]>
Commit | Line | Data |
---|---|---|
a7e5372d | 1 | /* Loop invariant motion. |
c75c517d SB |
2 | Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2010 |
3 | Free Software Foundation, Inc. | |
b8698a0f | 4 | |
a7e5372d | 5 | This file is part of GCC. |
b8698a0f | 6 | |
a7e5372d ZD |
7 | GCC is free software; you can redistribute it and/or modify it |
8 | under the terms of the GNU General Public License as published by the | |
9dcd6f09 | 9 | Free Software Foundation; either version 3, or (at your option) any |
a7e5372d | 10 | later version. |
b8698a0f | 11 | |
a7e5372d ZD |
12 | GCC is distributed in the hope that it will be useful, but WITHOUT |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
b8698a0f | 16 | |
a7e5372d | 17 | You should have received a copy of the GNU General Public License |
9dcd6f09 NC |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
a7e5372d ZD |
20 | |
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "tm.h" | |
25 | #include "tree.h" | |
a7e5372d | 26 | #include "tm_p.h" |
a7e5372d | 27 | #include "basic-block.h" |
2eb79bbb | 28 | #include "gimple-pretty-print.h" |
a7e5372d | 29 | #include "tree-flow.h" |
a7e5372d ZD |
30 | #include "cfgloop.h" |
31 | #include "domwalk.h" | |
32 | #include "params.h" | |
33 | #include "tree-pass.h" | |
34 | #include "flags.h" | |
01fd257a | 35 | #include "hashtab.h" |
72425608 ZD |
36 | #include "tree-affine.h" |
37 | #include "pointer-set.h" | |
8ded35f9 | 38 | #include "tree-ssa-propagate.h" |
a7e5372d | 39 | |
f10a6654 ZD |
40 | /* TODO: Support for predicated code motion. I.e. |
41 | ||
42 | while (1) | |
43 | { | |
44 | if (cond) | |
45 | { | |
46 | a = inv; | |
47 | something; | |
48 | } | |
49 | } | |
50 | ||
039496da | 51 | Where COND and INV are invariants, but evaluating INV may trap or be |
f10a6654 ZD |
52 | invalid from some other reason if !COND. This may be transformed to |
53 | ||
54 | if (cond) | |
55 | a = inv; | |
56 | while (1) | |
57 | { | |
58 | if (cond) | |
59 | something; | |
60 | } */ | |
61 | ||
a7e5372d ZD |
62 | /* A type for the list of statements that have to be moved in order to be able |
63 | to hoist an invariant computation. */ | |
64 | ||
65 | struct depend | |
66 | { | |
726a989a | 67 | gimple stmt; |
a7e5372d ZD |
68 | struct depend *next; |
69 | }; | |
70 | ||
a7e5372d ZD |
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 | ||
a7e5372d ZD |
86 | unsigned cost; /* Cost of the computation performed by the |
87 | statement. */ | |
88 | ||
89 | struct depend *depends; /* List of statements that must be also hoisted | |
90 | out of the loop when this statement is | |
91 | hoisted; i.e. those that define the operands | |
92 | of the statement and are inside of the | |
93 | MAX_LOOP loop. */ | |
94 | }; | |
95 | ||
726a989a RB |
96 | /* Maps statements to their lim_aux_data. */ |
97 | ||
98 | static struct pointer_map_t *lim_aux_data_map; | |
a7e5372d | 99 | |
72425608 | 100 | /* Description of a memory reference location. */ |
a7e5372d | 101 | |
72425608 | 102 | typedef struct mem_ref_loc |
a7e5372d ZD |
103 | { |
104 | tree *ref; /* The reference itself. */ | |
726a989a | 105 | gimple stmt; /* The statement in that it occurs. */ |
72425608 ZD |
106 | } *mem_ref_loc_p; |
107 | ||
108 | DEF_VEC_P(mem_ref_loc_p); | |
109 | DEF_VEC_ALLOC_P(mem_ref_loc_p, heap); | |
110 | ||
111 | /* The list of memory reference locations in a loop. */ | |
01fd257a | 112 | |
72425608 ZD |
113 | typedef struct mem_ref_locs |
114 | { | |
115 | VEC (mem_ref_loc_p, heap) *locs; | |
116 | } *mem_ref_locs_p; | |
117 | ||
118 | DEF_VEC_P(mem_ref_locs_p); | |
119 | DEF_VEC_ALLOC_P(mem_ref_locs_p, heap); | |
01fd257a | 120 | |
72425608 ZD |
121 | /* Description of a memory reference. */ |
122 | ||
123 | typedef struct mem_ref | |
01fd257a ZD |
124 | { |
125 | tree mem; /* The memory itself. */ | |
72425608 ZD |
126 | unsigned id; /* ID assigned to the memory reference |
127 | (its index in memory_accesses.refs_list) */ | |
01fd257a | 128 | hashval_t hash; /* Its hash value. */ |
fa10beec | 129 | bitmap stored; /* The set of loops in that this memory location |
72425608 ZD |
130 | is stored to. */ |
131 | VEC (mem_ref_locs_p, heap) *accesses_in_loop; | |
132 | /* The locations of the accesses. Vector | |
133 | indexed by the loop number. */ | |
72425608 ZD |
134 | |
135 | /* The following sets are computed on demand. We keep both set and | |
136 | its complement, so that we know whether the information was | |
137 | already computed or not. */ | |
138 | bitmap indep_loop; /* The set of loops in that the memory | |
139 | reference is independent, meaning: | |
140 | If it is stored in the loop, this store | |
141 | is independent on all other loads and | |
142 | stores. | |
143 | If it is only loaded, then it is independent | |
144 | on all stores in the loop. */ | |
145 | bitmap dep_loop; /* The complement of INDEP_LOOP. */ | |
146 | ||
147 | bitmap indep_ref; /* The set of memory references on that | |
148 | this reference is independent. */ | |
19c0d7df | 149 | bitmap dep_ref; /* The complement of INDEP_REF. */ |
72425608 ZD |
150 | } *mem_ref_p; |
151 | ||
152 | DEF_VEC_P(mem_ref_p); | |
153 | DEF_VEC_ALLOC_P(mem_ref_p, heap); | |
154 | ||
155 | DEF_VEC_P(bitmap); | |
156 | DEF_VEC_ALLOC_P(bitmap, heap); | |
157 | ||
158 | DEF_VEC_P(htab_t); | |
159 | DEF_VEC_ALLOC_P(htab_t, heap); | |
160 | ||
161 | /* Description of memory accesses in loops. */ | |
162 | ||
163 | static struct | |
164 | { | |
165 | /* The hash table of memory references accessed in loops. */ | |
166 | htab_t refs; | |
167 | ||
168 | /* The list of memory references. */ | |
169 | VEC (mem_ref_p, heap) *refs_list; | |
170 | ||
171 | /* The set of memory references accessed in each loop. */ | |
172 | VEC (bitmap, heap) *refs_in_loop; | |
173 | ||
174 | /* The set of memory references accessed in each loop, including | |
175 | subloops. */ | |
176 | VEC (bitmap, heap) *all_refs_in_loop; | |
177 | ||
546d314c RG |
178 | /* The set of memory references stored in each loop, including |
179 | subloops. */ | |
180 | VEC (bitmap, heap) *all_refs_stored_in_loop; | |
72425608 ZD |
181 | |
182 | /* Cache for expanding memory addresses. */ | |
183 | struct pointer_map_t *ttae_cache; | |
184 | } memory_accesses; | |
185 | ||
186 | static bool ref_indep_loop_p (struct loop *, mem_ref_p); | |
a7e5372d ZD |
187 | |
188 | /* Minimum cost of an expensive expression. */ | |
189 | #define LIM_EXPENSIVE ((unsigned) PARAM_VALUE (PARAM_LIM_EXPENSIVE)) | |
190 | ||
8a519095 | 191 | /* The outermost loop for which execution of the header guarantees that the |
a7e5372d ZD |
192 | block will be executed. */ |
193 | #define ALWAYS_EXECUTED_IN(BB) ((struct loop *) (BB)->aux) | |
8a519095 | 194 | #define SET_ALWAYS_EXECUTED_IN(BB, VAL) ((BB)->aux = (void *) (VAL)) |
a7e5372d | 195 | |
546d314c RG |
196 | /* Whether the reference was analyzable. */ |
197 | #define MEM_ANALYZABLE(REF) ((REF)->mem != error_mark_node) | |
198 | ||
726a989a RB |
199 | static struct lim_aux_data * |
200 | init_lim_data (gimple stmt) | |
201 | { | |
202 | void **p = pointer_map_insert (lim_aux_data_map, stmt); | |
203 | ||
204 | *p = XCNEW (struct lim_aux_data); | |
205 | return (struct lim_aux_data *) *p; | |
206 | } | |
207 | ||
208 | static struct lim_aux_data * | |
209 | get_lim_data (gimple stmt) | |
210 | { | |
211 | void **p = pointer_map_contains (lim_aux_data_map, stmt); | |
212 | if (!p) | |
213 | return NULL; | |
214 | ||
215 | return (struct lim_aux_data *) *p; | |
216 | } | |
217 | ||
218 | /* Releases the memory occupied by DATA. */ | |
219 | ||
220 | static void | |
221 | free_lim_aux_data (struct lim_aux_data *data) | |
222 | { | |
223 | struct depend *dep, *next; | |
224 | ||
225 | for (dep = data->depends; dep; dep = next) | |
226 | { | |
227 | next = dep->next; | |
228 | free (dep); | |
229 | } | |
230 | free (data); | |
231 | } | |
232 | ||
233 | static void | |
234 | clear_lim_data (gimple stmt) | |
235 | { | |
236 | void **p = pointer_map_contains (lim_aux_data_map, stmt); | |
237 | if (!p) | |
238 | return; | |
239 | ||
240 | free_lim_aux_data ((struct lim_aux_data *) *p); | |
241 | *p = NULL; | |
242 | } | |
243 | ||
a7e5372d ZD |
244 | /* Calls CBCK for each index in memory reference ADDR_P. There are two |
245 | kinds situations handled; in each of these cases, the memory reference | |
246 | and DATA are passed to the callback: | |
b8698a0f | 247 | |
a7e5372d ZD |
248 | Access to an array: ARRAY_{RANGE_}REF (base, index). In this case we also |
249 | pass the pointer to the index to the callback. | |
250 | ||
251 | Pointer dereference: INDIRECT_REF (addr). In this case we also pass the | |
252 | pointer to addr to the callback. | |
b8698a0f | 253 | |
a7e5372d ZD |
254 | If the callback returns false, the whole search stops and false is returned. |
255 | Otherwise the function returns true after traversing through the whole | |
256 | reference *ADDR_P. */ | |
257 | ||
258 | bool | |
259 | for_each_index (tree *addr_p, bool (*cbck) (tree, tree *, void *), void *data) | |
260 | { | |
be35cf60 | 261 | tree *nxt, *idx; |
a7e5372d ZD |
262 | |
263 | for (; ; addr_p = nxt) | |
264 | { | |
265 | switch (TREE_CODE (*addr_p)) | |
266 | { | |
267 | case SSA_NAME: | |
268 | return cbck (*addr_p, addr_p, data); | |
269 | ||
70f34814 | 270 | case MEM_REF: |
a7e5372d ZD |
271 | nxt = &TREE_OPERAND (*addr_p, 0); |
272 | return cbck (*addr_p, nxt, data); | |
273 | ||
274 | case BIT_FIELD_REF: | |
a7e5372d | 275 | case VIEW_CONVERT_EXPR: |
8b11a64c ZD |
276 | case REALPART_EXPR: |
277 | case IMAGPART_EXPR: | |
a7e5372d ZD |
278 | nxt = &TREE_OPERAND (*addr_p, 0); |
279 | break; | |
280 | ||
be35cf60 ZD |
281 | case COMPONENT_REF: |
282 | /* If the component has varying offset, it behaves like index | |
283 | as well. */ | |
284 | idx = &TREE_OPERAND (*addr_p, 2); | |
285 | if (*idx | |
286 | && !cbck (*addr_p, idx, data)) | |
287 | return false; | |
288 | ||
289 | nxt = &TREE_OPERAND (*addr_p, 0); | |
290 | break; | |
291 | ||
a7e5372d | 292 | case ARRAY_REF: |
61c25908 | 293 | case ARRAY_RANGE_REF: |
a7e5372d ZD |
294 | nxt = &TREE_OPERAND (*addr_p, 0); |
295 | if (!cbck (*addr_p, &TREE_OPERAND (*addr_p, 1), data)) | |
296 | return false; | |
297 | break; | |
298 | ||
299 | case VAR_DECL: | |
300 | case PARM_DECL: | |
301 | case STRING_CST: | |
302 | case RESULT_DECL: | |
60407f7a | 303 | case VECTOR_CST: |
33674347 | 304 | case COMPLEX_CST: |
e2889823 PB |
305 | case INTEGER_CST: |
306 | case REAL_CST: | |
325217ed | 307 | case FIXED_CST: |
bb0c55f6 | 308 | case CONSTRUCTOR: |
a7e5372d ZD |
309 | return true; |
310 | ||
3d45dd59 RG |
311 | case ADDR_EXPR: |
312 | gcc_assert (is_gimple_min_invariant (*addr_p)); | |
313 | return true; | |
314 | ||
ac182688 ZD |
315 | case TARGET_MEM_REF: |
316 | idx = &TMR_BASE (*addr_p); | |
317 | if (*idx | |
318 | && !cbck (*addr_p, idx, data)) | |
319 | return false; | |
320 | idx = &TMR_INDEX (*addr_p); | |
4d948885 RG |
321 | if (*idx |
322 | && !cbck (*addr_p, idx, data)) | |
323 | return false; | |
324 | idx = &TMR_INDEX2 (*addr_p); | |
ac182688 ZD |
325 | if (*idx |
326 | && !cbck (*addr_p, idx, data)) | |
327 | return false; | |
328 | return true; | |
329 | ||
a7e5372d | 330 | default: |
1e128c5f | 331 | gcc_unreachable (); |
a7e5372d ZD |
332 | } |
333 | } | |
334 | } | |
335 | ||
336 | /* If it is possible to hoist the statement STMT unconditionally, | |
337 | returns MOVE_POSSIBLE. | |
338 | If it is possible to hoist the statement STMT, but we must avoid making | |
339 | it executed if it would not be executed in the original program (e.g. | |
340 | because it may trap), return MOVE_PRESERVE_EXECUTION. | |
341 | Otherwise return MOVE_IMPOSSIBLE. */ | |
342 | ||
40923b20 | 343 | enum move_pos |
726a989a | 344 | movement_possibility (gimple stmt) |
a7e5372d | 345 | { |
726a989a RB |
346 | tree lhs; |
347 | enum move_pos ret = MOVE_POSSIBLE; | |
a7e5372d ZD |
348 | |
349 | if (flag_unswitch_loops | |
726a989a | 350 | && gimple_code (stmt) == GIMPLE_COND) |
a7e5372d ZD |
351 | { |
352 | /* If we perform unswitching, force the operands of the invariant | |
353 | condition to be moved out of the loop. */ | |
a7e5372d ZD |
354 | return MOVE_POSSIBLE; |
355 | } | |
356 | ||
e3bdfed6 RG |
357 | if (gimple_code (stmt) == GIMPLE_PHI |
358 | && gimple_phi_num_args (stmt) <= 2 | |
359 | && is_gimple_reg (gimple_phi_result (stmt)) | |
360 | && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_phi_result (stmt))) | |
361 | return MOVE_POSSIBLE; | |
362 | ||
726a989a | 363 | if (gimple_get_lhs (stmt) == NULL_TREE) |
a7e5372d ZD |
364 | return MOVE_IMPOSSIBLE; |
365 | ||
5006671f | 366 | if (gimple_vdef (stmt)) |
72425608 ZD |
367 | return MOVE_IMPOSSIBLE; |
368 | ||
726a989a RB |
369 | if (stmt_ends_bb_p (stmt) |
370 | || gimple_has_volatile_ops (stmt) | |
371 | || gimple_has_side_effects (stmt) | |
372 | || stmt_could_throw_p (stmt)) | |
a7e5372d ZD |
373 | return MOVE_IMPOSSIBLE; |
374 | ||
726a989a | 375 | if (is_gimple_call (stmt)) |
f10a6654 ZD |
376 | { |
377 | /* While pure or const call is guaranteed to have no side effects, we | |
378 | cannot move it arbitrarily. Consider code like | |
379 | ||
380 | char *s = something (); | |
381 | ||
382 | while (1) | |
383 | { | |
384 | if (s) | |
385 | t = strlen (s); | |
386 | else | |
387 | t = 0; | |
388 | } | |
389 | ||
390 | Here the strlen call cannot be moved out of the loop, even though | |
391 | s is invariant. In addition to possibly creating a call with | |
392 | invalid arguments, moving out a function call that is not executed | |
393 | may cause performance regressions in case the call is costly and | |
394 | not executed at all. */ | |
726a989a RB |
395 | ret = MOVE_PRESERVE_EXECUTION; |
396 | lhs = gimple_call_lhs (stmt); | |
f10a6654 | 397 | } |
726a989a RB |
398 | else if (is_gimple_assign (stmt)) |
399 | lhs = gimple_assign_lhs (stmt); | |
400 | else | |
401 | return MOVE_IMPOSSIBLE; | |
402 | ||
403 | if (TREE_CODE (lhs) == SSA_NAME | |
404 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)) | |
405 | return MOVE_IMPOSSIBLE; | |
406 | ||
407 | if (TREE_CODE (lhs) != SSA_NAME | |
408 | || gimple_could_trap_p (stmt)) | |
409 | return MOVE_PRESERVE_EXECUTION; | |
410 | ||
19c0d7df AH |
411 | /* Non local loads in a transaction cannot be hoisted out. Well, |
412 | unless the load happens on every path out of the loop, but we | |
413 | don't take this into account yet. */ | |
414 | if (flag_tm | |
415 | && gimple_in_transaction (stmt) | |
416 | && gimple_assign_single_p (stmt)) | |
417 | { | |
418 | tree rhs = gimple_assign_rhs1 (stmt); | |
419 | if (DECL_P (rhs) && is_global_var (rhs)) | |
420 | { | |
421 | if (dump_file) | |
422 | { | |
423 | fprintf (dump_file, "Cannot hoist conditional load of "); | |
424 | print_generic_expr (dump_file, rhs, TDF_SLIM); | |
425 | fprintf (dump_file, " because it is in a transaction.\n"); | |
426 | } | |
427 | return MOVE_IMPOSSIBLE; | |
428 | } | |
429 | } | |
430 | ||
726a989a | 431 | return ret; |
a7e5372d ZD |
432 | } |
433 | ||
434 | /* Suppose that operand DEF is used inside the LOOP. Returns the outermost | |
2a7e31df | 435 | loop to that we could move the expression using DEF if it did not have |
a7e5372d ZD |
436 | other operands, i.e. the outermost loop enclosing LOOP in that the value |
437 | of DEF is invariant. */ | |
438 | ||
439 | static struct loop * | |
440 | outermost_invariant_loop (tree def, struct loop *loop) | |
441 | { | |
726a989a | 442 | gimple def_stmt; |
a7e5372d ZD |
443 | basic_block def_bb; |
444 | struct loop *max_loop; | |
726a989a | 445 | struct lim_aux_data *lim_data; |
a7e5372d | 446 | |
726a989a | 447 | if (!def) |
a7e5372d ZD |
448 | return superloop_at_depth (loop, 1); |
449 | ||
726a989a RB |
450 | if (TREE_CODE (def) != SSA_NAME) |
451 | { | |
452 | gcc_assert (is_gimple_min_invariant (def)); | |
453 | return superloop_at_depth (loop, 1); | |
454 | } | |
455 | ||
a7e5372d | 456 | def_stmt = SSA_NAME_DEF_STMT (def); |
726a989a | 457 | def_bb = gimple_bb (def_stmt); |
a7e5372d ZD |
458 | if (!def_bb) |
459 | return superloop_at_depth (loop, 1); | |
460 | ||
461 | max_loop = find_common_loop (loop, def_bb->loop_father); | |
462 | ||
726a989a RB |
463 | lim_data = get_lim_data (def_stmt); |
464 | if (lim_data != NULL && lim_data->max_loop != NULL) | |
a7e5372d | 465 | max_loop = find_common_loop (max_loop, |
726a989a | 466 | loop_outer (lim_data->max_loop)); |
a7e5372d ZD |
467 | if (max_loop == loop) |
468 | return NULL; | |
9ba025a2 | 469 | max_loop = superloop_at_depth (loop, loop_depth (max_loop) + 1); |
a7e5372d ZD |
470 | |
471 | return max_loop; | |
472 | } | |
473 | ||
a7e5372d ZD |
474 | /* DATA is a structure containing information associated with a statement |
475 | inside LOOP. DEF is one of the operands of this statement. | |
b8698a0f | 476 | |
a7e5372d ZD |
477 | Find the outermost loop enclosing LOOP in that value of DEF is invariant |
478 | and record this in DATA->max_loop field. If DEF itself is defined inside | |
479 | this loop as well (i.e. we need to hoist it out of the loop if we want | |
480 | to hoist the statement represented by DATA), record the statement in that | |
481 | DEF is defined to the DATA->depends list. Additionally if ADD_COST is true, | |
482 | add the cost of the computation of DEF to the DATA->cost. | |
b8698a0f | 483 | |
a7e5372d ZD |
484 | If DEF is not invariant in LOOP, return false. Otherwise return TRUE. */ |
485 | ||
486 | static bool | |
487 | add_dependency (tree def, struct lim_aux_data *data, struct loop *loop, | |
488 | bool add_cost) | |
489 | { | |
726a989a RB |
490 | gimple def_stmt = SSA_NAME_DEF_STMT (def); |
491 | basic_block def_bb = gimple_bb (def_stmt); | |
a7e5372d ZD |
492 | struct loop *max_loop; |
493 | struct depend *dep; | |
726a989a | 494 | struct lim_aux_data *def_data; |
a7e5372d ZD |
495 | |
496 | if (!def_bb) | |
497 | return true; | |
498 | ||
499 | max_loop = outermost_invariant_loop (def, loop); | |
500 | if (!max_loop) | |
501 | return false; | |
502 | ||
503 | if (flow_loop_nested_p (data->max_loop, max_loop)) | |
504 | data->max_loop = max_loop; | |
505 | ||
726a989a RB |
506 | def_data = get_lim_data (def_stmt); |
507 | if (!def_data) | |
a7e5372d ZD |
508 | return true; |
509 | ||
510 | if (add_cost | |
511 | /* Only add the cost if the statement defining DEF is inside LOOP, | |
512 | i.e. if it is likely that by moving the invariants dependent | |
513 | on it, we will be able to avoid creating a new register for | |
514 | it (since it will be only used in these dependent invariants). */ | |
515 | && def_bb->loop_father == loop) | |
726a989a | 516 | data->cost += def_data->cost; |
a7e5372d | 517 | |
5ed6ace5 | 518 | dep = XNEW (struct depend); |
a7e5372d ZD |
519 | dep->stmt = def_stmt; |
520 | dep->next = data->depends; | |
521 | data->depends = dep; | |
522 | ||
523 | return true; | |
524 | } | |
525 | ||
546d314c RG |
526 | /* Returns an estimate for a cost of statement STMT. The values here |
527 | are just ad-hoc constants, similar to costs for inlining. */ | |
a7e5372d ZD |
528 | |
529 | static unsigned | |
726a989a | 530 | stmt_cost (gimple stmt) |
a7e5372d | 531 | { |
a7e5372d | 532 | /* Always try to create possibilities for unswitching. */ |
e3bdfed6 RG |
533 | if (gimple_code (stmt) == GIMPLE_COND |
534 | || gimple_code (stmt) == GIMPLE_PHI) | |
a7e5372d ZD |
535 | return LIM_EXPENSIVE; |
536 | ||
546d314c | 537 | /* We should be hoisting calls if possible. */ |
726a989a | 538 | if (is_gimple_call (stmt)) |
a7e5372d | 539 | { |
546d314c | 540 | tree fndecl; |
a7e5372d ZD |
541 | |
542 | /* Unless the call is a builtin_constant_p; this always folds to a | |
543 | constant, so moving it is useless. */ | |
726a989a RB |
544 | fndecl = gimple_call_fndecl (stmt); |
545 | if (fndecl | |
546 | && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL | |
547 | && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CONSTANT_P) | |
a7e5372d ZD |
548 | return 0; |
549 | ||
546d314c | 550 | return LIM_EXPENSIVE; |
726a989a RB |
551 | } |
552 | ||
546d314c RG |
553 | /* Hoisting memory references out should almost surely be a win. */ |
554 | if (gimple_references_memory_p (stmt)) | |
555 | return LIM_EXPENSIVE; | |
556 | ||
726a989a | 557 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
546d314c | 558 | return 1; |
a7e5372d | 559 | |
726a989a RB |
560 | switch (gimple_assign_rhs_code (stmt)) |
561 | { | |
a7e5372d | 562 | case MULT_EXPR: |
67af611e RG |
563 | case WIDEN_MULT_EXPR: |
564 | case WIDEN_MULT_PLUS_EXPR: | |
565 | case WIDEN_MULT_MINUS_EXPR: | |
566 | case DOT_PROD_EXPR: | |
567 | case FMA_EXPR: | |
a7e5372d ZD |
568 | case TRUNC_DIV_EXPR: |
569 | case CEIL_DIV_EXPR: | |
570 | case FLOOR_DIV_EXPR: | |
571 | case ROUND_DIV_EXPR: | |
572 | case EXACT_DIV_EXPR: | |
573 | case CEIL_MOD_EXPR: | |
574 | case FLOOR_MOD_EXPR: | |
575 | case ROUND_MOD_EXPR: | |
576 | case TRUNC_MOD_EXPR: | |
b4852851 | 577 | case RDIV_EXPR: |
a7e5372d | 578 | /* Division and multiplication are usually expensive. */ |
546d314c | 579 | return LIM_EXPENSIVE; |
a7e5372d | 580 | |
e0a60731 RG |
581 | case LSHIFT_EXPR: |
582 | case RSHIFT_EXPR: | |
67af611e RG |
583 | case WIDEN_LSHIFT_EXPR: |
584 | case LROTATE_EXPR: | |
585 | case RROTATE_EXPR: | |
546d314c RG |
586 | /* Shifts and rotates are usually expensive. */ |
587 | return LIM_EXPENSIVE; | |
588 | ||
589 | case CONSTRUCTOR: | |
590 | /* Make vector construction cost proportional to the number | |
591 | of elements. */ | |
592 | return CONSTRUCTOR_NELTS (gimple_assign_rhs1 (stmt)); | |
593 | ||
594 | case SSA_NAME: | |
595 | case PAREN_EXPR: | |
596 | /* Whether or not something is wrapped inside a PAREN_EXPR | |
597 | should not change move cost. Nor should an intermediate | |
598 | unpropagated SSA name copy. */ | |
599 | return 0; | |
e0a60731 | 600 | |
a7e5372d | 601 | default: |
546d314c | 602 | return 1; |
a7e5372d | 603 | } |
a7e5372d ZD |
604 | } |
605 | ||
72425608 ZD |
606 | /* Finds the outermost loop between OUTER and LOOP in that the memory reference |
607 | REF is independent. If REF is not independent in LOOP, NULL is returned | |
608 | instead. */ | |
609 | ||
610 | static struct loop * | |
611 | outermost_indep_loop (struct loop *outer, struct loop *loop, mem_ref_p ref) | |
612 | { | |
613 | struct loop *aloop; | |
614 | ||
615 | if (bitmap_bit_p (ref->stored, loop->num)) | |
616 | return NULL; | |
617 | ||
618 | for (aloop = outer; | |
619 | aloop != loop; | |
620 | aloop = superloop_at_depth (loop, loop_depth (aloop) + 1)) | |
621 | if (!bitmap_bit_p (ref->stored, aloop->num) | |
622 | && ref_indep_loop_p (aloop, ref)) | |
623 | return aloop; | |
624 | ||
625 | if (ref_indep_loop_p (loop, ref)) | |
626 | return loop; | |
627 | else | |
628 | return NULL; | |
629 | } | |
630 | ||
631 | /* If there is a simple load or store to a memory reference in STMT, returns | |
fa10beec | 632 | the location of the memory reference, and sets IS_STORE according to whether |
72425608 ZD |
633 | it is a store or load. Otherwise, returns NULL. */ |
634 | ||
635 | static tree * | |
726a989a | 636 | simple_mem_ref_in_stmt (gimple stmt, bool *is_store) |
72425608 | 637 | { |
726a989a RB |
638 | tree *lhs; |
639 | enum tree_code code; | |
72425608 ZD |
640 | |
641 | /* Recognize MEM = (SSA_NAME | invariant) and SSA_NAME = MEM patterns. */ | |
726a989a | 642 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
72425608 ZD |
643 | return NULL; |
644 | ||
726a989a RB |
645 | code = gimple_assign_rhs_code (stmt); |
646 | ||
647 | lhs = gimple_assign_lhs_ptr (stmt); | |
72425608 ZD |
648 | |
649 | if (TREE_CODE (*lhs) == SSA_NAME) | |
650 | { | |
726a989a RB |
651 | if (get_gimple_rhs_class (code) != GIMPLE_SINGLE_RHS |
652 | || !is_gimple_addressable (gimple_assign_rhs1 (stmt))) | |
72425608 ZD |
653 | return NULL; |
654 | ||
655 | *is_store = false; | |
726a989a | 656 | return gimple_assign_rhs1_ptr (stmt); |
72425608 | 657 | } |
726a989a RB |
658 | else if (code == SSA_NAME |
659 | || (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS | |
660 | && is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))) | |
72425608 ZD |
661 | { |
662 | *is_store = true; | |
663 | return lhs; | |
664 | } | |
665 | else | |
666 | return NULL; | |
667 | } | |
668 | ||
669 | /* Returns the memory reference contained in STMT. */ | |
670 | ||
671 | static mem_ref_p | |
726a989a | 672 | mem_ref_in_stmt (gimple stmt) |
72425608 ZD |
673 | { |
674 | bool store; | |
675 | tree *mem = simple_mem_ref_in_stmt (stmt, &store); | |
676 | hashval_t hash; | |
677 | mem_ref_p ref; | |
678 | ||
679 | if (!mem) | |
680 | return NULL; | |
681 | gcc_assert (!store); | |
682 | ||
683 | hash = iterative_hash_expr (*mem, 0); | |
3d9a9f94 | 684 | ref = (mem_ref_p) htab_find_with_hash (memory_accesses.refs, *mem, hash); |
72425608 ZD |
685 | |
686 | gcc_assert (ref != NULL); | |
687 | return ref; | |
688 | } | |
689 | ||
e3bdfed6 RG |
690 | /* From a controlling predicate in DOM determine the arguments from |
691 | the PHI node PHI that are chosen if the predicate evaluates to | |
692 | true and false and store them to *TRUE_ARG_P and *FALSE_ARG_P if | |
693 | they are non-NULL. Returns true if the arguments can be determined, | |
694 | else return false. */ | |
695 | ||
696 | static bool | |
697 | extract_true_false_args_from_phi (basic_block dom, gimple phi, | |
698 | tree *true_arg_p, tree *false_arg_p) | |
699 | { | |
700 | basic_block bb = gimple_bb (phi); | |
701 | edge true_edge, false_edge, tem; | |
702 | tree arg0 = NULL_TREE, arg1 = NULL_TREE; | |
703 | ||
704 | /* We have to verify that one edge into the PHI node is dominated | |
705 | by the true edge of the predicate block and the other edge | |
706 | dominated by the false edge. This ensures that the PHI argument | |
707 | we are going to take is completely determined by the path we | |
12d80acc RG |
708 | take from the predicate block. |
709 | We can only use BB dominance checks below if the destination of | |
710 | the true/false edges are dominated by their edge, thus only | |
711 | have a single predecessor. */ | |
e3bdfed6 RG |
712 | extract_true_false_edges_from_block (dom, &true_edge, &false_edge); |
713 | tem = EDGE_PRED (bb, 0); | |
714 | if (tem == true_edge | |
12d80acc RG |
715 | || (single_pred_p (true_edge->dest) |
716 | && (tem->src == true_edge->dest | |
717 | || dominated_by_p (CDI_DOMINATORS, | |
718 | tem->src, true_edge->dest)))) | |
e3bdfed6 RG |
719 | arg0 = PHI_ARG_DEF (phi, tem->dest_idx); |
720 | else if (tem == false_edge | |
12d80acc RG |
721 | || (single_pred_p (false_edge->dest) |
722 | && (tem->src == false_edge->dest | |
723 | || dominated_by_p (CDI_DOMINATORS, | |
724 | tem->src, false_edge->dest)))) | |
e3bdfed6 RG |
725 | arg1 = PHI_ARG_DEF (phi, tem->dest_idx); |
726 | else | |
727 | return false; | |
728 | tem = EDGE_PRED (bb, 1); | |
729 | if (tem == true_edge | |
12d80acc RG |
730 | || (single_pred_p (true_edge->dest) |
731 | && (tem->src == true_edge->dest | |
732 | || dominated_by_p (CDI_DOMINATORS, | |
733 | tem->src, true_edge->dest)))) | |
e3bdfed6 RG |
734 | arg0 = PHI_ARG_DEF (phi, tem->dest_idx); |
735 | else if (tem == false_edge | |
12d80acc RG |
736 | || (single_pred_p (false_edge->dest) |
737 | && (tem->src == false_edge->dest | |
738 | || dominated_by_p (CDI_DOMINATORS, | |
739 | tem->src, false_edge->dest)))) | |
e3bdfed6 RG |
740 | arg1 = PHI_ARG_DEF (phi, tem->dest_idx); |
741 | else | |
742 | return false; | |
743 | if (!arg0 || !arg1) | |
744 | return false; | |
745 | ||
746 | if (true_arg_p) | |
747 | *true_arg_p = arg0; | |
748 | if (false_arg_p) | |
749 | *false_arg_p = arg1; | |
750 | ||
751 | return true; | |
752 | } | |
753 | ||
a7e5372d ZD |
754 | /* Determine the outermost loop to that it is possible to hoist a statement |
755 | STMT and store it to LIM_DATA (STMT)->max_loop. To do this we determine | |
756 | the outermost loop in that the value computed by STMT is invariant. | |
757 | If MUST_PRESERVE_EXEC is true, additionally choose such a loop that | |
758 | we preserve the fact whether STMT is executed. It also fills other related | |
759 | information to LIM_DATA (STMT). | |
b8698a0f | 760 | |
a7e5372d ZD |
761 | The function returns false if STMT cannot be hoisted outside of the loop it |
762 | is defined in, and true otherwise. */ | |
763 | ||
764 | static bool | |
726a989a | 765 | determine_max_movement (gimple stmt, bool must_preserve_exec) |
a7e5372d | 766 | { |
726a989a | 767 | basic_block bb = gimple_bb (stmt); |
a7e5372d ZD |
768 | struct loop *loop = bb->loop_father; |
769 | struct loop *level; | |
726a989a | 770 | struct lim_aux_data *lim_data = get_lim_data (stmt); |
4c124b4c AM |
771 | tree val; |
772 | ssa_op_iter iter; | |
b8698a0f | 773 | |
a7e5372d ZD |
774 | if (must_preserve_exec) |
775 | level = ALWAYS_EXECUTED_IN (bb); | |
776 | else | |
777 | level = superloop_at_depth (loop, 1); | |
778 | lim_data->max_loop = level; | |
779 | ||
e3bdfed6 RG |
780 | if (gimple_code (stmt) == GIMPLE_PHI) |
781 | { | |
782 | use_operand_p use_p; | |
783 | unsigned min_cost = UINT_MAX; | |
784 | unsigned total_cost = 0; | |
785 | struct lim_aux_data *def_data; | |
786 | ||
787 | /* We will end up promoting dependencies to be unconditionally | |
788 | evaluated. For this reason the PHI cost (and thus the | |
789 | cost we remove from the loop by doing the invariant motion) | |
790 | is that of the cheapest PHI argument dependency chain. */ | |
791 | FOR_EACH_PHI_ARG (use_p, stmt, iter, SSA_OP_USE) | |
792 | { | |
793 | val = USE_FROM_PTR (use_p); | |
794 | if (TREE_CODE (val) != SSA_NAME) | |
795 | continue; | |
796 | if (!add_dependency (val, lim_data, loop, false)) | |
797 | return false; | |
798 | def_data = get_lim_data (SSA_NAME_DEF_STMT (val)); | |
799 | if (def_data) | |
800 | { | |
801 | min_cost = MIN (min_cost, def_data->cost); | |
802 | total_cost += def_data->cost; | |
803 | } | |
804 | } | |
805 | ||
806 | lim_data->cost += min_cost; | |
807 | ||
808 | if (gimple_phi_num_args (stmt) > 1) | |
809 | { | |
810 | basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb); | |
811 | gimple cond; | |
812 | if (gsi_end_p (gsi_last_bb (dom))) | |
813 | return false; | |
814 | cond = gsi_stmt (gsi_last_bb (dom)); | |
815 | if (gimple_code (cond) != GIMPLE_COND) | |
816 | return false; | |
817 | /* Verify that this is an extended form of a diamond and | |
818 | the PHI arguments are completely controlled by the | |
819 | predicate in DOM. */ | |
820 | if (!extract_true_false_args_from_phi (dom, stmt, NULL, NULL)) | |
821 | return false; | |
822 | ||
823 | /* Fold in dependencies and cost of the condition. */ | |
824 | FOR_EACH_SSA_TREE_OPERAND (val, cond, iter, SSA_OP_USE) | |
825 | { | |
826 | if (!add_dependency (val, lim_data, loop, false)) | |
827 | return false; | |
828 | def_data = get_lim_data (SSA_NAME_DEF_STMT (val)); | |
829 | if (def_data) | |
830 | total_cost += def_data->cost; | |
831 | } | |
832 | ||
833 | /* We want to avoid unconditionally executing very expensive | |
834 | operations. As costs for our dependencies cannot be | |
835 | negative just claim we are not invariand for this case. | |
836 | We also are not sure whether the control-flow inside the | |
837 | loop will vanish. */ | |
838 | if (total_cost - min_cost >= 2 * LIM_EXPENSIVE | |
839 | && !(min_cost != 0 | |
840 | && total_cost / min_cost <= 2)) | |
841 | return false; | |
842 | ||
843 | /* Assume that the control-flow in the loop will vanish. | |
844 | ??? We should verify this and not artificially increase | |
845 | the cost if that is not the case. */ | |
846 | lim_data->cost += stmt_cost (stmt); | |
847 | } | |
848 | ||
849 | return true; | |
850 | } | |
851 | else | |
852 | FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_USE) | |
853 | if (!add_dependency (val, lim_data, loop, true)) | |
854 | return false; | |
a7e5372d | 855 | |
5006671f | 856 | if (gimple_vuse (stmt)) |
72425608 ZD |
857 | { |
858 | mem_ref_p ref = mem_ref_in_stmt (stmt); | |
859 | ||
860 | if (ref) | |
861 | { | |
862 | lim_data->max_loop | |
863 | = outermost_indep_loop (lim_data->max_loop, loop, ref); | |
864 | if (!lim_data->max_loop) | |
865 | return false; | |
866 | } | |
867 | else | |
868 | { | |
5006671f | 869 | if ((val = gimple_vuse (stmt)) != NULL_TREE) |
72425608 ZD |
870 | { |
871 | if (!add_dependency (val, lim_data, loop, false)) | |
872 | return false; | |
873 | } | |
874 | } | |
875 | } | |
a7e5372d ZD |
876 | |
877 | lim_data->cost += stmt_cost (stmt); | |
878 | ||
879 | return true; | |
880 | } | |
881 | ||
882 | /* Suppose that some statement in ORIG_LOOP is hoisted to the loop LEVEL, | |
883 | and that one of the operands of this statement is computed by STMT. | |
884 | Ensure that STMT (together with all the statements that define its | |
885 | operands) is hoisted at least out of the loop LEVEL. */ | |
886 | ||
887 | static void | |
726a989a | 888 | set_level (gimple stmt, struct loop *orig_loop, struct loop *level) |
a7e5372d | 889 | { |
726a989a | 890 | struct loop *stmt_loop = gimple_bb (stmt)->loop_father; |
a7e5372d | 891 | struct depend *dep; |
726a989a | 892 | struct lim_aux_data *lim_data; |
a7e5372d ZD |
893 | |
894 | stmt_loop = find_common_loop (orig_loop, stmt_loop); | |
726a989a RB |
895 | lim_data = get_lim_data (stmt); |
896 | if (lim_data != NULL && lim_data->tgt_loop != NULL) | |
a7e5372d | 897 | stmt_loop = find_common_loop (stmt_loop, |
726a989a | 898 | loop_outer (lim_data->tgt_loop)); |
a7e5372d ZD |
899 | if (flow_loop_nested_p (stmt_loop, level)) |
900 | return; | |
901 | ||
726a989a RB |
902 | gcc_assert (level == lim_data->max_loop |
903 | || flow_loop_nested_p (lim_data->max_loop, level)); | |
a7e5372d | 904 | |
726a989a RB |
905 | lim_data->tgt_loop = level; |
906 | for (dep = lim_data->depends; dep; dep = dep->next) | |
a7e5372d ZD |
907 | set_level (dep->stmt, orig_loop, level); |
908 | } | |
909 | ||
910 | /* Determines an outermost loop from that we want to hoist the statement STMT. | |
911 | For now we chose the outermost possible loop. TODO -- use profiling | |
912 | information to set it more sanely. */ | |
913 | ||
914 | static void | |
726a989a | 915 | set_profitable_level (gimple stmt) |
a7e5372d | 916 | { |
726a989a | 917 | set_level (stmt, gimple_bb (stmt)->loop_father, get_lim_data (stmt)->max_loop); |
a7e5372d ZD |
918 | } |
919 | ||
726a989a | 920 | /* Returns true if STMT is a call that has side effects. */ |
a7e5372d ZD |
921 | |
922 | static bool | |
726a989a | 923 | nonpure_call_p (gimple stmt) |
a7e5372d | 924 | { |
726a989a | 925 | if (gimple_code (stmt) != GIMPLE_CALL) |
a7e5372d ZD |
926 | return false; |
927 | ||
726a989a | 928 | return gimple_has_side_effects (stmt); |
a7e5372d ZD |
929 | } |
930 | ||
e0a60731 RG |
931 | /* Rewrite a/b to a*(1/b). Return the invariant stmt to process. */ |
932 | ||
726a989a RB |
933 | static gimple |
934 | rewrite_reciprocal (gimple_stmt_iterator *bsi) | |
e0a60731 | 935 | { |
726a989a RB |
936 | gimple stmt, stmt1, stmt2; |
937 | tree var, name, lhs, type; | |
f50d67f6 | 938 | tree real_one; |
0ca5af51 | 939 | gimple_stmt_iterator gsi; |
e0a60731 | 940 | |
726a989a RB |
941 | stmt = gsi_stmt (*bsi); |
942 | lhs = gimple_assign_lhs (stmt); | |
943 | type = TREE_TYPE (lhs); | |
e0a60731 | 944 | |
46eb666a | 945 | var = create_tmp_reg (type, "reciptmp"); |
f50d67f6 | 946 | |
8e8e423f | 947 | real_one = build_one_cst (type); |
e0a60731 | 948 | |
726a989a | 949 | stmt1 = gimple_build_assign_with_ops (RDIV_EXPR, |
f50d67f6 | 950 | var, real_one, gimple_assign_rhs2 (stmt)); |
e0a60731 | 951 | name = make_ssa_name (var, stmt1); |
726a989a RB |
952 | gimple_assign_set_lhs (stmt1, name); |
953 | ||
954 | stmt2 = gimple_build_assign_with_ops (MULT_EXPR, lhs, name, | |
955 | gimple_assign_rhs1 (stmt)); | |
e0a60731 RG |
956 | |
957 | /* Replace division stmt with reciprocal and multiply stmts. | |
958 | The multiply stmt is not invariant, so update iterator | |
959 | and avoid rescanning. */ | |
0ca5af51 AO |
960 | gsi = *bsi; |
961 | gsi_insert_before (bsi, stmt1, GSI_NEW_STMT); | |
962 | gsi_replace (&gsi, stmt2, true); | |
e0a60731 RG |
963 | |
964 | /* Continue processing with invariant reciprocal statement. */ | |
965 | return stmt1; | |
966 | } | |
967 | ||
968 | /* Check if the pattern at *BSI is a bittest of the form | |
969 | (A >> B) & 1 != 0 and in this case rewrite it to A & (1 << B) != 0. */ | |
970 | ||
726a989a RB |
971 | static gimple |
972 | rewrite_bittest (gimple_stmt_iterator *bsi) | |
e0a60731 | 973 | { |
726a989a RB |
974 | gimple stmt, use_stmt, stmt1, stmt2; |
975 | tree lhs, var, name, t, a, b; | |
e0a60731 RG |
976 | use_operand_p use; |
977 | ||
726a989a RB |
978 | stmt = gsi_stmt (*bsi); |
979 | lhs = gimple_assign_lhs (stmt); | |
e0a60731 RG |
980 | |
981 | /* Verify that the single use of lhs is a comparison against zero. */ | |
982 | if (TREE_CODE (lhs) != SSA_NAME | |
5c7ec4f0 | 983 | || !single_imm_use (lhs, &use, &use_stmt) |
726a989a | 984 | || gimple_code (use_stmt) != GIMPLE_COND) |
e0a60731 | 985 | return stmt; |
726a989a RB |
986 | if (gimple_cond_lhs (use_stmt) != lhs |
987 | || (gimple_cond_code (use_stmt) != NE_EXPR | |
988 | && gimple_cond_code (use_stmt) != EQ_EXPR) | |
989 | || !integer_zerop (gimple_cond_rhs (use_stmt))) | |
e0a60731 RG |
990 | return stmt; |
991 | ||
992 | /* Get at the operands of the shift. The rhs is TMP1 & 1. */ | |
726a989a RB |
993 | stmt1 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt)); |
994 | if (gimple_code (stmt1) != GIMPLE_ASSIGN) | |
e0a60731 RG |
995 | return stmt; |
996 | ||
0d52bcc1 | 997 | /* There is a conversion in between possibly inserted by fold. */ |
1a87cf0c | 998 | if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt1))) |
e0a60731 | 999 | { |
726a989a | 1000 | t = gimple_assign_rhs1 (stmt1); |
e0a60731 RG |
1001 | if (TREE_CODE (t) != SSA_NAME |
1002 | || !has_single_use (t)) | |
1003 | return stmt; | |
1004 | stmt1 = SSA_NAME_DEF_STMT (t); | |
726a989a | 1005 | if (gimple_code (stmt1) != GIMPLE_ASSIGN) |
e0a60731 | 1006 | return stmt; |
e0a60731 RG |
1007 | } |
1008 | ||
1009 | /* Verify that B is loop invariant but A is not. Verify that with | |
1010 | all the stmt walking we are still in the same loop. */ | |
726a989a RB |
1011 | if (gimple_assign_rhs_code (stmt1) != RSHIFT_EXPR |
1012 | || loop_containing_stmt (stmt1) != loop_containing_stmt (stmt)) | |
1013 | return stmt; | |
e0a60731 | 1014 | |
726a989a RB |
1015 | a = gimple_assign_rhs1 (stmt1); |
1016 | b = gimple_assign_rhs2 (stmt1); | |
1017 | ||
1018 | if (outermost_invariant_loop (b, loop_containing_stmt (stmt1)) != NULL | |
1019 | && outermost_invariant_loop (a, loop_containing_stmt (stmt1)) == NULL) | |
1020 | { | |
0ca5af51 AO |
1021 | gimple_stmt_iterator rsi; |
1022 | ||
e0a60731 RG |
1023 | /* 1 << B */ |
1024 | var = create_tmp_var (TREE_TYPE (a), "shifttmp"); | |
e0a60731 RG |
1025 | t = fold_build2 (LSHIFT_EXPR, TREE_TYPE (a), |
1026 | build_int_cst (TREE_TYPE (a), 1), b); | |
726a989a | 1027 | stmt1 = gimple_build_assign (var, t); |
e0a60731 | 1028 | name = make_ssa_name (var, stmt1); |
726a989a | 1029 | gimple_assign_set_lhs (stmt1, name); |
e0a60731 RG |
1030 | |
1031 | /* A & (1 << B) */ | |
1032 | t = fold_build2 (BIT_AND_EXPR, TREE_TYPE (a), a, name); | |
726a989a | 1033 | stmt2 = gimple_build_assign (var, t); |
5c7ec4f0 | 1034 | name = make_ssa_name (var, stmt2); |
726a989a | 1035 | gimple_assign_set_lhs (stmt2, name); |
a6e2d112 UB |
1036 | |
1037 | /* Replace the SSA_NAME we compare against zero. Adjust | |
1038 | the type of zero accordingly. */ | |
5c7ec4f0 | 1039 | SET_USE (use, name); |
726a989a | 1040 | gimple_cond_set_rhs (use_stmt, build_int_cst_type (TREE_TYPE (name), 0)); |
e0a60731 | 1041 | |
0ca5af51 AO |
1042 | /* Don't use gsi_replace here, none of the new assignments sets |
1043 | the variable originally set in stmt. Move bsi to stmt1, and | |
1044 | then remove the original stmt, so that we get a chance to | |
1045 | retain debug info for it. */ | |
1046 | rsi = *bsi; | |
1047 | gsi_insert_before (bsi, stmt1, GSI_NEW_STMT); | |
1048 | gsi_insert_before (&rsi, stmt2, GSI_SAME_STMT); | |
1049 | gsi_remove (&rsi, true); | |
e0a60731 RG |
1050 | |
1051 | return stmt1; | |
1052 | } | |
1053 | ||
1054 | return stmt; | |
1055 | } | |
1056 | ||
1057 | ||
a7e5372d ZD |
1058 | /* Determine the outermost loops in that statements in basic block BB are |
1059 | invariant, and record them to the LIM_DATA associated with the statements. | |
1060 | Callback for walk_dominator_tree. */ | |
1061 | ||
1062 | static void | |
1063 | determine_invariantness_stmt (struct dom_walk_data *dw_data ATTRIBUTE_UNUSED, | |
1064 | basic_block bb) | |
1065 | { | |
1066 | enum move_pos pos; | |
726a989a RB |
1067 | gimple_stmt_iterator bsi; |
1068 | gimple stmt; | |
a7e5372d ZD |
1069 | bool maybe_never = ALWAYS_EXECUTED_IN (bb) == NULL; |
1070 | struct loop *outermost = ALWAYS_EXECUTED_IN (bb); | |
726a989a | 1071 | struct lim_aux_data *lim_data; |
a7e5372d | 1072 | |
9ba025a2 | 1073 | if (!loop_outer (bb->loop_father)) |
a7e5372d ZD |
1074 | return; |
1075 | ||
1076 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1077 | fprintf (dump_file, "Basic block %d (loop %d -- depth %d):\n\n", | |
9ba025a2 | 1078 | bb->index, bb->loop_father->num, loop_depth (bb->loop_father)); |
a7e5372d | 1079 | |
e3bdfed6 RG |
1080 | /* Look at PHI nodes, but only if there is at most two. |
1081 | ??? We could relax this further by post-processing the inserted | |
1082 | code and transforming adjacent cond-exprs with the same predicate | |
1083 | to control flow again. */ | |
1084 | bsi = gsi_start_phis (bb); | |
1085 | if (!gsi_end_p (bsi) | |
1086 | && ((gsi_next (&bsi), gsi_end_p (bsi)) | |
1087 | || (gsi_next (&bsi), gsi_end_p (bsi)))) | |
1088 | for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
1089 | { | |
1090 | stmt = gsi_stmt (bsi); | |
1091 | ||
1092 | pos = movement_possibility (stmt); | |
1093 | if (pos == MOVE_IMPOSSIBLE) | |
1094 | continue; | |
1095 | ||
1096 | lim_data = init_lim_data (stmt); | |
1097 | lim_data->always_executed_in = outermost; | |
1098 | ||
1099 | if (!determine_max_movement (stmt, false)) | |
1100 | { | |
1101 | lim_data->max_loop = NULL; | |
1102 | continue; | |
1103 | } | |
1104 | ||
1105 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1106 | { | |
1107 | print_gimple_stmt (dump_file, stmt, 2, 0); | |
1108 | fprintf (dump_file, " invariant up to level %d, cost %d.\n\n", | |
1109 | loop_depth (lim_data->max_loop), | |
1110 | lim_data->cost); | |
1111 | } | |
1112 | ||
1113 | if (lim_data->cost >= LIM_EXPENSIVE) | |
1114 | set_profitable_level (stmt); | |
1115 | } | |
1116 | ||
726a989a | 1117 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
a7e5372d | 1118 | { |
726a989a | 1119 | stmt = gsi_stmt (bsi); |
a7e5372d ZD |
1120 | |
1121 | pos = movement_possibility (stmt); | |
1122 | if (pos == MOVE_IMPOSSIBLE) | |
1123 | { | |
1124 | if (nonpure_call_p (stmt)) | |
1125 | { | |
1126 | maybe_never = true; | |
1127 | outermost = NULL; | |
1128 | } | |
8ded35f9 RG |
1129 | /* Make sure to note always_executed_in for stores to make |
1130 | store-motion work. */ | |
1131 | else if (stmt_makes_single_store (stmt)) | |
1132 | { | |
726a989a RB |
1133 | struct lim_aux_data *lim_data = init_lim_data (stmt); |
1134 | lim_data->always_executed_in = outermost; | |
8ded35f9 | 1135 | } |
a7e5372d ZD |
1136 | continue; |
1137 | } | |
1138 | ||
726a989a RB |
1139 | if (is_gimple_assign (stmt) |
1140 | && (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)) | |
1141 | == GIMPLE_BINARY_RHS)) | |
a9b77cd1 | 1142 | { |
726a989a RB |
1143 | tree op0 = gimple_assign_rhs1 (stmt); |
1144 | tree op1 = gimple_assign_rhs2 (stmt); | |
1145 | struct loop *ol1 = outermost_invariant_loop (op1, | |
1146 | loop_containing_stmt (stmt)); | |
a9b77cd1 ZD |
1147 | |
1148 | /* If divisor is invariant, convert a/b to a*(1/b), allowing reciprocal | |
1149 | to be hoisted out of loop, saving expensive divide. */ | |
1150 | if (pos == MOVE_POSSIBLE | |
726a989a | 1151 | && gimple_assign_rhs_code (stmt) == RDIV_EXPR |
a9b77cd1 ZD |
1152 | && flag_unsafe_math_optimizations |
1153 | && !flag_trapping_math | |
726a989a RB |
1154 | && ol1 != NULL |
1155 | && outermost_invariant_loop (op0, ol1) == NULL) | |
a9b77cd1 ZD |
1156 | stmt = rewrite_reciprocal (&bsi); |
1157 | ||
1158 | /* If the shift count is invariant, convert (A >> B) & 1 to | |
1159 | A & (1 << B) allowing the bit mask to be hoisted out of the loop | |
1160 | saving an expensive shift. */ | |
1161 | if (pos == MOVE_POSSIBLE | |
726a989a RB |
1162 | && gimple_assign_rhs_code (stmt) == BIT_AND_EXPR |
1163 | && integer_onep (op1) | |
1164 | && TREE_CODE (op0) == SSA_NAME | |
1165 | && has_single_use (op0)) | |
a9b77cd1 ZD |
1166 | stmt = rewrite_bittest (&bsi); |
1167 | } | |
37cca405 | 1168 | |
726a989a RB |
1169 | lim_data = init_lim_data (stmt); |
1170 | lim_data->always_executed_in = outermost; | |
a7e5372d ZD |
1171 | |
1172 | if (maybe_never && pos == MOVE_PRESERVE_EXECUTION) | |
1173 | continue; | |
1174 | ||
1175 | if (!determine_max_movement (stmt, pos == MOVE_PRESERVE_EXECUTION)) | |
1176 | { | |
726a989a | 1177 | lim_data->max_loop = NULL; |
a7e5372d ZD |
1178 | continue; |
1179 | } | |
1180 | ||
1181 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1182 | { | |
726a989a | 1183 | print_gimple_stmt (dump_file, stmt, 2, 0); |
a7e5372d | 1184 | fprintf (dump_file, " invariant up to level %d, cost %d.\n\n", |
726a989a RB |
1185 | loop_depth (lim_data->max_loop), |
1186 | lim_data->cost); | |
a7e5372d ZD |
1187 | } |
1188 | ||
726a989a | 1189 | if (lim_data->cost >= LIM_EXPENSIVE) |
a7e5372d ZD |
1190 | set_profitable_level (stmt); |
1191 | } | |
1192 | } | |
1193 | ||
1194 | /* For each statement determines the outermost loop in that it is invariant, | |
1195 | statements on whose motion it depends and the cost of the computation. | |
1196 | This information is stored to the LIM_DATA structure associated with | |
1197 | each statement. */ | |
1198 | ||
1199 | static void | |
1200 | determine_invariantness (void) | |
1201 | { | |
1202 | struct dom_walk_data walk_data; | |
1203 | ||
1204 | memset (&walk_data, 0, sizeof (struct dom_walk_data)); | |
2b28c07a | 1205 | walk_data.dom_direction = CDI_DOMINATORS; |
ccf5c864 | 1206 | walk_data.before_dom_children = determine_invariantness_stmt; |
a7e5372d ZD |
1207 | |
1208 | init_walk_dominator_tree (&walk_data); | |
1209 | walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR); | |
1210 | fini_walk_dominator_tree (&walk_data); | |
1211 | } | |
1212 | ||
a7e5372d | 1213 | /* Hoist the statements in basic block BB out of the loops prescribed by |
2a7e31df | 1214 | data stored in LIM_DATA structures associated with each statement. Callback |
a7e5372d ZD |
1215 | for walk_dominator_tree. */ |
1216 | ||
1217 | static void | |
e3bdfed6 | 1218 | move_computations_stmt (struct dom_walk_data *dw_data, |
a7e5372d ZD |
1219 | basic_block bb) |
1220 | { | |
1221 | struct loop *level; | |
726a989a RB |
1222 | gimple_stmt_iterator bsi; |
1223 | gimple stmt; | |
a7e5372d | 1224 | unsigned cost = 0; |
726a989a | 1225 | struct lim_aux_data *lim_data; |
a7e5372d | 1226 | |
9ba025a2 | 1227 | if (!loop_outer (bb->loop_father)) |
a7e5372d ZD |
1228 | return; |
1229 | ||
e3bdfed6 RG |
1230 | for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); ) |
1231 | { | |
1232 | gimple new_stmt; | |
1233 | stmt = gsi_stmt (bsi); | |
1234 | ||
1235 | lim_data = get_lim_data (stmt); | |
1236 | if (lim_data == NULL) | |
1237 | { | |
1238 | gsi_next (&bsi); | |
1239 | continue; | |
1240 | } | |
1241 | ||
1242 | cost = lim_data->cost; | |
1243 | level = lim_data->tgt_loop; | |
1244 | clear_lim_data (stmt); | |
1245 | ||
1246 | if (!level) | |
1247 | { | |
1248 | gsi_next (&bsi); | |
1249 | continue; | |
1250 | } | |
1251 | ||
1252 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1253 | { | |
1254 | fprintf (dump_file, "Moving PHI node\n"); | |
1255 | print_gimple_stmt (dump_file, stmt, 0, 0); | |
1256 | fprintf (dump_file, "(cost %u) out of loop %d.\n\n", | |
1257 | cost, level->num); | |
1258 | } | |
1259 | ||
1260 | if (gimple_phi_num_args (stmt) == 1) | |
1261 | { | |
1262 | tree arg = PHI_ARG_DEF (stmt, 0); | |
1263 | new_stmt = gimple_build_assign_with_ops (TREE_CODE (arg), | |
1264 | gimple_phi_result (stmt), | |
1265 | arg, NULL_TREE); | |
1266 | SSA_NAME_DEF_STMT (gimple_phi_result (stmt)) = new_stmt; | |
1267 | } | |
1268 | else | |
1269 | { | |
1270 | basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb); | |
1271 | gimple cond = gsi_stmt (gsi_last_bb (dom)); | |
1272 | tree arg0 = NULL_TREE, arg1 = NULL_TREE, t; | |
1273 | /* Get the PHI arguments corresponding to the true and false | |
1274 | edges of COND. */ | |
1275 | extract_true_false_args_from_phi (dom, stmt, &arg0, &arg1); | |
1276 | gcc_assert (arg0 && arg1); | |
1277 | t = build2 (gimple_cond_code (cond), boolean_type_node, | |
1278 | gimple_cond_lhs (cond), gimple_cond_rhs (cond)); | |
4e71066d RG |
1279 | new_stmt = gimple_build_assign_with_ops3 (COND_EXPR, |
1280 | gimple_phi_result (stmt), | |
1281 | t, arg0, arg1); | |
e3bdfed6 RG |
1282 | SSA_NAME_DEF_STMT (gimple_phi_result (stmt)) = new_stmt; |
1283 | *((unsigned int *)(dw_data->global_data)) |= TODO_cleanup_cfg; | |
1284 | } | |
1285 | gsi_insert_on_edge (loop_preheader_edge (level), new_stmt); | |
1286 | remove_phi_node (&bsi, false); | |
1287 | } | |
1288 | ||
726a989a | 1289 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); ) |
a7e5372d | 1290 | { |
13714310 RG |
1291 | edge e; |
1292 | ||
726a989a | 1293 | stmt = gsi_stmt (bsi); |
a7e5372d | 1294 | |
726a989a RB |
1295 | lim_data = get_lim_data (stmt); |
1296 | if (lim_data == NULL) | |
a7e5372d | 1297 | { |
726a989a | 1298 | gsi_next (&bsi); |
a7e5372d ZD |
1299 | continue; |
1300 | } | |
1301 | ||
726a989a RB |
1302 | cost = lim_data->cost; |
1303 | level = lim_data->tgt_loop; | |
1304 | clear_lim_data (stmt); | |
a7e5372d ZD |
1305 | |
1306 | if (!level) | |
1307 | { | |
726a989a | 1308 | gsi_next (&bsi); |
a7e5372d ZD |
1309 | continue; |
1310 | } | |
1311 | ||
1312 | /* We do not really want to move conditionals out of the loop; we just | |
1313 | placed it here to force its operands to be moved if necessary. */ | |
726a989a | 1314 | if (gimple_code (stmt) == GIMPLE_COND) |
a7e5372d ZD |
1315 | continue; |
1316 | ||
1317 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1318 | { | |
1319 | fprintf (dump_file, "Moving statement\n"); | |
726a989a | 1320 | print_gimple_stmt (dump_file, stmt, 0, 0); |
a7e5372d ZD |
1321 | fprintf (dump_file, "(cost %u) out of loop %d.\n\n", |
1322 | cost, level->num); | |
1323 | } | |
72425608 | 1324 | |
13714310 RG |
1325 | e = loop_preheader_edge (level); |
1326 | gcc_assert (!gimple_vdef (stmt)); | |
1327 | if (gimple_vuse (stmt)) | |
1328 | { | |
1329 | /* The new VUSE is the one from the virtual PHI in the loop | |
1330 | header or the one already present. */ | |
1331 | gimple_stmt_iterator gsi2; | |
1332 | for (gsi2 = gsi_start_phis (e->dest); | |
1333 | !gsi_end_p (gsi2); gsi_next (&gsi2)) | |
1334 | { | |
1335 | gimple phi = gsi_stmt (gsi2); | |
1336 | if (!is_gimple_reg (gimple_phi_result (phi))) | |
1337 | { | |
1338 | gimple_set_vuse (stmt, PHI_ARG_DEF_FROM_EDGE (phi, e)); | |
1339 | break; | |
1340 | } | |
1341 | } | |
1342 | } | |
726a989a | 1343 | gsi_remove (&bsi, false); |
13714310 | 1344 | gsi_insert_on_edge (e, stmt); |
a7e5372d ZD |
1345 | } |
1346 | } | |
1347 | ||
1348 | /* Hoist the statements out of the loops prescribed by data stored in | |
2a7e31df | 1349 | LIM_DATA structures associated with each statement.*/ |
a7e5372d | 1350 | |
e3bdfed6 | 1351 | static unsigned int |
a7e5372d ZD |
1352 | move_computations (void) |
1353 | { | |
1354 | struct dom_walk_data walk_data; | |
e3bdfed6 | 1355 | unsigned int todo = 0; |
a7e5372d ZD |
1356 | |
1357 | memset (&walk_data, 0, sizeof (struct dom_walk_data)); | |
e3bdfed6 | 1358 | walk_data.global_data = &todo; |
2b28c07a | 1359 | walk_data.dom_direction = CDI_DOMINATORS; |
ccf5c864 | 1360 | walk_data.before_dom_children = move_computations_stmt; |
a7e5372d ZD |
1361 | |
1362 | init_walk_dominator_tree (&walk_data); | |
1363 | walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR); | |
1364 | fini_walk_dominator_tree (&walk_data); | |
1365 | ||
726a989a | 1366 | gsi_commit_edge_inserts (); |
5006671f | 1367 | if (need_ssa_update_p (cfun)) |
84d65814 | 1368 | rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); |
e3bdfed6 RG |
1369 | |
1370 | return todo; | |
a7e5372d ZD |
1371 | } |
1372 | ||
1373 | /* Checks whether the statement defining variable *INDEX can be hoisted | |
1374 | out of the loop passed in DATA. Callback for for_each_index. */ | |
1375 | ||
1376 | static bool | |
1377 | may_move_till (tree ref, tree *index, void *data) | |
1378 | { | |
726a989a | 1379 | struct loop *loop = (struct loop *) data, *max_loop; |
a7e5372d ZD |
1380 | |
1381 | /* If REF is an array reference, check also that the step and the lower | |
1382 | bound is invariant in LOOP. */ | |
1383 | if (TREE_CODE (ref) == ARRAY_REF) | |
1384 | { | |
726a989a RB |
1385 | tree step = TREE_OPERAND (ref, 3); |
1386 | tree lbound = TREE_OPERAND (ref, 2); | |
a7e5372d | 1387 | |
726a989a | 1388 | max_loop = outermost_invariant_loop (step, loop); |
a7e5372d ZD |
1389 | if (!max_loop) |
1390 | return false; | |
1391 | ||
726a989a | 1392 | max_loop = outermost_invariant_loop (lbound, loop); |
a7e5372d ZD |
1393 | if (!max_loop) |
1394 | return false; | |
1395 | } | |
1396 | ||
1397 | max_loop = outermost_invariant_loop (*index, loop); | |
1398 | if (!max_loop) | |
1399 | return false; | |
1400 | ||
1401 | return true; | |
1402 | } | |
1403 | ||
726a989a | 1404 | /* If OP is SSA NAME, force the statement that defines it to be |
b4042a03 | 1405 | moved out of the LOOP. ORIG_LOOP is the loop in that EXPR is used. */ |
a7e5372d ZD |
1406 | |
1407 | static void | |
726a989a | 1408 | force_move_till_op (tree op, struct loop *orig_loop, struct loop *loop) |
a7e5372d | 1409 | { |
726a989a | 1410 | gimple stmt; |
a7e5372d | 1411 | |
726a989a RB |
1412 | if (!op |
1413 | || is_gimple_min_invariant (op)) | |
1414 | return; | |
a7e5372d | 1415 | |
726a989a | 1416 | gcc_assert (TREE_CODE (op) == SSA_NAME); |
b8698a0f | 1417 | |
726a989a RB |
1418 | stmt = SSA_NAME_DEF_STMT (op); |
1419 | if (gimple_nop_p (stmt)) | |
a7e5372d ZD |
1420 | return; |
1421 | ||
726a989a | 1422 | set_level (stmt, orig_loop, loop); |
a7e5372d ZD |
1423 | } |
1424 | ||
1425 | /* Forces statement defining invariants in REF (and *INDEX) to be moved out of | |
b4042a03 ZD |
1426 | the LOOP. The reference REF is used in the loop ORIG_LOOP. Callback for |
1427 | for_each_index. */ | |
1428 | ||
1429 | struct fmt_data | |
1430 | { | |
1431 | struct loop *loop; | |
1432 | struct loop *orig_loop; | |
1433 | }; | |
a7e5372d ZD |
1434 | |
1435 | static bool | |
1436 | force_move_till (tree ref, tree *index, void *data) | |
1437 | { | |
c22940cd | 1438 | struct fmt_data *fmt_data = (struct fmt_data *) data; |
a7e5372d ZD |
1439 | |
1440 | if (TREE_CODE (ref) == ARRAY_REF) | |
1441 | { | |
726a989a RB |
1442 | tree step = TREE_OPERAND (ref, 3); |
1443 | tree lbound = TREE_OPERAND (ref, 2); | |
a7e5372d | 1444 | |
726a989a RB |
1445 | force_move_till_op (step, fmt_data->orig_loop, fmt_data->loop); |
1446 | force_move_till_op (lbound, fmt_data->orig_loop, fmt_data->loop); | |
a7e5372d ZD |
1447 | } |
1448 | ||
726a989a | 1449 | force_move_till_op (*index, fmt_data->orig_loop, fmt_data->loop); |
a7e5372d ZD |
1450 | |
1451 | return true; | |
1452 | } | |
1453 | ||
72425608 ZD |
1454 | /* A hash function for struct mem_ref object OBJ. */ |
1455 | ||
1456 | static hashval_t | |
1457 | memref_hash (const void *obj) | |
1458 | { | |
3d9a9f94 | 1459 | const struct mem_ref *const mem = (const struct mem_ref *) obj; |
72425608 ZD |
1460 | |
1461 | return mem->hash; | |
1462 | } | |
1463 | ||
1464 | /* An equality function for struct mem_ref object OBJ1 with | |
1465 | memory reference OBJ2. */ | |
1466 | ||
1467 | static int | |
1468 | memref_eq (const void *obj1, const void *obj2) | |
1469 | { | |
3d9a9f94 | 1470 | const struct mem_ref *const mem1 = (const struct mem_ref *) obj1; |
72425608 | 1471 | |
3d9a9f94 | 1472 | return operand_equal_p (mem1->mem, (const_tree) obj2, 0); |
72425608 ZD |
1473 | } |
1474 | ||
1475 | /* Releases list of memory reference locations ACCS. */ | |
1476 | ||
1477 | static void | |
1478 | free_mem_ref_locs (mem_ref_locs_p accs) | |
1479 | { | |
1480 | unsigned i; | |
1481 | mem_ref_loc_p loc; | |
1482 | ||
1483 | if (!accs) | |
1484 | return; | |
1485 | ||
ac47786e | 1486 | FOR_EACH_VEC_ELT (mem_ref_loc_p, accs->locs, i, loc) |
72425608 ZD |
1487 | free (loc); |
1488 | VEC_free (mem_ref_loc_p, heap, accs->locs); | |
1489 | free (accs); | |
1490 | } | |
1491 | ||
1492 | /* A function to free the mem_ref object OBJ. */ | |
1493 | ||
1494 | static void | |
1495 | memref_free (void *obj) | |
1496 | { | |
3d9a9f94 | 1497 | struct mem_ref *const mem = (struct mem_ref *) obj; |
72425608 ZD |
1498 | unsigned i; |
1499 | mem_ref_locs_p accs; | |
1500 | ||
1501 | BITMAP_FREE (mem->stored); | |
1502 | BITMAP_FREE (mem->indep_loop); | |
1503 | BITMAP_FREE (mem->dep_loop); | |
1504 | BITMAP_FREE (mem->indep_ref); | |
1505 | BITMAP_FREE (mem->dep_ref); | |
1506 | ||
ac47786e | 1507 | FOR_EACH_VEC_ELT (mem_ref_locs_p, mem->accesses_in_loop, i, accs) |
72425608 ZD |
1508 | free_mem_ref_locs (accs); |
1509 | VEC_free (mem_ref_locs_p, heap, mem->accesses_in_loop); | |
1510 | ||
72425608 ZD |
1511 | free (mem); |
1512 | } | |
1513 | ||
1514 | /* Allocates and returns a memory reference description for MEM whose hash | |
1515 | value is HASH and id is ID. */ | |
1516 | ||
1517 | static mem_ref_p | |
1518 | mem_ref_alloc (tree mem, unsigned hash, unsigned id) | |
1519 | { | |
1520 | mem_ref_p ref = XNEW (struct mem_ref); | |
1521 | ref->mem = mem; | |
1522 | ref->id = id; | |
1523 | ref->hash = hash; | |
1524 | ref->stored = BITMAP_ALLOC (NULL); | |
1525 | ref->indep_loop = BITMAP_ALLOC (NULL); | |
1526 | ref->dep_loop = BITMAP_ALLOC (NULL); | |
1527 | ref->indep_ref = BITMAP_ALLOC (NULL); | |
1528 | ref->dep_ref = BITMAP_ALLOC (NULL); | |
1529 | ref->accesses_in_loop = NULL; | |
72425608 ZD |
1530 | |
1531 | return ref; | |
1532 | } | |
1533 | ||
1534 | /* Allocates and returns the new list of locations. */ | |
1535 | ||
1536 | static mem_ref_locs_p | |
1537 | mem_ref_locs_alloc (void) | |
1538 | { | |
1539 | mem_ref_locs_p accs = XNEW (struct mem_ref_locs); | |
1540 | accs->locs = NULL; | |
1541 | return accs; | |
1542 | } | |
1543 | ||
1544 | /* Records memory reference location *LOC in LOOP to the memory reference | |
1545 | description REF. The reference occurs in statement STMT. */ | |
a7e5372d ZD |
1546 | |
1547 | static void | |
726a989a | 1548 | record_mem_ref_loc (mem_ref_p ref, struct loop *loop, gimple stmt, tree *loc) |
a7e5372d | 1549 | { |
72425608 ZD |
1550 | mem_ref_loc_p aref = XNEW (struct mem_ref_loc); |
1551 | mem_ref_locs_p accs; | |
1552 | bitmap ril = VEC_index (bitmap, memory_accesses.refs_in_loop, loop->num); | |
1553 | ||
1554 | if (VEC_length (mem_ref_locs_p, ref->accesses_in_loop) | |
1555 | <= (unsigned) loop->num) | |
1556 | VEC_safe_grow_cleared (mem_ref_locs_p, heap, ref->accesses_in_loop, | |
1557 | loop->num + 1); | |
1558 | accs = VEC_index (mem_ref_locs_p, ref->accesses_in_loop, loop->num); | |
1559 | if (!accs) | |
1560 | { | |
1561 | accs = mem_ref_locs_alloc (); | |
1562 | VEC_replace (mem_ref_locs_p, ref->accesses_in_loop, loop->num, accs); | |
1563 | } | |
a7e5372d ZD |
1564 | |
1565 | aref->stmt = stmt; | |
72425608 | 1566 | aref->ref = loc; |
a7e5372d | 1567 | |
72425608 ZD |
1568 | VEC_safe_push (mem_ref_loc_p, heap, accs->locs, aref); |
1569 | bitmap_set_bit (ril, ref->id); | |
a7e5372d ZD |
1570 | } |
1571 | ||
72425608 | 1572 | /* Marks reference REF as stored in LOOP. */ |
a7e5372d ZD |
1573 | |
1574 | static void | |
72425608 | 1575 | mark_ref_stored (mem_ref_p ref, struct loop *loop) |
a7e5372d | 1576 | { |
72425608 ZD |
1577 | for (; |
1578 | loop != current_loops->tree_root | |
1579 | && !bitmap_bit_p (ref->stored, loop->num); | |
1580 | loop = loop_outer (loop)) | |
1581 | bitmap_set_bit (ref->stored, loop->num); | |
1582 | } | |
1583 | ||
1584 | /* Gathers memory references in statement STMT in LOOP, storing the | |
1585 | information about them in the memory_accesses structure. Marks | |
1586 | the vops accessed through unrecognized statements there as | |
1587 | well. */ | |
1588 | ||
1589 | static void | |
726a989a | 1590 | gather_mem_refs_stmt (struct loop *loop, gimple stmt) |
72425608 ZD |
1591 | { |
1592 | tree *mem = NULL; | |
1593 | hashval_t hash; | |
1594 | PTR *slot; | |
1595 | mem_ref_p ref; | |
72425608 | 1596 | bool is_stored; |
72425608 | 1597 | unsigned id; |
a7e5372d | 1598 | |
5006671f | 1599 | if (!gimple_vuse (stmt)) |
72425608 ZD |
1600 | return; |
1601 | ||
1602 | mem = simple_mem_ref_in_stmt (stmt, &is_stored); | |
1603 | if (!mem) | |
546d314c RG |
1604 | { |
1605 | id = VEC_length (mem_ref_p, memory_accesses.refs_list); | |
1606 | ref = mem_ref_alloc (error_mark_node, 0, id); | |
1607 | VEC_safe_push (mem_ref_p, heap, memory_accesses.refs_list, ref); | |
1608 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1609 | { | |
1610 | fprintf (dump_file, "Unanalyzed memory reference %u: ", id); | |
1611 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); | |
1612 | } | |
1613 | if (gimple_vdef (stmt)) | |
1614 | mark_ref_stored (ref, loop); | |
1615 | record_mem_ref_loc (ref, loop, stmt, mem); | |
1616 | return; | |
1617 | } | |
72425608 ZD |
1618 | |
1619 | hash = iterative_hash_expr (*mem, 0); | |
1620 | slot = htab_find_slot_with_hash (memory_accesses.refs, *mem, hash, INSERT); | |
1621 | ||
1622 | if (*slot) | |
1623 | { | |
3d9a9f94 | 1624 | ref = (mem_ref_p) *slot; |
72425608 ZD |
1625 | id = ref->id; |
1626 | } | |
1627 | else | |
a7e5372d | 1628 | { |
72425608 ZD |
1629 | id = VEC_length (mem_ref_p, memory_accesses.refs_list); |
1630 | ref = mem_ref_alloc (*mem, hash, id); | |
1631 | VEC_safe_push (mem_ref_p, heap, memory_accesses.refs_list, ref); | |
1632 | *slot = ref; | |
1633 | ||
1634 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1635 | { | |
1636 | fprintf (dump_file, "Memory reference %u: ", id); | |
1637 | print_generic_expr (dump_file, ref->mem, TDF_SLIM); | |
1638 | fprintf (dump_file, "\n"); | |
1639 | } | |
a7e5372d | 1640 | } |
039496da | 1641 | |
72425608 ZD |
1642 | if (is_stored) |
1643 | mark_ref_stored (ref, loop); | |
1644 | ||
72425608 ZD |
1645 | record_mem_ref_loc (ref, loop, stmt, mem); |
1646 | return; | |
a7e5372d ZD |
1647 | } |
1648 | ||
72425608 | 1649 | /* Gathers memory references in loops. */ |
a7e5372d ZD |
1650 | |
1651 | static void | |
72425608 | 1652 | gather_mem_refs_in_loops (void) |
a7e5372d | 1653 | { |
726a989a | 1654 | gimple_stmt_iterator bsi; |
72425608 ZD |
1655 | basic_block bb; |
1656 | struct loop *loop; | |
1657 | loop_iterator li; | |
72425608 ZD |
1658 | bitmap lrefs, alrefs, alrefso; |
1659 | ||
1660 | FOR_EACH_BB (bb) | |
1661 | { | |
1662 | loop = bb->loop_father; | |
1663 | if (loop == current_loops->tree_root) | |
1664 | continue; | |
1665 | ||
726a989a RB |
1666 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
1667 | gather_mem_refs_stmt (loop, gsi_stmt (bsi)); | |
72425608 ZD |
1668 | } |
1669 | ||
546d314c RG |
1670 | /* Propagate the information about accessed memory references up |
1671 | the loop hierarchy. */ | |
72425608 ZD |
1672 | FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST) |
1673 | { | |
1674 | lrefs = VEC_index (bitmap, memory_accesses.refs_in_loop, loop->num); | |
1675 | alrefs = VEC_index (bitmap, memory_accesses.all_refs_in_loop, loop->num); | |
1676 | bitmap_ior_into (alrefs, lrefs); | |
1677 | ||
1678 | if (loop_outer (loop) == current_loops->tree_root) | |
1679 | continue; | |
1680 | ||
72425608 ZD |
1681 | alrefso = VEC_index (bitmap, memory_accesses.all_refs_in_loop, |
1682 | loop_outer (loop)->num); | |
1683 | bitmap_ior_into (alrefso, alrefs); | |
1684 | } | |
1685 | } | |
1686 | ||
72425608 ZD |
1687 | /* Create a mapping from virtual operands to references that touch them |
1688 | in LOOP. */ | |
1689 | ||
1690 | static void | |
1691 | create_vop_ref_mapping_loop (struct loop *loop) | |
1692 | { | |
1693 | bitmap refs = VEC_index (bitmap, memory_accesses.refs_in_loop, loop->num); | |
1694 | struct loop *sloop; | |
1695 | bitmap_iterator bi; | |
1696 | unsigned i; | |
1697 | mem_ref_p ref; | |
1698 | ||
1699 | EXECUTE_IF_SET_IN_BITMAP (refs, 0, i, bi) | |
1700 | { | |
1701 | ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i); | |
546d314c RG |
1702 | for (sloop = loop; sloop != current_loops->tree_root; |
1703 | sloop = loop_outer (sloop)) | |
1704 | if (bitmap_bit_p (ref->stored, loop->num)) | |
1705 | { | |
1706 | bitmap refs_stored | |
1707 | = VEC_index (bitmap, memory_accesses.all_refs_stored_in_loop, | |
1708 | sloop->num); | |
1709 | bitmap_set_bit (refs_stored, ref->id); | |
1710 | } | |
72425608 ZD |
1711 | } |
1712 | } | |
1713 | ||
1714 | /* For each non-clobbered virtual operand and each loop, record the memory | |
1715 | references in this loop that touch the operand. */ | |
1716 | ||
1717 | static void | |
1718 | create_vop_ref_mapping (void) | |
1719 | { | |
1720 | loop_iterator li; | |
1721 | struct loop *loop; | |
1722 | ||
1723 | FOR_EACH_LOOP (li, loop, 0) | |
1724 | { | |
1725 | create_vop_ref_mapping_loop (loop); | |
1726 | } | |
1727 | } | |
1728 | ||
1729 | /* Gathers information about memory accesses in the loops. */ | |
1730 | ||
1731 | static void | |
1732 | analyze_memory_references (void) | |
1733 | { | |
1734 | unsigned i; | |
1735 | bitmap empty; | |
72425608 ZD |
1736 | |
1737 | memory_accesses.refs | |
1738 | = htab_create (100, memref_hash, memref_eq, memref_free); | |
1739 | memory_accesses.refs_list = NULL; | |
1740 | memory_accesses.refs_in_loop = VEC_alloc (bitmap, heap, | |
1741 | number_of_loops ()); | |
1742 | memory_accesses.all_refs_in_loop = VEC_alloc (bitmap, heap, | |
1743 | number_of_loops ()); | |
546d314c RG |
1744 | memory_accesses.all_refs_stored_in_loop = VEC_alloc (bitmap, heap, |
1745 | number_of_loops ()); | |
72425608 ZD |
1746 | |
1747 | for (i = 0; i < number_of_loops (); i++) | |
1748 | { | |
1749 | empty = BITMAP_ALLOC (NULL); | |
1750 | VEC_quick_push (bitmap, memory_accesses.refs_in_loop, empty); | |
1751 | empty = BITMAP_ALLOC (NULL); | |
1752 | VEC_quick_push (bitmap, memory_accesses.all_refs_in_loop, empty); | |
1753 | empty = BITMAP_ALLOC (NULL); | |
546d314c | 1754 | VEC_quick_push (bitmap, memory_accesses.all_refs_stored_in_loop, empty); |
72425608 ZD |
1755 | } |
1756 | ||
1757 | memory_accesses.ttae_cache = NULL; | |
1758 | ||
1759 | gather_mem_refs_in_loops (); | |
1760 | create_vop_ref_mapping (); | |
1761 | } | |
1762 | ||
72425608 ZD |
1763 | /* Returns true if MEM1 and MEM2 may alias. TTAE_CACHE is used as a cache in |
1764 | tree_to_aff_combination_expand. */ | |
1765 | ||
1766 | static bool | |
1767 | mem_refs_may_alias_p (tree mem1, tree mem2, struct pointer_map_t **ttae_cache) | |
1768 | { | |
1769 | /* Perform BASE + OFFSET analysis -- if MEM1 and MEM2 are based on the same | |
1770 | object and their offset differ in such a way that the locations cannot | |
1771 | overlap, then they cannot alias. */ | |
72425608 | 1772 | double_int size1, size2; |
1842e4d4 | 1773 | aff_tree off1, off2; |
72425608 | 1774 | |
1842e4d4 RG |
1775 | /* Perform basic offset and type-based disambiguation. */ |
1776 | if (!refs_may_alias_p (mem1, mem2)) | |
72425608 | 1777 | return false; |
a7e5372d | 1778 | |
72425608 ZD |
1779 | /* The expansion of addresses may be a bit expensive, thus we only do |
1780 | the check at -O2 and higher optimization levels. */ | |
1781 | if (optimize < 2) | |
1782 | return true; | |
1783 | ||
1784 | get_inner_reference_aff (mem1, &off1, &size1); | |
1785 | get_inner_reference_aff (mem2, &off2, &size2); | |
1786 | aff_combination_expand (&off1, ttae_cache); | |
1787 | aff_combination_expand (&off2, ttae_cache); | |
1788 | aff_combination_scale (&off1, double_int_minus_one); | |
1789 | aff_combination_add (&off2, &off1); | |
1790 | ||
02f5d6c5 | 1791 | if (aff_comb_cannot_overlap_p (&off2, size1, size2)) |
72425608 ZD |
1792 | return false; |
1793 | ||
1794 | return true; | |
1795 | } | |
1796 | ||
1797 | /* Rewrites location LOC by TMP_VAR. */ | |
1798 | ||
1799 | static void | |
1800 | rewrite_mem_ref_loc (mem_ref_loc_p loc, tree tmp_var) | |
1801 | { | |
72425608 ZD |
1802 | *loc->ref = tmp_var; |
1803 | update_stmt (loc->stmt); | |
1804 | } | |
1805 | ||
1806 | /* Adds all locations of REF in LOOP and its subloops to LOCS. */ | |
1807 | ||
1808 | static void | |
1809 | get_all_locs_in_loop (struct loop *loop, mem_ref_p ref, | |
1810 | VEC (mem_ref_loc_p, heap) **locs) | |
1811 | { | |
1812 | mem_ref_locs_p accs; | |
1813 | unsigned i; | |
1814 | mem_ref_loc_p loc; | |
1815 | bitmap refs = VEC_index (bitmap, memory_accesses.all_refs_in_loop, | |
1816 | loop->num); | |
1817 | struct loop *subloop; | |
1818 | ||
1819 | if (!bitmap_bit_p (refs, ref->id)) | |
1820 | return; | |
1821 | ||
1822 | if (VEC_length (mem_ref_locs_p, ref->accesses_in_loop) | |
1823 | > (unsigned) loop->num) | |
1824 | { | |
1825 | accs = VEC_index (mem_ref_locs_p, ref->accesses_in_loop, loop->num); | |
1826 | if (accs) | |
1827 | { | |
ac47786e | 1828 | FOR_EACH_VEC_ELT (mem_ref_loc_p, accs->locs, i, loc) |
72425608 ZD |
1829 | VEC_safe_push (mem_ref_loc_p, heap, *locs, loc); |
1830 | } | |
1831 | } | |
1832 | ||
1833 | for (subloop = loop->inner; subloop != NULL; subloop = subloop->next) | |
1834 | get_all_locs_in_loop (subloop, ref, locs); | |
1835 | } | |
1836 | ||
1837 | /* Rewrites all references to REF in LOOP by variable TMP_VAR. */ | |
1838 | ||
1839 | static void | |
1840 | rewrite_mem_refs (struct loop *loop, mem_ref_p ref, tree tmp_var) | |
1841 | { | |
1842 | unsigned i; | |
1843 | mem_ref_loc_p loc; | |
1844 | VEC (mem_ref_loc_p, heap) *locs = NULL; | |
1845 | ||
1846 | get_all_locs_in_loop (loop, ref, &locs); | |
ac47786e | 1847 | FOR_EACH_VEC_ELT (mem_ref_loc_p, locs, i, loc) |
72425608 ZD |
1848 | rewrite_mem_ref_loc (loc, tmp_var); |
1849 | VEC_free (mem_ref_loc_p, heap, locs); | |
a7e5372d ZD |
1850 | } |
1851 | ||
d28cbb07 ZD |
1852 | /* The name and the length of the currently generated variable |
1853 | for lsm. */ | |
1854 | #define MAX_LSM_NAME_LENGTH 40 | |
1855 | static char lsm_tmp_name[MAX_LSM_NAME_LENGTH + 1]; | |
1856 | static int lsm_tmp_name_length; | |
1857 | ||
1858 | /* Adds S to lsm_tmp_name. */ | |
1859 | ||
1860 | static void | |
1861 | lsm_tmp_name_add (const char *s) | |
1862 | { | |
1863 | int l = strlen (s) + lsm_tmp_name_length; | |
1864 | if (l > MAX_LSM_NAME_LENGTH) | |
1865 | return; | |
1866 | ||
1867 | strcpy (lsm_tmp_name + lsm_tmp_name_length, s); | |
1868 | lsm_tmp_name_length = l; | |
1869 | } | |
1870 | ||
1871 | /* Stores the name for temporary variable that replaces REF to | |
1872 | lsm_tmp_name. */ | |
1873 | ||
1874 | static void | |
1875 | gen_lsm_tmp_name (tree ref) | |
1876 | { | |
1877 | const char *name; | |
1878 | ||
1879 | switch (TREE_CODE (ref)) | |
1880 | { | |
70f34814 | 1881 | case MEM_REF: |
d5fed62d | 1882 | case TARGET_MEM_REF: |
d28cbb07 ZD |
1883 | gen_lsm_tmp_name (TREE_OPERAND (ref, 0)); |
1884 | lsm_tmp_name_add ("_"); | |
1885 | break; | |
1886 | ||
70f34814 RG |
1887 | case ADDR_EXPR: |
1888 | gen_lsm_tmp_name (TREE_OPERAND (ref, 0)); | |
1889 | break; | |
1890 | ||
d28cbb07 ZD |
1891 | case BIT_FIELD_REF: |
1892 | case VIEW_CONVERT_EXPR: | |
1893 | case ARRAY_RANGE_REF: | |
1894 | gen_lsm_tmp_name (TREE_OPERAND (ref, 0)); | |
1895 | break; | |
1896 | ||
1897 | case REALPART_EXPR: | |
1898 | gen_lsm_tmp_name (TREE_OPERAND (ref, 0)); | |
1899 | lsm_tmp_name_add ("_RE"); | |
1900 | break; | |
b8698a0f | 1901 | |
d28cbb07 ZD |
1902 | case IMAGPART_EXPR: |
1903 | gen_lsm_tmp_name (TREE_OPERAND (ref, 0)); | |
1904 | lsm_tmp_name_add ("_IM"); | |
1905 | break; | |
1906 | ||
1907 | case COMPONENT_REF: | |
1908 | gen_lsm_tmp_name (TREE_OPERAND (ref, 0)); | |
1909 | lsm_tmp_name_add ("_"); | |
1910 | name = get_name (TREE_OPERAND (ref, 1)); | |
1911 | if (!name) | |
1912 | name = "F"; | |
d28cbb07 | 1913 | lsm_tmp_name_add (name); |
cbe80ff8 | 1914 | break; |
d28cbb07 ZD |
1915 | |
1916 | case ARRAY_REF: | |
1917 | gen_lsm_tmp_name (TREE_OPERAND (ref, 0)); | |
1918 | lsm_tmp_name_add ("_I"); | |
1919 | break; | |
1920 | ||
1921 | case SSA_NAME: | |
1922 | ref = SSA_NAME_VAR (ref); | |
1923 | /* Fallthru. */ | |
1924 | ||
1925 | case VAR_DECL: | |
1926 | case PARM_DECL: | |
1927 | name = get_name (ref); | |
1928 | if (!name) | |
1929 | name = "D"; | |
1930 | lsm_tmp_name_add (name); | |
1931 | break; | |
1932 | ||
1933 | case STRING_CST: | |
1934 | lsm_tmp_name_add ("S"); | |
1935 | break; | |
1936 | ||
1937 | case RESULT_DECL: | |
1938 | lsm_tmp_name_add ("R"); | |
1939 | break; | |
1940 | ||
150e3929 RG |
1941 | case INTEGER_CST: |
1942 | /* Nothing. */ | |
1943 | break; | |
1944 | ||
d28cbb07 ZD |
1945 | default: |
1946 | gcc_unreachable (); | |
1947 | } | |
1948 | } | |
1949 | ||
1950 | /* Determines name for temporary variable that replaces REF. | |
bbc8a8dc ZD |
1951 | The name is accumulated into the lsm_tmp_name variable. |
1952 | N is added to the name of the temporary. */ | |
d28cbb07 | 1953 | |
bbc8a8dc ZD |
1954 | char * |
1955 | get_lsm_tmp_name (tree ref, unsigned n) | |
d28cbb07 | 1956 | { |
bbc8a8dc ZD |
1957 | char ns[2]; |
1958 | ||
d28cbb07 ZD |
1959 | lsm_tmp_name_length = 0; |
1960 | gen_lsm_tmp_name (ref); | |
1961 | lsm_tmp_name_add ("_lsm"); | |
bbc8a8dc ZD |
1962 | if (n < 10) |
1963 | { | |
1964 | ns[0] = '0' + n; | |
1965 | ns[1] = 0; | |
1966 | lsm_tmp_name_add (ns); | |
1967 | } | |
d28cbb07 ZD |
1968 | return lsm_tmp_name; |
1969 | } | |
1970 | ||
039496da AH |
1971 | struct prev_flag_edges { |
1972 | /* Edge to insert new flag comparison code. */ | |
1973 | edge append_cond_position; | |
1974 | ||
1975 | /* Edge for fall through from previous flag comparison. */ | |
1976 | edge last_cond_fallthru; | |
1977 | }; | |
1978 | ||
1979 | /* Helper function for execute_sm. Emit code to store TMP_VAR into | |
1980 | MEM along edge EX. | |
1981 | ||
1982 | The store is only done if MEM has changed. We do this so no | |
1983 | changes to MEM occur on code paths that did not originally store | |
1984 | into it. | |
1985 | ||
1986 | The common case for execute_sm will transform: | |
1987 | ||
1988 | for (...) { | |
1989 | if (foo) | |
1990 | stuff; | |
1991 | else | |
1992 | MEM = TMP_VAR; | |
1993 | } | |
1994 | ||
1995 | into: | |
1996 | ||
1997 | lsm = MEM; | |
1998 | for (...) { | |
1999 | if (foo) | |
2000 | stuff; | |
2001 | else | |
2002 | lsm = TMP_VAR; | |
2003 | } | |
2004 | MEM = lsm; | |
2005 | ||
2006 | This function will generate: | |
2007 | ||
2008 | lsm = MEM; | |
2009 | ||
2010 | lsm_flag = false; | |
2011 | ... | |
2012 | for (...) { | |
2013 | if (foo) | |
2014 | stuff; | |
2015 | else { | |
2016 | lsm = TMP_VAR; | |
2017 | lsm_flag = true; | |
2018 | } | |
2019 | } | |
2020 | if (lsm_flag) <-- | |
2021 | MEM = lsm; <-- | |
2022 | */ | |
2023 | ||
2024 | static void | |
2025 | execute_sm_if_changed (edge ex, tree mem, tree tmp_var, tree flag) | |
2026 | { | |
2027 | basic_block new_bb, then_bb, old_dest; | |
2028 | bool loop_has_only_one_exit; | |
2029 | edge then_old_edge, orig_ex = ex; | |
2030 | gimple_stmt_iterator gsi; | |
2031 | gimple stmt; | |
2032 | struct prev_flag_edges *prev_edges = (struct prev_flag_edges *) ex->aux; | |
2033 | ||
2034 | /* ?? Insert store after previous store if applicable. See note | |
2035 | below. */ | |
2036 | if (prev_edges) | |
2037 | ex = prev_edges->append_cond_position; | |
2038 | ||
2039 | loop_has_only_one_exit = single_pred_p (ex->dest); | |
2040 | ||
2041 | if (loop_has_only_one_exit) | |
2042 | ex = split_block_after_labels (ex->dest); | |
2043 | ||
2044 | old_dest = ex->dest; | |
2045 | new_bb = split_edge (ex); | |
2046 | then_bb = create_empty_bb (new_bb); | |
2047 | if (current_loops && new_bb->loop_father) | |
2048 | add_bb_to_loop (then_bb, new_bb->loop_father); | |
2049 | ||
2050 | gsi = gsi_start_bb (new_bb); | |
2051 | stmt = gimple_build_cond (NE_EXPR, flag, boolean_false_node, | |
2052 | NULL_TREE, NULL_TREE); | |
2053 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2054 | ||
2055 | gsi = gsi_start_bb (then_bb); | |
2056 | /* Insert actual store. */ | |
2057 | stmt = gimple_build_assign (unshare_expr (mem), tmp_var); | |
2058 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2059 | ||
2060 | make_edge (new_bb, then_bb, EDGE_TRUE_VALUE); | |
2061 | make_edge (new_bb, old_dest, EDGE_FALSE_VALUE); | |
2062 | then_old_edge = make_edge (then_bb, old_dest, EDGE_FALLTHRU); | |
2063 | ||
2064 | set_immediate_dominator (CDI_DOMINATORS, then_bb, new_bb); | |
2065 | ||
2066 | if (prev_edges) | |
2067 | { | |
2068 | basic_block prevbb = prev_edges->last_cond_fallthru->src; | |
2069 | redirect_edge_succ (prev_edges->last_cond_fallthru, new_bb); | |
2070 | set_immediate_dominator (CDI_DOMINATORS, new_bb, prevbb); | |
2071 | set_immediate_dominator (CDI_DOMINATORS, old_dest, | |
2072 | recompute_dominator (CDI_DOMINATORS, old_dest)); | |
2073 | } | |
2074 | ||
2075 | /* ?? Because stores may alias, they must happen in the exact | |
2076 | sequence they originally happened. Save the position right after | |
2077 | the (_lsm) store we just created so we can continue appending after | |
2078 | it and maintain the original order. */ | |
2079 | { | |
2080 | struct prev_flag_edges *p; | |
2081 | ||
2082 | if (orig_ex->aux) | |
2083 | orig_ex->aux = NULL; | |
2084 | alloc_aux_for_edge (orig_ex, sizeof (struct prev_flag_edges)); | |
2085 | p = (struct prev_flag_edges *) orig_ex->aux; | |
2086 | p->append_cond_position = then_old_edge; | |
2087 | p->last_cond_fallthru = find_edge (new_bb, old_dest); | |
2088 | orig_ex->aux = (void *) p; | |
2089 | } | |
2090 | ||
2091 | if (!loop_has_only_one_exit) | |
2092 | for (gsi = gsi_start_phis (old_dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2093 | { | |
2094 | gimple phi = gsi_stmt (gsi); | |
2095 | unsigned i; | |
2096 | ||
2097 | for (i = 0; i < gimple_phi_num_args (phi); i++) | |
2098 | if (gimple_phi_arg_edge (phi, i)->src == new_bb) | |
2099 | { | |
2100 | tree arg = gimple_phi_arg_def (phi, i); | |
9e227d60 | 2101 | add_phi_arg (phi, arg, then_old_edge, UNKNOWN_LOCATION); |
039496da AH |
2102 | update_stmt (phi); |
2103 | } | |
2104 | } | |
2105 | /* Remove the original fall through edge. This was the | |
2106 | single_succ_edge (new_bb). */ | |
2107 | EDGE_SUCC (new_bb, 0)->flags &= ~EDGE_FALLTHRU; | |
2108 | } | |
2109 | ||
2110 | /* Helper function for execute_sm. On every location where REF is | |
2111 | set, set an appropriate flag indicating the store. */ | |
2112 | ||
2113 | static tree | |
2114 | execute_sm_if_changed_flag_set (struct loop *loop, mem_ref_p ref) | |
2115 | { | |
2116 | unsigned i; | |
2117 | mem_ref_loc_p loc; | |
2118 | tree flag; | |
2119 | VEC (mem_ref_loc_p, heap) *locs = NULL; | |
2120 | char *str = get_lsm_tmp_name (ref->mem, ~0); | |
2121 | ||
2122 | lsm_tmp_name_add ("_flag"); | |
7cc434a3 | 2123 | flag = create_tmp_reg (boolean_type_node, str); |
039496da AH |
2124 | get_all_locs_in_loop (loop, ref, &locs); |
2125 | FOR_EACH_VEC_ELT (mem_ref_loc_p, locs, i, loc) | |
2126 | { | |
2127 | gimple_stmt_iterator gsi; | |
2128 | gimple stmt; | |
2129 | ||
2130 | gsi = gsi_for_stmt (loc->stmt); | |
2131 | stmt = gimple_build_assign (flag, boolean_true_node); | |
2132 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2133 | } | |
2134 | VEC_free (mem_ref_loc_p, heap, locs); | |
2135 | return flag; | |
2136 | } | |
2137 | ||
72425608 | 2138 | /* Executes store motion of memory reference REF from LOOP. |
ca83d385 ZD |
2139 | Exits from the LOOP are stored in EXITS. The initialization of the |
2140 | temporary variable is put to the preheader of the loop, and assignments | |
2141 | to the reference from the temporary variable are emitted to exits. */ | |
a7e5372d ZD |
2142 | |
2143 | static void | |
72425608 | 2144 | execute_sm (struct loop *loop, VEC (edge, heap) *exits, mem_ref_p ref) |
a7e5372d | 2145 | { |
039496da | 2146 | tree tmp_var, store_flag; |
a7e5372d | 2147 | unsigned i; |
039496da | 2148 | gimple load; |
b4042a03 | 2149 | struct fmt_data fmt_data; |
039496da | 2150 | edge ex, latch_edge; |
726a989a | 2151 | struct lim_aux_data *lim_data; |
039496da | 2152 | bool multi_threaded_model_p = false; |
a7e5372d | 2153 | |
a3631d97 ZD |
2154 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2155 | { | |
2156 | fprintf (dump_file, "Executing store motion of "); | |
72425608 | 2157 | print_generic_expr (dump_file, ref->mem, 0); |
a3631d97 ZD |
2158 | fprintf (dump_file, " from loop %d\n", loop->num); |
2159 | } | |
2160 | ||
7cc434a3 | 2161 | tmp_var = create_tmp_reg (TREE_TYPE (ref->mem), |
72425608 | 2162 | get_lsm_tmp_name (ref->mem, ~0)); |
a7e5372d | 2163 | |
b4042a03 ZD |
2164 | fmt_data.loop = loop; |
2165 | fmt_data.orig_loop = loop; | |
72425608 | 2166 | for_each_index (&ref->mem, force_move_till, &fmt_data); |
a7e5372d | 2167 | |
874a3589 | 2168 | if (block_in_transaction (loop_preheader_edge (loop)->src) |
039496da AH |
2169 | || !PARAM_VALUE (PARAM_ALLOW_STORE_DATA_RACES)) |
2170 | multi_threaded_model_p = true; | |
2171 | ||
2172 | if (multi_threaded_model_p) | |
2173 | store_flag = execute_sm_if_changed_flag_set (loop, ref); | |
2174 | ||
72425608 | 2175 | rewrite_mem_refs (loop, ref, tmp_var); |
a7e5372d | 2176 | |
039496da AH |
2177 | /* Emit the load code into the latch, so that we are sure it will |
2178 | be processed after all dependencies. */ | |
2179 | latch_edge = loop_latch_edge (loop); | |
2180 | ||
2181 | /* FIXME/TODO: For the multi-threaded variant, we could avoid this | |
2182 | load altogether, since the store is predicated by a flag. We | |
2183 | could, do the load only if it was originally in the loop. */ | |
726a989a RB |
2184 | load = gimple_build_assign (tmp_var, unshare_expr (ref->mem)); |
2185 | lim_data = init_lim_data (load); | |
2186 | lim_data->max_loop = loop; | |
2187 | lim_data->tgt_loop = loop; | |
039496da | 2188 | gsi_insert_on_edge (latch_edge, load); |
a7e5372d | 2189 | |
039496da | 2190 | if (multi_threaded_model_p) |
a7e5372d | 2191 | { |
039496da AH |
2192 | load = gimple_build_assign (store_flag, boolean_false_node); |
2193 | lim_data = init_lim_data (load); | |
2194 | lim_data->max_loop = loop; | |
2195 | lim_data->tgt_loop = loop; | |
2196 | gsi_insert_on_edge (latch_edge, load); | |
a7e5372d | 2197 | } |
039496da AH |
2198 | |
2199 | /* Sink the store to every exit from the loop. */ | |
2200 | FOR_EACH_VEC_ELT (edge, exits, i, ex) | |
2201 | if (!multi_threaded_model_p) | |
2202 | { | |
2203 | gimple store; | |
2204 | store = gimple_build_assign (unshare_expr (ref->mem), tmp_var); | |
2205 | gsi_insert_on_edge (ex, store); | |
2206 | } | |
2207 | else | |
2208 | execute_sm_if_changed (ex, ref->mem, tmp_var, store_flag); | |
a7e5372d ZD |
2209 | } |
2210 | ||
72425608 ZD |
2211 | /* Hoists memory references MEM_REFS out of LOOP. EXITS is the list of exit |
2212 | edges of the LOOP. */ | |
a7e5372d ZD |
2213 | |
2214 | static void | |
72425608 ZD |
2215 | hoist_memory_references (struct loop *loop, bitmap mem_refs, |
2216 | VEC (edge, heap) *exits) | |
a7e5372d | 2217 | { |
72425608 ZD |
2218 | mem_ref_p ref; |
2219 | unsigned i; | |
2220 | bitmap_iterator bi; | |
a3631d97 | 2221 | |
72425608 | 2222 | EXECUTE_IF_SET_IN_BITMAP (mem_refs, 0, i, bi) |
a7e5372d | 2223 | { |
72425608 ZD |
2224 | ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i); |
2225 | execute_sm (loop, exits, ref); | |
a7e5372d | 2226 | } |
01fd257a ZD |
2227 | } |
2228 | ||
58adb739 RG |
2229 | /* Returns true if REF is always accessed in LOOP. If STORED_P is true |
2230 | make sure REF is always stored to in LOOP. */ | |
a7e5372d ZD |
2231 | |
2232 | static bool | |
58adb739 | 2233 | ref_always_accessed_p (struct loop *loop, mem_ref_p ref, bool stored_p) |
a7e5372d | 2234 | { |
72425608 | 2235 | VEC (mem_ref_loc_p, heap) *locs = NULL; |
a7e5372d | 2236 | unsigned i; |
72425608 ZD |
2237 | mem_ref_loc_p loc; |
2238 | bool ret = false; | |
2239 | struct loop *must_exec; | |
58adb739 RG |
2240 | tree base; |
2241 | ||
2242 | base = get_base_address (ref->mem); | |
70f34814 RG |
2243 | if (INDIRECT_REF_P (base) |
2244 | || TREE_CODE (base) == MEM_REF) | |
58adb739 | 2245 | base = TREE_OPERAND (base, 0); |
a7e5372d | 2246 | |
72425608 | 2247 | get_all_locs_in_loop (loop, ref, &locs); |
ac47786e | 2248 | FOR_EACH_VEC_ELT (mem_ref_loc_p, locs, i, loc) |
72425608 | 2249 | { |
726a989a | 2250 | if (!get_lim_data (loc->stmt)) |
72425608 | 2251 | continue; |
a7e5372d | 2252 | |
58adb739 RG |
2253 | /* If we require an always executed store make sure the statement |
2254 | stores to the reference. */ | |
2255 | if (stored_p) | |
2256 | { | |
2257 | tree lhs; | |
2258 | if (!gimple_get_lhs (loc->stmt)) | |
2259 | continue; | |
2260 | lhs = get_base_address (gimple_get_lhs (loc->stmt)); | |
2261 | if (!lhs) | |
2262 | continue; | |
70f34814 RG |
2263 | if (INDIRECT_REF_P (lhs) |
2264 | || TREE_CODE (lhs) == MEM_REF) | |
58adb739 RG |
2265 | lhs = TREE_OPERAND (lhs, 0); |
2266 | if (lhs != base) | |
2267 | continue; | |
2268 | } | |
2269 | ||
726a989a | 2270 | must_exec = get_lim_data (loc->stmt)->always_executed_in; |
72425608 ZD |
2271 | if (!must_exec) |
2272 | continue; | |
a7e5372d | 2273 | |
72425608 ZD |
2274 | if (must_exec == loop |
2275 | || flow_loop_nested_p (must_exec, loop)) | |
2276 | { | |
2277 | ret = true; | |
2278 | break; | |
2279 | } | |
2280 | } | |
2281 | VEC_free (mem_ref_loc_p, heap, locs); | |
01fd257a | 2282 | |
72425608 | 2283 | return ret; |
01fd257a ZD |
2284 | } |
2285 | ||
72425608 | 2286 | /* Returns true if REF1 and REF2 are independent. */ |
01fd257a | 2287 | |
72425608 ZD |
2288 | static bool |
2289 | refs_independent_p (mem_ref_p ref1, mem_ref_p ref2) | |
01fd257a | 2290 | { |
72425608 ZD |
2291 | if (ref1 == ref2 |
2292 | || bitmap_bit_p (ref1->indep_ref, ref2->id)) | |
2293 | return true; | |
2294 | if (bitmap_bit_p (ref1->dep_ref, ref2->id)) | |
2295 | return false; | |
546d314c RG |
2296 | if (!MEM_ANALYZABLE (ref1) |
2297 | || !MEM_ANALYZABLE (ref2)) | |
2298 | return false; | |
01fd257a | 2299 | |
72425608 ZD |
2300 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2301 | fprintf (dump_file, "Querying dependency of refs %u and %u: ", | |
2302 | ref1->id, ref2->id); | |
2303 | ||
2304 | if (mem_refs_may_alias_p (ref1->mem, ref2->mem, | |
2305 | &memory_accesses.ttae_cache)) | |
2306 | { | |
2307 | bitmap_set_bit (ref1->dep_ref, ref2->id); | |
2308 | bitmap_set_bit (ref2->dep_ref, ref1->id); | |
2309 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2310 | fprintf (dump_file, "dependent.\n"); | |
2311 | return false; | |
2312 | } | |
2313 | else | |
2314 | { | |
2315 | bitmap_set_bit (ref1->indep_ref, ref2->id); | |
2316 | bitmap_set_bit (ref2->indep_ref, ref1->id); | |
2317 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2318 | fprintf (dump_file, "independent.\n"); | |
2319 | return true; | |
2320 | } | |
01fd257a ZD |
2321 | } |
2322 | ||
72425608 ZD |
2323 | /* Records the information whether REF is independent in LOOP (according |
2324 | to INDEP). */ | |
01fd257a ZD |
2325 | |
2326 | static void | |
72425608 | 2327 | record_indep_loop (struct loop *loop, mem_ref_p ref, bool indep) |
01fd257a | 2328 | { |
72425608 ZD |
2329 | if (indep) |
2330 | bitmap_set_bit (ref->indep_loop, loop->num); | |
2331 | else | |
2332 | bitmap_set_bit (ref->dep_loop, loop->num); | |
2333 | } | |
01fd257a | 2334 | |
72425608 ZD |
2335 | /* Returns true if REF is independent on all other memory references in |
2336 | LOOP. */ | |
01fd257a | 2337 | |
72425608 ZD |
2338 | static bool |
2339 | ref_indep_loop_p_1 (struct loop *loop, mem_ref_p ref) | |
2340 | { | |
546d314c | 2341 | bitmap refs_to_check; |
72425608 ZD |
2342 | unsigned i; |
2343 | bitmap_iterator bi; | |
2344 | bool ret = true, stored = bitmap_bit_p (ref->stored, loop->num); | |
72425608 ZD |
2345 | mem_ref_p aref; |
2346 | ||
546d314c RG |
2347 | if (stored) |
2348 | refs_to_check = VEC_index (bitmap, | |
2349 | memory_accesses.all_refs_in_loop, loop->num); | |
2350 | else | |
2351 | refs_to_check = VEC_index (bitmap, | |
2352 | memory_accesses.all_refs_stored_in_loop, | |
2353 | loop->num); | |
01fd257a | 2354 | |
72425608 | 2355 | EXECUTE_IF_SET_IN_BITMAP (refs_to_check, 0, i, bi) |
01fd257a | 2356 | { |
72425608 | 2357 | aref = VEC_index (mem_ref_p, memory_accesses.refs_list, i); |
546d314c RG |
2358 | if (!MEM_ANALYZABLE (aref) |
2359 | || !refs_independent_p (ref, aref)) | |
72425608 ZD |
2360 | { |
2361 | ret = false; | |
2362 | record_indep_loop (loop, aref, false); | |
2363 | break; | |
2364 | } | |
01fd257a | 2365 | } |
01fd257a | 2366 | |
72425608 | 2367 | return ret; |
01fd257a ZD |
2368 | } |
2369 | ||
72425608 ZD |
2370 | /* Returns true if REF is independent on all other memory references in |
2371 | LOOP. Wrapper over ref_indep_loop_p_1, caching its results. */ | |
01fd257a | 2372 | |
72425608 ZD |
2373 | static bool |
2374 | ref_indep_loop_p (struct loop *loop, mem_ref_p ref) | |
01fd257a | 2375 | { |
72425608 | 2376 | bool ret; |
01fd257a | 2377 | |
72425608 ZD |
2378 | if (bitmap_bit_p (ref->indep_loop, loop->num)) |
2379 | return true; | |
2380 | if (bitmap_bit_p (ref->dep_loop, loop->num)) | |
2381 | return false; | |
01fd257a | 2382 | |
72425608 | 2383 | ret = ref_indep_loop_p_1 (loop, ref); |
ed9c043b | 2384 | |
72425608 ZD |
2385 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2386 | fprintf (dump_file, "Querying dependencies of ref %u in loop %d: %s\n", | |
2387 | ref->id, loop->num, ret ? "independent" : "dependent"); | |
2388 | ||
2389 | record_indep_loop (loop, ref, ret); | |
2390 | ||
2391 | return ret; | |
01fd257a ZD |
2392 | } |
2393 | ||
72425608 | 2394 | /* Returns true if we can perform store motion of REF from LOOP. */ |
01fd257a | 2395 | |
72425608 ZD |
2396 | static bool |
2397 | can_sm_ref_p (struct loop *loop, mem_ref_p ref) | |
01fd257a | 2398 | { |
58adb739 RG |
2399 | tree base; |
2400 | ||
546d314c RG |
2401 | /* Can't hoist unanalyzable refs. */ |
2402 | if (!MEM_ANALYZABLE (ref)) | |
2403 | return false; | |
2404 | ||
72425608 ZD |
2405 | /* Unless the reference is stored in the loop, there is nothing to do. */ |
2406 | if (!bitmap_bit_p (ref->stored, loop->num)) | |
2407 | return false; | |
01fd257a | 2408 | |
72425608 ZD |
2409 | /* It should be movable. */ |
2410 | if (!is_gimple_reg_type (TREE_TYPE (ref->mem)) | |
2411 | || TREE_THIS_VOLATILE (ref->mem) | |
2412 | || !for_each_index (&ref->mem, may_move_till, loop)) | |
2413 | return false; | |
ed9c043b | 2414 | |
9939e416 RG |
2415 | /* If it can throw fail, we do not properly update EH info. */ |
2416 | if (tree_could_throw_p (ref->mem)) | |
2417 | return false; | |
2418 | ||
58adb739 RG |
2419 | /* If it can trap, it must be always executed in LOOP. |
2420 | Readonly memory locations may trap when storing to them, but | |
2421 | tree_could_trap_p is a predicate for rvalues, so check that | |
2422 | explicitly. */ | |
2423 | base = get_base_address (ref->mem); | |
2424 | if ((tree_could_trap_p (ref->mem) | |
2425 | || (DECL_P (base) && TREE_READONLY (base))) | |
2426 | && !ref_always_accessed_p (loop, ref, true)) | |
72425608 | 2427 | return false; |
ed9c043b | 2428 | |
72425608 ZD |
2429 | /* And it must be independent on all other memory references |
2430 | in LOOP. */ | |
2431 | if (!ref_indep_loop_p (loop, ref)) | |
2432 | return false; | |
ed9c043b | 2433 | |
72425608 | 2434 | return true; |
ed9c043b ZD |
2435 | } |
2436 | ||
72425608 ZD |
2437 | /* Marks the references in LOOP for that store motion should be performed |
2438 | in REFS_TO_SM. SM_EXECUTED is the set of references for that store | |
2439 | motion was performed in one of the outer loops. */ | |
ed9c043b ZD |
2440 | |
2441 | static void | |
72425608 | 2442 | find_refs_for_sm (struct loop *loop, bitmap sm_executed, bitmap refs_to_sm) |
01fd257a | 2443 | { |
72425608 ZD |
2444 | bitmap refs = VEC_index (bitmap, memory_accesses.all_refs_in_loop, |
2445 | loop->num); | |
2446 | unsigned i; | |
2447 | bitmap_iterator bi; | |
2448 | mem_ref_p ref; | |
2449 | ||
2450 | EXECUTE_IF_AND_COMPL_IN_BITMAP (refs, sm_executed, 0, i, bi) | |
2451 | { | |
2452 | ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i); | |
2453 | if (can_sm_ref_p (loop, ref)) | |
2454 | bitmap_set_bit (refs_to_sm, i); | |
2455 | } | |
ed9c043b | 2456 | } |
01fd257a | 2457 | |
72425608 ZD |
2458 | /* Checks whether LOOP (with exits stored in EXITS array) is suitable |
2459 | for a store motion optimization (i.e. whether we can insert statement | |
2460 | on its exits). */ | |
ed9c043b | 2461 | |
72425608 ZD |
2462 | static bool |
2463 | loop_suitable_for_sm (struct loop *loop ATTRIBUTE_UNUSED, | |
2464 | VEC (edge, heap) *exits) | |
ed9c043b | 2465 | { |
72425608 ZD |
2466 | unsigned i; |
2467 | edge ex; | |
01fd257a | 2468 | |
ac47786e | 2469 | FOR_EACH_VEC_ELT (edge, exits, i, ex) |
6391db68 | 2470 | if (ex->flags & (EDGE_ABNORMAL | EDGE_EH)) |
72425608 ZD |
2471 | return false; |
2472 | ||
2473 | return true; | |
01fd257a ZD |
2474 | } |
2475 | ||
a7e5372d | 2476 | /* Try to perform store motion for all memory references modified inside |
72425608 ZD |
2477 | LOOP. SM_EXECUTED is the bitmap of the memory references for that |
2478 | store motion was executed in one of the outer loops. */ | |
a7e5372d ZD |
2479 | |
2480 | static void | |
72425608 | 2481 | store_motion_loop (struct loop *loop, bitmap sm_executed) |
a7e5372d | 2482 | { |
ca83d385 | 2483 | VEC (edge, heap) *exits = get_loop_exit_edges (loop); |
72425608 ZD |
2484 | struct loop *subloop; |
2485 | bitmap sm_in_loop = BITMAP_ALLOC (NULL); | |
a7e5372d | 2486 | |
72425608 | 2487 | if (loop_suitable_for_sm (loop, exits)) |
a7e5372d | 2488 | { |
72425608 ZD |
2489 | find_refs_for_sm (loop, sm_executed, sm_in_loop); |
2490 | hoist_memory_references (loop, sm_in_loop, exits); | |
a7e5372d | 2491 | } |
ca83d385 | 2492 | VEC_free (edge, heap, exits); |
72425608 ZD |
2493 | |
2494 | bitmap_ior_into (sm_executed, sm_in_loop); | |
2495 | for (subloop = loop->inner; subloop != NULL; subloop = subloop->next) | |
2496 | store_motion_loop (subloop, sm_executed); | |
2497 | bitmap_and_compl_into (sm_executed, sm_in_loop); | |
2498 | BITMAP_FREE (sm_in_loop); | |
a7e5372d ZD |
2499 | } |
2500 | ||
2501 | /* Try to perform store motion for all memory references modified inside | |
d73be268 | 2502 | loops. */ |
a7e5372d ZD |
2503 | |
2504 | static void | |
72425608 | 2505 | store_motion (void) |
a7e5372d ZD |
2506 | { |
2507 | struct loop *loop; | |
72425608 | 2508 | bitmap sm_executed = BITMAP_ALLOC (NULL); |
d16464bb | 2509 | |
72425608 ZD |
2510 | for (loop = current_loops->tree_root->inner; loop != NULL; loop = loop->next) |
2511 | store_motion_loop (loop, sm_executed); | |
42fd6772 | 2512 | |
72425608 | 2513 | BITMAP_FREE (sm_executed); |
726a989a | 2514 | gsi_commit_edge_inserts (); |
a7e5372d ZD |
2515 | } |
2516 | ||
2517 | /* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e. | |
2518 | for each such basic block bb records the outermost loop for that execution | |
2519 | of its header implies execution of bb. CONTAINS_CALL is the bitmap of | |
2520 | blocks that contain a nonpure call. */ | |
2521 | ||
2522 | static void | |
2523 | fill_always_executed_in (struct loop *loop, sbitmap contains_call) | |
2524 | { | |
2525 | basic_block bb = NULL, *bbs, last = NULL; | |
2526 | unsigned i; | |
2527 | edge e; | |
2528 | struct loop *inn_loop = loop; | |
2529 | ||
8a519095 | 2530 | if (ALWAYS_EXECUTED_IN (loop->header) == NULL) |
a7e5372d ZD |
2531 | { |
2532 | bbs = get_loop_body_in_dom_order (loop); | |
2533 | ||
2534 | for (i = 0; i < loop->num_nodes; i++) | |
2535 | { | |
628f6a4e | 2536 | edge_iterator ei; |
a7e5372d ZD |
2537 | bb = bbs[i]; |
2538 | ||
2539 | if (dominated_by_p (CDI_DOMINATORS, loop->latch, bb)) | |
2540 | last = bb; | |
2541 | ||
2542 | if (TEST_BIT (contains_call, bb->index)) | |
2543 | break; | |
2544 | ||
628f6a4e | 2545 | FOR_EACH_EDGE (e, ei, bb->succs) |
a7e5372d ZD |
2546 | if (!flow_bb_inside_loop_p (loop, e->dest)) |
2547 | break; | |
2548 | if (e) | |
2549 | break; | |
2550 | ||
2551 | /* A loop might be infinite (TODO use simple loop analysis | |
2552 | to disprove this if possible). */ | |
2553 | if (bb->flags & BB_IRREDUCIBLE_LOOP) | |
2554 | break; | |
2555 | ||
2556 | if (!flow_bb_inside_loop_p (inn_loop, bb)) | |
2557 | break; | |
2558 | ||
2559 | if (bb->loop_father->header == bb) | |
2560 | { | |
2561 | if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb)) | |
2562 | break; | |
2563 | ||
2564 | /* In a loop that is always entered we may proceed anyway. | |
2565 | But record that we entered it and stop once we leave it. */ | |
2566 | inn_loop = bb->loop_father; | |
2567 | } | |
2568 | } | |
2569 | ||
2570 | while (1) | |
2571 | { | |
8a519095 | 2572 | SET_ALWAYS_EXECUTED_IN (last, loop); |
a7e5372d ZD |
2573 | if (last == loop->header) |
2574 | break; | |
2575 | last = get_immediate_dominator (CDI_DOMINATORS, last); | |
2576 | } | |
2577 | ||
2578 | free (bbs); | |
2579 | } | |
2580 | ||
2581 | for (loop = loop->inner; loop; loop = loop->next) | |
2582 | fill_always_executed_in (loop, contains_call); | |
2583 | } | |
2584 | ||
d73be268 | 2585 | /* Compute the global information needed by the loop invariant motion pass. */ |
a7e5372d ZD |
2586 | |
2587 | static void | |
d73be268 | 2588 | tree_ssa_lim_initialize (void) |
a7e5372d ZD |
2589 | { |
2590 | sbitmap contains_call = sbitmap_alloc (last_basic_block); | |
726a989a | 2591 | gimple_stmt_iterator bsi; |
a7e5372d ZD |
2592 | struct loop *loop; |
2593 | basic_block bb; | |
2594 | ||
2595 | sbitmap_zero (contains_call); | |
2596 | FOR_EACH_BB (bb) | |
2597 | { | |
726a989a | 2598 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
a7e5372d | 2599 | { |
726a989a | 2600 | if (nonpure_call_p (gsi_stmt (bsi))) |
a7e5372d ZD |
2601 | break; |
2602 | } | |
2603 | ||
726a989a | 2604 | if (!gsi_end_p (bsi)) |
a7e5372d ZD |
2605 | SET_BIT (contains_call, bb->index); |
2606 | } | |
2607 | ||
d73be268 | 2608 | for (loop = current_loops->tree_root->inner; loop; loop = loop->next) |
a7e5372d ZD |
2609 | fill_always_executed_in (loop, contains_call); |
2610 | ||
2611 | sbitmap_free (contains_call); | |
726a989a RB |
2612 | |
2613 | lim_aux_data_map = pointer_map_create (); | |
19c0d7df AH |
2614 | |
2615 | if (flag_tm) | |
2616 | compute_transaction_bits (); | |
039496da AH |
2617 | |
2618 | alloc_aux_for_edges (0); | |
a7e5372d ZD |
2619 | } |
2620 | ||
2621 | /* Cleans up after the invariant motion pass. */ | |
2622 | ||
2623 | static void | |
2624 | tree_ssa_lim_finalize (void) | |
2625 | { | |
2626 | basic_block bb; | |
72425608 ZD |
2627 | unsigned i; |
2628 | bitmap b; | |
a7e5372d | 2629 | |
039496da AH |
2630 | free_aux_for_edges (); |
2631 | ||
a7e5372d | 2632 | FOR_EACH_BB (bb) |
8a519095 | 2633 | SET_ALWAYS_EXECUTED_IN (bb, NULL); |
72425608 | 2634 | |
726a989a RB |
2635 | pointer_map_destroy (lim_aux_data_map); |
2636 | ||
72425608 ZD |
2637 | VEC_free (mem_ref_p, heap, memory_accesses.refs_list); |
2638 | htab_delete (memory_accesses.refs); | |
2639 | ||
ac47786e | 2640 | FOR_EACH_VEC_ELT (bitmap, memory_accesses.refs_in_loop, i, b) |
72425608 ZD |
2641 | BITMAP_FREE (b); |
2642 | VEC_free (bitmap, heap, memory_accesses.refs_in_loop); | |
2643 | ||
ac47786e | 2644 | FOR_EACH_VEC_ELT (bitmap, memory_accesses.all_refs_in_loop, i, b) |
72425608 ZD |
2645 | BITMAP_FREE (b); |
2646 | VEC_free (bitmap, heap, memory_accesses.all_refs_in_loop); | |
2647 | ||
546d314c | 2648 | FOR_EACH_VEC_ELT (bitmap, memory_accesses.all_refs_stored_in_loop, i, b) |
72425608 | 2649 | BITMAP_FREE (b); |
546d314c | 2650 | VEC_free (bitmap, heap, memory_accesses.all_refs_stored_in_loop); |
72425608 ZD |
2651 | |
2652 | if (memory_accesses.ttae_cache) | |
2653 | pointer_map_destroy (memory_accesses.ttae_cache); | |
a7e5372d ZD |
2654 | } |
2655 | ||
d73be268 | 2656 | /* Moves invariants from loops. Only "expensive" invariants are moved out -- |
a7e5372d ZD |
2657 | i.e. those that are likely to be win regardless of the register pressure. */ |
2658 | ||
e3bdfed6 | 2659 | unsigned int |
d73be268 | 2660 | tree_ssa_lim (void) |
a7e5372d | 2661 | { |
e3bdfed6 RG |
2662 | unsigned int todo; |
2663 | ||
d73be268 | 2664 | tree_ssa_lim_initialize (); |
a7e5372d | 2665 | |
72425608 ZD |
2666 | /* Gathers information about memory accesses in the loops. */ |
2667 | analyze_memory_references (); | |
2668 | ||
a7e5372d ZD |
2669 | /* For each statement determine the outermost loop in that it is |
2670 | invariant and cost for computing the invariant. */ | |
2671 | determine_invariantness (); | |
2672 | ||
72425608 ZD |
2673 | /* Execute store motion. Force the necessary invariants to be moved |
2674 | out of the loops as well. */ | |
2675 | store_motion (); | |
a7e5372d ZD |
2676 | |
2677 | /* Move the expressions that are expensive enough. */ | |
e3bdfed6 | 2678 | todo = move_computations (); |
a7e5372d ZD |
2679 | |
2680 | tree_ssa_lim_finalize (); | |
e3bdfed6 RG |
2681 | |
2682 | return todo; | |
a7e5372d | 2683 | } |