Line data Source code
1 : /* Loop invariant motion.
2 : Copyright (C) 2003-2026 Free Software Foundation, Inc.
3 :
4 : This file is part of GCC.
5 :
6 : GCC is free software; you can redistribute it and/or modify it
7 : under the terms of the GNU General Public License as published by the
8 : Free Software Foundation; either version 3, or (at your option) any
9 : later version.
10 :
11 : GCC is distributed in the hope that it will be useful, but WITHOUT
12 : ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 : FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 : for more details.
15 :
16 : You should have received a copy of the GNU General Public License
17 : along with GCC; see the file COPYING3. If not see
18 : <http://www.gnu.org/licenses/>. */
19 :
20 : #include "config.h"
21 : #include "system.h"
22 : #include "coretypes.h"
23 : #include "backend.h"
24 : #include "tree.h"
25 : #include "gimple.h"
26 : #include "cfghooks.h"
27 : #include "tree-pass.h"
28 : #include "ssa.h"
29 : #include "gimple-pretty-print.h"
30 : #include "fold-const.h"
31 : #include "cfganal.h"
32 : #include "tree-eh.h"
33 : #include "gimple-iterator.h"
34 : #include "tree-cfg.h"
35 : #include "tree-ssa-loop-manip.h"
36 : #include "tree-ssa-loop.h"
37 : #include "tree-into-ssa.h"
38 : #include "cfgloop.h"
39 : #include "tree-affine.h"
40 : #include "tree-ssa-propagate.h"
41 : #include "trans-mem.h"
42 : #include "gimple-fold.h"
43 : #include "tree-scalar-evolution.h"
44 : #include "tree-ssa-loop-niter.h"
45 : #include "alias.h"
46 : #include "builtins.h"
47 : #include "tree-dfa.h"
48 : #include "tree-ssa.h"
49 : #include "dbgcnt.h"
50 : #include "insn-codes.h"
51 : #include "optabs-tree.h"
52 :
53 : /* TODO: Support for predicated code motion. I.e.
54 :
55 : while (1)
56 : {
57 : if (cond)
58 : {
59 : a = inv;
60 : something;
61 : }
62 : }
63 :
64 : Where COND and INV are invariants, but evaluating INV may trap or be
65 : invalid from some other reason if !COND. This may be transformed to
66 :
67 : if (cond)
68 : a = inv;
69 : while (1)
70 : {
71 : if (cond)
72 : something;
73 : } */
74 :
75 : /* The auxiliary data kept for each statement. */
76 :
77 : struct lim_aux_data
78 : {
79 : class loop *max_loop; /* The outermost loop in that the statement
80 : is invariant. */
81 :
82 : class loop *tgt_loop; /* The loop out of that we want to move the
83 : invariant. */
84 :
85 : class loop *always_executed_in;
86 : /* The outermost loop for that we are sure
87 : the statement is executed if the loop
88 : is entered. */
89 :
90 : unsigned cost; /* Cost of the computation performed by the
91 : statement. */
92 :
93 : unsigned ref; /* The simple_mem_ref in this stmt or 0. */
94 :
95 : vec<gimple *> depends; /* Vector of statements that must be also
96 : hoisted out of the loop when this statement
97 : is hoisted; i.e. those that define the
98 : operands of the statement and are inside of
99 : the MAX_LOOP loop. */
100 : };
101 :
102 : /* Maps statements to their lim_aux_data. */
103 :
104 : static hash_map<gimple *, lim_aux_data *> *lim_aux_data_map;
105 :
106 : /* Description of a memory reference location. */
107 :
108 : struct mem_ref_loc
109 : {
110 : tree *ref; /* The reference itself. */
111 : gimple *stmt; /* The statement in that it occurs. */
112 : };
113 :
114 :
115 : /* Description of a memory reference. */
116 :
117 : class im_mem_ref
118 : {
119 : public:
120 : unsigned id : 30; /* ID assigned to the memory reference
121 : (its index in memory_accesses.refs_list) */
122 : unsigned ref_canonical : 1; /* Whether mem.ref was canonicalized. */
123 : unsigned ref_decomposed : 1; /* Whether the ref was hashed from mem. */
124 : hashval_t hash; /* Its hash value. */
125 :
126 : /* The memory access itself and associated caching of alias-oracle
127 : query meta-data. We are using mem.ref == error_mark_node for the
128 : case the reference is represented by its single access stmt
129 : in accesses_in_loop[0]. */
130 : ao_ref mem;
131 :
132 : bitmap stored; /* The set of loops in that this memory location
133 : is stored to. */
134 : bitmap loaded; /* The set of loops in that this memory location
135 : is loaded from. */
136 : vec<mem_ref_loc> accesses_in_loop;
137 : /* The locations of the accesses. */
138 :
139 : /* The following set is computed on demand. */
140 : bitmap_head dep_loop; /* The set of loops in that the memory
141 : reference is {in,}dependent in
142 : different modes. */
143 : };
144 :
145 : static bool in_loop_pipeline;
146 :
147 : /* We use six bits per loop in the ref->dep_loop bitmap to record
148 : the dep_kind x dep_state combinations. */
149 :
150 : enum dep_kind { lim_raw, sm_war, sm_waw };
151 : enum dep_state { dep_unknown, dep_independent, dep_dependent };
152 :
153 : /* coldest outermost loop for given loop. */
154 : vec<class loop *> coldest_outermost_loop;
155 : /* hotter outer loop nearest to given loop. */
156 : vec<class loop *> hotter_than_inner_loop;
157 :
158 : /* Populate the loop dependence cache of REF for LOOP, KIND with STATE. */
159 :
160 : static void
161 2101414 : record_loop_dependence (class loop *loop, im_mem_ref *ref,
162 : dep_kind kind, dep_state state)
163 : {
164 2101414 : gcc_assert (state != dep_unknown);
165 2101414 : unsigned bit = 6 * loop->num + kind * 2 + state == dep_dependent ? 1 : 0;
166 2101414 : bitmap_set_bit (&ref->dep_loop, bit);
167 2101414 : }
168 :
169 : /* Query the loop dependence cache of REF for LOOP, KIND. */
170 :
171 : static dep_state
172 664427 : query_loop_dependence (class loop *loop, im_mem_ref *ref, dep_kind kind)
173 : {
174 664427 : unsigned first_bit = 6 * loop->num + kind * 2;
175 664427 : if (bitmap_bit_p (&ref->dep_loop, first_bit))
176 : return dep_independent;
177 664427 : else if (bitmap_bit_p (&ref->dep_loop, first_bit + 1))
178 0 : return dep_dependent;
179 : return dep_unknown;
180 : }
181 :
182 : /* Mem_ref hashtable helpers. */
183 :
184 : struct mem_ref_hasher : nofree_ptr_hash <im_mem_ref>
185 : {
186 : typedef ao_ref *compare_type;
187 : static inline hashval_t hash (const im_mem_ref *);
188 : static inline bool equal (const im_mem_ref *, const ao_ref *);
189 : };
190 :
191 : /* A hash function for class im_mem_ref object OBJ. */
192 :
193 : inline hashval_t
194 11527668 : mem_ref_hasher::hash (const im_mem_ref *mem)
195 : {
196 11527668 : return mem->hash;
197 : }
198 :
199 : /* An equality function for class im_mem_ref object MEM1 with
200 : memory reference OBJ2. */
201 :
202 : inline bool
203 13377824 : mem_ref_hasher::equal (const im_mem_ref *mem1, const ao_ref *obj2)
204 : {
205 13377824 : if (!obj2->max_size_known_p ())
206 1465896 : return operand_equal_p (mem1->mem.ref, obj2->ref, 0);
207 :
208 11911928 : if (!mem1->ref_decomposed)
209 : return false;
210 :
211 : /* obj2 is now known decomposable and the hash is based on offset,
212 : size and base, matching gather_mem_refs_stmt. */
213 11378211 : bool refs_equal
214 11378211 : = ((TREE_CODE (mem1->mem.base) == MEM_REF
215 5044105 : && TREE_CODE (obj2->base) == MEM_REF
216 3361961 : && operand_equal_p (TREE_OPERAND (mem1->mem.base, 0),
217 3361961 : TREE_OPERAND (obj2->base, 0), 0)
218 2000133 : && known_eq (mem_ref_offset (mem1->mem.base) * BITS_PER_UNIT
219 : + mem1->mem.offset,
220 : mem_ref_offset (obj2->base) * BITS_PER_UNIT
221 : + obj2->offset))
222 15962025 : || (operand_equal_p (mem1->mem.base, obj2->base, 0)
223 878914 : && known_eq (mem1->mem.offset, obj2->offset)));
224 :
225 11378211 : return (refs_equal
226 959515 : && known_eq (mem1->mem.size, obj2->size)
227 957175 : && known_eq (mem1->mem.max_size, obj2->max_size)
228 957175 : && mem1->mem.volatile_p == obj2->volatile_p
229 957159 : && (mem1->mem.ref_alias_set == obj2->ref_alias_set
230 : /* We are not canonicalizing alias-sets but for the
231 : special-case we didn't canonicalize yet and the
232 : incoming ref is a alias-set zero MEM we pick
233 : the correct one already. */
234 15469 : || (!mem1->ref_canonical
235 13877 : && (TREE_CODE (obj2->ref) == MEM_REF
236 13877 : || TREE_CODE (obj2->ref) == TARGET_MEM_REF)
237 8769 : && obj2->ref_alias_set == 0)
238 : /* Likewise if there's a canonical ref with alias-set zero. */
239 12392 : || (mem1->ref_canonical && mem1->mem.ref_alias_set == 0))
240 945653 : && types_compatible_p (TREE_TYPE (mem1->mem.ref),
241 945653 : TREE_TYPE (obj2->ref)));
242 : }
243 :
244 :
245 : /* Description of memory accesses in loops. */
246 :
247 : static struct
248 : {
249 : /* The hash table of memory references accessed in loops. */
250 : hash_table<mem_ref_hasher> *refs;
251 :
252 : /* The list of memory references. */
253 : vec<im_mem_ref *> refs_list;
254 :
255 : /* The set of memory references accessed in each loop. */
256 : vec<bitmap_head> refs_loaded_in_loop;
257 :
258 : /* The set of memory references stored in each loop. */
259 : vec<bitmap_head> refs_stored_in_loop;
260 :
261 : /* The set of memory references stored in each loop, including subloops . */
262 : vec<bitmap_head> all_refs_stored_in_loop;
263 :
264 : /* Cache for expanding memory addresses. */
265 : hash_map<tree, name_expansion *> *ttae_cache;
266 : } memory_accesses;
267 :
268 : /* Obstack for the bitmaps in the above data structures. */
269 : static bitmap_obstack lim_bitmap_obstack;
270 : static obstack mem_ref_obstack;
271 :
272 : static bool ref_indep_loop_p (class loop *, im_mem_ref *, dep_kind);
273 : static bool ref_always_stored_p (class loop *, im_mem_ref *);
274 : static bool refs_independent_p (im_mem_ref *, im_mem_ref *, bool = true);
275 :
276 : /* Minimum cost of an expensive expression. */
277 : #define LIM_EXPENSIVE ((unsigned) param_lim_expensive)
278 :
279 : /* The outermost loop for which execution of the header guarantees that the
280 : block will be executed. */
281 : #define ALWAYS_EXECUTED_IN(BB) ((class loop *) (BB)->aux)
282 : #define SET_ALWAYS_EXECUTED_IN(BB, VAL) ((BB)->aux = (void *) (VAL))
283 :
284 : /* ID of the shared unanalyzable mem. */
285 : #define UNANALYZABLE_MEM_ID 0
286 :
287 : /* Whether the reference was analyzable. */
288 : #define MEM_ANALYZABLE(REF) ((REF)->id != UNANALYZABLE_MEM_ID)
289 :
290 : static struct lim_aux_data *
291 18244502 : init_lim_data (gimple *stmt)
292 : {
293 18244502 : lim_aux_data *p = XCNEW (struct lim_aux_data);
294 18244502 : lim_aux_data_map->put (stmt, p);
295 :
296 18244502 : return p;
297 : }
298 :
299 : static struct lim_aux_data *
300 100119548 : get_lim_data (gimple *stmt)
301 : {
302 100119548 : lim_aux_data **p = lim_aux_data_map->get (stmt);
303 100119548 : if (!p)
304 : return NULL;
305 :
306 49717300 : return *p;
307 : }
308 :
309 : /* Releases the memory occupied by DATA. */
310 :
311 : static void
312 18244319 : free_lim_aux_data (struct lim_aux_data *data)
313 : {
314 0 : data->depends.release ();
315 18244319 : free (data);
316 0 : }
317 :
318 : static void
319 18244319 : clear_lim_data (gimple *stmt)
320 : {
321 18244319 : lim_aux_data **p = lim_aux_data_map->get (stmt);
322 18244319 : if (!p)
323 : return;
324 :
325 18244319 : free_lim_aux_data (*p);
326 18244319 : *p = NULL;
327 : }
328 :
329 :
330 : /* The possibilities of statement movement. */
331 : enum move_pos
332 : {
333 : MOVE_IMPOSSIBLE, /* No movement -- side effect expression. */
334 : MOVE_PRESERVE_EXECUTION, /* Must not cause the non-executed statement
335 : become executed -- memory accesses, ... */
336 : MOVE_POSSIBLE /* Unlimited movement. */
337 : };
338 :
339 :
340 : /* If it is possible to hoist the statement STMT unconditionally,
341 : returns MOVE_POSSIBLE.
342 : If it is possible to hoist the statement STMT, but we must avoid making
343 : it executed if it would not be executed in the original program (e.g.
344 : because it may trap), return MOVE_PRESERVE_EXECUTION.
345 : Otherwise return MOVE_IMPOSSIBLE. */
346 :
347 : static enum move_pos
348 50883709 : movement_possibility_1 (gimple *stmt)
349 : {
350 50883709 : tree lhs;
351 50883709 : enum move_pos ret = MOVE_POSSIBLE;
352 :
353 50883709 : if (flag_unswitch_loops
354 50883709 : && gimple_code (stmt) == GIMPLE_COND)
355 : {
356 : /* If we perform unswitching, force the operands of the invariant
357 : condition to be moved out of the loop. */
358 : return MOVE_POSSIBLE;
359 : }
360 :
361 50510320 : if (gimple_code (stmt) == GIMPLE_PHI
362 2678121 : && gimple_phi_num_args (stmt) <= 2
363 4991510 : && !virtual_operand_p (gimple_phi_result (stmt))
364 51994429 : && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_phi_result (stmt)))
365 : return MOVE_POSSIBLE;
366 :
367 49026473 : if (gimple_get_lhs (stmt) == NULL_TREE)
368 : return MOVE_IMPOSSIBLE;
369 :
370 32465706 : if (gimple_vdef (stmt))
371 : return MOVE_IMPOSSIBLE;
372 :
373 13707073 : if (stmt_ends_bb_p (stmt)
374 12485060 : || gimple_has_volatile_ops (stmt)
375 13651604 : || gimple_has_side_effects (stmt)
376 27352775 : || stmt_could_throw_p (cfun, stmt))
377 : return MOVE_IMPOSSIBLE;
378 :
379 13248553 : if (is_gimple_call (stmt))
380 : {
381 : /* While pure or const call is guaranteed to have no side effects, we
382 : cannot move it arbitrarily. Consider code like
383 :
384 : char *s = something ();
385 :
386 : while (1)
387 : {
388 : if (s)
389 : t = strlen (s);
390 : else
391 : t = 0;
392 : }
393 :
394 : Here the strlen call cannot be moved out of the loop, even though
395 : s is invariant. In addition to possibly creating a call with
396 : invalid arguments, moving out a function call that is not executed
397 : may cause performance regressions in case the call is costly and
398 : not executed at all. */
399 130137 : ret = MOVE_PRESERVE_EXECUTION;
400 130137 : lhs = gimple_call_lhs (stmt);
401 : }
402 13118416 : else if (is_gimple_assign (stmt))
403 11924142 : lhs = gimple_assign_lhs (stmt);
404 : else
405 : return MOVE_IMPOSSIBLE;
406 :
407 12054279 : if (TREE_CODE (lhs) == SSA_NAME
408 12054279 : && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
409 : return MOVE_IMPOSSIBLE;
410 :
411 12054064 : if (TREE_CODE (lhs) != SSA_NAME
412 12054064 : || gimple_could_trap_p (stmt))
413 : return MOVE_PRESERVE_EXECUTION;
414 :
415 9584351 : if (is_gimple_assign (stmt))
416 : {
417 9462400 : auto code = gimple_assign_rhs_code (stmt);
418 9462400 : tree type = TREE_TYPE (gimple_assign_rhs1 (stmt));
419 : /* A shift or rotate by a possibly out-of-bound count moves only where
420 : rewrite_to_defined_unconditional can reduce the count, and not at
421 : all when optimizing for size. */
422 9462400 : if ((code == LSHIFT_EXPR
423 : || code == RSHIFT_EXPR
424 : || code == LROTATE_EXPR
425 9462400 : || code == RROTATE_EXPR)
426 169135 : && (TREE_CODE (gimple_assign_rhs2 (stmt)) != INTEGER_CST
427 : /* We cannot use ranges at 'stmt' here. */
428 126335 : || wi::geu_p (wi::to_wide (gimple_assign_rhs2 (stmt)),
429 9588735 : element_precision (type)))
430 9505201 : && (!gimple_needing_rewrite_undefined (stmt, true)
431 41784 : || optimize_loop_for_size_p (loop_containing_stmt (stmt))))
432 : ret = MOVE_PRESERVE_EXECUTION;
433 : }
434 :
435 : /* Non local loads in a transaction cannot be hoisted out. Well,
436 : unless the load happens on every path out of the loop, but we
437 : don't take this into account yet. */
438 9584351 : if (flag_tm
439 434 : && gimple_in_transaction (stmt)
440 9584454 : && gimple_assign_single_p (stmt))
441 : {
442 21 : tree rhs = gimple_assign_rhs1 (stmt);
443 21 : if (DECL_P (rhs) && is_global_var (rhs))
444 : {
445 18 : if (dump_file)
446 : {
447 2 : fprintf (dump_file, "Cannot hoist conditional load of ");
448 2 : print_generic_expr (dump_file, rhs, TDF_SLIM);
449 2 : fprintf (dump_file, " because it is in a transaction.\n");
450 : }
451 : return MOVE_IMPOSSIBLE;
452 : }
453 : }
454 :
455 : return ret;
456 : }
457 :
458 : static enum move_pos
459 50883709 : movement_possibility (gimple *stmt)
460 : {
461 50883709 : enum move_pos pos = movement_possibility_1 (stmt);
462 50883709 : if (pos == MOVE_POSSIBLE)
463 : {
464 11317722 : use_operand_p use_p;
465 11317722 : ssa_op_iter ssa_iter;
466 36675894 : FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, ssa_iter, SSA_OP_USE)
467 14060467 : if (TREE_CODE (USE_FROM_PTR (use_p)) == SSA_NAME
468 14060467 : && ssa_name_maybe_undef_p (USE_FROM_PTR (use_p)))
469 20017 : return MOVE_PRESERVE_EXECUTION;
470 : }
471 : return pos;
472 : }
473 :
474 :
475 : /* Compare the profile count inequality of bb and loop's preheader, it is
476 : three-state as stated in profile-count.h, FALSE is returned if inequality
477 : cannot be decided. */
478 : bool
479 14468416 : bb_colder_than_loop_preheader (basic_block bb, class loop *loop)
480 : {
481 14468416 : gcc_assert (bb && loop);
482 14468416 : return bb->count < loop_preheader_edge (loop)->src->count;
483 : }
484 :
485 : /* Check coldest loop between OUTERMOST_LOOP and LOOP by comparing profile
486 : count.
487 : It does three steps check:
488 : 1) Check whether CURR_BB is cold in it's own loop_father, if it is cold, just
489 : return NULL which means it should not be moved out at all;
490 : 2) CURR_BB is NOT cold, check if pre-computed COLDEST_LOOP is outside of
491 : OUTERMOST_LOOP, if it is inside of OUTERMOST_LOOP, return the COLDEST_LOOP;
492 : 3) If COLDEST_LOOP is outside of OUTERMOST_LOOP, check whether there is a
493 : hotter loop between OUTERMOST_LOOP and loop in pre-computed
494 : HOTTER_THAN_INNER_LOOP, return it's nested inner loop, otherwise return
495 : OUTERMOST_LOOP.
496 : At last, the coldest_loop is inside of OUTERMOST_LOOP, just return it as
497 : the hoist target. */
498 :
499 : static class loop *
500 12804667 : get_coldest_out_loop (class loop *outermost_loop, class loop *loop,
501 : basic_block curr_bb)
502 : {
503 12804667 : gcc_assert (outermost_loop == loop
504 : || flow_loop_nested_p (outermost_loop, loop));
505 :
506 : /* If bb_colder_than_loop_preheader returns false due to three-state
507 : comparison, OUTERMOST_LOOP is returned finally to preserve the behavior.
508 : Otherwise, return the coldest loop between OUTERMOST_LOOP and LOOP. */
509 12804667 : if (curr_bb && bb_colder_than_loop_preheader (curr_bb, loop))
510 : return NULL;
511 :
512 11726598 : class loop *coldest_loop = coldest_outermost_loop[loop->num];
513 23453196 : if (loop_depth (coldest_loop) < loop_depth (outermost_loop))
514 : {
515 229128 : class loop *hotter_loop = hotter_than_inner_loop[loop->num];
516 229128 : if (!hotter_loop
517 231217 : || loop_depth (hotter_loop) < loop_depth (outermost_loop))
518 : return outermost_loop;
519 :
520 : /* hotter_loop is between OUTERMOST_LOOP and LOOP like:
521 : [loop tree root, ..., coldest_loop, ..., outermost_loop, ...,
522 : hotter_loop, second_coldest_loop, ..., loop]
523 : return second_coldest_loop to be the hoist target. */
524 38 : class loop *aloop;
525 38 : for (aloop = hotter_loop->inner; aloop; aloop = aloop->next)
526 38 : if (aloop == loop || flow_loop_nested_p (aloop, loop))
527 : return aloop;
528 : }
529 : return coldest_loop;
530 : }
531 :
532 : /* Suppose that operand DEF is used inside the LOOP. Returns the outermost
533 : loop to that we could move the expression using DEF if it did not have
534 : other operands, i.e. the outermost loop enclosing LOOP in that the value
535 : of DEF is invariant. */
536 :
537 : static class loop *
538 20242948 : outermost_invariant_loop (tree def, class loop *loop)
539 : {
540 20242948 : gimple *def_stmt;
541 20242948 : basic_block def_bb;
542 20242948 : class loop *max_loop;
543 20242948 : struct lim_aux_data *lim_data;
544 :
545 20242948 : if (!def)
546 937693 : return superloop_at_depth (loop, 1);
547 :
548 19305255 : if (TREE_CODE (def) != SSA_NAME)
549 : {
550 4032203 : gcc_assert (is_gimple_min_invariant (def));
551 4032203 : return superloop_at_depth (loop, 1);
552 : }
553 :
554 15273052 : def_stmt = SSA_NAME_DEF_STMT (def);
555 15273052 : def_bb = gimple_bb (def_stmt);
556 15273052 : if (!def_bb)
557 76371 : return superloop_at_depth (loop, 1);
558 :
559 15196681 : max_loop = find_common_loop (loop, def_bb->loop_father);
560 :
561 15196681 : lim_data = get_lim_data (def_stmt);
562 15196681 : if (lim_data != NULL && lim_data->max_loop != NULL)
563 729870 : max_loop = find_common_loop (max_loop,
564 : loop_outer (lim_data->max_loop));
565 15196681 : if (max_loop == loop)
566 : return NULL;
567 1915313 : max_loop = superloop_at_depth (loop, loop_depth (max_loop) + 1);
568 :
569 1915313 : return max_loop;
570 : }
571 :
572 : /* DATA is a structure containing information associated with a statement
573 : inside LOOP. DEF is one of the operands of this statement.
574 :
575 : Find the outermost loop enclosing LOOP in that value of DEF is invariant
576 : and record this in DATA->max_loop field. If DEF itself is defined inside
577 : this loop as well (i.e. we need to hoist it out of the loop if we want
578 : to hoist the statement represented by DATA), record the statement in that
579 : DEF is defined to the DATA->depends list. Additionally if ADD_COST is true,
580 : add the cost of the computation of DEF to the DATA->cost.
581 :
582 : If DEF is not invariant in LOOP, return false. Otherwise return TRUE. */
583 :
584 : static bool
585 11920245 : add_dependency (tree def, struct lim_aux_data *data, class loop *loop,
586 : bool add_cost)
587 : {
588 11920245 : gimple *def_stmt = SSA_NAME_DEF_STMT (def);
589 11920245 : basic_block def_bb = gimple_bb (def_stmt);
590 11920245 : class loop *max_loop;
591 11920245 : struct lim_aux_data *def_data;
592 :
593 11920245 : if (!def_bb)
594 : return true;
595 :
596 11550738 : max_loop = outermost_invariant_loop (def, loop);
597 11550738 : if (!max_loop)
598 : return false;
599 :
600 1442678 : if (flow_loop_nested_p (data->max_loop, max_loop))
601 432353 : data->max_loop = max_loop;
602 :
603 1442678 : def_data = get_lim_data (def_stmt);
604 1442678 : if (!def_data)
605 : return true;
606 :
607 718968 : if (add_cost
608 : /* Only add the cost if the statement defining DEF is inside LOOP,
609 : i.e. if it is likely that by moving the invariants dependent
610 : on it, we will be able to avoid creating a new register for
611 : it (since it will be only used in these dependent invariants). */
612 693897 : && def_bb->loop_father == loop)
613 491257 : data->cost += def_data->cost;
614 :
615 718968 : data->depends.safe_push (def_stmt);
616 :
617 718968 : return true;
618 : }
619 :
620 : /* Returns an estimate for a cost of statement STMT. The values here
621 : are just ad-hoc constants, similar to costs for inlining. */
622 :
623 : static unsigned
624 822969 : stmt_cost (gimple *stmt)
625 : {
626 : /* Always try to create possibilities for unswitching. */
627 822969 : if (gimple_code (stmt) == GIMPLE_COND
628 822969 : || gimple_code (stmt) == GIMPLE_PHI)
629 17128 : return LIM_EXPENSIVE;
630 :
631 : /* We should be hoisting calls if possible. */
632 805841 : if (is_gimple_call (stmt))
633 : {
634 886 : tree fndecl;
635 :
636 : /* Unless the call is a builtin_constant_p; this always folds to a
637 : constant, so moving it is useless. */
638 886 : fndecl = gimple_call_fndecl (stmt);
639 886 : if (fndecl && fndecl_built_in_p (fndecl, BUILT_IN_CONSTANT_P))
640 : return 0;
641 :
642 886 : return LIM_EXPENSIVE;
643 : }
644 :
645 : /* Hoisting memory references out should almost surely be a win. */
646 804955 : if (gimple_references_memory_p (stmt))
647 70696 : return LIM_EXPENSIVE;
648 :
649 734259 : if (gimple_code (stmt) != GIMPLE_ASSIGN)
650 : return 1;
651 :
652 734259 : enum tree_code code = gimple_assign_rhs_code (stmt);
653 734259 : switch (code)
654 : {
655 121632 : case MULT_EXPR:
656 121632 : case WIDEN_MULT_EXPR:
657 121632 : case WIDEN_MULT_PLUS_EXPR:
658 121632 : case WIDEN_MULT_MINUS_EXPR:
659 121632 : case DOT_PROD_EXPR:
660 121632 : case TRUNC_DIV_EXPR:
661 121632 : case CEIL_DIV_EXPR:
662 121632 : case FLOOR_DIV_EXPR:
663 121632 : case ROUND_DIV_EXPR:
664 121632 : case EXACT_DIV_EXPR:
665 121632 : case CEIL_MOD_EXPR:
666 121632 : case FLOOR_MOD_EXPR:
667 121632 : case ROUND_MOD_EXPR:
668 121632 : case TRUNC_MOD_EXPR:
669 121632 : case RDIV_EXPR:
670 : /* Division and multiplication are usually expensive. */
671 121632 : return LIM_EXPENSIVE;
672 :
673 9021 : case LSHIFT_EXPR:
674 9021 : case RSHIFT_EXPR:
675 9021 : case WIDEN_LSHIFT_EXPR:
676 9021 : case LROTATE_EXPR:
677 9021 : case RROTATE_EXPR:
678 : /* Shifts and rotates are usually expensive. */
679 9021 : return LIM_EXPENSIVE;
680 :
681 483 : case COND_EXPR:
682 483 : case VEC_COND_EXPR:
683 : /* Conditionals are expensive. */
684 483 : return LIM_EXPENSIVE;
685 :
686 1477 : case CONSTRUCTOR:
687 : /* Make vector construction cost proportional to the number
688 : of elements. */
689 1477 : return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt));
690 :
691 : case SSA_NAME:
692 : case PAREN_EXPR:
693 : /* Whether or not something is wrapped inside a PAREN_EXPR
694 : should not change move cost. Nor should an intermediate
695 : unpropagated SSA name copy. */
696 : return 0;
697 :
698 565366 : default:
699 : /* Comparisons are usually expensive. */
700 565366 : if (TREE_CODE_CLASS (code) == tcc_comparison)
701 8521 : return LIM_EXPENSIVE;
702 : return 1;
703 : }
704 : }
705 :
706 : /* Finds the outermost loop between OUTER and LOOP in that the memory reference
707 : REF is independent. If REF is not independent in LOOP, NULL is returned
708 : instead. */
709 :
710 : static class loop *
711 810005 : outermost_indep_loop (class loop *outer, class loop *loop, im_mem_ref *ref)
712 : {
713 810005 : class loop *aloop;
714 :
715 810005 : if (ref->stored && bitmap_bit_p (ref->stored, loop->num))
716 : return NULL;
717 :
718 85746 : for (aloop = outer;
719 726554 : aloop != loop;
720 85746 : aloop = superloop_at_depth (loop, loop_depth (aloop) + 1))
721 10403 : if ((!ref->stored || !bitmap_bit_p (ref->stored, aloop->num))
722 96010 : && ref_indep_loop_p (aloop, ref, lim_raw))
723 : return aloop;
724 :
725 632380 : if (ref_indep_loop_p (loop, ref, lim_raw))
726 : return loop;
727 : else
728 570112 : return NULL;
729 : }
730 :
731 : /* If there is a simple load or store to a memory reference in STMT, returns
732 : the location of the memory reference, and sets IS_STORE according to whether
733 : it is a store or load. Otherwise, returns NULL. */
734 :
735 : static tree *
736 7271046 : simple_mem_ref_in_stmt (gimple *stmt, bool *is_store)
737 : {
738 7271046 : tree *lhs, *rhs;
739 :
740 : /* Recognize SSA_NAME = MEM and MEM = (SSA_NAME | invariant) patterns. */
741 7271046 : if (!gimple_assign_single_p (stmt))
742 : return NULL;
743 :
744 5956541 : lhs = gimple_assign_lhs_ptr (stmt);
745 5956541 : rhs = gimple_assign_rhs1_ptr (stmt);
746 :
747 9298107 : if (TREE_CODE (*lhs) == SSA_NAME && gimple_vuse (stmt))
748 : {
749 3341566 : *is_store = false;
750 3341566 : return rhs;
751 : }
752 2614975 : else if (gimple_vdef (stmt)
753 2614975 : && (TREE_CODE (*rhs) == SSA_NAME || is_gimple_min_invariant (*rhs)))
754 : {
755 2140650 : *is_store = true;
756 2140650 : return lhs;
757 : }
758 : else
759 : return NULL;
760 : }
761 :
762 : /* From a controlling predicate in DOM determine the arguments from
763 : the PHI node PHI that are chosen if the predicate evaluates to
764 : true and false and store them to *TRUE_ARG_P and *FALSE_ARG_P if
765 : they are non-NULL. Returns true if the arguments can be determined,
766 : else return false. */
767 :
768 : static bool
769 13801 : extract_true_false_args_from_phi (basic_block dom, gphi *phi,
770 : tree *true_arg_p, tree *false_arg_p)
771 : {
772 13801 : edge te, fe;
773 13801 : if (! extract_true_false_controlled_edges (dom, gimple_bb (phi),
774 : &te, &fe))
775 : return false;
776 :
777 12904 : if (true_arg_p)
778 1417 : *true_arg_p = PHI_ARG_DEF (phi, te->dest_idx);
779 12904 : if (false_arg_p)
780 1417 : *false_arg_p = PHI_ARG_DEF (phi, fe->dest_idx);
781 :
782 : return true;
783 : }
784 :
785 : /* Return true if STMT is a shift or rotate whose count would have to be
786 : masked when moved out of LOOP, because STMT is not always executed
787 : when LOOP is entered. An out-of-range constant count folds instead. */
788 :
789 : static bool
790 226970 : shift_needs_mask_p (gimple *stmt, class loop *loop)
791 : {
792 226970 : if (!gimple_needing_rewrite_undefined (stmt, true))
793 : return false;
794 25457 : enum tree_code code = gimple_assign_rhs_code (stmt);
795 25457 : if ((code != LSHIFT_EXPR
796 : && code != RSHIFT_EXPR
797 : && code != LROTATE_EXPR
798 25457 : && code != RROTATE_EXPR)
799 25457 : || TREE_CODE (gimple_assign_rhs2 (stmt)) == INTEGER_CST)
800 : return false;
801 765 : class loop *always = get_lim_data (stmt)->always_executed_in;
802 765 : return !always || !(always == loop || flow_loop_nested_p (always, loop));
803 : }
804 :
805 : /* Determine the outermost loop to that it is possible to hoist a statement
806 : STMT and store it to LIM_DATA (STMT)->max_loop. To do this we determine
807 : the outermost loop in that the value computed by STMT is invariant.
808 : If MUST_PRESERVE_EXEC is true, additionally choose such a loop that
809 : we preserve the fact whether STMT is executed. It also fills other related
810 : information to LIM_DATA (STMT).
811 :
812 : The function returns false if STMT cannot be hoisted outside of the loop it
813 : is defined in, and true otherwise. */
814 :
815 : static bool
816 12762142 : determine_max_movement (gimple *stmt, bool must_preserve_exec)
817 : {
818 12762142 : basic_block bb = gimple_bb (stmt);
819 12762142 : class loop *loop = bb->loop_father;
820 12762142 : class loop *level;
821 12762142 : struct lim_aux_data *lim_data = get_lim_data (stmt);
822 12762142 : tree val;
823 12762142 : ssa_op_iter iter;
824 :
825 12762142 : if (must_preserve_exec)
826 1458407 : level = ALWAYS_EXECUTED_IN (bb);
827 : else
828 11303735 : level = superloop_at_depth (loop, 1);
829 12762142 : lim_data->max_loop = get_coldest_out_loop (level, loop, bb);
830 12762142 : if (!lim_data->max_loop)
831 : return false;
832 :
833 11685752 : if (gphi *phi = dyn_cast <gphi *> (stmt))
834 : {
835 1426704 : use_operand_p use_p;
836 1426704 : unsigned min_cost = UINT_MAX;
837 1426704 : unsigned total_cost = 0;
838 1426704 : struct lim_aux_data *def_data;
839 :
840 : /* We will end up promoting dependencies to be unconditionally
841 : evaluated. For this reason the PHI cost (and thus the
842 : cost we remove from the loop by doing the invariant motion)
843 : is that of the cheapest PHI argument dependency chain. */
844 1641617 : FOR_EACH_PHI_ARG (use_p, phi, iter, SSA_OP_USE)
845 : {
846 1615765 : val = USE_FROM_PTR (use_p);
847 :
848 1615765 : if (TREE_CODE (val) != SSA_NAME)
849 : {
850 : /* Assign const 1 to constants. */
851 132117 : min_cost = MIN (min_cost, 1);
852 132117 : total_cost += 1;
853 132117 : continue;
854 : }
855 1483648 : if (!add_dependency (val, lim_data, loop, false))
856 : return false;
857 :
858 82796 : gimple *def_stmt = SSA_NAME_DEF_STMT (val);
859 82796 : if (gimple_bb (def_stmt)
860 82796 : && gimple_bb (def_stmt)->loop_father == loop)
861 : {
862 2996 : def_data = get_lim_data (def_stmt);
863 2996 : if (def_data)
864 : {
865 2996 : min_cost = MIN (min_cost, def_data->cost);
866 2996 : total_cost += def_data->cost;
867 : }
868 : }
869 : }
870 :
871 25852 : min_cost = MIN (min_cost, total_cost);
872 25852 : lim_data->cost += min_cost;
873 :
874 25852 : if (gimple_phi_num_args (phi) > 1)
875 : {
876 25599 : basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb);
877 25599 : gimple *cond;
878 51198 : if (gsi_end_p (gsi_last_bb (dom)))
879 : return false;
880 17483 : cond = gsi_stmt (gsi_last_bb (dom));
881 17483 : if (gimple_code (cond) != GIMPLE_COND)
882 : return false;
883 : /* Verify that this is an extended form of a diamond and
884 : the PHI arguments are completely controlled by the
885 : predicate in DOM. */
886 12384 : if (!extract_true_false_args_from_phi (dom, phi, NULL, NULL))
887 : return false;
888 :
889 : /* Check if one of the dependent statement is a vector compare whether
890 : the target supports it, otherwise it's invalid to hoist it out of
891 : the gcond it belonged to. */
892 11487 : if (VECTOR_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond))))
893 : {
894 5 : tree type = TREE_TYPE (gimple_cond_lhs (cond));
895 5 : auto code = gimple_cond_code (cond);
896 5 : if (!target_supports_op_p (type, code, optab_vector))
897 : return false;
898 : }
899 :
900 : /* Fold in dependencies and cost of the condition. */
901 14709 : FOR_EACH_SSA_TREE_OPERAND (val, cond, iter, SSA_OP_USE)
902 : {
903 12250 : if (!add_dependency (val, lim_data, loop, false))
904 : return false;
905 3227 : def_data = get_lim_data (SSA_NAME_DEF_STMT (val));
906 3227 : if (def_data)
907 1917 : lim_data->cost += def_data->cost;
908 : }
909 :
910 : /* We want to avoid unconditionally executing very expensive
911 : operations. As costs for our dependencies cannot be
912 : negative just claim we are not invariand for this case.
913 : We also are not sure whether the control-flow inside the
914 : loop will vanish. */
915 2459 : if (total_cost - min_cost >= 2 * LIM_EXPENSIVE
916 1088 : && !(min_cost != 0
917 1088 : && total_cost / min_cost <= 2))
918 : return false;
919 :
920 : /* Assume that the control-flow in the loop will vanish.
921 : ??? We should verify this and not artificially increase
922 : the cost if that is not the case. */
923 1417 : lim_data->cost += stmt_cost (stmt);
924 : }
925 :
926 : return true;
927 : }
928 :
929 : /* A stmt that receives abnormal edges cannot be hoisted. */
930 10259048 : if (is_a <gcall *> (stmt)
931 10259048 : && (gimple_call_flags (stmt) & ECF_RETURNS_TWICE))
932 : return false;
933 :
934 11985183 : FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_USE)
935 10423242 : if (!add_dependency (val, lim_data, loop, true))
936 : return false;
937 :
938 3108171 : if (gimple_vuse (stmt))
939 : {
940 811110 : im_mem_ref *ref
941 811110 : = lim_data ? memory_accesses.refs_list[lim_data->ref] : NULL;
942 811110 : if (ref
943 811110 : && MEM_ANALYZABLE (ref))
944 : {
945 810005 : lim_data->max_loop = outermost_indep_loop (lim_data->max_loop,
946 : loop, ref);
947 810005 : if (!lim_data->max_loop)
948 : return false;
949 : }
950 1105 : else if (! add_dependency (gimple_vuse (stmt), lim_data, loop, false))
951 : return false;
952 : }
953 :
954 : /* A shift that needs its count masked is not worth moving on its own. */
955 821552 : unsigned cost = stmt_cost (stmt);
956 821552 : if (cost >= LIM_EXPENSIVE && shift_needs_mask_p (stmt, lim_data->max_loop))
957 : cost = 1;
958 821552 : lim_data->cost += cost;
959 :
960 821552 : return true;
961 : }
962 :
963 : /* Suppose that some statement in ORIG_LOOP is hoisted to the loop LEVEL,
964 : and that one of the operands of this statement is computed by STMT.
965 : Ensure that STMT (together with all the statements that define its
966 : operands) is hoisted at least out of the loop LEVEL. */
967 :
968 : static void
969 760057 : set_level (gimple *stmt, class loop *orig_loop, class loop *level)
970 : {
971 760057 : class loop *stmt_loop = gimple_bb (stmt)->loop_father;
972 760057 : struct lim_aux_data *lim_data;
973 760057 : gimple *dep_stmt;
974 760057 : unsigned i;
975 :
976 760057 : stmt_loop = find_common_loop (orig_loop, stmt_loop);
977 760057 : lim_data = get_lim_data (stmt);
978 760057 : if (lim_data != NULL && lim_data->tgt_loop != NULL)
979 197071 : stmt_loop = find_common_loop (stmt_loop,
980 : loop_outer (lim_data->tgt_loop));
981 760057 : if (flow_loop_nested_p (stmt_loop, level))
982 760057 : return;
983 :
984 498899 : gcc_assert (level == lim_data->max_loop
985 : || flow_loop_nested_p (lim_data->max_loop, level));
986 :
987 498899 : lim_data->tgt_loop = level;
988 867745 : FOR_EACH_VEC_ELT (lim_data->depends, i, dep_stmt)
989 368846 : set_level (dep_stmt, orig_loop, level);
990 : }
991 :
992 : /* Determines an outermost loop from that we want to hoist the statement STMT.
993 : For now we chose the outermost possible loop. TODO -- use profiling
994 : information to set it more sanely. */
995 :
996 : static void
997 327771 : set_profitable_level (gimple *stmt)
998 : {
999 327771 : set_level (stmt, gimple_bb (stmt)->loop_father, get_lim_data (stmt)->max_loop);
1000 327771 : }
1001 :
1002 : /* Returns true if STMT is a call that has side effects. */
1003 :
1004 : static bool
1005 78875322 : nonpure_call_p (gimple *stmt)
1006 : {
1007 0 : if (gimple_code (stmt) != GIMPLE_CALL)
1008 : return false;
1009 :
1010 2371116 : return gimple_has_side_effects (stmt);
1011 : }
1012 :
1013 : /* Rewrite a/b to a*(1/b). Return the invariant stmt to process. */
1014 :
1015 : static gimple *
1016 35 : rewrite_reciprocal (gimple_stmt_iterator *bsi)
1017 : {
1018 35 : gassign *stmt, *stmt1, *stmt2;
1019 35 : tree name, lhs, type;
1020 35 : tree real_one;
1021 35 : gimple_stmt_iterator gsi;
1022 :
1023 35 : stmt = as_a <gassign *> (gsi_stmt (*bsi));
1024 35 : lhs = gimple_assign_lhs (stmt);
1025 35 : type = TREE_TYPE (lhs);
1026 :
1027 35 : real_one = build_one_cst (type);
1028 :
1029 35 : name = make_temp_ssa_name (type, NULL, "reciptmp");
1030 70 : stmt1 = gimple_build_assign (name, RDIV_EXPR, real_one,
1031 : gimple_assign_rhs2 (stmt));
1032 35 : stmt2 = gimple_build_assign (lhs, MULT_EXPR, name,
1033 : gimple_assign_rhs1 (stmt));
1034 :
1035 : /* Replace division stmt with reciprocal and multiply stmts.
1036 : The multiply stmt is not invariant, so update iterator
1037 : and avoid rescanning. */
1038 35 : gsi = *bsi;
1039 35 : gsi_insert_before (bsi, stmt1, GSI_NEW_STMT);
1040 35 : gsi_replace (&gsi, stmt2, true);
1041 :
1042 : /* Continue processing with invariant reciprocal statement. */
1043 35 : return stmt1;
1044 : }
1045 :
1046 : /* Check if the pattern at *BSI is a bittest of the form
1047 : (A >> B) & 1 != 0 and in this case rewrite it to A & (1 << B) != 0. */
1048 :
1049 : static gimple *
1050 9625 : rewrite_bittest (gimple_stmt_iterator *bsi)
1051 : {
1052 9625 : gassign *stmt;
1053 9625 : gimple *stmt1;
1054 9625 : gassign *stmt2;
1055 9625 : gimple *use_stmt;
1056 9625 : gcond *cond_stmt;
1057 9625 : tree lhs, name, t, a, b;
1058 9625 : use_operand_p use;
1059 :
1060 9625 : stmt = as_a <gassign *> (gsi_stmt (*bsi));
1061 9625 : lhs = gimple_assign_lhs (stmt);
1062 :
1063 : /* Verify that the single use of lhs is a comparison against zero. */
1064 9625 : if (TREE_CODE (lhs) != SSA_NAME
1065 9625 : || !single_imm_use (lhs, &use, &use_stmt))
1066 : return stmt;
1067 9173 : cond_stmt = dyn_cast <gcond *> (use_stmt);
1068 5319 : if (!cond_stmt)
1069 : return stmt;
1070 5319 : if (gimple_cond_lhs (cond_stmt) != lhs
1071 5141 : || (gimple_cond_code (cond_stmt) != NE_EXPR
1072 2227 : && gimple_cond_code (cond_stmt) != EQ_EXPR)
1073 10460 : || !integer_zerop (gimple_cond_rhs (cond_stmt)))
1074 : return stmt;
1075 :
1076 : /* Get at the operands of the shift. The rhs is TMP1 & 1. */
1077 5109 : stmt1 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt));
1078 5109 : if (gimple_code (stmt1) != GIMPLE_ASSIGN)
1079 : return stmt;
1080 :
1081 : /* There is a conversion in between possibly inserted by fold. */
1082 5029 : if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt1)))
1083 : {
1084 221 : t = gimple_assign_rhs1 (stmt1);
1085 221 : if (TREE_CODE (t) != SSA_NAME
1086 221 : || !has_single_use (t))
1087 : return stmt;
1088 151 : stmt1 = SSA_NAME_DEF_STMT (t);
1089 151 : if (gimple_code (stmt1) != GIMPLE_ASSIGN)
1090 : return stmt;
1091 : }
1092 :
1093 : /* Verify that B is loop invariant but A is not. Verify that with
1094 : all the stmt walking we are still in the same loop. */
1095 4917 : if (gimple_assign_rhs_code (stmt1) != RSHIFT_EXPR
1096 11357 : || loop_containing_stmt (stmt1) != loop_containing_stmt (stmt))
1097 : return stmt;
1098 :
1099 3220 : a = gimple_assign_rhs1 (stmt1);
1100 3220 : b = gimple_assign_rhs2 (stmt1);
1101 :
1102 3220 : if (outermost_invariant_loop (b, loop_containing_stmt (stmt1)) != NULL
1103 3249 : && outermost_invariant_loop (a, loop_containing_stmt (stmt1)) == NULL)
1104 : {
1105 29 : gimple_stmt_iterator rsi;
1106 :
1107 : /* 1 << B */
1108 29 : t = fold_build2 (LSHIFT_EXPR, TREE_TYPE (a),
1109 : build_int_cst (TREE_TYPE (a), 1), b);
1110 29 : name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp");
1111 29 : stmt1 = gimple_build_assign (name, t);
1112 :
1113 : /* A & (1 << B) */
1114 29 : t = fold_build2 (BIT_AND_EXPR, TREE_TYPE (a), a, name);
1115 29 : name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp");
1116 29 : stmt2 = gimple_build_assign (name, t);
1117 :
1118 : /* Replace the SSA_NAME we compare against zero. Adjust
1119 : the type of zero accordingly. */
1120 29 : SET_USE (use, name);
1121 29 : gimple_cond_set_rhs (cond_stmt,
1122 29 : build_int_cst_type (TREE_TYPE (name),
1123 : 0));
1124 :
1125 : /* Don't use gsi_replace here, none of the new assignments sets
1126 : the variable originally set in stmt. Move bsi to stmt1, and
1127 : then remove the original stmt, so that we get a chance to
1128 : retain debug info for it. */
1129 29 : rsi = *bsi;
1130 29 : gsi_insert_before (bsi, stmt1, GSI_NEW_STMT);
1131 29 : gsi_insert_before (&rsi, stmt2, GSI_SAME_STMT);
1132 29 : gimple *to_release = gsi_stmt (rsi);
1133 29 : gsi_remove (&rsi, true);
1134 29 : release_defs (to_release);
1135 :
1136 29 : return stmt1;
1137 : }
1138 :
1139 : return stmt;
1140 : }
1141 :
1142 : /* Determine the outermost loops in that statements in basic block BB are
1143 : invariant, and record them to the LIM_DATA associated with the
1144 : statements. */
1145 :
1146 : static void
1147 14975448 : compute_invariantness (basic_block bb)
1148 : {
1149 14975448 : enum move_pos pos;
1150 14975448 : gimple_stmt_iterator bsi;
1151 14975448 : gimple *stmt;
1152 14975448 : bool maybe_never = ALWAYS_EXECUTED_IN (bb) == NULL;
1153 14975448 : class loop *outermost = ALWAYS_EXECUTED_IN (bb);
1154 14975448 : struct lim_aux_data *lim_data;
1155 :
1156 14975448 : if (!loop_outer (bb->loop_father))
1157 8687793 : return;
1158 :
1159 6287655 : if (dump_file && (dump_flags & TDF_DETAILS))
1160 585 : fprintf (dump_file, "Basic block %d (loop %d -- depth %d):\n\n",
1161 585 : bb->index, bb->loop_father->num, loop_depth (bb->loop_father));
1162 :
1163 : /* Look at PHI nodes, but only if there is at most two.
1164 : ??? We could relax this further by post-processing the inserted
1165 : code and transforming adjacent cond-exprs with the same predicate
1166 : to control flow again. */
1167 6287655 : bsi = gsi_start_phis (bb);
1168 6287655 : if (!gsi_end_p (bsi)
1169 6287655 : && ((gsi_next (&bsi), gsi_end_p (bsi))
1170 1411184 : || (gsi_next (&bsi), gsi_end_p (bsi))))
1171 4446780 : for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
1172 : {
1173 2678121 : stmt = gsi_stmt (bsi);
1174 :
1175 2678121 : pos = movement_possibility (stmt);
1176 2678121 : if (pos == MOVE_IMPOSSIBLE)
1177 1194274 : continue;
1178 :
1179 1483847 : lim_data = get_lim_data (stmt);
1180 1483847 : if (! lim_data)
1181 1483847 : lim_data = init_lim_data (stmt);
1182 1483847 : lim_data->always_executed_in = outermost;
1183 :
1184 1483847 : if (!determine_max_movement (stmt, false))
1185 : {
1186 1482177 : lim_data->max_loop = NULL;
1187 1482177 : continue;
1188 : }
1189 :
1190 1670 : if (dump_file && (dump_flags & TDF_DETAILS))
1191 : {
1192 2 : print_gimple_stmt (dump_file, stmt, 2);
1193 2 : fprintf (dump_file, " invariant up to level %d, cost %d.\n\n",
1194 2 : loop_depth (lim_data->max_loop),
1195 : lim_data->cost);
1196 : }
1197 :
1198 1670 : if (lim_data->cost >= LIM_EXPENSIVE)
1199 1445 : set_profitable_level (stmt);
1200 : }
1201 :
1202 60780898 : for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
1203 : {
1204 48205588 : stmt = gsi_stmt (bsi);
1205 :
1206 48205588 : pos = movement_possibility (stmt);
1207 48205588 : if (pos == MOVE_IMPOSSIBLE)
1208 : {
1209 37018111 : if (nonpure_call_p (stmt))
1210 : {
1211 : maybe_never = true;
1212 : outermost = NULL;
1213 : }
1214 : /* Make sure to note always_executed_in for stores to make
1215 : store-motion work. */
1216 34543601 : else if (stmt_makes_single_store (stmt))
1217 : {
1218 2603848 : struct lim_aux_data *lim_data = get_lim_data (stmt);
1219 2603848 : if (! lim_data)
1220 0 : lim_data = init_lim_data (stmt);
1221 2603848 : lim_data->always_executed_in = outermost;
1222 : }
1223 35778153 : continue;
1224 35778153 : }
1225 :
1226 12427435 : if (is_gimple_assign (stmt)
1227 12427435 : && (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))
1228 : == GIMPLE_BINARY_RHS))
1229 : {
1230 5959416 : tree op0 = gimple_assign_rhs1 (stmt);
1231 5959416 : tree op1 = gimple_assign_rhs2 (stmt);
1232 11918832 : class loop *ol1 = outermost_invariant_loop (op1,
1233 : loop_containing_stmt (stmt));
1234 :
1235 : /* If divisor is invariant, convert a/b to a*(1/b), allowing reciprocal
1236 : to be hoisted out of loop, saving expensive divide. */
1237 5959416 : if (pos == MOVE_POSSIBLE
1238 5524470 : && gimple_assign_rhs_code (stmt) == RDIV_EXPR
1239 983 : && flag_unsafe_math_optimizations
1240 423 : && !flag_trapping_math
1241 423 : && ol1 != NULL
1242 5959454 : && outermost_invariant_loop (op0, ol1) == NULL)
1243 35 : stmt = rewrite_reciprocal (&bsi);
1244 :
1245 : /* If the shift count is invariant, convert (A >> B) & 1 to
1246 : A & (1 << B) allowing the bit mask to be hoisted out of the loop
1247 : saving an expensive shift. */
1248 5959416 : if (pos == MOVE_POSSIBLE
1249 5524470 : && gimple_assign_rhs_code (stmt) == BIT_AND_EXPR
1250 196171 : && integer_onep (op1)
1251 19709 : && TREE_CODE (op0) == SSA_NAME
1252 5979125 : && has_single_use (op0))
1253 9625 : stmt = rewrite_bittest (&bsi);
1254 : }
1255 :
1256 12427435 : lim_data = get_lim_data (stmt);
1257 12427435 : if (! lim_data)
1258 9448575 : lim_data = init_lim_data (stmt);
1259 12427435 : lim_data->always_executed_in = outermost;
1260 :
1261 12427435 : if (maybe_never && pos == MOVE_PRESERVE_EXECUTION)
1262 1149140 : continue;
1263 :
1264 11278295 : if (!determine_max_movement (stmt, pos == MOVE_PRESERVE_EXECUTION))
1265 : {
1266 10456743 : lim_data->max_loop = NULL;
1267 10456743 : continue;
1268 : }
1269 :
1270 821552 : if (dump_file && (dump_flags & TDF_DETAILS))
1271 : {
1272 318 : print_gimple_stmt (dump_file, stmt, 2);
1273 318 : fprintf (dump_file, " invariant up to level %d, cost %d.\n\n",
1274 318 : loop_depth (lim_data->max_loop),
1275 : lim_data->cost);
1276 : }
1277 :
1278 821552 : if (lim_data->cost >= LIM_EXPENSIVE)
1279 326326 : set_profitable_level (stmt);
1280 : /* When we run before PRE and PRE is active hoist all expressions
1281 : to the always executed loop since PRE would do so anyway
1282 : and we can preserve range info while PRE cannot. */
1283 495226 : else if (flag_tree_pre && !in_loop_pipeline
1284 105065 : && outermost)
1285 : {
1286 60494 : class loop *mloop = lim_data->max_loop;
1287 181482 : if (loop_depth (outermost) > loop_depth (mloop))
1288 : {
1289 6502 : mloop = outermost;
1290 6502 : if (dump_file && (dump_flags & TDF_DETAILS))
1291 1 : fprintf (dump_file, " constraining to loop depth %d\n\n\n",
1292 : loop_depth (mloop));
1293 : }
1294 60494 : set_level (stmt, bb->loop_father, mloop);
1295 : }
1296 : }
1297 : }
1298 :
1299 : /* Hoist the statements in basic block BB out of the loops prescribed by
1300 : data stored in LIM_DATA structures associated with each statement. Callback
1301 : for walk_dominator_tree. */
1302 :
1303 : unsigned int
1304 15014523 : move_computations_worker (basic_block bb)
1305 : {
1306 15014523 : class loop *level;
1307 15014523 : unsigned cost = 0;
1308 15014523 : struct lim_aux_data *lim_data;
1309 15014523 : unsigned int todo = 0;
1310 :
1311 15014523 : if (!loop_outer (bb->loop_father))
1312 : return todo;
1313 :
1314 10974552 : for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); )
1315 : {
1316 4680689 : gassign *new_stmt;
1317 4680689 : gphi *stmt = bsi.phi ();
1318 :
1319 4680689 : lim_data = get_lim_data (stmt);
1320 4680689 : if (lim_data == NULL)
1321 : {
1322 3196842 : gsi_next (&bsi);
1323 3196842 : continue;
1324 : }
1325 :
1326 1483847 : cost = lim_data->cost;
1327 1483847 : level = lim_data->tgt_loop;
1328 1483847 : clear_lim_data (stmt);
1329 :
1330 1483847 : if (!level)
1331 : {
1332 1482392 : gsi_next (&bsi);
1333 1482392 : continue;
1334 : }
1335 :
1336 1455 : if (dump_file && (dump_flags & TDF_DETAILS))
1337 : {
1338 2 : fprintf (dump_file, "Moving PHI node\n");
1339 2 : print_gimple_stmt (dump_file, stmt, 0);
1340 2 : fprintf (dump_file, "(cost %u) out of loop %d.\n\n",
1341 : cost, level->num);
1342 : }
1343 :
1344 1455 : if (gimple_phi_num_args (stmt) == 1)
1345 : {
1346 38 : tree arg = PHI_ARG_DEF (stmt, 0);
1347 38 : new_stmt = gimple_build_assign (gimple_phi_result (stmt),
1348 38 : TREE_CODE (arg), arg);
1349 : }
1350 : else
1351 : {
1352 1417 : basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb);
1353 1417 : gimple *cond = gsi_stmt (gsi_last_bb (dom));
1354 1417 : tree arg0 = NULL_TREE, arg1 = NULL_TREE, t;
1355 : /* Get the PHI arguments corresponding to the true and false
1356 : edges of COND. */
1357 1417 : extract_true_false_args_from_phi (dom, stmt, &arg0, &arg1);
1358 1417 : gcc_assert (arg0 && arg1);
1359 : /* For `bool_val != 0`, reuse bool_val. */
1360 1417 : if (gimple_cond_code (cond) == NE_EXPR
1361 631 : && integer_zerop (gimple_cond_rhs (cond))
1362 1973 : && types_compatible_p (TREE_TYPE (gimple_cond_lhs (cond)),
1363 : boolean_type_node))
1364 : {
1365 54 : t = gimple_cond_lhs (cond);
1366 : }
1367 : else
1368 : {
1369 1363 : t = make_ssa_name (boolean_type_node);
1370 1363 : new_stmt = gimple_build_assign (t, gimple_cond_code (cond),
1371 : gimple_cond_lhs (cond),
1372 : gimple_cond_rhs (cond));
1373 1363 : gsi_insert_on_edge (loop_preheader_edge (level), new_stmt);
1374 : }
1375 1417 : new_stmt = gimple_build_assign (gimple_phi_result (stmt),
1376 : COND_EXPR, t, arg0, arg1);
1377 1417 : todo |= TODO_cleanup_cfg;
1378 : }
1379 1455 : if (!ALWAYS_EXECUTED_IN (bb)
1380 1455 : || (ALWAYS_EXECUTED_IN (bb) != level
1381 100 : && !flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level)))
1382 889 : reset_flow_sensitive_info (gimple_assign_lhs (new_stmt));
1383 1455 : gsi_insert_on_edge (loop_preheader_edge (level), new_stmt);
1384 1455 : remove_phi_node (&bsi, false);
1385 : }
1386 :
1387 60855970 : for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); )
1388 : {
1389 48268244 : edge e;
1390 :
1391 48268244 : gimple *stmt = gsi_stmt (bsi);
1392 :
1393 48268244 : lim_data = get_lim_data (stmt);
1394 48268244 : if (lim_data == NULL)
1395 : {
1396 31507772 : gsi_next (&bsi);
1397 31507772 : continue;
1398 : }
1399 :
1400 16760472 : cost = lim_data->cost;
1401 16760472 : level = lim_data->tgt_loop;
1402 16760472 : clear_lim_data (stmt);
1403 :
1404 16760472 : if (!level)
1405 : {
1406 16223341 : gsi_next (&bsi);
1407 16223341 : continue;
1408 : }
1409 :
1410 : /* We do not really want to move conditionals out of the loop; we just
1411 : placed it here to force its operands to be moved if necessary. */
1412 537131 : if (gimple_code (stmt) == GIMPLE_COND)
1413 : {
1414 15711 : gsi_next (&bsi);
1415 15711 : continue;
1416 : }
1417 :
1418 521420 : if (dump_file && (dump_flags & TDF_DETAILS))
1419 : {
1420 387 : fprintf (dump_file, "Moving statement ");
1421 387 : print_gimple_stmt (dump_file, stmt, 0);
1422 387 : fprintf (dump_file, "(cost %u) out of loop %d.\n\n",
1423 : cost, level->num);
1424 : }
1425 :
1426 521420 : e = loop_preheader_edge (level);
1427 1042840 : gcc_assert (!gimple_vdef (stmt));
1428 1042840 : if (gimple_vuse (stmt))
1429 : {
1430 : /* The new VUSE is the one from the virtual PHI in the loop
1431 : header or the one already present. */
1432 88044 : gphi_iterator gsi2;
1433 88044 : for (gsi2 = gsi_start_phis (e->dest);
1434 205491 : !gsi_end_p (gsi2); gsi_next (&gsi2))
1435 : {
1436 186039 : gphi *phi = gsi2.phi ();
1437 372078 : if (virtual_operand_p (gimple_phi_result (phi)))
1438 : {
1439 68592 : SET_USE (gimple_vuse_op (stmt),
1440 : PHI_ARG_DEF_FROM_EDGE (phi, e));
1441 68592 : break;
1442 : }
1443 : }
1444 : }
1445 521420 : gsi_remove (&bsi, false);
1446 521420 : if (gimple_has_lhs (stmt)
1447 521420 : && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME
1448 481272 : && (!ALWAYS_EXECUTED_IN (bb)
1449 370183 : || !(ALWAYS_EXECUTED_IN (bb) == level
1450 77856 : || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level))))
1451 210985 : reset_flow_sensitive_info (gimple_get_lhs (stmt));
1452 : /* In case this is a stmt that is not unconditionally executed
1453 : when the target loop header is executed and the stmt may
1454 : invoke undefined integer or pointer overflow rewrite it to
1455 : unsigned arithmetic. */
1456 521420 : if (gimple_needing_rewrite_undefined (stmt, true)
1457 521420 : && (!ALWAYS_EXECUTED_IN (bb)
1458 73662 : || !(ALWAYS_EXECUTED_IN (bb) == level
1459 19070 : || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level))))
1460 20283 : gsi_insert_seq_on_edge (e, rewrite_to_defined_unconditional (stmt));
1461 : else
1462 501137 : gsi_insert_on_edge (e, stmt);
1463 : }
1464 :
1465 6293863 : return todo;
1466 : }
1467 :
1468 : /* Checks whether the statement defining variable *INDEX can be hoisted
1469 : out of the loop passed in DATA. Callback for for_each_index. */
1470 :
1471 : static bool
1472 1788963 : may_move_till (tree ref, tree *index, void *data)
1473 : {
1474 1788963 : class loop *loop = (class loop *) data, *max_loop;
1475 :
1476 : /* If REF is an array reference, check also that the step and the lower
1477 : bound is invariant in LOOP. */
1478 1788963 : if (TREE_CODE (ref) == ARRAY_REF)
1479 : {
1480 470272 : tree step = TREE_OPERAND (ref, 3);
1481 470272 : tree lbound = TREE_OPERAND (ref, 2);
1482 :
1483 470272 : max_loop = outermost_invariant_loop (step, loop);
1484 470272 : if (!max_loop)
1485 : return false;
1486 :
1487 470272 : max_loop = outermost_invariant_loop (lbound, loop);
1488 470272 : if (!max_loop)
1489 : return false;
1490 : }
1491 :
1492 1788963 : max_loop = outermost_invariant_loop (*index, loop);
1493 1788963 : if (!max_loop)
1494 1148063 : return false;
1495 :
1496 : return true;
1497 : }
1498 :
1499 : /* If OP is SSA NAME, force the statement that defines it to be
1500 : moved out of the LOOP. ORIG_LOOP is the loop in that EXPR is used. */
1501 :
1502 : static void
1503 36985 : force_move_till_op (tree op, class loop *orig_loop, class loop *loop)
1504 : {
1505 36985 : gimple *stmt;
1506 :
1507 36985 : if (!op
1508 36985 : || is_gimple_min_invariant (op))
1509 : return;
1510 :
1511 4546 : gcc_assert (TREE_CODE (op) == SSA_NAME);
1512 :
1513 4546 : stmt = SSA_NAME_DEF_STMT (op);
1514 4546 : if (gimple_nop_p (stmt))
1515 : return;
1516 :
1517 2946 : set_level (stmt, orig_loop, loop);
1518 : }
1519 :
1520 : /* Forces statement defining invariants in REF (and *INDEX) to be moved out of
1521 : the LOOP. The reference REF is used in the loop ORIG_LOOP. Callback for
1522 : for_each_index. */
1523 :
1524 : struct fmt_data
1525 : {
1526 : class loop *loop;
1527 : class loop *orig_loop;
1528 : };
1529 :
1530 : static bool
1531 19157 : force_move_till (tree ref, tree *index, void *data)
1532 : {
1533 19157 : struct fmt_data *fmt_data = (struct fmt_data *) data;
1534 :
1535 19157 : if (TREE_CODE (ref) == ARRAY_REF)
1536 : {
1537 8914 : tree step = TREE_OPERAND (ref, 3);
1538 8914 : tree lbound = TREE_OPERAND (ref, 2);
1539 :
1540 8914 : force_move_till_op (step, fmt_data->orig_loop, fmt_data->loop);
1541 8914 : force_move_till_op (lbound, fmt_data->orig_loop, fmt_data->loop);
1542 : }
1543 :
1544 19157 : force_move_till_op (*index, fmt_data->orig_loop, fmt_data->loop);
1545 :
1546 19157 : return true;
1547 : }
1548 :
1549 : /* A function to free the mem_ref object OBJ. */
1550 :
1551 : static void
1552 5444730 : memref_free (class im_mem_ref *mem)
1553 : {
1554 0 : mem->accesses_in_loop.release ();
1555 0 : }
1556 :
1557 : /* Allocates and returns a memory reference description for MEM whose hash
1558 : value is HASH and id is ID. */
1559 :
1560 : static im_mem_ref *
1561 5444730 : mem_ref_alloc (ao_ref *mem, unsigned hash, unsigned id)
1562 : {
1563 5444730 : im_mem_ref *ref = XOBNEW (&mem_ref_obstack, class im_mem_ref);
1564 5444730 : if (mem)
1565 4472163 : ref->mem = *mem;
1566 : else
1567 972567 : ao_ref_init (&ref->mem, error_mark_node);
1568 5444730 : ref->id = id;
1569 5444730 : ref->ref_canonical = false;
1570 5444730 : ref->ref_decomposed = false;
1571 5444730 : ref->hash = hash;
1572 5444730 : ref->stored = NULL;
1573 5444730 : ref->loaded = NULL;
1574 5444730 : bitmap_initialize (&ref->dep_loop, &lim_bitmap_obstack);
1575 5444730 : ref->accesses_in_loop.create (1);
1576 :
1577 5444730 : return ref;
1578 : }
1579 :
1580 : /* Records memory reference location *LOC in LOOP to the memory reference
1581 : description REF. The reference occurs in statement STMT. */
1582 :
1583 : static void
1584 5956541 : record_mem_ref_loc (im_mem_ref *ref, gimple *stmt, tree *loc)
1585 : {
1586 5956541 : mem_ref_loc aref;
1587 5956541 : aref.stmt = stmt;
1588 5956541 : aref.ref = loc;
1589 0 : ref->accesses_in_loop.safe_push (aref);
1590 0 : }
1591 :
1592 : /* Set the LOOP bit in REF stored bitmap and allocate that if
1593 : necessary. Return whether a bit was changed. */
1594 :
1595 : static bool
1596 4594052 : set_ref_stored_in_loop (im_mem_ref *ref, class loop *loop)
1597 : {
1598 4594052 : if (!ref->stored)
1599 2600865 : ref->stored = BITMAP_ALLOC (&lim_bitmap_obstack);
1600 4594052 : return bitmap_set_bit (ref->stored, loop->num);
1601 : }
1602 :
1603 : /* Marks reference REF as stored in LOOP. */
1604 :
1605 : static void
1606 3844829 : mark_ref_stored (im_mem_ref *ref, class loop *loop)
1607 : {
1608 3844829 : while (loop != current_loops->tree_root
1609 7368319 : && set_ref_stored_in_loop (ref, loop))
1610 3523490 : loop = loop_outer (loop);
1611 3844829 : }
1612 :
1613 : /* Set the LOOP bit in REF loaded bitmap and allocate that if
1614 : necessary. Return whether a bit was changed. */
1615 :
1616 : static bool
1617 6400771 : set_ref_loaded_in_loop (im_mem_ref *ref, class loop *loop)
1618 : {
1619 6400771 : if (!ref->loaded)
1620 3514190 : ref->loaded = BITMAP_ALLOC (&lim_bitmap_obstack);
1621 6400771 : return bitmap_set_bit (ref->loaded, loop->num);
1622 : }
1623 :
1624 : /* Marks reference REF as loaded in LOOP. */
1625 :
1626 : static void
1627 5130396 : mark_ref_loaded (im_mem_ref *ref, class loop *loop)
1628 : {
1629 5130396 : while (loop != current_loops->tree_root
1630 10144241 : && set_ref_loaded_in_loop (ref, loop))
1631 5013845 : loop = loop_outer (loop);
1632 5130396 : }
1633 :
1634 : /* Gathers memory references in statement STMT in LOOP, storing the
1635 : information about them in the memory_accesses structure. Marks
1636 : the vops accessed through unrecognized statements there as
1637 : well. */
1638 :
1639 : static void
1640 48205524 : gather_mem_refs_stmt (class loop *loop, gimple *stmt)
1641 : {
1642 48205524 : tree *mem = NULL;
1643 48205524 : hashval_t hash;
1644 48205524 : im_mem_ref **slot;
1645 48205524 : im_mem_ref *ref;
1646 48205524 : bool is_stored;
1647 48205524 : unsigned id;
1648 :
1649 64575278 : if (!gimple_vuse (stmt))
1650 : return;
1651 :
1652 7271046 : mem = simple_mem_ref_in_stmt (stmt, &is_stored);
1653 7271046 : if (!mem && is_gimple_assign (stmt))
1654 : {
1655 : /* For aggregate copies record distinct references but use them
1656 : only for disambiguation purposes. */
1657 474325 : id = memory_accesses.refs_list.length ();
1658 474325 : ref = mem_ref_alloc (NULL, 0, id);
1659 474325 : memory_accesses.refs_list.safe_push (ref);
1660 474325 : if (dump_file && (dump_flags & TDF_DETAILS))
1661 : {
1662 30 : fprintf (dump_file, "Unhandled memory reference %u: ", id);
1663 30 : print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1664 : }
1665 474325 : record_mem_ref_loc (ref, stmt, mem);
1666 948650 : is_stored = gimple_vdef (stmt);
1667 : }
1668 6796721 : else if (!mem)
1669 : {
1670 : /* We use the shared mem_ref for all unanalyzable refs. */
1671 1314505 : id = UNANALYZABLE_MEM_ID;
1672 1314505 : ref = memory_accesses.refs_list[id];
1673 1314505 : if (dump_file && (dump_flags & TDF_DETAILS))
1674 : {
1675 21 : fprintf (dump_file, "Unanalyzed memory reference %u: ", id);
1676 21 : print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1677 : }
1678 2629010 : is_stored = gimple_vdef (stmt);
1679 : }
1680 : else
1681 : {
1682 : /* We are looking for equal refs that might differ in structure
1683 : such as a.b vs. MEM[&a + 4]. So we key off the ao_ref but
1684 : make sure we can canonicalize the ref in the hashtable if
1685 : non-operand_equal_p refs are found. For the lookup we mark
1686 : the case we want strict equality with aor.max_size == -1. */
1687 5482216 : ao_ref aor;
1688 5482216 : ao_ref_init (&aor, *mem);
1689 5482216 : ao_ref_base (&aor);
1690 5482216 : ao_ref_alias_set (&aor);
1691 5482216 : HOST_WIDE_INT offset, size, max_size;
1692 5482216 : poly_int64 saved_maxsize = aor.max_size, mem_off;
1693 5482216 : tree mem_base;
1694 5482216 : bool ref_decomposed;
1695 5482216 : if (aor.max_size_known_p ()
1696 5298528 : && aor.offset.is_constant (&offset)
1697 5298528 : && aor.size.is_constant (&size)
1698 5298528 : && aor.max_size.is_constant (&max_size)
1699 5298528 : && size == max_size
1700 4703034 : && (size % BITS_PER_UNIT) == 0
1701 : /* We're canonicalizing to a MEM where TYPE_SIZE specifies the
1702 : size. Make sure this is consistent with the extraction. */
1703 4691389 : && poly_int_tree_p (TYPE_SIZE (TREE_TYPE (*mem)))
1704 4691389 : && known_eq (wi::to_poly_offset (TYPE_SIZE (TREE_TYPE (*mem))),
1705 : aor.size)
1706 10173513 : && (mem_base = get_addr_base_and_unit_offset (aor.ref, &mem_off)))
1707 : {
1708 4690163 : ref_decomposed = true;
1709 4690163 : tree base = ao_ref_base (&aor);
1710 4690163 : poly_int64 moffset;
1711 4690163 : HOST_WIDE_INT mcoffset;
1712 4690163 : if (TREE_CODE (base) == MEM_REF
1713 4690163 : && (mem_ref_offset (base) * BITS_PER_UNIT + offset).to_shwi (&moffset)
1714 4690163 : && moffset.is_constant (&mcoffset))
1715 : {
1716 2148177 : hash = iterative_hash_expr (TREE_OPERAND (base, 0), 0);
1717 2148177 : hash = iterative_hash_host_wide_int (mcoffset, hash);
1718 : }
1719 : else
1720 : {
1721 2541986 : hash = iterative_hash_expr (base, 0);
1722 2541986 : hash = iterative_hash_host_wide_int (offset, hash);
1723 : }
1724 4690163 : hash = iterative_hash_host_wide_int (size, hash);
1725 : }
1726 : else
1727 : {
1728 792053 : ref_decomposed = false;
1729 792053 : hash = iterative_hash_expr (aor.ref, 0);
1730 792053 : aor.max_size = -1;
1731 : }
1732 5482216 : slot = memory_accesses.refs->find_slot_with_hash (&aor, hash, INSERT);
1733 5482216 : aor.max_size = saved_maxsize;
1734 5482216 : if (*slot)
1735 : {
1736 1010053 : if (!(*slot)->ref_canonical
1737 1010053 : && !operand_equal_p (*mem, (*slot)->mem.ref, 0))
1738 : {
1739 : /* If we didn't yet canonicalize the hashtable ref (which
1740 : we'll end up using for code insertion) and hit a second
1741 : equal ref that is not structurally equivalent create
1742 : a canonical ref which is a bare MEM_REF. */
1743 60258 : if (TREE_CODE (*mem) == MEM_REF
1744 60258 : || TREE_CODE (*mem) == TARGET_MEM_REF)
1745 17860 : (*slot)->mem = aor;
1746 : else
1747 : {
1748 42398 : tree ref_alias_type = reference_alias_ptr_type (*mem);
1749 42398 : unsigned int ref_align = get_object_alignment (*mem);
1750 42398 : tree ref_type = TREE_TYPE (*mem);
1751 42398 : tree tmp = build1 (ADDR_EXPR, ptr_type_node,
1752 : unshare_expr (mem_base));
1753 42398 : if (TYPE_ALIGN (ref_type) != ref_align)
1754 12343 : ref_type = build_aligned_type (ref_type, ref_align);
1755 42398 : tree new_ref
1756 42398 : = fold_build2 (MEM_REF, ref_type, tmp,
1757 : build_int_cst (ref_alias_type, mem_off));
1758 42398 : if ((*slot)->mem.volatile_p)
1759 14 : TREE_THIS_VOLATILE (new_ref) = 1;
1760 42398 : (*slot)->mem.ref = new_ref;
1761 : /* Make sure the recorded base and offset are consistent
1762 : with the newly built ref. */
1763 42398 : if (TREE_CODE (TREE_OPERAND (new_ref, 0)) == ADDR_EXPR)
1764 : ;
1765 : else
1766 : {
1767 22495 : (*slot)->mem.base = new_ref;
1768 22495 : (*slot)->mem.offset = 0;
1769 : }
1770 42398 : gcc_checking_assert (TREE_CODE ((*slot)->mem.ref) == MEM_REF
1771 : && is_gimple_mem_ref_addr
1772 : (TREE_OPERAND ((*slot)->mem.ref,
1773 : 0)));
1774 42398 : (*slot)->mem.base_alias_set = (*slot)->mem.ref_alias_set;
1775 : }
1776 60258 : (*slot)->ref_canonical = true;
1777 : }
1778 1010053 : ref = *slot;
1779 1010053 : id = ref->id;
1780 : }
1781 : else
1782 : {
1783 4472163 : id = memory_accesses.refs_list.length ();
1784 4472163 : ref = mem_ref_alloc (&aor, hash, id);
1785 4472163 : ref->ref_decomposed = ref_decomposed;
1786 4472163 : memory_accesses.refs_list.safe_push (ref);
1787 4472163 : *slot = ref;
1788 :
1789 4472163 : if (dump_file && (dump_flags & TDF_DETAILS))
1790 : {
1791 550 : fprintf (dump_file, "Memory reference %u: ", id);
1792 550 : print_generic_expr (dump_file, ref->mem.ref, TDF_SLIM);
1793 550 : fprintf (dump_file, "\n");
1794 : }
1795 : }
1796 :
1797 5482216 : record_mem_ref_loc (ref, stmt, mem);
1798 : }
1799 7271046 : if (is_stored)
1800 : {
1801 3844829 : bitmap_set_bit (&memory_accesses.refs_stored_in_loop[loop->num], ref->id);
1802 3844829 : mark_ref_stored (ref, loop);
1803 : }
1804 : /* A not simple memory op is also a read when it is a write. */
1805 7271046 : if (!is_stored || id == UNANALYZABLE_MEM_ID
1806 2614975 : || ref->mem.ref == error_mark_node)
1807 : {
1808 5130396 : bitmap_set_bit (&memory_accesses.refs_loaded_in_loop[loop->num], ref->id);
1809 5130396 : mark_ref_loaded (ref, loop);
1810 : }
1811 7271046 : init_lim_data (stmt)->ref = ref->id;
1812 7271046 : return;
1813 : }
1814 :
1815 : static unsigned *bb_loop_postorder;
1816 :
1817 : /* qsort sort function to sort blocks after their loop fathers postorder. */
1818 :
1819 : static int
1820 164998334 : sort_bbs_in_loop_postorder_cmp (const void *bb1_, const void *bb2_,
1821 : void *bb_loop_postorder_)
1822 : {
1823 164998334 : unsigned *bb_loop_postorder = (unsigned *)bb_loop_postorder_;
1824 164998334 : basic_block bb1 = *(const basic_block *)bb1_;
1825 164998334 : basic_block bb2 = *(const basic_block *)bb2_;
1826 164998334 : class loop *loop1 = bb1->loop_father;
1827 164998334 : class loop *loop2 = bb2->loop_father;
1828 164998334 : if (loop1->num == loop2->num)
1829 78722346 : return bb1->index - bb2->index;
1830 86275988 : return bb_loop_postorder[loop1->num] < bb_loop_postorder[loop2->num] ? -1 : 1;
1831 : }
1832 :
1833 : /* qsort sort function to sort ref locs after their loop fathers postorder. */
1834 :
1835 : static int
1836 1010038 : sort_locs_in_loop_postorder_cmp (const void *loc1_, const void *loc2_,
1837 : void *bb_loop_postorder_)
1838 : {
1839 1010038 : unsigned *bb_loop_postorder = (unsigned *)bb_loop_postorder_;
1840 1010038 : const mem_ref_loc *loc1 = (const mem_ref_loc *)loc1_;
1841 1010038 : const mem_ref_loc *loc2 = (const mem_ref_loc *)loc2_;
1842 1010038 : class loop *loop1 = gimple_bb (loc1->stmt)->loop_father;
1843 1010038 : class loop *loop2 = gimple_bb (loc2->stmt)->loop_father;
1844 1010038 : if (loop1->num == loop2->num)
1845 : return 0;
1846 149238 : return bb_loop_postorder[loop1->num] < bb_loop_postorder[loop2->num] ? -1 : 1;
1847 : }
1848 :
1849 : /* Gathers memory references in loops. */
1850 :
1851 : static void
1852 498242 : analyze_memory_references (bool store_motion)
1853 : {
1854 498242 : gimple_stmt_iterator bsi;
1855 498242 : basic_block bb, *bbs;
1856 498242 : class loop *outer;
1857 498242 : unsigned i, n;
1858 :
1859 : /* Collect all basic-blocks in loops and sort them after their
1860 : loops postorder. */
1861 498242 : i = 0;
1862 498242 : bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS);
1863 15473690 : FOR_EACH_BB_FN (bb, cfun)
1864 14975448 : if (bb->loop_father != current_loops->tree_root)
1865 6287655 : bbs[i++] = bb;
1866 498242 : n = i;
1867 498242 : gcc_sort_r (bbs, n, sizeof (basic_block), sort_bbs_in_loop_postorder_cmp,
1868 : bb_loop_postorder);
1869 :
1870 : /* Visit blocks in loop postorder and assign mem-ref IDs in that order.
1871 : That results in better locality for all the bitmaps. It also
1872 : automatically sorts the location list of gathered memory references
1873 : after their loop postorder number allowing to binary-search it. */
1874 7284139 : for (i = 0; i < n; ++i)
1875 : {
1876 6287655 : basic_block bb = bbs[i];
1877 60780834 : for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
1878 48205524 : gather_mem_refs_stmt (bb->loop_father, gsi_stmt (bsi));
1879 : }
1880 :
1881 : /* Verify the list of gathered memory references is sorted after their
1882 : loop postorder number. */
1883 498242 : if (flag_checking)
1884 : {
1885 : im_mem_ref *ref;
1886 11387624 : FOR_EACH_VEC_ELT (memory_accesses.refs_list, i, ref)
1887 6454736 : for (unsigned j = 1; j < ref->accesses_in_loop.length (); ++j)
1888 1010038 : gcc_assert (sort_locs_in_loop_postorder_cmp
1889 : (&ref->accesses_in_loop[j-1], &ref->accesses_in_loop[j],
1890 : bb_loop_postorder) <= 0);
1891 : }
1892 :
1893 498242 : free (bbs);
1894 :
1895 498242 : if (!store_motion)
1896 498242 : return;
1897 :
1898 : /* Propagate the information about accessed memory references up
1899 : the loop hierarchy. */
1900 2833522 : for (auto loop : loops_list (cfun, LI_FROM_INNERMOST))
1901 : {
1902 : /* Finalize the overall touched references (including subloops). */
1903 1343287 : bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[loop->num],
1904 1343287 : &memory_accesses.refs_stored_in_loop[loop->num]);
1905 :
1906 : /* Propagate the information about accessed memory references up
1907 : the loop hierarchy. */
1908 1343287 : outer = loop_outer (loop);
1909 1343287 : if (outer == current_loops->tree_root)
1910 1039648 : continue;
1911 :
1912 303639 : bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[outer->num],
1913 303639 : &memory_accesses.all_refs_stored_in_loop[loop->num]);
1914 496745 : }
1915 : }
1916 :
1917 : /* Returns true if MEM1 and MEM2 may alias. TTAE_CACHE is used as a cache in
1918 : tree_to_aff_combination_expand. */
1919 :
1920 : static bool
1921 547389 : mem_refs_may_alias_p (im_mem_ref *mem1, im_mem_ref *mem2,
1922 : hash_map<tree, name_expansion *> **ttae_cache,
1923 : bool tbaa_p)
1924 : {
1925 547389 : gcc_checking_assert (mem1->mem.ref != error_mark_node
1926 : && mem2->mem.ref != error_mark_node);
1927 :
1928 : /* Perform BASE + OFFSET analysis -- if MEM1 and MEM2 are based on the same
1929 : object and their offset differ in such a way that the locations cannot
1930 : overlap, then they cannot alias. */
1931 547389 : poly_widest_int size1, size2;
1932 1094778 : aff_tree off1, off2;
1933 :
1934 : /* Perform basic offset and type-based disambiguation. */
1935 547389 : if (!refs_may_alias_p_1 (&mem1->mem, &mem2->mem, tbaa_p))
1936 : return false;
1937 :
1938 : /* The expansion of addresses may be a bit expensive, thus we only do
1939 : the check at -O2 and higher optimization levels. */
1940 65993 : if (optimize < 2)
1941 : return true;
1942 :
1943 55867 : get_inner_reference_aff (mem1->mem.ref, &off1, &size1);
1944 55867 : get_inner_reference_aff (mem2->mem.ref, &off2, &size2);
1945 55867 : aff_combination_expand (&off1, ttae_cache);
1946 55867 : aff_combination_expand (&off2, ttae_cache);
1947 55867 : aff_combination_scale (&off1, -1);
1948 55867 : aff_combination_add (&off2, &off1);
1949 :
1950 55867 : if (aff_comb_cannot_overlap_p (&off2, size1, size2))
1951 2407 : return false;
1952 :
1953 : return true;
1954 547389 : }
1955 :
1956 : /* Compare function for bsearch searching for reference locations
1957 : in a loop. */
1958 :
1959 : static int
1960 270563 : find_ref_loc_in_loop_cmp (const void *loop_, const void *loc_,
1961 : void *bb_loop_postorder_)
1962 : {
1963 270563 : unsigned *bb_loop_postorder = (unsigned *)bb_loop_postorder_;
1964 270563 : class loop *loop = (class loop *)const_cast<void *>(loop_);
1965 270563 : mem_ref_loc *loc = (mem_ref_loc *)const_cast<void *>(loc_);
1966 270563 : class loop *loc_loop = gimple_bb (loc->stmt)->loop_father;
1967 270563 : if (loop->num == loc_loop->num
1968 270563 : || flow_loop_nested_p (loop, loc_loop))
1969 : return 0;
1970 61487 : return (bb_loop_postorder[loop->num] < bb_loop_postorder[loc_loop->num]
1971 61487 : ? -1 : 1);
1972 : }
1973 :
1974 : /* Iterates over all locations of REF in LOOP and its subloops calling
1975 : fn.operator() with the location as argument. When that operator
1976 : returns true the iteration is stopped and true is returned.
1977 : Otherwise false is returned. */
1978 :
1979 : template <typename FN>
1980 : static bool
1981 209076 : for_all_locs_in_loop (class loop *loop, im_mem_ref *ref, FN fn)
1982 : {
1983 : unsigned i;
1984 : mem_ref_loc *loc;
1985 :
1986 : /* Search for the cluster of locs in the accesses_in_loop vector
1987 : which is sorted after postorder index of the loop father. */
1988 209076 : loc = ref->accesses_in_loop.bsearch (loop, find_ref_loc_in_loop_cmp,
1989 : bb_loop_postorder);
1990 209076 : if (!loc)
1991 : return false;
1992 :
1993 : /* We have found one location inside loop or its sub-loops. Iterate
1994 : both forward and backward to cover the whole cluster. */
1995 209076 : i = loc - ref->accesses_in_loop.address ();
1996 299926 : while (i > 0)
1997 : {
1998 148874 : --i;
1999 148874 : mem_ref_loc *l = &ref->accesses_in_loop[i];
2000 148874 : if (!flow_bb_inside_loop_p (loop, gimple_bb (l->stmt)))
2001 : break;
2002 124536 : if (fn (l))
2003 : return true;
2004 : }
2005 487208 : for (i = loc - ref->accesses_in_loop.address ();
2006 311818 : i < ref->accesses_in_loop.length (); ++i)
2007 : {
2008 231823 : mem_ref_loc *l = &ref->accesses_in_loop[i];
2009 231823 : if (!flow_bb_inside_loop_p (loop, gimple_bb (l->stmt)))
2010 : break;
2011 212191 : if (fn (l))
2012 : return true;
2013 : }
2014 :
2015 : return false;
2016 : }
2017 :
2018 : /* Rewrites location LOC by TMP_VAR. */
2019 :
2020 : class rewrite_mem_ref_loc
2021 : {
2022 : public:
2023 34299 : rewrite_mem_ref_loc (tree tmp_var_) : tmp_var (tmp_var_) {}
2024 : bool operator () (mem_ref_loc *loc);
2025 : tree tmp_var;
2026 : };
2027 :
2028 : bool
2029 54196 : rewrite_mem_ref_loc::operator () (mem_ref_loc *loc)
2030 : {
2031 54196 : *loc->ref = tmp_var;
2032 54196 : update_stmt (loc->stmt);
2033 54196 : return false;
2034 : }
2035 :
2036 : /* Rewrites all references to REF in LOOP by variable TMP_VAR. */
2037 :
2038 : static void
2039 34299 : rewrite_mem_refs (class loop *loop, im_mem_ref *ref, tree tmp_var)
2040 : {
2041 0 : for_all_locs_in_loop (loop, ref, rewrite_mem_ref_loc (tmp_var));
2042 0 : }
2043 :
2044 : /* Stores the first reference location in LOCP. */
2045 :
2046 : class first_mem_ref_loc_1
2047 : {
2048 : public:
2049 34299 : first_mem_ref_loc_1 (mem_ref_loc **locp_) : locp (locp_) {}
2050 : bool operator () (mem_ref_loc *loc);
2051 : mem_ref_loc **locp;
2052 : };
2053 :
2054 : bool
2055 34299 : first_mem_ref_loc_1::operator () (mem_ref_loc *loc)
2056 : {
2057 34299 : *locp = loc;
2058 34299 : return true;
2059 : }
2060 :
2061 : /* Returns the first reference location to REF in LOOP. */
2062 :
2063 : static mem_ref_loc *
2064 34299 : first_mem_ref_loc (class loop *loop, im_mem_ref *ref)
2065 : {
2066 34299 : mem_ref_loc *locp = NULL;
2067 0 : for_all_locs_in_loop (loop, ref, first_mem_ref_loc_1 (&locp));
2068 34299 : return locp;
2069 : }
2070 :
2071 : /* Helper function for execute_sm. Emit code to store TMP_VAR into
2072 : MEM along edge EX.
2073 :
2074 : The store is only done if MEM has changed. We do this so no
2075 : changes to MEM occur on code paths that did not originally store
2076 : into it.
2077 :
2078 : The common case for execute_sm will transform:
2079 :
2080 : for (...) {
2081 : if (foo)
2082 : stuff;
2083 : else
2084 : MEM = TMP_VAR;
2085 : }
2086 :
2087 : into:
2088 :
2089 : lsm = MEM;
2090 : for (...) {
2091 : if (foo)
2092 : stuff;
2093 : else
2094 : lsm = TMP_VAR;
2095 : }
2096 : MEM = lsm;
2097 :
2098 : This function will generate:
2099 :
2100 : lsm = MEM;
2101 :
2102 : lsm_flag = false;
2103 : ...
2104 : for (...) {
2105 : if (foo)
2106 : stuff;
2107 : else {
2108 : lsm = TMP_VAR;
2109 : lsm_flag = true;
2110 : }
2111 : }
2112 : if (lsm_flag) <--
2113 : MEM = lsm; <-- (X)
2114 :
2115 : In case MEM and TMP_VAR are NULL the function will return the then
2116 : block so the caller can insert (X) and other related stmts.
2117 : */
2118 :
2119 : static basic_block
2120 11850 : execute_sm_if_changed (edge ex, tree mem, tree tmp_var, tree flag,
2121 : edge preheader, hash_set <basic_block> *flag_bbs,
2122 : edge &append_cond_position, edge &last_cond_fallthru)
2123 : {
2124 11850 : basic_block new_bb, then_bb, old_dest;
2125 11850 : bool loop_has_only_one_exit;
2126 11850 : edge then_old_edge;
2127 11850 : gimple_stmt_iterator gsi;
2128 11850 : gimple *stmt;
2129 11850 : bool irr = ex->flags & EDGE_IRREDUCIBLE_LOOP;
2130 :
2131 11850 : profile_count count_sum = profile_count::zero ();
2132 11850 : int nbbs = 0, ncount = 0;
2133 11850 : profile_probability flag_probability = profile_probability::uninitialized ();
2134 :
2135 : /* Flag is set in FLAG_BBS. Determine probability that flag will be true
2136 : at loop exit.
2137 :
2138 : This code may look fancy, but it cannot update profile very realistically
2139 : because we do not know the probability that flag will be true at given
2140 : loop exit.
2141 :
2142 : We look for two interesting extremes
2143 : - when exit is dominated by block setting the flag, we know it will
2144 : always be true. This is a common case.
2145 : - when all blocks setting the flag have very low frequency we know
2146 : it will likely be false.
2147 : In all other cases we default to 2/3 for flag being true. */
2148 :
2149 11850 : for (hash_set<basic_block>::iterator it = flag_bbs->begin ();
2150 26723 : it != flag_bbs->end (); ++it)
2151 : {
2152 14873 : if ((*it)->count.initialized_p ())
2153 14860 : count_sum += (*it)->count, ncount ++;
2154 14873 : if (dominated_by_p (CDI_DOMINATORS, ex->src, *it))
2155 2437 : flag_probability = profile_probability::always ();
2156 14873 : nbbs++;
2157 : }
2158 :
2159 11850 : profile_probability cap
2160 11850 : = profile_probability::guessed_always ().apply_scale (2, 3);
2161 :
2162 11850 : if (flag_probability.initialized_p ())
2163 : ;
2164 9520 : else if (ncount == nbbs
2165 19231 : && preheader->count () >= count_sum && preheader->count ().nonzero_p ())
2166 : {
2167 172 : flag_probability = count_sum.probability_in (preheader->count ());
2168 172 : if (flag_probability > cap)
2169 77 : flag_probability = cap;
2170 : }
2171 :
2172 11850 : if (!flag_probability.initialized_p ())
2173 9348 : flag_probability = cap;
2174 :
2175 : /* ?? Insert store after previous store if applicable. See note
2176 : below. */
2177 11850 : if (append_cond_position)
2178 4131 : ex = append_cond_position;
2179 :
2180 11850 : loop_has_only_one_exit = single_pred_p (ex->dest);
2181 :
2182 11850 : if (loop_has_only_one_exit)
2183 5012 : ex = split_block_after_labels (ex->dest);
2184 : else
2185 : {
2186 6838 : for (gphi_iterator gpi = gsi_start_phis (ex->dest);
2187 10540 : !gsi_end_p (gpi); gsi_next (&gpi))
2188 : {
2189 4476 : gphi *phi = gpi.phi ();
2190 8952 : if (virtual_operand_p (gimple_phi_result (phi)))
2191 3702 : continue;
2192 :
2193 : /* When the destination has a non-virtual PHI node with multiple
2194 : predecessors make sure we preserve the PHI structure by
2195 : forcing a forwarder block so that hoisting of that PHI will
2196 : still work. */
2197 774 : split_edge (ex);
2198 774 : break;
2199 : }
2200 : }
2201 :
2202 11850 : old_dest = ex->dest;
2203 11850 : new_bb = split_edge (ex);
2204 11850 : if (append_cond_position)
2205 4131 : new_bb->count += last_cond_fallthru->count ();
2206 11850 : then_bb = create_empty_bb (new_bb);
2207 11850 : then_bb->count = new_bb->count.apply_probability (flag_probability);
2208 11850 : if (irr)
2209 49 : then_bb->flags = BB_IRREDUCIBLE_LOOP;
2210 11850 : add_bb_to_loop (then_bb, new_bb->loop_father);
2211 :
2212 11850 : gsi = gsi_start_bb (new_bb);
2213 11850 : stmt = gimple_build_cond (NE_EXPR, flag, boolean_false_node,
2214 : NULL_TREE, NULL_TREE);
2215 11850 : gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
2216 :
2217 : /* Insert actual store. */
2218 11850 : if (mem)
2219 : {
2220 9602 : gsi = gsi_start_bb (then_bb);
2221 9602 : stmt = gimple_build_assign (unshare_expr (mem), tmp_var);
2222 9602 : gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
2223 : }
2224 :
2225 11850 : edge e1 = single_succ_edge (new_bb);
2226 23651 : edge e2 = make_edge (new_bb, then_bb,
2227 : EDGE_TRUE_VALUE | (irr ? EDGE_IRREDUCIBLE_LOOP : 0));
2228 11850 : e2->probability = flag_probability;
2229 :
2230 11850 : e1->flags |= EDGE_FALSE_VALUE | (irr ? EDGE_IRREDUCIBLE_LOOP : 0);
2231 11850 : e1->flags &= ~EDGE_FALLTHRU;
2232 :
2233 11850 : e1->probability = flag_probability.invert ();
2234 :
2235 23651 : then_old_edge = make_single_succ_edge (then_bb, old_dest,
2236 : EDGE_FALLTHRU | (irr ? EDGE_IRREDUCIBLE_LOOP : 0));
2237 :
2238 11850 : set_immediate_dominator (CDI_DOMINATORS, then_bb, new_bb);
2239 :
2240 11850 : if (append_cond_position)
2241 : {
2242 4131 : basic_block prevbb = last_cond_fallthru->src;
2243 4131 : redirect_edge_succ (last_cond_fallthru, new_bb);
2244 4131 : set_immediate_dominator (CDI_DOMINATORS, new_bb, prevbb);
2245 4131 : set_immediate_dominator (CDI_DOMINATORS, old_dest,
2246 : recompute_dominator (CDI_DOMINATORS, old_dest));
2247 : }
2248 :
2249 : /* ?? Because stores may alias, they must happen in the exact
2250 : sequence they originally happened. Save the position right after
2251 : the (_lsm) store we just created so we can continue appending after
2252 : it and maintain the original order. */
2253 11850 : append_cond_position = then_old_edge;
2254 11850 : last_cond_fallthru = find_edge (new_bb, old_dest);
2255 :
2256 11850 : if (!loop_has_only_one_exit)
2257 6838 : for (gphi_iterator gpi = gsi_start_phis (old_dest);
2258 10386 : !gsi_end_p (gpi); gsi_next (&gpi))
2259 : {
2260 3548 : gphi *phi = gpi.phi ();
2261 3548 : unsigned i;
2262 :
2263 23492 : for (i = 0; i < gimple_phi_num_args (phi); i++)
2264 16396 : if (gimple_phi_arg_edge (phi, i)->src == new_bb)
2265 : {
2266 3548 : tree arg = gimple_phi_arg_def (phi, i);
2267 3548 : add_phi_arg (phi, arg, then_old_edge, UNKNOWN_LOCATION);
2268 3548 : update_stmt (phi);
2269 : }
2270 : }
2271 :
2272 11850 : return then_bb;
2273 : }
2274 :
2275 : /* When REF is set on the location, set flag indicating the store. */
2276 :
2277 : class sm_set_flag_if_changed
2278 : {
2279 : public:
2280 6735 : sm_set_flag_if_changed (tree flag_, hash_set <basic_block> *bbs_)
2281 6735 : : flag (flag_), bbs (bbs_) {}
2282 : bool operator () (mem_ref_loc *loc);
2283 : tree flag;
2284 : hash_set <basic_block> *bbs;
2285 : };
2286 :
2287 : bool
2288 12493 : sm_set_flag_if_changed::operator () (mem_ref_loc *loc)
2289 : {
2290 : /* Only set the flag for writes. */
2291 12493 : if (is_gimple_assign (loc->stmt)
2292 12493 : && gimple_assign_lhs_ptr (loc->stmt) == loc->ref)
2293 : {
2294 7755 : gimple_stmt_iterator gsi = gsi_for_stmt (loc->stmt);
2295 7755 : gimple *stmt = gimple_build_assign (flag, boolean_true_node);
2296 7755 : gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING);
2297 7755 : bbs->add (gimple_bb (stmt));
2298 : }
2299 12493 : return false;
2300 : }
2301 :
2302 : /* Helper function for execute_sm. On every location where REF is
2303 : set, set an appropriate flag indicating the store. */
2304 :
2305 : static tree
2306 6735 : execute_sm_if_changed_flag_set (class loop *loop, im_mem_ref *ref,
2307 : hash_set <basic_block> *bbs)
2308 : {
2309 6735 : tree flag;
2310 6735 : char *str = get_lsm_tmp_name (ref->mem.ref, ~0, "_flag");
2311 6735 : flag = create_tmp_reg (boolean_type_node, str);
2312 6735 : for_all_locs_in_loop (loop, ref, sm_set_flag_if_changed (flag, bbs));
2313 6735 : return flag;
2314 : }
2315 :
2316 34299 : struct sm_aux
2317 : {
2318 : tree tmp_var;
2319 : tree store_flag;
2320 : hash_set <basic_block> flag_bbs;
2321 : };
2322 :
2323 : /* Executes store motion of memory reference REF from LOOP.
2324 : Exits from the LOOP are stored in EXITS. The initialization of the
2325 : temporary variable is put to the preheader of the loop, and assignments
2326 : to the reference from the temporary variable are emitted to exits. */
2327 :
2328 : static sm_aux *
2329 34299 : execute_sm (class loop *loop, im_mem_ref *ref,
2330 : hash_map<im_mem_ref *, sm_aux *> &aux_map, bool maybe_mt,
2331 : bool use_other_flag_var)
2332 : {
2333 34299 : gassign *load;
2334 34299 : struct fmt_data fmt_data;
2335 34299 : struct lim_aux_data *lim_data;
2336 34299 : bool multi_threaded_model_p = false;
2337 34299 : gimple_stmt_iterator gsi;
2338 34299 : sm_aux *aux = new sm_aux;
2339 :
2340 34299 : if (dump_file && (dump_flags & TDF_DETAILS))
2341 : {
2342 82 : fprintf (dump_file, "Executing store motion of ");
2343 82 : print_generic_expr (dump_file, ref->mem.ref);
2344 82 : fprintf (dump_file, " from loop %d\n", loop->num);
2345 : }
2346 :
2347 34299 : aux->tmp_var = create_tmp_reg (TREE_TYPE (ref->mem.ref),
2348 34299 : get_lsm_tmp_name (ref->mem.ref, ~0));
2349 :
2350 34299 : fmt_data.loop = loop;
2351 34299 : fmt_data.orig_loop = loop;
2352 34299 : for_each_index (&ref->mem.ref, force_move_till, &fmt_data);
2353 :
2354 68598 : bool always_stored = ref_always_stored_p (loop, ref);
2355 34299 : if (maybe_mt
2356 10792 : && (bb_in_transaction (loop_preheader_edge (loop)->src)
2357 10791 : || (ref_can_have_store_data_races (ref->mem.ref) && ! always_stored)
2358 : /* Do not speculate a load/store when that's not a noop, either
2359 : because the mode cannot be transferred or because there's
2360 : UB involved for out-of-bound values. */
2361 3794 : || !mode_can_transfer_bits (TYPE_MODE (TREE_TYPE (ref->mem.ref)))
2362 3794 : || TREE_CODE (TREE_TYPE (ref->mem.ref)) == BOOLEAN_TYPE
2363 3783 : || (TREE_CODE (ref->mem.ref) == COMPONENT_REF
2364 882 : && DECL_BIT_FIELD (TREE_OPERAND (ref->mem.ref, 1)))))
2365 : multi_threaded_model_p = true;
2366 :
2367 34299 : if (multi_threaded_model_p && !use_other_flag_var)
2368 6735 : aux->store_flag
2369 6735 : = execute_sm_if_changed_flag_set (loop, ref, &aux->flag_bbs);
2370 : else
2371 : aux->store_flag = NULL_TREE;
2372 :
2373 : /* Remember variable setup. */
2374 34299 : aux_map.put (ref, aux);
2375 :
2376 34299 : rewrite_mem_refs (loop, ref, aux->tmp_var);
2377 :
2378 : /* Emit the load code on a random exit edge or into the latch if
2379 : the loop does not exit, so that we are sure it will be processed
2380 : by move_computations after all dependencies. */
2381 34299 : gsi = gsi_for_stmt (first_mem_ref_loc (loop, ref)->stmt);
2382 :
2383 : /* Avoid doing a load if there was no load of the ref in the loop.
2384 : Esp. when the ref is not always stored we cannot optimize it
2385 : away later. But when it is not always stored we must use a conditional
2386 : store then. */
2387 34299 : if ((!always_stored && !multi_threaded_model_p)
2388 34299 : || (ref->loaded && bitmap_bit_p (ref->loaded, loop->num)))
2389 17542 : load = gimple_build_assign (aux->tmp_var, unshare_expr (ref->mem.ref));
2390 : else
2391 : {
2392 : /* If not emitting a load mark the uninitialized state on the
2393 : loop entry as not to be warned for. */
2394 16757 : tree uninit = create_tmp_reg (TREE_TYPE (aux->tmp_var));
2395 16757 : suppress_warning (uninit, OPT_Wuninitialized);
2396 16757 : uninit = get_or_create_ssa_default_def (cfun, uninit);
2397 16757 : load = gimple_build_assign (aux->tmp_var, uninit);
2398 : }
2399 34299 : lim_data = init_lim_data (load);
2400 34299 : lim_data->max_loop = loop;
2401 34299 : lim_data->tgt_loop = loop;
2402 34299 : gsi_insert_before (&gsi, load, GSI_SAME_STMT);
2403 :
2404 34299 : if (aux->store_flag)
2405 : {
2406 6735 : load = gimple_build_assign (aux->store_flag, boolean_false_node);
2407 6735 : lim_data = init_lim_data (load);
2408 6735 : lim_data->max_loop = loop;
2409 6735 : lim_data->tgt_loop = loop;
2410 6735 : gsi_insert_before (&gsi, load, GSI_SAME_STMT);
2411 : }
2412 :
2413 34299 : return aux;
2414 : }
2415 :
2416 : /* sm_ord is used for ordinary stores we can retain order with respect
2417 : to other stores
2418 : sm_unord is used for conditional executed stores which need to be
2419 : able to execute in arbitrary order with respect to other stores
2420 : sm_other is used for stores we do not try to apply store motion to. */
2421 : enum sm_kind { sm_ord, sm_unord, sm_other };
2422 : struct seq_entry
2423 : {
2424 : seq_entry () = default;
2425 50761 : seq_entry (unsigned f, sm_kind k, tree fr = NULL)
2426 50761 : : first (f), second (k), from (fr) {}
2427 : unsigned first;
2428 : sm_kind second;
2429 : tree from;
2430 : };
2431 :
2432 : static void
2433 31707 : execute_sm_exit (class loop *loop, edge ex, vec<seq_entry> &seq,
2434 : hash_map<im_mem_ref *, sm_aux *> &aux_map, sm_kind kind,
2435 : edge &append_cond_position, edge &last_cond_fallthru,
2436 : bitmap clobbers_to_prune)
2437 : {
2438 : /* Sink the stores to exit from the loop. */
2439 111148 : for (unsigned i = seq.length (); i > 0; --i)
2440 : {
2441 47734 : im_mem_ref *ref = memory_accesses.refs_list[seq[i-1].first];
2442 47734 : if (seq[i-1].second == sm_other)
2443 : {
2444 4845 : gcc_assert (kind == sm_ord && seq[i-1].from != NULL_TREE);
2445 4845 : gassign *store;
2446 4845 : if (ref->mem.ref == error_mark_node)
2447 : {
2448 183 : tree lhs = gimple_assign_lhs (ref->accesses_in_loop[0].stmt);
2449 183 : if (dump_file && (dump_flags & TDF_DETAILS))
2450 : {
2451 3 : fprintf (dump_file, "Re-issuing dependent ");
2452 3 : print_generic_expr (dump_file, unshare_expr (seq[i-1].from));
2453 3 : fprintf (dump_file, " of ");
2454 3 : print_generic_expr (dump_file, lhs);
2455 3 : fprintf (dump_file, " from loop %d on exit %d -> %d\n",
2456 3 : loop->num, ex->src->index, ex->dest->index);
2457 : }
2458 183 : store = gimple_build_assign (unshare_expr (lhs),
2459 183 : unshare_expr (seq[i-1].from));
2460 183 : bitmap_set_bit (clobbers_to_prune, seq[i-1].first);
2461 : }
2462 : else
2463 : {
2464 4662 : if (dump_file && (dump_flags & TDF_DETAILS))
2465 : {
2466 0 : fprintf (dump_file, "Re-issuing dependent store of ");
2467 0 : print_generic_expr (dump_file, ref->mem.ref);
2468 0 : fprintf (dump_file, " from loop %d on exit %d -> %d\n",
2469 0 : loop->num, ex->src->index, ex->dest->index);
2470 : }
2471 4662 : store = gimple_build_assign (unshare_expr (ref->mem.ref),
2472 4662 : seq[i-1].from);
2473 : }
2474 4845 : gsi_insert_on_edge (ex, store);
2475 : }
2476 : else
2477 : {
2478 42889 : sm_aux *aux = *aux_map.get (ref);
2479 42889 : if (!aux->store_flag || kind == sm_ord)
2480 : {
2481 33287 : gassign *store;
2482 33287 : store = gimple_build_assign (unshare_expr (ref->mem.ref),
2483 : aux->tmp_var);
2484 33287 : gsi_insert_on_edge (ex, store);
2485 33287 : }
2486 : else
2487 9602 : execute_sm_if_changed (ex, ref->mem.ref, aux->tmp_var,
2488 : aux->store_flag,
2489 : loop_preheader_edge (loop), &aux->flag_bbs,
2490 : append_cond_position, last_cond_fallthru);
2491 : }
2492 : }
2493 31707 : }
2494 :
2495 : /* Push the SM candidate at index PTR in the sequence SEQ down until
2496 : we hit the next SM candidate. Return true if that went OK and
2497 : false if we could not disambiguate against another unrelated ref.
2498 : Update *AT to the index where the candidate now resides. */
2499 :
2500 : static bool
2501 37440 : sm_seq_push_down (vec<seq_entry> &seq, unsigned ptr, unsigned *at)
2502 : {
2503 37440 : *at = ptr;
2504 38181 : for (; ptr > 0; --ptr)
2505 : {
2506 17612 : seq_entry &new_cand = seq[ptr];
2507 17612 : seq_entry &against = seq[ptr-1];
2508 17612 : if (against.second == sm_ord
2509 5722 : || (against.second == sm_other && against.from != NULL_TREE))
2510 : /* Found the tail of the sequence. */
2511 : break;
2512 : /* We may not ignore self-dependences here. */
2513 1048 : if (new_cand.first == against.first
2514 : /* ??? We could actually handle clobbers here, but not easily
2515 : with LIMs dependence analysis. */
2516 826 : || (memory_accesses.refs_list[new_cand.first]->mem.ref
2517 826 : == error_mark_node)
2518 821 : || (memory_accesses.refs_list[against.first]->mem.ref
2519 : == error_mark_node)
2520 1845 : || !refs_independent_p (memory_accesses.refs_list[new_cand.first],
2521 797 : memory_accesses.refs_list[against.first],
2522 : false))
2523 : /* ??? Prune new_cand from the list of refs to apply SM to. */
2524 : return false;
2525 741 : std::swap (new_cand, against);
2526 741 : *at = ptr - 1;
2527 : }
2528 : return true;
2529 : }
2530 :
2531 : /* Computes the sequence of stores from candidates in REFS_NOT_IN_SEQ to SEQ
2532 : walking backwards from VDEF (or the end of BB if VDEF is NULL). */
2533 :
2534 : static int
2535 29707 : sm_seq_valid_bb (class loop *loop, basic_block bb, tree vdef,
2536 : vec<seq_entry> &seq, bitmap refs_not_in_seq,
2537 : bitmap refs_not_supported, bool forked,
2538 : bitmap fully_visited)
2539 : {
2540 34525 : if (!vdef)
2541 139118 : for (gimple_stmt_iterator gsi = gsi_last_bb (bb); !gsi_end_p (gsi);
2542 87244 : gsi_prev (&gsi))
2543 : {
2544 138800 : vdef = gimple_vdef (gsi_stmt (gsi));
2545 51556 : if (vdef)
2546 : break;
2547 : }
2548 25937 : if (!vdef)
2549 : {
2550 8019 : gphi *vphi = get_virtual_phi (bb);
2551 8019 : if (vphi)
2552 3952 : vdef = gimple_phi_result (vphi);
2553 : }
2554 34525 : if (!vdef)
2555 : {
2556 4067 : if (single_pred_p (bb))
2557 : /* This handles the perfect nest case. */
2558 3289 : return sm_seq_valid_bb (loop, single_pred (bb), vdef,
2559 : seq, refs_not_in_seq, refs_not_supported,
2560 3289 : forked, fully_visited);
2561 : return 0;
2562 : }
2563 77404 : do
2564 : {
2565 53931 : gimple *def = SSA_NAME_DEF_STMT (vdef);
2566 53931 : if (gimple_bb (def) != bb)
2567 : {
2568 : /* If we forked by processing a PHI do not allow our walk to
2569 : merge again until we handle that robustly. */
2570 2968 : if (forked)
2571 : {
2572 : /* Mark refs_not_in_seq as unsupported. */
2573 1261 : bitmap_ior_into (refs_not_supported, refs_not_in_seq);
2574 1261 : return 1;
2575 : }
2576 : /* Otherwise it doesn't really matter if we end up in different
2577 : BBs. */
2578 : bb = gimple_bb (def);
2579 : }
2580 52670 : if (gphi *phi = dyn_cast <gphi *> (def))
2581 : {
2582 : /* Handle CFG merges. Until we handle forks (gimple_bb (def) != bb)
2583 : this is still linear.
2584 : Eventually we want to cache intermediate results per BB
2585 : (but we can't easily cache for different exits?). */
2586 : /* Stop at PHIs with possible backedges. */
2587 10080 : if (bb == bb->loop_father->header
2588 5099 : || bb->flags & BB_IRREDUCIBLE_LOOP)
2589 : {
2590 : /* Mark refs_not_in_seq as unsupported. */
2591 4986 : bitmap_ior_into (refs_not_supported, refs_not_in_seq);
2592 4986 : return 1;
2593 : }
2594 5094 : if (gimple_phi_num_args (phi) == 1)
2595 1529 : return sm_seq_valid_bb (loop, gimple_phi_arg_edge (phi, 0)->src,
2596 : gimple_phi_arg_def (phi, 0), seq,
2597 : refs_not_in_seq, refs_not_supported,
2598 1529 : false, fully_visited);
2599 7130 : if (bitmap_bit_p (fully_visited,
2600 3565 : SSA_NAME_VERSION (gimple_phi_result (phi))))
2601 : return 1;
2602 3461 : auto_vec<seq_entry> first_edge_seq;
2603 3461 : auto_bitmap tem_refs_not_in_seq (&lim_bitmap_obstack);
2604 3461 : int eret;
2605 3461 : bitmap_copy (tem_refs_not_in_seq, refs_not_in_seq);
2606 3461 : eret = sm_seq_valid_bb (loop, gimple_phi_arg_edge (phi, 0)->src,
2607 : gimple_phi_arg_def (phi, 0),
2608 : first_edge_seq,
2609 : tem_refs_not_in_seq, refs_not_supported,
2610 : true, fully_visited);
2611 3461 : if (eret != 1)
2612 : return -1;
2613 : /* Simplify our lives by pruning the sequence of !sm_ord. */
2614 3279 : while (!first_edge_seq.is_empty ()
2615 7922 : && first_edge_seq.last ().second != sm_ord)
2616 1182 : first_edge_seq.pop ();
2617 7059 : for (unsigned int i = 1; i < gimple_phi_num_args (phi); ++i)
2618 : {
2619 5513 : tree vuse = gimple_phi_arg_def (phi, i);
2620 5513 : edge e = gimple_phi_arg_edge (phi, i);
2621 5513 : auto_vec<seq_entry> edge_seq;
2622 5513 : bitmap_and_compl (tem_refs_not_in_seq,
2623 : refs_not_in_seq, refs_not_supported);
2624 : /* If we've marked all refs we search for as unsupported
2625 : we can stop processing and use the sequence as before
2626 : the PHI. */
2627 5513 : if (bitmap_empty_p (tem_refs_not_in_seq))
2628 : return 1;
2629 3598 : eret = sm_seq_valid_bb (loop, e->src, vuse, edge_seq,
2630 : tem_refs_not_in_seq, refs_not_supported,
2631 : true, fully_visited);
2632 3598 : if (eret != 1)
2633 : return -1;
2634 : /* Simplify our lives by pruning the sequence of !sm_ord. */
2635 3162 : while (!edge_seq.is_empty ()
2636 7810 : && edge_seq.last ().second != sm_ord)
2637 1050 : edge_seq.pop ();
2638 7196 : unsigned min_len = MIN(first_edge_seq.length (),
2639 : edge_seq.length ());
2640 : /* Incrementally merge seqs into first_edge_seq. */
2641 3598 : int first_uneq = -1;
2642 3598 : auto_vec<seq_entry, 2> extra_refs;
2643 5758 : for (unsigned int i = 0; i < min_len; ++i)
2644 : {
2645 : /* ??? We can more intelligently merge when we face different
2646 : order by additional sinking operations in one sequence.
2647 : For now we simply mark them as to be processed by the
2648 : not order-preserving SM code. */
2649 2160 : if (first_edge_seq[i].first != edge_seq[i].first)
2650 : {
2651 34 : if (first_edge_seq[i].second == sm_ord)
2652 21 : bitmap_set_bit (refs_not_supported,
2653 21 : first_edge_seq[i].first);
2654 34 : if (edge_seq[i].second == sm_ord)
2655 26 : bitmap_set_bit (refs_not_supported, edge_seq[i].first);
2656 34 : first_edge_seq[i].second = sm_other;
2657 34 : first_edge_seq[i].from = NULL_TREE;
2658 : /* Record the dropped refs for later processing. */
2659 34 : if (first_uneq == -1)
2660 32 : first_uneq = i;
2661 68 : extra_refs.safe_push (seq_entry (edge_seq[i].first,
2662 34 : sm_other, NULL_TREE));
2663 : }
2664 : /* sm_other prevails. */
2665 2126 : else if (first_edge_seq[i].second != edge_seq[i].second)
2666 : {
2667 : /* Make sure the ref is marked as not supported. */
2668 0 : bitmap_set_bit (refs_not_supported,
2669 0 : first_edge_seq[i].first);
2670 0 : first_edge_seq[i].second = sm_other;
2671 0 : first_edge_seq[i].from = NULL_TREE;
2672 : }
2673 2126 : else if (first_edge_seq[i].second == sm_other
2674 4 : && first_edge_seq[i].from != NULL_TREE
2675 2130 : && (edge_seq[i].from == NULL_TREE
2676 4 : || !operand_equal_p (first_edge_seq[i].from,
2677 4 : edge_seq[i].from, 0)))
2678 4 : first_edge_seq[i].from = NULL_TREE;
2679 : }
2680 : /* Any excess elements become sm_other since they are now
2681 : coonditionally executed. */
2682 9890 : if (first_edge_seq.length () > edge_seq.length ())
2683 : {
2684 6464 : for (unsigned i = edge_seq.length ();
2685 8269 : i < first_edge_seq.length (); ++i)
2686 : {
2687 6464 : if (first_edge_seq[i].second == sm_ord)
2688 3342 : bitmap_set_bit (refs_not_supported,
2689 3342 : first_edge_seq[i].first);
2690 6464 : first_edge_seq[i].second = sm_other;
2691 : }
2692 : }
2693 1793 : else if (edge_seq.length () > first_edge_seq.length ())
2694 : {
2695 5 : if (first_uneq == -1)
2696 0 : first_uneq = first_edge_seq.length ();
2697 5 : for (unsigned i = first_edge_seq.length ();
2698 20 : i < edge_seq.length (); ++i)
2699 : {
2700 15 : if (edge_seq[i].second == sm_ord)
2701 5 : bitmap_set_bit (refs_not_supported, edge_seq[i].first);
2702 30 : extra_refs.safe_push (seq_entry (edge_seq[i].first,
2703 15 : sm_other, NULL_TREE));
2704 : }
2705 : }
2706 : /* Put unmerged refs at first_uneq to force dependence checking
2707 : on them. */
2708 3598 : if (first_uneq != -1)
2709 : {
2710 : /* Missing ordered_splice_at. */
2711 64 : if ((unsigned)first_uneq == first_edge_seq.length ())
2712 0 : first_edge_seq.safe_splice (extra_refs);
2713 : else
2714 : {
2715 32 : unsigned fes_length = first_edge_seq.length ();
2716 32 : first_edge_seq.safe_grow (fes_length
2717 32 : + extra_refs.length ());
2718 32 : memmove (&first_edge_seq[first_uneq + extra_refs.length ()],
2719 32 : &first_edge_seq[first_uneq],
2720 32 : (fes_length - first_uneq) * sizeof (seq_entry));
2721 64 : memcpy (&first_edge_seq[first_uneq],
2722 32 : extra_refs.address (),
2723 32 : extra_refs.length () * sizeof (seq_entry));
2724 : }
2725 : }
2726 5513 : }
2727 : /* Use the sequence from the first edge and push SMs down. */
2728 4345 : for (unsigned i = 0; i < first_edge_seq.length (); ++i)
2729 : {
2730 2799 : unsigned id = first_edge_seq[i].first;
2731 2799 : seq.safe_push (first_edge_seq[i]);
2732 2799 : unsigned new_idx;
2733 2799 : if ((first_edge_seq[i].second == sm_ord
2734 2178 : || (first_edge_seq[i].second == sm_other
2735 2178 : && first_edge_seq[i].from != NULL_TREE))
2736 3770 : && !sm_seq_push_down (seq, seq.length () - 1, &new_idx))
2737 : {
2738 125 : if (first_edge_seq[i].second == sm_ord)
2739 0 : bitmap_set_bit (refs_not_supported, id);
2740 : /* Mark it sm_other. */
2741 125 : seq[new_idx].second = sm_other;
2742 125 : seq[new_idx].from = NULL_TREE;
2743 : }
2744 : }
2745 1546 : bitmap_set_bit (fully_visited,
2746 1546 : SSA_NAME_VERSION (gimple_phi_result (phi)));
2747 1546 : return 1;
2748 3461 : }
2749 42590 : lim_aux_data *data = get_lim_data (def);
2750 42590 : im_mem_ref *ref = memory_accesses.refs_list[data->ref];
2751 42590 : if (data->ref == UNANALYZABLE_MEM_ID)
2752 : return -1;
2753 : /* Stop at memory references which we can't move. */
2754 42590 : else if ((ref->mem.ref == error_mark_node
2755 : /* We can move end-of-storage/object down. */
2756 379 : && !gimple_clobber_p (ref->accesses_in_loop[0].stmt,
2757 : CLOBBER_STORAGE_END)
2758 244 : && !gimple_clobber_p (ref->accesses_in_loop[0].stmt,
2759 : CLOBBER_OBJECT_END))
2760 42881 : || TREE_THIS_VOLATILE (ref->mem.ref))
2761 : {
2762 : /* Mark refs_not_in_seq as unsupported. */
2763 120 : bitmap_ior_into (refs_not_supported, refs_not_in_seq);
2764 120 : return 1;
2765 : }
2766 : /* One of the stores we want to apply SM to and we've not yet seen. */
2767 42470 : else if (bitmap_clear_bit (refs_not_in_seq, data->ref))
2768 : {
2769 33512 : seq.safe_push (seq_entry (data->ref, sm_ord));
2770 :
2771 : /* 1) push it down the queue until a SMed
2772 : and not ignored ref is reached, skipping all not SMed refs
2773 : and ignored refs via non-TBAA disambiguation. */
2774 33512 : unsigned new_idx;
2775 67024 : if (!sm_seq_push_down (seq, seq.length () - 1, &new_idx)
2776 : /* If that fails but we did not fork yet continue, we'll see
2777 : to re-materialize all of the stores in the sequence then.
2778 : Further stores will only be pushed up to this one. */
2779 33512 : && forked)
2780 : {
2781 0 : bitmap_set_bit (refs_not_supported, data->ref);
2782 : /* Mark it sm_other. */
2783 0 : seq[new_idx].second = sm_other;
2784 : }
2785 :
2786 : /* 2) check whether we've seen all refs we want to SM and if so
2787 : declare success for the active exit */
2788 33512 : if (bitmap_empty_p (refs_not_in_seq))
2789 18997 : return 1;
2790 : }
2791 : else
2792 : /* Another store not part of the final sequence. Simply push it. */
2793 8958 : seq.safe_push (seq_entry (data->ref, sm_other,
2794 8958 : gimple_assign_rhs1 (def)));
2795 :
2796 23473 : vdef = gimple_vuse (def);
2797 23473 : }
2798 : while (1);
2799 : }
2800 :
2801 : /* Hoists memory references MEM_REFS out of LOOP. EXITS is the list of exit
2802 : edges of the LOOP. */
2803 :
2804 : static void
2805 21148 : hoist_memory_references (class loop *loop, bitmap mem_refs,
2806 : const vec<edge> &exits)
2807 : {
2808 21148 : im_mem_ref *ref;
2809 21148 : unsigned i;
2810 21148 : bitmap_iterator bi;
2811 :
2812 : /* There's a special case we can use ordered re-materialization for
2813 : conditionally executed stores which is when all stores in the loop
2814 : happen in the same basic-block. In that case we know we'll reach
2815 : all stores and thus can simply process that BB and emit a single
2816 : conditional block of ordered materializations. See PR102436. */
2817 21148 : basic_block single_store_bb = NULL;
2818 30509 : EXECUTE_IF_SET_IN_BITMAP (&memory_accesses.all_refs_stored_in_loop[loop->num],
2819 : 0, i, bi)
2820 : {
2821 28369 : bool fail = false;
2822 28369 : ref = memory_accesses.refs_list[i];
2823 130143 : for (auto loc : ref->accesses_in_loop)
2824 140363 : if (!gimple_vdef (loc.stmt))
2825 : ;
2826 31283 : else if (!single_store_bb)
2827 : {
2828 21148 : single_store_bb = gimple_bb (loc.stmt);
2829 21148 : bool conditional = false;
2830 76244 : for (edge e : exits)
2831 22685 : if (!dominated_by_p (CDI_DOMINATORS, e->src, single_store_bb))
2832 : {
2833 : /* Conditional as seen from e. */
2834 : conditional = true;
2835 : break;
2836 : }
2837 21148 : if (!conditional)
2838 : {
2839 : fail = true;
2840 : break;
2841 : }
2842 : }
2843 10135 : else if (single_store_bb != gimple_bb (loc.stmt))
2844 : {
2845 : fail = true;
2846 : break;
2847 : }
2848 28369 : if (fail)
2849 : {
2850 : single_store_bb = NULL;
2851 : break;
2852 : }
2853 : }
2854 21148 : if (single_store_bb)
2855 : {
2856 : /* Analyze the single block with stores. */
2857 2140 : auto_bitmap fully_visited;
2858 2140 : auto_bitmap refs_not_supported;
2859 2140 : auto_bitmap refs_not_in_seq;
2860 2140 : auto_vec<seq_entry> seq;
2861 2140 : bitmap_copy (refs_not_in_seq, mem_refs);
2862 2140 : int res = sm_seq_valid_bb (loop, single_store_bb, NULL_TREE,
2863 : seq, refs_not_in_seq, refs_not_supported,
2864 : false, fully_visited);
2865 2140 : if (res != 1)
2866 : {
2867 : /* Unhandled refs can still fail this. */
2868 0 : bitmap_clear (mem_refs);
2869 0 : return;
2870 : }
2871 :
2872 : /* We cannot handle sm_other since we neither remember the
2873 : stored location nor the value at the point we execute them. */
2874 5173 : for (unsigned i = 0; i < seq.length (); ++i)
2875 : {
2876 3033 : unsigned new_i;
2877 3033 : if (seq[i].second == sm_other
2878 3033 : && seq[i].from != NULL_TREE)
2879 414 : seq[i].from = NULL_TREE;
2880 2619 : else if ((seq[i].second == sm_ord
2881 0 : || (seq[i].second == sm_other
2882 0 : && seq[i].from != NULL_TREE))
2883 2619 : && !sm_seq_push_down (seq, i, &new_i))
2884 : {
2885 69 : bitmap_set_bit (refs_not_supported, seq[new_i].first);
2886 69 : seq[new_i].second = sm_other;
2887 69 : seq[new_i].from = NULL_TREE;
2888 : }
2889 : }
2890 2140 : bitmap_and_compl_into (mem_refs, refs_not_supported);
2891 2140 : if (bitmap_empty_p (mem_refs))
2892 : return;
2893 :
2894 : /* Prune seq. */
2895 2460 : while (seq.last ().second == sm_other
2896 2460 : && seq.last ().from == NULL_TREE)
2897 372 : seq.pop ();
2898 :
2899 2088 : hash_map<im_mem_ref *, sm_aux *> aux_map;
2900 :
2901 : /* Execute SM but delay the store materialization for ordered
2902 : sequences on exit. Remember a created flag var and make
2903 : sure to re-use it. */
2904 2088 : sm_aux *flag_var_aux = nullptr;
2905 4638 : EXECUTE_IF_SET_IN_BITMAP (mem_refs, 0, i, bi)
2906 : {
2907 2550 : ref = memory_accesses.refs_list[i];
2908 2550 : sm_aux *aux = execute_sm (loop, ref, aux_map, true,
2909 : flag_var_aux != nullptr);
2910 2550 : if (aux->store_flag)
2911 1453 : flag_var_aux = aux;
2912 : }
2913 :
2914 : /* Materialize ordered store sequences on exits. */
2915 2088 : edge e;
2916 2088 : auto_bitmap clobbers_to_prune;
2917 5655 : FOR_EACH_VEC_ELT (exits, i, e)
2918 : {
2919 3567 : edge append_cond_position = NULL;
2920 3567 : edge last_cond_fallthru = NULL;
2921 3567 : edge insert_e = e;
2922 : /* Construct the single flag variable control flow and insert
2923 : the ordered seq of stores in the then block. With
2924 : -fstore-data-races we can do the stores unconditionally. */
2925 3567 : if (flag_var_aux)
2926 2248 : insert_e
2927 : = single_pred_edge
2928 2248 : (execute_sm_if_changed (e, NULL_TREE, NULL_TREE,
2929 : flag_var_aux->store_flag,
2930 : loop_preheader_edge (loop),
2931 : &flag_var_aux->flag_bbs,
2932 : append_cond_position,
2933 : last_cond_fallthru));
2934 3567 : execute_sm_exit (loop, insert_e, seq, aux_map, sm_ord,
2935 : append_cond_position, last_cond_fallthru,
2936 : clobbers_to_prune);
2937 3567 : gsi_commit_one_edge_insert (insert_e, NULL);
2938 : }
2939 :
2940 : /* Remove clobbers inside the loop we re-materialized on exits. */
2941 2088 : EXECUTE_IF_SET_IN_BITMAP (clobbers_to_prune, 0, i, bi)
2942 : {
2943 0 : gimple *stmt = memory_accesses.refs_list[i]->accesses_in_loop[0].stmt;
2944 0 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
2945 0 : unlink_stmt_vdef (stmt);
2946 0 : release_defs (stmt);
2947 0 : gimple_set_vdef (stmt, NULL_TREE);
2948 0 : gsi_remove (&gsi, true);
2949 : }
2950 :
2951 4638 : for (hash_map<im_mem_ref *, sm_aux *>::iterator iter = aux_map.begin ();
2952 4638 : iter != aux_map.end (); ++iter)
2953 5100 : delete (*iter).second;
2954 :
2955 2088 : return;
2956 4228 : }
2957 :
2958 : /* To address PR57359 before actually applying store-motion check
2959 : the candidates found for validity with regards to reordering
2960 : relative to other stores which we until here disambiguated using
2961 : TBAA which isn't valid.
2962 : What matters is the order of the last stores to the mem_refs
2963 : with respect to the other stores of the loop at the point of the
2964 : loop exits. */
2965 :
2966 : /* For each exit compute the store order, pruning from mem_refs
2967 : on the fly. */
2968 : /* The complexity of this is at least
2969 : O(number of exits * number of SM refs) but more approaching
2970 : O(number of exits * number of SM refs * number of stores). */
2971 : /* ??? Somehow do this in a single sweep over the loop body. */
2972 19008 : auto_vec<std::pair<edge, vec<seq_entry> > > sms;
2973 19008 : auto_bitmap refs_not_supported (&lim_bitmap_obstack);
2974 19008 : edge e;
2975 38738 : FOR_EACH_VEC_ELT (exits, i, e)
2976 : {
2977 21628 : vec<seq_entry> seq;
2978 21628 : seq.create (4);
2979 21628 : auto_bitmap refs_not_in_seq (&lim_bitmap_obstack);
2980 21628 : bitmap_and_compl (refs_not_in_seq, mem_refs, refs_not_supported);
2981 21628 : if (bitmap_empty_p (refs_not_in_seq))
2982 : {
2983 1120 : seq.release ();
2984 1120 : break;
2985 : }
2986 20508 : auto_bitmap fully_visited;
2987 20508 : int res = sm_seq_valid_bb (loop, e->src, NULL_TREE,
2988 : seq, refs_not_in_seq,
2989 : refs_not_supported, false,
2990 : fully_visited);
2991 20508 : if (res != 1)
2992 : {
2993 778 : bitmap_copy (refs_not_supported, mem_refs);
2994 778 : seq.release ();
2995 778 : break;
2996 : }
2997 19730 : sms.safe_push (std::make_pair (e, seq));
2998 22406 : }
2999 :
3000 : /* Prune pruned mem_refs from earlier processed exits. */
3001 19008 : bool changed = !bitmap_empty_p (refs_not_supported);
3002 19008 : while (changed)
3003 : {
3004 5670 : changed = false;
3005 5670 : std::pair<edge, vec<seq_entry> > *seq;
3006 38672 : FOR_EACH_VEC_ELT (sms, i, seq)
3007 : {
3008 : bool need_to_push = false;
3009 13035 : for (unsigned i = 0; i < seq->second.length (); ++i)
3010 : {
3011 6038 : sm_kind kind = seq->second[i].second;
3012 6038 : if (kind == sm_other && seq->second[i].from == NULL_TREE)
3013 : break;
3014 5125 : unsigned id = seq->second[i].first;
3015 5125 : unsigned new_idx;
3016 5125 : if (kind == sm_ord
3017 5125 : && bitmap_bit_p (refs_not_supported, id))
3018 : {
3019 2546 : seq->second[i].second = sm_other;
3020 2546 : gcc_assert (seq->second[i].from == NULL_TREE);
3021 : need_to_push = true;
3022 : }
3023 2579 : else if (need_to_push
3024 2579 : && !sm_seq_push_down (seq->second, i, &new_idx))
3025 : {
3026 : /* We need to push down both sm_ord and sm_other
3027 : but for the latter we need to disqualify all
3028 : following refs. */
3029 113 : if (kind == sm_ord)
3030 : {
3031 7 : if (bitmap_set_bit (refs_not_supported, id))
3032 7 : changed = true;
3033 7 : seq->second[new_idx].second = sm_other;
3034 : }
3035 : else
3036 : {
3037 411 : for (unsigned j = seq->second.length () - 1;
3038 305 : j > new_idx; --j)
3039 199 : if (seq->second[j].second == sm_ord
3040 321 : && bitmap_set_bit (refs_not_supported,
3041 122 : seq->second[j].first))
3042 : changed = true;
3043 106 : seq->second.truncate (new_idx);
3044 106 : break;
3045 : }
3046 : }
3047 : }
3048 : }
3049 : }
3050 19008 : std::pair<edge, vec<seq_entry> > *seq;
3051 58468 : FOR_EACH_VEC_ELT (sms, i, seq)
3052 : {
3053 : /* Prune sm_other from the end. */
3054 63018 : while (!seq->second.is_empty ()
3055 43288 : && seq->second.last ().second == sm_other)
3056 4907 : seq->second.pop ();
3057 : /* Prune duplicates from the start. */
3058 19730 : auto_bitmap seen (&lim_bitmap_obstack);
3059 19730 : unsigned j, k;
3060 47386 : for (j = k = 0; j < seq->second.length (); ++j)
3061 27656 : if (bitmap_set_bit (seen, seq->second[j].first))
3062 : {
3063 27547 : if (k != j)
3064 77 : seq->second[k] = seq->second[j];
3065 27547 : ++k;
3066 : }
3067 19730 : seq->second.truncate (k);
3068 : /* And verify. */
3069 19730 : seq_entry *e;
3070 86737 : FOR_EACH_VEC_ELT (seq->second, j, e)
3071 27547 : gcc_assert (e->second == sm_ord
3072 : || (e->second == sm_other && e->from != NULL_TREE));
3073 19730 : }
3074 :
3075 : /* Verify dependence for refs we cannot handle with the order preserving
3076 : code (refs_not_supported) or prune them from mem_refs. */
3077 19008 : auto_vec<seq_entry> unord_refs;
3078 33726 : EXECUTE_IF_SET_IN_BITMAP (refs_not_supported, 0, i, bi)
3079 : {
3080 14718 : ref = memory_accesses.refs_list[i];
3081 14718 : if (!ref_indep_loop_p (loop, ref, sm_waw))
3082 6476 : bitmap_clear_bit (mem_refs, i);
3083 : /* We've now verified store order for ref with respect to all other
3084 : stores in the loop does not matter. */
3085 : else
3086 8242 : unord_refs.safe_push (seq_entry (i, sm_unord));
3087 : }
3088 :
3089 19008 : hash_map<im_mem_ref *, sm_aux *> aux_map;
3090 :
3091 : /* Execute SM but delay the store materialization for ordered
3092 : sequences on exit. */
3093 50757 : EXECUTE_IF_SET_IN_BITMAP (mem_refs, 0, i, bi)
3094 : {
3095 31749 : ref = memory_accesses.refs_list[i];
3096 31749 : execute_sm (loop, ref, aux_map, bitmap_bit_p (refs_not_supported, i),
3097 : false);
3098 : }
3099 :
3100 : /* Materialize ordered store sequences on exits. */
3101 19008 : auto_bitmap clobbers_to_prune;
3102 41810 : FOR_EACH_VEC_ELT (exits, i, e)
3103 : {
3104 22802 : edge append_cond_position = NULL;
3105 22802 : edge last_cond_fallthru = NULL;
3106 22802 : if (i < sms.length ())
3107 : {
3108 19730 : gcc_assert (sms[i].first == e);
3109 19730 : execute_sm_exit (loop, e, sms[i].second, aux_map, sm_ord,
3110 : append_cond_position, last_cond_fallthru,
3111 : clobbers_to_prune);
3112 19730 : sms[i].second.release ();
3113 : }
3114 22802 : if (!unord_refs.is_empty ())
3115 8410 : execute_sm_exit (loop, e, unord_refs, aux_map, sm_unord,
3116 : append_cond_position, last_cond_fallthru,
3117 : clobbers_to_prune);
3118 : /* Commit edge inserts here to preserve the order of stores
3119 : when an exit exits multiple loops. */
3120 22802 : gsi_commit_one_edge_insert (e, NULL);
3121 : }
3122 :
3123 : /* Remove clobbers inside the loop we re-materialized on exits. */
3124 19191 : EXECUTE_IF_SET_IN_BITMAP (clobbers_to_prune, 0, i, bi)
3125 : {
3126 183 : gimple *stmt = memory_accesses.refs_list[i]->accesses_in_loop[0].stmt;
3127 183 : gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
3128 183 : unlink_stmt_vdef (stmt);
3129 183 : release_defs (stmt);
3130 183 : gimple_set_vdef (stmt, NULL_TREE);
3131 183 : gsi_remove (&gsi, true);
3132 : }
3133 :
3134 50757 : for (hash_map<im_mem_ref *, sm_aux *>::iterator iter = aux_map.begin ();
3135 69765 : iter != aux_map.end (); ++iter)
3136 63498 : delete (*iter).second;
3137 19008 : }
3138 :
3139 : class ref_always_stored
3140 : {
3141 : public:
3142 91911 : ref_always_stored (class loop *loop_)
3143 91911 : : loop (loop_) {}
3144 : bool operator () (mem_ref_loc *loc);
3145 : class loop *loop;
3146 : };
3147 :
3148 : bool
3149 193214 : ref_always_stored::operator () (mem_ref_loc *loc)
3150 : {
3151 : /* Make sure the statement is a store. */
3152 193214 : tree lhs = gimple_get_lhs (loc->stmt);
3153 193214 : if (!lhs
3154 193214 : || !(DECL_P (lhs) || REFERENCE_CLASS_P (lhs)))
3155 : return false;
3156 :
3157 116578 : struct lim_aux_data *lim_data = get_lim_data (loc->stmt);
3158 116578 : if (!lim_data)
3159 : return false;
3160 :
3161 116578 : class loop *must_exec = lim_data->always_executed_in;
3162 116578 : if (!must_exec)
3163 : return false;
3164 :
3165 37364 : if (must_exec == loop
3166 37364 : || flow_loop_nested_p (must_exec, loop))
3167 34304 : return true;
3168 :
3169 : return false;
3170 : }
3171 :
3172 : /* Returns true if REF is always stored in LOOP. */
3173 :
3174 : static bool
3175 91911 : ref_always_stored_p (class loop *loop, im_mem_ref *ref)
3176 : {
3177 91911 : return for_all_locs_in_loop (loop, ref, ref_always_stored (loop));
3178 : }
3179 :
3180 : /* Returns true if LOAD_REF and STORE_REF form a "self write" pattern
3181 : where the stored value comes from the loaded value via SSA.
3182 : Example: a[i] = a[0] is safe to hoist a[0] even when i==0. */
3183 :
3184 : static bool
3185 325688 : is_self_write (im_mem_ref *load_ref, im_mem_ref *store_ref)
3186 : {
3187 : /* Only handle the simple case with a single access per ref.
3188 : Bail out on multiple accesses to be conservative. */
3189 325688 : if (load_ref->accesses_in_loop.length () != 1
3190 410501 : || store_ref->accesses_in_loop.length () != 1)
3191 : return false;
3192 :
3193 84830 : gimple *load_stmt = load_ref->accesses_in_loop[0].stmt;
3194 84830 : gimple *store_stmt = store_ref->accesses_in_loop[0].stmt;
3195 :
3196 84830 : if (!is_gimple_assign (load_stmt) || !is_gimple_assign (store_stmt))
3197 : return false;
3198 :
3199 84830 : tree loaded_val = gimple_assign_lhs (load_stmt);
3200 84830 : tree stored_val = gimple_assign_rhs1 (store_stmt);
3201 :
3202 84830 : if (TREE_CODE (loaded_val) != SSA_NAME || TREE_CODE (stored_val) != SSA_NAME)
3203 : return false;
3204 :
3205 : /* Self write: stored value is the loaded value. */
3206 61634 : if (stored_val != loaded_val)
3207 : return false;
3208 :
3209 :
3210 : /* TODO: Try to factor this out with mem_ref_hasher::equal
3211 : into im_compare_access_position_and_size
3212 : or a similar helper to centralize this delicate handling
3213 : complete for MEM_REF offsets and base pointer equality.
3214 :
3215 : TODO: Also ensure max_size_known_p agrees or resort to
3216 : alignment considerations to rule out partial overlaps.
3217 :
3218 : See:
3219 : https://gcc.gnu.org/pipermail/gcc-patches/2025-December/704155.html
3220 : For more context on TODOs above. */
3221 :
3222 : /* They may alias. Verify exact same location. */
3223 1735 : return (operand_equal_p (load_ref->mem.base, store_ref->mem.base, 0)
3224 46 : && known_eq (load_ref->mem.size, store_ref->mem.size)
3225 1781 : && known_eq (load_ref->mem.offset, store_ref->mem.offset));
3226 :
3227 : }
3228 :
3229 : /* Returns true if REF1 and REF2 are independent. */
3230 :
3231 : static bool
3232 604692 : refs_independent_p (im_mem_ref *ref1, im_mem_ref *ref2, bool tbaa_p)
3233 : {
3234 604692 : if (ref1 == ref2)
3235 : return true;
3236 :
3237 547389 : if (dump_file && (dump_flags & TDF_DETAILS))
3238 3043 : fprintf (dump_file, "Querying dependency of refs %u and %u: ",
3239 3043 : ref1->id, ref2->id);
3240 :
3241 547389 : if (mem_refs_may_alias_p (ref1, ref2, &memory_accesses.ttae_cache, tbaa_p))
3242 : {
3243 63586 : if (dump_file && (dump_flags & TDF_DETAILS))
3244 96 : fprintf (dump_file, "dependent.\n");
3245 : return false;
3246 : }
3247 : else
3248 : {
3249 483803 : if (dump_file && (dump_flags & TDF_DETAILS))
3250 2947 : fprintf (dump_file, "independent.\n");
3251 : return true;
3252 : }
3253 : }
3254 :
3255 : /* Returns true if REF is independent on all other accessess in LOOP.
3256 : KIND specifies the kind of dependence to consider.
3257 : lim_raw assumes REF is not stored in LOOP and disambiguates RAW
3258 : dependences so if true REF can be hoisted out of LOOP
3259 : sm_war disambiguates a store REF against all other loads to see
3260 : whether the store can be sunk across loads out of LOOP
3261 : sm_waw disambiguates a store REF against all other stores to see
3262 : whether the store can be sunk across stores out of LOOP. */
3263 :
3264 : static bool
3265 2101414 : ref_indep_loop_p (class loop *loop, im_mem_ref *ref, dep_kind kind)
3266 : {
3267 2101414 : bool indep_p = true;
3268 2101414 : bitmap refs_to_check;
3269 :
3270 2101414 : if (kind == sm_war)
3271 984363 : refs_to_check = &memory_accesses.refs_loaded_in_loop[loop->num];
3272 : else
3273 1117051 : refs_to_check = &memory_accesses.refs_stored_in_loop[loop->num];
3274 :
3275 2101414 : if (bitmap_bit_p (refs_to_check, UNANALYZABLE_MEM_ID)
3276 2101414 : || ref->mem.ref == error_mark_node)
3277 : indep_p = false;
3278 : else
3279 : {
3280 : /* tri-state, { unknown, independent, dependent } */
3281 664427 : dep_state state = query_loop_dependence (loop, ref, kind);
3282 664427 : if (state != dep_unknown)
3283 0 : return state == dep_independent;
3284 :
3285 664427 : class loop *inner = loop->inner;
3286 820338 : while (inner)
3287 : {
3288 453639 : if (!ref_indep_loop_p (inner, ref, kind))
3289 : {
3290 : indep_p = false;
3291 : break;
3292 : }
3293 155911 : inner = inner->next;
3294 : }
3295 :
3296 664427 : if (indep_p)
3297 : {
3298 366699 : unsigned i;
3299 366699 : bitmap_iterator bi;
3300 929125 : EXECUTE_IF_SET_IN_BITMAP (refs_to_check, 0, i, bi)
3301 : {
3302 630280 : im_mem_ref *aref = memory_accesses.refs_list[i];
3303 630280 : if (aref->mem.ref == error_mark_node)
3304 : {
3305 26368 : gimple *stmt = aref->accesses_in_loop[0].stmt;
3306 26368 : if ((kind == sm_war
3307 9477 : && ref_maybe_used_by_stmt_p (stmt, &ref->mem,
3308 : kind != sm_waw))
3309 35221 : || stmt_may_clobber_ref_p_1 (stmt, &ref->mem,
3310 : kind != sm_waw))
3311 : {
3312 : indep_p = false;
3313 : break;
3314 : }
3315 : }
3316 : /* For hoisting loads (lim_raw), allow "self write": the store
3317 : writes back the loaded value. Example: a[i] = a[0]
3318 : is safe even when i==0 causes aliasing. */
3319 603912 : else if (kind == lim_raw
3320 325688 : && ref->loaded && aref->stored
3321 929600 : && is_self_write (ref, aref))
3322 : {
3323 17 : if (dump_file && (dump_flags & TDF_DETAILS))
3324 2 : fprintf (dump_file,
3325 : "Dependency of refs %u and %u: "
3326 : "independent (self write).\n",
3327 2 : ref->id, aref->id);
3328 : }
3329 :
3330 603895 : else if (!refs_independent_p (ref, aref, kind != sm_waw))
3331 : {
3332 : indep_p = false;
3333 : break;
3334 : }
3335 : }
3336 : }
3337 : }
3338 :
3339 2101414 : if (dump_file && (dump_flags & TDF_DETAILS))
3340 769 : fprintf (dump_file, "Querying %s dependencies of ref %u in loop %d: %s\n",
3341 : kind == lim_raw ? "RAW" : (kind == sm_war ? "SM WAR" : "SM WAW"),
3342 374 : ref->id, loop->num, indep_p ? "independent" : "dependent");
3343 :
3344 : /* Record the computed result in the cache. */
3345 2101414 : record_loop_dependence (loop, ref, kind,
3346 : indep_p ? dep_independent : dep_dependent);
3347 :
3348 2101414 : return indep_p;
3349 : }
3350 :
3351 : class ref_in_loop_hot_body
3352 : {
3353 : public:
3354 41832 : ref_in_loop_hot_body (class loop *loop_) : l (loop_) {}
3355 : bool operator () (mem_ref_loc *loc);
3356 : class loop *l;
3357 : };
3358 :
3359 : /* Check the coldest loop between loop L and innermost loop. If there is one
3360 : cold loop between L and INNER_LOOP, store motion can be performed, otherwise
3361 : no cold loop means no store motion. get_coldest_out_loop also handles cases
3362 : when l is inner_loop. */
3363 : bool
3364 42525 : ref_in_loop_hot_body::operator () (mem_ref_loc *loc)
3365 : {
3366 42525 : basic_block curr_bb = gimple_bb (loc->stmt);
3367 42525 : class loop *inner_loop = curr_bb->loop_father;
3368 42525 : return get_coldest_out_loop (l, inner_loop, curr_bb);
3369 : }
3370 :
3371 :
3372 : /* Returns true if we can perform store motion of REF from LOOP. */
3373 :
3374 : static bool
3375 2937752 : can_sm_ref_p (class loop *loop, im_mem_ref *ref)
3376 : {
3377 2937752 : bool store_could_trap;
3378 :
3379 : /* Can't hoist unanalyzable refs. */
3380 2937752 : if (!MEM_ANALYZABLE (ref))
3381 : return false;
3382 :
3383 : /* Can't hoist/sink aggregate copies. */
3384 2574853 : if (ref->mem.ref == error_mark_node)
3385 : return false;
3386 :
3387 : /* It should be movable. */
3388 2131001 : if (!is_gimple_reg_type (TREE_TYPE (ref->mem.ref))
3389 2130715 : || TREE_THIS_VOLATILE (ref->mem.ref)
3390 4241976 : || !for_each_index (&ref->mem.ref, may_move_till, loop))
3391 : return false;
3392 :
3393 962912 : store_could_trap = lhs_could_trap_p (ref->mem.ref);
3394 :
3395 : /* Do not move a load or store that can throw. Store motion does not
3396 : update EH information. */
3397 962912 : if (tree_could_throw_p (ref->mem.ref)
3398 962912 : || (flag_exceptions
3399 689179 : && cfun->can_throw_non_call_exceptions
3400 541292 : && store_could_trap))
3401 : return false;
3402 :
3403 : /* If a store can trap, it must be always executed in LOOP. */
3404 939611 : if (store_could_trap
3405 : /* ??? We can at least use false here, allowing loads? We
3406 : are forcing conditional stores if the ref is not always
3407 : stored to later anyway. So this would only guard
3408 : the load we need to emit. Thus when the ref is not
3409 : loaded we can elide this completely? */
3410 997223 : && !ref_always_stored_p (loop, ref))
3411 : return false;
3412 :
3413 : /* Verify all loads of ref can be hoisted. */
3414 892906 : if (ref->loaded
3415 90352 : && bitmap_bit_p (ref->loaded, loop->num)
3416 978102 : && !ref_indep_loop_p (loop, ref, lim_raw))
3417 : return false;
3418 :
3419 : /* Verify the candidate can be disambiguated against all loads,
3420 : that is, we can elide all in-loop stores. Disambiguation
3421 : against stores is done later when we cannot guarantee preserving
3422 : the order of stores. */
3423 829874 : if (!ref_indep_loop_p (loop, ref, sm_war))
3424 : return false;
3425 :
3426 : /* Verify whether the candidate is hot for LOOP. Only do store motion if the
3427 : candidate's profile count is hot. Statement in cold BB shouldn't be moved
3428 : out of it's loop_father. */
3429 41832 : if (!for_all_locs_in_loop (loop, ref, ref_in_loop_hot_body (loop)))
3430 : return false;
3431 :
3432 : return true;
3433 : }
3434 :
3435 : /* Marks the references in LOOP for that store motion should be performed
3436 : in REFS_TO_SM. SM_EXECUTED is the set of references for that store
3437 : motion was performed in one of the outer loops. */
3438 :
3439 : static void
3440 1297601 : find_refs_for_sm (class loop *loop, bitmap sm_executed, bitmap refs_to_sm)
3441 : {
3442 1297601 : bitmap refs = &memory_accesses.all_refs_stored_in_loop[loop->num];
3443 1297601 : unsigned i;
3444 1297601 : bitmap_iterator bi;
3445 1297601 : im_mem_ref *ref;
3446 :
3447 4235353 : EXECUTE_IF_AND_COMPL_IN_BITMAP (refs, sm_executed, 0, i, bi)
3448 : {
3449 2937752 : ref = memory_accesses.refs_list[i];
3450 2937752 : if (can_sm_ref_p (loop, ref) && dbg_cnt (lim))
3451 40846 : bitmap_set_bit (refs_to_sm, i);
3452 : }
3453 1297601 : }
3454 :
3455 : /* Checks whether LOOP (with exits stored in EXITS array) is suitable
3456 : for a store motion optimization (i.e. whether we can insert statement
3457 : on its exits). */
3458 :
3459 : static bool
3460 1343287 : loop_suitable_for_sm (class loop *loop ATTRIBUTE_UNUSED,
3461 : const vec<edge> &exits)
3462 : {
3463 1343287 : unsigned i;
3464 1343287 : edge ex;
3465 :
3466 3488400 : FOR_EACH_VEC_ELT (exits, i, ex)
3467 2190799 : if (ex->flags & (EDGE_ABNORMAL | EDGE_EH))
3468 : return false;
3469 :
3470 : return true;
3471 : }
3472 :
3473 : /* Try to perform store motion for all memory references modified inside
3474 : LOOP. SM_EXECUTED is the bitmap of the memory references for that
3475 : store motion was executed in one of the outer loops. */
3476 :
3477 : static void
3478 1343287 : store_motion_loop (class loop *loop, bitmap sm_executed)
3479 : {
3480 1343287 : auto_vec<edge> exits = get_loop_exit_edges (loop);
3481 1343287 : class loop *subloop;
3482 1343287 : bitmap sm_in_loop = BITMAP_ALLOC (&lim_bitmap_obstack);
3483 :
3484 1343287 : if (loop_suitable_for_sm (loop, exits))
3485 : {
3486 1297601 : find_refs_for_sm (loop, sm_executed, sm_in_loop);
3487 1297601 : if (!bitmap_empty_p (sm_in_loop))
3488 21148 : hoist_memory_references (loop, sm_in_loop, exits);
3489 : }
3490 :
3491 1343287 : bitmap_ior_into (sm_executed, sm_in_loop);
3492 1646926 : for (subloop = loop->inner; subloop != NULL; subloop = subloop->next)
3493 303639 : store_motion_loop (subloop, sm_executed);
3494 1343287 : bitmap_and_compl_into (sm_executed, sm_in_loop);
3495 1343287 : BITMAP_FREE (sm_in_loop);
3496 1343287 : }
3497 :
3498 : /* Try to perform store motion for all memory references modified inside
3499 : loops. */
3500 :
3501 : static void
3502 496745 : do_store_motion (void)
3503 : {
3504 496745 : class loop *loop;
3505 496745 : bitmap sm_executed = BITMAP_ALLOC (&lim_bitmap_obstack);
3506 :
3507 1536393 : for (loop = current_loops->tree_root->inner; loop != NULL; loop = loop->next)
3508 1039648 : store_motion_loop (loop, sm_executed);
3509 :
3510 496745 : BITMAP_FREE (sm_executed);
3511 496745 : }
3512 :
3513 : /* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e.
3514 : for each such basic block bb records the outermost loop for that execution
3515 : of its header implies execution of bb. CONTAINS_CALL is the bitmap of
3516 : blocks that contain a nonpure call. */
3517 :
3518 : static void
3519 1350631 : fill_always_executed_in_1 (class loop *loop, sbitmap contains_call)
3520 : {
3521 1350631 : basic_block bb = NULL, last = NULL;
3522 1350631 : edge e;
3523 1350631 : class loop *inn_loop = loop;
3524 :
3525 1656635 : for (class loop *inner = loop->inner; inner; inner = inner->next)
3526 306004 : fill_always_executed_in_1 (inner, contains_call);
3527 :
3528 1350631 : auto_vec<basic_block, 64> worklist;
3529 1350631 : worklist.reserve_exact (loop->num_nodes);
3530 1350631 : worklist.quick_push (loop->header);
3531 3034530 : do
3532 : {
3533 3034530 : edge_iterator ei;
3534 3034530 : bb = worklist.pop ();
3535 :
3536 3034530 : if (!flow_bb_inside_loop_p (inn_loop, bb))
3537 : {
3538 : /* When we are leaving a possibly infinite inner loop
3539 : we have to stop processing. */
3540 188342 : if (!finite_loop_p (inn_loop))
3541 : break;
3542 : /* If the loop was finite we can continue with processing
3543 : the loop we exited to. */
3544 179065 : inn_loop = bb->loop_father;
3545 : }
3546 :
3547 3025253 : if (dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
3548 1744141 : last = bb;
3549 :
3550 3025253 : if (bitmap_bit_p (contains_call, bb->index))
3551 : break;
3552 :
3553 : /* If LOOP exits from this BB stop processing. */
3554 5858422 : FOR_EACH_EDGE (e, ei, bb->succs)
3555 4163708 : if (!flow_bb_inside_loop_p (loop, e->dest))
3556 : break;
3557 2683084 : if (e)
3558 : break;
3559 :
3560 : /* A loop might be infinite (TODO use simple loop analysis
3561 : to disprove this if possible). */
3562 1694714 : if (bb->flags & BB_IRREDUCIBLE_LOOP)
3563 : break;
3564 :
3565 1694586 : if (bb->loop_father->header == bb)
3566 : /* Record that we enter into a subloop since it might not
3567 : be finite. */
3568 : /* ??? Entering into a not always executed subloop makes
3569 : fill_always_executed_in quadratic in loop depth since
3570 : we walk those loops N times. This is not a problem
3571 : in practice though, see PR102253 for a worst-case testcase. */
3572 632504 : inn_loop = bb->loop_father;
3573 :
3574 : /* Walk the body of LOOP sorted by dominance relation. Additionally,
3575 : if a basic block S dominates the latch, then only blocks dominated
3576 : by S are after it.
3577 : This is get_loop_body_in_dom_order using a worklist algorithm and
3578 : stopping once we are no longer interested in visiting further
3579 : blocks. */
3580 1694586 : unsigned old_len = worklist.length ();
3581 1694586 : unsigned postpone = 0;
3582 1694586 : for (basic_block son = first_dom_son (CDI_DOMINATORS, bb);
3583 3674721 : son;
3584 1980135 : son = next_dom_son (CDI_DOMINATORS, son))
3585 : {
3586 1980135 : if (!flow_bb_inside_loop_p (loop, son))
3587 22377 : continue;
3588 1957758 : if (dominated_by_p (CDI_DOMINATORS, loop->latch, son))
3589 540413 : postpone = worklist.length ();
3590 1957758 : worklist.quick_push (son);
3591 : }
3592 1694586 : if (postpone)
3593 : /* Postponing the block that dominates the latch means
3594 : processing it last and thus putting it earliest in the
3595 : worklist. */
3596 110544 : std::swap (worklist[old_len], worklist[postpone]);
3597 : }
3598 3389172 : while (!worklist.is_empty ());
3599 :
3600 2137651 : while (1)
3601 : {
3602 1744141 : if (last->loop_father == loop
3603 1744141 : || (ALWAYS_EXECUTED_IN (last)
3604 158966 : && loop_outer (ALWAYS_EXECUTED_IN (last)) == loop))
3605 : {
3606 1743465 : if (dump_enabled_p ())
3607 2477 : dump_printf (MSG_NOTE, "BB %d is always executed in loop %d\n",
3608 : last->index, loop->num);
3609 1743465 : SET_ALWAYS_EXECUTED_IN (last, loop);
3610 : }
3611 1744141 : if (last == loop->header)
3612 : break;
3613 393510 : last = get_immediate_dominator (CDI_DOMINATORS, last);
3614 : }
3615 1350631 : }
3616 :
3617 : /* Fills ALWAYS_EXECUTED_IN information for basic blocks, i.e.
3618 : for each such basic block bb records the outermost loop for that execution
3619 : of its header implies execution of bb. */
3620 :
3621 : static void
3622 498242 : fill_always_executed_in (void)
3623 : {
3624 498242 : basic_block bb;
3625 498242 : class loop *loop;
3626 :
3627 498242 : auto_sbitmap contains_call (last_basic_block_for_fn (cfun));
3628 498242 : bitmap_clear (contains_call);
3629 15473690 : FOR_EACH_BB_FN (bb, cfun)
3630 : {
3631 14975448 : if (loop_depth (bb->loop_father) == 0)
3632 : {
3633 : /* Outside of loops we can skip scanning all stmts. */
3634 8687793 : bitmap_set_bit (contains_call, bb->index);
3635 8687793 : continue;
3636 : }
3637 6287655 : gimple_stmt_iterator gsi;
3638 54666858 : for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
3639 : {
3640 44228327 : if (nonpure_call_p (gsi_stmt (gsi)))
3641 : break;
3642 : }
3643 :
3644 6287655 : if (!gsi_end_p (gsi))
3645 1005621 : bitmap_set_bit (contains_call, bb->index);
3646 : }
3647 :
3648 1542869 : for (loop = current_loops->tree_root->inner; loop; loop = loop->next)
3649 1044627 : fill_always_executed_in_1 (loop, contains_call);
3650 498242 : }
3651 :
3652 : /* Find the coldest loop preheader for LOOP, also find the nearest hotter loop
3653 : to LOOP. Then recursively iterate each inner loop. */
3654 :
3655 : void
3656 1350631 : fill_coldest_and_hotter_out_loop (class loop *coldest_loop,
3657 : class loop *hotter_loop, class loop *loop)
3658 : {
3659 1350631 : if (bb_colder_than_loop_preheader (loop_preheader_edge (loop)->src,
3660 : coldest_loop))
3661 36189 : coldest_loop = loop;
3662 :
3663 1350631 : coldest_outermost_loop[loop->num] = coldest_loop;
3664 :
3665 1350631 : hotter_than_inner_loop[loop->num] = NULL;
3666 1350631 : class loop *outer_loop = loop_outer (loop);
3667 1350631 : if (hotter_loop
3668 1350631 : && bb_colder_than_loop_preheader (loop_preheader_edge (loop)->src,
3669 : hotter_loop))
3670 4448 : hotter_than_inner_loop[loop->num] = hotter_loop;
3671 :
3672 1350631 : if (outer_loop && outer_loop != current_loops->tree_root
3673 1656635 : && bb_colder_than_loop_preheader (loop_preheader_edge (loop)->src,
3674 : outer_loop))
3675 38222 : hotter_than_inner_loop[loop->num] = outer_loop;
3676 :
3677 1350631 : if (dump_enabled_p ())
3678 : {
3679 1948 : dump_printf (MSG_NOTE, "loop %d's coldest_outermost_loop is %d, ",
3680 : loop->num, coldest_loop->num);
3681 1948 : if (hotter_than_inner_loop[loop->num])
3682 319 : dump_printf (MSG_NOTE, "hotter_than_inner_loop is %d\n",
3683 319 : hotter_than_inner_loop[loop->num]->num);
3684 : else
3685 1629 : dump_printf (MSG_NOTE, "hotter_than_inner_loop is NULL\n");
3686 : }
3687 :
3688 1350631 : class loop *inner_loop;
3689 1656635 : for (inner_loop = loop->inner; inner_loop; inner_loop = inner_loop->next)
3690 612008 : fill_coldest_and_hotter_out_loop (coldest_loop,
3691 306004 : hotter_than_inner_loop[loop->num],
3692 : inner_loop);
3693 1350631 : }
3694 :
3695 : /* Compute the global information needed by the loop invariant motion pass. */
3696 :
3697 : static void
3698 498242 : tree_ssa_lim_initialize (bool store_motion)
3699 : {
3700 498242 : unsigned i;
3701 :
3702 498242 : bitmap_obstack_initialize (&lim_bitmap_obstack);
3703 498242 : gcc_obstack_init (&mem_ref_obstack);
3704 498242 : lim_aux_data_map = new hash_map<gimple *, lim_aux_data *>;
3705 :
3706 498242 : if (flag_tm)
3707 104 : compute_transaction_bits ();
3708 :
3709 498242 : memory_accesses.refs = new hash_table<mem_ref_hasher> (100);
3710 498242 : memory_accesses.refs_list.create (100);
3711 : /* Allocate a special, unanalyzable mem-ref with ID zero. */
3712 498242 : memory_accesses.refs_list.quick_push
3713 498242 : (mem_ref_alloc (NULL, 0, UNANALYZABLE_MEM_ID));
3714 :
3715 996484 : memory_accesses.refs_loaded_in_loop.create (number_of_loops (cfun));
3716 996484 : memory_accesses.refs_loaded_in_loop.quick_grow_cleared (number_of_loops (cfun));
3717 996484 : memory_accesses.refs_stored_in_loop.create (number_of_loops (cfun));
3718 996484 : memory_accesses.refs_stored_in_loop.quick_grow_cleared (number_of_loops (cfun));
3719 498242 : if (store_motion)
3720 : {
3721 993490 : memory_accesses.all_refs_stored_in_loop.create (number_of_loops (cfun));
3722 496745 : memory_accesses.all_refs_stored_in_loop.quick_grow_cleared
3723 993490 : (number_of_loops (cfun));
3724 : }
3725 :
3726 6187092 : for (i = 0; i < number_of_loops (cfun); i++)
3727 : {
3728 2595304 : bitmap_initialize (&memory_accesses.refs_loaded_in_loop[i],
3729 : &lim_bitmap_obstack);
3730 2595304 : bitmap_initialize (&memory_accesses.refs_stored_in_loop[i],
3731 : &lim_bitmap_obstack);
3732 2595304 : if (store_motion)
3733 2582944 : bitmap_initialize (&memory_accesses.all_refs_stored_in_loop[i],
3734 : &lim_bitmap_obstack);
3735 : }
3736 :
3737 498242 : memory_accesses.ttae_cache = NULL;
3738 :
3739 : /* Initialize bb_loop_postorder with a mapping from loop->num to
3740 : its postorder index. */
3741 498242 : i = 0;
3742 996484 : bb_loop_postorder = XNEWVEC (unsigned, number_of_loops (cfun));
3743 2845357 : for (auto loop : loops_list (cfun, LI_FROM_INNERMOST))
3744 1350631 : bb_loop_postorder[loop->num] = i++;
3745 498242 : }
3746 :
3747 : /* Cleans up after the invariant motion pass. */
3748 :
3749 : static void
3750 498242 : tree_ssa_lim_finalize (void)
3751 : {
3752 498242 : basic_block bb;
3753 498242 : unsigned i;
3754 498242 : im_mem_ref *ref;
3755 :
3756 15512765 : FOR_EACH_BB_FN (bb, cfun)
3757 15014523 : SET_ALWAYS_EXECUTED_IN (bb, NULL);
3758 :
3759 498242 : bitmap_obstack_release (&lim_bitmap_obstack);
3760 996484 : delete lim_aux_data_map;
3761 :
3762 498242 : delete memory_accesses.refs;
3763 498242 : memory_accesses.refs = NULL;
3764 :
3765 5942972 : FOR_EACH_VEC_ELT (memory_accesses.refs_list, i, ref)
3766 5444730 : memref_free (ref);
3767 498242 : memory_accesses.refs_list.release ();
3768 498242 : obstack_free (&mem_ref_obstack, NULL);
3769 :
3770 498242 : memory_accesses.refs_loaded_in_loop.release ();
3771 498242 : memory_accesses.refs_stored_in_loop.release ();
3772 498242 : memory_accesses.all_refs_stored_in_loop.release ();
3773 :
3774 498242 : if (memory_accesses.ttae_cache)
3775 4877 : free_affine_expand_cache (&memory_accesses.ttae_cache);
3776 :
3777 498242 : free (bb_loop_postorder);
3778 :
3779 498242 : coldest_outermost_loop.release ();
3780 498242 : hotter_than_inner_loop.release ();
3781 498242 : }
3782 :
3783 : /* Moves invariants from loops. Only "expensive" invariants are moved out --
3784 : i.e. those that are likely to be win regardless of the register pressure.
3785 : Only perform store motion if STORE_MOTION is true. */
3786 :
3787 : unsigned int
3788 498242 : loop_invariant_motion_in_fun (function *fun, bool store_motion)
3789 : {
3790 498242 : unsigned int todo = 0;
3791 :
3792 498242 : tree_ssa_lim_initialize (store_motion);
3793 :
3794 498242 : mark_ssa_maybe_undefs ();
3795 :
3796 : /* Gathers information about memory accesses in the loops. */
3797 498242 : analyze_memory_references (store_motion);
3798 :
3799 : /* Fills ALWAYS_EXECUTED_IN information for basic blocks. */
3800 498242 : fill_always_executed_in ();
3801 :
3802 : /* Pre-compute coldest outermost loop and nearest hotter loop of each loop.
3803 : */
3804 498242 : class loop *loop;
3805 996484 : coldest_outermost_loop.create (number_of_loops (cfun));
3806 996484 : coldest_outermost_loop.safe_grow_cleared (number_of_loops (cfun));
3807 996484 : hotter_than_inner_loop.create (number_of_loops (cfun));
3808 996484 : hotter_than_inner_loop.safe_grow_cleared (number_of_loops (cfun));
3809 1542869 : for (loop = current_loops->tree_root->inner; loop != NULL; loop = loop->next)
3810 1044627 : fill_coldest_and_hotter_out_loop (loop, NULL, loop);
3811 :
3812 498242 : int *rpo = XNEWVEC (int, last_basic_block_for_fn (fun));
3813 498242 : int n = pre_and_rev_post_order_compute_fn (fun, NULL, rpo, false);
3814 :
3815 : /* For each statement determine the outermost loop in that it is
3816 : invariant and cost for computing the invariant. */
3817 15971932 : for (int i = 0; i < n; ++i)
3818 14975448 : compute_invariantness (BASIC_BLOCK_FOR_FN (fun, rpo[i]));
3819 :
3820 : /* Execute store motion. Force the necessary invariants to be moved
3821 : out of the loops as well. */
3822 498242 : if (store_motion)
3823 496745 : do_store_motion ();
3824 :
3825 498242 : free (rpo);
3826 498242 : rpo = XNEWVEC (int, last_basic_block_for_fn (fun));
3827 498242 : n = pre_and_rev_post_order_compute_fn (fun, NULL, rpo, false);
3828 :
3829 : /* Move the expressions that are expensive enough. */
3830 16011007 : for (int i = 0; i < n; ++i)
3831 15014523 : todo |= move_computations_worker (BASIC_BLOCK_FOR_FN (fun, rpo[i]));
3832 :
3833 498242 : free (rpo);
3834 :
3835 498242 : gsi_commit_edge_inserts ();
3836 498242 : if (need_ssa_update_p (fun))
3837 14680 : rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
3838 :
3839 498242 : tree_ssa_lim_finalize ();
3840 :
3841 498242 : return todo;
3842 : }
3843 :
3844 : /* Loop invariant motion pass. */
3845 :
3846 : namespace {
3847 :
3848 : const pass_data pass_data_lim =
3849 : {
3850 : GIMPLE_PASS, /* type */
3851 : "lim", /* name */
3852 : OPTGROUP_LOOP, /* optinfo_flags */
3853 : TV_LIM, /* tv_id */
3854 : PROP_cfg, /* properties_required */
3855 : 0, /* properties_provided */
3856 : 0, /* properties_destroyed */
3857 : 0, /* todo_flags_start */
3858 : 0, /* todo_flags_finish */
3859 : };
3860 :
3861 : class pass_lim : public gimple_opt_pass
3862 : {
3863 : public:
3864 1178144 : pass_lim (gcc::context *ctxt)
3865 2356288 : : gimple_opt_pass (pass_data_lim, ctxt)
3866 : {}
3867 :
3868 : /* opt_pass methods: */
3869 883608 : opt_pass * clone () final override { return new pass_lim (m_ctxt); }
3870 1317559 : bool gate (function *) final override { return flag_tree_loop_im != 0; }
3871 : unsigned int execute (function *) final override;
3872 :
3873 : }; // class pass_lim
3874 :
3875 : unsigned int
3876 1317196 : pass_lim::execute (function *fun)
3877 : {
3878 1317196 : in_loop_pipeline = scev_initialized_p ();
3879 1317196 : if (!in_loop_pipeline)
3880 1069032 : loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
3881 :
3882 2634392 : if (number_of_loops (fun) <= 1)
3883 : return 0;
3884 496763 : unsigned int todo = loop_invariant_motion_in_fun (fun, flag_move_loop_stores);
3885 :
3886 496763 : if (!in_loop_pipeline)
3887 248599 : loop_optimizer_finalize ();
3888 : else
3889 248164 : scev_reset ();
3890 : return todo;
3891 : }
3892 :
3893 : } // anon namespace
3894 :
3895 : gimple_opt_pass *
3896 294536 : make_pass_lim (gcc::context *ctxt)
3897 : {
3898 294536 : return new pass_lim (ctxt);
3899 : }
3900 :
|
