]>
Commit | Line | Data |
---|---|---|
0516f6fe | 1 | /* Post reload partially redundant load elimination |
5624e564 | 2 | Copyright (C) 2004-2015 Free Software Foundation, Inc. |
0516f6fe SB |
3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 8 | Software Foundation; either version 3, or (at your option) any later |
0516f6fe SB |
9 | version. |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
0516f6fe SB |
19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
c7131fb2 AM |
23 | #include "backend.h" |
24 | #include "tree.h" | |
25 | #include "rtl.h" | |
26 | #include "df.h" | |
718f9c0f | 27 | #include "diagnostic-core.h" |
0516f6fe | 28 | |
40e23961 | 29 | #include "alias.h" |
0516f6fe SB |
30 | #include "tm_p.h" |
31 | #include "regs.h" | |
0516f6fe | 32 | #include "flags.h" |
0516f6fe SB |
33 | #include "insn-config.h" |
34 | #include "recog.h" | |
60393bbc | 35 | #include "cfgrtl.h" |
59f2e9d8 | 36 | #include "profile.h" |
36566b39 PK |
37 | #include "expmed.h" |
38 | #include "dojump.h" | |
39 | #include "explow.h" | |
40 | #include "calls.h" | |
41 | #include "emit-rtl.h" | |
42 | #include "varasm.h" | |
43 | #include "stmt.h" | |
0516f6fe SB |
44 | #include "expr.h" |
45 | #include "except.h" | |
46 | #include "intl.h" | |
47 | #include "obstack.h" | |
0516f6fe | 48 | #include "params.h" |
3f55b339 | 49 | #include "target.h" |
ef330312 | 50 | #include "tree-pass.h" |
6fb5fa3c | 51 | #include "dbgcnt.h" |
af3eb110 | 52 | #include "gcse-common.h" |
0516f6fe SB |
53 | |
54 | /* The following code implements gcse after reload, the purpose of this | |
55 | pass is to cleanup redundant loads generated by reload and other | |
56 | optimizations that come after gcse. It searches for simple inter-block | |
57 | redundancies and tries to eliminate them by adding moves and loads | |
58 | in cold places. | |
59 | ||
60 | Perform partially redundant load elimination, try to eliminate redundant | |
61 | loads created by the reload pass. We try to look for full or partial | |
62 | redundant loads fed by one or more loads/stores in predecessor BBs, | |
63 | and try adding loads to make them fully redundant. We also check if | |
64 | it's worth adding loads to be able to delete the redundant load. | |
65 | ||
66 | Algorithm: | |
67 | 1. Build available expressions hash table: | |
68 | For each load/store instruction, if the loaded/stored memory didn't | |
69 | change until the end of the basic block add this memory expression to | |
70 | the hash table. | |
71 | 2. Perform Redundancy elimination: | |
72 | For each load instruction do the following: | |
73 | perform partial redundancy elimination, check if it's worth adding | |
74 | loads to make the load fully redundant. If so add loads and | |
75 | register copies and delete the load. | |
76 | 3. Delete instructions made redundant in step 2. | |
77 | ||
78 | Future enhancement: | |
79 | If the loaded register is used/defined between load and some store, | |
80 | look for some other free register between load and all its stores, | |
81 | and replace the load with a copy from this register to the loaded | |
82 | register. | |
83 | */ | |
84 | \f | |
85 | ||
86 | /* Keep statistics of this pass. */ | |
87 | static struct | |
88 | { | |
89 | int moves_inserted; | |
90 | int copies_inserted; | |
91 | int insns_deleted; | |
92 | } stats; | |
93 | ||
94 | /* We need to keep a hash table of expressions. The table entries are of | |
95 | type 'struct expr', and for each expression there is a single linked | |
2a7e31df | 96 | list of occurrences. */ |
0516f6fe | 97 | |
0516f6fe SB |
98 | /* Expression elements in the hash table. */ |
99 | struct expr | |
100 | { | |
101 | /* The expression (SET_SRC for expressions, PATTERN for assignments). */ | |
102 | rtx expr; | |
103 | ||
104 | /* The same hash for this entry. */ | |
105 | hashval_t hash; | |
106 | ||
af3eb110 JL |
107 | /* Index in the transparent bitmaps. */ |
108 | unsigned int bitmap_index; | |
109 | ||
0516f6fe SB |
110 | /* List of available occurrence in basic blocks in the function. */ |
111 | struct occr *avail_occr; | |
112 | }; | |
113 | ||
4a8fb1a1 LC |
114 | /* Hashtable helpers. */ |
115 | ||
8d67ee55 | 116 | struct expr_hasher : nofree_ptr_hash <expr> |
4a8fb1a1 | 117 | { |
67f58944 TS |
118 | static inline hashval_t hash (const expr *); |
119 | static inline bool equal (const expr *, const expr *); | |
4a8fb1a1 LC |
120 | }; |
121 | ||
122 | ||
123 | /* Hash expression X. | |
124 | DO_NOT_RECORD_P is a boolean indicating if a volatile operand is found | |
125 | or if the expression contains something we don't want to insert in the | |
126 | table. */ | |
127 | ||
128 | static hashval_t | |
129 | hash_expr (rtx x, int *do_not_record_p) | |
130 | { | |
131 | *do_not_record_p = 0; | |
132 | return hash_rtx (x, GET_MODE (x), do_not_record_p, | |
133 | NULL, /*have_reg_qty=*/false); | |
134 | } | |
135 | ||
136 | /* Callback for hashtab. | |
137 | Return the hash value for expression EXP. We don't actually hash | |
138 | here, we just return the cached hash value. */ | |
139 | ||
140 | inline hashval_t | |
67f58944 | 141 | expr_hasher::hash (const expr *exp) |
4a8fb1a1 LC |
142 | { |
143 | return exp->hash; | |
144 | } | |
145 | ||
146 | /* Callback for hashtab. | |
147 | Return nonzero if exp1 is equivalent to exp2. */ | |
148 | ||
149 | inline bool | |
67f58944 | 150 | expr_hasher::equal (const expr *exp1, const expr *exp2) |
4a8fb1a1 LC |
151 | { |
152 | int equiv_p = exp_equiv_p (exp1->expr, exp2->expr, 0, true); | |
153 | ||
154 | gcc_assert (!equiv_p || exp1->hash == exp2->hash); | |
155 | return equiv_p; | |
156 | } | |
157 | ||
158 | /* The table itself. */ | |
c203e8a7 | 159 | static hash_table<expr_hasher> *expr_table; |
4a8fb1a1 LC |
160 | \f |
161 | ||
0516f6fe SB |
162 | static struct obstack expr_obstack; |
163 | ||
164 | /* Occurrence of an expression. | |
2a7e31df | 165 | There is at most one occurrence per basic block. If a pattern appears |
0516f6fe SB |
166 | more than once, the last appearance is used. */ |
167 | ||
168 | struct occr | |
169 | { | |
170 | /* Next occurrence of this expression. */ | |
171 | struct occr *next; | |
172 | /* The insn that computes the expression. */ | |
6c6d76be | 173 | rtx_insn *insn; |
0516f6fe SB |
174 | /* Nonzero if this [anticipatable] occurrence has been deleted. */ |
175 | char deleted_p; | |
176 | }; | |
177 | ||
178 | static struct obstack occr_obstack; | |
179 | ||
180 | /* The following structure holds the information about the occurrences of | |
181 | the redundant instructions. */ | |
182 | struct unoccr | |
183 | { | |
184 | struct unoccr *next; | |
185 | edge pred; | |
6c6d76be | 186 | rtx_insn *insn; |
0516f6fe SB |
187 | }; |
188 | ||
189 | static struct obstack unoccr_obstack; | |
190 | ||
191 | /* Array where each element is the CUID if the insn that last set the hard | |
192 | register with the number of the element, since the start of the current | |
c93320c4 SB |
193 | basic block. |
194 | ||
195 | This array is used during the building of the hash table (step 1) to | |
196 | determine if a reg is killed before the end of a basic block. | |
197 | ||
198 | It is also used when eliminating partial redundancies (step 2) to see | |
199 | if a reg was modified since the start of a basic block. */ | |
0516f6fe SB |
200 | static int *reg_avail_info; |
201 | ||
202 | /* A list of insns that may modify memory within the current basic block. */ | |
203 | struct modifies_mem | |
204 | { | |
6c6d76be | 205 | rtx_insn *insn; |
0516f6fe SB |
206 | struct modifies_mem *next; |
207 | }; | |
208 | static struct modifies_mem *modifies_mem_list; | |
209 | ||
210 | /* The modifies_mem structs also go on an obstack, only this obstack is | |
211 | freed each time after completing the analysis or transformations on | |
212 | a basic block. So we allocate a dummy modifies_mem_obstack_bottom | |
213 | object on the obstack to keep track of the bottom of the obstack. */ | |
214 | static struct obstack modifies_mem_obstack; | |
215 | static struct modifies_mem *modifies_mem_obstack_bottom; | |
216 | ||
217 | /* Mapping of insn UIDs to CUIDs. | |
218 | CUIDs are like UIDs except they increase monotonically in each basic | |
219 | block, have no gaps, and only apply to real insns. */ | |
220 | static int *uid_cuid; | |
221 | #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)]) | |
af3eb110 JL |
222 | |
223 | /* Bitmap of blocks which have memory stores. */ | |
224 | static bitmap modify_mem_list_set; | |
225 | ||
226 | /* Bitmap of blocks which have calls. */ | |
227 | static bitmap blocks_with_calls; | |
228 | ||
229 | /* Vector indexed by block # with a list of all the insns that | |
230 | modify memory within the block. */ | |
231 | static vec<rtx_insn *> *modify_mem_list; | |
232 | ||
233 | /* Vector indexed by block # with a canonicalized list of insns | |
234 | that modify memory in the block. */ | |
235 | static vec<modify_pair> *canon_modify_mem_list; | |
236 | ||
237 | /* Vector of simple bitmaps indexed by block number. Each component sbitmap | |
238 | indicates which expressions are transparent through the block. */ | |
239 | static sbitmap *transp; | |
0516f6fe SB |
240 | \f |
241 | ||
242 | /* Helpers for memory allocation/freeing. */ | |
243 | static void alloc_mem (void); | |
244 | static void free_mem (void); | |
245 | ||
246 | /* Support for hash table construction and transformations. */ | |
6c6d76be DM |
247 | static bool oprs_unchanged_p (rtx, rtx_insn *, bool); |
248 | static void record_last_reg_set_info (rtx_insn *, rtx); | |
249 | static void record_last_reg_set_info_regno (rtx_insn *, int); | |
250 | static void record_last_mem_set_info (rtx_insn *); | |
7bc980e1 | 251 | static void record_last_set_info (rtx, const_rtx, void *); |
6c6d76be | 252 | static void record_opr_changes (rtx_insn *); |
0516f6fe | 253 | |
7bc980e1 | 254 | static void find_mem_conflicts (rtx, const_rtx, void *); |
0516f6fe SB |
255 | static int load_killed_in_block_p (int, rtx, bool); |
256 | static void reset_opr_set_tables (void); | |
257 | ||
258 | /* Hash table support. */ | |
259 | static hashval_t hash_expr (rtx, int *); | |
6c6d76be | 260 | static void insert_expr_in_table (rtx, rtx_insn *); |
0516f6fe | 261 | static struct expr *lookup_expr_in_table (rtx); |
0516f6fe SB |
262 | static void dump_hash_table (FILE *); |
263 | ||
264 | /* Helpers for eliminate_partially_redundant_load. */ | |
265 | static bool reg_killed_on_edge (rtx, edge); | |
266 | static bool reg_used_on_edge (rtx, edge); | |
267 | ||
6c6d76be | 268 | static rtx get_avail_load_store_reg (rtx_insn *); |
0516f6fe SB |
269 | |
270 | static bool bb_has_well_behaved_predecessors (basic_block); | |
af3eb110 | 271 | static struct occr* get_bb_avail_insn (basic_block, struct occr *, int); |
6c6d76be | 272 | static void hash_scan_set (rtx_insn *); |
0516f6fe SB |
273 | static void compute_hash_table (void); |
274 | ||
275 | /* The work horses of this pass. */ | |
276 | static void eliminate_partially_redundant_load (basic_block, | |
6c6d76be | 277 | rtx_insn *, |
0516f6fe SB |
278 | struct expr *); |
279 | static void eliminate_partially_redundant_loads (void); | |
280 | \f | |
281 | ||
282 | /* Allocate memory for the CUID mapping array and register/memory | |
283 | tracking tables. */ | |
284 | ||
285 | static void | |
286 | alloc_mem (void) | |
287 | { | |
288 | int i; | |
289 | basic_block bb; | |
6c6d76be | 290 | rtx_insn *insn; |
0516f6fe SB |
291 | |
292 | /* Find the largest UID and create a mapping from UIDs to CUIDs. */ | |
5ed6ace5 | 293 | uid_cuid = XCNEWVEC (int, get_max_uid () + 1); |
576a4795 | 294 | i = 1; |
11cd3bed | 295 | FOR_EACH_BB_FN (bb, cfun) |
0516f6fe SB |
296 | FOR_BB_INSNS (bb, insn) |
297 | { | |
298 | if (INSN_P (insn)) | |
299 | uid_cuid[INSN_UID (insn)] = i++; | |
300 | else | |
301 | uid_cuid[INSN_UID (insn)] = i; | |
302 | } | |
303 | ||
304 | /* Allocate the available expressions hash table. We don't want to | |
305 | make the hash table too small, but unnecessarily making it too large | |
306 | also doesn't help. The i/4 is a gcse.c relic, and seems like a | |
307 | reasonable choice. */ | |
c203e8a7 | 308 | expr_table = new hash_table<expr_hasher> (MAX (i / 4, 13)); |
0516f6fe SB |
309 | |
310 | /* We allocate everything on obstacks because we often can roll back | |
311 | the whole obstack to some point. Freeing obstacks is very fast. */ | |
312 | gcc_obstack_init (&expr_obstack); | |
313 | gcc_obstack_init (&occr_obstack); | |
314 | gcc_obstack_init (&unoccr_obstack); | |
315 | gcc_obstack_init (&modifies_mem_obstack); | |
316 | ||
317 | /* Working array used to track the last set for each register | |
318 | in the current block. */ | |
319 | reg_avail_info = (int *) xmalloc (FIRST_PSEUDO_REGISTER * sizeof (int)); | |
320 | ||
321 | /* Put a dummy modifies_mem object on the modifies_mem_obstack, so we | |
322 | can roll it back in reset_opr_set_tables. */ | |
323 | modifies_mem_obstack_bottom = | |
324 | (struct modifies_mem *) obstack_alloc (&modifies_mem_obstack, | |
325 | sizeof (struct modifies_mem)); | |
af3eb110 JL |
326 | |
327 | blocks_with_calls = BITMAP_ALLOC (NULL); | |
328 | modify_mem_list_set = BITMAP_ALLOC (NULL); | |
329 | ||
330 | modify_mem_list = (vec_rtx_heap *) xcalloc (last_basic_block_for_fn (cfun), | |
331 | sizeof (vec_rtx_heap)); | |
332 | canon_modify_mem_list | |
333 | = (vec_modify_pair_heap *) xcalloc (last_basic_block_for_fn (cfun), | |
334 | sizeof (vec_modify_pair_heap)); | |
0516f6fe SB |
335 | } |
336 | ||
337 | /* Free memory allocated by alloc_mem. */ | |
338 | ||
339 | static void | |
340 | free_mem (void) | |
341 | { | |
342 | free (uid_cuid); | |
343 | ||
c203e8a7 TS |
344 | delete expr_table; |
345 | expr_table = NULL; | |
0516f6fe SB |
346 | |
347 | obstack_free (&expr_obstack, NULL); | |
348 | obstack_free (&occr_obstack, NULL); | |
349 | obstack_free (&unoccr_obstack, NULL); | |
350 | obstack_free (&modifies_mem_obstack, NULL); | |
351 | ||
af3eb110 JL |
352 | unsigned i; |
353 | bitmap_iterator bi; | |
354 | EXECUTE_IF_SET_IN_BITMAP (modify_mem_list_set, 0, i, bi) | |
355 | { | |
356 | modify_mem_list[i].release (); | |
357 | canon_modify_mem_list[i].release (); | |
358 | } | |
359 | ||
360 | BITMAP_FREE (blocks_with_calls); | |
361 | BITMAP_FREE (modify_mem_list_set); | |
0516f6fe SB |
362 | free (reg_avail_info); |
363 | } | |
364 | \f | |
365 | ||
0516f6fe SB |
366 | /* Insert expression X in INSN in the hash TABLE. |
367 | If it is already present, record it as the last occurrence in INSN's | |
368 | basic block. */ | |
369 | ||
370 | static void | |
6c6d76be | 371 | insert_expr_in_table (rtx x, rtx_insn *insn) |
0516f6fe SB |
372 | { |
373 | int do_not_record_p; | |
374 | hashval_t hash; | |
375 | struct expr *cur_expr, **slot; | |
376 | struct occr *avail_occr, *last_occr = NULL; | |
377 | ||
378 | hash = hash_expr (x, &do_not_record_p); | |
379 | ||
380 | /* Do not insert expression in the table if it contains volatile operands, | |
381 | or if hash_expr determines the expression is something we don't want | |
382 | to or can't handle. */ | |
383 | if (do_not_record_p) | |
384 | return; | |
385 | ||
386 | /* We anticipate that redundant expressions are rare, so for convenience | |
387 | allocate a new hash table element here already and set its fields. | |
388 | If we don't do this, we need a hack with a static struct expr. Anyway, | |
389 | obstack_free is really fast and one more obstack_alloc doesn't hurt if | |
390 | we're going to see more expressions later on. */ | |
391 | cur_expr = (struct expr *) obstack_alloc (&expr_obstack, | |
392 | sizeof (struct expr)); | |
393 | cur_expr->expr = x; | |
394 | cur_expr->hash = hash; | |
395 | cur_expr->avail_occr = NULL; | |
396 | ||
c203e8a7 | 397 | slot = expr_table->find_slot_with_hash (cur_expr, hash, INSERT); |
b8698a0f | 398 | |
0516f6fe | 399 | if (! (*slot)) |
af3eb110 JL |
400 | { |
401 | /* The expression isn't found, so insert it. */ | |
402 | *slot = cur_expr; | |
403 | ||
404 | /* Anytime we add an entry to the table, record the index | |
405 | of the new entry. The bitmap index starts counting | |
406 | at zero. */ | |
407 | cur_expr->bitmap_index = expr_table->elements () - 1; | |
408 | } | |
0516f6fe SB |
409 | else |
410 | { | |
411 | /* The expression is already in the table, so roll back the | |
412 | obstack and use the existing table entry. */ | |
413 | obstack_free (&expr_obstack, cur_expr); | |
414 | cur_expr = *slot; | |
415 | } | |
416 | ||
417 | /* Search for another occurrence in the same basic block. */ | |
418 | avail_occr = cur_expr->avail_occr; | |
b0de17ef SB |
419 | while (avail_occr |
420 | && BLOCK_FOR_INSN (avail_occr->insn) != BLOCK_FOR_INSN (insn)) | |
0516f6fe SB |
421 | { |
422 | /* If an occurrence isn't found, save a pointer to the end of | |
423 | the list. */ | |
424 | last_occr = avail_occr; | |
425 | avail_occr = avail_occr->next; | |
426 | } | |
427 | ||
428 | if (avail_occr) | |
429 | /* Found another instance of the expression in the same basic block. | |
430 | Prefer this occurrence to the currently recorded one. We want | |
431 | the last one in the block and the block is scanned from start | |
432 | to end. */ | |
433 | avail_occr->insn = insn; | |
434 | else | |
435 | { | |
436 | /* First occurrence of this expression in this basic block. */ | |
437 | avail_occr = (struct occr *) obstack_alloc (&occr_obstack, | |
438 | sizeof (struct occr)); | |
439 | ||
440 | /* First occurrence of this expression in any block? */ | |
441 | if (cur_expr->avail_occr == NULL) | |
442 | cur_expr->avail_occr = avail_occr; | |
443 | else | |
444 | last_occr->next = avail_occr; | |
445 | ||
446 | avail_occr->insn = insn; | |
447 | avail_occr->next = NULL; | |
448 | avail_occr->deleted_p = 0; | |
449 | } | |
450 | } | |
451 | \f | |
452 | ||
453 | /* Lookup pattern PAT in the expression hash table. | |
454 | The result is a pointer to the table entry, or NULL if not found. */ | |
455 | ||
456 | static struct expr * | |
457 | lookup_expr_in_table (rtx pat) | |
458 | { | |
459 | int do_not_record_p; | |
460 | struct expr **slot, *tmp_expr; | |
461 | hashval_t hash = hash_expr (pat, &do_not_record_p); | |
462 | ||
463 | if (do_not_record_p) | |
464 | return NULL; | |
465 | ||
466 | tmp_expr = (struct expr *) obstack_alloc (&expr_obstack, | |
467 | sizeof (struct expr)); | |
468 | tmp_expr->expr = pat; | |
469 | tmp_expr->hash = hash; | |
470 | tmp_expr->avail_occr = NULL; | |
471 | ||
c203e8a7 | 472 | slot = expr_table->find_slot_with_hash (tmp_expr, hash, INSERT); |
0516f6fe SB |
473 | obstack_free (&expr_obstack, tmp_expr); |
474 | ||
475 | if (!slot) | |
476 | return NULL; | |
477 | else | |
478 | return (*slot); | |
479 | } | |
480 | \f | |
481 | ||
2a7e31df | 482 | /* Dump all expressions and occurrences that are currently in the |
0516f6fe SB |
483 | expression hash table to FILE. */ |
484 | ||
485 | /* This helper is called via htab_traverse. */ | |
4a8fb1a1 LC |
486 | int |
487 | dump_expr_hash_table_entry (expr **slot, FILE *file) | |
0516f6fe | 488 | { |
4a8fb1a1 | 489 | struct expr *exprs = *slot; |
0516f6fe SB |
490 | struct occr *occr; |
491 | ||
492 | fprintf (file, "expr: "); | |
4a8fb1a1 LC |
493 | print_rtl (file, exprs->expr); |
494 | fprintf (file,"\nhashcode: %u\n", exprs->hash); | |
cc9795d4 | 495 | fprintf (file,"list of occurrences:\n"); |
4a8fb1a1 | 496 | occr = exprs->avail_occr; |
0516f6fe SB |
497 | while (occr) |
498 | { | |
6c6d76be | 499 | rtx_insn *insn = occr->insn; |
0516f6fe SB |
500 | print_rtl_single (file, insn); |
501 | fprintf (file, "\n"); | |
502 | occr = occr->next; | |
503 | } | |
504 | fprintf (file, "\n"); | |
505 | return 1; | |
506 | } | |
507 | ||
508 | static void | |
509 | dump_hash_table (FILE *file) | |
510 | { | |
511 | fprintf (file, "\n\nexpression hash table\n"); | |
512 | fprintf (file, "size %ld, %ld elements, %f collision/search ratio\n", | |
c203e8a7 TS |
513 | (long) expr_table->size (), |
514 | (long) expr_table->elements (), | |
515 | expr_table->collisions ()); | |
516 | if (expr_table->elements () > 0) | |
0516f6fe SB |
517 | { |
518 | fprintf (file, "\n\ntable entries:\n"); | |
c203e8a7 | 519 | expr_table->traverse <FILE *, dump_expr_hash_table_entry> (file); |
0516f6fe SB |
520 | } |
521 | fprintf (file, "\n"); | |
522 | } | |
523 | \f | |
5da20cfe RS |
524 | /* Return true if register X is recorded as being set by an instruction |
525 | whose CUID is greater than the one given. */ | |
526 | ||
527 | static bool | |
528 | reg_changed_after_insn_p (rtx x, int cuid) | |
529 | { | |
530 | unsigned int regno, end_regno; | |
531 | ||
532 | regno = REGNO (x); | |
72d19505 | 533 | end_regno = END_REGNO (x); |
5da20cfe RS |
534 | do |
535 | if (reg_avail_info[regno] > cuid) | |
536 | return true; | |
537 | while (++regno < end_regno); | |
538 | return false; | |
539 | } | |
0516f6fe | 540 | |
c93320c4 SB |
541 | /* Return nonzero if the operands of expression X are unchanged |
542 | 1) from the start of INSN's basic block up to but not including INSN | |
543 | if AFTER_INSN is false, or | |
544 | 2) from INSN to the end of INSN's basic block if AFTER_INSN is true. */ | |
0516f6fe SB |
545 | |
546 | static bool | |
6c6d76be | 547 | oprs_unchanged_p (rtx x, rtx_insn *insn, bool after_insn) |
0516f6fe SB |
548 | { |
549 | int i, j; | |
550 | enum rtx_code code; | |
551 | const char *fmt; | |
552 | ||
553 | if (x == 0) | |
554 | return 1; | |
555 | ||
556 | code = GET_CODE (x); | |
557 | switch (code) | |
558 | { | |
559 | case REG: | |
0516f6fe | 560 | /* We are called after register allocation. */ |
e16acfcd | 561 | gcc_assert (REGNO (x) < FIRST_PSEUDO_REGISTER); |
0516f6fe | 562 | if (after_insn) |
5da20cfe | 563 | return !reg_changed_after_insn_p (x, INSN_CUID (insn) - 1); |
0516f6fe | 564 | else |
5da20cfe | 565 | return !reg_changed_after_insn_p (x, 0); |
0516f6fe SB |
566 | |
567 | case MEM: | |
568 | if (load_killed_in_block_p (INSN_CUID (insn), x, after_insn)) | |
569 | return 0; | |
570 | else | |
571 | return oprs_unchanged_p (XEXP (x, 0), insn, after_insn); | |
572 | ||
573 | case PC: | |
574 | case CC0: /*FIXME*/ | |
575 | case CONST: | |
d8116890 | 576 | CASE_CONST_ANY: |
0516f6fe SB |
577 | case SYMBOL_REF: |
578 | case LABEL_REF: | |
579 | case ADDR_VEC: | |
580 | case ADDR_DIFF_VEC: | |
581 | return 1; | |
582 | ||
583 | case PRE_DEC: | |
584 | case PRE_INC: | |
585 | case POST_DEC: | |
586 | case POST_INC: | |
587 | case PRE_MODIFY: | |
588 | case POST_MODIFY: | |
589 | if (after_insn) | |
590 | return 0; | |
591 | break; | |
592 | ||
593 | default: | |
594 | break; | |
595 | } | |
596 | ||
597 | for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--) | |
598 | { | |
599 | if (fmt[i] == 'e') | |
600 | { | |
601 | if (! oprs_unchanged_p (XEXP (x, i), insn, after_insn)) | |
602 | return 0; | |
603 | } | |
604 | else if (fmt[i] == 'E') | |
605 | for (j = 0; j < XVECLEN (x, i); j++) | |
606 | if (! oprs_unchanged_p (XVECEXP (x, i, j), insn, after_insn)) | |
607 | return 0; | |
608 | } | |
609 | ||
610 | return 1; | |
611 | } | |
612 | \f | |
613 | ||
614 | /* Used for communication between find_mem_conflicts and | |
615 | load_killed_in_block_p. Nonzero if find_mem_conflicts finds a | |
616 | conflict between two memory references. | |
617 | This is a bit of a hack to work around the limitations of note_stores. */ | |
618 | static int mems_conflict_p; | |
619 | ||
620 | /* DEST is the output of an instruction. If it is a memory reference, and | |
621 | possibly conflicts with the load found in DATA, then set mems_conflict_p | |
622 | to a nonzero value. */ | |
623 | ||
624 | static void | |
7bc980e1 | 625 | find_mem_conflicts (rtx dest, const_rtx setter ATTRIBUTE_UNUSED, |
0516f6fe SB |
626 | void *data) |
627 | { | |
628 | rtx mem_op = (rtx) data; | |
629 | ||
630 | while (GET_CODE (dest) == SUBREG | |
631 | || GET_CODE (dest) == ZERO_EXTRACT | |
0516f6fe SB |
632 | || GET_CODE (dest) == STRICT_LOW_PART) |
633 | dest = XEXP (dest, 0); | |
634 | ||
635 | /* If DEST is not a MEM, then it will not conflict with the load. Note | |
636 | that function calls are assumed to clobber memory, but are handled | |
637 | elsewhere. */ | |
638 | if (! MEM_P (dest)) | |
639 | return; | |
640 | ||
53d9622b | 641 | if (true_dependence (dest, GET_MODE (dest), mem_op)) |
0516f6fe SB |
642 | mems_conflict_p = 1; |
643 | } | |
644 | \f | |
645 | ||
646 | /* Return nonzero if the expression in X (a memory reference) is killed | |
c93320c4 SB |
647 | in the current basic block before (if AFTER_INSN is false) or after |
648 | (if AFTER_INSN is true) the insn with the CUID in UID_LIMIT. | |
649 | ||
650 | This function assumes that the modifies_mem table is flushed when | |
651 | the hash table construction or redundancy elimination phases start | |
652 | processing a new basic block. */ | |
0516f6fe SB |
653 | |
654 | static int | |
655 | load_killed_in_block_p (int uid_limit, rtx x, bool after_insn) | |
656 | { | |
657 | struct modifies_mem *list_entry = modifies_mem_list; | |
658 | ||
659 | while (list_entry) | |
660 | { | |
6c6d76be | 661 | rtx_insn *setter = list_entry->insn; |
0516f6fe SB |
662 | |
663 | /* Ignore entries in the list that do not apply. */ | |
664 | if ((after_insn | |
665 | && INSN_CUID (setter) < uid_limit) | |
666 | || (! after_insn | |
667 | && INSN_CUID (setter) > uid_limit)) | |
668 | { | |
669 | list_entry = list_entry->next; | |
670 | continue; | |
671 | } | |
672 | ||
673 | /* If SETTER is a call everything is clobbered. Note that calls | |
674 | to pure functions are never put on the list, so we need not | |
675 | worry about them. */ | |
676 | if (CALL_P (setter)) | |
677 | return 1; | |
678 | ||
679 | /* SETTER must be an insn of some kind that sets memory. Call | |
680 | note_stores to examine each hunk of memory that is modified. | |
681 | It will set mems_conflict_p to nonzero if there may be a | |
682 | conflict between X and SETTER. */ | |
683 | mems_conflict_p = 0; | |
684 | note_stores (PATTERN (setter), find_mem_conflicts, x); | |
685 | if (mems_conflict_p) | |
686 | return 1; | |
687 | ||
688 | list_entry = list_entry->next; | |
689 | } | |
690 | return 0; | |
691 | } | |
692 | \f | |
693 | ||
694 | /* Record register first/last/block set information for REGNO in INSN. */ | |
695 | ||
c93320c4 | 696 | static inline void |
6c6d76be | 697 | record_last_reg_set_info (rtx_insn *insn, rtx reg) |
6994f254 MM |
698 | { |
699 | unsigned int regno, end_regno; | |
700 | ||
701 | regno = REGNO (reg); | |
72d19505 | 702 | end_regno = END_REGNO (reg); |
6994f254 MM |
703 | do |
704 | reg_avail_info[regno] = INSN_CUID (insn); | |
705 | while (++regno < end_regno); | |
706 | } | |
707 | ||
708 | static inline void | |
6c6d76be | 709 | record_last_reg_set_info_regno (rtx_insn *insn, int regno) |
0516f6fe SB |
710 | { |
711 | reg_avail_info[regno] = INSN_CUID (insn); | |
712 | } | |
713 | ||
714 | ||
715 | /* Record memory modification information for INSN. We do not actually care | |
716 | about the memory location(s) that are set, or even how they are set (consider | |
717 | a CALL_INSN). We merely need to record which insns modify memory. */ | |
718 | ||
719 | static void | |
6c6d76be | 720 | record_last_mem_set_info (rtx_insn *insn) |
0516f6fe SB |
721 | { |
722 | struct modifies_mem *list_entry; | |
723 | ||
724 | list_entry = (struct modifies_mem *) obstack_alloc (&modifies_mem_obstack, | |
725 | sizeof (struct modifies_mem)); | |
726 | list_entry->insn = insn; | |
727 | list_entry->next = modifies_mem_list; | |
728 | modifies_mem_list = list_entry; | |
af3eb110 JL |
729 | |
730 | record_last_mem_set_info_common (insn, modify_mem_list, | |
731 | canon_modify_mem_list, | |
732 | modify_mem_list_set, | |
733 | blocks_with_calls); | |
0516f6fe SB |
734 | } |
735 | ||
736 | /* Called from compute_hash_table via note_stores to handle one | |
737 | SET or CLOBBER in an insn. DATA is really the instruction in which | |
738 | the SET is taking place. */ | |
739 | ||
740 | static void | |
7bc980e1 | 741 | record_last_set_info (rtx dest, const_rtx setter ATTRIBUTE_UNUSED, void *data) |
0516f6fe | 742 | { |
6c6d76be | 743 | rtx_insn *last_set_insn = (rtx_insn *) data; |
0516f6fe SB |
744 | |
745 | if (GET_CODE (dest) == SUBREG) | |
746 | dest = SUBREG_REG (dest); | |
747 | ||
748 | if (REG_P (dest)) | |
6994f254 | 749 | record_last_reg_set_info (last_set_insn, dest); |
56038245 SB |
750 | else if (MEM_P (dest)) |
751 | { | |
752 | /* Ignore pushes, they don't clobber memory. They may still | |
753 | clobber the stack pointer though. Some targets do argument | |
754 | pushes without adding REG_INC notes. See e.g. PR25196, | |
755 | where a pushsi2 on i386 doesn't have REG_INC notes. Note | |
756 | such changes here too. */ | |
757 | if (! push_operand (dest, GET_MODE (dest))) | |
758 | record_last_mem_set_info (last_set_insn); | |
759 | else | |
6994f254 | 760 | record_last_reg_set_info_regno (last_set_insn, STACK_POINTER_REGNUM); |
56038245 | 761 | } |
0516f6fe | 762 | } |
c93320c4 | 763 | |
0516f6fe SB |
764 | |
765 | /* Reset tables used to keep track of what's still available since the | |
766 | start of the block. */ | |
767 | ||
768 | static void | |
769 | reset_opr_set_tables (void) | |
770 | { | |
771 | memset (reg_avail_info, 0, FIRST_PSEUDO_REGISTER * sizeof (int)); | |
772 | obstack_free (&modifies_mem_obstack, modifies_mem_obstack_bottom); | |
773 | modifies_mem_list = NULL; | |
774 | } | |
c93320c4 | 775 | \f |
0516f6fe SB |
776 | |
777 | /* Record things set by INSN. | |
778 | This data is used by oprs_unchanged_p. */ | |
779 | ||
780 | static void | |
6c6d76be | 781 | record_opr_changes (rtx_insn *insn) |
0516f6fe | 782 | { |
c93320c4 | 783 | rtx note; |
0516f6fe | 784 | |
c93320c4 SB |
785 | /* Find all stores and record them. */ |
786 | note_stores (PATTERN (insn), record_last_set_info, insn); | |
0516f6fe | 787 | |
c93320c4 SB |
788 | /* Also record autoincremented REGs for this insn as changed. */ |
789 | for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) | |
790 | if (REG_NOTE_KIND (note) == REG_INC) | |
6994f254 | 791 | record_last_reg_set_info (insn, XEXP (note, 0)); |
0516f6fe | 792 | |
c93320c4 SB |
793 | /* Finally, if this is a call, record all call clobbers. */ |
794 | if (CALL_P (insn)) | |
795 | { | |
6994f254 | 796 | unsigned int regno; |
5da20cfe | 797 | rtx link, x; |
c7fb4c7a SB |
798 | hard_reg_set_iterator hrsi; |
799 | EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call, 0, regno, hrsi) | |
800 | record_last_reg_set_info_regno (insn, regno); | |
0516f6fe | 801 | |
5da20cfe RS |
802 | for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1)) |
803 | if (GET_CODE (XEXP (link, 0)) == CLOBBER) | |
804 | { | |
805 | x = XEXP (XEXP (link, 0), 0); | |
806 | if (REG_P (x)) | |
807 | { | |
808 | gcc_assert (HARD_REGISTER_P (x)); | |
6994f254 | 809 | record_last_reg_set_info (insn, x); |
5da20cfe RS |
810 | } |
811 | } | |
812 | ||
becfd6e5 | 813 | if (! RTL_CONST_OR_PURE_CALL_P (insn)) |
c93320c4 SB |
814 | record_last_mem_set_info (insn); |
815 | } | |
0516f6fe SB |
816 | } |
817 | \f | |
818 | ||
819 | /* Scan the pattern of INSN and add an entry to the hash TABLE. | |
820 | After reload we are interested in loads/stores only. */ | |
821 | ||
822 | static void | |
6c6d76be | 823 | hash_scan_set (rtx_insn *insn) |
0516f6fe SB |
824 | { |
825 | rtx pat = PATTERN (insn); | |
826 | rtx src = SET_SRC (pat); | |
827 | rtx dest = SET_DEST (pat); | |
828 | ||
829 | /* We are only interested in loads and stores. */ | |
830 | if (! MEM_P (src) && ! MEM_P (dest)) | |
831 | return; | |
832 | ||
833 | /* Don't mess with jumps and nops. */ | |
834 | if (JUMP_P (insn) || set_noop_p (pat)) | |
835 | return; | |
836 | ||
0516f6fe SB |
837 | if (REG_P (dest)) |
838 | { | |
c93320c4 | 839 | if (/* Don't CSE something if we can't do a reg/reg copy. */ |
0516f6fe SB |
840 | can_copy_p (GET_MODE (dest)) |
841 | /* Is SET_SRC something we want to gcse? */ | |
842 | && general_operand (src, GET_MODE (src)) | |
a3f4b7d8 SB |
843 | #ifdef STACK_REGS |
844 | /* Never consider insns touching the register stack. It may | |
845 | create situations that reg-stack cannot handle (e.g. a stack | |
846 | register live across an abnormal edge). */ | |
847 | && (REGNO (dest) < FIRST_STACK_REG || REGNO (dest) > LAST_STACK_REG) | |
848 | #endif | |
0516f6fe SB |
849 | /* An expression is not available if its operands are |
850 | subsequently modified, including this insn. */ | |
851 | && oprs_unchanged_p (src, insn, true)) | |
852 | { | |
853 | insert_expr_in_table (src, insn); | |
854 | } | |
855 | } | |
856 | else if (REG_P (src)) | |
857 | { | |
858 | /* Only record sets of pseudo-regs in the hash table. */ | |
c93320c4 | 859 | if (/* Don't CSE something if we can't do a reg/reg copy. */ |
0516f6fe SB |
860 | can_copy_p (GET_MODE (src)) |
861 | /* Is SET_DEST something we want to gcse? */ | |
862 | && general_operand (dest, GET_MODE (dest)) | |
a3f4b7d8 SB |
863 | #ifdef STACK_REGS |
864 | /* As above for STACK_REGS. */ | |
865 | && (REGNO (src) < FIRST_STACK_REG || REGNO (src) > LAST_STACK_REG) | |
866 | #endif | |
0516f6fe SB |
867 | && ! (flag_float_store && FLOAT_MODE_P (GET_MODE (dest))) |
868 | /* Check if the memory expression is killed after insn. */ | |
869 | && ! load_killed_in_block_p (INSN_CUID (insn) + 1, dest, true) | |
870 | && oprs_unchanged_p (XEXP (dest, 0), insn, true)) | |
871 | { | |
872 | insert_expr_in_table (dest, insn); | |
873 | } | |
874 | } | |
875 | } | |
876 | \f | |
c93320c4 | 877 | |
0516f6fe | 878 | /* Create hash table of memory expressions available at end of basic |
c93320c4 SB |
879 | blocks. Basically you should think of this hash table as the |
880 | representation of AVAIL_OUT. This is the set of expressions that | |
881 | is generated in a basic block and not killed before the end of the | |
882 | same basic block. Notice that this is really a local computation. */ | |
0516f6fe SB |
883 | |
884 | static void | |
885 | compute_hash_table (void) | |
886 | { | |
887 | basic_block bb; | |
888 | ||
11cd3bed | 889 | FOR_EACH_BB_FN (bb, cfun) |
0516f6fe | 890 | { |
6c6d76be | 891 | rtx_insn *insn; |
0516f6fe SB |
892 | |
893 | /* First pass over the instructions records information used to | |
c93320c4 SB |
894 | determine when registers and memory are last set. |
895 | Since we compute a "local" AVAIL_OUT, reset the tables that | |
896 | help us keep track of what has been modified since the start | |
897 | of the block. */ | |
898 | reset_opr_set_tables (); | |
0516f6fe SB |
899 | FOR_BB_INSNS (bb, insn) |
900 | { | |
c93320c4 SB |
901 | if (INSN_P (insn)) |
902 | record_opr_changes (insn); | |
903 | } | |
0516f6fe | 904 | |
c93320c4 | 905 | /* The next pass actually builds the hash table. */ |
0516f6fe SB |
906 | FOR_BB_INSNS (bb, insn) |
907 | if (INSN_P (insn) && GET_CODE (PATTERN (insn)) == SET) | |
908 | hash_scan_set (insn); | |
909 | } | |
910 | } | |
911 | \f | |
912 | ||
913 | /* Check if register REG is killed in any insn waiting to be inserted on | |
914 | edge E. This function is required to check that our data flow analysis | |
915 | is still valid prior to commit_edge_insertions. */ | |
916 | ||
917 | static bool | |
918 | reg_killed_on_edge (rtx reg, edge e) | |
919 | { | |
3ffa95c2 | 920 | rtx_insn *insn; |
0516f6fe SB |
921 | |
922 | for (insn = e->insns.r; insn; insn = NEXT_INSN (insn)) | |
923 | if (INSN_P (insn) && reg_set_p (reg, insn)) | |
924 | return true; | |
925 | ||
926 | return false; | |
927 | } | |
928 | ||
929 | /* Similar to above - check if register REG is used in any insn waiting | |
930 | to be inserted on edge E. | |
931 | Assumes no such insn can be a CALL_INSN; if so call reg_used_between_p | |
932 | with PREV(insn),NEXT(insn) instead of calling reg_overlap_mentioned_p. */ | |
933 | ||
934 | static bool | |
935 | reg_used_on_edge (rtx reg, edge e) | |
936 | { | |
3ffa95c2 | 937 | rtx_insn *insn; |
0516f6fe SB |
938 | |
939 | for (insn = e->insns.r; insn; insn = NEXT_INSN (insn)) | |
940 | if (INSN_P (insn) && reg_overlap_mentioned_p (reg, PATTERN (insn))) | |
941 | return true; | |
942 | ||
943 | return false; | |
944 | } | |
945 | \f | |
0516f6fe SB |
946 | /* Return the loaded/stored register of a load/store instruction. */ |
947 | ||
948 | static rtx | |
6c6d76be | 949 | get_avail_load_store_reg (rtx_insn *insn) |
0516f6fe | 950 | { |
e16acfcd NS |
951 | if (REG_P (SET_DEST (PATTERN (insn)))) |
952 | /* A load. */ | |
c3284718 | 953 | return SET_DEST (PATTERN (insn)); |
e16acfcd NS |
954 | else |
955 | { | |
956 | /* A store. */ | |
957 | gcc_assert (REG_P (SET_SRC (PATTERN (insn)))); | |
958 | return SET_SRC (PATTERN (insn)); | |
959 | } | |
0516f6fe SB |
960 | } |
961 | ||
962 | /* Return nonzero if the predecessors of BB are "well behaved". */ | |
963 | ||
964 | static bool | |
965 | bb_has_well_behaved_predecessors (basic_block bb) | |
966 | { | |
967 | edge pred; | |
628f6a4e | 968 | edge_iterator ei; |
0516f6fe | 969 | |
76015c34 | 970 | if (EDGE_COUNT (bb->preds) == 0) |
0516f6fe SB |
971 | return false; |
972 | ||
628f6a4e | 973 | FOR_EACH_EDGE (pred, ei, bb->preds) |
0516f6fe SB |
974 | { |
975 | if ((pred->flags & EDGE_ABNORMAL) && EDGE_CRITICAL_P (pred)) | |
976 | return false; | |
977 | ||
76015c34 EB |
978 | if ((pred->flags & EDGE_ABNORMAL_CALL) && cfun->has_nonlocal_label) |
979 | return false; | |
980 | ||
39718607 | 981 | if (tablejump_p (BB_END (pred->src), NULL, NULL)) |
0516f6fe SB |
982 | return false; |
983 | } | |
984 | return true; | |
985 | } | |
986 | ||
987 | ||
988 | /* Search for the occurrences of expression in BB. */ | |
989 | ||
990 | static struct occr* | |
af3eb110 | 991 | get_bb_avail_insn (basic_block bb, struct occr *orig_occr, int bitmap_index) |
0516f6fe | 992 | { |
af3eb110 JL |
993 | struct occr *occr = orig_occr; |
994 | ||
0516f6fe SB |
995 | for (; occr != NULL; occr = occr->next) |
996 | if (BLOCK_FOR_INSN (occr->insn) == bb) | |
997 | return occr; | |
af3eb110 JL |
998 | |
999 | /* If we could not find an occurrence in BB, see if BB | |
1000 | has a single predecessor with an occurrence that is | |
1001 | transparent through BB. */ | |
1002 | if (single_pred_p (bb) | |
1003 | && bitmap_bit_p (transp[bb->index], bitmap_index) | |
1004 | && (occr = get_bb_avail_insn (single_pred (bb), orig_occr, bitmap_index))) | |
1005 | { | |
1006 | rtx avail_reg = get_avail_load_store_reg (occr->insn); | |
1007 | if (!reg_set_between_p (avail_reg, | |
1008 | PREV_INSN (BB_HEAD (bb)), | |
1009 | NEXT_INSN (BB_END (bb))) | |
1010 | && !reg_killed_on_edge (avail_reg, single_pred_edge (bb))) | |
1011 | return occr; | |
1012 | } | |
1013 | ||
0516f6fe SB |
1014 | return NULL; |
1015 | } | |
1016 | ||
1017 | ||
af3eb110 JL |
1018 | /* This helper is called via htab_traverse. */ |
1019 | int | |
1020 | compute_expr_transp (expr **slot, FILE *dump_file ATTRIBUTE_UNUSED) | |
1021 | { | |
1022 | struct expr *expr = *slot; | |
1023 | ||
1024 | compute_transp (expr->expr, expr->bitmap_index, transp, | |
1025 | blocks_with_calls, modify_mem_list_set, | |
1026 | canon_modify_mem_list); | |
1027 | return 1; | |
1028 | } | |
1029 | ||
0516f6fe SB |
1030 | /* This handles the case where several stores feed a partially redundant |
1031 | load. It checks if the redundancy elimination is possible and if it's | |
c93320c4 SB |
1032 | worth it. |
1033 | ||
1034 | Redundancy elimination is possible if, | |
1035 | 1) None of the operands of an insn have been modified since the start | |
1036 | of the current basic block. | |
1037 | 2) In any predecessor of the current basic block, the same expression | |
1038 | is generated. | |
1039 | ||
1040 | See the function body for the heuristics that determine if eliminating | |
1041 | a redundancy is also worth doing, assuming it is possible. */ | |
0516f6fe SB |
1042 | |
1043 | static void | |
6c6d76be | 1044 | eliminate_partially_redundant_load (basic_block bb, rtx_insn *insn, |
0516f6fe SB |
1045 | struct expr *expr) |
1046 | { | |
1047 | edge pred; | |
6c6d76be | 1048 | rtx_insn *avail_insn = NULL; |
0516f6fe SB |
1049 | rtx avail_reg; |
1050 | rtx dest, pat; | |
1051 | struct occr *a_occr; | |
1052 | struct unoccr *occr, *avail_occrs = NULL; | |
1053 | struct unoccr *unoccr, *unavail_occrs = NULL, *rollback_unoccr = NULL; | |
1054 | int npred_ok = 0; | |
1055 | gcov_type ok_count = 0; /* Redundant load execution count. */ | |
1056 | gcov_type critical_count = 0; /* Execution count of critical edges. */ | |
628f6a4e | 1057 | edge_iterator ei; |
303f6390 | 1058 | bool critical_edge_split = false; |
0516f6fe SB |
1059 | |
1060 | /* The execution count of the loads to be added to make the | |
1061 | load fully redundant. */ | |
1062 | gcov_type not_ok_count = 0; | |
1063 | basic_block pred_bb; | |
1064 | ||
1065 | pat = PATTERN (insn); | |
1066 | dest = SET_DEST (pat); | |
1067 | ||
1068 | /* Check that the loaded register is not used, set, or killed from the | |
1069 | beginning of the block. */ | |
5da20cfe RS |
1070 | if (reg_changed_after_insn_p (dest, 0) |
1071 | || reg_used_between_p (dest, PREV_INSN (BB_HEAD (bb)), insn)) | |
0516f6fe SB |
1072 | return; |
1073 | ||
1074 | /* Check potential for replacing load with copy for predecessors. */ | |
628f6a4e | 1075 | FOR_EACH_EDGE (pred, ei, bb->preds) |
0516f6fe | 1076 | { |
6c6d76be | 1077 | rtx_insn *next_pred_bb_end; |
0516f6fe | 1078 | |
6c6d76be | 1079 | avail_insn = NULL; |
303f6390 | 1080 | avail_reg = NULL_RTX; |
0516f6fe | 1081 | pred_bb = pred->src; |
af3eb110 JL |
1082 | for (a_occr = get_bb_avail_insn (pred_bb, |
1083 | expr->avail_occr, | |
1084 | expr->bitmap_index); | |
1085 | a_occr; | |
1086 | a_occr = get_bb_avail_insn (pred_bb, | |
1087 | a_occr->next, | |
1088 | expr->bitmap_index)) | |
0516f6fe SB |
1089 | { |
1090 | /* Check if the loaded register is not used. */ | |
1091 | avail_insn = a_occr->insn; | |
e16acfcd NS |
1092 | avail_reg = get_avail_load_store_reg (avail_insn); |
1093 | gcc_assert (avail_reg); | |
b8698a0f | 1094 | |
0516f6fe SB |
1095 | /* Make sure we can generate a move from register avail_reg to |
1096 | dest. */ | |
1476d1bd MM |
1097 | rtx_insn *move = gen_move_insn (copy_rtx (dest), |
1098 | copy_rtx (avail_reg)); | |
daca1a96 RS |
1099 | extract_insn (move); |
1100 | if (! constrain_operands (1, get_preferred_alternatives (insn, | |
1101 | pred_bb)) | |
0516f6fe SB |
1102 | || reg_killed_on_edge (avail_reg, pred) |
1103 | || reg_used_on_edge (dest, pred)) | |
1104 | { | |
1105 | avail_insn = NULL; | |
1106 | continue; | |
1107 | } | |
af3eb110 | 1108 | next_pred_bb_end = NEXT_INSN (BB_END (BLOCK_FOR_INSN (avail_insn))); |
5da20cfe | 1109 | if (!reg_set_between_p (avail_reg, avail_insn, next_pred_bb_end)) |
0516f6fe SB |
1110 | /* AVAIL_INSN remains non-null. */ |
1111 | break; | |
1112 | else | |
1113 | avail_insn = NULL; | |
1114 | } | |
1115 | ||
1116 | if (EDGE_CRITICAL_P (pred)) | |
1117 | critical_count += pred->count; | |
1118 | ||
1119 | if (avail_insn != NULL_RTX) | |
1120 | { | |
1121 | npred_ok++; | |
1122 | ok_count += pred->count; | |
303f6390 MH |
1123 | if (! set_noop_p (PATTERN (gen_move_insn (copy_rtx (dest), |
1124 | copy_rtx (avail_reg))))) | |
1125 | { | |
1126 | /* Check if there is going to be a split. */ | |
1127 | if (EDGE_CRITICAL_P (pred)) | |
1128 | critical_edge_split = true; | |
1129 | } | |
1130 | else /* Its a dead move no need to generate. */ | |
1131 | continue; | |
0516f6fe | 1132 | occr = (struct unoccr *) obstack_alloc (&unoccr_obstack, |
9275de65 | 1133 | sizeof (struct unoccr)); |
0516f6fe SB |
1134 | occr->insn = avail_insn; |
1135 | occr->pred = pred; | |
1136 | occr->next = avail_occrs; | |
1137 | avail_occrs = occr; | |
1138 | if (! rollback_unoccr) | |
1139 | rollback_unoccr = occr; | |
1140 | } | |
1141 | else | |
1142 | { | |
c83eecad | 1143 | /* Adding a load on a critical edge will cause a split. */ |
303f6390 MH |
1144 | if (EDGE_CRITICAL_P (pred)) |
1145 | critical_edge_split = true; | |
0516f6fe SB |
1146 | not_ok_count += pred->count; |
1147 | unoccr = (struct unoccr *) obstack_alloc (&unoccr_obstack, | |
1148 | sizeof (struct unoccr)); | |
6c6d76be | 1149 | unoccr->insn = NULL; |
0516f6fe SB |
1150 | unoccr->pred = pred; |
1151 | unoccr->next = unavail_occrs; | |
1152 | unavail_occrs = unoccr; | |
1153 | if (! rollback_unoccr) | |
1154 | rollback_unoccr = unoccr; | |
1155 | } | |
1156 | } | |
1157 | ||
1158 | if (/* No load can be replaced by copy. */ | |
1159 | npred_ok == 0 | |
b8698a0f | 1160 | /* Prevent exploding the code. */ |
efd8f750 | 1161 | || (optimize_bb_for_size_p (bb) && npred_ok > 1) |
b8698a0f | 1162 | /* If we don't have profile information we cannot tell if splitting |
303f6390 MH |
1163 | a critical edge is profitable or not so don't do it. */ |
1164 | || ((! profile_info || ! flag_branch_probabilities | |
1165 | || targetm.cannot_modify_jumps_p ()) | |
1166 | && critical_edge_split)) | |
0516f6fe SB |
1167 | goto cleanup; |
1168 | ||
1169 | /* Check if it's worth applying the partial redundancy elimination. */ | |
1170 | if (ok_count < GCSE_AFTER_RELOAD_PARTIAL_FRACTION * not_ok_count) | |
1171 | goto cleanup; | |
1172 | if (ok_count < GCSE_AFTER_RELOAD_CRITICAL_FRACTION * critical_count) | |
1173 | goto cleanup; | |
1174 | ||
1175 | /* Generate moves to the loaded register from where | |
1176 | the memory is available. */ | |
1177 | for (occr = avail_occrs; occr; occr = occr->next) | |
1178 | { | |
1179 | avail_insn = occr->insn; | |
1180 | pred = occr->pred; | |
1181 | /* Set avail_reg to be the register having the value of the | |
1182 | memory. */ | |
1183 | avail_reg = get_avail_load_store_reg (avail_insn); | |
e16acfcd | 1184 | gcc_assert (avail_reg); |
0516f6fe SB |
1185 | |
1186 | insert_insn_on_edge (gen_move_insn (copy_rtx (dest), | |
1187 | copy_rtx (avail_reg)), | |
1188 | pred); | |
1189 | stats.moves_inserted++; | |
1190 | ||
1191 | if (dump_file) | |
1192 | fprintf (dump_file, | |
1193 | "generating move from %d to %d on edge from %d to %d\n", | |
1194 | REGNO (avail_reg), | |
1195 | REGNO (dest), | |
1196 | pred->src->index, | |
1197 | pred->dest->index); | |
1198 | } | |
1199 | ||
1200 | /* Regenerate loads where the memory is unavailable. */ | |
1201 | for (unoccr = unavail_occrs; unoccr; unoccr = unoccr->next) | |
1202 | { | |
1203 | pred = unoccr->pred; | |
1204 | insert_insn_on_edge (copy_insn (PATTERN (insn)), pred); | |
1205 | stats.copies_inserted++; | |
1206 | ||
1207 | if (dump_file) | |
1208 | { | |
1209 | fprintf (dump_file, | |
1210 | "generating on edge from %d to %d a copy of load: ", | |
1211 | pred->src->index, | |
1212 | pred->dest->index); | |
1213 | print_rtl (dump_file, PATTERN (insn)); | |
1214 | fprintf (dump_file, "\n"); | |
1215 | } | |
1216 | } | |
1217 | ||
1218 | /* Delete the insn if it is not available in this block and mark it | |
1219 | for deletion if it is available. If insn is available it may help | |
1220 | discover additional redundancies, so mark it for later deletion. */ | |
af3eb110 | 1221 | for (a_occr = get_bb_avail_insn (bb, expr->avail_occr, expr->bitmap_index); |
0516f6fe | 1222 | a_occr && (a_occr->insn != insn); |
af3eb110 | 1223 | a_occr = get_bb_avail_insn (bb, a_occr->next, expr->bitmap_index)) |
e84a58ff | 1224 | ; |
0516f6fe SB |
1225 | |
1226 | if (!a_occr) | |
303f6390 MH |
1227 | { |
1228 | stats.insns_deleted++; | |
1229 | ||
1230 | if (dump_file) | |
1231 | { | |
1232 | fprintf (dump_file, "deleting insn:\n"); | |
1233 | print_rtl_single (dump_file, insn); | |
1234 | fprintf (dump_file, "\n"); | |
1235 | } | |
1236 | delete_insn (insn); | |
1237 | } | |
0516f6fe SB |
1238 | else |
1239 | a_occr->deleted_p = 1; | |
1240 | ||
1241 | cleanup: | |
1242 | if (rollback_unoccr) | |
1243 | obstack_free (&unoccr_obstack, rollback_unoccr); | |
1244 | } | |
1245 | ||
1246 | /* Performing the redundancy elimination as described before. */ | |
1247 | ||
1248 | static void | |
1249 | eliminate_partially_redundant_loads (void) | |
1250 | { | |
6c6d76be | 1251 | rtx_insn *insn; |
0516f6fe SB |
1252 | basic_block bb; |
1253 | ||
1254 | /* Note we start at block 1. */ | |
1255 | ||
fefa31b5 | 1256 | if (ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
0516f6fe SB |
1257 | return; |
1258 | ||
1259 | FOR_BB_BETWEEN (bb, | |
fefa31b5 DM |
1260 | ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb->next_bb, |
1261 | EXIT_BLOCK_PTR_FOR_FN (cfun), | |
0516f6fe SB |
1262 | next_bb) |
1263 | { | |
c93320c4 | 1264 | /* Don't try anything on basic blocks with strange predecessors. */ |
0516f6fe SB |
1265 | if (! bb_has_well_behaved_predecessors (bb)) |
1266 | continue; | |
1267 | ||
c93320c4 | 1268 | /* Do not try anything on cold basic blocks. */ |
fb2fed03 | 1269 | if (optimize_bb_for_size_p (bb)) |
0516f6fe SB |
1270 | continue; |
1271 | ||
c93320c4 SB |
1272 | /* Reset the table of things changed since the start of the current |
1273 | basic block. */ | |
0516f6fe SB |
1274 | reset_opr_set_tables (); |
1275 | ||
c93320c4 SB |
1276 | /* Look at all insns in the current basic block and see if there are |
1277 | any loads in it that we can record. */ | |
0516f6fe SB |
1278 | FOR_BB_INSNS (bb, insn) |
1279 | { | |
1280 | /* Is it a load - of the form (set (reg) (mem))? */ | |
1281 | if (NONJUMP_INSN_P (insn) | |
1282 | && GET_CODE (PATTERN (insn)) == SET | |
1283 | && REG_P (SET_DEST (PATTERN (insn))) | |
1284 | && MEM_P (SET_SRC (PATTERN (insn)))) | |
1285 | { | |
1286 | rtx pat = PATTERN (insn); | |
1287 | rtx src = SET_SRC (pat); | |
1288 | struct expr *expr; | |
1289 | ||
1290 | if (!MEM_VOLATILE_P (src) | |
1291 | && GET_MODE (src) != BLKmode | |
1292 | && general_operand (src, GET_MODE (src)) | |
1293 | /* Are the operands unchanged since the start of the | |
1294 | block? */ | |
1295 | && oprs_unchanged_p (src, insn, false) | |
8f4f502f | 1296 | && !(cfun->can_throw_non_call_exceptions && may_trap_p (src)) |
0516f6fe SB |
1297 | && !side_effects_p (src) |
1298 | /* Is the expression recorded? */ | |
1299 | && (expr = lookup_expr_in_table (src)) != NULL) | |
1300 | { | |
1301 | /* We now have a load (insn) and an available memory at | |
1302 | its BB start (expr). Try to remove the loads if it is | |
1303 | redundant. */ | |
1304 | eliminate_partially_redundant_load (bb, insn, expr); | |
1305 | } | |
1306 | } | |
1307 | ||
c93320c4 SB |
1308 | /* Keep track of everything modified by this insn, so that we |
1309 | know what has been modified since the start of the current | |
1310 | basic block. */ | |
0516f6fe | 1311 | if (INSN_P (insn)) |
c93320c4 | 1312 | record_opr_changes (insn); |
0516f6fe SB |
1313 | } |
1314 | } | |
1315 | ||
1316 | commit_edge_insertions (); | |
1317 | } | |
1318 | ||
1319 | /* Go over the expression hash table and delete insns that were | |
1320 | marked for later deletion. */ | |
1321 | ||
1322 | /* This helper is called via htab_traverse. */ | |
4a8fb1a1 LC |
1323 | int |
1324 | delete_redundant_insns_1 (expr **slot, void *data ATTRIBUTE_UNUSED) | |
0516f6fe | 1325 | { |
4a8fb1a1 | 1326 | struct expr *exprs = *slot; |
0516f6fe SB |
1327 | struct occr *occr; |
1328 | ||
4a8fb1a1 | 1329 | for (occr = exprs->avail_occr; occr != NULL; occr = occr->next) |
0516f6fe | 1330 | { |
6fb5fa3c | 1331 | if (occr->deleted_p && dbg_cnt (gcse2_delete)) |
0516f6fe SB |
1332 | { |
1333 | delete_insn (occr->insn); | |
1334 | stats.insns_deleted++; | |
1335 | ||
1336 | if (dump_file) | |
1337 | { | |
1338 | fprintf (dump_file, "deleting insn:\n"); | |
1339 | print_rtl_single (dump_file, occr->insn); | |
1340 | fprintf (dump_file, "\n"); | |
1341 | } | |
1342 | } | |
1343 | } | |
1344 | ||
1345 | return 1; | |
1346 | } | |
1347 | ||
1348 | static void | |
1349 | delete_redundant_insns (void) | |
1350 | { | |
c203e8a7 | 1351 | expr_table->traverse <void *, delete_redundant_insns_1> (NULL); |
0516f6fe SB |
1352 | if (dump_file) |
1353 | fprintf (dump_file, "\n"); | |
1354 | } | |
1355 | ||
1356 | /* Main entry point of the GCSE after reload - clean some redundant loads | |
1357 | due to spilling. */ | |
1358 | ||
6e9ca1fa | 1359 | static void |
0516f6fe SB |
1360 | gcse_after_reload_main (rtx f ATTRIBUTE_UNUSED) |
1361 | { | |
3f55b339 | 1362 | |
0516f6fe SB |
1363 | memset (&stats, 0, sizeof (stats)); |
1364 | ||
fa10beec | 1365 | /* Allocate memory for this pass. |
0516f6fe SB |
1366 | Also computes and initializes the insns' CUIDs. */ |
1367 | alloc_mem (); | |
1368 | ||
1369 | /* We need alias analysis. */ | |
1370 | init_alias_analysis (); | |
1371 | ||
1372 | compute_hash_table (); | |
1373 | ||
1374 | if (dump_file) | |
1375 | dump_hash_table (dump_file); | |
1376 | ||
c203e8a7 | 1377 | if (expr_table->elements () > 0) |
0516f6fe | 1378 | { |
af3eb110 JL |
1379 | /* Knowing which MEMs are transparent through a block can signifiantly |
1380 | increase the number of redundant loads found. So compute transparency | |
1381 | information for each memory expression in the hash table. */ | |
1382 | df_analyze (); | |
1383 | /* This can not be part of the normal allocation routine because | |
1384 | we have to know the number of elements in the hash table. */ | |
1385 | transp = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), | |
1386 | expr_table->elements ()); | |
1387 | bitmap_vector_ones (transp, last_basic_block_for_fn (cfun)); | |
1388 | expr_table->traverse <FILE *, compute_expr_transp> (dump_file); | |
0516f6fe SB |
1389 | eliminate_partially_redundant_loads (); |
1390 | delete_redundant_insns (); | |
af3eb110 | 1391 | sbitmap_vector_free (transp); |
0516f6fe SB |
1392 | |
1393 | if (dump_file) | |
1394 | { | |
1395 | fprintf (dump_file, "GCSE AFTER RELOAD stats:\n"); | |
1396 | fprintf (dump_file, "copies inserted: %d\n", stats.copies_inserted); | |
1397 | fprintf (dump_file, "moves inserted: %d\n", stats.moves_inserted); | |
1398 | fprintf (dump_file, "insns deleted: %d\n", stats.insns_deleted); | |
1399 | fprintf (dump_file, "\n\n"); | |
1400 | } | |
4da3b811 NF |
1401 | |
1402 | statistics_counter_event (cfun, "copies inserted", | |
1403 | stats.copies_inserted); | |
1404 | statistics_counter_event (cfun, "moves inserted", | |
1405 | stats.moves_inserted); | |
1406 | statistics_counter_event (cfun, "insns deleted", | |
1407 | stats.insns_deleted); | |
0516f6fe | 1408 | } |
b8698a0f | 1409 | |
0516f6fe SB |
1410 | /* We are finished with alias. */ |
1411 | end_alias_analysis (); | |
1412 | ||
1413 | free_mem (); | |
1414 | } | |
1415 | ||
ef330312 | 1416 | \f |
ef330312 | 1417 | |
c2924966 | 1418 | static unsigned int |
ef330312 PB |
1419 | rest_of_handle_gcse2 (void) |
1420 | { | |
1421 | gcse_after_reload_main (get_insns ()); | |
1422 | rebuild_jump_labels (get_insns ()); | |
c2924966 | 1423 | return 0; |
ef330312 PB |
1424 | } |
1425 | ||
27a4cd48 DM |
1426 | namespace { |
1427 | ||
1428 | const pass_data pass_data_gcse2 = | |
ef330312 | 1429 | { |
27a4cd48 DM |
1430 | RTL_PASS, /* type */ |
1431 | "gcse2", /* name */ | |
1432 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
1433 | TV_GCSE_AFTER_RELOAD, /* tv_id */ |
1434 | 0, /* properties_required */ | |
1435 | 0, /* properties_provided */ | |
1436 | 0, /* properties_destroyed */ | |
1437 | 0, /* todo_flags_start */ | |
3bea341f | 1438 | 0, /* todo_flags_finish */ |
ef330312 | 1439 | }; |
27a4cd48 DM |
1440 | |
1441 | class pass_gcse2 : public rtl_opt_pass | |
1442 | { | |
1443 | public: | |
c3284718 RS |
1444 | pass_gcse2 (gcc::context *ctxt) |
1445 | : rtl_opt_pass (pass_data_gcse2, ctxt) | |
27a4cd48 DM |
1446 | {} |
1447 | ||
1448 | /* opt_pass methods: */ | |
1a3d085c TS |
1449 | virtual bool gate (function *fun) |
1450 | { | |
1451 | return (optimize > 0 && flag_gcse_after_reload | |
1452 | && optimize_function_for_speed_p (fun)); | |
1453 | } | |
1454 | ||
be55bfe6 | 1455 | virtual unsigned int execute (function *) { return rest_of_handle_gcse2 (); } |
27a4cd48 DM |
1456 | |
1457 | }; // class pass_gcse2 | |
1458 | ||
1459 | } // anon namespace | |
1460 | ||
1461 | rtl_opt_pass * | |
1462 | make_pass_gcse2 (gcc::context *ctxt) | |
1463 | { | |
1464 | return new pass_gcse2 (ctxt); | |
1465 | } |