions, and see
1772 : : whether any of them has the value we are looking for, we look
1773 : : at the reverse mapping, which tells us the set of expressions
1774 : : that have a given value (IE value->expressions with that
1775 : : value) and see if any of those expressions are in our set.
1776 : : The number of expressions per value is usually significantly
1777 : : less than the number of expressions in the set. In fact, for
1778 : : large testcases, doing it this way is roughly 5-10x faster
1779 : : than walking the bitmap.
1780 : : If this is somehow a significant lose for some cases, we can
1781 : : choose which set to walk based on which set is smaller. */
1782 : 24589189 : unsigned int i;
1783 : 24589189 : bitmap_iterator bi;
1784 : 24589189 : bitmap exprset = value_expressions[val];
1785 : :
1786 : 24589189 : if (!exprset->first->next)
1787 : 31377907 : EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
1788 : 29062068 : if (bitmap_bit_p (&set->expressions, i))
1789 : 22221094 : return expression_for_id (i);
1790 : :
1791 : 6002514 : EXECUTE_IF_AND_IN_BITMAP (exprset, &set->expressions, 0, i, bi)
1792 : 3634419 : return expression_for_id (i);
1793 : : }
1794 : : return NULL;
1795 : : }
1796 : :
1797 : : /* Determine if EXPR, a memory expression, is ANTIC_IN at the top of
1798 : : BLOCK by seeing if it is not killed in the block. Note that we are
1799 : : only determining whether there is a store that kills it. Because
1800 : : of the order in which clean iterates over values, we are guaranteed
1801 : : that altered operands will have caused us to be eliminated from the
1802 : : ANTIC_IN set already. */
1803 : :
1804 : : static bool
1805 : 1561268 : value_dies_in_block_x (pre_expr expr, basic_block block)
1806 : : {
1807 : 1561268 : tree vuse = PRE_EXPR_REFERENCE (expr)->vuse;
1808 : 1561268 : vn_reference_t refx = PRE_EXPR_REFERENCE (expr);
1809 : 1561268 : gimple *def;
1810 : 1561268 : gimple_stmt_iterator gsi;
1811 : 1561268 : unsigned id = get_expression_id (expr);
1812 : 1561268 : bool res = false;
1813 : 1561268 : ao_ref ref;
1814 : :
1815 : 1561268 : if (!vuse)
1816 : : return false;
1817 : :
1818 : : /* Lookup a previously calculated result. */
1819 : 1561268 : if (EXPR_DIES (block)
1820 : 1561268 : && bitmap_bit_p (EXPR_DIES (block), id * 2))
1821 : 136750 : return bitmap_bit_p (EXPR_DIES (block), id * 2 + 1);
1822 : :
1823 : : /* A memory expression {e, VUSE} dies in the block if there is a
1824 : : statement that may clobber e. If, starting statement walk from the
1825 : : top of the basic block, a statement uses VUSE there can be no kill
1826 : : inbetween that use and the original statement that loaded {e, VUSE},
1827 : : so we can stop walking. */
1828 : 1424518 : ref.base = NULL_TREE;
1829 : 12335855 : for (gsi = gsi_start_bb (block); !gsi_end_p (gsi); gsi_next (&gsi))
1830 : : {
1831 : 10474579 : tree def_vuse, def_vdef;
1832 : 10474579 : def = gsi_stmt (gsi);
1833 : 10474579 : def_vuse = gimple_vuse (def);
1834 : 10474579 : def_vdef = gimple_vdef (def);
1835 : :
1836 : : /* Not a memory statement. */
1837 : 10474579 : if (!def_vuse)
1838 : 7357349 : continue;
1839 : :
1840 : : /* Not a may-def. */
1841 : 3117230 : if (!def_vdef)
1842 : : {
1843 : : /* A load with the same VUSE, we're done. */
1844 : 906828 : if (def_vuse == vuse)
1845 : : break;
1846 : :
1847 : 631088 : continue;
1848 : : }
1849 : :
1850 : : /* Init ref only if we really need it. */
1851 : 2210402 : if (ref.base == NULL_TREE
1852 : 3291615 : && !ao_ref_init_from_vn_reference (&ref, refx->set, refx->base_set,
1853 : 1081213 : refx->type, refx->operands))
1854 : : {
1855 : : res = true;
1856 : : break;
1857 : : }
1858 : : /* If the statement may clobber expr, it dies. */
1859 : 2177852 : if (stmt_may_clobber_ref_p_1 (def, &ref))
1860 : : {
1861 : : res = true;
1862 : : break;
1863 : : }
1864 : : }
1865 : :
1866 : : /* Remember the result. */
1867 : 1424518 : if (!EXPR_DIES (block))
1868 : 687927 : EXPR_DIES (block) = BITMAP_ALLOC (&grand_bitmap_obstack);
1869 : 1424518 : bitmap_set_bit (EXPR_DIES (block), id * 2);
1870 : 1424518 : if (res)
1871 : 712020 : bitmap_set_bit (EXPR_DIES (block), id * 2 + 1);
1872 : :
1873 : : return res;
1874 : : }
1875 : :
1876 : :
1877 : : /* Determine if OP is valid in SET1 U SET2, which it is when the union
1878 : : contains its value-id. */
1879 : :
1880 : : static bool
1881 : 255758159 : op_valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, tree op)
1882 : : {
1883 : 255758159 : if (op && TREE_CODE (op) == SSA_NAME)
1884 : : {
1885 : 75372366 : unsigned int value_id = VN_INFO (op)->value_id;
1886 : 150742640 : if (!(bitmap_set_contains_value (set1, value_id)
1887 : 2029340 : || (set2 && bitmap_set_contains_value (set2, value_id))))
1888 : 2378034 : return false;
1889 : : }
1890 : : return true;
1891 : : }
1892 : :
1893 : : /* Determine if the expression EXPR is valid in SET1 U SET2.
1894 : : ONLY SET2 CAN BE NULL.
1895 : : This means that we have a leader for each part of the expression
1896 : : (if it consists of values), or the expression is an SSA_NAME.
1897 : : For loads/calls, we also see if the vuse is killed in this block. */
1898 : :
1899 : : static bool
1900 : 120087449 : valid_in_sets (bitmap_set_t set1, bitmap_set_t set2, pre_expr expr)
1901 : : {
1902 : 120087449 : switch (expr->kind)
1903 : : {
1904 : : case NAME:
1905 : : /* By construction all NAMEs are available. Non-available
1906 : : NAMEs are removed by subtracting TMP_GEN from the sets. */
1907 : : return true;
1908 : 54626209 : case NARY:
1909 : 54626209 : {
1910 : 54626209 : unsigned int i;
1911 : 54626209 : vn_nary_op_t nary = PRE_EXPR_NARY (expr);
1912 : 142683190 : for (i = 0; i < nary->length; i++)
1913 : 90218542 : if (!op_valid_in_sets (set1, set2, nary->op[i]))
1914 : : return false;
1915 : : return true;
1916 : : }
1917 : 18931375 : break;
1918 : 18931375 : case REFERENCE:
1919 : 18931375 : {
1920 : 18931375 : vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1921 : 18931375 : vn_reference_op_t vro;
1922 : 18931375 : unsigned int i;
1923 : :
1924 : 74039063 : FOR_EACH_VEC_ELT (ref->operands, i, vro)
1925 : : {
1926 : 55324161 : if (!op_valid_in_sets (set1, set2, vro->op0)
1927 : 55107728 : || !op_valid_in_sets (set1, set2, vro->op1)
1928 : 110431889 : || !op_valid_in_sets (set1, set2, vro->op2))
1929 : 216473 : return false;
1930 : : }
1931 : : return true;
1932 : : }
1933 : 0 : default:
1934 : 0 : gcc_unreachable ();
1935 : : }
1936 : : }
1937 : :
1938 : : /* Clean the set of expressions SET1 that are no longer valid in SET1 or SET2.
1939 : : This means expressions that are made up of values we have no leaders for
1940 : : in SET1 or SET2. */
1941 : :
1942 : : static void
1943 : 14463532 : clean (bitmap_set_t set1, bitmap_set_t set2 = NULL)
1944 : : {
1945 : 14463532 : vec<pre_expr> exprs = sorted_array_from_bitmap_set (set1);
1946 : 14463532 : pre_expr expr;
1947 : 14463532 : int i;
1948 : :
1949 : 75584758 : FOR_EACH_VEC_ELT (exprs, i, expr)
1950 : : {
1951 : 61121226 : if (!valid_in_sets (set1, set2, expr))
1952 : : {
1953 : 2378016 : unsigned int val = get_expr_value_id (expr);
1954 : 2378016 : bitmap_clear_bit (&set1->expressions, get_expression_id (expr));
1955 : : /* We are entered with possibly multiple expressions for a value
1956 : : so before removing a value from the set see if there's an
1957 : : expression for it left. */
1958 : 2378016 : if (! bitmap_find_leader (set1, val))
1959 : 2368095 : bitmap_clear_bit (&set1->values, val);
1960 : : }
1961 : : }
1962 : 14463532 : exprs.release ();
1963 : :
1964 : 14463532 : if (flag_checking)
1965 : : {
1966 : 14463345 : unsigned j;
1967 : 14463345 : bitmap_iterator bi;
1968 : 73206356 : FOR_EACH_EXPR_ID_IN_SET (set1, j, bi)
1969 : 58743011 : gcc_assert (valid_in_sets (set1, set2, expression_for_id (j)));
1970 : : }
1971 : 14463532 : }
1972 : :
1973 : : /* Clean the set of expressions that are no longer valid in SET because
1974 : : they are clobbered in BLOCK or because they trap and may not be executed.
1975 : : When CLEAN_TRAPS is true remove all possibly trapping expressions. */
1976 : :
1977 : : static void
1978 : 16689604 : prune_clobbered_mems (bitmap_set_t set, basic_block block, bool clean_traps)
1979 : : {
1980 : 16689604 : bitmap_iterator bi;
1981 : 16689604 : unsigned i;
1982 : 16689604 : unsigned to_remove = -1U;
1983 : 16689604 : bool any_removed = false;
1984 : :
1985 : 73787051 : FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
1986 : : {
1987 : : /* Remove queued expr. */
1988 : 57097447 : if (to_remove != -1U)
1989 : : {
1990 : 565994 : bitmap_clear_bit (&set->expressions, to_remove);
1991 : 565994 : any_removed = true;
1992 : 565994 : to_remove = -1U;
1993 : : }
1994 : :
1995 : 57097447 : pre_expr expr = expression_for_id (i);
1996 : 57097447 : if (expr->kind == REFERENCE)
1997 : : {
1998 : 7695965 : vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
1999 : 7695965 : if (ref->vuse)
2000 : : {
2001 : 6966221 : gimple *def_stmt = SSA_NAME_DEF_STMT (ref->vuse);
2002 : 6966221 : if (!gimple_nop_p (def_stmt)
2003 : : /* If value-numbering provided a memory state for this
2004 : : that dominates BLOCK we're done, otherwise we have
2005 : : to check if the value dies in BLOCK. */
2006 : 8622025 : && !(gimple_bb (def_stmt) != block
2007 : 3648164 : && dominated_by_p (CDI_DOMINATORS,
2008 : 3648164 : block, gimple_bb (def_stmt)))
2009 : 8527489 : && value_dies_in_block_x (expr, block))
2010 : : to_remove = i;
2011 : : }
2012 : : /* If the REFERENCE may trap make sure the block does not contain
2013 : : a possible exit point.
2014 : : ??? This is overly conservative if we translate AVAIL_OUT
2015 : : as the available expression might be after the exit point. */
2016 : 7068740 : if ((BB_MAY_NOTRETURN (block) || clean_traps)
2017 : 7958499 : && vn_reference_may_trap (ref))
2018 : : to_remove = i;
2019 : : }
2020 : 49401482 : else if (expr->kind == NARY)
2021 : : {
2022 : 26339661 : vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2023 : : /* If the NARY may trap make sure the block does not contain
2024 : : a possible exit point.
2025 : : ??? This is overly conservative if we translate AVAIL_OUT
2026 : : as the available expression might be after the exit point. */
2027 : 22502993 : if ((BB_MAY_NOTRETURN (block) || clean_traps)
2028 : 27204936 : && vn_nary_may_trap (nary))
2029 : : to_remove = i;
2030 : : }
2031 : : }
2032 : :
2033 : : /* Remove queued expr. */
2034 : 16689604 : if (to_remove != -1U)
2035 : : {
2036 : 403438 : bitmap_clear_bit (&set->expressions, to_remove);
2037 : 403438 : any_removed = true;
2038 : : }
2039 : :
2040 : : /* Above we only removed expressions, now clean the set of values
2041 : : which no longer have any corresponding expression. We cannot
2042 : : clear the value at the time we remove an expression since there
2043 : : may be multiple expressions per value.
2044 : : If we'd queue possibly to be removed values we could use
2045 : : the bitmap_find_leader way to see if there's still an expression
2046 : : for it. For some ratio of to be removed values and number of
2047 : : values/expressions in the set this might be faster than rebuilding
2048 : : the value-set. */
2049 : 16689604 : if (any_removed)
2050 : : {
2051 : 585694 : bitmap_clear (&set->values);
2052 : 3297038 : FOR_EACH_EXPR_ID_IN_SET (set, i, bi)
2053 : : {
2054 : 2711344 : pre_expr expr = expression_for_id (i);
2055 : 2711344 : unsigned int value_id = get_expr_value_id (expr);
2056 : 2711344 : bitmap_set_bit (&set->values, value_id);
2057 : : }
2058 : : }
2059 : 16689604 : }
2060 : :
2061 : : /* Compute the ANTIC set for BLOCK.
2062 : :
2063 : : If succs(BLOCK) > 1 then
2064 : : ANTIC_OUT[BLOCK] = intersection of ANTIC_IN[b] for all succ(BLOCK)
2065 : : else if succs(BLOCK) == 1 then
2066 : : ANTIC_OUT[BLOCK] = phi_translate (ANTIC_IN[succ(BLOCK)])
2067 : :
2068 : : ANTIC_IN[BLOCK] = clean(ANTIC_OUT[BLOCK] U EXP_GEN[BLOCK] - TMP_GEN[BLOCK])
2069 : :
2070 : : Note that clean() is deferred until after the iteration. */
2071 : :
2072 : : static bool
2073 : 15561125 : compute_antic_aux (basic_block block, bool block_has_abnormal_pred_edge)
2074 : : {
2075 : 15561125 : bitmap_set_t S, old, ANTIC_OUT;
2076 : 15561125 : edge e;
2077 : 15561125 : edge_iterator ei;
2078 : :
2079 : 15561125 : bool was_visited = BB_VISITED (block);
2080 : 15561125 : bool changed = ! BB_VISITED (block);
2081 : 15561125 : bool any_max_on_edge = false;
2082 : :
2083 : 15561125 : BB_VISITED (block) = 1;
2084 : 15561125 : old = ANTIC_OUT = S = NULL;
2085 : :
2086 : : /* If any edges from predecessors are abnormal, antic_in is empty,
2087 : : so do nothing. */
2088 : 15561125 : if (block_has_abnormal_pred_edge)
2089 : 4219 : goto maybe_dump_sets;
2090 : :
2091 : 15556906 : old = ANTIC_IN (block);
2092 : 15556906 : ANTIC_OUT = bitmap_set_new ();
2093 : :
2094 : : /* If the block has no successors, ANTIC_OUT is empty. */
2095 : 15556906 : if (EDGE_COUNT (block->succs) == 0)
2096 : : ;
2097 : : /* If we have one successor, we could have some phi nodes to
2098 : : translate through. */
2099 : 15556906 : else if (single_succ_p (block))
2100 : : {
2101 : 9879214 : e = single_succ_edge (block);
2102 : 9879214 : gcc_assert (BB_VISITED (e->dest));
2103 : 9879214 : phi_translate_set (ANTIC_OUT, ANTIC_IN (e->dest), e);
2104 : : }
2105 : : /* If we have multiple successors, we take the intersection of all of
2106 : : them. Note that in the case of loop exit phi nodes, we may have
2107 : : phis to translate through. */
2108 : : else
2109 : : {
2110 : 5677692 : size_t i;
2111 : 5677692 : edge first = NULL;
2112 : :
2113 : 5677692 : auto_vec<edge> worklist (EDGE_COUNT (block->succs));
2114 : 17150024 : FOR_EACH_EDGE (e, ei, block->succs)
2115 : : {
2116 : 11472332 : if (!first
2117 : 6137630 : && BB_VISITED (e->dest))
2118 : : first = e;
2119 : 5794640 : else if (BB_VISITED (e->dest))
2120 : 5147725 : worklist.quick_push (e);
2121 : : else
2122 : : {
2123 : : /* Unvisited successors get their ANTIC_IN replaced by the
2124 : : maximal set to arrive at a maximum ANTIC_IN solution.
2125 : : We can ignore them in the intersection operation and thus
2126 : : need not explicitely represent that maximum solution. */
2127 : 646915 : any_max_on_edge = true;
2128 : 646915 : if (dump_file && (dump_flags & TDF_DETAILS))
2129 : 18 : fprintf (dump_file, "ANTIC_IN is MAX on %d->%d\n",
2130 : 18 : e->src->index, e->dest->index);
2131 : : }
2132 : : }
2133 : :
2134 : : /* Of multiple successors we have to have visited one already
2135 : : which is guaranteed by iteration order. */
2136 : 5677692 : gcc_assert (first != NULL);
2137 : :
2138 : 5677692 : phi_translate_set (ANTIC_OUT, ANTIC_IN (first->dest), first);
2139 : :
2140 : : /* If we have multiple successors we need to intersect the ANTIC_OUT
2141 : : sets. For values that's a simple intersection but for
2142 : : expressions it is a union. Given we want to have a single
2143 : : expression per value in our sets we have to canonicalize.
2144 : : Avoid randomness and running into cycles like for PR82129 and
2145 : : canonicalize the expression we choose to the one with the
2146 : : lowest id. This requires we actually compute the union first. */
2147 : 10825417 : FOR_EACH_VEC_ELT (worklist, i, e)
2148 : : {
2149 : 5147725 : if (!gimple_seq_empty_p (phi_nodes (e->dest)))
2150 : : {
2151 : 2254 : bitmap_set_t tmp = bitmap_set_new ();
2152 : 2254 : phi_translate_set (tmp, ANTIC_IN (e->dest), e);
2153 : 2254 : bitmap_and_into (&ANTIC_OUT->values, &tmp->values);
2154 : 2254 : bitmap_ior_into (&ANTIC_OUT->expressions, &tmp->expressions);
2155 : 2254 : bitmap_set_free (tmp);
2156 : : }
2157 : : else
2158 : : {
2159 : 5145471 : bitmap_and_into (&ANTIC_OUT->values, &ANTIC_IN (e->dest)->values);
2160 : 5145471 : bitmap_ior_into (&ANTIC_OUT->expressions,
2161 : 5145471 : &ANTIC_IN (e->dest)->expressions);
2162 : : }
2163 : : }
2164 : 11355384 : if (! worklist.is_empty ())
2165 : : {
2166 : : /* Prune expressions not in the value set. */
2167 : 5035150 : bitmap_iterator bi;
2168 : 5035150 : unsigned int i;
2169 : 5035150 : unsigned int to_clear = -1U;
2170 : 35794547 : FOR_EACH_EXPR_ID_IN_SET (ANTIC_OUT, i, bi)
2171 : : {
2172 : 30759397 : if (to_clear != -1U)
2173 : : {
2174 : 16374630 : bitmap_clear_bit (&ANTIC_OUT->expressions, to_clear);
2175 : 16374630 : to_clear = -1U;
2176 : : }
2177 : 30759397 : pre_expr expr = expression_for_id (i);
2178 : 30759397 : unsigned int value_id = get_expr_value_id (expr);
2179 : 30759397 : if (!bitmap_bit_p (&ANTIC_OUT->values, value_id))
2180 : 20128305 : to_clear = i;
2181 : : }
2182 : 5035150 : if (to_clear != -1U)
2183 : 3753675 : bitmap_clear_bit (&ANTIC_OUT->expressions, to_clear);
2184 : : }
2185 : 5677692 : }
2186 : :
2187 : : /* Dump ANTIC_OUT before it's pruned. */
2188 : 15556906 : if (dump_file && (dump_flags & TDF_DETAILS))
2189 : 140 : print_bitmap_set (dump_file, ANTIC_OUT, "ANTIC_OUT", block->index);
2190 : :
2191 : : /* Prune expressions that are clobbered in block and thus become
2192 : : invalid if translated from ANTIC_OUT to ANTIC_IN. */
2193 : 15556906 : prune_clobbered_mems (ANTIC_OUT, block, any_max_on_edge);
2194 : :
2195 : : /* Generate ANTIC_OUT - TMP_GEN. */
2196 : 15556906 : S = bitmap_set_subtract_expressions (ANTIC_OUT, TMP_GEN (block));
2197 : :
2198 : : /* Start ANTIC_IN with EXP_GEN - TMP_GEN. */
2199 : 31113812 : ANTIC_IN (block) = bitmap_set_subtract_expressions (EXP_GEN (block),
2200 : 15556906 : TMP_GEN (block));
2201 : :
2202 : : /* Then union in the ANTIC_OUT - TMP_GEN values,
2203 : : to get ANTIC_OUT U EXP_GEN - TMP_GEN */
2204 : 15556906 : bitmap_ior_into (&ANTIC_IN (block)->values, &S->values);
2205 : 15556906 : bitmap_ior_into (&ANTIC_IN (block)->expressions, &S->expressions);
2206 : :
2207 : : /* clean (ANTIC_IN (block)) is defered to after the iteration converged
2208 : : because it can cause non-convergence, see for example PR81181. */
2209 : :
2210 : : /* Intersect ANTIC_IN with the old ANTIC_IN. This is required until
2211 : : we properly represent the maximum expression set, thus not prune
2212 : : values without expressions during the iteration. */
2213 : 15556906 : if (was_visited
2214 : 15556906 : && bitmap_and_into (&ANTIC_IN (block)->values, &old->values))
2215 : : {
2216 : 3703 : if (dump_file && (dump_flags & TDF_DETAILS))
2217 : 0 : fprintf (dump_file, "warning: intersecting with old ANTIC_IN "
2218 : : "shrinks the set\n");
2219 : : /* Prune expressions not in the value set. */
2220 : 3703 : bitmap_iterator bi;
2221 : 3703 : unsigned int i;
2222 : 3703 : unsigned int to_clear = -1U;
2223 : 38581 : FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (block), i, bi)
2224 : : {
2225 : 34878 : if (to_clear != -1U)
2226 : : {
2227 : 2021 : bitmap_clear_bit (&ANTIC_IN (block)->expressions, to_clear);
2228 : 2021 : to_clear = -1U;
2229 : : }
2230 : 34878 : pre_expr expr = expression_for_id (i);
2231 : 34878 : unsigned int value_id = get_expr_value_id (expr);
2232 : 34878 : if (!bitmap_bit_p (&ANTIC_IN (block)->values, value_id))
2233 : 4500 : to_clear = i;
2234 : : }
2235 : 3703 : if (to_clear != -1U)
2236 : 2479 : bitmap_clear_bit (&ANTIC_IN (block)->expressions, to_clear);
2237 : : }
2238 : :
2239 : 15556906 : if (!bitmap_set_equal (old, ANTIC_IN (block)))
2240 : 10066610 : changed = true;
2241 : :
2242 : 5490296 : maybe_dump_sets:
2243 : 15561125 : if (dump_file && (dump_flags & TDF_DETAILS))
2244 : : {
2245 : 140 : if (changed)
2246 : 122 : fprintf (dump_file, "[changed] ");
2247 : 140 : print_bitmap_set (dump_file, ANTIC_IN (block), "ANTIC_IN",
2248 : : block->index);
2249 : :
2250 : 140 : if (S)
2251 : 140 : print_bitmap_set (dump_file, S, "S", block->index);
2252 : : }
2253 : 15561125 : if (old)
2254 : 15556906 : bitmap_set_free (old);
2255 : 15561125 : if (S)
2256 : 15556906 : bitmap_set_free (S);
2257 : 15561125 : if (ANTIC_OUT)
2258 : 15556906 : bitmap_set_free (ANTIC_OUT);
2259 : 15561125 : return changed;
2260 : : }
2261 : :
2262 : : /* Compute PARTIAL_ANTIC for BLOCK.
2263 : :
2264 : : If succs(BLOCK) > 1 then
2265 : : PA_OUT[BLOCK] = value wise union of PA_IN[b] + all ANTIC_IN not
2266 : : in ANTIC_OUT for all succ(BLOCK)
2267 : : else if succs(BLOCK) == 1 then
2268 : : PA_OUT[BLOCK] = phi_translate (PA_IN[succ(BLOCK)])
2269 : :
2270 : : PA_IN[BLOCK] = clean(PA_OUT[BLOCK] - TMP_GEN[BLOCK] - ANTIC_IN[BLOCK])
2271 : :
2272 : : */
2273 : : static void
2274 : 1133898 : compute_partial_antic_aux (basic_block block,
2275 : : bool block_has_abnormal_pred_edge)
2276 : : {
2277 : 1133898 : bitmap_set_t old_PA_IN;
2278 : 1133898 : bitmap_set_t PA_OUT;
2279 : 1133898 : edge e;
2280 : 1133898 : edge_iterator ei;
2281 : 1133898 : unsigned long max_pa = param_max_partial_antic_length;
2282 : :
2283 : 1133898 : old_PA_IN = PA_OUT = NULL;
2284 : :
2285 : : /* If any edges from predecessors are abnormal, antic_in is empty,
2286 : : so do nothing. */
2287 : 1133898 : if (block_has_abnormal_pred_edge)
2288 : 745 : goto maybe_dump_sets;
2289 : :
2290 : : /* If there are too many partially anticipatable values in the
2291 : : block, phi_translate_set can take an exponential time: stop
2292 : : before the translation starts. */
2293 : 1133153 : if (max_pa
2294 : 1048412 : && single_succ_p (block)
2295 : 1841730 : && bitmap_count_bits (&PA_IN (single_succ (block))->values) > max_pa)
2296 : 455 : goto maybe_dump_sets;
2297 : :
2298 : 1132698 : old_PA_IN = PA_IN (block);
2299 : 1132698 : PA_OUT = bitmap_set_new ();
2300 : :
2301 : : /* If the block has no successors, ANTIC_OUT is empty. */
2302 : 1132698 : if (EDGE_COUNT (block->succs) == 0)
2303 : : ;
2304 : : /* If we have one successor, we could have some phi nodes to
2305 : : translate through. Note that we can't phi translate across DFS
2306 : : back edges in partial antic, because it uses a union operation on
2307 : : the successors. For recurrences like IV's, we will end up
2308 : : generating a new value in the set on each go around (i + 3 (VH.1)
2309 : : VH.1 + 1 (VH.2), VH.2 + 1 (VH.3), etc), forever. */
2310 : 1047959 : else if (single_succ_p (block))
2311 : : {
2312 : 708124 : e = single_succ_edge (block);
2313 : 708124 : if (!(e->flags & EDGE_DFS_BACK))
2314 : 634034 : phi_translate_set (PA_OUT, PA_IN (e->dest), e);
2315 : : }
2316 : : /* If we have multiple successors, we take the union of all of
2317 : : them. */
2318 : : else
2319 : : {
2320 : 339835 : size_t i;
2321 : :
2322 : 339835 : auto_vec<edge> worklist (EDGE_COUNT (block->succs));
2323 : 1024748 : FOR_EACH_EDGE (e, ei, block->succs)
2324 : : {
2325 : 684913 : if (e->flags & EDGE_DFS_BACK)
2326 : 295 : continue;
2327 : 684618 : worklist.quick_push (e);
2328 : : }
2329 : 339835 : if (worklist.length () > 0)
2330 : : {
2331 : 1024453 : FOR_EACH_VEC_ELT (worklist, i, e)
2332 : : {
2333 : 684618 : unsigned int i;
2334 : 684618 : bitmap_iterator bi;
2335 : :
2336 : 684618 : if (!gimple_seq_empty_p (phi_nodes (e->dest)))
2337 : : {
2338 : 700 : bitmap_set_t antic_in = bitmap_set_new ();
2339 : 700 : phi_translate_set (antic_in, ANTIC_IN (e->dest), e);
2340 : 1472 : FOR_EACH_EXPR_ID_IN_SET (antic_in, i, bi)
2341 : 772 : bitmap_value_insert_into_set (PA_OUT,
2342 : : expression_for_id (i));
2343 : 700 : bitmap_set_free (antic_in);
2344 : 700 : bitmap_set_t pa_in = bitmap_set_new ();
2345 : 700 : phi_translate_set (pa_in, PA_IN (e->dest), e);
2346 : 700 : FOR_EACH_EXPR_ID_IN_SET (pa_in, i, bi)
2347 : 0 : bitmap_value_insert_into_set (PA_OUT,
2348 : : expression_for_id (i));
2349 : 700 : bitmap_set_free (pa_in);
2350 : : }
2351 : : else
2352 : : {
2353 : 4393879 : FOR_EACH_EXPR_ID_IN_SET (ANTIC_IN (e->dest), i, bi)
2354 : 3709961 : bitmap_value_insert_into_set (PA_OUT,
2355 : : expression_for_id (i));
2356 : 7065751 : FOR_EACH_EXPR_ID_IN_SET (PA_IN (e->dest), i, bi)
2357 : 6381833 : bitmap_value_insert_into_set (PA_OUT,
2358 : : expression_for_id (i));
2359 : : }
2360 : : }
2361 : : }
2362 : 339835 : }
2363 : :
2364 : : /* Prune expressions that are clobbered in block and thus become
2365 : : invalid if translated from PA_OUT to PA_IN. */
2366 : 1132698 : prune_clobbered_mems (PA_OUT, block, false);
2367 : :
2368 : : /* PA_IN starts with PA_OUT - TMP_GEN.
2369 : : Then we subtract things from ANTIC_IN. */
2370 : 1132698 : PA_IN (block) = bitmap_set_subtract_expressions (PA_OUT, TMP_GEN (block));
2371 : :
2372 : : /* For partial antic, we want to put back in the phi results, since
2373 : : we will properly avoid making them partially antic over backedges. */
2374 : 1132698 : bitmap_ior_into (&PA_IN (block)->values, &PHI_GEN (block)->values);
2375 : 1132698 : bitmap_ior_into (&PA_IN (block)->expressions, &PHI_GEN (block)->expressions);
2376 : :
2377 : : /* PA_IN[block] = PA_IN[block] - ANTIC_IN[block] */
2378 : 1132698 : bitmap_set_subtract_values (PA_IN (block), ANTIC_IN (block));
2379 : :
2380 : 1132698 : clean (PA_IN (block), ANTIC_IN (block));
2381 : :
2382 : 1133898 : maybe_dump_sets:
2383 : 1133898 : if (dump_file && (dump_flags & TDF_DETAILS))
2384 : : {
2385 : 0 : if (PA_OUT)
2386 : 0 : print_bitmap_set (dump_file, PA_OUT, "PA_OUT", block->index);
2387 : :
2388 : 0 : print_bitmap_set (dump_file, PA_IN (block), "PA_IN", block->index);
2389 : : }
2390 : 1133898 : if (old_PA_IN)
2391 : 1132698 : bitmap_set_free (old_PA_IN);
2392 : 1133898 : if (PA_OUT)
2393 : 1132698 : bitmap_set_free (PA_OUT);
2394 : 1133898 : }
2395 : :
2396 : : /* Compute ANTIC and partial ANTIC sets. */
2397 : :
2398 : : static void
2399 : 953798 : compute_antic (void)
2400 : : {
2401 : 953798 : bool changed = true;
2402 : 953798 : int num_iterations = 0;
2403 : 953798 : basic_block block;
2404 : 953798 : int i;
2405 : 953798 : edge_iterator ei;
2406 : 953798 : edge e;
2407 : :
2408 : : /* If any predecessor edges are abnormal, we punt, so antic_in is empty.
2409 : : We pre-build the map of blocks with incoming abnormal edges here. */
2410 : 953798 : auto_sbitmap has_abnormal_preds (last_basic_block_for_fn (cfun));
2411 : 953798 : bitmap_clear (has_abnormal_preds);
2412 : :
2413 : 16192228 : FOR_ALL_BB_FN (block, cfun)
2414 : : {
2415 : 15238430 : BB_VISITED (block) = 0;
2416 : :
2417 : 34413487 : FOR_EACH_EDGE (e, ei, block->preds)
2418 : 19178323 : if (e->flags & EDGE_ABNORMAL)
2419 : : {
2420 : 3266 : bitmap_set_bit (has_abnormal_preds, block->index);
2421 : 3266 : break;
2422 : : }
2423 : :
2424 : : /* While we are here, give empty ANTIC_IN sets to each block. */
2425 : 15238430 : ANTIC_IN (block) = bitmap_set_new ();
2426 : 15238430 : if (do_partial_partial)
2427 : 1133898 : PA_IN (block) = bitmap_set_new ();
2428 : : }
2429 : :
2430 : : /* At the exit block we anticipate nothing. */
2431 : 953798 : BB_VISITED (EXIT_BLOCK_PTR_FOR_FN (cfun)) = 1;
2432 : :
2433 : : /* For ANTIC computation we need a postorder that also guarantees that
2434 : : a block with a single successor is visited after its successor.
2435 : : RPO on the inverted CFG has this property. */
2436 : 953798 : int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2437 : 953798 : int n = inverted_rev_post_order_compute (cfun, rpo);
2438 : :
2439 : 953798 : auto_sbitmap worklist (last_basic_block_for_fn (cfun) + 1);
2440 : 953798 : bitmap_clear (worklist);
2441 : 2768260 : FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)
2442 : 1814462 : bitmap_set_bit (worklist, e->src->index);
2443 : 2935371 : while (changed)
2444 : : {
2445 : 1981573 : if (dump_file && (dump_flags & TDF_DETAILS))
2446 : 28 : fprintf (dump_file, "Starting iteration %d\n", num_iterations);
2447 : : /* ??? We need to clear our PHI translation cache here as the
2448 : : ANTIC sets shrink and we restrict valid translations to
2449 : : those having operands with leaders in ANTIC. Same below
2450 : : for PA ANTIC computation. */
2451 : 1981573 : num_iterations++;
2452 : 1981573 : changed = false;
2453 : 37744535 : for (i = 0; i < n; ++i)
2454 : : {
2455 : 35762962 : if (bitmap_bit_p (worklist, rpo[i]))
2456 : : {
2457 : 15561125 : basic_block block = BASIC_BLOCK_FOR_FN (cfun, rpo[i]);
2458 : 15561125 : bitmap_clear_bit (worklist, block->index);
2459 : 15561125 : if (compute_antic_aux (block,
2460 : 15561125 : bitmap_bit_p (has_abnormal_preds,
2461 : : block->index)))
2462 : : {
2463 : 32908799 : FOR_EACH_EDGE (e, ei, block->preds)
2464 : 18099632 : bitmap_set_bit (worklist, e->src->index);
2465 : : changed = true;
2466 : : }
2467 : : }
2468 : : }
2469 : : /* Theoretically possible, but *highly* unlikely. */
2470 : 1981573 : gcc_checking_assert (num_iterations < 500);
2471 : : }
2472 : :
2473 : : /* We have to clean after the dataflow problem converged as cleaning
2474 : : can cause non-convergence because it is based on expressions
2475 : : rather than values. */
2476 : 14284632 : FOR_EACH_BB_FN (block, cfun)
2477 : 13330834 : clean (ANTIC_IN (block));
2478 : :
2479 : 953798 : statistics_histogram_event (cfun, "compute_antic iterations",
2480 : : num_iterations);
2481 : :
2482 : 953798 : if (do_partial_partial)
2483 : : {
2484 : : /* For partial antic we ignore backedges and thus we do not need
2485 : : to perform any iteration when we process blocks in rpo. */
2486 : 1218637 : for (i = 0; i < n; ++i)
2487 : : {
2488 : 1133898 : basic_block block = BASIC_BLOCK_FOR_FN (cfun, rpo[i]);
2489 : 1133898 : compute_partial_antic_aux (block,
2490 : 1133898 : bitmap_bit_p (has_abnormal_preds,
2491 : : block->index));
2492 : : }
2493 : : }
2494 : :
2495 : 953798 : free (rpo);
2496 : 953798 : }
2497 : :
2498 : :
2499 : : /* Inserted expressions are placed onto this worklist, which is used
2500 : : for performing quick dead code elimination of insertions we made
2501 : : that didn't turn out to be necessary. */
2502 : : static bitmap inserted_exprs;
2503 : :
2504 : : /* The actual worker for create_component_ref_by_pieces. */
2505 : :
2506 : : static tree
2507 : 1128564 : create_component_ref_by_pieces_1 (basic_block block, vn_reference_t ref,
2508 : : unsigned int *operand, gimple_seq *stmts)
2509 : : {
2510 : 1128564 : vn_reference_op_t currop = &ref->operands[*operand];
2511 : 1128564 : tree genop;
2512 : 1128564 : ++*operand;
2513 : 1128564 : switch (currop->opcode)
2514 : : {
2515 : 0 : case CALL_EXPR:
2516 : 0 : gcc_unreachable ();
2517 : :
2518 : 394883 : case MEM_REF:
2519 : 394883 : {
2520 : 394883 : tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2521 : : stmts);
2522 : 394883 : if (!baseop)
2523 : : return NULL_TREE;
2524 : 394883 : tree offset = currop->op0;
2525 : 394883 : if (TREE_CODE (baseop) == ADDR_EXPR
2526 : 394883 : && handled_component_p (TREE_OPERAND (baseop, 0)))
2527 : : {
2528 : 0 : poly_int64 off;
2529 : 0 : tree base;
2530 : 0 : base = get_addr_base_and_unit_offset (TREE_OPERAND (baseop, 0),
2531 : : &off);
2532 : 0 : gcc_assert (base);
2533 : 0 : offset = int_const_binop (PLUS_EXPR, offset,
2534 : 0 : build_int_cst (TREE_TYPE (offset),
2535 : : off));
2536 : 0 : baseop = build_fold_addr_expr (base);
2537 : : }
2538 : 394883 : genop = build2 (MEM_REF, currop->type, baseop, offset);
2539 : 394883 : MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2540 : 394883 : MR_DEPENDENCE_BASE (genop) = currop->base;
2541 : 394883 : REF_REVERSE_STORAGE_ORDER (genop) = currop->reverse;
2542 : 394883 : return genop;
2543 : : }
2544 : :
2545 : 0 : case TARGET_MEM_REF:
2546 : 0 : {
2547 : 0 : tree genop0 = NULL_TREE, genop1 = NULL_TREE;
2548 : 0 : vn_reference_op_t nextop = &ref->operands[(*operand)++];
2549 : 0 : tree baseop = create_component_ref_by_pieces_1 (block, ref, operand,
2550 : : stmts);
2551 : 0 : if (!baseop)
2552 : : return NULL_TREE;
2553 : 0 : if (currop->op0)
2554 : : {
2555 : 0 : genop0 = find_or_generate_expression (block, currop->op0, stmts);
2556 : 0 : if (!genop0)
2557 : : return NULL_TREE;
2558 : : }
2559 : 0 : if (nextop->op0)
2560 : : {
2561 : 0 : genop1 = find_or_generate_expression (block, nextop->op0, stmts);
2562 : 0 : if (!genop1)
2563 : : return NULL_TREE;
2564 : : }
2565 : 0 : genop = build5 (TARGET_MEM_REF, currop->type,
2566 : : baseop, currop->op2, genop0, currop->op1, genop1);
2567 : :
2568 : 0 : MR_DEPENDENCE_CLIQUE (genop) = currop->clique;
2569 : 0 : MR_DEPENDENCE_BASE (genop) = currop->base;
2570 : 0 : return genop;
2571 : : }
2572 : :
2573 : 254648 : case ADDR_EXPR:
2574 : 254648 : if (currop->op0)
2575 : : {
2576 : 252392 : gcc_assert (is_gimple_min_invariant (currop->op0));
2577 : 252392 : return currop->op0;
2578 : : }
2579 : : /* Fallthrough. */
2580 : 6361 : case REALPART_EXPR:
2581 : 6361 : case IMAGPART_EXPR:
2582 : 6361 : case VIEW_CONVERT_EXPR:
2583 : 6361 : {
2584 : 6361 : tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2585 : : stmts);
2586 : 6361 : if (!genop0)
2587 : : return NULL_TREE;
2588 : 6361 : return build1 (currop->opcode, currop->type, genop0);
2589 : : }
2590 : :
2591 : 4 : case WITH_SIZE_EXPR:
2592 : 4 : {
2593 : 4 : tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2594 : : stmts);
2595 : 4 : if (!genop0)
2596 : : return NULL_TREE;
2597 : 4 : tree genop1 = find_or_generate_expression (block, currop->op0, stmts);
2598 : 4 : if (!genop1)
2599 : : return NULL_TREE;
2600 : 4 : return build2 (currop->opcode, currop->type, genop0, genop1);
2601 : : }
2602 : :
2603 : 4563 : case BIT_FIELD_REF:
2604 : 4563 : {
2605 : 4563 : tree genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2606 : : stmts);
2607 : 4563 : if (!genop0)
2608 : : return NULL_TREE;
2609 : 4563 : tree op1 = currop->op0;
2610 : 4563 : tree op2 = currop->op1;
2611 : 4563 : tree t = build3 (BIT_FIELD_REF, currop->type, genop0, op1, op2);
2612 : 4563 : REF_REVERSE_STORAGE_ORDER (t) = currop->reverse;
2613 : 4563 : return t;
2614 : : }
2615 : :
2616 : : /* For array ref vn_reference_op's, operand 1 of the array ref
2617 : : is op0 of the reference op and operand 3 of the array ref is
2618 : : op1. */
2619 : 58403 : case ARRAY_RANGE_REF:
2620 : 58403 : case ARRAY_REF:
2621 : 58403 : {
2622 : 58403 : tree genop0;
2623 : 58403 : tree genop1 = currop->op0;
2624 : 58403 : tree genop2 = currop->op1;
2625 : 58403 : tree genop3 = currop->op2;
2626 : 58403 : genop0 = create_component_ref_by_pieces_1 (block, ref, operand,
2627 : : stmts);
2628 : 58403 : if (!genop0)
2629 : : return NULL_TREE;
2630 : 58403 : genop1 = find_or_generate_expression (block, genop1, stmts);
2631 : 58403 : if (!genop1)
2632 : : return NULL_TREE;
2633 : 58403 : if (genop2)
2634 : : {
2635 : 58403 : tree domain_type = TYPE_DOMAIN (TREE_TYPE (genop0));
2636 : : /* Drop zero minimum index if redundant. */
2637 : 58403 : if (integer_zerop (genop2)
2638 : 58403 : && (!domain_type
2639 : 57325 : || integer_zerop (TYPE_MIN_VALUE (domain_type))))
2640 : : genop2 = NULL_TREE;
2641 : : else
2642 : : {
2643 : 651 : genop2 = find_or_generate_expression (block, genop2, stmts);
2644 : 651 : if (!genop2)
2645 : : return NULL_TREE;
2646 : : }
2647 : : }
2648 : 58403 : if (genop3)
2649 : : {
2650 : 58403 : tree elmt_type = TREE_TYPE (TREE_TYPE (genop0));
2651 : : /* We can't always put a size in units of the element alignment
2652 : : here as the element alignment may be not visible. See
2653 : : PR43783. Simply drop the element size for constant
2654 : : sizes. */
2655 : 58403 : if ((TREE_CODE (genop3) == INTEGER_CST
2656 : 58399 : && TREE_CODE (TYPE_SIZE_UNIT (elmt_type)) == INTEGER_CST
2657 : 58399 : && wi::eq_p (wi::to_offset (TYPE_SIZE_UNIT (elmt_type)),
2658 : 58399 : (wi::to_offset (genop3) * vn_ref_op_align_unit (currop))))
2659 : 58403 : || (TREE_CODE (genop3) == EXACT_DIV_EXPR
2660 : 0 : && TREE_CODE (TREE_OPERAND (genop3, 1)) == INTEGER_CST
2661 : 0 : && operand_equal_p (TREE_OPERAND (genop3, 0), TYPE_SIZE_UNIT (elmt_type))
2662 : 0 : && wi::eq_p (wi::to_offset (TREE_OPERAND (genop3, 1)),
2663 : 58399 : vn_ref_op_align_unit (currop))))
2664 : : genop3 = NULL_TREE;
2665 : : else
2666 : : {
2667 : 4 : genop3 = find_or_generate_expression (block, genop3, stmts);
2668 : 4 : if (!genop3)
2669 : : return NULL_TREE;
2670 : : }
2671 : : }
2672 : 58403 : return build4 (currop->opcode, currop->type, genop0, genop1,
2673 : 58403 : genop2, genop3);
2674 : : }
2675 : 265542 : case COMPONENT_REF:
2676 : 265542 : {
2677 : 265542 : tree op0;
2678 : 265542 : tree op1;
2679 : 265542 : tree genop2 = currop->op1;
2680 : 265542 : op0 = create_component_ref_by_pieces_1 (block, ref, operand, stmts);
2681 : 265542 : if (!op0)
2682 : : return NULL_TREE;
2683 : : /* op1 should be a FIELD_DECL, which are represented by themselves. */
2684 : 265542 : op1 = currop->op0;
2685 : 265542 : if (genop2)
2686 : : {
2687 : 0 : genop2 = find_or_generate_expression (block, genop2, stmts);
2688 : 0 : if (!genop2)
2689 : : return NULL_TREE;
2690 : : }
2691 : 265542 : return build3 (COMPONENT_REF, TREE_TYPE (op1), op0, op1, genop2);
2692 : : }
2693 : :
2694 : 144722 : case SSA_NAME:
2695 : 144722 : {
2696 : 144722 : genop = find_or_generate_expression (block, currop->op0, stmts);
2697 : 144722 : return genop;
2698 : : }
2699 : 1694 : case STRING_CST:
2700 : 1694 : case INTEGER_CST:
2701 : 1694 : case POLY_INT_CST:
2702 : 1694 : case COMPLEX_CST:
2703 : 1694 : case VECTOR_CST:
2704 : 1694 : case REAL_CST:
2705 : 1694 : case CONSTRUCTOR:
2706 : 1694 : case VAR_DECL:
2707 : 1694 : case PARM_DECL:
2708 : 1694 : case CONST_DECL:
2709 : 1694 : case RESULT_DECL:
2710 : 1694 : case FUNCTION_DECL:
2711 : 1694 : return currop->op0;
2712 : :
2713 : 0 : default:
2714 : 0 : gcc_unreachable ();
2715 : : }
2716 : : }
2717 : :
2718 : : /* For COMPONENT_REF's and ARRAY_REF's, we can't have any intermediates for the
2719 : : COMPONENT_REF or MEM_REF or ARRAY_REF portion, because we'd end up with
2720 : : trying to rename aggregates into ssa form directly, which is a no no.
2721 : :
2722 : : Thus, this routine doesn't create temporaries, it just builds a
2723 : : single access expression for the array, calling
2724 : : find_or_generate_expression to build the innermost pieces.
2725 : :
2726 : : This function is a subroutine of create_expression_by_pieces, and
2727 : : should not be called on it's own unless you really know what you
2728 : : are doing. */
2729 : :
2730 : : static tree
2731 : 394912 : create_component_ref_by_pieces (basic_block block, vn_reference_t ref,
2732 : : gimple_seq *stmts)
2733 : : {
2734 : 394912 : unsigned int op = 0;
2735 : 394912 : return create_component_ref_by_pieces_1 (block, ref, &op, stmts);
2736 : : }
2737 : :
2738 : : /* Find a simple leader for an expression, or generate one using
2739 : : create_expression_by_pieces from a NARY expression for the value.
2740 : : BLOCK is the basic_block we are looking for leaders in.
2741 : : OP is the tree expression to find a leader for or generate.
2742 : : Returns the leader or NULL_TREE on failure. */
2743 : :
2744 : : static tree
2745 : 818837 : find_or_generate_expression (basic_block block, tree op, gimple_seq *stmts)
2746 : : {
2747 : : /* Constants are always leaders. */
2748 : 818837 : if (is_gimple_min_invariant (op))
2749 : : return op;
2750 : :
2751 : 633546 : gcc_assert (TREE_CODE (op) == SSA_NAME);
2752 : 633546 : vn_ssa_aux_t info = VN_INFO (op);
2753 : 633546 : unsigned int lookfor = info->value_id;
2754 : 633546 : if (value_id_constant_p (lookfor))
2755 : 3 : return info->valnum;
2756 : :
2757 : 633543 : pre_expr leader = bitmap_find_leader (AVAIL_OUT (block), lookfor);
2758 : 633543 : if (leader)
2759 : : {
2760 : 602218 : if (leader->kind == NAME)
2761 : 602218 : return PRE_EXPR_NAME (leader);
2762 : 0 : else if (leader->kind == CONSTANT)
2763 : 0 : return PRE_EXPR_CONSTANT (leader);
2764 : :
2765 : : /* Defer. */
2766 : : return NULL_TREE;
2767 : : }
2768 : 31325 : gcc_assert (!value_id_constant_p (lookfor));
2769 : :
2770 : : /* It must be a complex expression, so generate it recursively. Note
2771 : : that this is only necessary to handle gcc.dg/tree-ssa/ssa-pre28.c
2772 : : where the insert algorithm fails to insert a required expression. */
2773 : 31325 : bitmap exprset = value_expressions[lookfor];
2774 : 31325 : bitmap_iterator bi;
2775 : 31325 : unsigned int i;
2776 : 31325 : if (exprset)
2777 : 44675 : EXECUTE_IF_SET_IN_BITMAP (exprset, 0, i, bi)
2778 : : {
2779 : 41748 : pre_expr temp = expression_for_id (i);
2780 : : /* We cannot insert random REFERENCE expressions at arbitrary
2781 : : places. We can insert NARYs which eventually re-materializes
2782 : : its operand values. */
2783 : 41748 : if (temp->kind == NARY)
2784 : 28394 : return create_expression_by_pieces (block, temp, stmts,
2785 : 56788 : TREE_TYPE (op));
2786 : : }
2787 : :
2788 : : /* Defer. */
2789 : : return NULL_TREE;
2790 : : }
2791 : :
2792 : : /* Create an expression in pieces, so that we can handle very complex
2793 : : expressions that may be ANTIC, but not necessary GIMPLE.
2794 : : BLOCK is the basic block the expression will be inserted into,
2795 : : EXPR is the expression to insert (in value form)
2796 : : STMTS is a statement list to append the necessary insertions into.
2797 : :
2798 : : This function will die if we hit some value that shouldn't be
2799 : : ANTIC but is (IE there is no leader for it, or its components).
2800 : : The function returns NULL_TREE in case a different antic expression
2801 : : has to be inserted first.
2802 : : This function may also generate expressions that are themselves
2803 : : partially or fully redundant. Those that are will be either made
2804 : : fully redundant during the next iteration of insert (for partially
2805 : : redundant ones), or eliminated by eliminate (for fully redundant
2806 : : ones). */
2807 : :
2808 : : static tree
2809 : 2801561 : create_expression_by_pieces (basic_block block, pre_expr expr,
2810 : : gimple_seq *stmts, tree type)
2811 : : {
2812 : 2801561 : tree name;
2813 : 2801561 : tree folded;
2814 : 2801561 : gimple_seq forced_stmts = NULL;
2815 : 2801561 : unsigned int value_id;
2816 : 2801561 : gimple_stmt_iterator gsi;
2817 : 2801561 : tree exprtype = type ? type : get_expr_type (expr);
2818 : 2801561 : pre_expr nameexpr;
2819 : 2801561 : gassign *newstmt;
2820 : :
2821 : 2801561 : switch (expr->kind)
2822 : : {
2823 : : /* We may hit the NAME/CONSTANT case if we have to convert types
2824 : : that value numbering saw through. */
2825 : 716601 : case NAME:
2826 : 716601 : folded = PRE_EXPR_NAME (expr);
2827 : 716601 : if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (folded))
2828 : : return NULL_TREE;
2829 : 716587 : if (useless_type_conversion_p (exprtype, TREE_TYPE (folded)))
2830 : : return folded;
2831 : : break;
2832 : 1320276 : case CONSTANT:
2833 : 1320276 : {
2834 : 1320276 : folded = PRE_EXPR_CONSTANT (expr);
2835 : 1320276 : tree tem = fold_convert (exprtype, folded);
2836 : 1320276 : if (is_gimple_min_invariant (tem))
2837 : : return tem;
2838 : : break;
2839 : : }
2840 : 397357 : case REFERENCE:
2841 : 397357 : if (PRE_EXPR_REFERENCE (expr)->operands[0].opcode == CALL_EXPR)
2842 : : {
2843 : 2445 : vn_reference_t ref = PRE_EXPR_REFERENCE (expr);
2844 : 2445 : unsigned int operand = 1;
2845 : 2445 : vn_reference_op_t currop = &ref->operands[0];
2846 : 2445 : tree sc = NULL_TREE;
2847 : 2445 : tree fn = NULL_TREE;
2848 : 2445 : if (currop->op0)
2849 : : {
2850 : 2304 : fn = find_or_generate_expression (block, currop->op0, stmts);
2851 : 2304 : if (!fn)
2852 : 0 : return NULL_TREE;
2853 : : }
2854 : 2445 : if (currop->op1)
2855 : : {
2856 : 0 : sc = find_or_generate_expression (block, currop->op1, stmts);
2857 : 0 : if (!sc)
2858 : : return NULL_TREE;
2859 : : }
2860 : 4890 : auto_vec<tree> args (ref->operands.length () - 1);
2861 : 6341 : while (operand < ref->operands.length ())
2862 : : {
2863 : 3896 : tree arg = create_component_ref_by_pieces_1 (block, ref,
2864 : 3896 : &operand, stmts);
2865 : 3896 : if (!arg)
2866 : 0 : return NULL_TREE;
2867 : 3896 : args.quick_push (arg);
2868 : : }
2869 : 2445 : gcall *call;
2870 : 2445 : if (currop->op0)
2871 : : {
2872 : 2304 : call = gimple_build_call_vec (fn, args);
2873 : 2304 : gimple_call_set_fntype (call, currop->type);
2874 : : }
2875 : : else
2876 : 141 : call = gimple_build_call_internal_vec ((internal_fn)currop->clique,
2877 : : args);
2878 : 2445 : gimple_set_location (call, expr->loc);
2879 : 2445 : if (sc)
2880 : 0 : gimple_call_set_chain (call, sc);
2881 : 2445 : tree forcedname = make_ssa_name (ref->type);
2882 : 2445 : gimple_call_set_lhs (call, forcedname);
2883 : : /* There's no CCP pass after PRE which would re-compute alignment
2884 : : information so make sure we re-materialize this here. */
2885 : 2445 : if (gimple_call_builtin_p (call, BUILT_IN_ASSUME_ALIGNED)
2886 : 0 : && args.length () - 2 <= 1
2887 : 0 : && tree_fits_uhwi_p (args[1])
2888 : 2445 : && (args.length () != 3 || tree_fits_uhwi_p (args[2])))
2889 : : {
2890 : 0 : unsigned HOST_WIDE_INT halign = tree_to_uhwi (args[1]);
2891 : 0 : unsigned HOST_WIDE_INT hmisalign
2892 : 0 : = args.length () == 3 ? tree_to_uhwi (args[2]) : 0;
2893 : 0 : if ((halign & (halign - 1)) == 0
2894 : 0 : && (hmisalign & ~(halign - 1)) == 0
2895 : 0 : && (unsigned int)halign != 0)
2896 : 0 : set_ptr_info_alignment (get_ptr_info (forcedname),
2897 : : halign, hmisalign);
2898 : : }
2899 : 2445 : gimple_set_vuse (call, BB_LIVE_VOP_ON_EXIT (block));
2900 : 2445 : gimple_seq_add_stmt_without_update (&forced_stmts, call);
2901 : 2445 : folded = forcedname;
2902 : 2445 : }
2903 : : else
2904 : : {
2905 : 394912 : folded = create_component_ref_by_pieces (block,
2906 : : PRE_EXPR_REFERENCE (expr),
2907 : : stmts);
2908 : 394912 : if (!folded)
2909 : : return NULL_TREE;
2910 : 394912 : name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2911 : 394912 : newstmt = gimple_build_assign (name, folded);
2912 : 394912 : gimple_set_location (newstmt, expr->loc);
2913 : 394912 : gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2914 : 394912 : gimple_set_vuse (newstmt, BB_LIVE_VOP_ON_EXIT (block));
2915 : 394912 : folded = name;
2916 : : }
2917 : : break;
2918 : 367327 : case NARY:
2919 : 367327 : {
2920 : 367327 : vn_nary_op_t nary = PRE_EXPR_NARY (expr);
2921 : 367327 : tree *genop = XALLOCAVEC (tree, nary->length);
2922 : 367327 : unsigned i;
2923 : 970377 : for (i = 0; i < nary->length; ++i)
2924 : : {
2925 : 612749 : genop[i] = find_or_generate_expression (block, nary->op[i], stmts);
2926 : 612749 : if (!genop[i])
2927 : : return NULL_TREE;
2928 : : /* Ensure genop[] is properly typed for POINTER_PLUS_EXPR. It
2929 : : may have conversions stripped. */
2930 : 603050 : if (nary->opcode == POINTER_PLUS_EXPR)
2931 : : {
2932 : 100682 : if (i == 0)
2933 : 50358 : genop[i] = gimple_convert (&forced_stmts,
2934 : : nary->type, genop[i]);
2935 : 50324 : else if (i == 1)
2936 : 50324 : genop[i] = gimple_convert (&forced_stmts,
2937 : : sizetype, genop[i]);
2938 : : }
2939 : : else
2940 : 502368 : genop[i] = gimple_convert (&forced_stmts,
2941 : 502368 : TREE_TYPE (nary->op[i]), genop[i]);
2942 : : }
2943 : 357628 : if (nary->opcode == CONSTRUCTOR)
2944 : : {
2945 : 4 : vec<constructor_elt, va_gc> *elts = NULL;
2946 : 20 : for (i = 0; i < nary->length; ++i)
2947 : 16 : CONSTRUCTOR_APPEND_ELT (elts, NULL_TREE, genop[i]);
2948 : 4 : folded = build_constructor (nary->type, elts);
2949 : 4 : name = make_temp_ssa_name (exprtype, NULL, "pretmp");
2950 : 4 : newstmt = gimple_build_assign (name, folded);
2951 : 4 : gimple_set_location (newstmt, expr->loc);
2952 : 4 : gimple_seq_add_stmt_without_update (&forced_stmts, newstmt);
2953 : 4 : folded = name;
2954 : : }
2955 : : else
2956 : : {
2957 : 357624 : switch (nary->length)
2958 : : {
2959 : 113414 : case 1:
2960 : 113414 : folded = gimple_build (&forced_stmts, expr->loc,
2961 : : nary->opcode, nary->type, genop[0]);
2962 : 113414 : break;
2963 : 244067 : case 2:
2964 : 244067 : folded = gimple_build (&forced_stmts, expr->loc, nary->opcode,
2965 : : nary->type, genop[0], genop[1]);
2966 : 244067 : break;
2967 : 143 : case 3:
2968 : 143 : folded = gimple_build (&forced_stmts, expr->loc, nary->opcode,
2969 : : nary->type, genop[0], genop[1],
2970 : : genop[2]);
2971 : 143 : break;
2972 : 0 : default:
2973 : 0 : gcc_unreachable ();
2974 : : }
2975 : : }
2976 : : }
2977 : : break;
2978 : 0 : default:
2979 : 0 : gcc_unreachable ();
2980 : : }
2981 : :
2982 : 844447 : folded = gimple_convert (&forced_stmts, exprtype, folded);
2983 : :
2984 : : /* If there is nothing to insert, return the simplified result. */
2985 : 844447 : if (gimple_seq_empty_p (forced_stmts))
2986 : : return folded;
2987 : : /* If we simplified to a constant return it and discard eventually
2988 : : built stmts. */
2989 : 754947 : if (is_gimple_min_invariant (folded))
2990 : : {
2991 : 0 : gimple_seq_discard (forced_stmts);
2992 : 0 : return folded;
2993 : : }
2994 : : /* Likewise if we simplified to sth not queued for insertion. */
2995 : 754947 : bool found = false;
2996 : 754947 : gsi = gsi_last (forced_stmts);
2997 : 754947 : for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2998 : : {
2999 : 754947 : gimple *stmt = gsi_stmt (gsi);
3000 : 754947 : tree forcedname = gimple_get_lhs (stmt);
3001 : 754947 : if (forcedname == folded)
3002 : : {
3003 : : found = true;
3004 : : break;
3005 : : }
3006 : : }
3007 : 754947 : if (! found)
3008 : : {
3009 : 0 : gimple_seq_discard (forced_stmts);
3010 : 0 : return folded;
3011 : : }
3012 : 754947 : gcc_assert (TREE_CODE (folded) == SSA_NAME);
3013 : :
3014 : : /* If we have any intermediate expressions to the value sets, add them
3015 : : to the value sets and chain them in the instruction stream. */
3016 : 754947 : if (forced_stmts)
3017 : : {
3018 : 754947 : gsi = gsi_start (forced_stmts);
3019 : 1510371 : for (; !gsi_end_p (gsi); gsi_next (&gsi))
3020 : : {
3021 : 755424 : gimple *stmt = gsi_stmt (gsi);
3022 : 755424 : tree forcedname = gimple_get_lhs (stmt);
3023 : 755424 : pre_expr nameexpr;
3024 : :
3025 : 755424 : if (forcedname != folded)
3026 : : {
3027 : 477 : vn_ssa_aux_t vn_info = VN_INFO (forcedname);
3028 : 477 : vn_info->valnum = forcedname;
3029 : 477 : vn_info->value_id = get_next_value_id ();
3030 : 477 : nameexpr = get_or_alloc_expr_for_name (forcedname);
3031 : 477 : add_to_value (vn_info->value_id, nameexpr);
3032 : 477 : if (NEW_SETS (block))
3033 : 477 : bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3034 : 477 : bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3035 : : }
3036 : :
3037 : 755424 : bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (forcedname));
3038 : : }
3039 : 754947 : gimple_seq_add_seq (stmts, forced_stmts);
3040 : : }
3041 : :
3042 : 754947 : name = folded;
3043 : :
3044 : : /* Fold the last statement. */
3045 : 754947 : gsi = gsi_last (*stmts);
3046 : 754947 : if (fold_stmt_inplace (&gsi))
3047 : 217569 : update_stmt (gsi_stmt (gsi));
3048 : :
3049 : : /* Add a value number to the temporary.
3050 : : The value may already exist in either NEW_SETS, or AVAIL_OUT, because
3051 : : we are creating the expression by pieces, and this particular piece of
3052 : : the expression may have been represented. There is no harm in replacing
3053 : : here. */
3054 : 754947 : value_id = get_expr_value_id (expr);
3055 : 754947 : vn_ssa_aux_t vn_info = VN_INFO (name);
3056 : 754947 : vn_info->value_id = value_id;
3057 : 754947 : vn_info->valnum = vn_valnum_from_value_id (value_id);
3058 : 754947 : if (vn_info->valnum == NULL_TREE)
3059 : 233447 : vn_info->valnum = name;
3060 : 754947 : gcc_assert (vn_info->valnum != NULL_TREE);
3061 : 754947 : nameexpr = get_or_alloc_expr_for_name (name);
3062 : 754947 : add_to_value (value_id, nameexpr);
3063 : 754947 : if (NEW_SETS (block))
3064 : 531753 : bitmap_value_replace_in_set (NEW_SETS (block), nameexpr);
3065 : 754947 : bitmap_value_replace_in_set (AVAIL_OUT (block), nameexpr);
3066 : :
3067 : 754947 : pre_stats.insertions++;
3068 : 754947 : if (dump_file && (dump_flags & TDF_DETAILS))
3069 : : {
3070 : 22 : fprintf (dump_file, "Inserted ");
3071 : 44 : print_gimple_stmt (dump_file, gsi_stmt (gsi_last (*stmts)), 0);
3072 : 22 : fprintf (dump_file, " in predecessor %d (%04d)\n",
3073 : : block->index, value_id);
3074 : : }
3075 : :
3076 : : return name;
3077 : : }
3078 : :
3079 : :
3080 : : /* Insert the to-be-made-available values of expression EXPRNUM for each
3081 : : predecessor, stored in AVAIL, into the predecessors of BLOCK, and
3082 : : merge the result with a phi node, given the same value number as
3083 : : NODE. Return true if we have inserted new stuff. */
3084 : :
3085 : : static bool
3086 : 1862544 : insert_into_preds_of_block (basic_block block, unsigned int exprnum,
3087 : : vec<pre_expr> &avail)
3088 : : {
3089 : 1862544 : pre_expr expr = expression_for_id (exprnum);
3090 : 1862544 : pre_expr newphi;
3091 : 1862544 : unsigned int val = get_expr_value_id (expr);
3092 : 1862544 : edge pred;
3093 : 1862544 : bool insertions = false;
3094 : 1862544 : bool nophi = false;
3095 : 1862544 : basic_block bprime;
3096 : 1862544 : pre_expr eprime;
3097 : 1862544 : edge_iterator ei;
3098 : 1862544 : tree type = get_expr_type (expr);
3099 : 1862544 : tree temp;
3100 : 1862544 : gphi *phi;
3101 : :
3102 : : /* Make sure we aren't creating an induction variable. */
3103 : 1862544 : if (bb_loop_depth (block) > 0 && EDGE_COUNT (block->preds) == 2)
3104 : : {
3105 : 1525697 : bool firstinsideloop = false;
3106 : 1525697 : bool secondinsideloop = false;
3107 : 4577091 : firstinsideloop = flow_bb_inside_loop_p (block->loop_father,
3108 : 1525697 : EDGE_PRED (block, 0)->src);
3109 : 4577091 : secondinsideloop = flow_bb_inside_loop_p (block->loop_father,
3110 : 1525697 : EDGE_PRED (block, 1)->src);
3111 : : /* Induction variables only have one edge inside the loop. */
3112 : 1525697 : if ((firstinsideloop ^ secondinsideloop)
3113 : 1456723 : && expr->kind != REFERENCE)
3114 : : {
3115 : 1382016 : if (dump_file && (dump_flags & TDF_DETAILS))
3116 : 56 : fprintf (dump_file, "Skipping insertion of phi for partial "
3117 : : "redundancy: Looks like an induction variable\n");
3118 : : nophi = true;
3119 : : }
3120 : : }
3121 : :
3122 : : /* Make the necessary insertions. */
3123 : 5805657 : FOR_EACH_EDGE (pred, ei, block->preds)
3124 : : {
3125 : : /* When we are not inserting a PHI node do not bother inserting
3126 : : into places that do not dominate the anticipated computations. */
3127 : 3943113 : if (nophi && !dominated_by_p (CDI_DOMINATORS, block, pred->src))
3128 : 1396085 : continue;
3129 : 2549973 : gimple_seq stmts = NULL;
3130 : 2549973 : tree builtexpr;
3131 : 2549973 : bprime = pred->src;
3132 : 2549973 : eprime = avail[pred->dest_idx];
3133 : 2549973 : builtexpr = create_expression_by_pieces (bprime, eprime,
3134 : : &stmts, type);
3135 : 2549973 : gcc_assert (!(pred->flags & EDGE_ABNORMAL));
3136 : 2549973 : if (!gimple_seq_empty_p (stmts))
3137 : : {
3138 : 510137 : basic_block new_bb = gsi_insert_seq_on_edge_immediate (pred, stmts);
3139 : 510137 : gcc_assert (! new_bb);
3140 : : insertions = true;
3141 : : }
3142 : 2549973 : if (!builtexpr)
3143 : : {
3144 : : /* We cannot insert a PHI node if we failed to insert
3145 : : on one edge. */
3146 : 2945 : nophi = true;
3147 : 2945 : continue;
3148 : : }
3149 : 2547028 : if (is_gimple_min_invariant (builtexpr))
3150 : 1320302 : avail[pred->dest_idx] = get_or_alloc_expr_for_constant (builtexpr);
3151 : : else
3152 : 1226726 : avail[pred->dest_idx] = get_or_alloc_expr_for_name (builtexpr);
3153 : : }
3154 : : /* If we didn't want a phi node, and we made insertions, we still have
3155 : : inserted new stuff, and thus return true. If we didn't want a phi node,
3156 : : and didn't make insertions, we haven't added anything new, so return
3157 : : false. */
3158 : 1862544 : if (nophi && insertions)
3159 : : return true;
3160 : 1852759 : else if (nophi && !insertions)
3161 : : return false;
3162 : :
3163 : : /* Now build a phi for the new variable. */
3164 : 477588 : temp = make_temp_ssa_name (type, NULL, "prephitmp");
3165 : 477588 : phi = create_phi_node (temp, block);
3166 : :
3167 : 477588 : vn_ssa_aux_t vn_info = VN_INFO (temp);
3168 : 477588 : vn_info->value_id = val;
3169 : 477588 : vn_info->valnum = vn_valnum_from_value_id (val);
3170 : 477588 : if (vn_info->valnum == NULL_TREE)
3171 : 98640 : vn_info->valnum = temp;
3172 : 477588 : bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3173 : 1641832 : FOR_EACH_EDGE (pred, ei, block->preds)
3174 : : {
3175 : 1164244 : pre_expr ae = avail[pred->dest_idx];
3176 : 1164244 : gcc_assert (get_expr_type (ae) == type
3177 : : || useless_type_conversion_p (type, get_expr_type (ae)));
3178 : 1164244 : if (ae->kind == CONSTANT)
3179 : 181980 : add_phi_arg (phi, unshare_expr (PRE_EXPR_CONSTANT (ae)),
3180 : : pred, UNKNOWN_LOCATION);
3181 : : else
3182 : 982264 : add_phi_arg (phi, PRE_EXPR_NAME (ae), pred, UNKNOWN_LOCATION);
3183 : : }
3184 : :
3185 : 477588 : newphi = get_or_alloc_expr_for_name (temp);
3186 : 477588 : add_to_value (val, newphi);
3187 : :
3188 : : /* The value should *not* exist in PHI_GEN, or else we wouldn't be doing
3189 : : this insertion, since we test for the existence of this value in PHI_GEN
3190 : : before proceeding with the partial redundancy checks in insert_aux.
3191 : :
3192 : : The value may exist in AVAIL_OUT, in particular, it could be represented
3193 : : by the expression we are trying to eliminate, in which case we want the
3194 : : replacement to occur. If it's not existing in AVAIL_OUT, we want it
3195 : : inserted there.
3196 : :
3197 : : Similarly, to the PHI_GEN case, the value should not exist in NEW_SETS of
3198 : : this block, because if it did, it would have existed in our dominator's
3199 : : AVAIL_OUT, and would have been skipped due to the full redundancy check.
3200 : : */
3201 : :
3202 : 477588 : bitmap_insert_into_set (PHI_GEN (block), newphi);
3203 : 477588 : bitmap_value_replace_in_set (AVAIL_OUT (block),
3204 : : newphi);
3205 : 477588 : if (NEW_SETS (block))
3206 : 477588 : bitmap_insert_into_set (NEW_SETS (block), newphi);
3207 : :
3208 : : /* If we insert a PHI node for a conversion of another PHI node
3209 : : in the same basic-block try to preserve range information.
3210 : : This is important so that followup loop passes receive optimal
3211 : : number of iteration analysis results. See PR61743. */
3212 : 477588 : if (expr->kind == NARY
3213 : 178096 : && CONVERT_EXPR_CODE_P (expr->u.nary->opcode)
3214 : 57490 : && TREE_CODE (expr->u.nary->op[0]) == SSA_NAME
3215 : 57277 : && gimple_bb (SSA_NAME_DEF_STMT (expr->u.nary->op[0])) == block
3216 : 45820 : && INTEGRAL_TYPE_P (type)
3217 : 44873 : && INTEGRAL_TYPE_P (TREE_TYPE (expr->u.nary->op[0]))
3218 : 43867 : && (TYPE_PRECISION (type)
3219 : 43867 : >= TYPE_PRECISION (TREE_TYPE (expr->u.nary->op[0])))
3220 : 512674 : && SSA_NAME_RANGE_INFO (expr->u.nary->op[0]))
3221 : : {
3222 : 21534 : int_range_max r;
3223 : 43068 : if (get_range_query (cfun)->range_of_expr (r, expr->u.nary->op[0])
3224 : 21534 : && !r.undefined_p ()
3225 : 21534 : && !r.varying_p ()
3226 : 43068 : && !wi::neg_p (r.lower_bound (), SIGNED)
3227 : 59023 : && !wi::neg_p (r.upper_bound (), SIGNED))
3228 : : {
3229 : : /* Just handle extension and sign-changes of all-positive ranges. */
3230 : 15289 : range_cast (r, type);
3231 : 15289 : set_range_info (temp, r);
3232 : : }
3233 : 21534 : }
3234 : :
3235 : 477588 : if (dump_file && (dump_flags & TDF_DETAILS))
3236 : : {
3237 : 10 : fprintf (dump_file, "Created phi ");
3238 : 10 : print_gimple_stmt (dump_file, phi, 0);
3239 : 10 : fprintf (dump_file, " in block %d (%04d)\n", block->index, val);
3240 : : }
3241 : 477588 : pre_stats.phis++;
3242 : 477588 : return true;
3243 : : }
3244 : :
3245 : :
3246 : :
3247 : : /* Perform insertion of partially redundant or hoistable values.
3248 : : For BLOCK, do the following:
3249 : : 1. Propagate the NEW_SETS of the dominator into the current block.
3250 : : If the block has multiple predecessors,
3251 : : 2a. Iterate over the ANTIC expressions for the block to see if
3252 : : any of them are partially redundant.
3253 : : 2b. If so, insert them into the necessary predecessors to make
3254 : : the expression fully redundant.
3255 : : 2c. Insert a new PHI merging the values of the predecessors.
3256 : : 2d. Insert the new PHI, and the new expressions, into the
3257 : : NEW_SETS set.
3258 : : If the block has multiple successors,
3259 : : 3a. Iterate over the ANTIC values for the block to see if
3260 : : any of them are good candidates for hoisting.
3261 : : 3b. If so, insert expressions computing the values in BLOCK,
3262 : : and add the new expressions into the NEW_SETS set.
3263 : : 4. Recursively call ourselves on the dominator children of BLOCK.
3264 : :
3265 : : Steps 1, 2a, and 4 are done by insert_aux. 2b, 2c and 2d are done by
3266 : : do_pre_regular_insertion and do_partial_insertion. 3a and 3b are
3267 : : done in do_hoist_insertion.
3268 : : */
3269 : :
3270 : : static bool
3271 : 3704479 : do_pre_regular_insertion (basic_block block, basic_block dom,
3272 : : vec<pre_expr> exprs)
3273 : : {
3274 : 3704479 : bool new_stuff = false;
3275 : 3704479 : pre_expr expr;
3276 : 3704479 : auto_vec<pre_expr, 2> avail;
3277 : 3704479 : int i;
3278 : :
3279 : 3704479 : avail.safe_grow (EDGE_COUNT (block->preds), true);
3280 : :
3281 : 25402274 : FOR_EACH_VEC_ELT (exprs, i, expr)
3282 : : {
3283 : 21697795 : if (expr->kind == NARY
3284 : 21697795 : || expr->kind == REFERENCE)
3285 : : {
3286 : 12334979 : unsigned int val;
3287 : 12334979 : bool by_some = false;
3288 : 12334979 : bool cant_insert = false;
3289 : 12334979 : bool all_same = true;
3290 : 12334979 : unsigned num_inserts = 0;
3291 : 12334979 : unsigned num_const = 0;
3292 : 12334979 : pre_expr first_s = NULL;
3293 : 12334979 : edge pred;
3294 : 12334979 : basic_block bprime;
3295 : 12334979 : pre_expr eprime = NULL;
3296 : 12334979 : edge_iterator ei;
3297 : 12334979 : pre_expr edoubleprime = NULL;
3298 : 12334979 : bool do_insertion = false;
3299 : :
3300 : 12334979 : val = get_expr_value_id (expr);
3301 : 24669958 : if (bitmap_set_contains_value (PHI_GEN (block), val))
3302 : 1067286 : continue;
3303 : 11559572 : if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3304 : : {
3305 : 291879 : if (dump_file && (dump_flags & TDF_DETAILS))
3306 : : {
3307 : 7 : fprintf (dump_file, "Found fully redundant value: ");
3308 : 7 : print_pre_expr (dump_file, expr);
3309 : 7 : fprintf (dump_file, "\n");
3310 : : }
3311 : 291879 : continue;
3312 : : }
3313 : :
3314 : 36643715 : FOR_EACH_EDGE (pred, ei, block->preds)
3315 : : {
3316 : 25376834 : unsigned int vprime;
3317 : :
3318 : : /* We should never run insertion for the exit block
3319 : : and so not come across fake pred edges. */
3320 : 25376834 : gcc_assert (!(pred->flags & EDGE_FAKE));
3321 : 25376834 : bprime = pred->src;
3322 : : /* We are looking at ANTIC_OUT of bprime. */
3323 : 25376834 : eprime = phi_translate (NULL, expr, ANTIC_IN (block), NULL, pred);
3324 : :
3325 : : /* eprime will generally only be NULL if the
3326 : : value of the expression, translated
3327 : : through the PHI for this predecessor, is
3328 : : undefined. If that is the case, we can't
3329 : : make the expression fully redundant,
3330 : : because its value is undefined along a
3331 : : predecessor path. We can thus break out
3332 : : early because it doesn't matter what the
3333 : : rest of the results are. */
3334 : 25376834 : if (eprime == NULL)
3335 : : {
3336 : 812 : avail[pred->dest_idx] = NULL;
3337 : 812 : cant_insert = true;
3338 : 812 : break;
3339 : : }
3340 : :
3341 : 25376022 : vprime = get_expr_value_id (eprime);
3342 : 25376022 : edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime),
3343 : : vprime);
3344 : 25376022 : if (edoubleprime == NULL)
3345 : : {
3346 : 22860037 : avail[pred->dest_idx] = eprime;
3347 : 22860037 : all_same = false;
3348 : 22860037 : num_inserts++;
3349 : : }
3350 : : else
3351 : : {
3352 : 2515985 : avail[pred->dest_idx] = edoubleprime;
3353 : 2515985 : by_some = true;
3354 : 2515985 : if (edoubleprime->kind == CONSTANT)
3355 : 1659277 : num_const++;
3356 : : /* We want to perform insertions to remove a redundancy on
3357 : : a path in the CFG we want to optimize for speed. */
3358 : 2515985 : if (optimize_edge_for_speed_p (pred))
3359 : 2094877 : do_insertion = true;
3360 : 2515985 : if (first_s == NULL)
3361 : : first_s = edoubleprime;
3362 : 280923 : else if (!pre_expr_d::equal (first_s, edoubleprime))
3363 : 217444 : all_same = false;
3364 : : }
3365 : : }
3366 : : /* If we can insert it, it's not the same value
3367 : : already existing along every predecessor, and
3368 : : it's defined by some predecessor, it is
3369 : : partially redundant. */
3370 : 11267693 : if (!cant_insert && !all_same && by_some)
3371 : : {
3372 : : /* If the expression is redundant on all edges and we need
3373 : : to at most insert one copy from a constant do the PHI
3374 : : insertion even when not optimizing a path that's to be
3375 : : optimized for speed. */
3376 : 2233005 : if (num_inserts == 0 && num_const <= 1)
3377 : : do_insertion = true;
3378 : 2093846 : if (!do_insertion)
3379 : : {
3380 : 376461 : if (dump_file && (dump_flags & TDF_DETAILS))
3381 : : {
3382 : 0 : fprintf (dump_file, "Skipping partial redundancy for "
3383 : : "expression ");
3384 : 0 : print_pre_expr (dump_file, expr);
3385 : 0 : fprintf (dump_file, " (%04d), no redundancy on to be "
3386 : : "optimized for speed edge\n", val);
3387 : : }
3388 : : }
3389 : 1856544 : else if (dbg_cnt (treepre_insert))
3390 : : {
3391 : 1856544 : if (dump_file && (dump_flags & TDF_DETAILS))
3392 : : {
3393 : 66 : fprintf (dump_file, "Found partial redundancy for "
3394 : : "expression ");
3395 : 66 : print_pre_expr (dump_file, expr);
3396 : 66 : fprintf (dump_file, " (%04d)\n",
3397 : : get_expr_value_id (expr));
3398 : : }
3399 : 1856544 : if (insert_into_preds_of_block (block,
3400 : : get_expression_id (expr),
3401 : : avail))
3402 : 11267693 : new_stuff = true;
3403 : : }
3404 : : }
3405 : : /* If all edges produce the same value and that value is
3406 : : an invariant, then the PHI has the same value on all
3407 : : edges. Note this. */
3408 : 9034688 : else if (!cant_insert
3409 : 9034688 : && all_same
3410 : 9034688 : && (edoubleprime->kind != NAME
3411 : 1180 : || !SSA_NAME_OCCURS_IN_ABNORMAL_PHI
3412 : : (PRE_EXPR_NAME (edoubleprime))))
3413 : : {
3414 : 2027 : gcc_assert (edoubleprime->kind == CONSTANT
3415 : : || edoubleprime->kind == NAME);
3416 : :
3417 : 2027 : tree temp = make_temp_ssa_name (get_expr_type (expr),
3418 : : NULL, "pretmp");
3419 : 2027 : gassign *assign
3420 : 2027 : = gimple_build_assign (temp,
3421 : 2027 : edoubleprime->kind == CONSTANT ?
3422 : : PRE_EXPR_CONSTANT (edoubleprime) :
3423 : : PRE_EXPR_NAME (edoubleprime));
3424 : 2027 : gimple_stmt_iterator gsi = gsi_after_labels (block);
3425 : 2027 : gsi_insert_before (&gsi, assign, GSI_NEW_STMT);
3426 : :
3427 : 2027 : vn_ssa_aux_t vn_info = VN_INFO (temp);
3428 : 2027 : vn_info->value_id = val;
3429 : 2027 : vn_info->valnum = vn_valnum_from_value_id (val);
3430 : 2027 : if (vn_info->valnum == NULL_TREE)
3431 : 299 : vn_info->valnum = temp;
3432 : 2027 : bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (temp));
3433 : 2027 : pre_expr newe = get_or_alloc_expr_for_name (temp);
3434 : 2027 : add_to_value (val, newe);
3435 : 2027 : bitmap_value_replace_in_set (AVAIL_OUT (block), newe);
3436 : 2027 : bitmap_insert_into_set (NEW_SETS (block), newe);
3437 : 2027 : bitmap_insert_into_set (PHI_GEN (block), newe);
3438 : : }
3439 : : }
3440 : : }
3441 : :
3442 : 3704479 : return new_stuff;
3443 : 3704479 : }
3444 : :
3445 : :
3446 : : /* Perform insertion for partially anticipatable expressions. There
3447 : : is only one case we will perform insertion for these. This case is
3448 : : if the expression is partially anticipatable, and fully available.
3449 : : In this case, we know that putting it earlier will enable us to
3450 : : remove the later computation. */
3451 : :
3452 : : static bool
3453 : 298608 : do_pre_partial_partial_insertion (basic_block block, basic_block dom,
3454 : : vec<pre_expr> exprs)
3455 : : {
3456 : 298608 : bool new_stuff = false;
3457 : 298608 : pre_expr expr;
3458 : 298608 : auto_vec<pre_expr, 2> avail;
3459 : 298608 : int i;
3460 : :
3461 : 298608 : avail.safe_grow (EDGE_COUNT (block->preds), true);
3462 : :
3463 : 2799037 : FOR_EACH_VEC_ELT (exprs, i, expr)
3464 : : {
3465 : 2500429 : if (expr->kind == NARY
3466 : 2500429 : || expr->kind == REFERENCE)
3467 : : {
3468 : 1870836 : unsigned int val;
3469 : 1870836 : bool by_all = true;
3470 : 1870836 : bool cant_insert = false;
3471 : 1870836 : edge pred;
3472 : 1870836 : basic_block bprime;
3473 : 1870836 : pre_expr eprime = NULL;
3474 : 1870836 : edge_iterator ei;
3475 : :
3476 : 1870836 : val = get_expr_value_id (expr);
3477 : 3741672 : if (bitmap_set_contains_value (PHI_GEN (block), val))
3478 : 56993 : continue;
3479 : 1861742 : if (bitmap_set_contains_value (AVAIL_OUT (dom), val))
3480 : 47899 : continue;
3481 : :
3482 : 1893171 : FOR_EACH_EDGE (pred, ei, block->preds)
3483 : : {
3484 : 1883266 : unsigned int vprime;
3485 : 1883266 : pre_expr edoubleprime;
3486 : :
3487 : : /* We should never run insertion for the exit block
3488 : : and so not come across fake pred edges. */
3489 : 1883266 : gcc_assert (!(pred->flags & EDGE_FAKE));
3490 : 1883266 : bprime = pred->src;
3491 : 3766532 : eprime = phi_translate (NULL, expr, ANTIC_IN (block),
3492 : 1883266 : PA_IN (block), pred);
3493 : :
3494 : : /* eprime will generally only be NULL if the
3495 : : value of the expression, translated
3496 : : through the PHI for this predecessor, is
3497 : : undefined. If that is the case, we can't
3498 : : make the expression fully redundant,
3499 : : because its value is undefined along a
3500 : : predecessor path. We can thus break out
3501 : : early because it doesn't matter what the
3502 : : rest of the results are. */
3503 : 1883266 : if (eprime == NULL)
3504 : : {
3505 : 36 : avail[pred->dest_idx] = NULL;
3506 : 36 : cant_insert = true;
3507 : 36 : break;
3508 : : }
3509 : :
3510 : 1883230 : vprime = get_expr_value_id (eprime);
3511 : 1883230 : edoubleprime = bitmap_find_leader (AVAIL_OUT (bprime), vprime);
3512 : 1883230 : avail[pred->dest_idx] = edoubleprime;
3513 : 1883230 : if (edoubleprime == NULL)
3514 : : {
3515 : : by_all = false;
3516 : : break;
3517 : : }
3518 : : }
3519 : :
3520 : : /* If we can insert it, it's not the same value
3521 : : already existing along every predecessor, and
3522 : : it's defined by some predecessor, it is
3523 : : partially redundant. */
3524 : 1813843 : if (!cant_insert && by_all)
3525 : : {
3526 : 9905 : edge succ;
3527 : 9905 : bool do_insertion = false;
3528 : :
3529 : : /* Insert only if we can remove a later expression on a path
3530 : : that we want to optimize for speed.
3531 : : The phi node that we will be inserting in BLOCK is not free,
3532 : : and inserting it for the sake of !optimize_for_speed successor
3533 : : may cause regressions on the speed path. */
3534 : 26888 : FOR_EACH_EDGE (succ, ei, block->succs)
3535 : : {
3536 : 16983 : if (bitmap_set_contains_value (PA_IN (succ->dest), val)
3537 : 16983 : || bitmap_set_contains_value (ANTIC_IN (succ->dest), val))
3538 : : {
3539 : 8752 : if (optimize_edge_for_speed_p (succ))
3540 : 16983 : do_insertion = true;
3541 : : }
3542 : : }
3543 : :
3544 : 9905 : if (!do_insertion)
3545 : : {
3546 : 3905 : if (dump_file && (dump_flags & TDF_DETAILS))
3547 : : {
3548 : 0 : fprintf (dump_file, "Skipping partial partial redundancy "
3549 : : "for expression ");
3550 : 0 : print_pre_expr (dump_file, expr);
3551 : 0 : fprintf (dump_file, " (%04d), not (partially) anticipated "
3552 : : "on any to be optimized for speed edges\n", val);
3553 : : }
3554 : : }
3555 : 6000 : else if (dbg_cnt (treepre_insert))
3556 : : {
3557 : 6000 : pre_stats.pa_insert++;
3558 : 6000 : if (dump_file && (dump_flags & TDF_DETAILS))
3559 : : {
3560 : 0 : fprintf (dump_file, "Found partial partial redundancy "
3561 : : "for expression ");
3562 : 0 : print_pre_expr (dump_file, expr);
3563 : 0 : fprintf (dump_file, " (%04d)\n",
3564 : : get_expr_value_id (expr));
3565 : : }
3566 : 6000 : if (insert_into_preds_of_block (block,
3567 : : get_expression_id (expr),
3568 : : avail))
3569 : 9905 : new_stuff = true;
3570 : : }
3571 : : }
3572 : : }
3573 : : }
3574 : :
3575 : 298608 : return new_stuff;
3576 : 298608 : }
3577 : :
3578 : : /* Insert expressions in BLOCK to compute hoistable values up.
3579 : : Return TRUE if something was inserted, otherwise return FALSE.
3580 : : The caller has to make sure that BLOCK has at least two successors. */
3581 : :
3582 : : static bool
3583 : 4788431 : do_hoist_insertion (basic_block block)
3584 : : {
3585 : 4788431 : edge e;
3586 : 4788431 : edge_iterator ei;
3587 : 4788431 : bool new_stuff = false;
3588 : 4788431 : unsigned i;
3589 : 4788431 : gimple_stmt_iterator last;
3590 : :
3591 : : /* At least two successors, or else... */
3592 : 4788431 : gcc_assert (EDGE_COUNT (block->succs) >= 2);
3593 : :
3594 : : /* Check that all successors of BLOCK are dominated by block.
3595 : : We could use dominated_by_p() for this, but actually there is a much
3596 : : quicker check: any successor that is dominated by BLOCK can't have
3597 : : more than one predecessor edge. */
3598 : 14464209 : FOR_EACH_EDGE (e, ei, block->succs)
3599 : 14319231 : if (! single_pred_p (e->dest))
3600 : : return false;
3601 : :
3602 : : /* Determine the insertion point. If we cannot safely insert before
3603 : : the last stmt if we'd have to, bail out. */
3604 : 4781281 : last = gsi_last_bb (block);
3605 : 4781281 : if (!gsi_end_p (last)
3606 : 4780853 : && !is_ctrl_stmt (gsi_stmt (last))
3607 : 5435332 : && stmt_ends_bb_p (gsi_stmt (last)))
3608 : : return false;
3609 : :
3610 : : /* We have multiple successors, compute ANTIC_OUT by taking the intersection
3611 : : of all of ANTIC_IN translating through PHI nodes. Track the union
3612 : : of the expression sets so we can pick a representative that is
3613 : : fully generatable out of hoistable expressions. */
3614 : 4127817 : bitmap_set_t ANTIC_OUT = bitmap_set_new ();
3615 : 4127817 : bool first = true;
3616 : 12490543 : FOR_EACH_EDGE (e, ei, block->succs)
3617 : : {
3618 : 8362726 : if (first)
3619 : : {
3620 : 4127817 : phi_translate_set (ANTIC_OUT, ANTIC_IN (e->dest), e);
3621 : 4127817 : first = false;
3622 : : }
3623 : 4234909 : else if (!gimple_seq_empty_p (phi_nodes (e->dest)))
3624 : : {
3625 : 1 : bitmap_set_t tmp = bitmap_set_new ();
3626 : 1 : phi_translate_set (tmp, ANTIC_IN (e->dest), e);
3627 : 1 : bitmap_and_into (&ANTIC_OUT->values, &tmp->values);
3628 : 1 : bitmap_ior_into (&ANTIC_OUT->expressions, &tmp->expressions);
3629 : 1 : bitmap_set_free (tmp);
3630 : : }
3631 : : else
3632 : : {
3633 : 4234908 : bitmap_and_into (&ANTIC_OUT->values, &ANTIC_IN (e->dest)->values);
3634 : 4234908 : bitmap_ior_into (&ANTIC_OUT->expressions,
3635 : 4234908 : &ANTIC_IN (e->dest)->expressions);
3636 : : }
3637 : : }
3638 : :
3639 : : /* Compute the set of hoistable expressions from ANTIC_OUT. First compute
3640 : : hoistable values. */
3641 : 4127817 : bitmap_set hoistable_set;
3642 : :
3643 : : /* A hoistable value must be in ANTIC_OUT(block)
3644 : : but not in AVAIL_OUT(BLOCK). */
3645 : 4127817 : bitmap_initialize (&hoistable_set.values, &grand_bitmap_obstack);
3646 : 4127817 : bitmap_and_compl (&hoistable_set.values,
3647 : 4127817 : &ANTIC_OUT->values, &AVAIL_OUT (block)->values);
3648 : :
3649 : : /* Short-cut for a common case: hoistable_set is empty. */
3650 : 4127817 : if (bitmap_empty_p (&hoistable_set.values))
3651 : : {
3652 : 3401315 : bitmap_set_free (ANTIC_OUT);
3653 : 3401315 : return false;
3654 : : }
3655 : :
3656 : : /* Compute which of the hoistable values is in AVAIL_OUT of
3657 : : at least one of the successors of BLOCK. */
3658 : 726502 : bitmap_head availout_in_some;
3659 : 726502 : bitmap_initialize (&availout_in_some, &grand_bitmap_obstack);
3660 : 2185791 : FOR_EACH_EDGE (e, ei, block->succs)
3661 : : /* Do not consider expressions solely because their availability
3662 : : on loop exits. They'd be ANTIC-IN throughout the whole loop
3663 : : and thus effectively hoisted across loops by combination of
3664 : : PRE and hoisting. */
3665 : 1459289 : if (! loop_exit_edge_p (block->loop_father, e))
3666 : 1295283 : bitmap_ior_and_into (&availout_in_some, &hoistable_set.values,
3667 : 1295283 : &AVAIL_OUT (e->dest)->values);
3668 : 726502 : bitmap_clear (&hoistable_set.values);
3669 : :
3670 : : /* Short-cut for a common case: availout_in_some is empty. */
3671 : 726502 : if (bitmap_empty_p (&availout_in_some))
3672 : : {
3673 : 581524 : bitmap_set_free (ANTIC_OUT);
3674 : 581524 : return false;
3675 : : }
3676 : :
3677 : : /* Hack hoistable_set in-place so we can use sorted_array_from_bitmap_set. */
3678 : 144978 : bitmap_move (&hoistable_set.values, &availout_in_some);
3679 : 144978 : hoistable_set.expressions = ANTIC_OUT->expressions;
3680 : :
3681 : : /* Now finally construct the topological-ordered expression set. */
3682 : 144978 : vec<pre_expr> exprs = sorted_array_from_bitmap_set (&hoistable_set);
3683 : :
3684 : : /* If there are candidate values for hoisting, insert expressions
3685 : : strategically to make the hoistable expressions fully redundant. */
3686 : 144978 : pre_expr expr;
3687 : 428054 : FOR_EACH_VEC_ELT (exprs, i, expr)
3688 : : {
3689 : : /* While we try to sort expressions topologically above the
3690 : : sorting doesn't work out perfectly. Catch expressions we
3691 : : already inserted. */
3692 : 283076 : unsigned int value_id = get_expr_value_id (expr);
3693 : 566152 : if (bitmap_set_contains_value (AVAIL_OUT (block), value_id))
3694 : : {
3695 : 59864 : if (dump_file && (dump_flags & TDF_DETAILS))
3696 : : {
3697 : 1 : fprintf (dump_file,
3698 : : "Already inserted expression for ");
3699 : 1 : print_pre_expr (dump_file, expr);
3700 : 1 : fprintf (dump_file, " (%04d)\n", value_id);
3701 : : }
3702 : 59882 : continue;
3703 : : }
3704 : :
3705 : : /* If we end up with a punned expression representation and this
3706 : : happens to be a float typed one give up - we can't know for
3707 : : sure whether all paths perform the floating-point load we are
3708 : : about to insert and on some targets this can cause correctness
3709 : : issues. See PR88240. */
3710 : 223212 : if (expr->kind == REFERENCE
3711 : 101932 : && PRE_EXPR_REFERENCE (expr)->punned
3712 : 223402 : && FLOAT_TYPE_P (get_expr_type (expr)))
3713 : 0 : continue;
3714 : :
3715 : : /* Only hoist if the full expression is available for hoisting.
3716 : : This avoids hoisting values that are not common and for
3717 : : example evaluate an expression that's not valid to evaluate
3718 : : unconditionally (PR112310). */
3719 : 223212 : if (!valid_in_sets (&hoistable_set, AVAIL_OUT (block), expr))
3720 : 18 : continue;
3721 : :
3722 : : /* OK, we should hoist this value. Perform the transformation. */
3723 : 223194 : pre_stats.hoist_insert++;
3724 : 223194 : if (dump_file && (dump_flags & TDF_DETAILS))
3725 : : {
3726 : 4 : fprintf (dump_file,
3727 : : "Inserting expression in block %d for code hoisting: ",
3728 : : block->index);
3729 : 4 : print_pre_expr (dump_file, expr);
3730 : 4 : fprintf (dump_file, " (%04d)\n", value_id);
3731 : : }
3732 : :
3733 : 223194 : gimple_seq stmts = NULL;
3734 : 223194 : tree res = create_expression_by_pieces (block, expr, &stmts,
3735 : : get_expr_type (expr));
3736 : :
3737 : : /* Do not return true if expression creation ultimately
3738 : : did not insert any statements. */
3739 : 223194 : if (gimple_seq_empty_p (stmts))
3740 : : res = NULL_TREE;
3741 : : else
3742 : : {
3743 : 223194 : if (gsi_end_p (last) || is_ctrl_stmt (gsi_stmt (last)))
3744 : 223194 : gsi_insert_seq_before (&last, stmts, GSI_SAME_STMT);
3745 : : else
3746 : 0 : gsi_insert_seq_after (&last, stmts, GSI_NEW_STMT);
3747 : : }
3748 : :
3749 : : /* Make sure to not return true if expression creation ultimately
3750 : : failed but also make sure to insert any stmts produced as they
3751 : : are tracked in inserted_exprs. */
3752 : 223194 : if (! res)
3753 : 0 : continue;
3754 : :
3755 : 223194 : new_stuff = true;
3756 : : }
3757 : :
3758 : 144978 : exprs.release ();
3759 : 144978 : bitmap_clear (&hoistable_set.values);
3760 : 144978 : bitmap_set_free (ANTIC_OUT);
3761 : :
3762 : 144978 : return new_stuff;
3763 : : }
3764 : :
3765 : : /* Perform insertion of partially redundant and hoistable values. */
3766 : :
3767 : : static void
3768 : 953798 : insert (void)
3769 : : {
3770 : 953798 : basic_block bb;
3771 : :
3772 : 16192228 : FOR_ALL_BB_FN (bb, cfun)
3773 : 15238430 : NEW_SETS (bb) = bitmap_set_new ();
3774 : :
3775 : 953798 : int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
3776 : 953798 : int *bb_rpo = XNEWVEC (int, last_basic_block_for_fn (cfun) + 1);
3777 : 953798 : int rpo_num = pre_and_rev_post_order_compute (NULL, rpo, false);
3778 : 14284632 : for (int i = 0; i < rpo_num; ++i)
3779 : 13330834 : bb_rpo[rpo[i]] = i;
3780 : :
3781 : : int num_iterations = 0;
3782 : 1005743 : bool changed;
3783 : 1005743 : do
3784 : : {
3785 : 1005743 : num_iterations++;
3786 : 1005743 : if (dump_file && dump_flags & TDF_DETAILS)
3787 : 18 : fprintf (dump_file, "Starting insert iteration %d\n", num_iterations);
3788 : :
3789 : : changed = false;
3790 : 18517975 : for (int idx = 0; idx < rpo_num; ++idx)
3791 : : {
3792 : 17512232 : basic_block block = BASIC_BLOCK_FOR_FN (cfun, rpo[idx]);
3793 : 17512232 : basic_block dom = get_immediate_dominator (CDI_DOMINATORS, block);
3794 : 17512232 : if (dom)
3795 : : {
3796 : 17512232 : unsigned i;
3797 : 17512232 : bitmap_iterator bi;
3798 : 17512232 : bitmap_set_t newset;
3799 : :
3800 : : /* First, update the AVAIL_OUT set with anything we may have
3801 : : inserted higher up in the dominator tree. */
3802 : 17512232 : newset = NEW_SETS (dom);
3803 : :
3804 : : /* Note that we need to value_replace both NEW_SETS, and
3805 : : AVAIL_OUT. For both the case of NEW_SETS, the value may be
3806 : : represented by some non-simple expression here that we want
3807 : : to replace it with. */
3808 : 17512232 : bool avail_out_changed = false;
3809 : 32709258 : FOR_EACH_EXPR_ID_IN_SET (newset, i, bi)
3810 : : {
3811 : 15197026 : pre_expr expr = expression_for_id (i);
3812 : 15197026 : bitmap_value_replace_in_set (NEW_SETS (block), expr);
3813 : 15197026 : avail_out_changed
3814 : 15197026 : |= bitmap_value_replace_in_set (AVAIL_OUT (block), expr);
3815 : : }
3816 : : /* We need to iterate if AVAIL_OUT of an already processed
3817 : : block source changed. */
3818 : 17512232 : if (avail_out_changed && !changed)
3819 : : {
3820 : 1761063 : edge_iterator ei;
3821 : 1761063 : edge e;
3822 : 4194074 : FOR_EACH_EDGE (e, ei, block->succs)
3823 : 2433011 : if (e->dest->index != EXIT_BLOCK
3824 : 2322542 : && bb_rpo[e->dest->index] < idx)
3825 : 2433011 : changed = true;
3826 : : }
3827 : :
3828 : : /* Insert expressions for partial redundancies. */
3829 : 35023901 : if (flag_tree_pre && !single_pred_p (block))
3830 : : {
3831 : 3433049 : vec<pre_expr> exprs
3832 : 3433049 : = sorted_array_from_bitmap_set (ANTIC_IN (block));
3833 : : /* Sorting is not perfect, iterate locally. */
3834 : 7137528 : while (do_pre_regular_insertion (block, dom, exprs))
3835 : : ;
3836 : 3433049 : exprs.release ();
3837 : 3433049 : if (do_partial_partial)
3838 : : {
3839 : 295674 : exprs = sorted_array_from_bitmap_set (PA_IN (block));
3840 : 594282 : while (do_pre_partial_partial_insertion (block, dom,
3841 : : exprs))
3842 : : ;
3843 : 295674 : exprs.release ();
3844 : : }
3845 : : }
3846 : : }
3847 : : }
3848 : :
3849 : : /* Clear the NEW sets before the next iteration. We have already
3850 : : fully propagated its contents. */
3851 : 1005743 : if (changed)
3852 : 4337233 : FOR_ALL_BB_FN (bb, cfun)
3853 : 8570576 : bitmap_set_free (NEW_SETS (bb));
3854 : : }
3855 : : while (changed);
3856 : :
3857 : 953798 : statistics_histogram_event (cfun, "insert iterations", num_iterations);
3858 : :
3859 : : /* AVAIL_OUT is not needed after insertion so we don't have to
3860 : : propagate NEW_SETS from hoist insertion. */
3861 : 16192228 : FOR_ALL_BB_FN (bb, cfun)
3862 : : {
3863 : 15238430 : bitmap_set_free (NEW_SETS (bb));
3864 : 15238430 : bitmap_set_pool.remove (NEW_SETS (bb));
3865 : 15238430 : NEW_SETS (bb) = NULL;
3866 : : }
3867 : :
3868 : : /* Insert expressions for hoisting. Do a backward walk here since
3869 : : inserting into BLOCK exposes new opportunities in its predecessors.
3870 : : Since PRE and hoist insertions can cause back-to-back iteration
3871 : : and we are interested in PRE insertion exposed hoisting opportunities
3872 : : but not in hoisting exposed PRE ones do hoist insertion only after
3873 : : PRE insertion iteration finished and do not iterate it. */
3874 : 953798 : if (flag_code_hoisting)
3875 : 14284090 : for (int idx = rpo_num - 1; idx >= 0; --idx)
3876 : : {
3877 : 13330344 : basic_block block = BASIC_BLOCK_FOR_FN (cfun, rpo[idx]);
3878 : 18118775 : if (EDGE_COUNT (block->succs) >= 2)
3879 : 4788431 : changed |= do_hoist_insertion (block);
3880 : : }
3881 : :
3882 : 953798 : free (rpo);
3883 : 953798 : free (bb_rpo);
3884 : 953798 : }
3885 : :
3886 : :
3887 : : /* Compute the AVAIL set for all basic blocks.
3888 : :
3889 : : This function performs value numbering of the statements in each basic
3890 : : block. The AVAIL sets are built from information we glean while doing
3891 : : this value numbering, since the AVAIL sets contain only one entry per
3892 : : value.
3893 : :
3894 : : AVAIL_IN[BLOCK] = AVAIL_OUT[dom(BLOCK)].
3895 : : AVAIL_OUT[BLOCK] = AVAIL_IN[BLOCK] U PHI_GEN[BLOCK] U TMP_GEN[BLOCK]. */
3896 : :
3897 : : static void
3898 : 953798 : compute_avail (function *fun)
3899 : : {
3900 : :
3901 : 953798 : basic_block block, son;
3902 : 953798 : basic_block *worklist;
3903 : 953798 : size_t sp = 0;
3904 : 953798 : unsigned i;
3905 : 953798 : tree name;
3906 : :
3907 : : /* We pretend that default definitions are defined in the entry block.
3908 : : This includes function arguments and the static chain decl. */
3909 : 46736371 : FOR_EACH_SSA_NAME (i, name, fun)
3910 : : {
3911 : 33830671 : pre_expr e;
3912 : 33830671 : if (!SSA_NAME_IS_DEFAULT_DEF (name)
3913 : 2889615 : || has_zero_uses (name)
3914 : 36189721 : || virtual_operand_p (name))
3915 : 32424631 : continue;
3916 : :
3917 : 1406040 : e = get_or_alloc_expr_for_name (name);
3918 : 1406040 : add_to_value (get_expr_value_id (e), e);
3919 : 1406040 : bitmap_insert_into_set (TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (fun)), e);
3920 : 1406040 : bitmap_value_insert_into_set (AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (fun)),
3921 : : e);
3922 : : }
3923 : :
3924 : 953798 : if (dump_file && (dump_flags & TDF_DETAILS))
3925 : : {
3926 : 14 : print_bitmap_set (dump_file, TMP_GEN (ENTRY_BLOCK_PTR_FOR_FN (fun)),
3927 : : "tmp_gen", ENTRY_BLOCK);
3928 : 14 : print_bitmap_set (dump_file, AVAIL_OUT (ENTRY_BLOCK_PTR_FOR_FN (fun)),
3929 : : "avail_out", ENTRY_BLOCK);
3930 : : }
3931 : :
3932 : : /* Allocate the worklist. */
3933 : 953798 : worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (fun));
3934 : :
3935 : : /* Seed the algorithm by putting the dominator children of the entry
3936 : : block on the worklist. */
3937 : 953798 : for (son = first_dom_son (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (fun));
3938 : 1907596 : son;
3939 : 953798 : son = next_dom_son (CDI_DOMINATORS, son))
3940 : 953798 : worklist[sp++] = son;
3941 : :
3942 : 1907596 : BB_LIVE_VOP_ON_EXIT (ENTRY_BLOCK_PTR_FOR_FN (fun))
3943 : 953798 : = ssa_default_def (fun, gimple_vop (fun));
3944 : :
3945 : : /* Loop until the worklist is empty. */
3946 : 14284632 : while (sp)
3947 : : {
3948 : 13330834 : gimple *stmt;
3949 : 13330834 : basic_block dom;
3950 : :
3951 : : /* Pick a block from the worklist. */
3952 : 13330834 : block = worklist[--sp];
3953 : 13330834 : vn_context_bb = block;
3954 : :
3955 : : /* Initially, the set of available values in BLOCK is that of
3956 : : its immediate dominator. */
3957 : 13330834 : dom = get_immediate_dominator (CDI_DOMINATORS, block);
3958 : 13330834 : if (dom)
3959 : : {
3960 : 13330834 : bitmap_set_copy (AVAIL_OUT (block), AVAIL_OUT (dom));
3961 : 13330834 : BB_LIVE_VOP_ON_EXIT (block) = BB_LIVE_VOP_ON_EXIT (dom);
3962 : : }
3963 : :
3964 : : /* Generate values for PHI nodes. */
3965 : 17214717 : for (gphi_iterator gsi = gsi_start_phis (block); !gsi_end_p (gsi);
3966 : 3883883 : gsi_next (&gsi))
3967 : : {
3968 : 3883883 : tree result = gimple_phi_result (gsi.phi ());
3969 : :
3970 : : /* We have no need for virtual phis, as they don't represent
3971 : : actual computations. */
3972 : 7767766 : if (virtual_operand_p (result))
3973 : : {
3974 : 1765432 : BB_LIVE_VOP_ON_EXIT (block) = result;
3975 : 1765432 : continue;
3976 : : }
3977 : :
3978 : 2118451 : pre_expr e = get_or_alloc_expr_for_name (result);
3979 : 2118451 : add_to_value (get_expr_value_id (e), e);
3980 : 2118451 : bitmap_value_insert_into_set (AVAIL_OUT (block), e);
3981 : 2118451 : bitmap_insert_into_set (PHI_GEN (block), e);
3982 : : }
3983 : :
3984 : 13330834 : BB_MAY_NOTRETURN (block) = 0;
3985 : :
3986 : : /* Now compute value numbers and populate value sets with all
3987 : : the expressions computed in BLOCK. */
3988 : 13330834 : bool set_bb_may_notreturn = false;
3989 : 107157215 : for (gimple_stmt_iterator gsi = gsi_start_bb (block); !gsi_end_p (gsi);
3990 : 80495547 : gsi_next (&gsi))
3991 : : {
3992 : 80495547 : ssa_op_iter iter;
3993 : 80495547 : tree op;
3994 : :
3995 : 80495547 : stmt = gsi_stmt (gsi);
3996 : :
3997 : 80495547 : if (set_bb_may_notreturn)
3998 : : {
3999 : 2654063 : BB_MAY_NOTRETURN (block) = 1;
4000 : 2654063 : set_bb_may_notreturn = false;
4001 : : }
4002 : :
4003 : : /* Cache whether the basic-block has any non-visible side-effect
4004 : : or control flow.
4005 : : If this isn't a call or it is the last stmt in the
4006 : : basic-block then the CFG represents things correctly. */
4007 : 80495547 : if (is_gimple_call (stmt) && !stmt_ends_bb_p (stmt))
4008 : : {
4009 : : /* Non-looping const functions always return normally.
4010 : : Otherwise the call might not return or have side-effects
4011 : : that forbids hoisting possibly trapping expressions
4012 : : before it. */
4013 : 3705780 : int flags = gimple_call_flags (stmt);
4014 : 3705780 : if (!(flags & (ECF_CONST|ECF_PURE))
4015 : 561965 : || (flags & ECF_LOOPING_CONST_OR_PURE)
4016 : 4241532 : || stmt_can_throw_external (fun, stmt))
4017 : : /* Defer setting of BB_MAY_NOTRETURN to avoid it
4018 : : influencing the processing of the call itself. */
4019 : : set_bb_may_notreturn = true;
4020 : : }
4021 : :
4022 : 95419370 : FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_DEF)
4023 : : {
4024 : 14923823 : pre_expr e = get_or_alloc_expr_for_name (op);
4025 : 14923823 : add_to_value (get_expr_value_id (e), e);
4026 : 14923823 : bitmap_insert_into_set (TMP_GEN (block), e);
4027 : 14923823 : bitmap_value_insert_into_set (AVAIL_OUT (block), e);
4028 : : }
4029 : :
4030 : 107561762 : if (gimple_vdef (stmt))
4031 : 12042637 : BB_LIVE_VOP_ON_EXIT (block) = gimple_vdef (stmt);
4032 : :
4033 : 80495547 : if (gimple_has_side_effects (stmt)
4034 : 74075437 : || stmt_could_throw_p (fun, stmt)
4035 : 153410706 : || is_gimple_debug (stmt))
4036 : 75169397 : continue;
4037 : :
4038 : 47006150 : FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
4039 : : {
4040 : 22040583 : if (ssa_undefined_value_p (op))
4041 : 48780 : continue;
4042 : 21991803 : pre_expr e = get_or_alloc_expr_for_name (op);
4043 : 21991803 : bitmap_value_insert_into_set (EXP_GEN (block), e);
4044 : : }
4045 : :
4046 : 24965567 : switch (gimple_code (stmt))
4047 : : {
4048 : 928915 : case GIMPLE_RETURN:
4049 : 928915 : continue;
4050 : :
4051 : 533456 : case GIMPLE_CALL:
4052 : 533456 : {
4053 : 533456 : vn_reference_t ref;
4054 : 533456 : vn_reference_s ref1;
4055 : 533456 : pre_expr result = NULL;
4056 : :
4057 : 533456 : vn_reference_lookup_call (as_a <gcall *> (stmt), &ref, &ref1);
4058 : : /* There is no point to PRE a call without a value. */
4059 : 533456 : if (!ref || !ref->result)
4060 : 7901 : continue;
4061 : :
4062 : : /* If the value of the call is not invalidated in
4063 : : this block until it is computed, add the expression
4064 : : to EXP_GEN. */
4065 : 525555 : if ((!gimple_vuse (stmt)
4066 : 295212 : || gimple_code
4067 : 295212 : (SSA_NAME_DEF_STMT (gimple_vuse (stmt))) == GIMPLE_PHI
4068 : 268856 : || gimple_bb (SSA_NAME_DEF_STMT
4069 : : (gimple_vuse (stmt))) != block)
4070 : : /* If the REFERENCE traps and there was a preceding
4071 : : point in the block that might not return avoid
4072 : : adding the reference to EXP_GEN. */
4073 : 766835 : && (!BB_MAY_NOTRETURN (block)
4074 : 9974 : || !vn_reference_may_trap (ref)))
4075 : : {
4076 : 461649 : result = get_or_alloc_expr_for_reference
4077 : 461649 : (ref, gimple_location (stmt));
4078 : 461649 : add_to_value (get_expr_value_id (result), result);
4079 : 461649 : bitmap_value_insert_into_set (EXP_GEN (block), result);
4080 : : }
4081 : 525555 : continue;
4082 : 525555 : }
4083 : :
4084 : 18177046 : case GIMPLE_ASSIGN:
4085 : 18177046 : {
4086 : 18177046 : pre_expr result = NULL;
4087 : 18177046 : switch (vn_get_stmt_kind (stmt))
4088 : : {
4089 : 7499832 : case VN_NARY:
4090 : 7499832 : {
4091 : 7499832 : enum tree_code code = gimple_assign_rhs_code (stmt);
4092 : 7499832 : vn_nary_op_t nary;
4093 : :
4094 : : /* COND_EXPR is awkward in that it contains an
4095 : : embedded complex expression.
4096 : : Don't even try to shove it through PRE. */
4097 : 7499832 : if (code == COND_EXPR)
4098 : 138995 : continue;
4099 : :
4100 : 7496291 : vn_nary_op_lookup_stmt (stmt, &nary);
4101 : 7496291 : if (!nary || nary->predicated_values)
4102 : 107344 : continue;
4103 : :
4104 : 7388947 : unsigned value_id = nary->value_id;
4105 : 7388947 : if (value_id_constant_p (value_id))
4106 : 0 : continue;
4107 : :
4108 : : /* Record the un-valueized expression for EXP_GEN. */
4109 : 7388947 : nary = XALLOCAVAR (struct vn_nary_op_s,
4110 : : sizeof_vn_nary_op
4111 : : (vn_nary_length_from_stmt (stmt)));
4112 : 7388947 : init_vn_nary_op_from_stmt (nary, as_a <gassign *> (stmt));
4113 : :
4114 : : /* If the NARY traps and there was a preceding
4115 : : point in the block that might not return avoid
4116 : : adding the nary to EXP_GEN. */
4117 : 7417057 : if (BB_MAY_NOTRETURN (block)
4118 : 7388947 : && vn_nary_may_trap (nary))
4119 : 28110 : continue;
4120 : :
4121 : 7360837 : result = get_or_alloc_expr_for_nary
4122 : 7360837 : (nary, value_id, gimple_location (stmt));
4123 : 7360837 : break;
4124 : : }
4125 : :
4126 : 5212405 : case VN_REFERENCE:
4127 : 5212405 : {
4128 : 5212405 : tree rhs1 = gimple_assign_rhs1 (stmt);
4129 : 5212405 : ao_ref rhs1_ref;
4130 : 5212405 : ao_ref_init (&rhs1_ref, rhs1);
4131 : 5212405 : alias_set_type set = ao_ref_alias_set (&rhs1_ref);
4132 : 5212405 : alias_set_type base_set
4133 : 5212405 : = ao_ref_base_alias_set (&rhs1_ref);
4134 : 5212405 : vec<vn_reference_op_s> operands
4135 : 5212405 : = vn_reference_operands_for_lookup (rhs1);
4136 : 5212405 : vn_reference_t ref;
4137 : :
4138 : : /* We handle &MEM[ptr + 5].b[1].c as
4139 : : POINTER_PLUS_EXPR. */
4140 : 5212405 : if (operands[0].opcode == ADDR_EXPR
4141 : 5481698 : && operands.last ().opcode == SSA_NAME)
4142 : : {
4143 : 269285 : tree ops[2];
4144 : 269285 : if (vn_pp_nary_for_addr (operands, ops))
4145 : : {
4146 : 184114 : vn_nary_op_t nary;
4147 : 184114 : vn_nary_op_lookup_pieces (2, POINTER_PLUS_EXPR,
4148 : 184114 : TREE_TYPE (rhs1), ops,
4149 : : &nary);
4150 : 184114 : operands.release ();
4151 : 184114 : if (nary && !nary->predicated_values)
4152 : : {
4153 : 184105 : unsigned value_id = nary->value_id;
4154 : 184105 : if (value_id_constant_p (value_id))
4155 : 9 : continue;
4156 : 184105 : result = get_or_alloc_expr_for_nary
4157 : 184105 : (nary, value_id, gimple_location (stmt));
4158 : 184105 : break;
4159 : : }
4160 : 9 : continue;
4161 : 9 : }
4162 : : }
4163 : :
4164 : 10056582 : vn_reference_lookup_pieces (gimple_vuse (stmt), set,
4165 : 5028291 : base_set, TREE_TYPE (rhs1),
4166 : : operands, &ref, VN_WALK);
4167 : 5028291 : if (!ref)
4168 : : {
4169 : 369359 : operands.release ();
4170 : 369359 : continue;
4171 : : }
4172 : :
4173 : : /* If the REFERENCE traps and there was a preceding
4174 : : point in the block that might not return avoid
4175 : : adding the reference to EXP_GEN. */
4176 : 4885886 : if (BB_MAY_NOTRETURN (block)
4177 : 4658932 : && vn_reference_may_trap (ref))
4178 : : {
4179 : 226954 : operands.release ();
4180 : 226954 : continue;
4181 : : }
4182 : :
4183 : : /* If the value of the reference is not invalidated in
4184 : : this block until it is computed, add the expression
4185 : : to EXP_GEN. */
4186 : 8863956 : if (gimple_vuse (stmt))
4187 : : {
4188 : 4346812 : gimple *def_stmt;
4189 : 4346812 : bool ok = true;
4190 : 4346812 : def_stmt = SSA_NAME_DEF_STMT (gimple_vuse (stmt));
4191 : 7118892 : while (!gimple_nop_p (def_stmt)
4192 : 6151813 : && gimple_code (def_stmt) != GIMPLE_PHI
4193 : 12053608 : && gimple_bb (def_stmt) == block)
4194 : : {
4195 : 3657403 : if (stmt_may_clobber_ref_p
4196 : 3657403 : (def_stmt, gimple_assign_rhs1 (stmt)))
4197 : : {
4198 : : ok = false;
4199 : : break;
4200 : : }
4201 : 2772080 : def_stmt
4202 : 2772080 : = SSA_NAME_DEF_STMT (gimple_vuse (def_stmt));
4203 : : }
4204 : 4346812 : if (!ok)
4205 : : {
4206 : 885323 : operands.release ();
4207 : 885323 : continue;
4208 : : }
4209 : : }
4210 : :
4211 : : /* If the load was value-numbered to another
4212 : : load make sure we do not use its expression
4213 : : for insertion if it wouldn't be a valid
4214 : : replacement. */
4215 : : /* At the momemt we have a testcase
4216 : : for hoist insertion of aligned vs. misaligned
4217 : : variants in gcc.dg/torture/pr65270-1.c thus
4218 : : with just alignment to be considered we can
4219 : : simply replace the expression in the hashtable
4220 : : with the most conservative one. */
4221 : 3546655 : vn_reference_op_t ref1 = &ref->operands.last ();
4222 : 3546655 : while (ref1->opcode != TARGET_MEM_REF
4223 : 7093217 : && ref1->opcode != MEM_REF
4224 : 7093217 : && ref1 != &ref->operands[0])
4225 : 3546562 : --ref1;
4226 : 3546655 : vn_reference_op_t ref2 = &operands.last ();
4227 : 3546655 : while (ref2->opcode != TARGET_MEM_REF
4228 : 7093222 : && ref2->opcode != MEM_REF
4229 : 10640096 : && ref2 != &operands[0])
4230 : 3546567 : --ref2;
4231 : 3546655 : if ((ref1->opcode == TARGET_MEM_REF
4232 : : || ref1->opcode == MEM_REF)
4233 : 7092998 : && (TYPE_ALIGN (ref1->type)
4234 : 3546343 : > TYPE_ALIGN (ref2->type)))
4235 : 1141 : ref1->type
4236 : 1141 : = build_aligned_type (ref1->type,
4237 : 1141 : TYPE_ALIGN (ref2->type));
4238 : : /* TBAA behavior is an obvious part so make sure
4239 : : that the hashtable one covers this as well
4240 : : by adjusting the ref alias set and its base. */
4241 : 3546655 : if ((ref->set == set
4242 : 12831 : || alias_set_subset_of (set, ref->set))
4243 : 3552211 : && (ref->base_set == base_set
4244 : 11723 : || alias_set_subset_of (base_set, ref->base_set)))
4245 : : ;
4246 : 14085 : else if (ref1->opcode != ref2->opcode
4247 : 14080 : || (ref1->opcode != MEM_REF
4248 : 14080 : && ref1->opcode != TARGET_MEM_REF))
4249 : : {
4250 : : /* With mismatching base opcodes or bases
4251 : : other than MEM_REF or TARGET_MEM_REF we
4252 : : can't do any easy TBAA adjustment. */
4253 : 5 : operands.release ();
4254 : 5 : continue;
4255 : : }
4256 : 14080 : else if (ref->set == set
4257 : 14080 : || alias_set_subset_of (ref->set, set))
4258 : : {
4259 : 13530 : tree reft = reference_alias_ptr_type (rhs1);
4260 : 13530 : ref->set = set;
4261 : 13530 : ref->base_set = set;
4262 : 13530 : if (ref1->opcode == MEM_REF)
4263 : 13530 : ref1->op0
4264 : 27060 : = wide_int_to_tree (reft,
4265 : 13530 : wi::to_wide (ref1->op0));
4266 : : else
4267 : 0 : ref1->op2
4268 : 0 : = wide_int_to_tree (reft,
4269 : 0 : wi::to_wide (ref1->op2));
4270 : : }
4271 : : else
4272 : : {
4273 : 550 : ref->set = 0;
4274 : 550 : ref->base_set = 0;
4275 : 550 : if (ref1->opcode == MEM_REF)
4276 : 550 : ref1->op0
4277 : 1100 : = wide_int_to_tree (ptr_type_node,
4278 : 550 : wi::to_wide (ref1->op0));
4279 : : else
4280 : 0 : ref1->op2
4281 : 0 : = wide_int_to_tree (ptr_type_node,
4282 : 0 : wi::to_wide (ref1->op2));
4283 : : }
4284 : : /* We also need to make sure that the access path
4285 : : ends in an access of the same size as otherwise
4286 : : we might assume an access may not trap while in
4287 : : fact it might. That's independent of whether
4288 : : TBAA is in effect. */
4289 : 3546650 : if (TYPE_SIZE (ref1->type) != TYPE_SIZE (ref2->type)
4290 : 3546650 : && (! TYPE_SIZE (ref1->type)
4291 : 11448 : || ! TYPE_SIZE (ref2->type)
4292 : 11445 : || ! operand_equal_p (TYPE_SIZE (ref1->type),
4293 : 11445 : TYPE_SIZE (ref2->type))))
4294 : : {
4295 : 11456 : operands.release ();
4296 : 11456 : continue;
4297 : : }
4298 : 3535194 : operands.release ();
4299 : :
4300 : 3535194 : result = get_or_alloc_expr_for_reference
4301 : 3535194 : (ref, gimple_location (stmt));
4302 : 3535194 : break;
4303 : : }
4304 : :
4305 : 5464809 : default:
4306 : 5464809 : continue;
4307 : 5464809 : }
4308 : :
4309 : 11080136 : add_to_value (get_expr_value_id (result), result);
4310 : 11080136 : bitmap_value_insert_into_set (EXP_GEN (block), result);
4311 : 11080136 : continue;
4312 : 11080136 : }
4313 : 5326150 : default:
4314 : 5326150 : break;
4315 : 928915 : }
4316 : : }
4317 : 13330834 : if (set_bb_may_notreturn)
4318 : : {
4319 : 518261 : BB_MAY_NOTRETURN (block) = 1;
4320 : 518261 : set_bb_may_notreturn = false;
4321 : : }
4322 : :
4323 : 13330834 : if (dump_file && (dump_flags & TDF_DETAILS))
4324 : : {
4325 : 108 : print_bitmap_set (dump_file, EXP_GEN (block),
4326 : : "exp_gen", block->index);
4327 : 108 : print_bitmap_set (dump_file, PHI_GEN (block),
4328 : : "phi_gen", block->index);
4329 : 108 : print_bitmap_set (dump_file, TMP_GEN (block),
4330 : : "tmp_gen", block->index);
4331 : 108 : print_bitmap_set (dump_file, AVAIL_OUT (block),
4332 : : "avail_out", block->index);
4333 : : }
4334 : :
4335 : : /* Put the dominator children of BLOCK on the worklist of blocks
4336 : : to compute available sets for. */
4337 : 13330834 : for (son = first_dom_son (CDI_DOMINATORS, block);
4338 : 25707870 : son;
4339 : 12377036 : son = next_dom_son (CDI_DOMINATORS, son))
4340 : 12377036 : worklist[sp++] = son;
4341 : : }
4342 : 953798 : vn_context_bb = NULL;
4343 : :
4344 : 953798 : free (worklist);
4345 : 953798 : }
4346 : :
4347 : :
4348 : : /* Initialize data structures used by PRE. */
4349 : :
4350 : : static void
4351 : 953804 : init_pre (void)
4352 : : {
4353 : 953804 : basic_block bb;
4354 : :
4355 : 953804 : next_expression_id = 1;
4356 : 953804 : expressions.create (0);
4357 : 953804 : expressions.safe_push (NULL);
4358 : 953804 : value_expressions.create (get_max_value_id () + 1);
4359 : 953804 : value_expressions.quick_grow_cleared (get_max_value_id () + 1);
4360 : 953804 : constant_value_expressions.create (get_max_constant_value_id () + 1);
4361 : 953804 : constant_value_expressions.quick_grow_cleared (get_max_constant_value_id () + 1);
4362 : 953804 : name_to_id.create (0);
4363 : 953804 : gcc_obstack_init (&pre_expr_obstack);
4364 : :
4365 : 953804 : inserted_exprs = BITMAP_ALLOC (NULL);
4366 : :
4367 : 953804 : connect_infinite_loops_to_exit ();
4368 : 953804 : memset (&pre_stats, 0, sizeof (pre_stats));
4369 : :
4370 : 953804 : alloc_aux_for_blocks (sizeof (struct bb_bitmap_sets));
4371 : :
4372 : 953804 : calculate_dominance_info (CDI_DOMINATORS);
4373 : :
4374 : 953804 : bitmap_obstack_initialize (&grand_bitmap_obstack);
4375 : 1907608 : expression_to_id = new hash_table<pre_expr_d> (num_ssa_names * 3);
4376 : 16225244 : FOR_ALL_BB_FN (bb, cfun)
4377 : : {
4378 : 15271440 : EXP_GEN (bb) = bitmap_set_new ();
4379 : 15271440 : PHI_GEN (bb) = bitmap_set_new ();
4380 : 15271440 : TMP_GEN (bb) = bitmap_set_new ();
4381 : 15271440 : AVAIL_OUT (bb) = bitmap_set_new ();
4382 : 15271440 : PHI_TRANS_TABLE (bb) = NULL;
4383 : : }
4384 : 953804 : }
4385 : :
4386 : :
4387 : : /* Deallocate data structures used by PRE. */
4388 : :
4389 : : static void
4390 : 953804 : fini_pre ()
4391 : : {
4392 : 953804 : value_expressions.release ();
4393 : 953804 : constant_value_expressions.release ();
4394 : 953804 : expressions.release ();
4395 : 953804 : bitmap_obstack_release (&grand_bitmap_obstack);
4396 : 953804 : bitmap_set_pool.release ();
4397 : 953804 : pre_expr_pool.release ();
4398 : 953804 : delete expression_to_id;
4399 : 953804 : expression_to_id = NULL;
4400 : 953804 : name_to_id.release ();
4401 : 953804 : obstack_free (&pre_expr_obstack, NULL);
4402 : :
4403 : 953804 : basic_block bb;
4404 : 16224920 : FOR_ALL_BB_FN (bb, cfun)
4405 : 15271116 : if (bb->aux && PHI_TRANS_TABLE (bb))
4406 : 6127064 : delete PHI_TRANS_TABLE (bb);
4407 : 953804 : free_aux_for_blocks ();
4408 : 953804 : }
4409 : :
4410 : : namespace {
4411 : :
4412 : : const pass_data pass_data_pre =
4413 : : {
4414 : : GIMPLE_PASS, /* type */
4415 : : "pre", /* name */
4416 : : OPTGROUP_NONE, /* optinfo_flags */
4417 : : TV_TREE_PRE, /* tv_id */
4418 : : ( PROP_cfg | PROP_ssa ), /* properties_required */
4419 : : 0, /* properties_provided */
4420 : : 0, /* properties_destroyed */
4421 : : TODO_rebuild_alias, /* todo_flags_start */
4422 : : 0, /* todo_flags_finish */
4423 : : };
4424 : :
4425 : : class pass_pre : public gimple_opt_pass
4426 : : {
4427 : : public:
4428 : 287349 : pass_pre (gcc::context *ctxt)
4429 : 574698 : : gimple_opt_pass (pass_data_pre, ctxt)
4430 : : {}
4431 : :
4432 : : /* opt_pass methods: */
4433 : 1027737 : bool gate (function *) final override
4434 : 1027737 : { return flag_tree_pre != 0 || flag_code_hoisting != 0; }
4435 : : unsigned int execute (function *) final override;
4436 : :
4437 : : }; // class pass_pre
4438 : :
4439 : : /* Valueization hook for RPO VN when we are calling back to it
4440 : : at ANTIC compute time. */
4441 : :
4442 : : static tree
4443 : 101693554 : pre_valueize (tree name)
4444 : : {
4445 : 101693554 : if (TREE_CODE (name) == SSA_NAME)
4446 : : {
4447 : 101396283 : tree tem = VN_INFO (name)->valnum;
4448 : 101396283 : if (tem != VN_TOP && tem != name)
4449 : : {
4450 : 14009944 : if (TREE_CODE (tem) != SSA_NAME
4451 : 14009944 : || SSA_NAME_IS_DEFAULT_DEF (tem))
4452 : : return tem;
4453 : : /* We create temporary SSA names for representatives that
4454 : : do not have a definition (yet) but are not default defs either
4455 : : assume they are fine to use. */
4456 : 14005401 : basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (tem));
4457 : 14005401 : if (! def_bb
4458 : 14005401 : || dominated_by_p (CDI_DOMINATORS, vn_context_bb, def_bb))
4459 : 119664 : return tem;
4460 : : /* ??? Now we could look for a leader. Ideally we'd somehow
4461 : : expose RPO VN leaders and get rid of AVAIL_OUT as well... */
4462 : : }
4463 : : }
4464 : : return name;
4465 : : }
4466 : :
4467 : : unsigned int
4468 : 953804 : pass_pre::execute (function *fun)
4469 : : {
4470 : 953804 : unsigned int todo = 0;
4471 : :
4472 : 1907608 : do_partial_partial =
4473 : 953804 : flag_tree_partial_pre && optimize_function_for_speed_p (fun);
4474 : :
4475 : : /* This has to happen before VN runs because
4476 : : loop_optimizer_init may create new phis, etc. */
4477 : 953804 : loop_optimizer_init (LOOPS_NORMAL);
4478 : 953804 : split_edges_for_insertion ();
4479 : 953804 : scev_initialize ();
4480 : 953804 : calculate_dominance_info (CDI_DOMINATORS);
4481 : :
4482 : 953804 : run_rpo_vn (VN_WALK);
4483 : :
4484 : 953804 : init_pre ();
4485 : :
4486 : 953804 : vn_valueize = pre_valueize;
4487 : :
4488 : : /* Insert can get quite slow on an incredibly large number of basic
4489 : : blocks due to some quadratic behavior. Until this behavior is
4490 : : fixed, don't run it when he have an incredibly large number of
4491 : : bb's. If we aren't going to run insert, there is no point in
4492 : : computing ANTIC, either, even though it's plenty fast nor do
4493 : : we require AVAIL. */
4494 : 953804 : if (n_basic_blocks_for_fn (fun) < 4000)
4495 : : {
4496 : 953798 : compute_avail (fun);
4497 : 953798 : compute_antic ();
4498 : 953798 : insert ();
4499 : : }
4500 : :
4501 : : /* Make sure to remove fake edges before committing our inserts.
4502 : : This makes sure we don't end up with extra critical edges that
4503 : : we would need to split. */
4504 : 953804 : remove_fake_exit_edges ();
4505 : 953804 : gsi_commit_edge_inserts ();
4506 : :
4507 : : /* Eliminate folds statements which might (should not...) end up
4508 : : not keeping virtual operands up-to-date. */
4509 : 953804 : gcc_assert (!need_ssa_update_p (fun));
4510 : :
4511 : 953804 : statistics_counter_event (fun, "Insertions", pre_stats.insertions);
4512 : 953804 : statistics_counter_event (fun, "PA inserted", pre_stats.pa_insert);
4513 : 953804 : statistics_counter_event (fun, "HOIST inserted", pre_stats.hoist_insert);
4514 : 953804 : statistics_counter_event (fun, "New PHIs", pre_stats.phis);
4515 : :
4516 : 953804 : todo |= eliminate_with_rpo_vn (inserted_exprs);
4517 : :
4518 : 953804 : vn_valueize = NULL;
4519 : :
4520 : 953804 : fini_pre ();
4521 : :
4522 : 953804 : scev_finalize ();
4523 : 953804 : loop_optimizer_finalize ();
4524 : :
4525 : : /* Perform a CFG cleanup before we run simple_dce_from_worklist since
4526 : : unreachable code regions will have not up-to-date SSA form which
4527 : : confuses it. */
4528 : 953804 : bool need_crit_edge_split = false;
4529 : 953804 : if (todo & TODO_cleanup_cfg)
4530 : : {
4531 : 139803 : cleanup_tree_cfg ();
4532 : 139803 : need_crit_edge_split = true;
4533 : : }
4534 : :
4535 : : /* Because we don't follow exactly the standard PRE algorithm, and decide not
4536 : : to insert PHI nodes sometimes, and because value numbering of casts isn't
4537 : : perfect, we sometimes end up inserting dead code. This simple DCE-like
4538 : : pass removes any insertions we made that weren't actually used. */
4539 : 953804 : simple_dce_from_worklist (inserted_exprs);
4540 : 953804 : BITMAP_FREE (inserted_exprs);
4541 : :
4542 : : /* TODO: tail_merge_optimize may merge all predecessors of a block, in which
4543 : : case we can merge the block with the remaining predecessor of the block.
4544 : : It should either:
4545 : : - call merge_blocks after each tail merge iteration
4546 : : - call merge_blocks after all tail merge iterations
4547 : : - mark TODO_cleanup_cfg when necessary. */
4548 : 953804 : todo |= tail_merge_optimize (need_crit_edge_split);
4549 : :
4550 : 953804 : free_rpo_vn ();
4551 : :
4552 : : /* Tail merging invalidates the virtual SSA web, together with
4553 : : cfg-cleanup opportunities exposed by PRE this will wreck the
4554 : : SSA updating machinery. So make sure to run update-ssa
4555 : : manually, before eventually scheduling cfg-cleanup as part of
4556 : : the todo. */
4557 : 953804 : update_ssa (TODO_update_ssa_only_virtuals);
4558 : :
4559 : 953804 : return todo;
4560 : : }
4561 : :
4562 : : } // anon namespace
4563 : :
4564 : : gimple_opt_pass *
4565 : 287349 : make_pass_pre (gcc::context *ctxt)
4566 : : {
4567 : 287349 : return new pass_pre (ctxt);
4568 : : }
|
