Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-ssa-loop-im.c @ 146:351920fa3827
merge
author | anatofuz <anatofuz@cr.ie.u-ryukyu.ac.jp> |
---|---|
date | Sun, 01 Mar 2020 16:13:28 +0900 |
parents | 1830386684a0 |
children |
rev | line source |
---|---|
0 | 1 /* Loop invariant motion. |
145 | 2 Copyright (C) 2003-2020 Free Software Foundation, Inc. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
3 |
0 | 4 This file is part of GCC. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
5 |
0 | 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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
10 |
0 | 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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
15 |
0 | 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" | |
111 | 23 #include "backend.h" |
0 | 24 #include "tree.h" |
111 | 25 #include "gimple.h" |
26 #include "cfghooks.h" | |
27 #include "tree-pass.h" | |
28 #include "ssa.h" | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
29 #include "gimple-pretty-print.h" |
111 | 30 #include "fold-const.h" |
31 #include "cfganal.h" | |
32 #include "tree-eh.h" | |
33 #include "gimplify.h" | |
34 #include "gimple-iterator.h" | |
35 #include "tree-cfg.h" | |
36 #include "tree-ssa-loop-manip.h" | |
37 #include "tree-ssa-loop.h" | |
38 #include "tree-into-ssa.h" | |
0 | 39 #include "cfgloop.h" |
40 #include "domwalk.h" | |
41 #include "tree-affine.h" | |
42 #include "tree-ssa-propagate.h" | |
111 | 43 #include "trans-mem.h" |
44 #include "gimple-fold.h" | |
45 #include "tree-scalar-evolution.h" | |
46 #include "tree-ssa-loop-niter.h" | |
145 | 47 #include "alias.h" |
48 #include "builtins.h" | |
49 #include "tree-dfa.h" | |
0 | 50 |
51 /* TODO: Support for predicated code motion. I.e. | |
52 | |
53 while (1) | |
54 { | |
55 if (cond) | |
56 { | |
57 a = inv; | |
58 something; | |
59 } | |
60 } | |
61 | |
111 | 62 Where COND and INV are invariants, but evaluating INV may trap or be |
0 | 63 invalid from some other reason if !COND. This may be transformed to |
64 | |
65 if (cond) | |
66 a = inv; | |
67 while (1) | |
68 { | |
69 if (cond) | |
70 something; | |
71 } */ | |
72 | |
73 /* The auxiliary data kept for each statement. */ | |
74 | |
75 struct lim_aux_data | |
76 { | |
145 | 77 class loop *max_loop; /* The outermost loop in that the statement |
0 | 78 is invariant. */ |
79 | |
145 | 80 class loop *tgt_loop; /* The loop out of that we want to move the |
0 | 81 invariant. */ |
82 | |
145 | 83 class loop *always_executed_in; |
0 | 84 /* The outermost loop for that we are sure |
85 the statement is executed if the loop | |
86 is entered. */ | |
87 | |
88 unsigned cost; /* Cost of the computation performed by the | |
89 statement. */ | |
90 | |
111 | 91 unsigned ref; /* The simple_mem_ref in this stmt or 0. */ |
92 | |
93 vec<gimple *> depends; /* Vector of statements that must be also | |
94 hoisted out of the loop when this statement | |
95 is hoisted; i.e. those that define the | |
96 operands of the statement and are inside of | |
97 the MAX_LOOP loop. */ | |
0 | 98 }; |
99 | |
100 /* Maps statements to their lim_aux_data. */ | |
101 | |
111 | 102 static hash_map<gimple *, lim_aux_data *> *lim_aux_data_map; |
0 | 103 |
104 /* Description of a memory reference location. */ | |
105 | |
111 | 106 struct mem_ref_loc |
0 | 107 { |
108 tree *ref; /* The reference itself. */ | |
111 | 109 gimple *stmt; /* The statement in that it occurs. */ |
110 }; | |
111 | |
0 | 112 |
113 /* Description of a memory reference. */ | |
114 | |
145 | 115 class im_mem_ref |
0 | 116 { |
145 | 117 public: |
118 unsigned id : 30; /* ID assigned to the memory reference | |
0 | 119 (its index in memory_accesses.refs_list) */ |
145 | 120 unsigned ref_canonical : 1; /* Whether mem.ref was canonicalized. */ |
121 unsigned ref_decomposed : 1; /* Whether the ref was hashed from mem. */ | |
0 | 122 hashval_t hash; /* Its hash value. */ |
111 | 123 |
124 /* The memory access itself and associated caching of alias-oracle | |
125 query meta-data. */ | |
126 ao_ref mem; | |
127 | |
0 | 128 bitmap stored; /* The set of loops in that this memory location |
129 is stored to. */ | |
111 | 130 vec<mem_ref_loc> accesses_in_loop; |
0 | 131 /* The locations of the accesses. Vector |
132 indexed by the loop number. */ | |
133 | |
134 /* The following sets are computed on demand. We keep both set and | |
135 its complement, so that we know whether the information was | |
136 already computed or not. */ | |
111 | 137 bitmap_head indep_loop; /* The set of loops in that the memory |
0 | 138 reference is independent, meaning: |
139 If it is stored in the loop, this store | |
140 is independent on all other loads and | |
141 stores. | |
142 If it is only loaded, then it is independent | |
143 on all stores in the loop. */ | |
111 | 144 bitmap_head dep_loop; /* The complement of INDEP_LOOP. */ |
145 }; | |
146 | |
147 /* We use two bits per loop in the ref->{in,}dep_loop bitmaps, the first | |
148 to record (in)dependence against stores in the loop and its subloops, the | |
149 second to record (in)dependence against all references in the loop | |
150 and its subloops. */ | |
151 #define LOOP_DEP_BIT(loopnum, storedp) (2 * (loopnum) + (storedp ? 1 : 0)) | |
152 | |
153 /* Mem_ref hashtable helpers. */ | |
154 | |
155 struct mem_ref_hasher : nofree_ptr_hash <im_mem_ref> | |
156 { | |
145 | 157 typedef ao_ref *compare_type; |
111 | 158 static inline hashval_t hash (const im_mem_ref *); |
145 | 159 static inline bool equal (const im_mem_ref *, const ao_ref *); |
111 | 160 }; |
161 | |
145 | 162 /* A hash function for class im_mem_ref object OBJ. */ |
111 | 163 |
164 inline hashval_t | |
165 mem_ref_hasher::hash (const im_mem_ref *mem) | |
166 { | |
167 return mem->hash; | |
168 } | |
169 | |
145 | 170 /* An equality function for class im_mem_ref object MEM1 with |
111 | 171 memory reference OBJ2. */ |
172 | |
173 inline bool | |
145 | 174 mem_ref_hasher::equal (const im_mem_ref *mem1, const ao_ref *obj2) |
111 | 175 { |
145 | 176 if (obj2->max_size_known_p ()) |
177 return (mem1->ref_decomposed | |
178 && operand_equal_p (mem1->mem.base, obj2->base, 0) | |
179 && known_eq (mem1->mem.offset, obj2->offset) | |
180 && known_eq (mem1->mem.size, obj2->size) | |
181 && known_eq (mem1->mem.max_size, obj2->max_size) | |
182 && mem1->mem.volatile_p == obj2->volatile_p | |
183 && (mem1->mem.ref_alias_set == obj2->ref_alias_set | |
184 /* We are not canonicalizing alias-sets but for the | |
185 special-case we didn't canonicalize yet and the | |
186 incoming ref is a alias-set zero MEM we pick | |
187 the correct one already. */ | |
188 || (!mem1->ref_canonical | |
189 && (TREE_CODE (obj2->ref) == MEM_REF | |
190 || TREE_CODE (obj2->ref) == TARGET_MEM_REF) | |
191 && obj2->ref_alias_set == 0) | |
192 /* Likewise if there's a canonical ref with alias-set zero. */ | |
193 || (mem1->ref_canonical && mem1->mem.ref_alias_set == 0)) | |
194 && types_compatible_p (TREE_TYPE (mem1->mem.ref), | |
195 TREE_TYPE (obj2->ref))); | |
196 else | |
197 return operand_equal_p (mem1->mem.ref, obj2->ref, 0); | |
111 | 198 } |
199 | |
0 | 200 |
201 /* Description of memory accesses in loops. */ | |
202 | |
203 static struct | |
204 { | |
205 /* The hash table of memory references accessed in loops. */ | |
111 | 206 hash_table<mem_ref_hasher> *refs; |
0 | 207 |
208 /* The list of memory references. */ | |
111 | 209 vec<im_mem_ref *> refs_list; |
0 | 210 |
211 /* The set of memory references accessed in each loop. */ | |
111 | 212 vec<bitmap_head> refs_in_loop; |
213 | |
214 /* The set of memory references stored in each loop. */ | |
215 vec<bitmap_head> refs_stored_in_loop; | |
216 | |
217 /* The set of memory references stored in each loop, including subloops . */ | |
218 vec<bitmap_head> all_refs_stored_in_loop; | |
0 | 219 |
220 /* Cache for expanding memory addresses. */ | |
111 | 221 hash_map<tree, name_expansion *> *ttae_cache; |
0 | 222 } memory_accesses; |
223 | |
111 | 224 /* Obstack for the bitmaps in the above data structures. */ |
225 static bitmap_obstack lim_bitmap_obstack; | |
226 static obstack mem_ref_obstack; | |
227 | |
145 | 228 static bool ref_indep_loop_p (class loop *, im_mem_ref *); |
229 static bool ref_always_accessed_p (class loop *, im_mem_ref *, bool); | |
0 | 230 |
231 /* Minimum cost of an expensive expression. */ | |
145 | 232 #define LIM_EXPENSIVE ((unsigned) param_lim_expensive) |
0 | 233 |
111 | 234 /* The outermost loop for which execution of the header guarantees that the |
0 | 235 block will be executed. */ |
145 | 236 #define ALWAYS_EXECUTED_IN(BB) ((class loop *) (BB)->aux) |
111 | 237 #define SET_ALWAYS_EXECUTED_IN(BB, VAL) ((BB)->aux = (void *) (VAL)) |
238 | |
239 /* ID of the shared unanalyzable mem. */ | |
240 #define UNANALYZABLE_MEM_ID 0 | |
241 | |
242 /* Whether the reference was analyzable. */ | |
243 #define MEM_ANALYZABLE(REF) ((REF)->id != UNANALYZABLE_MEM_ID) | |
0 | 244 |
245 static struct lim_aux_data * | |
111 | 246 init_lim_data (gimple *stmt) |
0 | 247 { |
111 | 248 lim_aux_data *p = XCNEW (struct lim_aux_data); |
249 lim_aux_data_map->put (stmt, p); | |
250 | |
251 return p; | |
0 | 252 } |
253 | |
254 static struct lim_aux_data * | |
111 | 255 get_lim_data (gimple *stmt) |
0 | 256 { |
111 | 257 lim_aux_data **p = lim_aux_data_map->get (stmt); |
0 | 258 if (!p) |
259 return NULL; | |
260 | |
111 | 261 return *p; |
0 | 262 } |
263 | |
264 /* Releases the memory occupied by DATA. */ | |
265 | |
266 static void | |
267 free_lim_aux_data (struct lim_aux_data *data) | |
268 { | |
111 | 269 data->depends.release (); |
0 | 270 free (data); |
271 } | |
272 | |
273 static void | |
111 | 274 clear_lim_data (gimple *stmt) |
0 | 275 { |
111 | 276 lim_aux_data **p = lim_aux_data_map->get (stmt); |
0 | 277 if (!p) |
278 return; | |
279 | |
111 | 280 free_lim_aux_data (*p); |
0 | 281 *p = NULL; |
282 } | |
283 | |
111 | 284 |
285 /* The possibilities of statement movement. */ | |
286 enum move_pos | |
287 { | |
288 MOVE_IMPOSSIBLE, /* No movement -- side effect expression. */ | |
289 MOVE_PRESERVE_EXECUTION, /* Must not cause the non-executed statement | |
290 become executed -- memory accesses, ... */ | |
291 MOVE_POSSIBLE /* Unlimited movement. */ | |
292 }; | |
293 | |
0 | 294 |
295 /* If it is possible to hoist the statement STMT unconditionally, | |
296 returns MOVE_POSSIBLE. | |
297 If it is possible to hoist the statement STMT, but we must avoid making | |
298 it executed if it would not be executed in the original program (e.g. | |
299 because it may trap), return MOVE_PRESERVE_EXECUTION. | |
300 Otherwise return MOVE_IMPOSSIBLE. */ | |
301 | |
302 enum move_pos | |
111 | 303 movement_possibility (gimple *stmt) |
0 | 304 { |
305 tree lhs; | |
306 enum move_pos ret = MOVE_POSSIBLE; | |
307 | |
308 if (flag_unswitch_loops | |
309 && gimple_code (stmt) == GIMPLE_COND) | |
310 { | |
311 /* If we perform unswitching, force the operands of the invariant | |
312 condition to be moved out of the loop. */ | |
313 return MOVE_POSSIBLE; | |
314 } | |
315 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
316 if (gimple_code (stmt) == GIMPLE_PHI |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
317 && gimple_phi_num_args (stmt) <= 2 |
111 | 318 && !virtual_operand_p (gimple_phi_result (stmt)) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
319 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_phi_result (stmt))) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
320 return MOVE_POSSIBLE; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
321 |
0 | 322 if (gimple_get_lhs (stmt) == NULL_TREE) |
323 return MOVE_IMPOSSIBLE; | |
324 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
325 if (gimple_vdef (stmt)) |
0 | 326 return MOVE_IMPOSSIBLE; |
327 | |
328 if (stmt_ends_bb_p (stmt) | |
329 || gimple_has_volatile_ops (stmt) | |
330 || gimple_has_side_effects (stmt) | |
131 | 331 || stmt_could_throw_p (cfun, stmt)) |
0 | 332 return MOVE_IMPOSSIBLE; |
333 | |
334 if (is_gimple_call (stmt)) | |
335 { | |
336 /* While pure or const call is guaranteed to have no side effects, we | |
337 cannot move it arbitrarily. Consider code like | |
338 | |
339 char *s = something (); | |
340 | |
341 while (1) | |
342 { | |
343 if (s) | |
344 t = strlen (s); | |
345 else | |
346 t = 0; | |
347 } | |
348 | |
349 Here the strlen call cannot be moved out of the loop, even though | |
350 s is invariant. In addition to possibly creating a call with | |
351 invalid arguments, moving out a function call that is not executed | |
352 may cause performance regressions in case the call is costly and | |
353 not executed at all. */ | |
354 ret = MOVE_PRESERVE_EXECUTION; | |
355 lhs = gimple_call_lhs (stmt); | |
356 } | |
357 else if (is_gimple_assign (stmt)) | |
358 lhs = gimple_assign_lhs (stmt); | |
359 else | |
360 return MOVE_IMPOSSIBLE; | |
361 | |
362 if (TREE_CODE (lhs) == SSA_NAME | |
363 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)) | |
364 return MOVE_IMPOSSIBLE; | |
365 | |
366 if (TREE_CODE (lhs) != SSA_NAME | |
367 || gimple_could_trap_p (stmt)) | |
368 return MOVE_PRESERVE_EXECUTION; | |
369 | |
111 | 370 /* Non local loads in a transaction cannot be hoisted out. Well, |
371 unless the load happens on every path out of the loop, but we | |
372 don't take this into account yet. */ | |
373 if (flag_tm | |
374 && gimple_in_transaction (stmt) | |
375 && gimple_assign_single_p (stmt)) | |
376 { | |
377 tree rhs = gimple_assign_rhs1 (stmt); | |
378 if (DECL_P (rhs) && is_global_var (rhs)) | |
379 { | |
380 if (dump_file) | |
381 { | |
382 fprintf (dump_file, "Cannot hoist conditional load of "); | |
383 print_generic_expr (dump_file, rhs, TDF_SLIM); | |
384 fprintf (dump_file, " because it is in a transaction.\n"); | |
385 } | |
386 return MOVE_IMPOSSIBLE; | |
387 } | |
388 } | |
389 | |
0 | 390 return ret; |
391 } | |
392 | |
393 /* Suppose that operand DEF is used inside the LOOP. Returns the outermost | |
394 loop to that we could move the expression using DEF if it did not have | |
395 other operands, i.e. the outermost loop enclosing LOOP in that the value | |
396 of DEF is invariant. */ | |
397 | |
145 | 398 static class loop * |
399 outermost_invariant_loop (tree def, class loop *loop) | |
0 | 400 { |
111 | 401 gimple *def_stmt; |
0 | 402 basic_block def_bb; |
145 | 403 class loop *max_loop; |
0 | 404 struct lim_aux_data *lim_data; |
405 | |
406 if (!def) | |
407 return superloop_at_depth (loop, 1); | |
408 | |
409 if (TREE_CODE (def) != SSA_NAME) | |
410 { | |
411 gcc_assert (is_gimple_min_invariant (def)); | |
412 return superloop_at_depth (loop, 1); | |
413 } | |
414 | |
415 def_stmt = SSA_NAME_DEF_STMT (def); | |
416 def_bb = gimple_bb (def_stmt); | |
417 if (!def_bb) | |
418 return superloop_at_depth (loop, 1); | |
419 | |
420 max_loop = find_common_loop (loop, def_bb->loop_father); | |
421 | |
422 lim_data = get_lim_data (def_stmt); | |
423 if (lim_data != NULL && lim_data->max_loop != NULL) | |
424 max_loop = find_common_loop (max_loop, | |
425 loop_outer (lim_data->max_loop)); | |
426 if (max_loop == loop) | |
427 return NULL; | |
428 max_loop = superloop_at_depth (loop, loop_depth (max_loop) + 1); | |
429 | |
430 return max_loop; | |
431 } | |
432 | |
433 /* DATA is a structure containing information associated with a statement | |
434 inside LOOP. DEF is one of the operands of this statement. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
435 |
0 | 436 Find the outermost loop enclosing LOOP in that value of DEF is invariant |
437 and record this in DATA->max_loop field. If DEF itself is defined inside | |
438 this loop as well (i.e. we need to hoist it out of the loop if we want | |
439 to hoist the statement represented by DATA), record the statement in that | |
440 DEF is defined to the DATA->depends list. Additionally if ADD_COST is true, | |
441 add the cost of the computation of DEF to the DATA->cost. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
442 |
0 | 443 If DEF is not invariant in LOOP, return false. Otherwise return TRUE. */ |
444 | |
445 static bool | |
145 | 446 add_dependency (tree def, struct lim_aux_data *data, class loop *loop, |
0 | 447 bool add_cost) |
448 { | |
111 | 449 gimple *def_stmt = SSA_NAME_DEF_STMT (def); |
0 | 450 basic_block def_bb = gimple_bb (def_stmt); |
145 | 451 class loop *max_loop; |
0 | 452 struct lim_aux_data *def_data; |
453 | |
454 if (!def_bb) | |
455 return true; | |
456 | |
457 max_loop = outermost_invariant_loop (def, loop); | |
458 if (!max_loop) | |
459 return false; | |
460 | |
461 if (flow_loop_nested_p (data->max_loop, max_loop)) | |
462 data->max_loop = max_loop; | |
463 | |
464 def_data = get_lim_data (def_stmt); | |
465 if (!def_data) | |
466 return true; | |
467 | |
468 if (add_cost | |
469 /* Only add the cost if the statement defining DEF is inside LOOP, | |
470 i.e. if it is likely that by moving the invariants dependent | |
471 on it, we will be able to avoid creating a new register for | |
472 it (since it will be only used in these dependent invariants). */ | |
473 && def_bb->loop_father == loop) | |
474 data->cost += def_data->cost; | |
475 | |
111 | 476 data->depends.safe_push (def_stmt); |
0 | 477 |
478 return true; | |
479 } | |
480 | |
111 | 481 /* Returns an estimate for a cost of statement STMT. The values here |
482 are just ad-hoc constants, similar to costs for inlining. */ | |
0 | 483 |
484 static unsigned | |
111 | 485 stmt_cost (gimple *stmt) |
0 | 486 { |
487 /* Always try to create possibilities for unswitching. */ | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
488 if (gimple_code (stmt) == GIMPLE_COND |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
489 || gimple_code (stmt) == GIMPLE_PHI) |
0 | 490 return LIM_EXPENSIVE; |
491 | |
111 | 492 /* We should be hoisting calls if possible. */ |
0 | 493 if (is_gimple_call (stmt)) |
494 { | |
111 | 495 tree fndecl; |
0 | 496 |
497 /* Unless the call is a builtin_constant_p; this always folds to a | |
498 constant, so moving it is useless. */ | |
499 fndecl = gimple_call_fndecl (stmt); | |
131 | 500 if (fndecl && fndecl_built_in_p (fndecl, BUILT_IN_CONSTANT_P)) |
0 | 501 return 0; |
502 | |
111 | 503 return LIM_EXPENSIVE; |
0 | 504 } |
505 | |
111 | 506 /* Hoisting memory references out should almost surely be a win. */ |
507 if (gimple_references_memory_p (stmt)) | |
508 return LIM_EXPENSIVE; | |
509 | |
0 | 510 if (gimple_code (stmt) != GIMPLE_ASSIGN) |
111 | 511 return 1; |
0 | 512 |
513 switch (gimple_assign_rhs_code (stmt)) | |
514 { | |
515 case MULT_EXPR: | |
111 | 516 case WIDEN_MULT_EXPR: |
517 case WIDEN_MULT_PLUS_EXPR: | |
518 case WIDEN_MULT_MINUS_EXPR: | |
519 case DOT_PROD_EXPR: | |
0 | 520 case TRUNC_DIV_EXPR: |
521 case CEIL_DIV_EXPR: | |
522 case FLOOR_DIV_EXPR: | |
523 case ROUND_DIV_EXPR: | |
524 case EXACT_DIV_EXPR: | |
525 case CEIL_MOD_EXPR: | |
526 case FLOOR_MOD_EXPR: | |
527 case ROUND_MOD_EXPR: | |
528 case TRUNC_MOD_EXPR: | |
529 case RDIV_EXPR: | |
530 /* Division and multiplication are usually expensive. */ | |
111 | 531 return LIM_EXPENSIVE; |
0 | 532 |
533 case LSHIFT_EXPR: | |
534 case RSHIFT_EXPR: | |
111 | 535 case WIDEN_LSHIFT_EXPR: |
536 case LROTATE_EXPR: | |
537 case RROTATE_EXPR: | |
538 /* Shifts and rotates are usually expensive. */ | |
539 return LIM_EXPENSIVE; | |
540 | |
541 case CONSTRUCTOR: | |
542 /* Make vector construction cost proportional to the number | |
543 of elements. */ | |
544 return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt)); | |
545 | |
546 case SSA_NAME: | |
547 case PAREN_EXPR: | |
548 /* Whether or not something is wrapped inside a PAREN_EXPR | |
549 should not change move cost. Nor should an intermediate | |
550 unpropagated SSA name copy. */ | |
551 return 0; | |
0 | 552 |
553 default: | |
111 | 554 return 1; |
0 | 555 } |
556 } | |
557 | |
558 /* Finds the outermost loop between OUTER and LOOP in that the memory reference | |
559 REF is independent. If REF is not independent in LOOP, NULL is returned | |
560 instead. */ | |
561 | |
145 | 562 static class loop * |
563 outermost_indep_loop (class loop *outer, class loop *loop, im_mem_ref *ref) | |
0 | 564 { |
145 | 565 class loop *aloop; |
0 | 566 |
111 | 567 if (ref->stored && bitmap_bit_p (ref->stored, loop->num)) |
0 | 568 return NULL; |
569 | |
570 for (aloop = outer; | |
571 aloop != loop; | |
572 aloop = superloop_at_depth (loop, loop_depth (aloop) + 1)) | |
111 | 573 if ((!ref->stored || !bitmap_bit_p (ref->stored, aloop->num)) |
131 | 574 && ref_indep_loop_p (aloop, ref)) |
0 | 575 return aloop; |
576 | |
131 | 577 if (ref_indep_loop_p (loop, ref)) |
0 | 578 return loop; |
579 else | |
580 return NULL; | |
581 } | |
582 | |
583 /* If there is a simple load or store to a memory reference in STMT, returns | |
584 the location of the memory reference, and sets IS_STORE according to whether | |
585 it is a store or load. Otherwise, returns NULL. */ | |
586 | |
587 static tree * | |
111 | 588 simple_mem_ref_in_stmt (gimple *stmt, bool *is_store) |
0 | 589 { |
111 | 590 tree *lhs, *rhs; |
591 | |
592 /* Recognize SSA_NAME = MEM and MEM = (SSA_NAME | invariant) patterns. */ | |
593 if (!gimple_assign_single_p (stmt)) | |
0 | 594 return NULL; |
595 | |
596 lhs = gimple_assign_lhs_ptr (stmt); | |
111 | 597 rhs = gimple_assign_rhs1_ptr (stmt); |
598 | |
599 if (TREE_CODE (*lhs) == SSA_NAME && gimple_vuse (stmt)) | |
0 | 600 { |
601 *is_store = false; | |
111 | 602 return rhs; |
0 | 603 } |
111 | 604 else if (gimple_vdef (stmt) |
605 && (TREE_CODE (*rhs) == SSA_NAME || is_gimple_min_invariant (*rhs))) | |
0 | 606 { |
607 *is_store = true; | |
608 return lhs; | |
609 } | |
610 else | |
611 return NULL; | |
612 } | |
613 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
614 /* From a controlling predicate in DOM determine the arguments from |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
615 the PHI node PHI that are chosen if the predicate evaluates to |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
616 true and false and store them to *TRUE_ARG_P and *FALSE_ARG_P if |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
617 they are non-NULL. Returns true if the arguments can be determined, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
618 else return false. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
619 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
620 static bool |
111 | 621 extract_true_false_args_from_phi (basic_block dom, gphi *phi, |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
622 tree *true_arg_p, tree *false_arg_p) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
623 { |
111 | 624 edge te, fe; |
625 if (! extract_true_false_controlled_edges (dom, gimple_bb (phi), | |
626 &te, &fe)) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
627 return false; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
628 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
629 if (true_arg_p) |
111 | 630 *true_arg_p = PHI_ARG_DEF (phi, te->dest_idx); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
631 if (false_arg_p) |
111 | 632 *false_arg_p = PHI_ARG_DEF (phi, fe->dest_idx); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
633 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
634 return true; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
635 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
636 |
0 | 637 /* Determine the outermost loop to that it is possible to hoist a statement |
638 STMT and store it to LIM_DATA (STMT)->max_loop. To do this we determine | |
639 the outermost loop in that the value computed by STMT is invariant. | |
640 If MUST_PRESERVE_EXEC is true, additionally choose such a loop that | |
641 we preserve the fact whether STMT is executed. It also fills other related | |
642 information to LIM_DATA (STMT). | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
643 |
0 | 644 The function returns false if STMT cannot be hoisted outside of the loop it |
645 is defined in, and true otherwise. */ | |
646 | |
647 static bool | |
111 | 648 determine_max_movement (gimple *stmt, bool must_preserve_exec) |
0 | 649 { |
650 basic_block bb = gimple_bb (stmt); | |
145 | 651 class loop *loop = bb->loop_father; |
652 class loop *level; | |
0 | 653 struct lim_aux_data *lim_data = get_lim_data (stmt); |
654 tree val; | |
655 ssa_op_iter iter; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
656 |
0 | 657 if (must_preserve_exec) |
658 level = ALWAYS_EXECUTED_IN (bb); | |
659 else | |
660 level = superloop_at_depth (loop, 1); | |
661 lim_data->max_loop = level; | |
662 | |
111 | 663 if (gphi *phi = dyn_cast <gphi *> (stmt)) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
664 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
665 use_operand_p use_p; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
666 unsigned min_cost = UINT_MAX; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
667 unsigned total_cost = 0; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
668 struct lim_aux_data *def_data; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
669 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
670 /* We will end up promoting dependencies to be unconditionally |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
671 evaluated. For this reason the PHI cost (and thus the |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
672 cost we remove from the loop by doing the invariant motion) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
673 is that of the cheapest PHI argument dependency chain. */ |
111 | 674 FOR_EACH_PHI_ARG (use_p, phi, iter, SSA_OP_USE) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
675 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
676 val = USE_FROM_PTR (use_p); |
111 | 677 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
678 if (TREE_CODE (val) != SSA_NAME) |
111 | 679 { |
680 /* Assign const 1 to constants. */ | |
681 min_cost = MIN (min_cost, 1); | |
682 total_cost += 1; | |
683 continue; | |
684 } | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
685 if (!add_dependency (val, lim_data, loop, false)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
686 return false; |
111 | 687 |
688 gimple *def_stmt = SSA_NAME_DEF_STMT (val); | |
689 if (gimple_bb (def_stmt) | |
690 && gimple_bb (def_stmt)->loop_father == loop) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
691 { |
111 | 692 def_data = get_lim_data (def_stmt); |
693 if (def_data) | |
694 { | |
695 min_cost = MIN (min_cost, def_data->cost); | |
696 total_cost += def_data->cost; | |
697 } | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
698 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
699 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
700 |
111 | 701 min_cost = MIN (min_cost, total_cost); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
702 lim_data->cost += min_cost; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
703 |
111 | 704 if (gimple_phi_num_args (phi) > 1) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
705 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
706 basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb); |
111 | 707 gimple *cond; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
708 if (gsi_end_p (gsi_last_bb (dom))) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
709 return false; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
710 cond = gsi_stmt (gsi_last_bb (dom)); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
711 if (gimple_code (cond) != GIMPLE_COND) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
712 return false; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
713 /* Verify that this is an extended form of a diamond and |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
714 the PHI arguments are completely controlled by the |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
715 predicate in DOM. */ |
111 | 716 if (!extract_true_false_args_from_phi (dom, phi, NULL, NULL)) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
717 return false; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
718 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
719 /* Fold in dependencies and cost of the condition. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
720 FOR_EACH_SSA_TREE_OPERAND (val, cond, iter, SSA_OP_USE) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
721 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
722 if (!add_dependency (val, lim_data, loop, false)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
723 return false; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
724 def_data = get_lim_data (SSA_NAME_DEF_STMT (val)); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
725 if (def_data) |
111 | 726 lim_data->cost += def_data->cost; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
727 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
728 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
729 /* We want to avoid unconditionally executing very expensive |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
730 operations. As costs for our dependencies cannot be |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
731 negative just claim we are not invariand for this case. |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
732 We also are not sure whether the control-flow inside the |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
733 loop will vanish. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
734 if (total_cost - min_cost >= 2 * LIM_EXPENSIVE |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
735 && !(min_cost != 0 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
736 && total_cost / min_cost <= 2)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
737 return false; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
738 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
739 /* Assume that the control-flow in the loop will vanish. |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
740 ??? We should verify this and not artificially increase |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
741 the cost if that is not the case. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
742 lim_data->cost += stmt_cost (stmt); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
743 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
744 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
745 return true; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
746 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
747 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
748 FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_USE) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
749 if (!add_dependency (val, lim_data, loop, true)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
750 return false; |
0 | 751 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
752 if (gimple_vuse (stmt)) |
0 | 753 { |
111 | 754 im_mem_ref *ref |
755 = lim_data ? memory_accesses.refs_list[lim_data->ref] : NULL; | |
756 if (ref | |
757 && MEM_ANALYZABLE (ref)) | |
0 | 758 { |
111 | 759 lim_data->max_loop = outermost_indep_loop (lim_data->max_loop, |
760 loop, ref); | |
0 | 761 if (!lim_data->max_loop) |
762 return false; | |
763 } | |
111 | 764 else if (! add_dependency (gimple_vuse (stmt), lim_data, loop, false)) |
765 return false; | |
0 | 766 } |
767 | |
768 lim_data->cost += stmt_cost (stmt); | |
769 | |
770 return true; | |
771 } | |
772 | |
773 /* Suppose that some statement in ORIG_LOOP is hoisted to the loop LEVEL, | |
774 and that one of the operands of this statement is computed by STMT. | |
775 Ensure that STMT (together with all the statements that define its | |
776 operands) is hoisted at least out of the loop LEVEL. */ | |
777 | |
778 static void | |
145 | 779 set_level (gimple *stmt, class loop *orig_loop, class loop *level) |
0 | 780 { |
145 | 781 class loop *stmt_loop = gimple_bb (stmt)->loop_father; |
0 | 782 struct lim_aux_data *lim_data; |
111 | 783 gimple *dep_stmt; |
784 unsigned i; | |
0 | 785 |
786 stmt_loop = find_common_loop (orig_loop, stmt_loop); | |
787 lim_data = get_lim_data (stmt); | |
788 if (lim_data != NULL && lim_data->tgt_loop != NULL) | |
789 stmt_loop = find_common_loop (stmt_loop, | |
790 loop_outer (lim_data->tgt_loop)); | |
791 if (flow_loop_nested_p (stmt_loop, level)) | |
792 return; | |
793 | |
794 gcc_assert (level == lim_data->max_loop | |
795 || flow_loop_nested_p (lim_data->max_loop, level)); | |
796 | |
797 lim_data->tgt_loop = level; | |
111 | 798 FOR_EACH_VEC_ELT (lim_data->depends, i, dep_stmt) |
799 set_level (dep_stmt, orig_loop, level); | |
0 | 800 } |
801 | |
802 /* Determines an outermost loop from that we want to hoist the statement STMT. | |
803 For now we chose the outermost possible loop. TODO -- use profiling | |
804 information to set it more sanely. */ | |
805 | |
806 static void | |
111 | 807 set_profitable_level (gimple *stmt) |
0 | 808 { |
809 set_level (stmt, gimple_bb (stmt)->loop_father, get_lim_data (stmt)->max_loop); | |
810 } | |
811 | |
812 /* Returns true if STMT is a call that has side effects. */ | |
813 | |
814 static bool | |
111 | 815 nonpure_call_p (gimple *stmt) |
0 | 816 { |
817 if (gimple_code (stmt) != GIMPLE_CALL) | |
818 return false; | |
819 | |
820 return gimple_has_side_effects (stmt); | |
821 } | |
822 | |
823 /* Rewrite a/b to a*(1/b). Return the invariant stmt to process. */ | |
824 | |
111 | 825 static gimple * |
0 | 826 rewrite_reciprocal (gimple_stmt_iterator *bsi) |
827 { | |
111 | 828 gassign *stmt, *stmt1, *stmt2; |
829 tree name, lhs, type; | |
0 | 830 tree real_one; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
831 gimple_stmt_iterator gsi; |
0 | 832 |
111 | 833 stmt = as_a <gassign *> (gsi_stmt (*bsi)); |
0 | 834 lhs = gimple_assign_lhs (stmt); |
835 type = TREE_TYPE (lhs); | |
836 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
837 real_one = build_one_cst (type); |
0 | 838 |
111 | 839 name = make_temp_ssa_name (type, NULL, "reciptmp"); |
840 stmt1 = gimple_build_assign (name, RDIV_EXPR, real_one, | |
841 gimple_assign_rhs2 (stmt)); | |
842 stmt2 = gimple_build_assign (lhs, MULT_EXPR, name, | |
843 gimple_assign_rhs1 (stmt)); | |
0 | 844 |
845 /* Replace division stmt with reciprocal and multiply stmts. | |
846 The multiply stmt is not invariant, so update iterator | |
847 and avoid rescanning. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
848 gsi = *bsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
849 gsi_insert_before (bsi, stmt1, GSI_NEW_STMT); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
850 gsi_replace (&gsi, stmt2, true); |
0 | 851 |
852 /* Continue processing with invariant reciprocal statement. */ | |
853 return stmt1; | |
854 } | |
855 | |
856 /* Check if the pattern at *BSI is a bittest of the form | |
857 (A >> B) & 1 != 0 and in this case rewrite it to A & (1 << B) != 0. */ | |
858 | |
111 | 859 static gimple * |
0 | 860 rewrite_bittest (gimple_stmt_iterator *bsi) |
861 { | |
111 | 862 gassign *stmt; |
863 gimple *stmt1; | |
864 gassign *stmt2; | |
865 gimple *use_stmt; | |
866 gcond *cond_stmt; | |
867 tree lhs, name, t, a, b; | |
0 | 868 use_operand_p use; |
869 | |
111 | 870 stmt = as_a <gassign *> (gsi_stmt (*bsi)); |
0 | 871 lhs = gimple_assign_lhs (stmt); |
872 | |
873 /* Verify that the single use of lhs is a comparison against zero. */ | |
874 if (TREE_CODE (lhs) != SSA_NAME | |
111 | 875 || !single_imm_use (lhs, &use, &use_stmt)) |
876 return stmt; | |
877 cond_stmt = dyn_cast <gcond *> (use_stmt); | |
878 if (!cond_stmt) | |
0 | 879 return stmt; |
111 | 880 if (gimple_cond_lhs (cond_stmt) != lhs |
881 || (gimple_cond_code (cond_stmt) != NE_EXPR | |
882 && gimple_cond_code (cond_stmt) != EQ_EXPR) | |
883 || !integer_zerop (gimple_cond_rhs (cond_stmt))) | |
0 | 884 return stmt; |
885 | |
886 /* Get at the operands of the shift. The rhs is TMP1 & 1. */ | |
887 stmt1 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
888 if (gimple_code (stmt1) != GIMPLE_ASSIGN) | |
889 return stmt; | |
890 | |
891 /* There is a conversion in between possibly inserted by fold. */ | |
892 if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt1))) | |
893 { | |
894 t = gimple_assign_rhs1 (stmt1); | |
895 if (TREE_CODE (t) != SSA_NAME | |
896 || !has_single_use (t)) | |
897 return stmt; | |
898 stmt1 = SSA_NAME_DEF_STMT (t); | |
899 if (gimple_code (stmt1) != GIMPLE_ASSIGN) | |
900 return stmt; | |
901 } | |
902 | |
903 /* Verify that B is loop invariant but A is not. Verify that with | |
904 all the stmt walking we are still in the same loop. */ | |
905 if (gimple_assign_rhs_code (stmt1) != RSHIFT_EXPR | |
906 || loop_containing_stmt (stmt1) != loop_containing_stmt (stmt)) | |
907 return stmt; | |
908 | |
909 a = gimple_assign_rhs1 (stmt1); | |
910 b = gimple_assign_rhs2 (stmt1); | |
911 | |
912 if (outermost_invariant_loop (b, loop_containing_stmt (stmt1)) != NULL | |
913 && outermost_invariant_loop (a, loop_containing_stmt (stmt1)) == NULL) | |
914 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
915 gimple_stmt_iterator rsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
916 |
0 | 917 /* 1 << B */ |
918 t = fold_build2 (LSHIFT_EXPR, TREE_TYPE (a), | |
919 build_int_cst (TREE_TYPE (a), 1), b); | |
111 | 920 name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp"); |
921 stmt1 = gimple_build_assign (name, t); | |
0 | 922 |
923 /* A & (1 << B) */ | |
924 t = fold_build2 (BIT_AND_EXPR, TREE_TYPE (a), a, name); | |
111 | 925 name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp"); |
926 stmt2 = gimple_build_assign (name, t); | |
0 | 927 |
928 /* Replace the SSA_NAME we compare against zero. Adjust | |
929 the type of zero accordingly. */ | |
930 SET_USE (use, name); | |
111 | 931 gimple_cond_set_rhs (cond_stmt, |
932 build_int_cst_type (TREE_TYPE (name), | |
933 0)); | |
0 | 934 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
935 /* Don't use gsi_replace here, none of the new assignments sets |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
936 the variable originally set in stmt. Move bsi to stmt1, and |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
937 then remove the original stmt, so that we get a chance to |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
938 retain debug info for it. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
939 rsi = *bsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
940 gsi_insert_before (bsi, stmt1, GSI_NEW_STMT); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
941 gsi_insert_before (&rsi, stmt2, GSI_SAME_STMT); |
111 | 942 gimple *to_release = gsi_stmt (rsi); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
943 gsi_remove (&rsi, true); |
111 | 944 release_defs (to_release); |
0 | 945 |
946 return stmt1; | |
947 } | |
948 | |
949 return stmt; | |
950 } | |
951 | |
111 | 952 /* For each statement determines the outermost loop in that it is invariant, |
953 - statements on whose motion it depends and the cost of the computation. | |
954 - This information is stored to the LIM_DATA structure associated with | |
955 - each statement. */ | |
956 class invariantness_dom_walker : public dom_walker | |
957 { | |
958 public: | |
959 invariantness_dom_walker (cdi_direction direction) | |
960 : dom_walker (direction) {} | |
961 | |
962 virtual edge before_dom_children (basic_block); | |
963 }; | |
0 | 964 |
965 /* Determine the outermost loops in that statements in basic block BB are | |
966 invariant, and record them to the LIM_DATA associated with the statements. | |
111 | 967 Callback for dom_walker. */ |
968 | |
969 edge | |
970 invariantness_dom_walker::before_dom_children (basic_block bb) | |
0 | 971 { |
972 enum move_pos pos; | |
973 gimple_stmt_iterator bsi; | |
111 | 974 gimple *stmt; |
0 | 975 bool maybe_never = ALWAYS_EXECUTED_IN (bb) == NULL; |
145 | 976 class loop *outermost = ALWAYS_EXECUTED_IN (bb); |
0 | 977 struct lim_aux_data *lim_data; |
978 | |
979 if (!loop_outer (bb->loop_father)) | |
111 | 980 return NULL; |
0 | 981 |
982 if (dump_file && (dump_flags & TDF_DETAILS)) | |
983 fprintf (dump_file, "Basic block %d (loop %d -- depth %d):\n\n", | |
984 bb->index, bb->loop_father->num, loop_depth (bb->loop_father)); | |
985 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
986 /* Look at PHI nodes, but only if there is at most two. |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
987 ??? We could relax this further by post-processing the inserted |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
988 code and transforming adjacent cond-exprs with the same predicate |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
989 to control flow again. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
990 bsi = gsi_start_phis (bb); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
991 if (!gsi_end_p (bsi) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
992 && ((gsi_next (&bsi), gsi_end_p (bsi)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
993 || (gsi_next (&bsi), gsi_end_p (bsi)))) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
994 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
995 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
996 stmt = gsi_stmt (bsi); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
997 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
998 pos = movement_possibility (stmt); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
999 if (pos == MOVE_IMPOSSIBLE) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1000 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1001 |
111 | 1002 lim_data = get_lim_data (stmt); |
1003 if (! lim_data) | |
1004 lim_data = init_lim_data (stmt); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1005 lim_data->always_executed_in = outermost; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1006 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1007 if (!determine_max_movement (stmt, false)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1008 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1009 lim_data->max_loop = NULL; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1010 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1011 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1012 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1013 if (dump_file && (dump_flags & TDF_DETAILS)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1014 { |
111 | 1015 print_gimple_stmt (dump_file, stmt, 2); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1016 fprintf (dump_file, " invariant up to level %d, cost %d.\n\n", |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1017 loop_depth (lim_data->max_loop), |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1018 lim_data->cost); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1019 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1020 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1021 if (lim_data->cost >= LIM_EXPENSIVE) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1022 set_profitable_level (stmt); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1023 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1024 |
0 | 1025 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
1026 { | |
1027 stmt = gsi_stmt (bsi); | |
1028 | |
1029 pos = movement_possibility (stmt); | |
1030 if (pos == MOVE_IMPOSSIBLE) | |
1031 { | |
1032 if (nonpure_call_p (stmt)) | |
1033 { | |
1034 maybe_never = true; | |
1035 outermost = NULL; | |
1036 } | |
1037 /* Make sure to note always_executed_in for stores to make | |
1038 store-motion work. */ | |
1039 else if (stmt_makes_single_store (stmt)) | |
1040 { | |
111 | 1041 struct lim_aux_data *lim_data = get_lim_data (stmt); |
1042 if (! lim_data) | |
1043 lim_data = init_lim_data (stmt); | |
0 | 1044 lim_data->always_executed_in = outermost; |
1045 } | |
1046 continue; | |
1047 } | |
1048 | |
1049 if (is_gimple_assign (stmt) | |
1050 && (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)) | |
1051 == GIMPLE_BINARY_RHS)) | |
1052 { | |
1053 tree op0 = gimple_assign_rhs1 (stmt); | |
1054 tree op1 = gimple_assign_rhs2 (stmt); | |
145 | 1055 class loop *ol1 = outermost_invariant_loop (op1, |
0 | 1056 loop_containing_stmt (stmt)); |
1057 | |
1058 /* If divisor is invariant, convert a/b to a*(1/b), allowing reciprocal | |
1059 to be hoisted out of loop, saving expensive divide. */ | |
1060 if (pos == MOVE_POSSIBLE | |
1061 && gimple_assign_rhs_code (stmt) == RDIV_EXPR | |
1062 && flag_unsafe_math_optimizations | |
1063 && !flag_trapping_math | |
1064 && ol1 != NULL | |
1065 && outermost_invariant_loop (op0, ol1) == NULL) | |
1066 stmt = rewrite_reciprocal (&bsi); | |
1067 | |
1068 /* If the shift count is invariant, convert (A >> B) & 1 to | |
1069 A & (1 << B) allowing the bit mask to be hoisted out of the loop | |
1070 saving an expensive shift. */ | |
1071 if (pos == MOVE_POSSIBLE | |
1072 && gimple_assign_rhs_code (stmt) == BIT_AND_EXPR | |
1073 && integer_onep (op1) | |
1074 && TREE_CODE (op0) == SSA_NAME | |
1075 && has_single_use (op0)) | |
1076 stmt = rewrite_bittest (&bsi); | |
1077 } | |
1078 | |
111 | 1079 lim_data = get_lim_data (stmt); |
1080 if (! lim_data) | |
1081 lim_data = init_lim_data (stmt); | |
0 | 1082 lim_data->always_executed_in = outermost; |
1083 | |
1084 if (maybe_never && pos == MOVE_PRESERVE_EXECUTION) | |
1085 continue; | |
1086 | |
1087 if (!determine_max_movement (stmt, pos == MOVE_PRESERVE_EXECUTION)) | |
1088 { | |
1089 lim_data->max_loop = NULL; | |
1090 continue; | |
1091 } | |
1092 | |
1093 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1094 { | |
111 | 1095 print_gimple_stmt (dump_file, stmt, 2); |
0 | 1096 fprintf (dump_file, " invariant up to level %d, cost %d.\n\n", |
1097 loop_depth (lim_data->max_loop), | |
1098 lim_data->cost); | |
1099 } | |
1100 | |
1101 if (lim_data->cost >= LIM_EXPENSIVE) | |
1102 set_profitable_level (stmt); | |
1103 } | |
111 | 1104 return NULL; |
0 | 1105 } |
1106 | |
1107 /* Hoist the statements in basic block BB out of the loops prescribed by | |
1108 data stored in LIM_DATA structures associated with each statement. Callback | |
1109 for walk_dominator_tree. */ | |
1110 | |
111 | 1111 unsigned int |
1112 move_computations_worker (basic_block bb) | |
0 | 1113 { |
145 | 1114 class loop *level; |
0 | 1115 unsigned cost = 0; |
1116 struct lim_aux_data *lim_data; | |
111 | 1117 unsigned int todo = 0; |
0 | 1118 |
1119 if (!loop_outer (bb->loop_father)) | |
111 | 1120 return todo; |
1121 | |
1122 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); ) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1123 { |
111 | 1124 gassign *new_stmt; |
1125 gphi *stmt = bsi.phi (); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1126 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1127 lim_data = get_lim_data (stmt); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1128 if (lim_data == NULL) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1129 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1130 gsi_next (&bsi); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1131 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1132 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1133 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1134 cost = lim_data->cost; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1135 level = lim_data->tgt_loop; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1136 clear_lim_data (stmt); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1137 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1138 if (!level) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1139 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1140 gsi_next (&bsi); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1141 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1142 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1143 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1144 if (dump_file && (dump_flags & TDF_DETAILS)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1145 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1146 fprintf (dump_file, "Moving PHI node\n"); |
111 | 1147 print_gimple_stmt (dump_file, stmt, 0); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1148 fprintf (dump_file, "(cost %u) out of loop %d.\n\n", |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1149 cost, level->num); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1150 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1151 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1152 if (gimple_phi_num_args (stmt) == 1) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1153 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1154 tree arg = PHI_ARG_DEF (stmt, 0); |
111 | 1155 new_stmt = gimple_build_assign (gimple_phi_result (stmt), |
1156 TREE_CODE (arg), arg); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1157 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1158 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1159 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1160 basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb); |
111 | 1161 gimple *cond = gsi_stmt (gsi_last_bb (dom)); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1162 tree arg0 = NULL_TREE, arg1 = NULL_TREE, t; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1163 /* Get the PHI arguments corresponding to the true and false |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1164 edges of COND. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1165 extract_true_false_args_from_phi (dom, stmt, &arg0, &arg1); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1166 gcc_assert (arg0 && arg1); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1167 t = build2 (gimple_cond_code (cond), boolean_type_node, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1168 gimple_cond_lhs (cond), gimple_cond_rhs (cond)); |
111 | 1169 new_stmt = gimple_build_assign (gimple_phi_result (stmt), |
1170 COND_EXPR, t, arg0, arg1); | |
1171 todo |= TODO_cleanup_cfg; | |
1172 } | |
1173 if (INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (new_stmt))) | |
1174 && (!ALWAYS_EXECUTED_IN (bb) | |
1175 || (ALWAYS_EXECUTED_IN (bb) != level | |
1176 && !flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level)))) | |
1177 { | |
1178 tree lhs = gimple_assign_lhs (new_stmt); | |
1179 SSA_NAME_RANGE_INFO (lhs) = NULL; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1180 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1181 gsi_insert_on_edge (loop_preheader_edge (level), new_stmt); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1182 remove_phi_node (&bsi, false); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1183 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1184 |
111 | 1185 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); ) |
0 | 1186 { |
111 | 1187 edge e; |
1188 | |
1189 gimple *stmt = gsi_stmt (bsi); | |
0 | 1190 |
1191 lim_data = get_lim_data (stmt); | |
1192 if (lim_data == NULL) | |
1193 { | |
1194 gsi_next (&bsi); | |
1195 continue; | |
1196 } | |
1197 | |
1198 cost = lim_data->cost; | |
1199 level = lim_data->tgt_loop; | |
1200 clear_lim_data (stmt); | |
1201 | |
1202 if (!level) | |
1203 { | |
1204 gsi_next (&bsi); | |
1205 continue; | |
1206 } | |
1207 | |
1208 /* We do not really want to move conditionals out of the loop; we just | |
1209 placed it here to force its operands to be moved if necessary. */ | |
1210 if (gimple_code (stmt) == GIMPLE_COND) | |
1211 continue; | |
1212 | |
1213 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1214 { | |
1215 fprintf (dump_file, "Moving statement\n"); | |
111 | 1216 print_gimple_stmt (dump_file, stmt, 0); |
0 | 1217 fprintf (dump_file, "(cost %u) out of loop %d.\n\n", |
1218 cost, level->num); | |
1219 } | |
1220 | |
111 | 1221 e = loop_preheader_edge (level); |
1222 gcc_assert (!gimple_vdef (stmt)); | |
1223 if (gimple_vuse (stmt)) | |
1224 { | |
1225 /* The new VUSE is the one from the virtual PHI in the loop | |
1226 header or the one already present. */ | |
1227 gphi_iterator gsi2; | |
1228 for (gsi2 = gsi_start_phis (e->dest); | |
1229 !gsi_end_p (gsi2); gsi_next (&gsi2)) | |
1230 { | |
1231 gphi *phi = gsi2.phi (); | |
1232 if (virtual_operand_p (gimple_phi_result (phi))) | |
1233 { | |
145 | 1234 SET_USE (gimple_vuse_op (stmt), |
1235 PHI_ARG_DEF_FROM_EDGE (phi, e)); | |
111 | 1236 break; |
1237 } | |
1238 } | |
1239 } | |
0 | 1240 gsi_remove (&bsi, false); |
111 | 1241 if (gimple_has_lhs (stmt) |
1242 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME | |
1243 && INTEGRAL_TYPE_P (TREE_TYPE (gimple_get_lhs (stmt))) | |
1244 && (!ALWAYS_EXECUTED_IN (bb) | |
1245 || !(ALWAYS_EXECUTED_IN (bb) == level | |
1246 || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level)))) | |
1247 { | |
1248 tree lhs = gimple_get_lhs (stmt); | |
1249 SSA_NAME_RANGE_INFO (lhs) = NULL; | |
1250 } | |
1251 /* In case this is a stmt that is not unconditionally executed | |
1252 when the target loop header is executed and the stmt may | |
1253 invoke undefined integer or pointer overflow rewrite it to | |
1254 unsigned arithmetic. */ | |
1255 if (is_gimple_assign (stmt) | |
1256 && INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (stmt))) | |
1257 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (gimple_assign_lhs (stmt))) | |
1258 && arith_code_with_undefined_signed_overflow | |
1259 (gimple_assign_rhs_code (stmt)) | |
1260 && (!ALWAYS_EXECUTED_IN (bb) | |
1261 || !(ALWAYS_EXECUTED_IN (bb) == level | |
1262 || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level)))) | |
1263 gsi_insert_seq_on_edge (e, rewrite_to_defined_overflow (stmt)); | |
1264 else | |
1265 gsi_insert_on_edge (e, stmt); | |
0 | 1266 } |
111 | 1267 |
1268 return todo; | |
0 | 1269 } |
1270 | |
1271 /* Hoist the statements out of the loops prescribed by data stored in | |
1272 LIM_DATA structures associated with each statement.*/ | |
1273 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1274 static unsigned int |
0 | 1275 move_computations (void) |
1276 { | |
111 | 1277 int *rpo = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
1278 int n = pre_and_rev_post_order_compute_fn (cfun, NULL, rpo, false); | |
1279 unsigned todo = 0; | |
1280 | |
1281 for (int i = 0; i < n; ++i) | |
1282 todo |= move_computations_worker (BASIC_BLOCK_FOR_FN (cfun, rpo[i])); | |
1283 | |
1284 free (rpo); | |
0 | 1285 |
1286 gsi_commit_edge_inserts (); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
1287 if (need_ssa_update_p (cfun)) |
0 | 1288 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1289 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1290 return todo; |
0 | 1291 } |
1292 | |
1293 /* Checks whether the statement defining variable *INDEX can be hoisted | |
1294 out of the loop passed in DATA. Callback for for_each_index. */ | |
1295 | |
1296 static bool | |
1297 may_move_till (tree ref, tree *index, void *data) | |
1298 { | |
145 | 1299 class loop *loop = (class loop *) data, *max_loop; |
0 | 1300 |
1301 /* If REF is an array reference, check also that the step and the lower | |
1302 bound is invariant in LOOP. */ | |
1303 if (TREE_CODE (ref) == ARRAY_REF) | |
1304 { | |
1305 tree step = TREE_OPERAND (ref, 3); | |
1306 tree lbound = TREE_OPERAND (ref, 2); | |
1307 | |
1308 max_loop = outermost_invariant_loop (step, loop); | |
1309 if (!max_loop) | |
1310 return false; | |
1311 | |
1312 max_loop = outermost_invariant_loop (lbound, loop); | |
1313 if (!max_loop) | |
1314 return false; | |
1315 } | |
1316 | |
1317 max_loop = outermost_invariant_loop (*index, loop); | |
1318 if (!max_loop) | |
1319 return false; | |
1320 | |
1321 return true; | |
1322 } | |
1323 | |
1324 /* If OP is SSA NAME, force the statement that defines it to be | |
1325 moved out of the LOOP. ORIG_LOOP is the loop in that EXPR is used. */ | |
1326 | |
1327 static void | |
145 | 1328 force_move_till_op (tree op, class loop *orig_loop, class loop *loop) |
0 | 1329 { |
111 | 1330 gimple *stmt; |
0 | 1331 |
1332 if (!op | |
1333 || is_gimple_min_invariant (op)) | |
1334 return; | |
1335 | |
1336 gcc_assert (TREE_CODE (op) == SSA_NAME); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
1337 |
0 | 1338 stmt = SSA_NAME_DEF_STMT (op); |
1339 if (gimple_nop_p (stmt)) | |
1340 return; | |
1341 | |
1342 set_level (stmt, orig_loop, loop); | |
1343 } | |
1344 | |
1345 /* Forces statement defining invariants in REF (and *INDEX) to be moved out of | |
1346 the LOOP. The reference REF is used in the loop ORIG_LOOP. Callback for | |
1347 for_each_index. */ | |
1348 | |
1349 struct fmt_data | |
1350 { | |
145 | 1351 class loop *loop; |
1352 class loop *orig_loop; | |
0 | 1353 }; |
1354 | |
1355 static bool | |
1356 force_move_till (tree ref, tree *index, void *data) | |
1357 { | |
1358 struct fmt_data *fmt_data = (struct fmt_data *) data; | |
1359 | |
1360 if (TREE_CODE (ref) == ARRAY_REF) | |
1361 { | |
1362 tree step = TREE_OPERAND (ref, 3); | |
1363 tree lbound = TREE_OPERAND (ref, 2); | |
1364 | |
1365 force_move_till_op (step, fmt_data->orig_loop, fmt_data->loop); | |
1366 force_move_till_op (lbound, fmt_data->orig_loop, fmt_data->loop); | |
1367 } | |
1368 | |
1369 force_move_till_op (*index, fmt_data->orig_loop, fmt_data->loop); | |
1370 | |
1371 return true; | |
1372 } | |
1373 | |
1374 /* A function to free the mem_ref object OBJ. */ | |
1375 | |
1376 static void | |
145 | 1377 memref_free (class im_mem_ref *mem) |
0 | 1378 { |
111 | 1379 mem->accesses_in_loop.release (); |
0 | 1380 } |
1381 | |
1382 /* Allocates and returns a memory reference description for MEM whose hash | |
1383 value is HASH and id is ID. */ | |
1384 | |
111 | 1385 static im_mem_ref * |
145 | 1386 mem_ref_alloc (ao_ref *mem, unsigned hash, unsigned id) |
0 | 1387 { |
145 | 1388 im_mem_ref *ref = XOBNEW (&mem_ref_obstack, class im_mem_ref); |
1389 if (mem) | |
1390 ref->mem = *mem; | |
1391 else | |
1392 ao_ref_init (&ref->mem, error_mark_node); | |
0 | 1393 ref->id = id; |
145 | 1394 ref->ref_canonical = false; |
1395 ref->ref_decomposed = false; | |
0 | 1396 ref->hash = hash; |
111 | 1397 ref->stored = NULL; |
1398 bitmap_initialize (&ref->indep_loop, &lim_bitmap_obstack); | |
1399 bitmap_initialize (&ref->dep_loop, &lim_bitmap_obstack); | |
1400 ref->accesses_in_loop.create (1); | |
0 | 1401 |
1402 return ref; | |
1403 } | |
1404 | |
1405 /* Records memory reference location *LOC in LOOP to the memory reference | |
1406 description REF. The reference occurs in statement STMT. */ | |
1407 | |
1408 static void | |
111 | 1409 record_mem_ref_loc (im_mem_ref *ref, gimple *stmt, tree *loc) |
0 | 1410 { |
111 | 1411 mem_ref_loc aref; |
1412 aref.stmt = stmt; | |
1413 aref.ref = loc; | |
1414 ref->accesses_in_loop.safe_push (aref); | |
1415 } | |
1416 | |
1417 /* Set the LOOP bit in REF stored bitmap and allocate that if | |
1418 necessary. Return whether a bit was changed. */ | |
1419 | |
1420 static bool | |
145 | 1421 set_ref_stored_in_loop (im_mem_ref *ref, class loop *loop) |
111 | 1422 { |
1423 if (!ref->stored) | |
1424 ref->stored = BITMAP_ALLOC (&lim_bitmap_obstack); | |
1425 return bitmap_set_bit (ref->stored, loop->num); | |
0 | 1426 } |
1427 | |
1428 /* Marks reference REF as stored in LOOP. */ | |
1429 | |
1430 static void | |
145 | 1431 mark_ref_stored (im_mem_ref *ref, class loop *loop) |
0 | 1432 { |
111 | 1433 while (loop != current_loops->tree_root |
1434 && set_ref_stored_in_loop (ref, loop)) | |
1435 loop = loop_outer (loop); | |
0 | 1436 } |
1437 | |
1438 /* Gathers memory references in statement STMT in LOOP, storing the | |
1439 information about them in the memory_accesses structure. Marks | |
1440 the vops accessed through unrecognized statements there as | |
1441 well. */ | |
1442 | |
1443 static void | |
145 | 1444 gather_mem_refs_stmt (class loop *loop, gimple *stmt) |
0 | 1445 { |
1446 tree *mem = NULL; | |
1447 hashval_t hash; | |
111 | 1448 im_mem_ref **slot; |
1449 im_mem_ref *ref; | |
0 | 1450 bool is_stored; |
1451 unsigned id; | |
1452 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
1453 if (!gimple_vuse (stmt)) |
0 | 1454 return; |
1455 | |
1456 mem = simple_mem_ref_in_stmt (stmt, &is_stored); | |
1457 if (!mem) | |
1458 { | |
111 | 1459 /* We use the shared mem_ref for all unanalyzable refs. */ |
1460 id = UNANALYZABLE_MEM_ID; | |
1461 ref = memory_accesses.refs_list[id]; | |
1462 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1463 { | |
1464 fprintf (dump_file, "Unanalyzed memory reference %u: ", id); | |
1465 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); | |
1466 } | |
1467 is_stored = gimple_vdef (stmt); | |
0 | 1468 } |
1469 else | |
1470 { | |
145 | 1471 /* We are looking for equal refs that might differ in structure |
1472 such as a.b vs. MEM[&a + 4]. So we key off the ao_ref but | |
1473 make sure we can canonicalize the ref in the hashtable if | |
1474 non-operand_equal_p refs are found. For the lookup we mark | |
1475 the case we want strict equality with aor.max_size == -1. */ | |
1476 ao_ref aor; | |
1477 ao_ref_init (&aor, *mem); | |
1478 ao_ref_base (&aor); | |
1479 ao_ref_alias_set (&aor); | |
1480 HOST_WIDE_INT offset, size, max_size; | |
1481 poly_int64 saved_maxsize = aor.max_size, mem_off; | |
1482 tree mem_base; | |
1483 bool ref_decomposed; | |
1484 if (aor.max_size_known_p () | |
1485 && aor.offset.is_constant (&offset) | |
1486 && aor.size.is_constant (&size) | |
1487 && aor.max_size.is_constant (&max_size) | |
1488 && size == max_size | |
1489 && (size % BITS_PER_UNIT) == 0 | |
1490 /* We're canonicalizing to a MEM where TYPE_SIZE specifies the | |
1491 size. Make sure this is consistent with the extraction. */ | |
1492 && poly_int_tree_p (TYPE_SIZE (TREE_TYPE (*mem))) | |
1493 && known_eq (wi::to_poly_offset (TYPE_SIZE (TREE_TYPE (*mem))), | |
1494 aor.size) | |
1495 && (mem_base = get_addr_base_and_unit_offset (aor.ref, &mem_off))) | |
1496 { | |
1497 ref_decomposed = true; | |
1498 hash = iterative_hash_expr (ao_ref_base (&aor), 0); | |
1499 hash = iterative_hash_host_wide_int (offset, hash); | |
1500 hash = iterative_hash_host_wide_int (size, hash); | |
1501 } | |
1502 else | |
1503 { | |
1504 ref_decomposed = false; | |
1505 hash = iterative_hash_expr (aor.ref, 0); | |
1506 aor.max_size = -1; | |
1507 } | |
1508 slot = memory_accesses.refs->find_slot_with_hash (&aor, hash, INSERT); | |
1509 aor.max_size = saved_maxsize; | |
111 | 1510 if (*slot) |
1511 { | |
145 | 1512 if (!(*slot)->ref_canonical |
1513 && !operand_equal_p (*mem, (*slot)->mem.ref, 0)) | |
1514 { | |
1515 /* If we didn't yet canonicalize the hashtable ref (which | |
1516 we'll end up using for code insertion) and hit a second | |
1517 equal ref that is not structurally equivalent create | |
1518 a canonical ref which is a bare MEM_REF. */ | |
1519 if (TREE_CODE (*mem) == MEM_REF | |
1520 || TREE_CODE (*mem) == TARGET_MEM_REF) | |
1521 { | |
1522 (*slot)->mem.ref = *mem; | |
1523 (*slot)->mem.base_alias_set = ao_ref_base_alias_set (&aor); | |
1524 } | |
1525 else | |
1526 { | |
1527 tree ref_alias_type = reference_alias_ptr_type (*mem); | |
1528 unsigned int ref_align = get_object_alignment (*mem); | |
1529 tree ref_type = TREE_TYPE (*mem); | |
1530 tree tmp = build_fold_addr_expr (unshare_expr (mem_base)); | |
1531 if (TYPE_ALIGN (ref_type) != ref_align) | |
1532 ref_type = build_aligned_type (ref_type, ref_align); | |
1533 (*slot)->mem.ref | |
1534 = fold_build2 (MEM_REF, ref_type, tmp, | |
1535 build_int_cst (ref_alias_type, mem_off)); | |
1536 if ((*slot)->mem.volatile_p) | |
1537 TREE_THIS_VOLATILE ((*slot)->mem.ref) = 1; | |
1538 gcc_checking_assert (TREE_CODE ((*slot)->mem.ref) == MEM_REF | |
1539 && is_gimple_mem_ref_addr | |
1540 (TREE_OPERAND ((*slot)->mem.ref, | |
1541 0))); | |
1542 (*slot)->mem.base_alias_set = (*slot)->mem.ref_alias_set; | |
1543 } | |
1544 (*slot)->ref_canonical = true; | |
1545 } | |
111 | 1546 ref = *slot; |
1547 id = ref->id; | |
1548 } | |
1549 else | |
0 | 1550 { |
111 | 1551 id = memory_accesses.refs_list.length (); |
145 | 1552 ref = mem_ref_alloc (&aor, hash, id); |
1553 ref->ref_decomposed = ref_decomposed; | |
111 | 1554 memory_accesses.refs_list.safe_push (ref); |
1555 *slot = ref; | |
1556 | |
1557 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1558 { | |
1559 fprintf (dump_file, "Memory reference %u: ", id); | |
1560 print_generic_expr (dump_file, ref->mem.ref, TDF_SLIM); | |
1561 fprintf (dump_file, "\n"); | |
1562 } | |
0 | 1563 } |
111 | 1564 |
1565 record_mem_ref_loc (ref, stmt, mem); | |
1566 } | |
1567 bitmap_set_bit (&memory_accesses.refs_in_loop[loop->num], ref->id); | |
1568 if (is_stored) | |
1569 { | |
1570 bitmap_set_bit (&memory_accesses.refs_stored_in_loop[loop->num], ref->id); | |
1571 mark_ref_stored (ref, loop); | |
0 | 1572 } |
111 | 1573 init_lim_data (stmt)->ref = ref->id; |
0 | 1574 return; |
111 | 1575 } |
1576 | |
1577 static unsigned *bb_loop_postorder; | |
1578 | |
1579 /* qsort sort function to sort blocks after their loop fathers postorder. */ | |
1580 | |
1581 static int | |
145 | 1582 sort_bbs_in_loop_postorder_cmp (const void *bb1_, const void *bb2_, |
1583 void *bb_loop_postorder_) | |
111 | 1584 { |
145 | 1585 unsigned *bb_loop_postorder = (unsigned *)bb_loop_postorder_; |
1586 basic_block bb1 = *(const basic_block *)bb1_; | |
1587 basic_block bb2 = *(const basic_block *)bb2_; | |
1588 class loop *loop1 = bb1->loop_father; | |
1589 class loop *loop2 = bb2->loop_father; | |
111 | 1590 if (loop1->num == loop2->num) |
131 | 1591 return bb1->index - bb2->index; |
111 | 1592 return bb_loop_postorder[loop1->num] < bb_loop_postorder[loop2->num] ? -1 : 1; |
1593 } | |
1594 | |
1595 /* qsort sort function to sort ref locs after their loop fathers postorder. */ | |
1596 | |
1597 static int | |
145 | 1598 sort_locs_in_loop_postorder_cmp (const void *loc1_, const void *loc2_, |
1599 void *bb_loop_postorder_) | |
111 | 1600 { |
145 | 1601 unsigned *bb_loop_postorder = (unsigned *)bb_loop_postorder_; |
1602 const mem_ref_loc *loc1 = (const mem_ref_loc *)loc1_; | |
1603 const mem_ref_loc *loc2 = (const mem_ref_loc *)loc2_; | |
1604 class loop *loop1 = gimple_bb (loc1->stmt)->loop_father; | |
1605 class loop *loop2 = gimple_bb (loc2->stmt)->loop_father; | |
111 | 1606 if (loop1->num == loop2->num) |
1607 return 0; | |
1608 return bb_loop_postorder[loop1->num] < bb_loop_postorder[loop2->num] ? -1 : 1; | |
0 | 1609 } |
1610 | |
1611 /* Gathers memory references in loops. */ | |
1612 | |
1613 static void | |
111 | 1614 analyze_memory_references (void) |
0 | 1615 { |
1616 gimple_stmt_iterator bsi; | |
111 | 1617 basic_block bb, *bbs; |
145 | 1618 class loop *loop, *outer; |
111 | 1619 unsigned i, n; |
1620 | |
1621 /* Collect all basic-blocks in loops and sort them after their | |
1622 loops postorder. */ | |
1623 i = 0; | |
1624 bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS); | |
1625 FOR_EACH_BB_FN (bb, cfun) | |
1626 if (bb->loop_father != current_loops->tree_root) | |
1627 bbs[i++] = bb; | |
1628 n = i; | |
145 | 1629 gcc_sort_r (bbs, n, sizeof (basic_block), sort_bbs_in_loop_postorder_cmp, |
1630 bb_loop_postorder); | |
111 | 1631 |
1632 /* Visit blocks in loop postorder and assign mem-ref IDs in that order. | |
1633 That results in better locality for all the bitmaps. */ | |
1634 for (i = 0; i < n; ++i) | |
0 | 1635 { |
111 | 1636 basic_block bb = bbs[i]; |
1637 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1638 gather_mem_refs_stmt (bb->loop_father, gsi_stmt (bsi)); | |
0 | 1639 } |
111 | 1640 |
1641 /* Sort the location list of gathered memory references after their | |
1642 loop postorder number. */ | |
1643 im_mem_ref *ref; | |
1644 FOR_EACH_VEC_ELT (memory_accesses.refs_list, i, ref) | |
145 | 1645 ref->accesses_in_loop.sort (sort_locs_in_loop_postorder_cmp, |
1646 bb_loop_postorder); | |
111 | 1647 |
1648 free (bbs); | |
1649 | |
1650 /* Propagate the information about accessed memory references up | |
1651 the loop hierarchy. */ | |
1652 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) | |
0 | 1653 { |
111 | 1654 /* Finalize the overall touched references (including subloops). */ |
1655 bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[loop->num], | |
1656 &memory_accesses.refs_stored_in_loop[loop->num]); | |
1657 | |
1658 /* Propagate the information about accessed memory references up | |
1659 the loop hierarchy. */ | |
1660 outer = loop_outer (loop); | |
1661 if (outer == current_loops->tree_root) | |
1662 continue; | |
1663 | |
1664 bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[outer->num], | |
1665 &memory_accesses.all_refs_stored_in_loop[loop->num]); | |
0 | 1666 } |
1667 } | |
1668 | |
1669 /* Returns true if MEM1 and MEM2 may alias. TTAE_CACHE is used as a cache in | |
1670 tree_to_aff_combination_expand. */ | |
1671 | |
1672 static bool | |
111 | 1673 mem_refs_may_alias_p (im_mem_ref *mem1, im_mem_ref *mem2, |
1674 hash_map<tree, name_expansion *> **ttae_cache) | |
0 | 1675 { |
1676 /* Perform BASE + OFFSET analysis -- if MEM1 and MEM2 are based on the same | |
1677 object and their offset differ in such a way that the locations cannot | |
1678 overlap, then they cannot alias. */ | |
131 | 1679 poly_widest_int size1, size2; |
0 | 1680 aff_tree off1, off2; |
1681 | |
1682 /* Perform basic offset and type-based disambiguation. */ | |
111 | 1683 if (!refs_may_alias_p_1 (&mem1->mem, &mem2->mem, true)) |
0 | 1684 return false; |
1685 | |
1686 /* The expansion of addresses may be a bit expensive, thus we only do | |
1687 the check at -O2 and higher optimization levels. */ | |
1688 if (optimize < 2) | |
1689 return true; | |
1690 | |
111 | 1691 get_inner_reference_aff (mem1->mem.ref, &off1, &size1); |
1692 get_inner_reference_aff (mem2->mem.ref, &off2, &size2); | |
0 | 1693 aff_combination_expand (&off1, ttae_cache); |
1694 aff_combination_expand (&off2, ttae_cache); | |
111 | 1695 aff_combination_scale (&off1, -1); |
0 | 1696 aff_combination_add (&off2, &off1); |
1697 | |
111 | 1698 if (aff_comb_cannot_overlap_p (&off2, size1, size2)) |
0 | 1699 return false; |
1700 | |
1701 return true; | |
1702 } | |
1703 | |
111 | 1704 /* Compare function for bsearch searching for reference locations |
1705 in a loop. */ | |
1706 | |
1707 static int | |
145 | 1708 find_ref_loc_in_loop_cmp (const void *loop_, const void *loc_, |
1709 void *bb_loop_postorder_) | |
111 | 1710 { |
145 | 1711 unsigned *bb_loop_postorder = (unsigned *)bb_loop_postorder_; |
1712 class loop *loop = (class loop *)const_cast<void *>(loop_); | |
111 | 1713 mem_ref_loc *loc = (mem_ref_loc *)const_cast<void *>(loc_); |
145 | 1714 class loop *loc_loop = gimple_bb (loc->stmt)->loop_father; |
111 | 1715 if (loop->num == loc_loop->num |
1716 || flow_loop_nested_p (loop, loc_loop)) | |
1717 return 0; | |
1718 return (bb_loop_postorder[loop->num] < bb_loop_postorder[loc_loop->num] | |
1719 ? -1 : 1); | |
1720 } | |
1721 | |
1722 /* Iterates over all locations of REF in LOOP and its subloops calling | |
1723 fn.operator() with the location as argument. When that operator | |
1724 returns true the iteration is stopped and true is returned. | |
1725 Otherwise false is returned. */ | |
1726 | |
1727 template <typename FN> | |
1728 static bool | |
145 | 1729 for_all_locs_in_loop (class loop *loop, im_mem_ref *ref, FN fn) |
111 | 1730 { |
1731 unsigned i; | |
1732 mem_ref_loc *loc; | |
1733 | |
1734 /* Search for the cluster of locs in the accesses_in_loop vector | |
1735 which is sorted after postorder index of the loop father. */ | |
145 | 1736 loc = ref->accesses_in_loop.bsearch (loop, find_ref_loc_in_loop_cmp, |
1737 bb_loop_postorder); | |
111 | 1738 if (!loc) |
1739 return false; | |
1740 | |
1741 /* We have found one location inside loop or its sub-loops. Iterate | |
1742 both forward and backward to cover the whole cluster. */ | |
1743 i = loc - ref->accesses_in_loop.address (); | |
1744 while (i > 0) | |
1745 { | |
1746 --i; | |
1747 mem_ref_loc *l = &ref->accesses_in_loop[i]; | |
1748 if (!flow_bb_inside_loop_p (loop, gimple_bb (l->stmt))) | |
1749 break; | |
1750 if (fn (l)) | |
1751 return true; | |
1752 } | |
1753 for (i = loc - ref->accesses_in_loop.address (); | |
1754 i < ref->accesses_in_loop.length (); ++i) | |
1755 { | |
1756 mem_ref_loc *l = &ref->accesses_in_loop[i]; | |
1757 if (!flow_bb_inside_loop_p (loop, gimple_bb (l->stmt))) | |
1758 break; | |
1759 if (fn (l)) | |
1760 return true; | |
1761 } | |
1762 | |
1763 return false; | |
1764 } | |
1765 | |
0 | 1766 /* Rewrites location LOC by TMP_VAR. */ |
1767 | |
145 | 1768 class rewrite_mem_ref_loc |
0 | 1769 { |
145 | 1770 public: |
111 | 1771 rewrite_mem_ref_loc (tree tmp_var_) : tmp_var (tmp_var_) {} |
1772 bool operator () (mem_ref_loc *loc); | |
1773 tree tmp_var; | |
1774 }; | |
1775 | |
1776 bool | |
1777 rewrite_mem_ref_loc::operator () (mem_ref_loc *loc) | |
1778 { | |
0 | 1779 *loc->ref = tmp_var; |
1780 update_stmt (loc->stmt); | |
111 | 1781 return false; |
0 | 1782 } |
1783 | |
1784 /* Rewrites all references to REF in LOOP by variable TMP_VAR. */ | |
1785 | |
1786 static void | |
145 | 1787 rewrite_mem_refs (class loop *loop, im_mem_ref *ref, tree tmp_var) |
111 | 1788 { |
1789 for_all_locs_in_loop (loop, ref, rewrite_mem_ref_loc (tmp_var)); | |
1790 } | |
1791 | |
1792 /* Stores the first reference location in LOCP. */ | |
1793 | |
145 | 1794 class first_mem_ref_loc_1 |
0 | 1795 { |
145 | 1796 public: |
111 | 1797 first_mem_ref_loc_1 (mem_ref_loc **locp_) : locp (locp_) {} |
1798 bool operator () (mem_ref_loc *loc); | |
1799 mem_ref_loc **locp; | |
1800 }; | |
1801 | |
1802 bool | |
1803 first_mem_ref_loc_1::operator () (mem_ref_loc *loc) | |
1804 { | |
1805 *locp = loc; | |
1806 return true; | |
1807 } | |
1808 | |
1809 /* Returns the first reference location to REF in LOOP. */ | |
1810 | |
1811 static mem_ref_loc * | |
145 | 1812 first_mem_ref_loc (class loop *loop, im_mem_ref *ref) |
111 | 1813 { |
1814 mem_ref_loc *locp = NULL; | |
1815 for_all_locs_in_loop (loop, ref, first_mem_ref_loc_1 (&locp)); | |
1816 return locp; | |
0 | 1817 } |
1818 | |
111 | 1819 struct prev_flag_edges { |
1820 /* Edge to insert new flag comparison code. */ | |
1821 edge append_cond_position; | |
1822 | |
1823 /* Edge for fall through from previous flag comparison. */ | |
1824 edge last_cond_fallthru; | |
1825 }; | |
1826 | |
1827 /* Helper function for execute_sm. Emit code to store TMP_VAR into | |
1828 MEM along edge EX. | |
1829 | |
1830 The store is only done if MEM has changed. We do this so no | |
1831 changes to MEM occur on code paths that did not originally store | |
1832 into it. | |
1833 | |
1834 The common case for execute_sm will transform: | |
1835 | |
1836 for (...) { | |
1837 if (foo) | |
1838 stuff; | |
1839 else | |
1840 MEM = TMP_VAR; | |
1841 } | |
1842 | |
1843 into: | |
1844 | |
1845 lsm = MEM; | |
1846 for (...) { | |
1847 if (foo) | |
1848 stuff; | |
1849 else | |
1850 lsm = TMP_VAR; | |
1851 } | |
1852 MEM = lsm; | |
1853 | |
1854 This function will generate: | |
1855 | |
1856 lsm = MEM; | |
1857 | |
1858 lsm_flag = false; | |
1859 ... | |
1860 for (...) { | |
1861 if (foo) | |
1862 stuff; | |
1863 else { | |
1864 lsm = TMP_VAR; | |
1865 lsm_flag = true; | |
1866 } | |
1867 } | |
1868 if (lsm_flag) <-- | |
1869 MEM = lsm; <-- | |
1870 */ | |
0 | 1871 |
1872 static void | |
111 | 1873 execute_sm_if_changed (edge ex, tree mem, tree tmp_var, tree flag, |
1874 edge preheader, hash_set <basic_block> *flag_bbs) | |
0 | 1875 { |
111 | 1876 basic_block new_bb, then_bb, old_dest; |
1877 bool loop_has_only_one_exit; | |
1878 edge then_old_edge, orig_ex = ex; | |
1879 gimple_stmt_iterator gsi; | |
1880 gimple *stmt; | |
1881 struct prev_flag_edges *prev_edges = (struct prev_flag_edges *) ex->aux; | |
1882 bool irr = ex->flags & EDGE_IRREDUCIBLE_LOOP; | |
1883 | |
1884 profile_count count_sum = profile_count::zero (); | |
1885 int nbbs = 0, ncount = 0; | |
1886 profile_probability flag_probability = profile_probability::uninitialized (); | |
1887 | |
1888 /* Flag is set in FLAG_BBS. Determine probability that flag will be true | |
1889 at loop exit. | |
1890 | |
145 | 1891 This code may look fancy, but it cannot update profile very realistically |
111 | 1892 because we do not know the probability that flag will be true at given |
1893 loop exit. | |
1894 | |
1895 We look for two interesting extremes | |
1896 - when exit is dominated by block setting the flag, we know it will | |
1897 always be true. This is a common case. | |
1898 - when all blocks setting the flag have very low frequency we know | |
1899 it will likely be false. | |
1900 In all other cases we default to 2/3 for flag being true. */ | |
1901 | |
1902 for (hash_set<basic_block>::iterator it = flag_bbs->begin (); | |
1903 it != flag_bbs->end (); ++it) | |
1904 { | |
1905 if ((*it)->count.initialized_p ()) | |
1906 count_sum += (*it)->count, ncount ++; | |
1907 if (dominated_by_p (CDI_DOMINATORS, ex->src, *it)) | |
1908 flag_probability = profile_probability::always (); | |
1909 nbbs++; | |
1910 } | |
1911 | |
1912 profile_probability cap = profile_probability::always ().apply_scale (2, 3); | |
1913 | |
1914 if (flag_probability.initialized_p ()) | |
1915 ; | |
131 | 1916 else if (ncount == nbbs |
1917 && preheader->count () >= count_sum && preheader->count ().nonzero_p ()) | |
111 | 1918 { |
1919 flag_probability = count_sum.probability_in (preheader->count ()); | |
1920 if (flag_probability > cap) | |
1921 flag_probability = cap; | |
1922 } | |
131 | 1923 |
1924 if (!flag_probability.initialized_p ()) | |
111 | 1925 flag_probability = cap; |
1926 | |
1927 /* ?? Insert store after previous store if applicable. See note | |
1928 below. */ | |
1929 if (prev_edges) | |
1930 ex = prev_edges->append_cond_position; | |
1931 | |
1932 loop_has_only_one_exit = single_pred_p (ex->dest); | |
1933 | |
1934 if (loop_has_only_one_exit) | |
1935 ex = split_block_after_labels (ex->dest); | |
1936 else | |
0 | 1937 { |
111 | 1938 for (gphi_iterator gpi = gsi_start_phis (ex->dest); |
1939 !gsi_end_p (gpi); gsi_next (&gpi)) | |
1940 { | |
1941 gphi *phi = gpi.phi (); | |
1942 if (virtual_operand_p (gimple_phi_result (phi))) | |
1943 continue; | |
1944 | |
1945 /* When the destination has a non-virtual PHI node with multiple | |
1946 predecessors make sure we preserve the PHI structure by | |
1947 forcing a forwarder block so that hoisting of that PHI will | |
1948 still work. */ | |
1949 split_edge (ex); | |
1950 break; | |
1951 } | |
1952 } | |
1953 | |
1954 old_dest = ex->dest; | |
1955 new_bb = split_edge (ex); | |
1956 then_bb = create_empty_bb (new_bb); | |
1957 then_bb->count = new_bb->count.apply_probability (flag_probability); | |
1958 if (irr) | |
1959 then_bb->flags = BB_IRREDUCIBLE_LOOP; | |
1960 add_bb_to_loop (then_bb, new_bb->loop_father); | |
1961 | |
1962 gsi = gsi_start_bb (new_bb); | |
1963 stmt = gimple_build_cond (NE_EXPR, flag, boolean_false_node, | |
1964 NULL_TREE, NULL_TREE); | |
1965 gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
1966 | |
1967 gsi = gsi_start_bb (then_bb); | |
1968 /* Insert actual store. */ | |
1969 stmt = gimple_build_assign (unshare_expr (mem), tmp_var); | |
1970 gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
1971 | |
1972 edge e1 = single_succ_edge (new_bb); | |
1973 edge e2 = make_edge (new_bb, then_bb, | |
1974 EDGE_TRUE_VALUE | (irr ? EDGE_IRREDUCIBLE_LOOP : 0)); | |
1975 e2->probability = flag_probability; | |
1976 | |
1977 e1->flags |= EDGE_FALSE_VALUE | (irr ? EDGE_IRREDUCIBLE_LOOP : 0); | |
1978 e1->flags &= ~EDGE_FALLTHRU; | |
1979 | |
1980 e1->probability = flag_probability.invert (); | |
1981 | |
1982 then_old_edge = make_single_succ_edge (then_bb, old_dest, | |
1983 EDGE_FALLTHRU | (irr ? EDGE_IRREDUCIBLE_LOOP : 0)); | |
1984 | |
1985 set_immediate_dominator (CDI_DOMINATORS, then_bb, new_bb); | |
1986 | |
1987 if (prev_edges) | |
1988 { | |
1989 basic_block prevbb = prev_edges->last_cond_fallthru->src; | |
1990 redirect_edge_succ (prev_edges->last_cond_fallthru, new_bb); | |
1991 set_immediate_dominator (CDI_DOMINATORS, new_bb, prevbb); | |
1992 set_immediate_dominator (CDI_DOMINATORS, old_dest, | |
1993 recompute_dominator (CDI_DOMINATORS, old_dest)); | |
0 | 1994 } |
111 | 1995 |
1996 /* ?? Because stores may alias, they must happen in the exact | |
1997 sequence they originally happened. Save the position right after | |
1998 the (_lsm) store we just created so we can continue appending after | |
1999 it and maintain the original order. */ | |
2000 { | |
2001 struct prev_flag_edges *p; | |
2002 | |
2003 if (orig_ex->aux) | |
2004 orig_ex->aux = NULL; | |
2005 alloc_aux_for_edge (orig_ex, sizeof (struct prev_flag_edges)); | |
2006 p = (struct prev_flag_edges *) orig_ex->aux; | |
2007 p->append_cond_position = then_old_edge; | |
2008 p->last_cond_fallthru = find_edge (new_bb, old_dest); | |
2009 orig_ex->aux = (void *) p; | |
2010 } | |
2011 | |
2012 if (!loop_has_only_one_exit) | |
2013 for (gphi_iterator gpi = gsi_start_phis (old_dest); | |
2014 !gsi_end_p (gpi); gsi_next (&gpi)) | |
2015 { | |
2016 gphi *phi = gpi.phi (); | |
2017 unsigned i; | |
2018 | |
2019 for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2020 if (gimple_phi_arg_edge (phi, i)->src == new_bb) | |
2021 { | |
2022 tree arg = gimple_phi_arg_def (phi, i); | |
2023 add_phi_arg (phi, arg, then_old_edge, UNKNOWN_LOCATION); | |
2024 update_stmt (phi); | |
2025 } | |
2026 } | |
0 | 2027 } |
2028 | |
111 | 2029 /* When REF is set on the location, set flag indicating the store. */ |
2030 | |
145 | 2031 class sm_set_flag_if_changed |
0 | 2032 { |
145 | 2033 public: |
111 | 2034 sm_set_flag_if_changed (tree flag_, hash_set <basic_block> *bbs_) |
2035 : flag (flag_), bbs (bbs_) {} | |
2036 bool operator () (mem_ref_loc *loc); | |
2037 tree flag; | |
2038 hash_set <basic_block> *bbs; | |
2039 }; | |
2040 | |
2041 bool | |
2042 sm_set_flag_if_changed::operator () (mem_ref_loc *loc) | |
2043 { | |
2044 /* Only set the flag for writes. */ | |
2045 if (is_gimple_assign (loc->stmt) | |
2046 && gimple_assign_lhs_ptr (loc->stmt) == loc->ref) | |
0 | 2047 { |
111 | 2048 gimple_stmt_iterator gsi = gsi_for_stmt (loc->stmt); |
2049 gimple *stmt = gimple_build_assign (flag, boolean_true_node); | |
2050 gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2051 bbs->add (gimple_bb (stmt)); | |
0 | 2052 } |
111 | 2053 return false; |
2054 } | |
2055 | |
2056 /* Helper function for execute_sm. On every location where REF is | |
2057 set, set an appropriate flag indicating the store. */ | |
2058 | |
2059 static tree | |
145 | 2060 execute_sm_if_changed_flag_set (class loop *loop, im_mem_ref *ref, |
111 | 2061 hash_set <basic_block> *bbs) |
2062 { | |
2063 tree flag; | |
2064 char *str = get_lsm_tmp_name (ref->mem.ref, ~0, "_flag"); | |
2065 flag = create_tmp_reg (boolean_type_node, str); | |
2066 for_all_locs_in_loop (loop, ref, sm_set_flag_if_changed (flag, bbs)); | |
2067 return flag; | |
0 | 2068 } |
2069 | |
2070 /* Executes store motion of memory reference REF from LOOP. | |
2071 Exits from the LOOP are stored in EXITS. The initialization of the | |
2072 temporary variable is put to the preheader of the loop, and assignments | |
2073 to the reference from the temporary variable are emitted to exits. */ | |
2074 | |
2075 static void | |
145 | 2076 execute_sm (class loop *loop, vec<edge> exits, im_mem_ref *ref) |
0 | 2077 { |
111 | 2078 tree tmp_var, store_flag = NULL_TREE; |
0 | 2079 unsigned i; |
111 | 2080 gassign *load; |
0 | 2081 struct fmt_data fmt_data; |
2082 edge ex; | |
2083 struct lim_aux_data *lim_data; | |
111 | 2084 bool multi_threaded_model_p = false; |
2085 gimple_stmt_iterator gsi; | |
2086 hash_set<basic_block> flag_bbs; | |
0 | 2087 |
2088 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2089 { | |
2090 fprintf (dump_file, "Executing store motion of "); | |
111 | 2091 print_generic_expr (dump_file, ref->mem.ref); |
0 | 2092 fprintf (dump_file, " from loop %d\n", loop->num); |
2093 } | |
2094 | |
111 | 2095 tmp_var = create_tmp_reg (TREE_TYPE (ref->mem.ref), |
2096 get_lsm_tmp_name (ref->mem.ref, ~0)); | |
0 | 2097 |
2098 fmt_data.loop = loop; | |
2099 fmt_data.orig_loop = loop; | |
111 | 2100 for_each_index (&ref->mem.ref, force_move_till, &fmt_data); |
2101 | |
2102 if (bb_in_transaction (loop_preheader_edge (loop)->src) | |
145 | 2103 || (! flag_store_data_races |
111 | 2104 && ! ref_always_accessed_p (loop, ref, true))) |
2105 multi_threaded_model_p = true; | |
2106 | |
2107 if (multi_threaded_model_p) | |
2108 store_flag = execute_sm_if_changed_flag_set (loop, ref, &flag_bbs); | |
0 | 2109 |
2110 rewrite_mem_refs (loop, ref, tmp_var); | |
2111 | |
111 | 2112 /* Emit the load code on a random exit edge or into the latch if |
2113 the loop does not exit, so that we are sure it will be processed | |
2114 by move_computations after all dependencies. */ | |
2115 gsi = gsi_for_stmt (first_mem_ref_loc (loop, ref)->stmt); | |
2116 | |
2117 /* FIXME/TODO: For the multi-threaded variant, we could avoid this | |
2118 load altogether, since the store is predicated by a flag. We | |
2119 could, do the load only if it was originally in the loop. */ | |
2120 load = gimple_build_assign (tmp_var, unshare_expr (ref->mem.ref)); | |
0 | 2121 lim_data = init_lim_data (load); |
2122 lim_data->max_loop = loop; | |
2123 lim_data->tgt_loop = loop; | |
111 | 2124 gsi_insert_before (&gsi, load, GSI_SAME_STMT); |
2125 | |
2126 if (multi_threaded_model_p) | |
0 | 2127 { |
111 | 2128 load = gimple_build_assign (store_flag, boolean_false_node); |
2129 lim_data = init_lim_data (load); | |
2130 lim_data->max_loop = loop; | |
2131 lim_data->tgt_loop = loop; | |
2132 gsi_insert_before (&gsi, load, GSI_SAME_STMT); | |
0 | 2133 } |
111 | 2134 |
2135 /* Sink the store to every exit from the loop. */ | |
2136 FOR_EACH_VEC_ELT (exits, i, ex) | |
2137 if (!multi_threaded_model_p) | |
2138 { | |
2139 gassign *store; | |
2140 store = gimple_build_assign (unshare_expr (ref->mem.ref), tmp_var); | |
2141 gsi_insert_on_edge (ex, store); | |
2142 } | |
2143 else | |
2144 execute_sm_if_changed (ex, ref->mem.ref, tmp_var, store_flag, | |
2145 loop_preheader_edge (loop), &flag_bbs); | |
0 | 2146 } |
2147 | |
2148 /* Hoists memory references MEM_REFS out of LOOP. EXITS is the list of exit | |
2149 edges of the LOOP. */ | |
2150 | |
2151 static void | |
145 | 2152 hoist_memory_references (class loop *loop, bitmap mem_refs, |
111 | 2153 vec<edge> exits) |
0 | 2154 { |
111 | 2155 im_mem_ref *ref; |
0 | 2156 unsigned i; |
2157 bitmap_iterator bi; | |
2158 | |
2159 EXECUTE_IF_SET_IN_BITMAP (mem_refs, 0, i, bi) | |
2160 { | |
111 | 2161 ref = memory_accesses.refs_list[i]; |
0 | 2162 execute_sm (loop, exits, ref); |
2163 } | |
2164 } | |
2165 | |
145 | 2166 class ref_always_accessed |
111 | 2167 { |
145 | 2168 public: |
2169 ref_always_accessed (class loop *loop_, bool stored_p_) | |
111 | 2170 : loop (loop_), stored_p (stored_p_) {} |
2171 bool operator () (mem_ref_loc *loc); | |
145 | 2172 class loop *loop; |
111 | 2173 bool stored_p; |
2174 }; | |
2175 | |
2176 bool | |
2177 ref_always_accessed::operator () (mem_ref_loc *loc) | |
2178 { | |
145 | 2179 class loop *must_exec; |
111 | 2180 |
2181 if (!get_lim_data (loc->stmt)) | |
2182 return false; | |
2183 | |
2184 /* If we require an always executed store make sure the statement | |
2185 stores to the reference. */ | |
2186 if (stored_p) | |
2187 { | |
2188 tree lhs = gimple_get_lhs (loc->stmt); | |
2189 if (!lhs | |
2190 || lhs != *loc->ref) | |
2191 return false; | |
2192 } | |
2193 | |
2194 must_exec = get_lim_data (loc->stmt)->always_executed_in; | |
2195 if (!must_exec) | |
2196 return false; | |
2197 | |
2198 if (must_exec == loop | |
2199 || flow_loop_nested_p (must_exec, loop)) | |
2200 return true; | |
2201 | |
2202 return false; | |
2203 } | |
2204 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2205 /* Returns true if REF is always accessed in LOOP. If STORED_P is true |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2206 make sure REF is always stored to in LOOP. */ |
0 | 2207 |
2208 static bool | |
145 | 2209 ref_always_accessed_p (class loop *loop, im_mem_ref *ref, bool stored_p) |
0 | 2210 { |
111 | 2211 return for_all_locs_in_loop (loop, ref, |
2212 ref_always_accessed (loop, stored_p)); | |
0 | 2213 } |
2214 | |
2215 /* Returns true if REF1 and REF2 are independent. */ | |
2216 | |
2217 static bool | |
111 | 2218 refs_independent_p (im_mem_ref *ref1, im_mem_ref *ref2) |
0 | 2219 { |
111 | 2220 if (ref1 == ref2) |
0 | 2221 return true; |
2222 | |
2223 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2224 fprintf (dump_file, "Querying dependency of refs %u and %u: ", | |
2225 ref1->id, ref2->id); | |
2226 | |
111 | 2227 if (mem_refs_may_alias_p (ref1, ref2, &memory_accesses.ttae_cache)) |
0 | 2228 { |
2229 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2230 fprintf (dump_file, "dependent.\n"); | |
2231 return false; | |
2232 } | |
2233 else | |
2234 { | |
2235 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2236 fprintf (dump_file, "independent.\n"); | |
2237 return true; | |
2238 } | |
2239 } | |
2240 | |
111 | 2241 /* Mark REF dependent on stores or loads (according to STORED_P) in LOOP |
2242 and its super-loops. */ | |
0 | 2243 |
2244 static void | |
145 | 2245 record_dep_loop (class loop *loop, im_mem_ref *ref, bool stored_p) |
111 | 2246 { |
2247 /* We can propagate dependent-in-loop bits up the loop | |
2248 hierarchy to all outer loops. */ | |
2249 while (loop != current_loops->tree_root | |
2250 && bitmap_set_bit (&ref->dep_loop, LOOP_DEP_BIT (loop->num, stored_p))) | |
2251 loop = loop_outer (loop); | |
2252 } | |
2253 | |
2254 /* Returns true if REF is independent on all other memory | |
131 | 2255 references in LOOP. */ |
111 | 2256 |
2257 static bool | |
145 | 2258 ref_indep_loop_p_1 (class loop *loop, im_mem_ref *ref, bool stored_p) |
0 | 2259 { |
111 | 2260 stored_p |= (ref->stored && bitmap_bit_p (ref->stored, loop->num)); |
2261 | |
2262 bool indep_p = true; | |
2263 bitmap refs_to_check; | |
2264 | |
2265 if (stored_p) | |
2266 refs_to_check = &memory_accesses.refs_in_loop[loop->num]; | |
2267 else | |
2268 refs_to_check = &memory_accesses.refs_stored_in_loop[loop->num]; | |
2269 | |
2270 if (bitmap_bit_p (refs_to_check, UNANALYZABLE_MEM_ID)) | |
2271 indep_p = false; | |
0 | 2272 else |
111 | 2273 { |
2274 if (bitmap_bit_p (&ref->indep_loop, LOOP_DEP_BIT (loop->num, stored_p))) | |
2275 return true; | |
2276 if (bitmap_bit_p (&ref->dep_loop, LOOP_DEP_BIT (loop->num, stored_p))) | |
2277 return false; | |
2278 | |
145 | 2279 class loop *inner = loop->inner; |
111 | 2280 while (inner) |
2281 { | |
131 | 2282 if (!ref_indep_loop_p_1 (inner, ref, stored_p)) |
111 | 2283 { |
2284 indep_p = false; | |
2285 break; | |
2286 } | |
2287 inner = inner->next; | |
2288 } | |
2289 | |
2290 if (indep_p) | |
2291 { | |
2292 unsigned i; | |
2293 bitmap_iterator bi; | |
2294 EXECUTE_IF_SET_IN_BITMAP (refs_to_check, 0, i, bi) | |
2295 { | |
2296 im_mem_ref *aref = memory_accesses.refs_list[i]; | |
2297 if (!refs_independent_p (ref, aref)) | |
2298 { | |
2299 indep_p = false; | |
2300 break; | |
2301 } | |
2302 } | |
2303 } | |
2304 } | |
2305 | |
2306 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2307 fprintf (dump_file, "Querying dependencies of ref %u in loop %d: %s\n", | |
2308 ref->id, loop->num, indep_p ? "independent" : "dependent"); | |
2309 | |
2310 /* Record the computed result in the cache. */ | |
2311 if (indep_p) | |
2312 { | |
2313 if (bitmap_set_bit (&ref->indep_loop, LOOP_DEP_BIT (loop->num, stored_p)) | |
2314 && stored_p) | |
2315 { | |
2316 /* If it's independend against all refs then it's independent | |
2317 against stores, too. */ | |
2318 bitmap_set_bit (&ref->indep_loop, LOOP_DEP_BIT (loop->num, false)); | |
2319 } | |
2320 } | |
2321 else | |
2322 { | |
2323 record_dep_loop (loop, ref, stored_p); | |
2324 if (!stored_p) | |
2325 { | |
2326 /* If it's dependent against stores it's dependent against | |
2327 all refs, too. */ | |
2328 record_dep_loop (loop, ref, true); | |
2329 } | |
2330 } | |
2331 | |
2332 return indep_p; | |
0 | 2333 } |
2334 | |
2335 /* Returns true if REF is independent on all other memory references in | |
131 | 2336 LOOP. */ |
0 | 2337 |
2338 static bool | |
145 | 2339 ref_indep_loop_p (class loop *loop, im_mem_ref *ref) |
0 | 2340 { |
111 | 2341 gcc_checking_assert (MEM_ANALYZABLE (ref)); |
2342 | |
131 | 2343 return ref_indep_loop_p_1 (loop, ref, false); |
0 | 2344 } |
2345 | |
2346 /* Returns true if we can perform store motion of REF from LOOP. */ | |
2347 | |
2348 static bool | |
145 | 2349 can_sm_ref_p (class loop *loop, im_mem_ref *ref) |
0 | 2350 { |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2351 tree base; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2352 |
111 | 2353 /* Can't hoist unanalyzable refs. */ |
2354 if (!MEM_ANALYZABLE (ref)) | |
0 | 2355 return false; |
2356 | |
2357 /* It should be movable. */ | |
111 | 2358 if (!is_gimple_reg_type (TREE_TYPE (ref->mem.ref)) |
2359 || TREE_THIS_VOLATILE (ref->mem.ref) | |
2360 || !for_each_index (&ref->mem.ref, may_move_till, loop)) | |
0 | 2361 return false; |
2362 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2363 /* If it can throw fail, we do not properly update EH info. */ |
111 | 2364 if (tree_could_throw_p (ref->mem.ref)) |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2365 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2366 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2367 /* If it can trap, it must be always executed in LOOP. |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2368 Readonly memory locations may trap when storing to them, but |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2369 tree_could_trap_p is a predicate for rvalues, so check that |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2370 explicitly. */ |
111 | 2371 base = get_base_address (ref->mem.ref); |
2372 if ((tree_could_trap_p (ref->mem.ref) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2373 || (DECL_P (base) && TREE_READONLY (base))) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2374 && !ref_always_accessed_p (loop, ref, true)) |
0 | 2375 return false; |
2376 | |
2377 /* And it must be independent on all other memory references | |
2378 in LOOP. */ | |
131 | 2379 if (!ref_indep_loop_p (loop, ref)) |
0 | 2380 return false; |
2381 | |
2382 return true; | |
2383 } | |
2384 | |
2385 /* Marks the references in LOOP for that store motion should be performed | |
2386 in REFS_TO_SM. SM_EXECUTED is the set of references for that store | |
2387 motion was performed in one of the outer loops. */ | |
2388 | |
2389 static void | |
145 | 2390 find_refs_for_sm (class loop *loop, bitmap sm_executed, bitmap refs_to_sm) |
0 | 2391 { |
111 | 2392 bitmap refs = &memory_accesses.all_refs_stored_in_loop[loop->num]; |
0 | 2393 unsigned i; |
2394 bitmap_iterator bi; | |
111 | 2395 im_mem_ref *ref; |
0 | 2396 |
2397 EXECUTE_IF_AND_COMPL_IN_BITMAP (refs, sm_executed, 0, i, bi) | |
2398 { | |
111 | 2399 ref = memory_accesses.refs_list[i]; |
0 | 2400 if (can_sm_ref_p (loop, ref)) |
2401 bitmap_set_bit (refs_to_sm, i); | |
2402 } | |
2403 } | |
2404 | |
2405 /* Checks whether LOOP (with exits stored in EXITS array) is suitable | |
2406 for a store motion optimization (i.e. whether we can insert statement | |
2407 on its exits). */ | |
2408 | |
2409 static bool | |
145 | 2410 loop_suitable_for_sm (class loop *loop ATTRIBUTE_UNUSED, |
111 | 2411 vec<edge> exits) |
0 | 2412 { |
2413 unsigned i; | |
2414 edge ex; | |
2415 | |
111 | 2416 FOR_EACH_VEC_ELT (exits, i, ex) |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2417 if (ex->flags & (EDGE_ABNORMAL | EDGE_EH)) |
0 | 2418 return false; |
2419 | |
2420 return true; | |
2421 } | |
2422 | |
2423 /* Try to perform store motion for all memory references modified inside | |
2424 LOOP. SM_EXECUTED is the bitmap of the memory references for that | |
2425 store motion was executed in one of the outer loops. */ | |
2426 | |
2427 static void | |
145 | 2428 store_motion_loop (class loop *loop, bitmap sm_executed) |
0 | 2429 { |
111 | 2430 vec<edge> exits = get_loop_exit_edges (loop); |
145 | 2431 class loop *subloop; |
111 | 2432 bitmap sm_in_loop = BITMAP_ALLOC (&lim_bitmap_obstack); |
0 | 2433 |
2434 if (loop_suitable_for_sm (loop, exits)) | |
2435 { | |
2436 find_refs_for_sm (loop, sm_executed, sm_in_loop); | |
2437 hoist_memory_references (loop, sm_in_loop, exits); | |
2438 } | |
111 | 2439 exits.release (); |
0 | 2440 |
2441 bitmap_ior_into (sm_executed, sm_in_loop); | |
2442 for (subloop = loop->inner; subloop != NULL; subloop = subloop->next) | |
2443 store_motion_loop (subloop, sm_executed); | |
2444 bitmap_and_compl_into (sm_executed, sm_in_loop); | |
2445 BITMAP_FREE (sm_in_loop); | |
2446 } | |
2447 | |
2448 /* Try to perform store motion for all memory references modified inside | |
2449 loops. */ | |
2450 | |
2451 static void | |
2452 store_motion (void) | |
2453 { | |
145 | 2454 class loop *loop; |
111 | 2455 bitmap sm_executed = BITMAP_ALLOC (&lim_bitmap_obstack); |
0 | 2456 |
2457 for (loop = current_loops->tree_root->inner; loop != NULL; loop = loop->next) | |
2458 store_motion_loop (loop, sm_executed); | |
2459 | |
2460 BITMAP_FREE (sm_executed); | |
2461 gsi_commit_edge_inserts (); | |
2462 } | |
2463 | |
2464 /* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e. | |
2465 for each such basic block bb records the outermost loop for that execution | |
2466 of its header implies execution of bb. CONTAINS_CALL is the bitmap of | |
2467 blocks that contain a nonpure call. */ | |
2468 | |
2469 static void | |
145 | 2470 fill_always_executed_in_1 (class loop *loop, sbitmap contains_call) |
0 | 2471 { |
2472 basic_block bb = NULL, *bbs, last = NULL; | |
2473 unsigned i; | |
2474 edge e; | |
145 | 2475 class loop *inn_loop = loop; |
0 | 2476 |
111 | 2477 if (ALWAYS_EXECUTED_IN (loop->header) == NULL) |
0 | 2478 { |
2479 bbs = get_loop_body_in_dom_order (loop); | |
2480 | |
2481 for (i = 0; i < loop->num_nodes; i++) | |
2482 { | |
2483 edge_iterator ei; | |
2484 bb = bbs[i]; | |
2485 | |
2486 if (dominated_by_p (CDI_DOMINATORS, loop->latch, bb)) | |
2487 last = bb; | |
2488 | |
111 | 2489 if (bitmap_bit_p (contains_call, bb->index)) |
0 | 2490 break; |
2491 | |
2492 FOR_EACH_EDGE (e, ei, bb->succs) | |
111 | 2493 { |
2494 /* If there is an exit from this BB. */ | |
2495 if (!flow_bb_inside_loop_p (loop, e->dest)) | |
2496 break; | |
2497 /* Or we enter a possibly non-finite loop. */ | |
2498 if (flow_loop_nested_p (bb->loop_father, | |
2499 e->dest->loop_father) | |
2500 && ! finite_loop_p (e->dest->loop_father)) | |
2501 break; | |
2502 } | |
0 | 2503 if (e) |
2504 break; | |
2505 | |
2506 /* A loop might be infinite (TODO use simple loop analysis | |
2507 to disprove this if possible). */ | |
2508 if (bb->flags & BB_IRREDUCIBLE_LOOP) | |
2509 break; | |
2510 | |
2511 if (!flow_bb_inside_loop_p (inn_loop, bb)) | |
2512 break; | |
2513 | |
2514 if (bb->loop_father->header == bb) | |
2515 { | |
2516 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb)) | |
2517 break; | |
2518 | |
2519 /* In a loop that is always entered we may proceed anyway. | |
2520 But record that we entered it and stop once we leave it. */ | |
2521 inn_loop = bb->loop_father; | |
2522 } | |
2523 } | |
2524 | |
2525 while (1) | |
2526 { | |
111 | 2527 SET_ALWAYS_EXECUTED_IN (last, loop); |
0 | 2528 if (last == loop->header) |
2529 break; | |
2530 last = get_immediate_dominator (CDI_DOMINATORS, last); | |
2531 } | |
2532 | |
2533 free (bbs); | |
2534 } | |
2535 | |
2536 for (loop = loop->inner; loop; loop = loop->next) | |
111 | 2537 fill_always_executed_in_1 (loop, contains_call); |
0 | 2538 } |
2539 | |
111 | 2540 /* Fills ALWAYS_EXECUTED_IN information for basic blocks, i.e. |
2541 for each such basic block bb records the outermost loop for that execution | |
2542 of its header implies execution of bb. */ | |
2543 | |
2544 static void | |
2545 fill_always_executed_in (void) | |
2546 { | |
2547 basic_block bb; | |
145 | 2548 class loop *loop; |
111 | 2549 |
2550 auto_sbitmap contains_call (last_basic_block_for_fn (cfun)); | |
2551 bitmap_clear (contains_call); | |
2552 FOR_EACH_BB_FN (bb, cfun) | |
2553 { | |
2554 gimple_stmt_iterator gsi; | |
2555 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2556 { | |
2557 if (nonpure_call_p (gsi_stmt (gsi))) | |
2558 break; | |
2559 } | |
2560 | |
2561 if (!gsi_end_p (gsi)) | |
2562 bitmap_set_bit (contains_call, bb->index); | |
2563 } | |
2564 | |
2565 for (loop = current_loops->tree_root->inner; loop; loop = loop->next) | |
2566 fill_always_executed_in_1 (loop, contains_call); | |
2567 } | |
2568 | |
2569 | |
0 | 2570 /* Compute the global information needed by the loop invariant motion pass. */ |
2571 | |
2572 static void | |
2573 tree_ssa_lim_initialize (void) | |
2574 { | |
145 | 2575 class loop *loop; |
111 | 2576 unsigned i; |
2577 | |
2578 bitmap_obstack_initialize (&lim_bitmap_obstack); | |
2579 gcc_obstack_init (&mem_ref_obstack); | |
2580 lim_aux_data_map = new hash_map<gimple *, lim_aux_data *>; | |
2581 | |
2582 if (flag_tm) | |
2583 compute_transaction_bits (); | |
2584 | |
2585 alloc_aux_for_edges (0); | |
2586 | |
2587 memory_accesses.refs = new hash_table<mem_ref_hasher> (100); | |
2588 memory_accesses.refs_list.create (100); | |
2589 /* Allocate a special, unanalyzable mem-ref with ID zero. */ | |
2590 memory_accesses.refs_list.quick_push | |
145 | 2591 (mem_ref_alloc (NULL, 0, UNANALYZABLE_MEM_ID)); |
111 | 2592 |
2593 memory_accesses.refs_in_loop.create (number_of_loops (cfun)); | |
2594 memory_accesses.refs_in_loop.quick_grow (number_of_loops (cfun)); | |
2595 memory_accesses.refs_stored_in_loop.create (number_of_loops (cfun)); | |
2596 memory_accesses.refs_stored_in_loop.quick_grow (number_of_loops (cfun)); | |
2597 memory_accesses.all_refs_stored_in_loop.create (number_of_loops (cfun)); | |
2598 memory_accesses.all_refs_stored_in_loop.quick_grow (number_of_loops (cfun)); | |
2599 | |
2600 for (i = 0; i < number_of_loops (cfun); i++) | |
0 | 2601 { |
111 | 2602 bitmap_initialize (&memory_accesses.refs_in_loop[i], |
2603 &lim_bitmap_obstack); | |
2604 bitmap_initialize (&memory_accesses.refs_stored_in_loop[i], | |
2605 &lim_bitmap_obstack); | |
2606 bitmap_initialize (&memory_accesses.all_refs_stored_in_loop[i], | |
2607 &lim_bitmap_obstack); | |
0 | 2608 } |
2609 | |
111 | 2610 memory_accesses.ttae_cache = NULL; |
2611 | |
2612 /* Initialize bb_loop_postorder with a mapping from loop->num to | |
2613 its postorder index. */ | |
2614 i = 0; | |
2615 bb_loop_postorder = XNEWVEC (unsigned, number_of_loops (cfun)); | |
2616 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) | |
2617 bb_loop_postorder[loop->num] = i++; | |
0 | 2618 } |
2619 | |
2620 /* Cleans up after the invariant motion pass. */ | |
2621 | |
2622 static void | |
2623 tree_ssa_lim_finalize (void) | |
2624 { | |
2625 basic_block bb; | |
2626 unsigned i; | |
111 | 2627 im_mem_ref *ref; |
2628 | |
2629 free_aux_for_edges (); | |
2630 | |
2631 FOR_EACH_BB_FN (bb, cfun) | |
2632 SET_ALWAYS_EXECUTED_IN (bb, NULL); | |
2633 | |
2634 bitmap_obstack_release (&lim_bitmap_obstack); | |
2635 delete lim_aux_data_map; | |
2636 | |
2637 delete memory_accesses.refs; | |
2638 memory_accesses.refs = NULL; | |
2639 | |
2640 FOR_EACH_VEC_ELT (memory_accesses.refs_list, i, ref) | |
2641 memref_free (ref); | |
2642 memory_accesses.refs_list.release (); | |
2643 obstack_free (&mem_ref_obstack, NULL); | |
2644 | |
2645 memory_accesses.refs_in_loop.release (); | |
2646 memory_accesses.refs_stored_in_loop.release (); | |
2647 memory_accesses.all_refs_stored_in_loop.release (); | |
0 | 2648 |
2649 if (memory_accesses.ttae_cache) | |
111 | 2650 free_affine_expand_cache (&memory_accesses.ttae_cache); |
2651 | |
2652 free (bb_loop_postorder); | |
0 | 2653 } |
2654 | |
2655 /* Moves invariants from loops. Only "expensive" invariants are moved out -- | |
2656 i.e. those that are likely to be win regardless of the register pressure. */ | |
2657 | |
111 | 2658 static unsigned int |
0 | 2659 tree_ssa_lim (void) |
2660 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2661 unsigned int todo; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2662 |
0 | 2663 tree_ssa_lim_initialize (); |
2664 | |
2665 /* Gathers information about memory accesses in the loops. */ | |
2666 analyze_memory_references (); | |
2667 | |
111 | 2668 /* Fills ALWAYS_EXECUTED_IN information for basic blocks. */ |
2669 fill_always_executed_in (); | |
2670 | |
0 | 2671 /* For each statement determine the outermost loop in that it is |
2672 invariant and cost for computing the invariant. */ | |
111 | 2673 invariantness_dom_walker (CDI_DOMINATORS) |
2674 .walk (cfun->cfg->x_entry_block_ptr); | |
0 | 2675 |
2676 /* Execute store motion. Force the necessary invariants to be moved | |
2677 out of the loops as well. */ | |
2678 store_motion (); | |
2679 | |
2680 /* Move the expressions that are expensive enough. */ | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2681 todo = move_computations (); |
0 | 2682 |
2683 tree_ssa_lim_finalize (); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2684 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2685 return todo; |
0 | 2686 } |
111 | 2687 |
2688 /* Loop invariant motion pass. */ | |
2689 | |
2690 namespace { | |
2691 | |
2692 const pass_data pass_data_lim = | |
2693 { | |
2694 GIMPLE_PASS, /* type */ | |
2695 "lim", /* name */ | |
2696 OPTGROUP_LOOP, /* optinfo_flags */ | |
2697 TV_LIM, /* tv_id */ | |
2698 PROP_cfg, /* properties_required */ | |
2699 0, /* properties_provided */ | |
2700 0, /* properties_destroyed */ | |
2701 0, /* todo_flags_start */ | |
2702 0, /* todo_flags_finish */ | |
2703 }; | |
2704 | |
2705 class pass_lim : public gimple_opt_pass | |
2706 { | |
2707 public: | |
2708 pass_lim (gcc::context *ctxt) | |
2709 : gimple_opt_pass (pass_data_lim, ctxt) | |
2710 {} | |
2711 | |
2712 /* opt_pass methods: */ | |
2713 opt_pass * clone () { return new pass_lim (m_ctxt); } | |
2714 virtual bool gate (function *) { return flag_tree_loop_im != 0; } | |
2715 virtual unsigned int execute (function *); | |
2716 | |
2717 }; // class pass_lim | |
2718 | |
2719 unsigned int | |
2720 pass_lim::execute (function *fun) | |
2721 { | |
2722 bool in_loop_pipeline = scev_initialized_p (); | |
2723 if (!in_loop_pipeline) | |
2724 loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS); | |
2725 | |
2726 if (number_of_loops (fun) <= 1) | |
2727 return 0; | |
2728 unsigned int todo = tree_ssa_lim (); | |
2729 | |
2730 if (!in_loop_pipeline) | |
2731 loop_optimizer_finalize (); | |
131 | 2732 else |
2733 scev_reset (); | |
111 | 2734 return todo; |
2735 } | |
2736 | |
2737 } // anon namespace | |
2738 | |
2739 gimple_opt_pass * | |
2740 make_pass_lim (gcc::context *ctxt) | |
2741 { | |
2742 return new pass_lim (ctxt); | |
2743 } | |
2744 | |
2745 |