Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-ssa-loop-im.c @ 115:4cb7a319550d
fix c-parser.c
author | mir3636 |
---|---|
date | Tue, 28 Nov 2017 19:31:15 +0900 |
parents | 04ced10e8804 |
children | 84e7813d76e9 |
rev | line source |
---|---|
0 | 1 /* Loop invariant motion. |
111 | 2 Copyright (C) 2003-2017 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 "params.h" | |
42 #include "tree-affine.h" | |
43 #include "tree-ssa-propagate.h" | |
111 | 44 #include "trans-mem.h" |
45 #include "gimple-fold.h" | |
46 #include "tree-scalar-evolution.h" | |
47 #include "tree-ssa-loop-niter.h" | |
0 | 48 |
49 /* TODO: Support for predicated code motion. I.e. | |
50 | |
51 while (1) | |
52 { | |
53 if (cond) | |
54 { | |
55 a = inv; | |
56 something; | |
57 } | |
58 } | |
59 | |
111 | 60 Where COND and INV are invariants, but evaluating INV may trap or be |
0 | 61 invalid from some other reason if !COND. This may be transformed to |
62 | |
63 if (cond) | |
64 a = inv; | |
65 while (1) | |
66 { | |
67 if (cond) | |
68 something; | |
69 } */ | |
70 | |
71 /* The auxiliary data kept for each statement. */ | |
72 | |
73 struct lim_aux_data | |
74 { | |
75 struct loop *max_loop; /* The outermost loop in that the statement | |
76 is invariant. */ | |
77 | |
78 struct loop *tgt_loop; /* The loop out of that we want to move the | |
79 invariant. */ | |
80 | |
81 struct loop *always_executed_in; | |
82 /* The outermost loop for that we are sure | |
83 the statement is executed if the loop | |
84 is entered. */ | |
85 | |
86 unsigned cost; /* Cost of the computation performed by the | |
87 statement. */ | |
88 | |
111 | 89 unsigned ref; /* The simple_mem_ref in this stmt or 0. */ |
90 | |
91 vec<gimple *> depends; /* Vector of statements that must be also | |
92 hoisted out of the loop when this statement | |
93 is hoisted; i.e. those that define the | |
94 operands of the statement and are inside of | |
95 the MAX_LOOP loop. */ | |
0 | 96 }; |
97 | |
98 /* Maps statements to their lim_aux_data. */ | |
99 | |
111 | 100 static hash_map<gimple *, lim_aux_data *> *lim_aux_data_map; |
0 | 101 |
102 /* Description of a memory reference location. */ | |
103 | |
111 | 104 struct mem_ref_loc |
0 | 105 { |
106 tree *ref; /* The reference itself. */ | |
111 | 107 gimple *stmt; /* The statement in that it occurs. */ |
108 }; | |
109 | |
0 | 110 |
111 /* Description of a memory reference. */ | |
112 | |
111 | 113 struct im_mem_ref |
0 | 114 { |
115 unsigned id; /* ID assigned to the memory reference | |
116 (its index in memory_accesses.refs_list) */ | |
117 hashval_t hash; /* Its hash value. */ | |
111 | 118 |
119 /* The memory access itself and associated caching of alias-oracle | |
120 query meta-data. */ | |
121 ao_ref mem; | |
122 | |
0 | 123 bitmap stored; /* The set of loops in that this memory location |
124 is stored to. */ | |
111 | 125 vec<mem_ref_loc> accesses_in_loop; |
0 | 126 /* The locations of the accesses. Vector |
127 indexed by the loop number. */ | |
128 | |
129 /* The following sets are computed on demand. We keep both set and | |
130 its complement, so that we know whether the information was | |
131 already computed or not. */ | |
111 | 132 bitmap_head indep_loop; /* The set of loops in that the memory |
0 | 133 reference is independent, meaning: |
134 If it is stored in the loop, this store | |
135 is independent on all other loads and | |
136 stores. | |
137 If it is only loaded, then it is independent | |
138 on all stores in the loop. */ | |
111 | 139 bitmap_head dep_loop; /* The complement of INDEP_LOOP. */ |
140 }; | |
141 | |
142 /* We use two bits per loop in the ref->{in,}dep_loop bitmaps, the first | |
143 to record (in)dependence against stores in the loop and its subloops, the | |
144 second to record (in)dependence against all references in the loop | |
145 and its subloops. */ | |
146 #define LOOP_DEP_BIT(loopnum, storedp) (2 * (loopnum) + (storedp ? 1 : 0)) | |
147 | |
148 /* Mem_ref hashtable helpers. */ | |
149 | |
150 struct mem_ref_hasher : nofree_ptr_hash <im_mem_ref> | |
151 { | |
152 typedef tree_node *compare_type; | |
153 static inline hashval_t hash (const im_mem_ref *); | |
154 static inline bool equal (const im_mem_ref *, const tree_node *); | |
155 }; | |
156 | |
157 /* A hash function for struct im_mem_ref object OBJ. */ | |
158 | |
159 inline hashval_t | |
160 mem_ref_hasher::hash (const im_mem_ref *mem) | |
161 { | |
162 return mem->hash; | |
163 } | |
164 | |
165 /* An equality function for struct im_mem_ref object MEM1 with | |
166 memory reference OBJ2. */ | |
167 | |
168 inline bool | |
169 mem_ref_hasher::equal (const im_mem_ref *mem1, const tree_node *obj2) | |
170 { | |
171 return operand_equal_p (mem1->mem.ref, (const_tree) obj2, 0); | |
172 } | |
173 | |
0 | 174 |
175 /* Description of memory accesses in loops. */ | |
176 | |
177 static struct | |
178 { | |
179 /* The hash table of memory references accessed in loops. */ | |
111 | 180 hash_table<mem_ref_hasher> *refs; |
0 | 181 |
182 /* The list of memory references. */ | |
111 | 183 vec<im_mem_ref *> refs_list; |
0 | 184 |
185 /* The set of memory references accessed in each loop. */ | |
111 | 186 vec<bitmap_head> refs_in_loop; |
187 | |
188 /* The set of memory references stored in each loop. */ | |
189 vec<bitmap_head> refs_stored_in_loop; | |
190 | |
191 /* The set of memory references stored in each loop, including subloops . */ | |
192 vec<bitmap_head> all_refs_stored_in_loop; | |
0 | 193 |
194 /* Cache for expanding memory addresses. */ | |
111 | 195 hash_map<tree, name_expansion *> *ttae_cache; |
0 | 196 } memory_accesses; |
197 | |
111 | 198 /* Obstack for the bitmaps in the above data structures. */ |
199 static bitmap_obstack lim_bitmap_obstack; | |
200 static obstack mem_ref_obstack; | |
201 | |
202 static bool ref_indep_loop_p (struct loop *, im_mem_ref *, struct loop *); | |
203 static bool ref_always_accessed_p (struct loop *, im_mem_ref *, bool); | |
0 | 204 |
205 /* Minimum cost of an expensive expression. */ | |
206 #define LIM_EXPENSIVE ((unsigned) PARAM_VALUE (PARAM_LIM_EXPENSIVE)) | |
207 | |
111 | 208 /* The outermost loop for which execution of the header guarantees that the |
0 | 209 block will be executed. */ |
210 #define ALWAYS_EXECUTED_IN(BB) ((struct loop *) (BB)->aux) | |
111 | 211 #define SET_ALWAYS_EXECUTED_IN(BB, VAL) ((BB)->aux = (void *) (VAL)) |
212 | |
213 /* ID of the shared unanalyzable mem. */ | |
214 #define UNANALYZABLE_MEM_ID 0 | |
215 | |
216 /* Whether the reference was analyzable. */ | |
217 #define MEM_ANALYZABLE(REF) ((REF)->id != UNANALYZABLE_MEM_ID) | |
0 | 218 |
219 static struct lim_aux_data * | |
111 | 220 init_lim_data (gimple *stmt) |
0 | 221 { |
111 | 222 lim_aux_data *p = XCNEW (struct lim_aux_data); |
223 lim_aux_data_map->put (stmt, p); | |
224 | |
225 return p; | |
0 | 226 } |
227 | |
228 static struct lim_aux_data * | |
111 | 229 get_lim_data (gimple *stmt) |
0 | 230 { |
111 | 231 lim_aux_data **p = lim_aux_data_map->get (stmt); |
0 | 232 if (!p) |
233 return NULL; | |
234 | |
111 | 235 return *p; |
0 | 236 } |
237 | |
238 /* Releases the memory occupied by DATA. */ | |
239 | |
240 static void | |
241 free_lim_aux_data (struct lim_aux_data *data) | |
242 { | |
111 | 243 data->depends.release (); |
0 | 244 free (data); |
245 } | |
246 | |
247 static void | |
111 | 248 clear_lim_data (gimple *stmt) |
0 | 249 { |
111 | 250 lim_aux_data **p = lim_aux_data_map->get (stmt); |
0 | 251 if (!p) |
252 return; | |
253 | |
111 | 254 free_lim_aux_data (*p); |
0 | 255 *p = NULL; |
256 } | |
257 | |
111 | 258 |
259 /* The possibilities of statement movement. */ | |
260 enum move_pos | |
261 { | |
262 MOVE_IMPOSSIBLE, /* No movement -- side effect expression. */ | |
263 MOVE_PRESERVE_EXECUTION, /* Must not cause the non-executed statement | |
264 become executed -- memory accesses, ... */ | |
265 MOVE_POSSIBLE /* Unlimited movement. */ | |
266 }; | |
267 | |
0 | 268 |
269 /* If it is possible to hoist the statement STMT unconditionally, | |
270 returns MOVE_POSSIBLE. | |
271 If it is possible to hoist the statement STMT, but we must avoid making | |
272 it executed if it would not be executed in the original program (e.g. | |
273 because it may trap), return MOVE_PRESERVE_EXECUTION. | |
274 Otherwise return MOVE_IMPOSSIBLE. */ | |
275 | |
276 enum move_pos | |
111 | 277 movement_possibility (gimple *stmt) |
0 | 278 { |
279 tree lhs; | |
280 enum move_pos ret = MOVE_POSSIBLE; | |
281 | |
282 if (flag_unswitch_loops | |
283 && gimple_code (stmt) == GIMPLE_COND) | |
284 { | |
285 /* If we perform unswitching, force the operands of the invariant | |
286 condition to be moved out of the loop. */ | |
287 return MOVE_POSSIBLE; | |
288 } | |
289 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
290 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
|
291 && gimple_phi_num_args (stmt) <= 2 |
111 | 292 && !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
|
293 && !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
|
294 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
|
295 |
0 | 296 if (gimple_get_lhs (stmt) == NULL_TREE) |
297 return MOVE_IMPOSSIBLE; | |
298 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
299 if (gimple_vdef (stmt)) |
0 | 300 return MOVE_IMPOSSIBLE; |
301 | |
302 if (stmt_ends_bb_p (stmt) | |
303 || gimple_has_volatile_ops (stmt) | |
304 || gimple_has_side_effects (stmt) | |
305 || stmt_could_throw_p (stmt)) | |
306 return MOVE_IMPOSSIBLE; | |
307 | |
308 if (is_gimple_call (stmt)) | |
309 { | |
310 /* While pure or const call is guaranteed to have no side effects, we | |
311 cannot move it arbitrarily. Consider code like | |
312 | |
313 char *s = something (); | |
314 | |
315 while (1) | |
316 { | |
317 if (s) | |
318 t = strlen (s); | |
319 else | |
320 t = 0; | |
321 } | |
322 | |
323 Here the strlen call cannot be moved out of the loop, even though | |
324 s is invariant. In addition to possibly creating a call with | |
325 invalid arguments, moving out a function call that is not executed | |
326 may cause performance regressions in case the call is costly and | |
327 not executed at all. */ | |
328 ret = MOVE_PRESERVE_EXECUTION; | |
329 lhs = gimple_call_lhs (stmt); | |
330 } | |
331 else if (is_gimple_assign (stmt)) | |
332 lhs = gimple_assign_lhs (stmt); | |
333 else | |
334 return MOVE_IMPOSSIBLE; | |
335 | |
336 if (TREE_CODE (lhs) == SSA_NAME | |
337 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)) | |
338 return MOVE_IMPOSSIBLE; | |
339 | |
340 if (TREE_CODE (lhs) != SSA_NAME | |
341 || gimple_could_trap_p (stmt)) | |
342 return MOVE_PRESERVE_EXECUTION; | |
343 | |
111 | 344 /* Non local loads in a transaction cannot be hoisted out. Well, |
345 unless the load happens on every path out of the loop, but we | |
346 don't take this into account yet. */ | |
347 if (flag_tm | |
348 && gimple_in_transaction (stmt) | |
349 && gimple_assign_single_p (stmt)) | |
350 { | |
351 tree rhs = gimple_assign_rhs1 (stmt); | |
352 if (DECL_P (rhs) && is_global_var (rhs)) | |
353 { | |
354 if (dump_file) | |
355 { | |
356 fprintf (dump_file, "Cannot hoist conditional load of "); | |
357 print_generic_expr (dump_file, rhs, TDF_SLIM); | |
358 fprintf (dump_file, " because it is in a transaction.\n"); | |
359 } | |
360 return MOVE_IMPOSSIBLE; | |
361 } | |
362 } | |
363 | |
0 | 364 return ret; |
365 } | |
366 | |
367 /* Suppose that operand DEF is used inside the LOOP. Returns the outermost | |
368 loop to that we could move the expression using DEF if it did not have | |
369 other operands, i.e. the outermost loop enclosing LOOP in that the value | |
370 of DEF is invariant. */ | |
371 | |
372 static struct loop * | |
373 outermost_invariant_loop (tree def, struct loop *loop) | |
374 { | |
111 | 375 gimple *def_stmt; |
0 | 376 basic_block def_bb; |
377 struct loop *max_loop; | |
378 struct lim_aux_data *lim_data; | |
379 | |
380 if (!def) | |
381 return superloop_at_depth (loop, 1); | |
382 | |
383 if (TREE_CODE (def) != SSA_NAME) | |
384 { | |
385 gcc_assert (is_gimple_min_invariant (def)); | |
386 return superloop_at_depth (loop, 1); | |
387 } | |
388 | |
389 def_stmt = SSA_NAME_DEF_STMT (def); | |
390 def_bb = gimple_bb (def_stmt); | |
391 if (!def_bb) | |
392 return superloop_at_depth (loop, 1); | |
393 | |
394 max_loop = find_common_loop (loop, def_bb->loop_father); | |
395 | |
396 lim_data = get_lim_data (def_stmt); | |
397 if (lim_data != NULL && lim_data->max_loop != NULL) | |
398 max_loop = find_common_loop (max_loop, | |
399 loop_outer (lim_data->max_loop)); | |
400 if (max_loop == loop) | |
401 return NULL; | |
402 max_loop = superloop_at_depth (loop, loop_depth (max_loop) + 1); | |
403 | |
404 return max_loop; | |
405 } | |
406 | |
407 /* DATA is a structure containing information associated with a statement | |
408 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
|
409 |
0 | 410 Find the outermost loop enclosing LOOP in that value of DEF is invariant |
411 and record this in DATA->max_loop field. If DEF itself is defined inside | |
412 this loop as well (i.e. we need to hoist it out of the loop if we want | |
413 to hoist the statement represented by DATA), record the statement in that | |
414 DEF is defined to the DATA->depends list. Additionally if ADD_COST is true, | |
415 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
|
416 |
0 | 417 If DEF is not invariant in LOOP, return false. Otherwise return TRUE. */ |
418 | |
419 static bool | |
420 add_dependency (tree def, struct lim_aux_data *data, struct loop *loop, | |
421 bool add_cost) | |
422 { | |
111 | 423 gimple *def_stmt = SSA_NAME_DEF_STMT (def); |
0 | 424 basic_block def_bb = gimple_bb (def_stmt); |
425 struct loop *max_loop; | |
426 struct lim_aux_data *def_data; | |
427 | |
428 if (!def_bb) | |
429 return true; | |
430 | |
431 max_loop = outermost_invariant_loop (def, loop); | |
432 if (!max_loop) | |
433 return false; | |
434 | |
435 if (flow_loop_nested_p (data->max_loop, max_loop)) | |
436 data->max_loop = max_loop; | |
437 | |
438 def_data = get_lim_data (def_stmt); | |
439 if (!def_data) | |
440 return true; | |
441 | |
442 if (add_cost | |
443 /* Only add the cost if the statement defining DEF is inside LOOP, | |
444 i.e. if it is likely that by moving the invariants dependent | |
445 on it, we will be able to avoid creating a new register for | |
446 it (since it will be only used in these dependent invariants). */ | |
447 && def_bb->loop_father == loop) | |
448 data->cost += def_data->cost; | |
449 | |
111 | 450 data->depends.safe_push (def_stmt); |
0 | 451 |
452 return true; | |
453 } | |
454 | |
111 | 455 /* Returns an estimate for a cost of statement STMT. The values here |
456 are just ad-hoc constants, similar to costs for inlining. */ | |
0 | 457 |
458 static unsigned | |
111 | 459 stmt_cost (gimple *stmt) |
0 | 460 { |
461 /* 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
|
462 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
|
463 || gimple_code (stmt) == GIMPLE_PHI) |
0 | 464 return LIM_EXPENSIVE; |
465 | |
111 | 466 /* We should be hoisting calls if possible. */ |
0 | 467 if (is_gimple_call (stmt)) |
468 { | |
111 | 469 tree fndecl; |
0 | 470 |
471 /* Unless the call is a builtin_constant_p; this always folds to a | |
472 constant, so moving it is useless. */ | |
473 fndecl = gimple_call_fndecl (stmt); | |
474 if (fndecl | |
475 && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL | |
476 && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CONSTANT_P) | |
477 return 0; | |
478 | |
111 | 479 return LIM_EXPENSIVE; |
0 | 480 } |
481 | |
111 | 482 /* Hoisting memory references out should almost surely be a win. */ |
483 if (gimple_references_memory_p (stmt)) | |
484 return LIM_EXPENSIVE; | |
485 | |
0 | 486 if (gimple_code (stmt) != GIMPLE_ASSIGN) |
111 | 487 return 1; |
0 | 488 |
489 switch (gimple_assign_rhs_code (stmt)) | |
490 { | |
491 case MULT_EXPR: | |
111 | 492 case WIDEN_MULT_EXPR: |
493 case WIDEN_MULT_PLUS_EXPR: | |
494 case WIDEN_MULT_MINUS_EXPR: | |
495 case DOT_PROD_EXPR: | |
496 case FMA_EXPR: | |
0 | 497 case TRUNC_DIV_EXPR: |
498 case CEIL_DIV_EXPR: | |
499 case FLOOR_DIV_EXPR: | |
500 case ROUND_DIV_EXPR: | |
501 case EXACT_DIV_EXPR: | |
502 case CEIL_MOD_EXPR: | |
503 case FLOOR_MOD_EXPR: | |
504 case ROUND_MOD_EXPR: | |
505 case TRUNC_MOD_EXPR: | |
506 case RDIV_EXPR: | |
507 /* Division and multiplication are usually expensive. */ | |
111 | 508 return LIM_EXPENSIVE; |
0 | 509 |
510 case LSHIFT_EXPR: | |
511 case RSHIFT_EXPR: | |
111 | 512 case WIDEN_LSHIFT_EXPR: |
513 case LROTATE_EXPR: | |
514 case RROTATE_EXPR: | |
515 /* Shifts and rotates are usually expensive. */ | |
516 return LIM_EXPENSIVE; | |
517 | |
518 case CONSTRUCTOR: | |
519 /* Make vector construction cost proportional to the number | |
520 of elements. */ | |
521 return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt)); | |
522 | |
523 case SSA_NAME: | |
524 case PAREN_EXPR: | |
525 /* Whether or not something is wrapped inside a PAREN_EXPR | |
526 should not change move cost. Nor should an intermediate | |
527 unpropagated SSA name copy. */ | |
528 return 0; | |
0 | 529 |
530 default: | |
111 | 531 return 1; |
0 | 532 } |
533 } | |
534 | |
535 /* Finds the outermost loop between OUTER and LOOP in that the memory reference | |
536 REF is independent. If REF is not independent in LOOP, NULL is returned | |
537 instead. */ | |
538 | |
539 static struct loop * | |
111 | 540 outermost_indep_loop (struct loop *outer, struct loop *loop, im_mem_ref *ref) |
0 | 541 { |
542 struct loop *aloop; | |
543 | |
111 | 544 if (ref->stored && bitmap_bit_p (ref->stored, loop->num)) |
0 | 545 return NULL; |
546 | |
547 for (aloop = outer; | |
548 aloop != loop; | |
549 aloop = superloop_at_depth (loop, loop_depth (aloop) + 1)) | |
111 | 550 if ((!ref->stored || !bitmap_bit_p (ref->stored, aloop->num)) |
551 && ref_indep_loop_p (aloop, ref, loop)) | |
0 | 552 return aloop; |
553 | |
111 | 554 if (ref_indep_loop_p (loop, ref, loop)) |
0 | 555 return loop; |
556 else | |
557 return NULL; | |
558 } | |
559 | |
560 /* If there is a simple load or store to a memory reference in STMT, returns | |
561 the location of the memory reference, and sets IS_STORE according to whether | |
562 it is a store or load. Otherwise, returns NULL. */ | |
563 | |
564 static tree * | |
111 | 565 simple_mem_ref_in_stmt (gimple *stmt, bool *is_store) |
0 | 566 { |
111 | 567 tree *lhs, *rhs; |
568 | |
569 /* Recognize SSA_NAME = MEM and MEM = (SSA_NAME | invariant) patterns. */ | |
570 if (!gimple_assign_single_p (stmt)) | |
0 | 571 return NULL; |
572 | |
573 lhs = gimple_assign_lhs_ptr (stmt); | |
111 | 574 rhs = gimple_assign_rhs1_ptr (stmt); |
575 | |
576 if (TREE_CODE (*lhs) == SSA_NAME && gimple_vuse (stmt)) | |
0 | 577 { |
578 *is_store = false; | |
111 | 579 return rhs; |
0 | 580 } |
111 | 581 else if (gimple_vdef (stmt) |
582 && (TREE_CODE (*rhs) == SSA_NAME || is_gimple_min_invariant (*rhs))) | |
0 | 583 { |
584 *is_store = true; | |
585 return lhs; | |
586 } | |
587 else | |
588 return NULL; | |
589 } | |
590 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
591 /* 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
|
592 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
|
593 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
|
594 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
|
595 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
|
596 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
597 static bool |
111 | 598 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
|
599 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
|
600 { |
111 | 601 edge te, fe; |
602 if (! extract_true_false_controlled_edges (dom, gimple_bb (phi), | |
603 &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
|
604 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
|
605 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
606 if (true_arg_p) |
111 | 607 *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
|
608 if (false_arg_p) |
111 | 609 *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
|
610 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
611 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
|
612 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
613 |
0 | 614 /* Determine the outermost loop to that it is possible to hoist a statement |
615 STMT and store it to LIM_DATA (STMT)->max_loop. To do this we determine | |
616 the outermost loop in that the value computed by STMT is invariant. | |
617 If MUST_PRESERVE_EXEC is true, additionally choose such a loop that | |
618 we preserve the fact whether STMT is executed. It also fills other related | |
619 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
|
620 |
0 | 621 The function returns false if STMT cannot be hoisted outside of the loop it |
622 is defined in, and true otherwise. */ | |
623 | |
624 static bool | |
111 | 625 determine_max_movement (gimple *stmt, bool must_preserve_exec) |
0 | 626 { |
627 basic_block bb = gimple_bb (stmt); | |
628 struct loop *loop = bb->loop_father; | |
629 struct loop *level; | |
630 struct lim_aux_data *lim_data = get_lim_data (stmt); | |
631 tree val; | |
632 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
|
633 |
0 | 634 if (must_preserve_exec) |
635 level = ALWAYS_EXECUTED_IN (bb); | |
636 else | |
637 level = superloop_at_depth (loop, 1); | |
638 lim_data->max_loop = level; | |
639 | |
111 | 640 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
|
641 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
642 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
|
643 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
|
644 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
|
645 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
|
646 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
647 /* 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
|
648 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
|
649 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
|
650 is that of the cheapest PHI argument dependency chain. */ |
111 | 651 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
|
652 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
653 val = USE_FROM_PTR (use_p); |
111 | 654 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
655 if (TREE_CODE (val) != SSA_NAME) |
111 | 656 { |
657 /* Assign const 1 to constants. */ | |
658 min_cost = MIN (min_cost, 1); | |
659 total_cost += 1; | |
660 continue; | |
661 } | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
662 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
|
663 return false; |
111 | 664 |
665 gimple *def_stmt = SSA_NAME_DEF_STMT (val); | |
666 if (gimple_bb (def_stmt) | |
667 && 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
|
668 { |
111 | 669 def_data = get_lim_data (def_stmt); |
670 if (def_data) | |
671 { | |
672 min_cost = MIN (min_cost, def_data->cost); | |
673 total_cost += def_data->cost; | |
674 } | |
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 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
677 |
111 | 678 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
|
679 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
|
680 |
111 | 681 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
|
682 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
683 basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb); |
111 | 684 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
|
685 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
|
686 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
|
687 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
|
688 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
|
689 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
|
690 /* 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
|
691 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
|
692 predicate in DOM. */ |
111 | 693 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
|
694 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
|
695 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
696 /* 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
|
697 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
|
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 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
|
700 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
|
701 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
|
702 if (def_data) |
111 | 703 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
|
704 } |
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 /* 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
|
707 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
|
708 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
|
709 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
|
710 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
|
711 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
|
712 && !(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
|
713 && 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
|
714 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
|
715 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
716 /* 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
|
717 ??? 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
|
718 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
|
719 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
|
720 } |
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 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
|
723 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
724 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
725 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
|
726 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
|
727 return false; |
0 | 728 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
729 if (gimple_vuse (stmt)) |
0 | 730 { |
111 | 731 im_mem_ref *ref |
732 = lim_data ? memory_accesses.refs_list[lim_data->ref] : NULL; | |
733 if (ref | |
734 && MEM_ANALYZABLE (ref)) | |
0 | 735 { |
111 | 736 lim_data->max_loop = outermost_indep_loop (lim_data->max_loop, |
737 loop, ref); | |
0 | 738 if (!lim_data->max_loop) |
739 return false; | |
740 } | |
111 | 741 else if (! add_dependency (gimple_vuse (stmt), lim_data, loop, false)) |
742 return false; | |
0 | 743 } |
744 | |
745 lim_data->cost += stmt_cost (stmt); | |
746 | |
747 return true; | |
748 } | |
749 | |
750 /* Suppose that some statement in ORIG_LOOP is hoisted to the loop LEVEL, | |
751 and that one of the operands of this statement is computed by STMT. | |
752 Ensure that STMT (together with all the statements that define its | |
753 operands) is hoisted at least out of the loop LEVEL. */ | |
754 | |
755 static void | |
111 | 756 set_level (gimple *stmt, struct loop *orig_loop, struct loop *level) |
0 | 757 { |
758 struct loop *stmt_loop = gimple_bb (stmt)->loop_father; | |
759 struct lim_aux_data *lim_data; | |
111 | 760 gimple *dep_stmt; |
761 unsigned i; | |
0 | 762 |
763 stmt_loop = find_common_loop (orig_loop, stmt_loop); | |
764 lim_data = get_lim_data (stmt); | |
765 if (lim_data != NULL && lim_data->tgt_loop != NULL) | |
766 stmt_loop = find_common_loop (stmt_loop, | |
767 loop_outer (lim_data->tgt_loop)); | |
768 if (flow_loop_nested_p (stmt_loop, level)) | |
769 return; | |
770 | |
771 gcc_assert (level == lim_data->max_loop | |
772 || flow_loop_nested_p (lim_data->max_loop, level)); | |
773 | |
774 lim_data->tgt_loop = level; | |
111 | 775 FOR_EACH_VEC_ELT (lim_data->depends, i, dep_stmt) |
776 set_level (dep_stmt, orig_loop, level); | |
0 | 777 } |
778 | |
779 /* Determines an outermost loop from that we want to hoist the statement STMT. | |
780 For now we chose the outermost possible loop. TODO -- use profiling | |
781 information to set it more sanely. */ | |
782 | |
783 static void | |
111 | 784 set_profitable_level (gimple *stmt) |
0 | 785 { |
786 set_level (stmt, gimple_bb (stmt)->loop_father, get_lim_data (stmt)->max_loop); | |
787 } | |
788 | |
789 /* Returns true if STMT is a call that has side effects. */ | |
790 | |
791 static bool | |
111 | 792 nonpure_call_p (gimple *stmt) |
0 | 793 { |
794 if (gimple_code (stmt) != GIMPLE_CALL) | |
795 return false; | |
796 | |
797 return gimple_has_side_effects (stmt); | |
798 } | |
799 | |
800 /* Rewrite a/b to a*(1/b). Return the invariant stmt to process. */ | |
801 | |
111 | 802 static gimple * |
0 | 803 rewrite_reciprocal (gimple_stmt_iterator *bsi) |
804 { | |
111 | 805 gassign *stmt, *stmt1, *stmt2; |
806 tree name, lhs, type; | |
0 | 807 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
|
808 gimple_stmt_iterator gsi; |
0 | 809 |
111 | 810 stmt = as_a <gassign *> (gsi_stmt (*bsi)); |
0 | 811 lhs = gimple_assign_lhs (stmt); |
812 type = TREE_TYPE (lhs); | |
813 | |
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
|
814 real_one = build_one_cst (type); |
0 | 815 |
111 | 816 name = make_temp_ssa_name (type, NULL, "reciptmp"); |
817 stmt1 = gimple_build_assign (name, RDIV_EXPR, real_one, | |
818 gimple_assign_rhs2 (stmt)); | |
819 stmt2 = gimple_build_assign (lhs, MULT_EXPR, name, | |
820 gimple_assign_rhs1 (stmt)); | |
0 | 821 |
822 /* Replace division stmt with reciprocal and multiply stmts. | |
823 The multiply stmt is not invariant, so update iterator | |
824 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
|
825 gsi = *bsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
826 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
|
827 gsi_replace (&gsi, stmt2, true); |
0 | 828 |
829 /* Continue processing with invariant reciprocal statement. */ | |
830 return stmt1; | |
831 } | |
832 | |
833 /* Check if the pattern at *BSI is a bittest of the form | |
834 (A >> B) & 1 != 0 and in this case rewrite it to A & (1 << B) != 0. */ | |
835 | |
111 | 836 static gimple * |
0 | 837 rewrite_bittest (gimple_stmt_iterator *bsi) |
838 { | |
111 | 839 gassign *stmt; |
840 gimple *stmt1; | |
841 gassign *stmt2; | |
842 gimple *use_stmt; | |
843 gcond *cond_stmt; | |
844 tree lhs, name, t, a, b; | |
0 | 845 use_operand_p use; |
846 | |
111 | 847 stmt = as_a <gassign *> (gsi_stmt (*bsi)); |
0 | 848 lhs = gimple_assign_lhs (stmt); |
849 | |
850 /* Verify that the single use of lhs is a comparison against zero. */ | |
851 if (TREE_CODE (lhs) != SSA_NAME | |
111 | 852 || !single_imm_use (lhs, &use, &use_stmt)) |
853 return stmt; | |
854 cond_stmt = dyn_cast <gcond *> (use_stmt); | |
855 if (!cond_stmt) | |
0 | 856 return stmt; |
111 | 857 if (gimple_cond_lhs (cond_stmt) != lhs |
858 || (gimple_cond_code (cond_stmt) != NE_EXPR | |
859 && gimple_cond_code (cond_stmt) != EQ_EXPR) | |
860 || !integer_zerop (gimple_cond_rhs (cond_stmt))) | |
0 | 861 return stmt; |
862 | |
863 /* Get at the operands of the shift. The rhs is TMP1 & 1. */ | |
864 stmt1 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); | |
865 if (gimple_code (stmt1) != GIMPLE_ASSIGN) | |
866 return stmt; | |
867 | |
868 /* There is a conversion in between possibly inserted by fold. */ | |
869 if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt1))) | |
870 { | |
871 t = gimple_assign_rhs1 (stmt1); | |
872 if (TREE_CODE (t) != SSA_NAME | |
873 || !has_single_use (t)) | |
874 return stmt; | |
875 stmt1 = SSA_NAME_DEF_STMT (t); | |
876 if (gimple_code (stmt1) != GIMPLE_ASSIGN) | |
877 return stmt; | |
878 } | |
879 | |
880 /* Verify that B is loop invariant but A is not. Verify that with | |
881 all the stmt walking we are still in the same loop. */ | |
882 if (gimple_assign_rhs_code (stmt1) != RSHIFT_EXPR | |
883 || loop_containing_stmt (stmt1) != loop_containing_stmt (stmt)) | |
884 return stmt; | |
885 | |
886 a = gimple_assign_rhs1 (stmt1); | |
887 b = gimple_assign_rhs2 (stmt1); | |
888 | |
889 if (outermost_invariant_loop (b, loop_containing_stmt (stmt1)) != NULL | |
890 && outermost_invariant_loop (a, loop_containing_stmt (stmt1)) == NULL) | |
891 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
892 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
|
893 |
0 | 894 /* 1 << B */ |
895 t = fold_build2 (LSHIFT_EXPR, TREE_TYPE (a), | |
896 build_int_cst (TREE_TYPE (a), 1), b); | |
111 | 897 name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp"); |
898 stmt1 = gimple_build_assign (name, t); | |
0 | 899 |
900 /* A & (1 << B) */ | |
901 t = fold_build2 (BIT_AND_EXPR, TREE_TYPE (a), a, name); | |
111 | 902 name = make_temp_ssa_name (TREE_TYPE (a), NULL, "shifttmp"); |
903 stmt2 = gimple_build_assign (name, t); | |
0 | 904 |
905 /* Replace the SSA_NAME we compare against zero. Adjust | |
906 the type of zero accordingly. */ | |
907 SET_USE (use, name); | |
111 | 908 gimple_cond_set_rhs (cond_stmt, |
909 build_int_cst_type (TREE_TYPE (name), | |
910 0)); | |
0 | 911 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
912 /* 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
|
913 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
|
914 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
|
915 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
|
916 rsi = *bsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
917 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
|
918 gsi_insert_before (&rsi, stmt2, GSI_SAME_STMT); |
111 | 919 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
|
920 gsi_remove (&rsi, true); |
111 | 921 release_defs (to_release); |
0 | 922 |
923 return stmt1; | |
924 } | |
925 | |
926 return stmt; | |
927 } | |
928 | |
111 | 929 /* For each statement determines the outermost loop in that it is invariant, |
930 - statements on whose motion it depends and the cost of the computation. | |
931 - This information is stored to the LIM_DATA structure associated with | |
932 - each statement. */ | |
933 class invariantness_dom_walker : public dom_walker | |
934 { | |
935 public: | |
936 invariantness_dom_walker (cdi_direction direction) | |
937 : dom_walker (direction) {} | |
938 | |
939 virtual edge before_dom_children (basic_block); | |
940 }; | |
0 | 941 |
942 /* Determine the outermost loops in that statements in basic block BB are | |
943 invariant, and record them to the LIM_DATA associated with the statements. | |
111 | 944 Callback for dom_walker. */ |
945 | |
946 edge | |
947 invariantness_dom_walker::before_dom_children (basic_block bb) | |
0 | 948 { |
949 enum move_pos pos; | |
950 gimple_stmt_iterator bsi; | |
111 | 951 gimple *stmt; |
0 | 952 bool maybe_never = ALWAYS_EXECUTED_IN (bb) == NULL; |
953 struct loop *outermost = ALWAYS_EXECUTED_IN (bb); | |
954 struct lim_aux_data *lim_data; | |
955 | |
956 if (!loop_outer (bb->loop_father)) | |
111 | 957 return NULL; |
0 | 958 |
959 if (dump_file && (dump_flags & TDF_DETAILS)) | |
960 fprintf (dump_file, "Basic block %d (loop %d -- depth %d):\n\n", | |
961 bb->index, bb->loop_father->num, loop_depth (bb->loop_father)); | |
962 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
963 /* 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
|
964 ??? 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
|
965 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
|
966 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
|
967 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
|
968 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
|
969 && ((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
|
970 || (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
|
971 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
|
972 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
973 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
|
974 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
975 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
|
976 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
|
977 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
978 |
111 | 979 lim_data = get_lim_data (stmt); |
980 if (! lim_data) | |
981 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
|
982 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
|
983 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
984 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
|
985 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
986 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
|
987 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
988 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
989 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
990 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
|
991 { |
111 | 992 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
|
993 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
|
994 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
|
995 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
|
996 } |
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 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
|
999 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
|
1000 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1001 |
0 | 1002 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
1003 { | |
1004 stmt = gsi_stmt (bsi); | |
1005 | |
1006 pos = movement_possibility (stmt); | |
1007 if (pos == MOVE_IMPOSSIBLE) | |
1008 { | |
1009 if (nonpure_call_p (stmt)) | |
1010 { | |
1011 maybe_never = true; | |
1012 outermost = NULL; | |
1013 } | |
1014 /* Make sure to note always_executed_in for stores to make | |
1015 store-motion work. */ | |
1016 else if (stmt_makes_single_store (stmt)) | |
1017 { | |
111 | 1018 struct lim_aux_data *lim_data = get_lim_data (stmt); |
1019 if (! lim_data) | |
1020 lim_data = init_lim_data (stmt); | |
0 | 1021 lim_data->always_executed_in = outermost; |
1022 } | |
1023 continue; | |
1024 } | |
1025 | |
1026 if (is_gimple_assign (stmt) | |
1027 && (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)) | |
1028 == GIMPLE_BINARY_RHS)) | |
1029 { | |
1030 tree op0 = gimple_assign_rhs1 (stmt); | |
1031 tree op1 = gimple_assign_rhs2 (stmt); | |
1032 struct loop *ol1 = outermost_invariant_loop (op1, | |
1033 loop_containing_stmt (stmt)); | |
1034 | |
1035 /* If divisor is invariant, convert a/b to a*(1/b), allowing reciprocal | |
1036 to be hoisted out of loop, saving expensive divide. */ | |
1037 if (pos == MOVE_POSSIBLE | |
1038 && gimple_assign_rhs_code (stmt) == RDIV_EXPR | |
1039 && flag_unsafe_math_optimizations | |
1040 && !flag_trapping_math | |
1041 && ol1 != NULL | |
1042 && outermost_invariant_loop (op0, ol1) == NULL) | |
1043 stmt = rewrite_reciprocal (&bsi); | |
1044 | |
1045 /* If the shift count is invariant, convert (A >> B) & 1 to | |
1046 A & (1 << B) allowing the bit mask to be hoisted out of the loop | |
1047 saving an expensive shift. */ | |
1048 if (pos == MOVE_POSSIBLE | |
1049 && gimple_assign_rhs_code (stmt) == BIT_AND_EXPR | |
1050 && integer_onep (op1) | |
1051 && TREE_CODE (op0) == SSA_NAME | |
1052 && has_single_use (op0)) | |
1053 stmt = rewrite_bittest (&bsi); | |
1054 } | |
1055 | |
111 | 1056 lim_data = get_lim_data (stmt); |
1057 if (! lim_data) | |
1058 lim_data = init_lim_data (stmt); | |
0 | 1059 lim_data->always_executed_in = outermost; |
1060 | |
1061 if (maybe_never && pos == MOVE_PRESERVE_EXECUTION) | |
1062 continue; | |
1063 | |
1064 if (!determine_max_movement (stmt, pos == MOVE_PRESERVE_EXECUTION)) | |
1065 { | |
1066 lim_data->max_loop = NULL; | |
1067 continue; | |
1068 } | |
1069 | |
1070 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1071 { | |
111 | 1072 print_gimple_stmt (dump_file, stmt, 2); |
0 | 1073 fprintf (dump_file, " invariant up to level %d, cost %d.\n\n", |
1074 loop_depth (lim_data->max_loop), | |
1075 lim_data->cost); | |
1076 } | |
1077 | |
1078 if (lim_data->cost >= LIM_EXPENSIVE) | |
1079 set_profitable_level (stmt); | |
1080 } | |
111 | 1081 return NULL; |
0 | 1082 } |
1083 | |
111 | 1084 class move_computations_dom_walker : public dom_walker |
0 | 1085 { |
111 | 1086 public: |
1087 move_computations_dom_walker (cdi_direction direction) | |
1088 : dom_walker (direction), todo_ (0) {} | |
1089 | |
1090 virtual edge before_dom_children (basic_block); | |
1091 | |
1092 unsigned int todo_; | |
1093 }; | |
0 | 1094 |
1095 /* Hoist the statements in basic block BB out of the loops prescribed by | |
1096 data stored in LIM_DATA structures associated with each statement. Callback | |
1097 for walk_dominator_tree. */ | |
1098 | |
111 | 1099 unsigned int |
1100 move_computations_worker (basic_block bb) | |
0 | 1101 { |
1102 struct loop *level; | |
1103 unsigned cost = 0; | |
1104 struct lim_aux_data *lim_data; | |
111 | 1105 unsigned int todo = 0; |
0 | 1106 |
1107 if (!loop_outer (bb->loop_father)) | |
111 | 1108 return todo; |
1109 | |
1110 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
|
1111 { |
111 | 1112 gassign *new_stmt; |
1113 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
|
1114 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1115 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
|
1116 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
|
1117 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1118 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
|
1119 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1120 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1121 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1122 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
|
1123 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
|
1124 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
|
1125 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1126 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
|
1127 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1128 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
|
1129 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1130 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1131 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1132 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
|
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 fprintf (dump_file, "Moving PHI node\n"); |
111 | 1135 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
|
1136 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
|
1137 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
|
1138 } |
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 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
|
1141 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1142 tree arg = PHI_ARG_DEF (stmt, 0); |
111 | 1143 new_stmt = gimple_build_assign (gimple_phi_result (stmt), |
1144 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
|
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 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1147 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1148 basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb); |
111 | 1149 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
|
1150 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
|
1151 /* 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
|
1152 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
|
1153 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
|
1154 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
|
1155 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
|
1156 gimple_cond_lhs (cond), gimple_cond_rhs (cond)); |
111 | 1157 new_stmt = gimple_build_assign (gimple_phi_result (stmt), |
1158 COND_EXPR, t, arg0, arg1); | |
1159 todo |= TODO_cleanup_cfg; | |
1160 } | |
1161 if (INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (new_stmt))) | |
1162 && (!ALWAYS_EXECUTED_IN (bb) | |
1163 || (ALWAYS_EXECUTED_IN (bb) != level | |
1164 && !flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level)))) | |
1165 { | |
1166 tree lhs = gimple_assign_lhs (new_stmt); | |
1167 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
|
1168 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1169 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
|
1170 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
|
1171 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1172 |
111 | 1173 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); ) |
0 | 1174 { |
111 | 1175 edge e; |
1176 | |
1177 gimple *stmt = gsi_stmt (bsi); | |
0 | 1178 |
1179 lim_data = get_lim_data (stmt); | |
1180 if (lim_data == NULL) | |
1181 { | |
1182 gsi_next (&bsi); | |
1183 continue; | |
1184 } | |
1185 | |
1186 cost = lim_data->cost; | |
1187 level = lim_data->tgt_loop; | |
1188 clear_lim_data (stmt); | |
1189 | |
1190 if (!level) | |
1191 { | |
1192 gsi_next (&bsi); | |
1193 continue; | |
1194 } | |
1195 | |
1196 /* We do not really want to move conditionals out of the loop; we just | |
1197 placed it here to force its operands to be moved if necessary. */ | |
1198 if (gimple_code (stmt) == GIMPLE_COND) | |
1199 continue; | |
1200 | |
1201 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1202 { | |
1203 fprintf (dump_file, "Moving statement\n"); | |
111 | 1204 print_gimple_stmt (dump_file, stmt, 0); |
0 | 1205 fprintf (dump_file, "(cost %u) out of loop %d.\n\n", |
1206 cost, level->num); | |
1207 } | |
1208 | |
111 | 1209 e = loop_preheader_edge (level); |
1210 gcc_assert (!gimple_vdef (stmt)); | |
1211 if (gimple_vuse (stmt)) | |
1212 { | |
1213 /* The new VUSE is the one from the virtual PHI in the loop | |
1214 header or the one already present. */ | |
1215 gphi_iterator gsi2; | |
1216 for (gsi2 = gsi_start_phis (e->dest); | |
1217 !gsi_end_p (gsi2); gsi_next (&gsi2)) | |
1218 { | |
1219 gphi *phi = gsi2.phi (); | |
1220 if (virtual_operand_p (gimple_phi_result (phi))) | |
1221 { | |
1222 gimple_set_vuse (stmt, PHI_ARG_DEF_FROM_EDGE (phi, e)); | |
1223 break; | |
1224 } | |
1225 } | |
1226 } | |
0 | 1227 gsi_remove (&bsi, false); |
111 | 1228 if (gimple_has_lhs (stmt) |
1229 && TREE_CODE (gimple_get_lhs (stmt)) == SSA_NAME | |
1230 && INTEGRAL_TYPE_P (TREE_TYPE (gimple_get_lhs (stmt))) | |
1231 && (!ALWAYS_EXECUTED_IN (bb) | |
1232 || !(ALWAYS_EXECUTED_IN (bb) == level | |
1233 || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level)))) | |
1234 { | |
1235 tree lhs = gimple_get_lhs (stmt); | |
1236 SSA_NAME_RANGE_INFO (lhs) = NULL; | |
1237 } | |
1238 /* In case this is a stmt that is not unconditionally executed | |
1239 when the target loop header is executed and the stmt may | |
1240 invoke undefined integer or pointer overflow rewrite it to | |
1241 unsigned arithmetic. */ | |
1242 if (is_gimple_assign (stmt) | |
1243 && INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (stmt))) | |
1244 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (gimple_assign_lhs (stmt))) | |
1245 && arith_code_with_undefined_signed_overflow | |
1246 (gimple_assign_rhs_code (stmt)) | |
1247 && (!ALWAYS_EXECUTED_IN (bb) | |
1248 || !(ALWAYS_EXECUTED_IN (bb) == level | |
1249 || flow_loop_nested_p (ALWAYS_EXECUTED_IN (bb), level)))) | |
1250 gsi_insert_seq_on_edge (e, rewrite_to_defined_overflow (stmt)); | |
1251 else | |
1252 gsi_insert_on_edge (e, stmt); | |
0 | 1253 } |
111 | 1254 |
1255 return todo; | |
0 | 1256 } |
1257 | |
1258 /* Hoist the statements out of the loops prescribed by data stored in | |
1259 LIM_DATA structures associated with each statement.*/ | |
1260 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1261 static unsigned int |
0 | 1262 move_computations (void) |
1263 { | |
111 | 1264 int *rpo = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
1265 int n = pre_and_rev_post_order_compute_fn (cfun, NULL, rpo, false); | |
1266 unsigned todo = 0; | |
1267 | |
1268 for (int i = 0; i < n; ++i) | |
1269 todo |= move_computations_worker (BASIC_BLOCK_FOR_FN (cfun, rpo[i])); | |
1270 | |
1271 free (rpo); | |
0 | 1272 |
1273 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
|
1274 if (need_ssa_update_p (cfun)) |
0 | 1275 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
|
1276 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1277 return todo; |
0 | 1278 } |
1279 | |
1280 /* Checks whether the statement defining variable *INDEX can be hoisted | |
1281 out of the loop passed in DATA. Callback for for_each_index. */ | |
1282 | |
1283 static bool | |
1284 may_move_till (tree ref, tree *index, void *data) | |
1285 { | |
1286 struct loop *loop = (struct loop *) data, *max_loop; | |
1287 | |
1288 /* If REF is an array reference, check also that the step and the lower | |
1289 bound is invariant in LOOP. */ | |
1290 if (TREE_CODE (ref) == ARRAY_REF) | |
1291 { | |
1292 tree step = TREE_OPERAND (ref, 3); | |
1293 tree lbound = TREE_OPERAND (ref, 2); | |
1294 | |
1295 max_loop = outermost_invariant_loop (step, loop); | |
1296 if (!max_loop) | |
1297 return false; | |
1298 | |
1299 max_loop = outermost_invariant_loop (lbound, loop); | |
1300 if (!max_loop) | |
1301 return false; | |
1302 } | |
1303 | |
1304 max_loop = outermost_invariant_loop (*index, loop); | |
1305 if (!max_loop) | |
1306 return false; | |
1307 | |
1308 return true; | |
1309 } | |
1310 | |
1311 /* If OP is SSA NAME, force the statement that defines it to be | |
1312 moved out of the LOOP. ORIG_LOOP is the loop in that EXPR is used. */ | |
1313 | |
1314 static void | |
1315 force_move_till_op (tree op, struct loop *orig_loop, struct loop *loop) | |
1316 { | |
111 | 1317 gimple *stmt; |
0 | 1318 |
1319 if (!op | |
1320 || is_gimple_min_invariant (op)) | |
1321 return; | |
1322 | |
1323 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
|
1324 |
0 | 1325 stmt = SSA_NAME_DEF_STMT (op); |
1326 if (gimple_nop_p (stmt)) | |
1327 return; | |
1328 | |
1329 set_level (stmt, orig_loop, loop); | |
1330 } | |
1331 | |
1332 /* Forces statement defining invariants in REF (and *INDEX) to be moved out of | |
1333 the LOOP. The reference REF is used in the loop ORIG_LOOP. Callback for | |
1334 for_each_index. */ | |
1335 | |
1336 struct fmt_data | |
1337 { | |
1338 struct loop *loop; | |
1339 struct loop *orig_loop; | |
1340 }; | |
1341 | |
1342 static bool | |
1343 force_move_till (tree ref, tree *index, void *data) | |
1344 { | |
1345 struct fmt_data *fmt_data = (struct fmt_data *) data; | |
1346 | |
1347 if (TREE_CODE (ref) == ARRAY_REF) | |
1348 { | |
1349 tree step = TREE_OPERAND (ref, 3); | |
1350 tree lbound = TREE_OPERAND (ref, 2); | |
1351 | |
1352 force_move_till_op (step, fmt_data->orig_loop, fmt_data->loop); | |
1353 force_move_till_op (lbound, fmt_data->orig_loop, fmt_data->loop); | |
1354 } | |
1355 | |
1356 force_move_till_op (*index, fmt_data->orig_loop, fmt_data->loop); | |
1357 | |
1358 return true; | |
1359 } | |
1360 | |
1361 /* A function to free the mem_ref object OBJ. */ | |
1362 | |
1363 static void | |
111 | 1364 memref_free (struct im_mem_ref *mem) |
0 | 1365 { |
111 | 1366 mem->accesses_in_loop.release (); |
0 | 1367 } |
1368 | |
1369 /* Allocates and returns a memory reference description for MEM whose hash | |
1370 value is HASH and id is ID. */ | |
1371 | |
111 | 1372 static im_mem_ref * |
0 | 1373 mem_ref_alloc (tree mem, unsigned hash, unsigned id) |
1374 { | |
111 | 1375 im_mem_ref *ref = XOBNEW (&mem_ref_obstack, struct im_mem_ref); |
1376 ao_ref_init (&ref->mem, mem); | |
0 | 1377 ref->id = id; |
1378 ref->hash = hash; | |
111 | 1379 ref->stored = NULL; |
1380 bitmap_initialize (&ref->indep_loop, &lim_bitmap_obstack); | |
1381 bitmap_initialize (&ref->dep_loop, &lim_bitmap_obstack); | |
1382 ref->accesses_in_loop.create (1); | |
0 | 1383 |
1384 return ref; | |
1385 } | |
1386 | |
1387 /* Records memory reference location *LOC in LOOP to the memory reference | |
1388 description REF. The reference occurs in statement STMT. */ | |
1389 | |
1390 static void | |
111 | 1391 record_mem_ref_loc (im_mem_ref *ref, gimple *stmt, tree *loc) |
0 | 1392 { |
111 | 1393 mem_ref_loc aref; |
1394 aref.stmt = stmt; | |
1395 aref.ref = loc; | |
1396 ref->accesses_in_loop.safe_push (aref); | |
1397 } | |
1398 | |
1399 /* Set the LOOP bit in REF stored bitmap and allocate that if | |
1400 necessary. Return whether a bit was changed. */ | |
1401 | |
1402 static bool | |
1403 set_ref_stored_in_loop (im_mem_ref *ref, struct loop *loop) | |
1404 { | |
1405 if (!ref->stored) | |
1406 ref->stored = BITMAP_ALLOC (&lim_bitmap_obstack); | |
1407 return bitmap_set_bit (ref->stored, loop->num); | |
0 | 1408 } |
1409 | |
1410 /* Marks reference REF as stored in LOOP. */ | |
1411 | |
1412 static void | |
111 | 1413 mark_ref_stored (im_mem_ref *ref, struct loop *loop) |
0 | 1414 { |
111 | 1415 while (loop != current_loops->tree_root |
1416 && set_ref_stored_in_loop (ref, loop)) | |
1417 loop = loop_outer (loop); | |
0 | 1418 } |
1419 | |
1420 /* Gathers memory references in statement STMT in LOOP, storing the | |
1421 information about them in the memory_accesses structure. Marks | |
1422 the vops accessed through unrecognized statements there as | |
1423 well. */ | |
1424 | |
1425 static void | |
111 | 1426 gather_mem_refs_stmt (struct loop *loop, gimple *stmt) |
0 | 1427 { |
1428 tree *mem = NULL; | |
1429 hashval_t hash; | |
111 | 1430 im_mem_ref **slot; |
1431 im_mem_ref *ref; | |
0 | 1432 bool is_stored; |
1433 unsigned id; | |
1434 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
47
diff
changeset
|
1435 if (!gimple_vuse (stmt)) |
0 | 1436 return; |
1437 | |
1438 mem = simple_mem_ref_in_stmt (stmt, &is_stored); | |
1439 if (!mem) | |
1440 { | |
111 | 1441 /* We use the shared mem_ref for all unanalyzable refs. */ |
1442 id = UNANALYZABLE_MEM_ID; | |
1443 ref = memory_accesses.refs_list[id]; | |
1444 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1445 { | |
1446 fprintf (dump_file, "Unanalyzed memory reference %u: ", id); | |
1447 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); | |
1448 } | |
1449 is_stored = gimple_vdef (stmt); | |
0 | 1450 } |
1451 else | |
1452 { | |
111 | 1453 hash = iterative_hash_expr (*mem, 0); |
1454 slot = memory_accesses.refs->find_slot_with_hash (*mem, hash, INSERT); | |
1455 if (*slot) | |
1456 { | |
1457 ref = *slot; | |
1458 id = ref->id; | |
1459 } | |
1460 else | |
0 | 1461 { |
111 | 1462 id = memory_accesses.refs_list.length (); |
1463 ref = mem_ref_alloc (*mem, hash, id); | |
1464 memory_accesses.refs_list.safe_push (ref); | |
1465 *slot = ref; | |
1466 | |
1467 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1468 { | |
1469 fprintf (dump_file, "Memory reference %u: ", id); | |
1470 print_generic_expr (dump_file, ref->mem.ref, TDF_SLIM); | |
1471 fprintf (dump_file, "\n"); | |
1472 } | |
0 | 1473 } |
111 | 1474 |
1475 record_mem_ref_loc (ref, stmt, mem); | |
1476 } | |
1477 bitmap_set_bit (&memory_accesses.refs_in_loop[loop->num], ref->id); | |
1478 if (is_stored) | |
1479 { | |
1480 bitmap_set_bit (&memory_accesses.refs_stored_in_loop[loop->num], ref->id); | |
1481 mark_ref_stored (ref, loop); | |
0 | 1482 } |
111 | 1483 init_lim_data (stmt)->ref = ref->id; |
0 | 1484 return; |
111 | 1485 } |
1486 | |
1487 static unsigned *bb_loop_postorder; | |
1488 | |
1489 /* qsort sort function to sort blocks after their loop fathers postorder. */ | |
1490 | |
1491 static int | |
1492 sort_bbs_in_loop_postorder_cmp (const void *bb1_, const void *bb2_) | |
1493 { | |
1494 basic_block bb1 = *(basic_block *)const_cast<void *>(bb1_); | |
1495 basic_block bb2 = *(basic_block *)const_cast<void *>(bb2_); | |
1496 struct loop *loop1 = bb1->loop_father; | |
1497 struct loop *loop2 = bb2->loop_father; | |
1498 if (loop1->num == loop2->num) | |
1499 return 0; | |
1500 return bb_loop_postorder[loop1->num] < bb_loop_postorder[loop2->num] ? -1 : 1; | |
1501 } | |
1502 | |
1503 /* qsort sort function to sort ref locs after their loop fathers postorder. */ | |
1504 | |
1505 static int | |
1506 sort_locs_in_loop_postorder_cmp (const void *loc1_, const void *loc2_) | |
1507 { | |
1508 mem_ref_loc *loc1 = (mem_ref_loc *)const_cast<void *>(loc1_); | |
1509 mem_ref_loc *loc2 = (mem_ref_loc *)const_cast<void *>(loc2_); | |
1510 struct loop *loop1 = gimple_bb (loc1->stmt)->loop_father; | |
1511 struct loop *loop2 = gimple_bb (loc2->stmt)->loop_father; | |
1512 if (loop1->num == loop2->num) | |
1513 return 0; | |
1514 return bb_loop_postorder[loop1->num] < bb_loop_postorder[loop2->num] ? -1 : 1; | |
0 | 1515 } |
1516 | |
1517 /* Gathers memory references in loops. */ | |
1518 | |
1519 static void | |
111 | 1520 analyze_memory_references (void) |
0 | 1521 { |
1522 gimple_stmt_iterator bsi; | |
111 | 1523 basic_block bb, *bbs; |
1524 struct loop *loop, *outer; | |
1525 unsigned i, n; | |
1526 | |
1527 /* Collect all basic-blocks in loops and sort them after their | |
1528 loops postorder. */ | |
1529 i = 0; | |
1530 bbs = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS); | |
1531 FOR_EACH_BB_FN (bb, cfun) | |
1532 if (bb->loop_father != current_loops->tree_root) | |
1533 bbs[i++] = bb; | |
1534 n = i; | |
1535 qsort (bbs, n, sizeof (basic_block), sort_bbs_in_loop_postorder_cmp); | |
1536 | |
1537 /* Visit blocks in loop postorder and assign mem-ref IDs in that order. | |
1538 That results in better locality for all the bitmaps. */ | |
1539 for (i = 0; i < n; ++i) | |
0 | 1540 { |
111 | 1541 basic_block bb = bbs[i]; |
1542 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1543 gather_mem_refs_stmt (bb->loop_father, gsi_stmt (bsi)); | |
0 | 1544 } |
111 | 1545 |
1546 /* Sort the location list of gathered memory references after their | |
1547 loop postorder number. */ | |
1548 im_mem_ref *ref; | |
1549 FOR_EACH_VEC_ELT (memory_accesses.refs_list, i, ref) | |
1550 ref->accesses_in_loop.qsort (sort_locs_in_loop_postorder_cmp); | |
1551 | |
1552 free (bbs); | |
1553 // free (bb_loop_postorder); | |
1554 | |
1555 /* Propagate the information about accessed memory references up | |
1556 the loop hierarchy. */ | |
1557 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) | |
0 | 1558 { |
111 | 1559 /* Finalize the overall touched references (including subloops). */ |
1560 bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[loop->num], | |
1561 &memory_accesses.refs_stored_in_loop[loop->num]); | |
1562 | |
1563 /* Propagate the information about accessed memory references up | |
1564 the loop hierarchy. */ | |
1565 outer = loop_outer (loop); | |
1566 if (outer == current_loops->tree_root) | |
1567 continue; | |
1568 | |
1569 bitmap_ior_into (&memory_accesses.all_refs_stored_in_loop[outer->num], | |
1570 &memory_accesses.all_refs_stored_in_loop[loop->num]); | |
0 | 1571 } |
1572 } | |
1573 | |
1574 /* Returns true if MEM1 and MEM2 may alias. TTAE_CACHE is used as a cache in | |
1575 tree_to_aff_combination_expand. */ | |
1576 | |
1577 static bool | |
111 | 1578 mem_refs_may_alias_p (im_mem_ref *mem1, im_mem_ref *mem2, |
1579 hash_map<tree, name_expansion *> **ttae_cache) | |
0 | 1580 { |
1581 /* Perform BASE + OFFSET analysis -- if MEM1 and MEM2 are based on the same | |
1582 object and their offset differ in such a way that the locations cannot | |
1583 overlap, then they cannot alias. */ | |
111 | 1584 widest_int size1, size2; |
0 | 1585 aff_tree off1, off2; |
1586 | |
1587 /* Perform basic offset and type-based disambiguation. */ | |
111 | 1588 if (!refs_may_alias_p_1 (&mem1->mem, &mem2->mem, true)) |
0 | 1589 return false; |
1590 | |
1591 /* The expansion of addresses may be a bit expensive, thus we only do | |
1592 the check at -O2 and higher optimization levels. */ | |
1593 if (optimize < 2) | |
1594 return true; | |
1595 | |
111 | 1596 get_inner_reference_aff (mem1->mem.ref, &off1, &size1); |
1597 get_inner_reference_aff (mem2->mem.ref, &off2, &size2); | |
0 | 1598 aff_combination_expand (&off1, ttae_cache); |
1599 aff_combination_expand (&off2, ttae_cache); | |
111 | 1600 aff_combination_scale (&off1, -1); |
0 | 1601 aff_combination_add (&off2, &off1); |
1602 | |
111 | 1603 if (aff_comb_cannot_overlap_p (&off2, size1, size2)) |
0 | 1604 return false; |
1605 | |
1606 return true; | |
1607 } | |
1608 | |
111 | 1609 /* Compare function for bsearch searching for reference locations |
1610 in a loop. */ | |
1611 | |
1612 static int | |
1613 find_ref_loc_in_loop_cmp (const void *loop_, const void *loc_) | |
1614 { | |
1615 struct loop *loop = (struct loop *)const_cast<void *>(loop_); | |
1616 mem_ref_loc *loc = (mem_ref_loc *)const_cast<void *>(loc_); | |
1617 struct loop *loc_loop = gimple_bb (loc->stmt)->loop_father; | |
1618 if (loop->num == loc_loop->num | |
1619 || flow_loop_nested_p (loop, loc_loop)) | |
1620 return 0; | |
1621 return (bb_loop_postorder[loop->num] < bb_loop_postorder[loc_loop->num] | |
1622 ? -1 : 1); | |
1623 } | |
1624 | |
1625 /* Iterates over all locations of REF in LOOP and its subloops calling | |
1626 fn.operator() with the location as argument. When that operator | |
1627 returns true the iteration is stopped and true is returned. | |
1628 Otherwise false is returned. */ | |
1629 | |
1630 template <typename FN> | |
1631 static bool | |
1632 for_all_locs_in_loop (struct loop *loop, im_mem_ref *ref, FN fn) | |
1633 { | |
1634 unsigned i; | |
1635 mem_ref_loc *loc; | |
1636 | |
1637 /* Search for the cluster of locs in the accesses_in_loop vector | |
1638 which is sorted after postorder index of the loop father. */ | |
1639 loc = ref->accesses_in_loop.bsearch (loop, find_ref_loc_in_loop_cmp); | |
1640 if (!loc) | |
1641 return false; | |
1642 | |
1643 /* We have found one location inside loop or its sub-loops. Iterate | |
1644 both forward and backward to cover the whole cluster. */ | |
1645 i = loc - ref->accesses_in_loop.address (); | |
1646 while (i > 0) | |
1647 { | |
1648 --i; | |
1649 mem_ref_loc *l = &ref->accesses_in_loop[i]; | |
1650 if (!flow_bb_inside_loop_p (loop, gimple_bb (l->stmt))) | |
1651 break; | |
1652 if (fn (l)) | |
1653 return true; | |
1654 } | |
1655 for (i = loc - ref->accesses_in_loop.address (); | |
1656 i < ref->accesses_in_loop.length (); ++i) | |
1657 { | |
1658 mem_ref_loc *l = &ref->accesses_in_loop[i]; | |
1659 if (!flow_bb_inside_loop_p (loop, gimple_bb (l->stmt))) | |
1660 break; | |
1661 if (fn (l)) | |
1662 return true; | |
1663 } | |
1664 | |
1665 return false; | |
1666 } | |
1667 | |
0 | 1668 /* Rewrites location LOC by TMP_VAR. */ |
1669 | |
111 | 1670 struct rewrite_mem_ref_loc |
0 | 1671 { |
111 | 1672 rewrite_mem_ref_loc (tree tmp_var_) : tmp_var (tmp_var_) {} |
1673 bool operator () (mem_ref_loc *loc); | |
1674 tree tmp_var; | |
1675 }; | |
1676 | |
1677 bool | |
1678 rewrite_mem_ref_loc::operator () (mem_ref_loc *loc) | |
1679 { | |
0 | 1680 *loc->ref = tmp_var; |
1681 update_stmt (loc->stmt); | |
111 | 1682 return false; |
0 | 1683 } |
1684 | |
1685 /* Rewrites all references to REF in LOOP by variable TMP_VAR. */ | |
1686 | |
1687 static void | |
111 | 1688 rewrite_mem_refs (struct loop *loop, im_mem_ref *ref, tree tmp_var) |
1689 { | |
1690 for_all_locs_in_loop (loop, ref, rewrite_mem_ref_loc (tmp_var)); | |
1691 } | |
1692 | |
1693 /* Stores the first reference location in LOCP. */ | |
1694 | |
1695 struct first_mem_ref_loc_1 | |
0 | 1696 { |
111 | 1697 first_mem_ref_loc_1 (mem_ref_loc **locp_) : locp (locp_) {} |
1698 bool operator () (mem_ref_loc *loc); | |
1699 mem_ref_loc **locp; | |
1700 }; | |
1701 | |
1702 bool | |
1703 first_mem_ref_loc_1::operator () (mem_ref_loc *loc) | |
1704 { | |
1705 *locp = loc; | |
1706 return true; | |
1707 } | |
1708 | |
1709 /* Returns the first reference location to REF in LOOP. */ | |
1710 | |
1711 static mem_ref_loc * | |
1712 first_mem_ref_loc (struct loop *loop, im_mem_ref *ref) | |
1713 { | |
1714 mem_ref_loc *locp = NULL; | |
1715 for_all_locs_in_loop (loop, ref, first_mem_ref_loc_1 (&locp)); | |
1716 return locp; | |
0 | 1717 } |
1718 | |
111 | 1719 struct prev_flag_edges { |
1720 /* Edge to insert new flag comparison code. */ | |
1721 edge append_cond_position; | |
1722 | |
1723 /* Edge for fall through from previous flag comparison. */ | |
1724 edge last_cond_fallthru; | |
1725 }; | |
1726 | |
1727 /* Helper function for execute_sm. Emit code to store TMP_VAR into | |
1728 MEM along edge EX. | |
1729 | |
1730 The store is only done if MEM has changed. We do this so no | |
1731 changes to MEM occur on code paths that did not originally store | |
1732 into it. | |
1733 | |
1734 The common case for execute_sm will transform: | |
1735 | |
1736 for (...) { | |
1737 if (foo) | |
1738 stuff; | |
1739 else | |
1740 MEM = TMP_VAR; | |
1741 } | |
1742 | |
1743 into: | |
1744 | |
1745 lsm = MEM; | |
1746 for (...) { | |
1747 if (foo) | |
1748 stuff; | |
1749 else | |
1750 lsm = TMP_VAR; | |
1751 } | |
1752 MEM = lsm; | |
1753 | |
1754 This function will generate: | |
1755 | |
1756 lsm = MEM; | |
1757 | |
1758 lsm_flag = false; | |
1759 ... | |
1760 for (...) { | |
1761 if (foo) | |
1762 stuff; | |
1763 else { | |
1764 lsm = TMP_VAR; | |
1765 lsm_flag = true; | |
1766 } | |
1767 } | |
1768 if (lsm_flag) <-- | |
1769 MEM = lsm; <-- | |
1770 */ | |
0 | 1771 |
1772 static void | |
111 | 1773 execute_sm_if_changed (edge ex, tree mem, tree tmp_var, tree flag, |
1774 edge preheader, hash_set <basic_block> *flag_bbs) | |
0 | 1775 { |
111 | 1776 basic_block new_bb, then_bb, old_dest; |
1777 bool loop_has_only_one_exit; | |
1778 edge then_old_edge, orig_ex = ex; | |
1779 gimple_stmt_iterator gsi; | |
1780 gimple *stmt; | |
1781 struct prev_flag_edges *prev_edges = (struct prev_flag_edges *) ex->aux; | |
1782 bool irr = ex->flags & EDGE_IRREDUCIBLE_LOOP; | |
1783 | |
1784 int freq_sum = 0; | |
1785 profile_count count_sum = profile_count::zero (); | |
1786 int nbbs = 0, ncount = 0; | |
1787 profile_probability flag_probability = profile_probability::uninitialized (); | |
1788 | |
1789 /* Flag is set in FLAG_BBS. Determine probability that flag will be true | |
1790 at loop exit. | |
1791 | |
1792 This code may look fancy, but it can not update profile very realistically | |
1793 because we do not know the probability that flag will be true at given | |
1794 loop exit. | |
1795 | |
1796 We look for two interesting extremes | |
1797 - when exit is dominated by block setting the flag, we know it will | |
1798 always be true. This is a common case. | |
1799 - when all blocks setting the flag have very low frequency we know | |
1800 it will likely be false. | |
1801 In all other cases we default to 2/3 for flag being true. */ | |
1802 | |
1803 for (hash_set<basic_block>::iterator it = flag_bbs->begin (); | |
1804 it != flag_bbs->end (); ++it) | |
1805 { | |
1806 freq_sum += (*it)->frequency; | |
1807 if ((*it)->count.initialized_p ()) | |
1808 count_sum += (*it)->count, ncount ++; | |
1809 if (dominated_by_p (CDI_DOMINATORS, ex->src, *it)) | |
1810 flag_probability = profile_probability::always (); | |
1811 nbbs++; | |
1812 } | |
1813 | |
1814 profile_probability cap = profile_probability::always ().apply_scale (2, 3); | |
1815 | |
1816 if (flag_probability.initialized_p ()) | |
1817 ; | |
1818 else if (ncount == nbbs && count_sum > 0 && preheader->count () >= count_sum) | |
1819 { | |
1820 flag_probability = count_sum.probability_in (preheader->count ()); | |
1821 if (flag_probability > cap) | |
1822 flag_probability = cap; | |
1823 } | |
1824 else if (freq_sum > 0 && EDGE_FREQUENCY (preheader) >= freq_sum) | |
1825 { | |
1826 flag_probability = profile_probability::from_reg_br_prob_base | |
1827 (GCOV_COMPUTE_SCALE (freq_sum, EDGE_FREQUENCY (preheader))); | |
1828 if (flag_probability > cap) | |
1829 flag_probability = cap; | |
1830 } | |
1831 else | |
1832 flag_probability = cap; | |
1833 | |
1834 /* ?? Insert store after previous store if applicable. See note | |
1835 below. */ | |
1836 if (prev_edges) | |
1837 ex = prev_edges->append_cond_position; | |
1838 | |
1839 loop_has_only_one_exit = single_pred_p (ex->dest); | |
1840 | |
1841 if (loop_has_only_one_exit) | |
1842 ex = split_block_after_labels (ex->dest); | |
1843 else | |
0 | 1844 { |
111 | 1845 for (gphi_iterator gpi = gsi_start_phis (ex->dest); |
1846 !gsi_end_p (gpi); gsi_next (&gpi)) | |
1847 { | |
1848 gphi *phi = gpi.phi (); | |
1849 if (virtual_operand_p (gimple_phi_result (phi))) | |
1850 continue; | |
1851 | |
1852 /* When the destination has a non-virtual PHI node with multiple | |
1853 predecessors make sure we preserve the PHI structure by | |
1854 forcing a forwarder block so that hoisting of that PHI will | |
1855 still work. */ | |
1856 split_edge (ex); | |
1857 break; | |
1858 } | |
1859 } | |
1860 | |
1861 old_dest = ex->dest; | |
1862 new_bb = split_edge (ex); | |
1863 then_bb = create_empty_bb (new_bb); | |
1864 then_bb->frequency = flag_probability.apply (new_bb->frequency); | |
1865 then_bb->count = new_bb->count.apply_probability (flag_probability); | |
1866 if (irr) | |
1867 then_bb->flags = BB_IRREDUCIBLE_LOOP; | |
1868 add_bb_to_loop (then_bb, new_bb->loop_father); | |
1869 | |
1870 gsi = gsi_start_bb (new_bb); | |
1871 stmt = gimple_build_cond (NE_EXPR, flag, boolean_false_node, | |
1872 NULL_TREE, NULL_TREE); | |
1873 gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
1874 | |
1875 gsi = gsi_start_bb (then_bb); | |
1876 /* Insert actual store. */ | |
1877 stmt = gimple_build_assign (unshare_expr (mem), tmp_var); | |
1878 gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
1879 | |
1880 edge e1 = single_succ_edge (new_bb); | |
1881 edge e2 = make_edge (new_bb, then_bb, | |
1882 EDGE_TRUE_VALUE | (irr ? EDGE_IRREDUCIBLE_LOOP : 0)); | |
1883 e2->probability = flag_probability; | |
1884 | |
1885 e1->flags |= EDGE_FALSE_VALUE | (irr ? EDGE_IRREDUCIBLE_LOOP : 0); | |
1886 e1->flags &= ~EDGE_FALLTHRU; | |
1887 | |
1888 e1->probability = flag_probability.invert (); | |
1889 | |
1890 then_old_edge = make_single_succ_edge (then_bb, old_dest, | |
1891 EDGE_FALLTHRU | (irr ? EDGE_IRREDUCIBLE_LOOP : 0)); | |
1892 | |
1893 set_immediate_dominator (CDI_DOMINATORS, then_bb, new_bb); | |
1894 | |
1895 if (prev_edges) | |
1896 { | |
1897 basic_block prevbb = prev_edges->last_cond_fallthru->src; | |
1898 redirect_edge_succ (prev_edges->last_cond_fallthru, new_bb); | |
1899 set_immediate_dominator (CDI_DOMINATORS, new_bb, prevbb); | |
1900 set_immediate_dominator (CDI_DOMINATORS, old_dest, | |
1901 recompute_dominator (CDI_DOMINATORS, old_dest)); | |
0 | 1902 } |
111 | 1903 |
1904 /* ?? Because stores may alias, they must happen in the exact | |
1905 sequence they originally happened. Save the position right after | |
1906 the (_lsm) store we just created so we can continue appending after | |
1907 it and maintain the original order. */ | |
1908 { | |
1909 struct prev_flag_edges *p; | |
1910 | |
1911 if (orig_ex->aux) | |
1912 orig_ex->aux = NULL; | |
1913 alloc_aux_for_edge (orig_ex, sizeof (struct prev_flag_edges)); | |
1914 p = (struct prev_flag_edges *) orig_ex->aux; | |
1915 p->append_cond_position = then_old_edge; | |
1916 p->last_cond_fallthru = find_edge (new_bb, old_dest); | |
1917 orig_ex->aux = (void *) p; | |
1918 } | |
1919 | |
1920 if (!loop_has_only_one_exit) | |
1921 for (gphi_iterator gpi = gsi_start_phis (old_dest); | |
1922 !gsi_end_p (gpi); gsi_next (&gpi)) | |
1923 { | |
1924 gphi *phi = gpi.phi (); | |
1925 unsigned i; | |
1926 | |
1927 for (i = 0; i < gimple_phi_num_args (phi); i++) | |
1928 if (gimple_phi_arg_edge (phi, i)->src == new_bb) | |
1929 { | |
1930 tree arg = gimple_phi_arg_def (phi, i); | |
1931 add_phi_arg (phi, arg, then_old_edge, UNKNOWN_LOCATION); | |
1932 update_stmt (phi); | |
1933 } | |
1934 } | |
0 | 1935 } |
1936 | |
111 | 1937 /* When REF is set on the location, set flag indicating the store. */ |
1938 | |
1939 struct sm_set_flag_if_changed | |
0 | 1940 { |
111 | 1941 sm_set_flag_if_changed (tree flag_, hash_set <basic_block> *bbs_) |
1942 : flag (flag_), bbs (bbs_) {} | |
1943 bool operator () (mem_ref_loc *loc); | |
1944 tree flag; | |
1945 hash_set <basic_block> *bbs; | |
1946 }; | |
1947 | |
1948 bool | |
1949 sm_set_flag_if_changed::operator () (mem_ref_loc *loc) | |
1950 { | |
1951 /* Only set the flag for writes. */ | |
1952 if (is_gimple_assign (loc->stmt) | |
1953 && gimple_assign_lhs_ptr (loc->stmt) == loc->ref) | |
0 | 1954 { |
111 | 1955 gimple_stmt_iterator gsi = gsi_for_stmt (loc->stmt); |
1956 gimple *stmt = gimple_build_assign (flag, boolean_true_node); | |
1957 gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
1958 bbs->add (gimple_bb (stmt)); | |
0 | 1959 } |
111 | 1960 return false; |
1961 } | |
1962 | |
1963 /* Helper function for execute_sm. On every location where REF is | |
1964 set, set an appropriate flag indicating the store. */ | |
1965 | |
1966 static tree | |
1967 execute_sm_if_changed_flag_set (struct loop *loop, im_mem_ref *ref, | |
1968 hash_set <basic_block> *bbs) | |
1969 { | |
1970 tree flag; | |
1971 char *str = get_lsm_tmp_name (ref->mem.ref, ~0, "_flag"); | |
1972 flag = create_tmp_reg (boolean_type_node, str); | |
1973 for_all_locs_in_loop (loop, ref, sm_set_flag_if_changed (flag, bbs)); | |
1974 return flag; | |
0 | 1975 } |
1976 | |
1977 /* Executes store motion of memory reference REF from LOOP. | |
1978 Exits from the LOOP are stored in EXITS. The initialization of the | |
1979 temporary variable is put to the preheader of the loop, and assignments | |
1980 to the reference from the temporary variable are emitted to exits. */ | |
1981 | |
1982 static void | |
111 | 1983 execute_sm (struct loop *loop, vec<edge> exits, im_mem_ref *ref) |
0 | 1984 { |
111 | 1985 tree tmp_var, store_flag = NULL_TREE; |
0 | 1986 unsigned i; |
111 | 1987 gassign *load; |
0 | 1988 struct fmt_data fmt_data; |
1989 edge ex; | |
1990 struct lim_aux_data *lim_data; | |
111 | 1991 bool multi_threaded_model_p = false; |
1992 gimple_stmt_iterator gsi; | |
1993 hash_set<basic_block> flag_bbs; | |
0 | 1994 |
1995 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1996 { | |
1997 fprintf (dump_file, "Executing store motion of "); | |
111 | 1998 print_generic_expr (dump_file, ref->mem.ref); |
0 | 1999 fprintf (dump_file, " from loop %d\n", loop->num); |
2000 } | |
2001 | |
111 | 2002 tmp_var = create_tmp_reg (TREE_TYPE (ref->mem.ref), |
2003 get_lsm_tmp_name (ref->mem.ref, ~0)); | |
0 | 2004 |
2005 fmt_data.loop = loop; | |
2006 fmt_data.orig_loop = loop; | |
111 | 2007 for_each_index (&ref->mem.ref, force_move_till, &fmt_data); |
2008 | |
2009 if (bb_in_transaction (loop_preheader_edge (loop)->src) | |
2010 || (! PARAM_VALUE (PARAM_ALLOW_STORE_DATA_RACES) | |
2011 && ! ref_always_accessed_p (loop, ref, true))) | |
2012 multi_threaded_model_p = true; | |
2013 | |
2014 if (multi_threaded_model_p) | |
2015 store_flag = execute_sm_if_changed_flag_set (loop, ref, &flag_bbs); | |
0 | 2016 |
2017 rewrite_mem_refs (loop, ref, tmp_var); | |
2018 | |
111 | 2019 /* Emit the load code on a random exit edge or into the latch if |
2020 the loop does not exit, so that we are sure it will be processed | |
2021 by move_computations after all dependencies. */ | |
2022 gsi = gsi_for_stmt (first_mem_ref_loc (loop, ref)->stmt); | |
2023 | |
2024 /* FIXME/TODO: For the multi-threaded variant, we could avoid this | |
2025 load altogether, since the store is predicated by a flag. We | |
2026 could, do the load only if it was originally in the loop. */ | |
2027 load = gimple_build_assign (tmp_var, unshare_expr (ref->mem.ref)); | |
0 | 2028 lim_data = init_lim_data (load); |
2029 lim_data->max_loop = loop; | |
2030 lim_data->tgt_loop = loop; | |
111 | 2031 gsi_insert_before (&gsi, load, GSI_SAME_STMT); |
2032 | |
2033 if (multi_threaded_model_p) | |
0 | 2034 { |
111 | 2035 load = gimple_build_assign (store_flag, boolean_false_node); |
2036 lim_data = init_lim_data (load); | |
2037 lim_data->max_loop = loop; | |
2038 lim_data->tgt_loop = loop; | |
2039 gsi_insert_before (&gsi, load, GSI_SAME_STMT); | |
0 | 2040 } |
111 | 2041 |
2042 /* Sink the store to every exit from the loop. */ | |
2043 FOR_EACH_VEC_ELT (exits, i, ex) | |
2044 if (!multi_threaded_model_p) | |
2045 { | |
2046 gassign *store; | |
2047 store = gimple_build_assign (unshare_expr (ref->mem.ref), tmp_var); | |
2048 gsi_insert_on_edge (ex, store); | |
2049 } | |
2050 else | |
2051 execute_sm_if_changed (ex, ref->mem.ref, tmp_var, store_flag, | |
2052 loop_preheader_edge (loop), &flag_bbs); | |
0 | 2053 } |
2054 | |
2055 /* Hoists memory references MEM_REFS out of LOOP. EXITS is the list of exit | |
2056 edges of the LOOP. */ | |
2057 | |
2058 static void | |
2059 hoist_memory_references (struct loop *loop, bitmap mem_refs, | |
111 | 2060 vec<edge> exits) |
0 | 2061 { |
111 | 2062 im_mem_ref *ref; |
0 | 2063 unsigned i; |
2064 bitmap_iterator bi; | |
2065 | |
2066 EXECUTE_IF_SET_IN_BITMAP (mem_refs, 0, i, bi) | |
2067 { | |
111 | 2068 ref = memory_accesses.refs_list[i]; |
0 | 2069 execute_sm (loop, exits, ref); |
2070 } | |
2071 } | |
2072 | |
111 | 2073 struct ref_always_accessed |
2074 { | |
2075 ref_always_accessed (struct loop *loop_, bool stored_p_) | |
2076 : loop (loop_), stored_p (stored_p_) {} | |
2077 bool operator () (mem_ref_loc *loc); | |
2078 struct loop *loop; | |
2079 bool stored_p; | |
2080 }; | |
2081 | |
2082 bool | |
2083 ref_always_accessed::operator () (mem_ref_loc *loc) | |
2084 { | |
2085 struct loop *must_exec; | |
2086 | |
2087 if (!get_lim_data (loc->stmt)) | |
2088 return false; | |
2089 | |
2090 /* If we require an always executed store make sure the statement | |
2091 stores to the reference. */ | |
2092 if (stored_p) | |
2093 { | |
2094 tree lhs = gimple_get_lhs (loc->stmt); | |
2095 if (!lhs | |
2096 || lhs != *loc->ref) | |
2097 return false; | |
2098 } | |
2099 | |
2100 must_exec = get_lim_data (loc->stmt)->always_executed_in; | |
2101 if (!must_exec) | |
2102 return false; | |
2103 | |
2104 if (must_exec == loop | |
2105 || flow_loop_nested_p (must_exec, loop)) | |
2106 return true; | |
2107 | |
2108 return false; | |
2109 } | |
2110 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2111 /* 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
|
2112 make sure REF is always stored to in LOOP. */ |
0 | 2113 |
2114 static bool | |
111 | 2115 ref_always_accessed_p (struct loop *loop, im_mem_ref *ref, bool stored_p) |
0 | 2116 { |
111 | 2117 return for_all_locs_in_loop (loop, ref, |
2118 ref_always_accessed (loop, stored_p)); | |
0 | 2119 } |
2120 | |
2121 /* Returns true if REF1 and REF2 are independent. */ | |
2122 | |
2123 static bool | |
111 | 2124 refs_independent_p (im_mem_ref *ref1, im_mem_ref *ref2) |
0 | 2125 { |
111 | 2126 if (ref1 == ref2) |
0 | 2127 return true; |
2128 | |
2129 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2130 fprintf (dump_file, "Querying dependency of refs %u and %u: ", | |
2131 ref1->id, ref2->id); | |
2132 | |
111 | 2133 if (mem_refs_may_alias_p (ref1, ref2, &memory_accesses.ttae_cache)) |
0 | 2134 { |
2135 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2136 fprintf (dump_file, "dependent.\n"); | |
2137 return false; | |
2138 } | |
2139 else | |
2140 { | |
2141 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2142 fprintf (dump_file, "independent.\n"); | |
2143 return true; | |
2144 } | |
2145 } | |
2146 | |
111 | 2147 /* Mark REF dependent on stores or loads (according to STORED_P) in LOOP |
2148 and its super-loops. */ | |
0 | 2149 |
2150 static void | |
111 | 2151 record_dep_loop (struct loop *loop, im_mem_ref *ref, bool stored_p) |
2152 { | |
2153 /* We can propagate dependent-in-loop bits up the loop | |
2154 hierarchy to all outer loops. */ | |
2155 while (loop != current_loops->tree_root | |
2156 && bitmap_set_bit (&ref->dep_loop, LOOP_DEP_BIT (loop->num, stored_p))) | |
2157 loop = loop_outer (loop); | |
2158 } | |
2159 | |
2160 /* Returns true if REF is independent on all other memory | |
2161 references in LOOP. REF_LOOP is where REF is accessed, SAFELEN is the | |
2162 safelen to apply. */ | |
2163 | |
2164 static bool | |
2165 ref_indep_loop_p_1 (int safelen, struct loop *loop, im_mem_ref *ref, | |
2166 bool stored_p, struct loop *ref_loop) | |
0 | 2167 { |
111 | 2168 stored_p |= (ref->stored && bitmap_bit_p (ref->stored, loop->num)); |
2169 | |
2170 if (loop->safelen > safelen | |
2171 /* Check that REF is accessed inside LOOP. */ | |
2172 && (loop == ref_loop || flow_loop_nested_p (loop, ref_loop))) | |
2173 safelen = loop->safelen; | |
2174 | |
2175 bool indep_p = true; | |
2176 bitmap refs_to_check; | |
2177 | |
2178 if (stored_p) | |
2179 refs_to_check = &memory_accesses.refs_in_loop[loop->num]; | |
2180 else | |
2181 refs_to_check = &memory_accesses.refs_stored_in_loop[loop->num]; | |
2182 | |
2183 if (bitmap_bit_p (refs_to_check, UNANALYZABLE_MEM_ID)) | |
2184 indep_p = false; | |
2185 else if (safelen > 1) | |
2186 { | |
2187 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2188 { | |
2189 fprintf (dump_file,"REF is independent due to safelen %d\n", | |
2190 safelen); | |
2191 print_generic_expr (dump_file, ref->mem.ref, TDF_SLIM); | |
2192 fprintf (dump_file, "\n"); | |
2193 } | |
2194 | |
2195 /* We need to recurse to properly handle UNANALYZABLE_MEM_ID. */ | |
2196 struct loop *inner = loop->inner; | |
2197 while (inner) | |
2198 { | |
2199 if (!ref_indep_loop_p_1 (safelen, inner, ref, stored_p, ref_loop)) | |
2200 { | |
2201 indep_p = false; | |
2202 break; | |
2203 } | |
2204 inner = inner->next; | |
2205 } | |
2206 | |
2207 /* Avoid caching here as safelen depends on context and refs | |
2208 are shared between different contexts. */ | |
2209 return indep_p; | |
2210 } | |
0 | 2211 else |
111 | 2212 { |
2213 if (bitmap_bit_p (&ref->indep_loop, LOOP_DEP_BIT (loop->num, stored_p))) | |
2214 return true; | |
2215 if (bitmap_bit_p (&ref->dep_loop, LOOP_DEP_BIT (loop->num, stored_p))) | |
2216 return false; | |
2217 | |
2218 struct loop *inner = loop->inner; | |
2219 while (inner) | |
2220 { | |
2221 if (!ref_indep_loop_p_1 (safelen, inner, ref, stored_p, ref_loop)) | |
2222 { | |
2223 indep_p = false; | |
2224 break; | |
2225 } | |
2226 inner = inner->next; | |
2227 } | |
2228 | |
2229 if (indep_p) | |
2230 { | |
2231 unsigned i; | |
2232 bitmap_iterator bi; | |
2233 EXECUTE_IF_SET_IN_BITMAP (refs_to_check, 0, i, bi) | |
2234 { | |
2235 im_mem_ref *aref = memory_accesses.refs_list[i]; | |
2236 if (!refs_independent_p (ref, aref)) | |
2237 { | |
2238 indep_p = false; | |
2239 break; | |
2240 } | |
2241 } | |
2242 } | |
2243 } | |
2244 | |
2245 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2246 fprintf (dump_file, "Querying dependencies of ref %u in loop %d: %s\n", | |
2247 ref->id, loop->num, indep_p ? "independent" : "dependent"); | |
2248 | |
2249 /* Record the computed result in the cache. */ | |
2250 if (indep_p) | |
2251 { | |
2252 if (bitmap_set_bit (&ref->indep_loop, LOOP_DEP_BIT (loop->num, stored_p)) | |
2253 && stored_p) | |
2254 { | |
2255 /* If it's independend against all refs then it's independent | |
2256 against stores, too. */ | |
2257 bitmap_set_bit (&ref->indep_loop, LOOP_DEP_BIT (loop->num, false)); | |
2258 } | |
2259 } | |
2260 else | |
2261 { | |
2262 record_dep_loop (loop, ref, stored_p); | |
2263 if (!stored_p) | |
2264 { | |
2265 /* If it's dependent against stores it's dependent against | |
2266 all refs, too. */ | |
2267 record_dep_loop (loop, ref, true); | |
2268 } | |
2269 } | |
2270 | |
2271 return indep_p; | |
0 | 2272 } |
2273 | |
2274 /* Returns true if REF is independent on all other memory references in | |
111 | 2275 LOOP. REF_LOOP is the loop where REF is accessed. */ |
0 | 2276 |
2277 static bool | |
111 | 2278 ref_indep_loop_p (struct loop *loop, im_mem_ref *ref, struct loop *ref_loop) |
0 | 2279 { |
111 | 2280 gcc_checking_assert (MEM_ANALYZABLE (ref)); |
2281 | |
2282 return ref_indep_loop_p_1 (0, loop, ref, false, ref_loop); | |
0 | 2283 } |
2284 | |
2285 /* Returns true if we can perform store motion of REF from LOOP. */ | |
2286 | |
2287 static bool | |
111 | 2288 can_sm_ref_p (struct loop *loop, im_mem_ref *ref) |
0 | 2289 { |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2290 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
|
2291 |
111 | 2292 /* Can't hoist unanalyzable refs. */ |
2293 if (!MEM_ANALYZABLE (ref)) | |
0 | 2294 return false; |
2295 | |
2296 /* It should be movable. */ | |
111 | 2297 if (!is_gimple_reg_type (TREE_TYPE (ref->mem.ref)) |
2298 || TREE_THIS_VOLATILE (ref->mem.ref) | |
2299 || !for_each_index (&ref->mem.ref, may_move_till, loop)) | |
0 | 2300 return false; |
2301 | |
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
|
2302 /* If it can throw fail, we do not properly update EH info. */ |
111 | 2303 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
|
2304 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
|
2305 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2306 /* 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
|
2307 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
|
2308 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
|
2309 explicitly. */ |
111 | 2310 base = get_base_address (ref->mem.ref); |
2311 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
|
2312 || (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
|
2313 && !ref_always_accessed_p (loop, ref, true)) |
0 | 2314 return false; |
2315 | |
2316 /* And it must be independent on all other memory references | |
2317 in LOOP. */ | |
111 | 2318 if (!ref_indep_loop_p (loop, ref, loop)) |
0 | 2319 return false; |
2320 | |
2321 return true; | |
2322 } | |
2323 | |
2324 /* Marks the references in LOOP for that store motion should be performed | |
2325 in REFS_TO_SM. SM_EXECUTED is the set of references for that store | |
2326 motion was performed in one of the outer loops. */ | |
2327 | |
2328 static void | |
2329 find_refs_for_sm (struct loop *loop, bitmap sm_executed, bitmap refs_to_sm) | |
2330 { | |
111 | 2331 bitmap refs = &memory_accesses.all_refs_stored_in_loop[loop->num]; |
0 | 2332 unsigned i; |
2333 bitmap_iterator bi; | |
111 | 2334 im_mem_ref *ref; |
0 | 2335 |
2336 EXECUTE_IF_AND_COMPL_IN_BITMAP (refs, sm_executed, 0, i, bi) | |
2337 { | |
111 | 2338 ref = memory_accesses.refs_list[i]; |
0 | 2339 if (can_sm_ref_p (loop, ref)) |
2340 bitmap_set_bit (refs_to_sm, i); | |
2341 } | |
2342 } | |
2343 | |
2344 /* Checks whether LOOP (with exits stored in EXITS array) is suitable | |
2345 for a store motion optimization (i.e. whether we can insert statement | |
2346 on its exits). */ | |
2347 | |
2348 static bool | |
2349 loop_suitable_for_sm (struct loop *loop ATTRIBUTE_UNUSED, | |
111 | 2350 vec<edge> exits) |
0 | 2351 { |
2352 unsigned i; | |
2353 edge ex; | |
2354 | |
111 | 2355 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
|
2356 if (ex->flags & (EDGE_ABNORMAL | EDGE_EH)) |
0 | 2357 return false; |
2358 | |
2359 return true; | |
2360 } | |
2361 | |
2362 /* Try to perform store motion for all memory references modified inside | |
2363 LOOP. SM_EXECUTED is the bitmap of the memory references for that | |
2364 store motion was executed in one of the outer loops. */ | |
2365 | |
2366 static void | |
2367 store_motion_loop (struct loop *loop, bitmap sm_executed) | |
2368 { | |
111 | 2369 vec<edge> exits = get_loop_exit_edges (loop); |
0 | 2370 struct loop *subloop; |
111 | 2371 bitmap sm_in_loop = BITMAP_ALLOC (&lim_bitmap_obstack); |
0 | 2372 |
2373 if (loop_suitable_for_sm (loop, exits)) | |
2374 { | |
2375 find_refs_for_sm (loop, sm_executed, sm_in_loop); | |
2376 hoist_memory_references (loop, sm_in_loop, exits); | |
2377 } | |
111 | 2378 exits.release (); |
0 | 2379 |
2380 bitmap_ior_into (sm_executed, sm_in_loop); | |
2381 for (subloop = loop->inner; subloop != NULL; subloop = subloop->next) | |
2382 store_motion_loop (subloop, sm_executed); | |
2383 bitmap_and_compl_into (sm_executed, sm_in_loop); | |
2384 BITMAP_FREE (sm_in_loop); | |
2385 } | |
2386 | |
2387 /* Try to perform store motion for all memory references modified inside | |
2388 loops. */ | |
2389 | |
2390 static void | |
2391 store_motion (void) | |
2392 { | |
2393 struct loop *loop; | |
111 | 2394 bitmap sm_executed = BITMAP_ALLOC (&lim_bitmap_obstack); |
0 | 2395 |
2396 for (loop = current_loops->tree_root->inner; loop != NULL; loop = loop->next) | |
2397 store_motion_loop (loop, sm_executed); | |
2398 | |
2399 BITMAP_FREE (sm_executed); | |
2400 gsi_commit_edge_inserts (); | |
2401 } | |
2402 | |
2403 /* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e. | |
2404 for each such basic block bb records the outermost loop for that execution | |
2405 of its header implies execution of bb. CONTAINS_CALL is the bitmap of | |
2406 blocks that contain a nonpure call. */ | |
2407 | |
2408 static void | |
111 | 2409 fill_always_executed_in_1 (struct loop *loop, sbitmap contains_call) |
0 | 2410 { |
2411 basic_block bb = NULL, *bbs, last = NULL; | |
2412 unsigned i; | |
2413 edge e; | |
2414 struct loop *inn_loop = loop; | |
2415 | |
111 | 2416 if (ALWAYS_EXECUTED_IN (loop->header) == NULL) |
0 | 2417 { |
2418 bbs = get_loop_body_in_dom_order (loop); | |
2419 | |
2420 for (i = 0; i < loop->num_nodes; i++) | |
2421 { | |
2422 edge_iterator ei; | |
2423 bb = bbs[i]; | |
2424 | |
2425 if (dominated_by_p (CDI_DOMINATORS, loop->latch, bb)) | |
2426 last = bb; | |
2427 | |
111 | 2428 if (bitmap_bit_p (contains_call, bb->index)) |
0 | 2429 break; |
2430 | |
2431 FOR_EACH_EDGE (e, ei, bb->succs) | |
111 | 2432 { |
2433 /* If there is an exit from this BB. */ | |
2434 if (!flow_bb_inside_loop_p (loop, e->dest)) | |
2435 break; | |
2436 /* Or we enter a possibly non-finite loop. */ | |
2437 if (flow_loop_nested_p (bb->loop_father, | |
2438 e->dest->loop_father) | |
2439 && ! finite_loop_p (e->dest->loop_father)) | |
2440 break; | |
2441 } | |
0 | 2442 if (e) |
2443 break; | |
2444 | |
2445 /* A loop might be infinite (TODO use simple loop analysis | |
2446 to disprove this if possible). */ | |
2447 if (bb->flags & BB_IRREDUCIBLE_LOOP) | |
2448 break; | |
2449 | |
2450 if (!flow_bb_inside_loop_p (inn_loop, bb)) | |
2451 break; | |
2452 | |
2453 if (bb->loop_father->header == bb) | |
2454 { | |
2455 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb)) | |
2456 break; | |
2457 | |
2458 /* In a loop that is always entered we may proceed anyway. | |
2459 But record that we entered it and stop once we leave it. */ | |
2460 inn_loop = bb->loop_father; | |
2461 } | |
2462 } | |
2463 | |
2464 while (1) | |
2465 { | |
111 | 2466 SET_ALWAYS_EXECUTED_IN (last, loop); |
0 | 2467 if (last == loop->header) |
2468 break; | |
2469 last = get_immediate_dominator (CDI_DOMINATORS, last); | |
2470 } | |
2471 | |
2472 free (bbs); | |
2473 } | |
2474 | |
2475 for (loop = loop->inner; loop; loop = loop->next) | |
111 | 2476 fill_always_executed_in_1 (loop, contains_call); |
0 | 2477 } |
2478 | |
111 | 2479 /* Fills ALWAYS_EXECUTED_IN information for basic blocks, i.e. |
2480 for each such basic block bb records the outermost loop for that execution | |
2481 of its header implies execution of bb. */ | |
2482 | |
2483 static void | |
2484 fill_always_executed_in (void) | |
2485 { | |
2486 basic_block bb; | |
2487 struct loop *loop; | |
2488 | |
2489 auto_sbitmap contains_call (last_basic_block_for_fn (cfun)); | |
2490 bitmap_clear (contains_call); | |
2491 FOR_EACH_BB_FN (bb, cfun) | |
2492 { | |
2493 gimple_stmt_iterator gsi; | |
2494 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2495 { | |
2496 if (nonpure_call_p (gsi_stmt (gsi))) | |
2497 break; | |
2498 } | |
2499 | |
2500 if (!gsi_end_p (gsi)) | |
2501 bitmap_set_bit (contains_call, bb->index); | |
2502 } | |
2503 | |
2504 for (loop = current_loops->tree_root->inner; loop; loop = loop->next) | |
2505 fill_always_executed_in_1 (loop, contains_call); | |
2506 } | |
2507 | |
2508 | |
0 | 2509 /* Compute the global information needed by the loop invariant motion pass. */ |
2510 | |
2511 static void | |
2512 tree_ssa_lim_initialize (void) | |
2513 { | |
2514 struct loop *loop; | |
111 | 2515 unsigned i; |
2516 | |
2517 bitmap_obstack_initialize (&lim_bitmap_obstack); | |
2518 gcc_obstack_init (&mem_ref_obstack); | |
2519 lim_aux_data_map = new hash_map<gimple *, lim_aux_data *>; | |
2520 | |
2521 if (flag_tm) | |
2522 compute_transaction_bits (); | |
2523 | |
2524 alloc_aux_for_edges (0); | |
2525 | |
2526 memory_accesses.refs = new hash_table<mem_ref_hasher> (100); | |
2527 memory_accesses.refs_list.create (100); | |
2528 /* Allocate a special, unanalyzable mem-ref with ID zero. */ | |
2529 memory_accesses.refs_list.quick_push | |
2530 (mem_ref_alloc (error_mark_node, 0, UNANALYZABLE_MEM_ID)); | |
2531 | |
2532 memory_accesses.refs_in_loop.create (number_of_loops (cfun)); | |
2533 memory_accesses.refs_in_loop.quick_grow (number_of_loops (cfun)); | |
2534 memory_accesses.refs_stored_in_loop.create (number_of_loops (cfun)); | |
2535 memory_accesses.refs_stored_in_loop.quick_grow (number_of_loops (cfun)); | |
2536 memory_accesses.all_refs_stored_in_loop.create (number_of_loops (cfun)); | |
2537 memory_accesses.all_refs_stored_in_loop.quick_grow (number_of_loops (cfun)); | |
2538 | |
2539 for (i = 0; i < number_of_loops (cfun); i++) | |
0 | 2540 { |
111 | 2541 bitmap_initialize (&memory_accesses.refs_in_loop[i], |
2542 &lim_bitmap_obstack); | |
2543 bitmap_initialize (&memory_accesses.refs_stored_in_loop[i], | |
2544 &lim_bitmap_obstack); | |
2545 bitmap_initialize (&memory_accesses.all_refs_stored_in_loop[i], | |
2546 &lim_bitmap_obstack); | |
0 | 2547 } |
2548 | |
111 | 2549 memory_accesses.ttae_cache = NULL; |
2550 | |
2551 /* Initialize bb_loop_postorder with a mapping from loop->num to | |
2552 its postorder index. */ | |
2553 i = 0; | |
2554 bb_loop_postorder = XNEWVEC (unsigned, number_of_loops (cfun)); | |
2555 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) | |
2556 bb_loop_postorder[loop->num] = i++; | |
0 | 2557 } |
2558 | |
2559 /* Cleans up after the invariant motion pass. */ | |
2560 | |
2561 static void | |
2562 tree_ssa_lim_finalize (void) | |
2563 { | |
2564 basic_block bb; | |
2565 unsigned i; | |
111 | 2566 im_mem_ref *ref; |
2567 | |
2568 free_aux_for_edges (); | |
2569 | |
2570 FOR_EACH_BB_FN (bb, cfun) | |
2571 SET_ALWAYS_EXECUTED_IN (bb, NULL); | |
2572 | |
2573 bitmap_obstack_release (&lim_bitmap_obstack); | |
2574 delete lim_aux_data_map; | |
2575 | |
2576 delete memory_accesses.refs; | |
2577 memory_accesses.refs = NULL; | |
2578 | |
2579 FOR_EACH_VEC_ELT (memory_accesses.refs_list, i, ref) | |
2580 memref_free (ref); | |
2581 memory_accesses.refs_list.release (); | |
2582 obstack_free (&mem_ref_obstack, NULL); | |
2583 | |
2584 memory_accesses.refs_in_loop.release (); | |
2585 memory_accesses.refs_stored_in_loop.release (); | |
2586 memory_accesses.all_refs_stored_in_loop.release (); | |
0 | 2587 |
2588 if (memory_accesses.ttae_cache) | |
111 | 2589 free_affine_expand_cache (&memory_accesses.ttae_cache); |
2590 | |
2591 free (bb_loop_postorder); | |
0 | 2592 } |
2593 | |
2594 /* Moves invariants from loops. Only "expensive" invariants are moved out -- | |
2595 i.e. those that are likely to be win regardless of the register pressure. */ | |
2596 | |
111 | 2597 static unsigned int |
0 | 2598 tree_ssa_lim (void) |
2599 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2600 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
|
2601 |
0 | 2602 tree_ssa_lim_initialize (); |
2603 | |
2604 /* Gathers information about memory accesses in the loops. */ | |
2605 analyze_memory_references (); | |
2606 | |
111 | 2607 /* Fills ALWAYS_EXECUTED_IN information for basic blocks. */ |
2608 fill_always_executed_in (); | |
2609 | |
0 | 2610 /* For each statement determine the outermost loop in that it is |
2611 invariant and cost for computing the invariant. */ | |
111 | 2612 invariantness_dom_walker (CDI_DOMINATORS) |
2613 .walk (cfun->cfg->x_entry_block_ptr); | |
0 | 2614 |
2615 /* Execute store motion. Force the necessary invariants to be moved | |
2616 out of the loops as well. */ | |
2617 store_motion (); | |
2618 | |
2619 /* 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
|
2620 todo = move_computations (); |
0 | 2621 |
2622 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
|
2623 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2624 return todo; |
0 | 2625 } |
111 | 2626 |
2627 /* Loop invariant motion pass. */ | |
2628 | |
2629 namespace { | |
2630 | |
2631 const pass_data pass_data_lim = | |
2632 { | |
2633 GIMPLE_PASS, /* type */ | |
2634 "lim", /* name */ | |
2635 OPTGROUP_LOOP, /* optinfo_flags */ | |
2636 TV_LIM, /* tv_id */ | |
2637 PROP_cfg, /* properties_required */ | |
2638 0, /* properties_provided */ | |
2639 0, /* properties_destroyed */ | |
2640 0, /* todo_flags_start */ | |
2641 0, /* todo_flags_finish */ | |
2642 }; | |
2643 | |
2644 class pass_lim : public gimple_opt_pass | |
2645 { | |
2646 public: | |
2647 pass_lim (gcc::context *ctxt) | |
2648 : gimple_opt_pass (pass_data_lim, ctxt) | |
2649 {} | |
2650 | |
2651 /* opt_pass methods: */ | |
2652 opt_pass * clone () { return new pass_lim (m_ctxt); } | |
2653 virtual bool gate (function *) { return flag_tree_loop_im != 0; } | |
2654 virtual unsigned int execute (function *); | |
2655 | |
2656 }; // class pass_lim | |
2657 | |
2658 unsigned int | |
2659 pass_lim::execute (function *fun) | |
2660 { | |
2661 bool in_loop_pipeline = scev_initialized_p (); | |
2662 if (!in_loop_pipeline) | |
2663 loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS); | |
2664 | |
2665 if (number_of_loops (fun) <= 1) | |
2666 return 0; | |
2667 unsigned int todo = tree_ssa_lim (); | |
2668 | |
2669 if (!in_loop_pipeline) | |
2670 loop_optimizer_finalize (); | |
2671 return todo; | |
2672 } | |
2673 | |
2674 } // anon namespace | |
2675 | |
2676 gimple_opt_pass * | |
2677 make_pass_lim (gcc::context *ctxt) | |
2678 { | |
2679 return new pass_lim (ctxt); | |
2680 } | |
2681 | |
2682 |