]>
Commit | Line | Data |
---|---|---|
3072d30e | 1 | /* RTL dead store elimination. |
d353bf18 | 2 | Copyright (C) 2005-2015 Free Software Foundation, Inc. |
3072d30e | 3 | |
4 | Contributed by Richard Sandiford <rsandifor@codesourcery.com> | |
5 | and Kenneth Zadeck <zadeck@naturalbridge.com> | |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
8c4c00c1 | 11 | Software Foundation; either version 3, or (at your option) any later |
3072d30e | 12 | version. |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ | |
3072d30e | 22 | |
23 | #undef BASELINE | |
24 | ||
25 | #include "config.h" | |
26 | #include "system.h" | |
27 | #include "coretypes.h" | |
3072d30e | 28 | #include "tm.h" |
29 | #include "rtl.h" | |
b20a8bb4 | 30 | #include "alias.h" |
31 | #include "symtab.h" | |
3072d30e | 32 | #include "tree.h" |
b20a8bb4 | 33 | #include "fold-const.h" |
9ed99284 | 34 | #include "stor-layout.h" |
10d4de0e | 35 | #include "tm_p.h" |
3072d30e | 36 | #include "regs.h" |
37 | #include "hard-reg-set.h" | |
5a9ecd4a | 38 | #include "regset.h" |
3072d30e | 39 | #include "flags.h" |
94ea8568 | 40 | #include "dominance.h" |
41 | #include "cfg.h" | |
42 | #include "cfgrtl.h" | |
43 | #include "predict.h" | |
44 | #include "basic-block.h" | |
3072d30e | 45 | #include "df.h" |
46 | #include "cselib.h" | |
3072d30e | 47 | #include "tree-pass.h" |
48 | #include "alloc-pool.h" | |
3072d30e | 49 | #include "insn-config.h" |
d53441c8 | 50 | #include "function.h" |
d53441c8 | 51 | #include "expmed.h" |
52 | #include "dojump.h" | |
53 | #include "explow.h" | |
54 | #include "calls.h" | |
55 | #include "emit-rtl.h" | |
56 | #include "varasm.h" | |
57 | #include "stmt.h" | |
3072d30e | 58 | #include "expr.h" |
59 | #include "recog.h" | |
34517c64 | 60 | #include "insn-codes.h" |
5c9051a4 | 61 | #include "optabs.h" |
3072d30e | 62 | #include "dbgcnt.h" |
9311ed8a | 63 | #include "target.h" |
1242bee6 | 64 | #include "params.h" |
bc61cadb | 65 | #include "tree-ssa-alias.h" |
66 | #include "internal-fn.h" | |
67 | #include "gimple-expr.h" | |
073c1fd5 | 68 | #include "gimple.h" |
69 | #include "gimple-ssa.h" | |
ec1203cd | 70 | #include "rtl-iter.h" |
1f91a12d | 71 | #include "cfgcleanup.h" |
3072d30e | 72 | |
73 | /* This file contains three techniques for performing Dead Store | |
48e1416a | 74 | Elimination (dse). |
3072d30e | 75 | |
76 | * The first technique performs dse locally on any base address. It | |
77 | is based on the cselib which is a local value numbering technique. | |
78 | This technique is local to a basic block but deals with a fairly | |
79 | general addresses. | |
48e1416a | 80 | |
3072d30e | 81 | * The second technique performs dse globally but is restricted to |
82 | base addresses that are either constant or are relative to the | |
83 | frame_pointer. | |
84 | ||
85 | * The third technique, (which is only done after register allocation) | |
86 | processes the spill spill slots. This differs from the second | |
87 | technique because it takes advantage of the fact that spilling is | |
88 | completely free from the effects of aliasing. | |
89 | ||
90 | Logically, dse is a backwards dataflow problem. A store can be | |
91 | deleted if it if cannot be reached in the backward direction by any | |
92 | use of the value being stored. However, the local technique uses a | |
93 | forwards scan of the basic block because cselib requires that the | |
94 | block be processed in that order. | |
95 | ||
96 | The pass is logically broken into 7 steps: | |
97 | ||
98 | 0) Initialization. | |
99 | ||
100 | 1) The local algorithm, as well as scanning the insns for the two | |
101 | global algorithms. | |
102 | ||
103 | 2) Analysis to see if the global algs are necessary. In the case | |
104 | of stores base on a constant address, there must be at least two | |
105 | stores to that address, to make it possible to delete some of the | |
106 | stores. In the case of stores off of the frame or spill related | |
107 | stores, only one store to an address is necessary because those | |
108 | stores die at the end of the function. | |
109 | ||
48e1416a | 110 | 3) Set up the global dataflow equations based on processing the |
3072d30e | 111 | info parsed in the first step. |
112 | ||
113 | 4) Solve the dataflow equations. | |
114 | ||
115 | 5) Delete the insns that the global analysis has indicated are | |
116 | unnecessary. | |
117 | ||
9d75589a | 118 | 6) Delete insns that store the same value as preceding store |
aa140b76 | 119 | where the earlier store couldn't be eliminated. |
120 | ||
121 | 7) Cleanup. | |
3072d30e | 122 | |
123 | This step uses cselib and canon_rtx to build the largest expression | |
124 | possible for each address. This pass is a forwards pass through | |
125 | each basic block. From the point of view of the global technique, | |
126 | the first pass could examine a block in either direction. The | |
bef304b8 | 127 | forwards ordering is to accommodate cselib. |
3072d30e | 128 | |
2d0fd66d | 129 | We make a simplifying assumption: addresses fall into four broad |
3072d30e | 130 | categories: |
131 | ||
132 | 1) base has rtx_varies_p == false, offset is constant. | |
133 | 2) base has rtx_varies_p == false, offset variable. | |
134 | 3) base has rtx_varies_p == true, offset constant. | |
135 | 4) base has rtx_varies_p == true, offset variable. | |
136 | ||
137 | The local passes are able to process all 4 kinds of addresses. The | |
2d0fd66d | 138 | global pass only handles 1). |
3072d30e | 139 | |
140 | The global problem is formulated as follows: | |
141 | ||
142 | A store, S1, to address A, where A is not relative to the stack | |
143 | frame, can be eliminated if all paths from S1 to the end of the | |
2d0fd66d | 144 | function contain another store to A before a read to A. |
3072d30e | 145 | |
146 | If the address A is relative to the stack frame, a store S2 to A | |
2d0fd66d | 147 | can be eliminated if there are no paths from S2 that reach the |
3072d30e | 148 | end of the function that read A before another store to A. In |
2d0fd66d | 149 | this case S2 can be deleted if there are paths from S2 to the |
3072d30e | 150 | end of the function that have no reads or writes to A. This |
151 | second case allows stores to the stack frame to be deleted that | |
152 | would otherwise die when the function returns. This cannot be | |
153 | done if stores_off_frame_dead_at_return is not true. See the doc | |
154 | for that variable for when this variable is false. | |
155 | ||
156 | The global problem is formulated as a backwards set union | |
157 | dataflow problem where the stores are the gens and reads are the | |
158 | kills. Set union problems are rare and require some special | |
159 | handling given our representation of bitmaps. A straightforward | |
2d0fd66d | 160 | implementation requires a lot of bitmaps filled with 1s. |
3072d30e | 161 | These are expensive and cumbersome in our bitmap formulation so |
162 | care has been taken to avoid large vectors filled with 1s. See | |
163 | the comments in bb_info and in the dataflow confluence functions | |
48e1416a | 164 | for details. |
3072d30e | 165 | |
166 | There are two places for further enhancements to this algorithm: | |
48e1416a | 167 | |
3072d30e | 168 | 1) The original dse which was embedded in a pass called flow also |
169 | did local address forwarding. For example in | |
170 | ||
171 | A <- r100 | |
172 | ... <- A | |
173 | ||
174 | flow would replace the right hand side of the second insn with a | |
6dfdc153 | 175 | reference to r100. Most of the information is available to add this |
3072d30e | 176 | to this pass. It has not done it because it is a lot of work in |
177 | the case that either r100 is assigned to between the first and | |
178 | second insn and/or the second insn is a load of part of the value | |
179 | stored by the first insn. | |
180 | ||
181 | insn 5 in gcc.c-torture/compile/990203-1.c simple case. | |
182 | insn 15 in gcc.c-torture/execute/20001017-2.c simple case. | |
183 | insn 25 in gcc.c-torture/execute/20001026-1.c simple case. | |
184 | insn 44 in gcc.c-torture/execute/20010910-1.c simple case. | |
185 | ||
186 | 2) The cleaning up of spill code is quite profitable. It currently | |
187 | depends on reading tea leaves and chicken entrails left by reload. | |
188 | This pass depends on reload creating a singleton alias set for each | |
189 | spill slot and telling the next dse pass which of these alias sets | |
190 | are the singletons. Rather than analyze the addresses of the | |
191 | spills, dse's spill processing just does analysis of the loads and | |
192 | stores that use those alias sets. There are three cases where this | |
193 | falls short: | |
194 | ||
195 | a) Reload sometimes creates the slot for one mode of access, and | |
196 | then inserts loads and/or stores for a smaller mode. In this | |
197 | case, the current code just punts on the slot. The proper thing | |
198 | to do is to back out and use one bit vector position for each | |
199 | byte of the entity associated with the slot. This depends on | |
200 | KNOWING that reload always generates the accesses for each of the | |
201 | bytes in some canonical (read that easy to understand several | |
202 | passes after reload happens) way. | |
203 | ||
204 | b) Reload sometimes decides that spill slot it allocated was not | |
205 | large enough for the mode and goes back and allocates more slots | |
206 | with the same mode and alias set. The backout in this case is a | |
207 | little more graceful than (a). In this case the slot is unmarked | |
208 | as being a spill slot and if final address comes out to be based | |
48e1416a | 209 | off the frame pointer, the global algorithm handles this slot. |
3072d30e | 210 | |
211 | c) For any pass that may prespill, there is currently no | |
212 | mechanism to tell the dse pass that the slot being used has the | |
213 | special properties that reload uses. It may be that all that is | |
bef304b8 | 214 | required is to have those passes make the same calls that reload |
3072d30e | 215 | does, assuming that the alias sets can be manipulated in the same |
216 | way. */ | |
217 | ||
218 | /* There are limits to the size of constant offsets we model for the | |
219 | global problem. There are certainly test cases, that exceed this | |
220 | limit, however, it is unlikely that there are important programs | |
221 | that really have constant offsets this size. */ | |
222 | #define MAX_OFFSET (64 * 1024) | |
223 | ||
4fb07d00 | 224 | /* Obstack for the DSE dataflow bitmaps. We don't want to put these |
225 | on the default obstack because these bitmaps can grow quite large | |
226 | (~2GB for the small (!) test case of PR54146) and we'll hold on to | |
227 | all that memory until the end of the compiler run. | |
228 | As a bonus, delete_tree_live_info can destroy all the bitmaps by just | |
229 | releasing the whole obstack. */ | |
230 | static bitmap_obstack dse_bitmap_obstack; | |
231 | ||
232 | /* Obstack for other data. As for above: Kinda nice to be able to | |
233 | throw it all away at the end in one big sweep. */ | |
234 | static struct obstack dse_obstack; | |
235 | ||
236 | /* Scratch bitmap for cselib's cselib_expand_value_rtx. */ | |
3072d30e | 237 | static bitmap scratch = NULL; |
4fb07d00 | 238 | |
55c5ac9f | 239 | struct insn_info_type; |
3072d30e | 240 | |
241 | /* This structure holds information about a candidate store. */ | |
48e1416a | 242 | struct store_info |
3072d30e | 243 | { |
244 | ||
245 | /* False means this is a clobber. */ | |
246 | bool is_set; | |
247 | ||
aa140b76 | 248 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
249 | bool is_large; | |
250 | ||
3072d30e | 251 | /* The id of the mem group of the base address. If rtx_varies_p is |
252 | true, this is -1. Otherwise, it is the index into the group | |
253 | table. */ | |
254 | int group_id; | |
48e1416a | 255 | |
3072d30e | 256 | /* This is the cselib value. */ |
257 | cselib_val *cse_base; | |
258 | ||
259 | /* This canonized mem. */ | |
260 | rtx mem; | |
261 | ||
82d2c88b | 262 | /* Canonized MEM address for use by canon_true_dependence. */ |
3072d30e | 263 | rtx mem_addr; |
264 | ||
265 | /* If this is non-zero, it is the alias set of a spill location. */ | |
32c2fdea | 266 | alias_set_type alias_set; |
3072d30e | 267 | |
268 | /* The offset of the first and byte before the last byte associated | |
269 | with the operation. */ | |
aa140b76 | 270 | HOST_WIDE_INT begin, end; |
271 | ||
272 | union | |
273 | { | |
274 | /* A bitmask as wide as the number of bytes in the word that | |
275 | contains a 1 if the byte may be needed. The store is unused if | |
276 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
277 | unsigned HOST_WIDE_INT small_bitmask; | |
278 | ||
279 | struct | |
280 | { | |
281 | /* A bitmap with one bit per byte. Cleared bit means the position | |
282 | is needed. Used if IS_LARGE is false. */ | |
843bd2fa | 283 | bitmap bmap; |
3072d30e | 284 | |
aa140b76 | 285 | /* Number of set bits (i.e. unneeded bytes) in BITMAP. If it is |
286 | equal to END - BEGIN, the whole store is unused. */ | |
287 | int count; | |
288 | } large; | |
289 | } positions_needed; | |
3072d30e | 290 | |
291 | /* The next store info for this insn. */ | |
292 | struct store_info *next; | |
293 | ||
294 | /* The right hand side of the store. This is used if there is a | |
295 | subsequent reload of the mems address somewhere later in the | |
296 | basic block. */ | |
aa140b76 | 297 | rtx rhs; |
298 | ||
299 | /* If rhs is or holds a constant, this contains that constant, | |
300 | otherwise NULL. */ | |
301 | rtx const_rhs; | |
302 | ||
303 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
304 | insn stored. These aren't eliminated early, because doing that | |
305 | might prevent the earlier larger store to be eliminated. */ | |
55c5ac9f | 306 | struct insn_info_type *redundant_reason; |
3072d30e | 307 | }; |
308 | ||
4e43e20a | 309 | /* Return a bitmask with the first N low bits set. */ |
310 | ||
311 | static unsigned HOST_WIDE_INT | |
312 | lowpart_bitmask (int n) | |
313 | { | |
314 | unsigned HOST_WIDE_INT mask = ~(unsigned HOST_WIDE_INT) 0; | |
315 | return mask >> (HOST_BITS_PER_WIDE_INT - n); | |
316 | } | |
317 | ||
3072d30e | 318 | typedef struct store_info *store_info_t; |
55c5ac9f | 319 | static pool_allocator<store_info> cse_store_info_pool ("cse_store_info_pool", |
320 | 100); | |
321 | ||
322 | static pool_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool", | |
323 | 100); | |
3072d30e | 324 | |
325 | /* This structure holds information about a load. These are only | |
326 | built for rtx bases. */ | |
55c5ac9f | 327 | struct read_info_type |
3072d30e | 328 | { |
329 | /* The id of the mem group of the base address. */ | |
330 | int group_id; | |
331 | ||
332 | /* If this is non-zero, it is the alias set of a spill location. */ | |
32c2fdea | 333 | alias_set_type alias_set; |
3072d30e | 334 | |
335 | /* The offset of the first and byte after the last byte associated | |
336 | with the operation. If begin == end == 0, the read did not have | |
337 | a constant offset. */ | |
338 | int begin, end; | |
339 | ||
340 | /* The mem being read. */ | |
341 | rtx mem; | |
342 | ||
343 | /* The next read_info for this insn. */ | |
55c5ac9f | 344 | struct read_info_type *next; |
345 | ||
346 | /* Pool allocation new operator. */ | |
347 | inline void *operator new (size_t) | |
348 | { | |
349 | return pool.allocate (); | |
350 | } | |
351 | ||
352 | /* Delete operator utilizing pool allocation. */ | |
353 | inline void operator delete (void *ptr) | |
354 | { | |
355 | pool.remove ((read_info_type *) ptr); | |
356 | } | |
357 | ||
358 | /* Memory allocation pool. */ | |
359 | static pool_allocator<read_info_type> pool; | |
3072d30e | 360 | }; |
55c5ac9f | 361 | typedef struct read_info_type *read_info_t; |
3072d30e | 362 | |
55c5ac9f | 363 | pool_allocator<read_info_type> read_info_type::pool ("read_info_pool", 100); |
3072d30e | 364 | |
365 | /* One of these records is created for each insn. */ | |
366 | ||
55c5ac9f | 367 | struct insn_info_type |
3072d30e | 368 | { |
369 | /* Set true if the insn contains a store but the insn itself cannot | |
370 | be deleted. This is set if the insn is a parallel and there is | |
371 | more than one non dead output or if the insn is in some way | |
372 | volatile. */ | |
373 | bool cannot_delete; | |
374 | ||
375 | /* This field is only used by the global algorithm. It is set true | |
376 | if the insn contains any read of mem except for a (1). This is | |
377 | also set if the insn is a call or has a clobber mem. If the insn | |
378 | contains a wild read, the use_rec will be null. */ | |
379 | bool wild_read; | |
380 | ||
b4a708fb | 381 | /* This is true only for CALL instructions which could potentially read |
382 | any non-frame memory location. This field is used by the global | |
383 | algorithm. */ | |
384 | bool non_frame_wild_read; | |
385 | ||
17e1318c | 386 | /* This field is only used for the processing of const functions. |
387 | These functions cannot read memory, but they can read the stack | |
16bf64db | 388 | because that is where they may get their parms. We need to be |
389 | this conservative because, like the store motion pass, we don't | |
390 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
391 | Moreover, we need to distinguish two cases: | |
392 | 1. Before reload (register elimination), the stores related to | |
393 | outgoing arguments are stack pointer based and thus deemed | |
394 | of non-constant base in this pass. This requires special | |
395 | handling but also means that the frame pointer based stores | |
396 | need not be killed upon encountering a const function call. | |
397 | 2. After reload, the stores related to outgoing arguments can be | |
398 | either stack pointer or hard frame pointer based. This means | |
399 | that we have no other choice than also killing all the frame | |
400 | pointer based stores upon encountering a const function call. | |
17853422 | 401 | This field is set after reload for const function calls and before |
402 | reload for const tail function calls on targets where arg pointer | |
403 | is the frame pointer. Having this set is less severe than a wild | |
404 | read, it just means that all the frame related stores are killed | |
405 | rather than all the stores. */ | |
16bf64db | 406 | bool frame_read; |
407 | ||
408 | /* This field is only used for the processing of const functions. | |
409 | It is set if the insn may contain a stack pointer based store. */ | |
17e1318c | 410 | bool stack_pointer_based; |
3072d30e | 411 | |
412 | /* This is true if any of the sets within the store contains a | |
413 | cselib base. Such stores can only be deleted by the local | |
414 | algorithm. */ | |
415 | bool contains_cselib_groups; | |
416 | ||
417 | /* The insn. */ | |
ebabb7a3 | 418 | rtx_insn *insn; |
3072d30e | 419 | |
420 | /* The list of mem sets or mem clobbers that are contained in this | |
421 | insn. If the insn is deletable, it contains only one mem set. | |
422 | But it could also contain clobbers. Insns that contain more than | |
423 | one mem set are not deletable, but each of those mems are here in | |
6dfdc153 | 424 | order to provide info to delete other insns. */ |
3072d30e | 425 | store_info_t store_rec; |
426 | ||
427 | /* The linked list of mem uses in this insn. Only the reads from | |
428 | rtx bases are listed here. The reads to cselib bases are | |
429 | completely processed during the first scan and so are never | |
430 | created. */ | |
431 | read_info_t read_rec; | |
432 | ||
5a9ecd4a | 433 | /* The live fixed registers. We assume only fixed registers can |
434 | cause trouble by being clobbered from an expanded pattern; | |
435 | storing only the live fixed registers (rather than all registers) | |
436 | means less memory needs to be allocated / copied for the individual | |
437 | stores. */ | |
438 | regset fixed_regs_live; | |
439 | ||
3072d30e | 440 | /* The prev insn in the basic block. */ |
55c5ac9f | 441 | struct insn_info_type * prev_insn; |
3072d30e | 442 | |
443 | /* The linked list of insns that are in consideration for removal in | |
9d75589a | 444 | the forwards pass through the basic block. This pointer may be |
3072d30e | 445 | trash as it is not cleared when a wild read occurs. The only |
f0b5f617 | 446 | time it is guaranteed to be correct is when the traversal starts |
3072d30e | 447 | at active_local_stores. */ |
55c5ac9f | 448 | struct insn_info_type * next_local_store; |
449 | ||
450 | /* Pool allocation new operator. */ | |
451 | inline void *operator new (size_t) | |
452 | { | |
453 | return pool.allocate (); | |
454 | } | |
455 | ||
456 | /* Delete operator utilizing pool allocation. */ | |
457 | inline void operator delete (void *ptr) | |
458 | { | |
459 | pool.remove ((insn_info_type *) ptr); | |
460 | } | |
461 | ||
462 | /* Memory allocation pool. */ | |
463 | static pool_allocator<insn_info_type> pool; | |
3072d30e | 464 | }; |
55c5ac9f | 465 | typedef struct insn_info_type *insn_info_t; |
3072d30e | 466 | |
55c5ac9f | 467 | pool_allocator<insn_info_type> insn_info_type::pool ("insn_info_pool", 100); |
3072d30e | 468 | |
469 | /* The linked list of stores that are under consideration in this | |
48e1416a | 470 | basic block. */ |
3072d30e | 471 | static insn_info_t active_local_stores; |
1242bee6 | 472 | static int active_local_stores_len; |
3072d30e | 473 | |
55c5ac9f | 474 | struct dse_bb_info_type |
3072d30e | 475 | { |
3072d30e | 476 | /* Pointer to the insn info for the last insn in the block. These |
477 | are linked so this is how all of the insns are reached. During | |
478 | scanning this is the current insn being scanned. */ | |
479 | insn_info_t last_insn; | |
480 | ||
481 | /* The info for the global dataflow problem. */ | |
482 | ||
483 | ||
484 | /* This is set if the transfer function should and in the wild_read | |
485 | bitmap before applying the kill and gen sets. That vector knocks | |
486 | out most of the bits in the bitmap and thus speeds up the | |
487 | operations. */ | |
488 | bool apply_wild_read; | |
489 | ||
a1b0a968 | 490 | /* The following 4 bitvectors hold information about which positions |
491 | of which stores are live or dead. They are indexed by | |
492 | get_bitmap_index. */ | |
493 | ||
3072d30e | 494 | /* The set of store positions that exist in this block before a wild read. */ |
495 | bitmap gen; | |
48e1416a | 496 | |
3072d30e | 497 | /* The set of load positions that exist in this block above the |
498 | same position of a store. */ | |
499 | bitmap kill; | |
500 | ||
501 | /* The set of stores that reach the top of the block without being | |
502 | killed by a read. | |
503 | ||
504 | Do not represent the in if it is all ones. Note that this is | |
505 | what the bitvector should logically be initialized to for a set | |
506 | intersection problem. However, like the kill set, this is too | |
507 | expensive. So initially, the in set will only be created for the | |
508 | exit block and any block that contains a wild read. */ | |
509 | bitmap in; | |
510 | ||
511 | /* The set of stores that reach the bottom of the block from it's | |
512 | successors. | |
513 | ||
514 | Do not represent the in if it is all ones. Note that this is | |
515 | what the bitvector should logically be initialized to for a set | |
516 | intersection problem. However, like the kill and in set, this is | |
517 | too expensive. So what is done is that the confluence operator | |
518 | just initializes the vector from one of the out sets of the | |
519 | successors of the block. */ | |
520 | bitmap out; | |
a1b0a968 | 521 | |
522 | /* The following bitvector is indexed by the reg number. It | |
523 | contains the set of regs that are live at the current instruction | |
524 | being processed. While it contains info for all of the | |
5a9ecd4a | 525 | registers, only the hard registers are actually examined. It is used |
526 | to assure that shift and/or add sequences that are inserted do not | |
9d75589a | 527 | accidentally clobber live hard regs. */ |
a1b0a968 | 528 | bitmap regs_live; |
55c5ac9f | 529 | |
530 | /* Pool allocation new operator. */ | |
531 | inline void *operator new (size_t) | |
532 | { | |
533 | return pool.allocate (); | |
534 | } | |
535 | ||
536 | /* Delete operator utilizing pool allocation. */ | |
537 | inline void operator delete (void *ptr) | |
538 | { | |
539 | pool.remove ((dse_bb_info_type *) ptr); | |
540 | } | |
541 | ||
542 | /* Memory allocation pool. */ | |
543 | static pool_allocator<dse_bb_info_type> pool; | |
3072d30e | 544 | }; |
545 | ||
55c5ac9f | 546 | typedef struct dse_bb_info_type *bb_info_t; |
547 | pool_allocator<dse_bb_info_type> dse_bb_info_type::pool ("bb_info_pool", 100); | |
3072d30e | 548 | |
549 | /* Table to hold all bb_infos. */ | |
550 | static bb_info_t *bb_table; | |
551 | ||
552 | /* There is a group_info for each rtx base that is used to reference | |
553 | memory. There are also not many of the rtx bases because they are | |
554 | very limited in scope. */ | |
555 | ||
48e1416a | 556 | struct group_info |
3072d30e | 557 | { |
558 | /* The actual base of the address. */ | |
559 | rtx rtx_base; | |
560 | ||
561 | /* The sequential id of the base. This allows us to have a | |
562 | canonical ordering of these that is not based on addresses. */ | |
563 | int id; | |
564 | ||
0ac758f7 | 565 | /* True if there are any positions that are to be processed |
566 | globally. */ | |
567 | bool process_globally; | |
568 | ||
569 | /* True if the base of this group is either the frame_pointer or | |
570 | hard_frame_pointer. */ | |
571 | bool frame_related; | |
572 | ||
ec410bf1 | 573 | /* A mem wrapped around the base pointer for the group in order to do |
574 | read dependency. It must be given BLKmode in order to encompass all | |
575 | the possible offsets from the base. */ | |
3072d30e | 576 | rtx base_mem; |
48e1416a | 577 | |
82d2c88b | 578 | /* Canonized version of base_mem's address. */ |
579 | rtx canon_base_addr; | |
3072d30e | 580 | |
581 | /* These two sets of two bitmaps are used to keep track of how many | |
6dfdc153 | 582 | stores are actually referencing that position from this base. We |
3072d30e | 583 | only do this for rtx bases as this will be used to assign |
6dfdc153 | 584 | positions in the bitmaps for the global problem. Bit N is set in |
3072d30e | 585 | store1 on the first store for offset N. Bit N is set in store2 |
586 | for the second store to offset N. This is all we need since we | |
587 | only care about offsets that have two or more stores for them. | |
588 | ||
589 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
590 | for 0 and greater offsets. | |
591 | ||
592 | There is one special case here, for stores into the stack frame, | |
593 | we will or store1 into store2 before deciding which stores look | |
594 | at globally. This is because stores to the stack frame that have | |
595 | no other reads before the end of the function can also be | |
596 | deleted. */ | |
597 | bitmap store1_n, store1_p, store2_n, store2_p; | |
598 | ||
b4a708fb | 599 | /* These bitmaps keep track of offsets in this group escape this function. |
600 | An offset escapes if it corresponds to a named variable whose | |
601 | addressable flag is set. */ | |
602 | bitmap escaped_n, escaped_p; | |
603 | ||
6dfdc153 | 604 | /* The positions in this bitmap have the same assignments as the in, |
3072d30e | 605 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6dfdc153 | 606 | the positions that are occupied by stores for this group. */ |
3072d30e | 607 | bitmap group_kill; |
608 | ||
3072d30e | 609 | /* The offset_map is used to map the offsets from this base into |
6dfdc153 | 610 | positions in the global bitmaps. It is only created after all of |
3072d30e | 611 | the all of stores have been scanned and we know which ones we |
612 | care about. */ | |
48e1416a | 613 | int *offset_map_n, *offset_map_p; |
614 | int offset_map_size_n, offset_map_size_p; | |
55c5ac9f | 615 | |
616 | /* Pool allocation new operator. */ | |
617 | inline void *operator new (size_t) | |
618 | { | |
619 | return pool.allocate (); | |
620 | } | |
621 | ||
622 | /* Delete operator utilizing pool allocation. */ | |
623 | inline void operator delete (void *ptr) | |
624 | { | |
625 | pool.remove ((group_info *) ptr); | |
626 | } | |
627 | ||
628 | /* Memory allocation pool. */ | |
629 | static pool_allocator<group_info> pool; | |
3072d30e | 630 | }; |
631 | typedef struct group_info *group_info_t; | |
c1fdef8e | 632 | typedef const struct group_info *const_group_info_t; |
55c5ac9f | 633 | |
634 | pool_allocator<group_info> group_info::pool ("rtx_group_info_pool", 100); | |
3072d30e | 635 | |
3072d30e | 636 | /* Index into the rtx_group_vec. */ |
637 | static int rtx_group_next_id; | |
638 | ||
3072d30e | 639 | |
f1f41a6c | 640 | static vec<group_info_t> rtx_group_vec; |
3072d30e | 641 | |
642 | ||
643 | /* This structure holds the set of changes that are being deferred | |
644 | when removing read operation. See replace_read. */ | |
48e1416a | 645 | struct deferred_change |
3072d30e | 646 | { |
647 | ||
648 | /* The mem that is being replaced. */ | |
649 | rtx *loc; | |
650 | ||
651 | /* The reg it is being replaced with. */ | |
652 | rtx reg; | |
653 | ||
654 | struct deferred_change *next; | |
55c5ac9f | 655 | |
656 | /* Pool allocation new operator. */ | |
657 | inline void *operator new (size_t) | |
658 | { | |
659 | return pool.allocate (); | |
660 | } | |
661 | ||
662 | /* Delete operator utilizing pool allocation. */ | |
663 | inline void operator delete (void *ptr) | |
664 | { | |
665 | pool.remove ((deferred_change *) ptr); | |
666 | } | |
667 | ||
668 | /* Memory allocation pool. */ | |
669 | static pool_allocator<deferred_change> pool; | |
3072d30e | 670 | }; |
671 | ||
672 | typedef struct deferred_change *deferred_change_t; | |
55c5ac9f | 673 | |
674 | pool_allocator<deferred_change> deferred_change::pool | |
675 | ("deferred_change_pool", 10); | |
3072d30e | 676 | |
677 | static deferred_change_t deferred_change_list = NULL; | |
678 | ||
3072d30e | 679 | /* The group that holds all of the clear_alias_sets. */ |
680 | static group_info_t clear_alias_group; | |
681 | ||
682 | /* The modes of the clear_alias_sets. */ | |
683 | static htab_t clear_alias_mode_table; | |
684 | ||
685 | /* Hash table element to look up the mode for an alias set. */ | |
686 | struct clear_alias_mode_holder | |
687 | { | |
32c2fdea | 688 | alias_set_type alias_set; |
3754d046 | 689 | machine_mode mode; |
3072d30e | 690 | }; |
691 | ||
18d50ae6 | 692 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
ff3ae375 | 693 | this for vararg functions because they play games with the frame. */ |
3072d30e | 694 | static bool stores_off_frame_dead_at_return; |
695 | ||
696 | /* Counter for stats. */ | |
48e1416a | 697 | static int globally_deleted; |
698 | static int locally_deleted; | |
699 | static int spill_deleted; | |
700 | ||
3072d30e | 701 | static bitmap all_blocks; |
702 | ||
b4a708fb | 703 | /* Locations that are killed by calls in the global phase. */ |
704 | static bitmap kill_on_calls; | |
705 | ||
3072d30e | 706 | /* The number of bits used in the global bitmaps. */ |
707 | static unsigned int current_position; | |
3072d30e | 708 | \f |
709 | /*---------------------------------------------------------------------------- | |
710 | Zeroth step. | |
711 | ||
48e1416a | 712 | Initialization. |
3072d30e | 713 | ----------------------------------------------------------------------------*/ |
714 | ||
3072d30e | 715 | |
716 | /* Find the entry associated with ALIAS_SET. */ | |
717 | ||
718 | static struct clear_alias_mode_holder * | |
32c2fdea | 719 | clear_alias_set_lookup (alias_set_type alias_set) |
3072d30e | 720 | { |
721 | struct clear_alias_mode_holder tmp_holder; | |
722 | void **slot; | |
48e1416a | 723 | |
3072d30e | 724 | tmp_holder.alias_set = alias_set; |
725 | slot = htab_find_slot (clear_alias_mode_table, &tmp_holder, NO_INSERT); | |
726 | gcc_assert (*slot); | |
48e1416a | 727 | |
364c0c59 | 728 | return (struct clear_alias_mode_holder *) *slot; |
3072d30e | 729 | } |
730 | ||
731 | ||
732 | /* Hashtable callbacks for maintaining the "bases" field of | |
733 | store_group_info, given that the addresses are function invariants. */ | |
734 | ||
770ff93b | 735 | struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
d1455aa3 | 736 | { |
9969c043 | 737 | static inline hashval_t hash (const group_info *); |
738 | static inline bool equal (const group_info *, const group_info *); | |
d1455aa3 | 739 | }; |
740 | ||
741 | inline bool | |
9969c043 | 742 | invariant_group_base_hasher::equal (const group_info *gi1, |
743 | const group_info *gi2) | |
3072d30e | 744 | { |
3072d30e | 745 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
746 | } | |
747 | ||
d1455aa3 | 748 | inline hashval_t |
9969c043 | 749 | invariant_group_base_hasher::hash (const group_info *gi) |
3072d30e | 750 | { |
3072d30e | 751 | int do_not_record; |
752 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
753 | } | |
754 | ||
d1455aa3 | 755 | /* Tables of group_info structures, hashed by base value. */ |
c1f445d2 | 756 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
d1455aa3 | 757 | |
3072d30e | 758 | |
759 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
760 | ||
761 | static group_info_t | |
762 | get_group_info (rtx base) | |
763 | { | |
48e1416a | 764 | struct group_info tmp_gi; |
765 | group_info_t gi; | |
d1455aa3 | 766 | group_info **slot; |
3072d30e | 767 | |
768 | if (base) | |
769 | { | |
770 | /* Find the store_base_info structure for BASE, creating a new one | |
771 | if necessary. */ | |
772 | tmp_gi.rtx_base = base; | |
c1f445d2 | 773 | slot = rtx_group_table->find_slot (&tmp_gi, INSERT); |
3072d30e | 774 | gi = (group_info_t) *slot; |
775 | } | |
776 | else | |
777 | { | |
778 | if (!clear_alias_group) | |
779 | { | |
55c5ac9f | 780 | clear_alias_group = gi = new group_info; |
3072d30e | 781 | memset (gi, 0, sizeof (struct group_info)); |
782 | gi->id = rtx_group_next_id++; | |
4fb07d00 | 783 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
784 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
785 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
786 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
787 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
788 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
789 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
3072d30e | 790 | gi->process_globally = false; |
791 | gi->offset_map_size_n = 0; | |
792 | gi->offset_map_size_p = 0; | |
793 | gi->offset_map_n = NULL; | |
794 | gi->offset_map_p = NULL; | |
f1f41a6c | 795 | rtx_group_vec.safe_push (gi); |
3072d30e | 796 | } |
797 | return clear_alias_group; | |
798 | } | |
799 | ||
800 | if (gi == NULL) | |
801 | { | |
55c5ac9f | 802 | *slot = gi = new group_info; |
3072d30e | 803 | gi->rtx_base = base; |
804 | gi->id = rtx_group_next_id++; | |
ec410bf1 | 805 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
82d2c88b | 806 | gi->canon_base_addr = canon_rtx (base); |
4fb07d00 | 807 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
808 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
809 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
810 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
811 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
812 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
813 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
3072d30e | 814 | gi->process_globally = false; |
48e1416a | 815 | gi->frame_related = |
3072d30e | 816 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
817 | gi->offset_map_size_n = 0; | |
818 | gi->offset_map_size_p = 0; | |
819 | gi->offset_map_n = NULL; | |
820 | gi->offset_map_p = NULL; | |
f1f41a6c | 821 | rtx_group_vec.safe_push (gi); |
3072d30e | 822 | } |
823 | ||
824 | return gi; | |
825 | } | |
826 | ||
827 | ||
828 | /* Initialization of data structures. */ | |
829 | ||
830 | static void | |
831 | dse_step0 (void) | |
832 | { | |
833 | locally_deleted = 0; | |
834 | globally_deleted = 0; | |
835 | spill_deleted = 0; | |
836 | ||
4fb07d00 | 837 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
838 | gcc_obstack_init (&dse_obstack); | |
839 | ||
840 | scratch = BITMAP_ALLOC (®_obstack); | |
841 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
3072d30e | 842 | |
3072d30e | 843 | |
c1f445d2 | 844 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
3072d30e | 845 | |
fe672ac0 | 846 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
3072d30e | 847 | rtx_group_next_id = 0; |
848 | ||
18d50ae6 | 849 | stores_off_frame_dead_at_return = !cfun->stdarg; |
3072d30e | 850 | |
851 | init_alias_analysis (); | |
48e1416a | 852 | |
e85eaec5 | 853 | clear_alias_group = NULL; |
3072d30e | 854 | } |
855 | ||
856 | ||
857 | \f | |
858 | /*---------------------------------------------------------------------------- | |
859 | First step. | |
860 | ||
861 | Scan all of the insns. Any random ordering of the blocks is fine. | |
bef304b8 | 862 | Each block is scanned in forward order to accommodate cselib which |
3072d30e | 863 | is used to remove stores with non-constant bases. |
864 | ----------------------------------------------------------------------------*/ | |
865 | ||
866 | /* Delete all of the store_info recs from INSN_INFO. */ | |
867 | ||
48e1416a | 868 | static void |
3072d30e | 869 | free_store_info (insn_info_t insn_info) |
870 | { | |
871 | store_info_t store_info = insn_info->store_rec; | |
872 | while (store_info) | |
873 | { | |
874 | store_info_t next = store_info->next; | |
aa140b76 | 875 | if (store_info->is_large) |
843bd2fa | 876 | BITMAP_FREE (store_info->positions_needed.large.bmap); |
3072d30e | 877 | if (store_info->cse_base) |
55c5ac9f | 878 | cse_store_info_pool.remove (store_info); |
3072d30e | 879 | else |
55c5ac9f | 880 | rtx_store_info_pool.remove (store_info); |
3072d30e | 881 | store_info = next; |
882 | } | |
883 | ||
884 | insn_info->cannot_delete = true; | |
885 | insn_info->contains_cselib_groups = false; | |
886 | insn_info->store_rec = NULL; | |
887 | } | |
888 | ||
5a9ecd4a | 889 | typedef struct |
890 | { | |
4cd001d5 | 891 | rtx_insn *first, *current; |
5a9ecd4a | 892 | regset fixed_regs_live; |
893 | bool failure; | |
894 | } note_add_store_info; | |
895 | ||
896 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
897 | Check if a register is clobbered which is live afterwards. */ | |
898 | ||
899 | static void | |
900 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
901 | { | |
4cd001d5 | 902 | rtx_insn *insn; |
5a9ecd4a | 903 | note_add_store_info *info = (note_add_store_info *) data; |
5a9ecd4a | 904 | |
905 | if (!REG_P (loc)) | |
906 | return; | |
907 | ||
908 | /* If this register is referenced by the current or an earlier insn, | |
909 | that's OK. E.g. this applies to the register that is being incremented | |
910 | with this addition. */ | |
911 | for (insn = info->first; | |
912 | insn != NEXT_INSN (info->current); | |
913 | insn = NEXT_INSN (insn)) | |
914 | if (reg_referenced_p (loc, PATTERN (insn))) | |
915 | return; | |
916 | ||
917 | /* If we come here, we have a clobber of a register that's only OK | |
918 | if that register is not live. If we don't have liveness information | |
919 | available, fail now. */ | |
920 | if (!info->fixed_regs_live) | |
921 | { | |
6a298741 | 922 | info->failure = true; |
5a9ecd4a | 923 | return; |
924 | } | |
925 | /* Now check if this is a live fixed register. */ | |
6a298741 | 926 | unsigned int end_regno = END_REGNO (loc); |
927 | for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) | |
928 | if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) | |
929 | info->failure = true; | |
5a9ecd4a | 930 | } |
931 | ||
1f864115 | 932 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
933 | SRC + SRCOFF before insn ARG. */ | |
3072d30e | 934 | |
935 | static int | |
1f864115 | 936 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
937 | rtx op ATTRIBUTE_UNUSED, | |
938 | rtx dest, rtx src, rtx srcoff, void *arg) | |
3072d30e | 939 | { |
5a9ecd4a | 940 | insn_info_t insn_info = (insn_info_t) arg; |
4cd001d5 | 941 | rtx_insn *insn = insn_info->insn, *new_insn, *cur; |
5a9ecd4a | 942 | note_add_store_info info; |
48e1416a | 943 | |
1f864115 | 944 | /* We can reuse all operands without copying, because we are about |
945 | to delete the insn that contained it. */ | |
5a9ecd4a | 946 | if (srcoff) |
a280a5da | 947 | { |
948 | start_sequence (); | |
949 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
950 | new_insn = get_insns (); | |
951 | end_sequence (); | |
952 | } | |
5a9ecd4a | 953 | else |
f9a00e9e | 954 | new_insn = gen_move_insn (dest, src); |
5a9ecd4a | 955 | info.first = new_insn; |
956 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
957 | info.failure = false; | |
958 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
959 | { | |
960 | info.current = cur; | |
961 | note_stores (PATTERN (cur), note_add_store, &info); | |
962 | } | |
3072d30e | 963 | |
5a9ecd4a | 964 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
965 | return it immediately, communicating the failure to its caller. */ | |
966 | if (info.failure) | |
967 | return 1; | |
968 | ||
969 | emit_insn_before (new_insn, insn); | |
3072d30e | 970 | |
623ad592 | 971 | return 0; |
3072d30e | 972 | } |
973 | ||
5a9ecd4a | 974 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
975 | is there, is split into a separate insn. | |
976 | Return true on success (or if there was nothing to do), false on failure. */ | |
3072d30e | 977 | |
5a9ecd4a | 978 | static bool |
979 | check_for_inc_dec_1 (insn_info_t insn_info) | |
3072d30e | 980 | { |
ebabb7a3 | 981 | rtx_insn *insn = insn_info->insn; |
3072d30e | 982 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
983 | if (note) | |
623ad592 | 984 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
985 | insn_info) == 0; | |
5a9ecd4a | 986 | return true; |
3072d30e | 987 | } |
988 | ||
989 | ||
5a9ecd4a | 990 | /* Entry point for postreload. If you work on reload_cse, or you need this |
991 | anywhere else, consider if you can provide register liveness information | |
992 | and add a parameter to this function so that it can be passed down in | |
993 | insn_info.fixed_regs_live. */ | |
994 | bool | |
ebabb7a3 | 995 | check_for_inc_dec (rtx_insn *insn) |
5a9ecd4a | 996 | { |
55c5ac9f | 997 | insn_info_type insn_info; |
5a9ecd4a | 998 | rtx note; |
999 | ||
1000 | insn_info.insn = insn; | |
1001 | insn_info.fixed_regs_live = NULL; | |
1002 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
1003 | if (note) | |
623ad592 | 1004 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
1005 | &insn_info) == 0; | |
5a9ecd4a | 1006 | return true; |
1007 | } | |
1008 | ||
48e1416a | 1009 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
3072d30e | 1010 | |
1011 | static void | |
1012 | delete_dead_store_insn (insn_info_t insn_info) | |
1013 | { | |
1014 | read_info_t read_info; | |
1015 | ||
1016 | if (!dbg_cnt (dse)) | |
1017 | return; | |
1018 | ||
5a9ecd4a | 1019 | if (!check_for_inc_dec_1 (insn_info)) |
1020 | return; | |
1ca59310 | 1021 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1022 | { |
48e1416a | 1023 | fprintf (dump_file, "Locally deleting insn %d ", |
3072d30e | 1024 | INSN_UID (insn_info->insn)); |
1025 | if (insn_info->store_rec->alias_set) | |
48e1416a | 1026 | fprintf (dump_file, "alias set %d\n", |
32c2fdea | 1027 | (int) insn_info->store_rec->alias_set); |
3072d30e | 1028 | else |
1029 | fprintf (dump_file, "\n"); | |
1030 | } | |
1031 | ||
1032 | free_store_info (insn_info); | |
1033 | read_info = insn_info->read_rec; | |
48e1416a | 1034 | |
3072d30e | 1035 | while (read_info) |
1036 | { | |
1037 | read_info_t next = read_info->next; | |
55c5ac9f | 1038 | delete read_info; |
3072d30e | 1039 | read_info = next; |
1040 | } | |
1041 | insn_info->read_rec = NULL; | |
1042 | ||
1043 | delete_insn (insn_info->insn); | |
1044 | locally_deleted++; | |
1045 | insn_info->insn = NULL; | |
1046 | ||
1047 | insn_info->wild_read = false; | |
1048 | } | |
1049 | ||
f7b5f694 | 1050 | /* Return whether DECL, a local variable, can possibly escape the current |
1051 | function scope. */ | |
1052 | ||
1053 | static bool | |
1054 | local_variable_can_escape (tree decl) | |
1055 | { | |
1056 | if (TREE_ADDRESSABLE (decl)) | |
1057 | return true; | |
1058 | ||
1059 | /* If this is a partitioned variable, we need to consider all the variables | |
1060 | in the partition. This is necessary because a store into one of them can | |
1061 | be replaced with a store into another and this may not change the outcome | |
1062 | of the escape analysis. */ | |
1063 | if (cfun->gimple_df->decls_to_pointers != NULL) | |
1064 | { | |
5f8841a5 | 1065 | tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); |
f7b5f694 | 1066 | if (namep) |
5f8841a5 | 1067 | return TREE_ADDRESSABLE (*namep); |
f7b5f694 | 1068 | } |
1069 | ||
1070 | return false; | |
1071 | } | |
1072 | ||
1073 | /* Return whether EXPR can possibly escape the current function scope. */ | |
1074 | ||
b4a708fb | 1075 | static bool |
1076 | can_escape (tree expr) | |
1077 | { | |
1078 | tree base; | |
1079 | if (!expr) | |
1080 | return true; | |
1081 | base = get_base_address (expr); | |
1082 | if (DECL_P (base) | |
f7b5f694 | 1083 | && !may_be_aliased (base) |
1084 | && !(TREE_CODE (base) == VAR_DECL | |
1085 | && !DECL_EXTERNAL (base) | |
1086 | && !TREE_STATIC (base) | |
1087 | && local_variable_can_escape (base))) | |
b4a708fb | 1088 | return false; |
1089 | return true; | |
1090 | } | |
3072d30e | 1091 | |
1092 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
1093 | OFFSET and WIDTH. */ | |
1094 | ||
1095 | static void | |
b4a708fb | 1096 | set_usage_bits (group_info_t group, HOST_WIDE_INT offset, HOST_WIDE_INT width, |
1097 | tree expr) | |
3072d30e | 1098 | { |
1099 | HOST_WIDE_INT i; | |
b4a708fb | 1100 | bool expr_escapes = can_escape (expr); |
aa140b76 | 1101 | if (offset > -MAX_OFFSET && offset + width < MAX_OFFSET) |
3072d30e | 1102 | for (i=offset; i<offset+width; i++) |
1103 | { | |
1104 | bitmap store1; | |
1105 | bitmap store2; | |
b4a708fb | 1106 | bitmap escaped; |
3072d30e | 1107 | int ai; |
1108 | if (i < 0) | |
1109 | { | |
1110 | store1 = group->store1_n; | |
1111 | store2 = group->store2_n; | |
b4a708fb | 1112 | escaped = group->escaped_n; |
3072d30e | 1113 | ai = -i; |
1114 | } | |
1115 | else | |
1116 | { | |
1117 | store1 = group->store1_p; | |
1118 | store2 = group->store2_p; | |
b4a708fb | 1119 | escaped = group->escaped_p; |
3072d30e | 1120 | ai = i; |
1121 | } | |
48e1416a | 1122 | |
6ef9bbe0 | 1123 | if (!bitmap_set_bit (store1, ai)) |
3072d30e | 1124 | bitmap_set_bit (store2, ai); |
48e1416a | 1125 | else |
3072d30e | 1126 | { |
3072d30e | 1127 | if (i < 0) |
1128 | { | |
1129 | if (group->offset_map_size_n < ai) | |
1130 | group->offset_map_size_n = ai; | |
1131 | } | |
1132 | else | |
1133 | { | |
1134 | if (group->offset_map_size_p < ai) | |
1135 | group->offset_map_size_p = ai; | |
1136 | } | |
1137 | } | |
b4a708fb | 1138 | if (expr_escapes) |
1139 | bitmap_set_bit (escaped, ai); | |
3072d30e | 1140 | } |
1141 | } | |
1142 | ||
b4a708fb | 1143 | static void |
1144 | reset_active_stores (void) | |
1145 | { | |
1146 | active_local_stores = NULL; | |
1147 | active_local_stores_len = 0; | |
1148 | } | |
3072d30e | 1149 | |
b4a708fb | 1150 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
3072d30e | 1151 | |
1152 | static void | |
b4a708fb | 1153 | free_read_records (bb_info_t bb_info) |
3072d30e | 1154 | { |
1155 | insn_info_t insn_info = bb_info->last_insn; | |
1156 | read_info_t *ptr = &insn_info->read_rec; | |
3072d30e | 1157 | while (*ptr) |
1158 | { | |
1159 | read_info_t next = (*ptr)->next; | |
32c2fdea | 1160 | if ((*ptr)->alias_set == 0) |
3072d30e | 1161 | { |
55c5ac9f | 1162 | delete *ptr; |
3072d30e | 1163 | *ptr = next; |
b4a708fb | 1164 | } |
48e1416a | 1165 | else |
b4a708fb | 1166 | ptr = &(*ptr)->next; |
3072d30e | 1167 | } |
b4a708fb | 1168 | } |
1169 | ||
1170 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1171 | ||
1172 | static void | |
1173 | add_wild_read (bb_info_t bb_info) | |
1174 | { | |
1175 | insn_info_t insn_info = bb_info->last_insn; | |
3072d30e | 1176 | insn_info->wild_read = true; |
b4a708fb | 1177 | free_read_records (bb_info); |
1178 | reset_active_stores (); | |
3072d30e | 1179 | } |
1180 | ||
b4a708fb | 1181 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1182 | non-frame locations. */ | |
1183 | ||
1184 | static void | |
1185 | add_non_frame_wild_read (bb_info_t bb_info) | |
1186 | { | |
1187 | insn_info_t insn_info = bb_info->last_insn; | |
1188 | insn_info->non_frame_wild_read = true; | |
1189 | free_read_records (bb_info); | |
1190 | reset_active_stores (); | |
1191 | } | |
3072d30e | 1192 | |
17e1318c | 1193 | /* Return true if X is a constant or one of the registers that behave |
1194 | as a constant over the life of a function. This is equivalent to | |
1195 | !rtx_varies_p for memory addresses. */ | |
3072d30e | 1196 | |
1197 | static bool | |
1198 | const_or_frame_p (rtx x) | |
1199 | { | |
0349edce | 1200 | if (CONSTANT_P (x)) |
1201 | return true; | |
1202 | ||
1203 | if (GET_CODE (x) == REG) | |
3072d30e | 1204 | { |
3072d30e | 1205 | /* Note that we have to test for the actual rtx used for the frame |
1206 | and arg pointers and not just the register number in case we have | |
1207 | eliminated the frame and/or arg pointer and are using it | |
1208 | for pseudos. */ | |
1209 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1210 | /* The arg pointer varies if it is not a fixed register. */ | |
1211 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1212 | || x == pic_offset_table_rtx) | |
1213 | return true; | |
1214 | return false; | |
3072d30e | 1215 | } |
55c5ac9f | 1216 | |
0349edce | 1217 | return false; |
3072d30e | 1218 | } |
1219 | ||
48e1416a | 1220 | /* Take all reasonable action to put the address of MEM into the form |
1221 | that we can do analysis on. | |
3072d30e | 1222 | |
1223 | The gold standard is to get the address into the form: address + | |
1224 | OFFSET where address is something that rtx_varies_p considers a | |
1225 | constant. When we can get the address in this form, we can do | |
1226 | global analysis on it. Note that for constant bases, address is | |
1227 | not actually returned, only the group_id. The address can be | |
1228 | obtained from that. | |
1229 | ||
1230 | If that fails, we try cselib to get a value we can at least use | |
48e1416a | 1231 | locally. If that fails we return false. |
1232 | ||
3072d30e | 1233 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1234 | group for non_varying bases. | |
1235 | ||
1236 | FOR_READ is true if this is a mem read and false if not. */ | |
1237 | ||
1238 | static bool | |
1239 | canon_address (rtx mem, | |
32c2fdea | 1240 | alias_set_type *alias_set_out, |
3072d30e | 1241 | int *group_id, |
48e1416a | 1242 | HOST_WIDE_INT *offset, |
3072d30e | 1243 | cselib_val **base) |
1244 | { | |
3754d046 | 1245 | machine_mode address_mode = get_address_mode (mem); |
3072d30e | 1246 | rtx mem_address = XEXP (mem, 0); |
1247 | rtx expanded_address, address; | |
627540ce | 1248 | int expanded; |
1249 | ||
3072d30e | 1250 | *alias_set_out = 0; |
1251 | ||
1f864115 | 1252 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
3072d30e | 1253 | |
1ca59310 | 1254 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1255 | { |
1256 | fprintf (dump_file, " mem: "); | |
1257 | print_inline_rtx (dump_file, mem_address, 0); | |
1258 | fprintf (dump_file, "\n"); | |
1259 | } | |
1260 | ||
627540ce | 1261 | /* First see if just canon_rtx (mem_address) is const or frame, |
1262 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1263 | address = NULL_RTX; | |
1264 | for (expanded = 0; expanded < 2; expanded++) | |
1265 | { | |
1266 | if (expanded) | |
1267 | { | |
1268 | /* Use cselib to replace all of the reg references with the full | |
48e1416a | 1269 | expression. This will take care of the case where we have |
3072d30e | 1270 | |
627540ce | 1271 | r_x = base + offset; |
1272 | val = *r_x; | |
48e1416a | 1273 | |
1274 | by making it into | |
3072d30e | 1275 | |
627540ce | 1276 | val = *(base + offset); */ |
3072d30e | 1277 | |
627540ce | 1278 | expanded_address = cselib_expand_value_rtx (mem_address, |
1279 | scratch, 5); | |
3072d30e | 1280 | |
627540ce | 1281 | /* If this fails, just go with the address from first |
1282 | iteration. */ | |
1283 | if (!expanded_address) | |
1284 | break; | |
1285 | } | |
1286 | else | |
1287 | expanded_address = mem_address; | |
3072d30e | 1288 | |
627540ce | 1289 | /* Split the address into canonical BASE + OFFSET terms. */ |
1290 | address = canon_rtx (expanded_address); | |
3072d30e | 1291 | |
627540ce | 1292 | *offset = 0; |
3072d30e | 1293 | |
1ca59310 | 1294 | if (dump_file && (dump_flags & TDF_DETAILS)) |
627540ce | 1295 | { |
1296 | if (expanded) | |
1297 | { | |
1298 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1299 | print_inline_rtx (dump_file, expanded_address, 0); | |
1300 | fprintf (dump_file, "\n"); | |
1301 | } | |
3072d30e | 1302 | |
627540ce | 1303 | fprintf (dump_file, "\n after canon_rtx address: "); |
1304 | print_inline_rtx (dump_file, address, 0); | |
1305 | fprintf (dump_file, "\n"); | |
1306 | } | |
3072d30e | 1307 | |
627540ce | 1308 | if (GET_CODE (address) == CONST) |
1309 | address = XEXP (address, 0); | |
3072d30e | 1310 | |
627540ce | 1311 | if (GET_CODE (address) == PLUS |
1312 | && CONST_INT_P (XEXP (address, 1))) | |
1313 | { | |
1314 | *offset = INTVAL (XEXP (address, 1)); | |
1315 | address = XEXP (address, 0); | |
1316 | } | |
3072d30e | 1317 | |
bd1a81f7 | 1318 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1319 | && const_or_frame_p (address)) | |
3072d30e | 1320 | { |
627540ce | 1321 | group_info_t group = get_group_info (address); |
1322 | ||
1ca59310 | 1323 | if (dump_file && (dump_flags & TDF_DETAILS)) |
627540ce | 1324 | fprintf (dump_file, " gid=%d offset=%d \n", |
1325 | group->id, (int)*offset); | |
1326 | *base = NULL; | |
1327 | *group_id = group->id; | |
1328 | return true; | |
3072d30e | 1329 | } |
627540ce | 1330 | } |
1331 | ||
1f864115 | 1332 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
627540ce | 1333 | *group_id = -1; |
1334 | ||
1335 | if (*base == NULL) | |
1336 | { | |
1ca59310 | 1337 | if (dump_file && (dump_flags & TDF_DETAILS)) |
627540ce | 1338 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1339 | return false; | |
3072d30e | 1340 | } |
1ca59310 | 1341 | if (dump_file && (dump_flags & TDF_DETAILS)) |
01df1184 | 1342 | fprintf (dump_file, " varying cselib base=%u:%u offset = %d\n", |
1343 | (*base)->uid, (*base)->hash, (int)*offset); | |
3072d30e | 1344 | return true; |
1345 | } | |
1346 | ||
1347 | ||
1348 | /* Clear the rhs field from the active_local_stores array. */ | |
1349 | ||
1350 | static void | |
1351 | clear_rhs_from_active_local_stores (void) | |
1352 | { | |
1353 | insn_info_t ptr = active_local_stores; | |
1354 | ||
1355 | while (ptr) | |
1356 | { | |
1357 | store_info_t store_info = ptr->store_rec; | |
1358 | /* Skip the clobbers. */ | |
1359 | while (!store_info->is_set) | |
1360 | store_info = store_info->next; | |
1361 | ||
1362 | store_info->rhs = NULL; | |
aa140b76 | 1363 | store_info->const_rhs = NULL; |
3072d30e | 1364 | |
1365 | ptr = ptr->next_local_store; | |
1366 | } | |
1367 | } | |
1368 | ||
1369 | ||
aa140b76 | 1370 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1371 | ||
1372 | static inline void | |
1373 | set_position_unneeded (store_info_t s_info, int pos) | |
1374 | { | |
1375 | if (__builtin_expect (s_info->is_large, false)) | |
1376 | { | |
6ef9bbe0 | 1377 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1378 | s_info->positions_needed.large.count++; | |
aa140b76 | 1379 | } |
1380 | else | |
1381 | s_info->positions_needed.small_bitmask | |
1382 | &= ~(((unsigned HOST_WIDE_INT) 1) << pos); | |
1383 | } | |
1384 | ||
1385 | /* Mark the whole store S_INFO as unneeded. */ | |
1386 | ||
1387 | static inline void | |
1388 | set_all_positions_unneeded (store_info_t s_info) | |
1389 | { | |
1390 | if (__builtin_expect (s_info->is_large, false)) | |
1391 | { | |
1392 | int pos, end = s_info->end - s_info->begin; | |
1393 | for (pos = 0; pos < end; pos++) | |
843bd2fa | 1394 | bitmap_set_bit (s_info->positions_needed.large.bmap, pos); |
aa140b76 | 1395 | s_info->positions_needed.large.count = end; |
1396 | } | |
1397 | else | |
1398 | s_info->positions_needed.small_bitmask = (unsigned HOST_WIDE_INT) 0; | |
1399 | } | |
1400 | ||
1401 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1402 | ||
1403 | static inline bool | |
1404 | any_positions_needed_p (store_info_t s_info) | |
1405 | { | |
1406 | if (__builtin_expect (s_info->is_large, false)) | |
1407 | return (s_info->positions_needed.large.count | |
1408 | < s_info->end - s_info->begin); | |
1409 | else | |
1410 | return (s_info->positions_needed.small_bitmask | |
1411 | != (unsigned HOST_WIDE_INT) 0); | |
1412 | } | |
1413 | ||
1414 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
1415 | store are needed. */ | |
1416 | ||
1417 | static inline bool | |
1418 | all_positions_needed_p (store_info_t s_info, int start, int width) | |
1419 | { | |
1420 | if (__builtin_expect (s_info->is_large, false)) | |
1421 | { | |
1422 | int end = start + width; | |
1423 | while (start < end) | |
843bd2fa | 1424 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, start++)) |
aa140b76 | 1425 | return false; |
1426 | return true; | |
1427 | } | |
1428 | else | |
1429 | { | |
1430 | unsigned HOST_WIDE_INT mask = lowpart_bitmask (width) << start; | |
1431 | return (s_info->positions_needed.small_bitmask & mask) == mask; | |
1432 | } | |
1433 | } | |
1434 | ||
1435 | ||
3754d046 | 1436 | static rtx get_stored_val (store_info_t, machine_mode, HOST_WIDE_INT, |
aa140b76 | 1437 | HOST_WIDE_INT, basic_block, bool); |
1438 | ||
1439 | ||
3072d30e | 1440 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1441 | there is a candidate store, after adding it to the appropriate | |
1442 | local store group if so. */ | |
1443 | ||
1444 | static int | |
1445 | record_store (rtx body, bb_info_t bb_info) | |
1446 | { | |
82d2c88b | 1447 | rtx mem, rhs, const_rhs, mem_addr; |
3072d30e | 1448 | HOST_WIDE_INT offset = 0; |
1449 | HOST_WIDE_INT width = 0; | |
32c2fdea | 1450 | alias_set_type spill_alias_set; |
3072d30e | 1451 | insn_info_t insn_info = bb_info->last_insn; |
1452 | store_info_t store_info = NULL; | |
1453 | int group_id; | |
1454 | cselib_val *base = NULL; | |
aa140b76 | 1455 | insn_info_t ptr, last, redundant_reason; |
3072d30e | 1456 | bool store_is_unused; |
1457 | ||
1458 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1459 | return 0; | |
1460 | ||
aa140b76 | 1461 | mem = SET_DEST (body); |
1462 | ||
3072d30e | 1463 | /* If this is not used, then this cannot be used to keep the insn |
1464 | from being deleted. On the other hand, it does provide something | |
1465 | that can be used to prove that another store is dead. */ | |
1466 | store_is_unused | |
aa140b76 | 1467 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
3072d30e | 1468 | |
1469 | /* Check whether that value is a suitable memory location. */ | |
3072d30e | 1470 | if (!MEM_P (mem)) |
1471 | { | |
1472 | /* If the set or clobber is unused, then it does not effect our | |
1473 | ability to get rid of the entire insn. */ | |
1474 | if (!store_is_unused) | |
1475 | insn_info->cannot_delete = true; | |
1476 | return 0; | |
1477 | } | |
1478 | ||
1479 | /* At this point we know mem is a mem. */ | |
1480 | if (GET_MODE (mem) == BLKmode) | |
1481 | { | |
1482 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
1483 | { | |
1ca59310 | 1484 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1485 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1486 | add_wild_read (bb_info); | |
1487 | insn_info->cannot_delete = true; | |
aa140b76 | 1488 | return 0; |
3072d30e | 1489 | } |
aa140b76 | 1490 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1491 | as memset (addr, 0, 36); */ | |
5b2a69fa | 1492 | else if (!MEM_SIZE_KNOWN_P (mem) |
1493 | || MEM_SIZE (mem) <= 0 | |
1494 | || MEM_SIZE (mem) > MAX_OFFSET | |
aa140b76 | 1495 | || GET_CODE (body) != SET |
aa140b76 | 1496 | || !CONST_INT_P (SET_SRC (body))) |
3072d30e | 1497 | { |
aa140b76 | 1498 | if (!store_is_unused) |
1499 | { | |
1500 | /* If the set or clobber is unused, then it does not effect our | |
1501 | ability to get rid of the entire insn. */ | |
1502 | insn_info->cannot_delete = true; | |
1503 | clear_rhs_from_active_local_stores (); | |
1504 | } | |
1505 | return 0; | |
3072d30e | 1506 | } |
3072d30e | 1507 | } |
1508 | ||
1509 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1510 | if (MEM_VOLATILE_P (mem)) | |
aa140b76 | 1511 | insn_info->cannot_delete = true; |
3072d30e | 1512 | |
1513 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
1514 | { | |
1515 | clear_rhs_from_active_local_stores (); | |
1516 | return 0; | |
1517 | } | |
1518 | ||
aa140b76 | 1519 | if (GET_MODE (mem) == BLKmode) |
5b2a69fa | 1520 | width = MEM_SIZE (mem); |
aa140b76 | 1521 | else |
e4209874 | 1522 | width = GET_MODE_SIZE (GET_MODE (mem)); |
3072d30e | 1523 | |
1524 | if (spill_alias_set) | |
1525 | { | |
1526 | bitmap store1 = clear_alias_group->store1_p; | |
1527 | bitmap store2 = clear_alias_group->store2_p; | |
aa140b76 | 1528 | |
1529 | gcc_assert (GET_MODE (mem) != BLKmode); | |
48e1416a | 1530 | |
6ef9bbe0 | 1531 | if (!bitmap_set_bit (store1, spill_alias_set)) |
3072d30e | 1532 | bitmap_set_bit (store2, spill_alias_set); |
48e1416a | 1533 | |
3072d30e | 1534 | if (clear_alias_group->offset_map_size_p < spill_alias_set) |
1535 | clear_alias_group->offset_map_size_p = spill_alias_set; | |
48e1416a | 1536 | |
55c5ac9f | 1537 | store_info = rtx_store_info_pool.allocate (); |
3072d30e | 1538 | |
1ca59310 | 1539 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1540 | fprintf (dump_file, " processing spill store %d(%s)\n", |
32c2fdea | 1541 | (int) spill_alias_set, GET_MODE_NAME (GET_MODE (mem))); |
3072d30e | 1542 | } |
1543 | else if (group_id >= 0) | |
1544 | { | |
1545 | /* In the restrictive case where the base is a constant or the | |
1546 | frame pointer we can do global analysis. */ | |
48e1416a | 1547 | |
1548 | group_info_t group | |
f1f41a6c | 1549 | = rtx_group_vec[group_id]; |
b4a708fb | 1550 | tree expr = MEM_EXPR (mem); |
48e1416a | 1551 | |
55c5ac9f | 1552 | store_info = rtx_store_info_pool.allocate (); |
b4a708fb | 1553 | set_usage_bits (group, offset, width, expr); |
3072d30e | 1554 | |
1ca59310 | 1555 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1556 | fprintf (dump_file, " processing const base store gid=%d[%d..%d)\n", |
1557 | group_id, (int)offset, (int)(offset+width)); | |
1558 | } | |
1559 | else | |
1560 | { | |
86e87ef6 | 1561 | if (may_be_sp_based_p (XEXP (mem, 0))) |
17e1318c | 1562 | insn_info->stack_pointer_based = true; |
3072d30e | 1563 | insn_info->contains_cselib_groups = true; |
17e1318c | 1564 | |
55c5ac9f | 1565 | store_info = cse_store_info_pool.allocate (); |
3072d30e | 1566 | group_id = -1; |
1567 | ||
1ca59310 | 1568 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1569 | fprintf (dump_file, " processing cselib store [%d..%d)\n", |
1570 | (int)offset, (int)(offset+width)); | |
1571 | } | |
1572 | ||
aa140b76 | 1573 | const_rhs = rhs = NULL_RTX; |
1574 | if (GET_CODE (body) == SET | |
1575 | /* No place to keep the value after ra. */ | |
1576 | && !reload_completed | |
1577 | && (REG_P (SET_SRC (body)) | |
1578 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1579 | || CONSTANT_P (SET_SRC (body))) | |
1580 | && !MEM_VOLATILE_P (mem) | |
1581 | /* Sometimes the store and reload is used for truncation and | |
1582 | rounding. */ | |
1583 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1584 | { | |
1585 | rhs = SET_SRC (body); | |
1586 | if (CONSTANT_P (rhs)) | |
1587 | const_rhs = rhs; | |
1588 | else if (body == PATTERN (insn_info->insn)) | |
1589 | { | |
1590 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1591 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1592 | const_rhs = XEXP (tem, 0); | |
1593 | } | |
1594 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1595 | { | |
1596 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1597 | ||
1598 | if (tem && CONSTANT_P (tem)) | |
1599 | const_rhs = tem; | |
1600 | } | |
1601 | } | |
1602 | ||
3072d30e | 1603 | /* Check to see if this stores causes some other stores to be |
1604 | dead. */ | |
1605 | ptr = active_local_stores; | |
1606 | last = NULL; | |
aa140b76 | 1607 | redundant_reason = NULL; |
82d2c88b | 1608 | mem = canon_rtx (mem); |
1609 | /* For alias_set != 0 canon_true_dependence should be never called. */ | |
1610 | if (spill_alias_set) | |
1611 | mem_addr = NULL_RTX; | |
1612 | else | |
1613 | { | |
1614 | if (group_id < 0) | |
1615 | mem_addr = base->val_rtx; | |
1616 | else | |
1617 | { | |
1618 | group_info_t group | |
f1f41a6c | 1619 | = rtx_group_vec[group_id]; |
82d2c88b | 1620 | mem_addr = group->canon_base_addr; |
1621 | } | |
90f3e775 | 1622 | /* get_addr can only handle VALUE but cannot handle expr like: |
1623 | VALUE + OFFSET, so call get_addr to get original addr for | |
1624 | mem_addr before plus_constant. */ | |
1625 | mem_addr = get_addr (mem_addr); | |
82d2c88b | 1626 | if (offset) |
87cf5753 | 1627 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
82d2c88b | 1628 | } |
3072d30e | 1629 | |
1630 | while (ptr) | |
1631 | { | |
1632 | insn_info_t next = ptr->next_local_store; | |
1633 | store_info_t s_info = ptr->store_rec; | |
9ce37fa7 | 1634 | bool del = true; |
3072d30e | 1635 | |
1636 | /* Skip the clobbers. We delete the active insn if this insn | |
6dfdc153 | 1637 | shadows the set. To have been put on the active list, it |
3072d30e | 1638 | has exactly on set. */ |
1639 | while (!s_info->is_set) | |
1640 | s_info = s_info->next; | |
1641 | ||
1642 | if (s_info->alias_set != spill_alias_set) | |
9ce37fa7 | 1643 | del = false; |
3072d30e | 1644 | else if (s_info->alias_set) |
1645 | { | |
48e1416a | 1646 | struct clear_alias_mode_holder *entry |
3072d30e | 1647 | = clear_alias_set_lookup (s_info->alias_set); |
1648 | /* Generally, spills cannot be processed if and of the | |
1649 | references to the slot have a different mode. But if | |
1650 | we are in the same block and mode is exactly the same | |
1651 | between this store and one before in the same block, | |
1652 | we can still delete it. */ | |
1653 | if ((GET_MODE (mem) == GET_MODE (s_info->mem)) | |
1654 | && (GET_MODE (mem) == entry->mode)) | |
1655 | { | |
9ce37fa7 | 1656 | del = true; |
aa140b76 | 1657 | set_all_positions_unneeded (s_info); |
3072d30e | 1658 | } |
1ca59310 | 1659 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1660 | fprintf (dump_file, " trying spill store in insn=%d alias_set=%d\n", |
32c2fdea | 1661 | INSN_UID (ptr->insn), (int) s_info->alias_set); |
3072d30e | 1662 | } |
48e1416a | 1663 | else if ((s_info->group_id == group_id) |
3072d30e | 1664 | && (s_info->cse_base == base)) |
1665 | { | |
1666 | HOST_WIDE_INT i; | |
1ca59310 | 1667 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1668 | fprintf (dump_file, " trying store in insn=%d gid=%d[%d..%d)\n", |
48e1416a | 1669 | INSN_UID (ptr->insn), s_info->group_id, |
3072d30e | 1670 | (int)s_info->begin, (int)s_info->end); |
aa140b76 | 1671 | |
1672 | /* Even if PTR won't be eliminated as unneeded, if both | |
1673 | PTR and this insn store the same constant value, we might | |
1674 | eliminate this insn instead. */ | |
1675 | if (s_info->const_rhs | |
1676 | && const_rhs | |
1677 | && offset >= s_info->begin | |
1678 | && offset + width <= s_info->end | |
1679 | && all_positions_needed_p (s_info, offset - s_info->begin, | |
1680 | width)) | |
1681 | { | |
1682 | if (GET_MODE (mem) == BLKmode) | |
1683 | { | |
1684 | if (GET_MODE (s_info->mem) == BLKmode | |
1685 | && s_info->const_rhs == const_rhs) | |
1686 | redundant_reason = ptr; | |
1687 | } | |
1688 | else if (s_info->const_rhs == const0_rtx | |
1689 | && const_rhs == const0_rtx) | |
1690 | redundant_reason = ptr; | |
1691 | else | |
1692 | { | |
1693 | rtx val; | |
1694 | start_sequence (); | |
1695 | val = get_stored_val (s_info, GET_MODE (mem), | |
1696 | offset, offset + width, | |
1697 | BLOCK_FOR_INSN (insn_info->insn), | |
1698 | true); | |
1699 | if (get_insns () != NULL) | |
1700 | val = NULL_RTX; | |
1701 | end_sequence (); | |
1702 | if (val && rtx_equal_p (val, const_rhs)) | |
1703 | redundant_reason = ptr; | |
1704 | } | |
1705 | } | |
1706 | ||
1707 | for (i = MAX (offset, s_info->begin); | |
1708 | i < offset + width && i < s_info->end; | |
1709 | i++) | |
1710 | set_position_unneeded (s_info, i - s_info->begin); | |
3072d30e | 1711 | } |
1712 | else if (s_info->rhs) | |
1713 | /* Need to see if it is possible for this store to overwrite | |
1714 | the value of store_info. If it is, set the rhs to NULL to | |
1715 | keep it from being used to remove a load. */ | |
1716 | { | |
48e1416a | 1717 | if (canon_true_dependence (s_info->mem, |
3072d30e | 1718 | GET_MODE (s_info->mem), |
1719 | s_info->mem_addr, | |
376a287d | 1720 | mem, mem_addr)) |
aa140b76 | 1721 | { |
1722 | s_info->rhs = NULL; | |
1723 | s_info->const_rhs = NULL; | |
1724 | } | |
3072d30e | 1725 | } |
82d2c88b | 1726 | |
3072d30e | 1727 | /* An insn can be deleted if every position of every one of |
1728 | its s_infos is zero. */ | |
c75be2fe | 1729 | if (any_positions_needed_p (s_info)) |
9ce37fa7 | 1730 | del = false; |
aa140b76 | 1731 | |
9ce37fa7 | 1732 | if (del) |
3072d30e | 1733 | { |
1734 | insn_info_t insn_to_delete = ptr; | |
48e1416a | 1735 | |
1242bee6 | 1736 | active_local_stores_len--; |
3072d30e | 1737 | if (last) |
1738 | last->next_local_store = ptr->next_local_store; | |
1739 | else | |
1740 | active_local_stores = ptr->next_local_store; | |
48e1416a | 1741 | |
c75be2fe | 1742 | if (!insn_to_delete->cannot_delete) |
1743 | delete_dead_store_insn (insn_to_delete); | |
3072d30e | 1744 | } |
1745 | else | |
1746 | last = ptr; | |
48e1416a | 1747 | |
3072d30e | 1748 | ptr = next; |
1749 | } | |
48e1416a | 1750 | |
3072d30e | 1751 | /* Finish filling in the store_info. */ |
1752 | store_info->next = insn_info->store_rec; | |
1753 | insn_info->store_rec = store_info; | |
82d2c88b | 1754 | store_info->mem = mem; |
3072d30e | 1755 | store_info->alias_set = spill_alias_set; |
82d2c88b | 1756 | store_info->mem_addr = mem_addr; |
3072d30e | 1757 | store_info->cse_base = base; |
aa140b76 | 1758 | if (width > HOST_BITS_PER_WIDE_INT) |
1759 | { | |
1760 | store_info->is_large = true; | |
1761 | store_info->positions_needed.large.count = 0; | |
4fb07d00 | 1762 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
aa140b76 | 1763 | } |
1764 | else | |
1765 | { | |
1766 | store_info->is_large = false; | |
1767 | store_info->positions_needed.small_bitmask = lowpart_bitmask (width); | |
1768 | } | |
3072d30e | 1769 | store_info->group_id = group_id; |
1770 | store_info->begin = offset; | |
1771 | store_info->end = offset + width; | |
1772 | store_info->is_set = GET_CODE (body) == SET; | |
aa140b76 | 1773 | store_info->rhs = rhs; |
1774 | store_info->const_rhs = const_rhs; | |
1775 | store_info->redundant_reason = redundant_reason; | |
3072d30e | 1776 | |
3072d30e | 1777 | /* If this is a clobber, we return 0. We will only be able to |
1778 | delete this insn if there is only one store USED store, but we | |
1779 | can use the clobber to delete other stores earlier. */ | |
1780 | return store_info->is_set ? 1 : 0; | |
1781 | } | |
1782 | ||
1783 | ||
1784 | static void | |
1785 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1786 | { | |
48e1416a | 1787 | fprintf (dump_file, "%s insn=%d %s\n", start, |
3072d30e | 1788 | INSN_UID (insn_info->insn), |
1789 | insn_info->store_rec ? "has store" : "naked"); | |
1790 | } | |
1791 | ||
1792 | ||
5c9051a4 | 1793 | /* If the modes are different and the value's source and target do not |
1794 | line up, we need to extract the value from lower part of the rhs of | |
1795 | the store, shift it, and then put it into a form that can be shoved | |
1796 | into the read_insn. This function generates a right SHIFT of a | |
1797 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1798 | shift sequence is returned or NULL if we failed to find a | |
1799 | shift. */ | |
1800 | ||
1801 | static rtx | |
10d4de0e | 1802 | find_shift_sequence (int access_size, |
5c9051a4 | 1803 | store_info_t store_info, |
3754d046 | 1804 | machine_mode read_mode, |
aa140b76 | 1805 | int shift, bool speed, bool require_cst) |
5c9051a4 | 1806 | { |
3754d046 | 1807 | machine_mode store_mode = GET_MODE (store_info->mem); |
1808 | machine_mode new_mode; | |
10d4de0e | 1809 | rtx read_reg = NULL; |
5c9051a4 | 1810 | |
1811 | /* Some machines like the x86 have shift insns for each size of | |
1812 | operand. Other machines like the ppc or the ia-64 may only have | |
1813 | shift insns that shift values within 32 or 64 bit registers. | |
1814 | This loop tries to find the smallest shift insn that will right | |
1815 | justify the value we want to read but is available in one insn on | |
1816 | the machine. */ | |
1817 | ||
10d4de0e | 1818 | for (new_mode = smallest_mode_for_size (access_size * BITS_PER_UNIT, |
1819 | MODE_INT); | |
1820 | GET_MODE_BITSIZE (new_mode) <= BITS_PER_WORD; | |
1821 | new_mode = GET_MODE_WIDER_MODE (new_mode)) | |
5c9051a4 | 1822 | { |
4cd001d5 | 1823 | rtx target, new_reg, new_lhs; |
1824 | rtx_insn *shift_seq, *insn; | |
171557e8 | 1825 | int cost; |
af97461e | 1826 | |
4ed4afb9 | 1827 | /* If a constant was stored into memory, try to simplify it here, |
1828 | otherwise the cost of the shift might preclude this optimization | |
1829 | e.g. at -Os, even when no actual shift will be needed. */ | |
aa140b76 | 1830 | if (store_info->const_rhs) |
4ed4afb9 | 1831 | { |
1832 | unsigned int byte = subreg_lowpart_offset (new_mode, store_mode); | |
aa140b76 | 1833 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1834 | store_mode, byte); | |
4ed4afb9 | 1835 | if (ret && CONSTANT_P (ret)) |
1836 | { | |
1837 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, | |
1838 | ret, GEN_INT (shift)); | |
1839 | if (ret && CONSTANT_P (ret)) | |
1840 | { | |
1841 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1842 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1843 | if (ret && CONSTANT_P (ret) | |
7013e87c | 1844 | && set_src_cost (ret, speed) <= COSTS_N_INSNS (1)) |
4ed4afb9 | 1845 | return ret; |
1846 | } | |
1847 | } | |
1848 | } | |
1849 | ||
aa140b76 | 1850 | if (require_cst) |
1851 | return NULL_RTX; | |
1852 | ||
10d4de0e | 1853 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1854 | requires a real instruction. */ | |
1855 | if (GET_MODE_BITSIZE (new_mode) < GET_MODE_BITSIZE (store_mode) | |
396f2130 | 1856 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
af97461e | 1857 | continue; |
1858 | ||
10d4de0e | 1859 | /* Also try a wider mode if the necessary punning is either not |
1860 | desirable or not possible. */ | |
1861 | if (!CONSTANT_P (store_info->rhs) | |
1862 | && !MODES_TIEABLE_P (new_mode, store_mode)) | |
1863 | continue; | |
10d4de0e | 1864 | |
af97461e | 1865 | new_reg = gen_reg_rtx (new_mode); |
5c9051a4 | 1866 | |
1867 | start_sequence (); | |
1868 | ||
1869 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1870 | of one dsp where the cost of these two was not the same. But | |
1871 | this really is a rare case anyway. */ | |
1872 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
1873 | GEN_INT (shift), new_reg, 1, OPTAB_DIRECT); | |
1874 | ||
597d5470 | 1875 | shift_seq = get_insns (); |
1876 | end_sequence (); | |
5c9051a4 | 1877 | |
597d5470 | 1878 | if (target != new_reg || shift_seq == NULL) |
1879 | continue; | |
1880 | ||
1881 | cost = 0; | |
1882 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1883 | if (INSN_P (insn)) | |
f529eb25 | 1884 | cost += insn_rtx_cost (PATTERN (insn), speed); |
597d5470 | 1885 | |
1886 | /* The computation up to here is essentially independent | |
1887 | of the arguments and could be precomputed. It may | |
1888 | not be worth doing so. We could precompute if | |
1889 | worthwhile or at least cache the results. The result | |
5133fd21 | 1890 | technically depends on both SHIFT and ACCESS_SIZE, |
1891 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
597d5470 | 1892 | |
1893 | if (cost > COSTS_N_INSNS (1)) | |
1894 | continue; | |
1895 | ||
171557e8 | 1896 | new_lhs = extract_low_bits (new_mode, store_mode, |
1897 | copy_rtx (store_info->rhs)); | |
1898 | if (new_lhs == NULL_RTX) | |
1899 | continue; | |
1900 | ||
597d5470 | 1901 | /* We found an acceptable shift. Generate a move to |
1902 | take the value from the store and put it into the | |
1903 | shift pseudo, then shift it, then generate another | |
1904 | move to put in into the target of the read. */ | |
10d4de0e | 1905 | emit_move_insn (new_reg, new_lhs); |
597d5470 | 1906 | emit_insn (shift_seq); |
10d4de0e | 1907 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
597d5470 | 1908 | break; |
5c9051a4 | 1909 | } |
1910 | ||
10d4de0e | 1911 | return read_reg; |
5c9051a4 | 1912 | } |
1913 | ||
1914 | ||
a1b0a968 | 1915 | /* Call back for note_stores to find the hard regs set or clobbered by |
1916 | insn. Data is a bitmap of the hardregs set so far. */ | |
1917 | ||
1918 | static void | |
1919 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1920 | { | |
1921 | bitmap regs_set = (bitmap) data; | |
1922 | ||
1923 | if (REG_P (x) | |
771d4616 | 1924 | && HARD_REGISTER_P (x)) |
0933f1d9 | 1925 | bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); |
a1b0a968 | 1926 | } |
1927 | ||
aa140b76 | 1928 | /* Helper function for replace_read and record_store. |
1929 | Attempt to return a value stored in STORE_INFO, from READ_BEGIN | |
1930 | to one before READ_END bytes read in READ_MODE. Return NULL | |
1931 | if not successful. If REQUIRE_CST is true, return always constant. */ | |
1932 | ||
1933 | static rtx | |
3754d046 | 1934 | get_stored_val (store_info_t store_info, machine_mode read_mode, |
aa140b76 | 1935 | HOST_WIDE_INT read_begin, HOST_WIDE_INT read_end, |
1936 | basic_block bb, bool require_cst) | |
1937 | { | |
3754d046 | 1938 | machine_mode store_mode = GET_MODE (store_info->mem); |
aa140b76 | 1939 | int shift; |
1940 | int access_size; /* In bytes. */ | |
1941 | rtx read_reg; | |
1942 | ||
1943 | /* To get here the read is within the boundaries of the write so | |
1944 | shift will never be negative. Start out with the shift being in | |
1945 | bytes. */ | |
1946 | if (store_mode == BLKmode) | |
1947 | shift = 0; | |
1948 | else if (BYTES_BIG_ENDIAN) | |
1949 | shift = store_info->end - read_end; | |
1950 | else | |
1951 | shift = read_begin - store_info->begin; | |
1952 | ||
1953 | access_size = shift + GET_MODE_SIZE (read_mode); | |
1954 | ||
1955 | /* From now on it is bits. */ | |
1956 | shift *= BITS_PER_UNIT; | |
1957 | ||
1958 | if (shift) | |
1959 | read_reg = find_shift_sequence (access_size, store_info, read_mode, shift, | |
1960 | optimize_bb_for_speed_p (bb), | |
1961 | require_cst); | |
1962 | else if (store_mode == BLKmode) | |
1963 | { | |
1964 | /* The store is a memset (addr, const_val, const_size). */ | |
1965 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
1966 | store_mode = int_mode_for_mode (read_mode); | |
1967 | if (store_mode == BLKmode) | |
1968 | read_reg = NULL_RTX; | |
1969 | else if (store_info->rhs == const0_rtx) | |
1970 | read_reg = extract_low_bits (read_mode, store_mode, const0_rtx); | |
1971 | else if (GET_MODE_BITSIZE (store_mode) > HOST_BITS_PER_WIDE_INT | |
1972 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) | |
1973 | read_reg = NULL_RTX; | |
1974 | else | |
1975 | { | |
1976 | unsigned HOST_WIDE_INT c | |
1977 | = INTVAL (store_info->rhs) | |
1978 | & (((HOST_WIDE_INT) 1 << BITS_PER_UNIT) - 1); | |
1979 | int shift = BITS_PER_UNIT; | |
1980 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1981 | { | |
1982 | c |= (c << shift); | |
1983 | shift <<= 1; | |
1984 | } | |
f62058c3 | 1985 | read_reg = gen_int_mode (c, store_mode); |
aa140b76 | 1986 | read_reg = extract_low_bits (read_mode, store_mode, read_reg); |
1987 | } | |
1988 | } | |
1989 | else if (store_info->const_rhs | |
1990 | && (require_cst | |
1991 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1992 | read_reg = extract_low_bits (read_mode, store_mode, | |
1993 | copy_rtx (store_info->const_rhs)); | |
1994 | else | |
1995 | read_reg = extract_low_bits (read_mode, store_mode, | |
1996 | copy_rtx (store_info->rhs)); | |
1997 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1998 | read_reg = NULL_RTX; | |
1999 | return read_reg; | |
2000 | } | |
a1b0a968 | 2001 | |
3072d30e | 2002 | /* Take a sequence of: |
2003 | A <- r1 | |
2004 | ... | |
2005 | ... <- A | |
2006 | ||
48e1416a | 2007 | and change it into |
3072d30e | 2008 | r2 <- r1 |
2009 | A <- r1 | |
2010 | ... | |
2011 | ... <- r2 | |
2012 | ||
5c9051a4 | 2013 | or |
2014 | ||
2015 | r3 <- extract (r1) | |
2016 | r3 <- r3 >> shift | |
2017 | r2 <- extract (r3) | |
2018 | ... <- r2 | |
2019 | ||
2020 | or | |
2021 | ||
2022 | r2 <- extract (r1) | |
2023 | ... <- r2 | |
2024 | ||
2025 | Depending on the alignment and the mode of the store and | |
2026 | subsequent load. | |
2027 | ||
2028 | ||
2029 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
3072d30e | 2030 | and READ_INSN are for the read. Return true if the replacement |
2031 | went ok. */ | |
2032 | ||
2033 | static bool | |
48e1416a | 2034 | replace_read (store_info_t store_info, insn_info_t store_insn, |
aa140b76 | 2035 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
2036 | bitmap regs_live) | |
3072d30e | 2037 | { |
3754d046 | 2038 | machine_mode store_mode = GET_MODE (store_info->mem); |
2039 | machine_mode read_mode = GET_MODE (read_info->mem); | |
4cd001d5 | 2040 | rtx_insn *insns, *this_insn; |
2041 | rtx read_reg; | |
aa140b76 | 2042 | basic_block bb; |
5c9051a4 | 2043 | |
3072d30e | 2044 | if (!dbg_cnt (dse)) |
2045 | return false; | |
2046 | ||
10d4de0e | 2047 | /* Create a sequence of instructions to set up the read register. |
2048 | This sequence goes immediately before the store and its result | |
2049 | is read by the load. | |
2050 | ||
2051 | We need to keep this in perspective. We are replacing a read | |
5c9051a4 | 2052 | with a sequence of insns, but the read will almost certainly be |
2053 | in cache, so it is not going to be an expensive one. Thus, we | |
2054 | are not willing to do a multi insn shift or worse a subroutine | |
2055 | call to get rid of the read. */ | |
1ca59310 | 2056 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 2057 | fprintf (dump_file, "trying to replace %smode load in insn %d" |
2058 | " from %smode store in insn %d\n", | |
2059 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
2060 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
2061 | start_sequence (); | |
aa140b76 | 2062 | bb = BLOCK_FOR_INSN (read_insn->insn); |
2063 | read_reg = get_stored_val (store_info, | |
2064 | read_mode, read_info->begin, read_info->end, | |
2065 | bb, false); | |
10d4de0e | 2066 | if (read_reg == NULL_RTX) |
5c9051a4 | 2067 | { |
10d4de0e | 2068 | end_sequence (); |
1ca59310 | 2069 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 2070 | fprintf (dump_file, " -- could not extract bits of stored value\n"); |
2071 | return false; | |
5c9051a4 | 2072 | } |
10d4de0e | 2073 | /* Force the value into a new register so that it won't be clobbered |
2074 | between the store and the load. */ | |
2075 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
2076 | insns = get_insns (); | |
2077 | end_sequence (); | |
5c9051a4 | 2078 | |
a1b0a968 | 2079 | if (insns != NULL_RTX) |
2080 | { | |
2081 | /* Now we have to scan the set of new instructions to see if the | |
2082 | sequence contains and sets of hardregs that happened to be | |
2083 | live at this point. For instance, this can happen if one of | |
2084 | the insns sets the CC and the CC happened to be live at that | |
2085 | point. This does occasionally happen, see PR 37922. */ | |
4fb07d00 | 2086 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
a1b0a968 | 2087 | |
2088 | for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
2089 | note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
48e1416a | 2090 | |
a1b0a968 | 2091 | bitmap_and_into (regs_set, regs_live); |
2092 | if (!bitmap_empty_p (regs_set)) | |
2093 | { | |
1ca59310 | 2094 | if (dump_file && (dump_flags & TDF_DETAILS)) |
a1b0a968 | 2095 | { |
48e1416a | 2096 | fprintf (dump_file, |
a1b0a968 | 2097 | "abandoning replacement because sequence clobbers live hardregs:"); |
2098 | df_print_regset (dump_file, regs_set); | |
2099 | } | |
48e1416a | 2100 | |
a1b0a968 | 2101 | BITMAP_FREE (regs_set); |
2102 | return false; | |
2103 | } | |
2104 | BITMAP_FREE (regs_set); | |
2105 | } | |
2106 | ||
5c9051a4 | 2107 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
3072d30e | 2108 | { |
55c5ac9f | 2109 | deferred_change_t change = new deferred_change; |
48e1416a | 2110 | |
5c9051a4 | 2111 | /* Insert this right before the store insn where it will be safe |
2112 | from later insns that might change it before the read. */ | |
2113 | emit_insn_before (insns, store_insn->insn); | |
48e1416a | 2114 | |
5c9051a4 | 2115 | /* And now for the kludge part: cselib croaks if you just |
2116 | return at this point. There are two reasons for this: | |
48e1416a | 2117 | |
5c9051a4 | 2118 | 1) Cselib has an idea of how many pseudos there are and |
2119 | that does not include the new ones we just added. | |
48e1416a | 2120 | |
5c9051a4 | 2121 | 2) Cselib does not know about the move insn we added |
2122 | above the store_info, and there is no way to tell it | |
2123 | about it, because it has "moved on". | |
48e1416a | 2124 | |
5c9051a4 | 2125 | Problem (1) is fixable with a certain amount of engineering. |
2126 | Problem (2) is requires starting the bb from scratch. This | |
2127 | could be expensive. | |
48e1416a | 2128 | |
5c9051a4 | 2129 | So we are just going to have to lie. The move/extraction |
2130 | insns are not really an issue, cselib did not see them. But | |
2131 | the use of the new pseudo read_insn is a real problem because | |
2132 | cselib has not scanned this insn. The way that we solve this | |
2133 | problem is that we are just going to put the mem back for now | |
2134 | and when we are finished with the block, we undo this. We | |
2135 | keep a table of mems to get rid of. At the end of the basic | |
2136 | block we can put them back. */ | |
48e1416a | 2137 | |
5c9051a4 | 2138 | *loc = read_info->mem; |
55c5ac9f | 2139 | change->next = deferred_change_list; |
2140 | deferred_change_list = change; | |
2141 | change->loc = loc; | |
2142 | change->reg = read_reg; | |
48e1416a | 2143 | |
5c9051a4 | 2144 | /* Get rid of the read_info, from the point of view of the |
2145 | rest of dse, play like this read never happened. */ | |
2146 | read_insn->read_rec = read_info->next; | |
55c5ac9f | 2147 | delete read_info; |
1ca59310 | 2148 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 2149 | { |
2150 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2151 | print_simple_rtl (dump_file, read_reg); | |
2152 | fprintf (dump_file, "\n"); | |
2153 | } | |
5c9051a4 | 2154 | return true; |
3072d30e | 2155 | } |
48e1416a | 2156 | else |
3072d30e | 2157 | { |
1ca59310 | 2158 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 2159 | { |
2160 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2161 | print_simple_rtl (dump_file, read_reg); | |
2162 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2163 | } | |
3072d30e | 2164 | return false; |
2165 | } | |
2166 | } | |
2167 | ||
ec1203cd | 2168 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2169 | be active. */ | |
3072d30e | 2170 | |
ec1203cd | 2171 | static void |
2172 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info) | |
3072d30e | 2173 | { |
82d2c88b | 2174 | rtx mem = *loc, mem_addr; |
3072d30e | 2175 | insn_info_t insn_info; |
2176 | HOST_WIDE_INT offset = 0; | |
2177 | HOST_WIDE_INT width = 0; | |
32c2fdea | 2178 | alias_set_type spill_alias_set = 0; |
48e1416a | 2179 | cselib_val *base = NULL; |
3072d30e | 2180 | int group_id; |
2181 | read_info_t read_info; | |
2182 | ||
3072d30e | 2183 | insn_info = bb_info->last_insn; |
2184 | ||
2185 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2186 | || (MEM_VOLATILE_P (mem))) | |
2187 | { | |
1ca59310 | 2188 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2189 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
2190 | add_wild_read (bb_info); | |
2191 | insn_info->cannot_delete = true; | |
ec1203cd | 2192 | return; |
3072d30e | 2193 | } |
2194 | ||
2195 | /* If it is reading readonly mem, then there can be no conflict with | |
2196 | another write. */ | |
2197 | if (MEM_READONLY_P (mem)) | |
ec1203cd | 2198 | return; |
3072d30e | 2199 | |
2200 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
2201 | { | |
1ca59310 | 2202 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2203 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
2204 | add_wild_read (bb_info); | |
ec1203cd | 2205 | return; |
3072d30e | 2206 | } |
2207 | ||
2208 | if (GET_MODE (mem) == BLKmode) | |
2209 | width = -1; | |
2210 | else | |
2211 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2212 | ||
55c5ac9f | 2213 | read_info = new read_info_type; |
3072d30e | 2214 | read_info->group_id = group_id; |
2215 | read_info->mem = mem; | |
2216 | read_info->alias_set = spill_alias_set; | |
2217 | read_info->begin = offset; | |
2218 | read_info->end = offset + width; | |
2219 | read_info->next = insn_info->read_rec; | |
2220 | insn_info->read_rec = read_info; | |
82d2c88b | 2221 | /* For alias_set != 0 canon_true_dependence should be never called. */ |
2222 | if (spill_alias_set) | |
2223 | mem_addr = NULL_RTX; | |
2224 | else | |
2225 | { | |
2226 | if (group_id < 0) | |
2227 | mem_addr = base->val_rtx; | |
2228 | else | |
2229 | { | |
2230 | group_info_t group | |
f1f41a6c | 2231 | = rtx_group_vec[group_id]; |
82d2c88b | 2232 | mem_addr = group->canon_base_addr; |
2233 | } | |
90f3e775 | 2234 | /* get_addr can only handle VALUE but cannot handle expr like: |
2235 | VALUE + OFFSET, so call get_addr to get original addr for | |
2236 | mem_addr before plus_constant. */ | |
2237 | mem_addr = get_addr (mem_addr); | |
82d2c88b | 2238 | if (offset) |
87cf5753 | 2239 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
82d2c88b | 2240 | } |
3072d30e | 2241 | |
bef304b8 | 2242 | /* We ignore the clobbers in store_info. The is mildly aggressive, |
3072d30e | 2243 | but there really should not be a clobber followed by a read. */ |
2244 | ||
2245 | if (spill_alias_set) | |
2246 | { | |
2247 | insn_info_t i_ptr = active_local_stores; | |
2248 | insn_info_t last = NULL; | |
2249 | ||
1ca59310 | 2250 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2251 | fprintf (dump_file, " processing spill load %d\n", |
32c2fdea | 2252 | (int) spill_alias_set); |
3072d30e | 2253 | |
2254 | while (i_ptr) | |
2255 | { | |
2256 | store_info_t store_info = i_ptr->store_rec; | |
2257 | ||
2258 | /* Skip the clobbers. */ | |
2259 | while (!store_info->is_set) | |
2260 | store_info = store_info->next; | |
48e1416a | 2261 | |
3072d30e | 2262 | if (store_info->alias_set == spill_alias_set) |
2263 | { | |
1ca59310 | 2264 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2265 | dump_insn_info ("removing from active", i_ptr); |
2266 | ||
1242bee6 | 2267 | active_local_stores_len--; |
3072d30e | 2268 | if (last) |
2269 | last->next_local_store = i_ptr->next_local_store; | |
2270 | else | |
2271 | active_local_stores = i_ptr->next_local_store; | |
2272 | } | |
2273 | else | |
2274 | last = i_ptr; | |
2275 | i_ptr = i_ptr->next_local_store; | |
2276 | } | |
2277 | } | |
2278 | else if (group_id >= 0) | |
2279 | { | |
2280 | /* This is the restricted case where the base is a constant or | |
2281 | the frame pointer and offset is a constant. */ | |
2282 | insn_info_t i_ptr = active_local_stores; | |
2283 | insn_info_t last = NULL; | |
48e1416a | 2284 | |
1ca59310 | 2285 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2286 | { |
2287 | if (width == -1) | |
2288 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", | |
2289 | group_id); | |
2290 | else | |
2291 | fprintf (dump_file, " processing const load gid=%d[%d..%d)\n", | |
2292 | group_id, (int)offset, (int)(offset+width)); | |
2293 | } | |
2294 | ||
2295 | while (i_ptr) | |
2296 | { | |
2297 | bool remove = false; | |
2298 | store_info_t store_info = i_ptr->store_rec; | |
48e1416a | 2299 | |
3072d30e | 2300 | /* Skip the clobbers. */ |
2301 | while (!store_info->is_set) | |
2302 | store_info = store_info->next; | |
48e1416a | 2303 | |
3072d30e | 2304 | /* There are three cases here. */ |
2305 | if (store_info->group_id < 0) | |
2306 | /* We have a cselib store followed by a read from a | |
2307 | const base. */ | |
48e1416a | 2308 | remove |
2309 | = canon_true_dependence (store_info->mem, | |
3072d30e | 2310 | GET_MODE (store_info->mem), |
2311 | store_info->mem_addr, | |
376a287d | 2312 | mem, mem_addr); |
48e1416a | 2313 | |
3072d30e | 2314 | else if (group_id == store_info->group_id) |
2315 | { | |
2316 | /* This is a block mode load. We may get lucky and | |
2317 | canon_true_dependence may save the day. */ | |
2318 | if (width == -1) | |
48e1416a | 2319 | remove |
2320 | = canon_true_dependence (store_info->mem, | |
3072d30e | 2321 | GET_MODE (store_info->mem), |
2322 | store_info->mem_addr, | |
376a287d | 2323 | mem, mem_addr); |
48e1416a | 2324 | |
3072d30e | 2325 | /* If this read is just reading back something that we just |
2326 | stored, rewrite the read. */ | |
48e1416a | 2327 | else |
3072d30e | 2328 | { |
2329 | if (store_info->rhs | |
aa140b76 | 2330 | && offset >= store_info->begin |
2331 | && offset + width <= store_info->end | |
2332 | && all_positions_needed_p (store_info, | |
2333 | offset - store_info->begin, | |
2334 | width) | |
2335 | && replace_read (store_info, i_ptr, read_info, | |
2336 | insn_info, loc, bb_info->regs_live)) | |
ec1203cd | 2337 | return; |
aa140b76 | 2338 | |
3072d30e | 2339 | /* The bases are the same, just see if the offsets |
2340 | overlap. */ | |
48e1416a | 2341 | if ((offset < store_info->end) |
3072d30e | 2342 | && (offset + width > store_info->begin)) |
2343 | remove = true; | |
2344 | } | |
2345 | } | |
48e1416a | 2346 | |
2347 | /* else | |
3072d30e | 2348 | The else case that is missing here is that the |
2349 | bases are constant but different. There is nothing | |
2350 | to do here because there is no overlap. */ | |
48e1416a | 2351 | |
3072d30e | 2352 | if (remove) |
2353 | { | |
1ca59310 | 2354 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2355 | dump_insn_info ("removing from active", i_ptr); |
2356 | ||
1242bee6 | 2357 | active_local_stores_len--; |
3072d30e | 2358 | if (last) |
2359 | last->next_local_store = i_ptr->next_local_store; | |
2360 | else | |
2361 | active_local_stores = i_ptr->next_local_store; | |
2362 | } | |
2363 | else | |
2364 | last = i_ptr; | |
2365 | i_ptr = i_ptr->next_local_store; | |
2366 | } | |
2367 | } | |
48e1416a | 2368 | else |
3072d30e | 2369 | { |
2370 | insn_info_t i_ptr = active_local_stores; | |
2371 | insn_info_t last = NULL; | |
1ca59310 | 2372 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2373 | { |
2374 | fprintf (dump_file, " processing cselib load mem:"); | |
2375 | print_inline_rtx (dump_file, mem, 0); | |
2376 | fprintf (dump_file, "\n"); | |
2377 | } | |
2378 | ||
2379 | while (i_ptr) | |
2380 | { | |
2381 | bool remove = false; | |
2382 | store_info_t store_info = i_ptr->store_rec; | |
48e1416a | 2383 | |
1ca59310 | 2384 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2385 | fprintf (dump_file, " processing cselib load against insn %d\n", |
2386 | INSN_UID (i_ptr->insn)); | |
2387 | ||
2388 | /* Skip the clobbers. */ | |
2389 | while (!store_info->is_set) | |
2390 | store_info = store_info->next; | |
2391 | ||
2392 | /* If this read is just reading back something that we just | |
2393 | stored, rewrite the read. */ | |
2394 | if (store_info->rhs | |
2395 | && store_info->group_id == -1 | |
2396 | && store_info->cse_base == base | |
2ffe5515 | 2397 | && width != -1 |
aa140b76 | 2398 | && offset >= store_info->begin |
2399 | && offset + width <= store_info->end | |
2400 | && all_positions_needed_p (store_info, | |
2401 | offset - store_info->begin, width) | |
2402 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, | |
2403 | bb_info->regs_live)) | |
ec1203cd | 2404 | return; |
3072d30e | 2405 | |
2406 | if (!store_info->alias_set) | |
48e1416a | 2407 | remove = canon_true_dependence (store_info->mem, |
3072d30e | 2408 | GET_MODE (store_info->mem), |
2409 | store_info->mem_addr, | |
376a287d | 2410 | mem, mem_addr); |
48e1416a | 2411 | |
3072d30e | 2412 | if (remove) |
2413 | { | |
1ca59310 | 2414 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2415 | dump_insn_info ("removing from active", i_ptr); |
48e1416a | 2416 | |
1242bee6 | 2417 | active_local_stores_len--; |
3072d30e | 2418 | if (last) |
2419 | last->next_local_store = i_ptr->next_local_store; | |
2420 | else | |
2421 | active_local_stores = i_ptr->next_local_store; | |
2422 | } | |
2423 | else | |
2424 | last = i_ptr; | |
2425 | i_ptr = i_ptr->next_local_store; | |
2426 | } | |
2427 | } | |
3072d30e | 2428 | } |
2429 | ||
ec1203cd | 2430 | /* A note_uses callback in which DATA points the INSN_INFO for |
3072d30e | 2431 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2432 | true for any part of *LOC. */ | |
2433 | ||
2434 | static void | |
2435 | check_mem_read_use (rtx *loc, void *data) | |
2436 | { | |
ec1203cd | 2437 | subrtx_ptr_iterator::array_type array; |
2438 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
2439 | { | |
2440 | rtx *loc = *iter; | |
2441 | if (MEM_P (*loc)) | |
2442 | check_mem_read_rtx (loc, (bb_info_t) data); | |
2443 | } | |
3072d30e | 2444 | } |
2445 | ||
aa140b76 | 2446 | |
2447 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2448 | So far it only handles arguments passed in registers. */ | |
2449 | ||
2450 | static bool | |
2451 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2452 | { | |
39cba157 | 2453 | CUMULATIVE_ARGS args_so_far_v; |
2454 | cumulative_args_t args_so_far; | |
aa140b76 | 2455 | tree arg; |
2456 | int idx; | |
2457 | ||
39cba157 | 2458 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2459 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
aa140b76 | 2460 | |
2461 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2462 | for (idx = 0; | |
2463 | arg != void_list_node && idx < nargs; | |
2464 | arg = TREE_CHAIN (arg), idx++) | |
2465 | { | |
3754d046 | 2466 | machine_mode mode = TYPE_MODE (TREE_VALUE (arg)); |
f387af4f | 2467 | rtx reg, link, tmp; |
39cba157 | 2468 | reg = targetm.calls.function_arg (args_so_far, mode, NULL_TREE, true); |
aa140b76 | 2469 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode |
2470 | || GET_MODE_CLASS (mode) != MODE_INT) | |
2471 | return false; | |
2472 | ||
2473 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2474 | link; | |
2475 | link = XEXP (link, 1)) | |
2476 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2477 | { | |
2478 | args[idx] = XEXP (XEXP (link, 0), 0); | |
2479 | if (REG_P (args[idx]) | |
2480 | && REGNO (args[idx]) == REGNO (reg) | |
2481 | && (GET_MODE (args[idx]) == mode | |
2482 | || (GET_MODE_CLASS (GET_MODE (args[idx])) == MODE_INT | |
2483 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2484 | <= UNITS_PER_WORD) | |
2485 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2486 | > GET_MODE_SIZE (mode))))) | |
2487 | break; | |
2488 | } | |
2489 | if (!link) | |
2490 | return false; | |
2491 | ||
2492 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2493 | if (GET_MODE (args[idx]) != mode) | |
2494 | { | |
2495 | if (!tmp || !CONST_INT_P (tmp)) | |
2496 | return false; | |
f62058c3 | 2497 | tmp = gen_int_mode (INTVAL (tmp), mode); |
aa140b76 | 2498 | } |
2499 | if (tmp) | |
2500 | args[idx] = tmp; | |
2501 | ||
39cba157 | 2502 | targetm.calls.function_arg_advance (args_so_far, mode, NULL_TREE, true); |
aa140b76 | 2503 | } |
2504 | if (arg != void_list_node || idx != nargs) | |
2505 | return false; | |
2506 | return true; | |
2507 | } | |
2508 | ||
5a9ecd4a | 2509 | /* Return a bitmap of the fixed registers contained in IN. */ |
2510 | ||
2511 | static bitmap | |
2512 | copy_fixed_regs (const_bitmap in) | |
2513 | { | |
2514 | bitmap ret; | |
2515 | ||
2516 | ret = ALLOC_REG_SET (NULL); | |
2517 | bitmap_and (ret, in, fixed_reg_set_regset); | |
2518 | return ret; | |
2519 | } | |
aa140b76 | 2520 | |
3072d30e | 2521 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2522 | if some part of it is not a candidate store and assigns to a | |
2523 | non-register target. */ | |
2524 | ||
2525 | static void | |
ebabb7a3 | 2526 | scan_insn (bb_info_t bb_info, rtx_insn *insn) |
3072d30e | 2527 | { |
2528 | rtx body; | |
55c5ac9f | 2529 | insn_info_type *insn_info = new insn_info_type; |
3072d30e | 2530 | int mems_found = 0; |
55c5ac9f | 2531 | memset (insn_info, 0, sizeof (struct insn_info_type)); |
3072d30e | 2532 | |
1ca59310 | 2533 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2534 | fprintf (dump_file, "\n**scanning insn=%d\n", |
2535 | INSN_UID (insn)); | |
2536 | ||
2537 | insn_info->prev_insn = bb_info->last_insn; | |
2538 | insn_info->insn = insn; | |
2539 | bb_info->last_insn = insn_info; | |
48e1416a | 2540 | |
9845d120 | 2541 | if (DEBUG_INSN_P (insn)) |
2542 | { | |
2543 | insn_info->cannot_delete = true; | |
2544 | return; | |
2545 | } | |
3072d30e | 2546 | |
3072d30e | 2547 | /* Look at all of the uses in the insn. */ |
2548 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2549 | ||
2550 | if (CALL_P (insn)) | |
2551 | { | |
aa140b76 | 2552 | bool const_call; |
2553 | tree memset_call = NULL_TREE; | |
2554 | ||
3072d30e | 2555 | insn_info->cannot_delete = true; |
17e1318c | 2556 | |
3072d30e | 2557 | /* Const functions cannot do anything bad i.e. read memory, |
17e1318c | 2558 | however, they can read their parameters which may have |
aa140b76 | 2559 | been pushed onto the stack. |
2560 | memset and bzero don't read memory either. */ | |
2561 | const_call = RTL_CONST_CALL_P (insn); | |
2562 | if (!const_call) | |
2563 | { | |
cf7fb72d | 2564 | rtx call = get_call_rtx_from (insn); |
2565 | if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF) | |
aa140b76 | 2566 | { |
2567 | rtx symbol = XEXP (XEXP (call, 0), 0); | |
2568 | if (SYMBOL_REF_DECL (symbol) | |
2569 | && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL) | |
2570 | { | |
2571 | if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol)) | |
2572 | == BUILT_IN_NORMAL | |
2573 | && (DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol)) | |
2574 | == BUILT_IN_MEMSET)) | |
2575 | || SYMBOL_REF_DECL (symbol) == block_clear_fn) | |
2576 | memset_call = SYMBOL_REF_DECL (symbol); | |
2577 | } | |
2578 | } | |
2579 | } | |
2580 | if (const_call || memset_call) | |
3072d30e | 2581 | { |
2582 | insn_info_t i_ptr = active_local_stores; | |
2583 | insn_info_t last = NULL; | |
2584 | ||
1ca59310 | 2585 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 2586 | fprintf (dump_file, "%s call %d\n", |
2587 | const_call ? "const" : "memset", INSN_UID (insn)); | |
3072d30e | 2588 | |
16bf64db | 2589 | /* See the head comment of the frame_read field. */ |
17853422 | 2590 | if (reload_completed |
2591 | /* Tail calls are storing their arguments using | |
2592 | arg pointer. If it is a frame pointer on the target, | |
2593 | even before reload we need to kill frame pointer based | |
2594 | stores. */ | |
2595 | || (SIBLING_CALL_P (insn) | |
2596 | && HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
16bf64db | 2597 | insn_info->frame_read = true; |
2598 | ||
2599 | /* Loop over the active stores and remove those which are | |
2600 | killed by the const function call. */ | |
3072d30e | 2601 | while (i_ptr) |
2602 | { | |
16bf64db | 2603 | bool remove_store = false; |
2604 | ||
2605 | /* The stack pointer based stores are always killed. */ | |
17e1318c | 2606 | if (i_ptr->stack_pointer_based) |
16bf64db | 2607 | remove_store = true; |
2608 | ||
2609 | /* If the frame is read, the frame related stores are killed. */ | |
2610 | else if (insn_info->frame_read) | |
2611 | { | |
2612 | store_info_t store_info = i_ptr->store_rec; | |
2613 | ||
2614 | /* Skip the clobbers. */ | |
2615 | while (!store_info->is_set) | |
2616 | store_info = store_info->next; | |
2617 | ||
2618 | if (store_info->group_id >= 0 | |
f1f41a6c | 2619 | && rtx_group_vec[store_info->group_id]->frame_related) |
16bf64db | 2620 | remove_store = true; |
2621 | } | |
2622 | ||
2623 | if (remove_store) | |
3072d30e | 2624 | { |
1ca59310 | 2625 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2626 | dump_insn_info ("removing from active", i_ptr); |
48e1416a | 2627 | |
1242bee6 | 2628 | active_local_stores_len--; |
3072d30e | 2629 | if (last) |
2630 | last->next_local_store = i_ptr->next_local_store; | |
2631 | else | |
2632 | active_local_stores = i_ptr->next_local_store; | |
2633 | } | |
2634 | else | |
2635 | last = i_ptr; | |
16bf64db | 2636 | |
3072d30e | 2637 | i_ptr = i_ptr->next_local_store; |
2638 | } | |
aa140b76 | 2639 | |
2640 | if (memset_call) | |
2641 | { | |
2642 | rtx args[3]; | |
2643 | if (get_call_args (insn, memset_call, args, 3) | |
2644 | && CONST_INT_P (args[1]) | |
2645 | && CONST_INT_P (args[2]) | |
2646 | && INTVAL (args[2]) > 0) | |
2647 | { | |
2648 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
5b2a69fa | 2649 | set_mem_size (mem, INTVAL (args[2])); |
d1f9b275 | 2650 | body = gen_rtx_SET (mem, args[1]); |
aa140b76 | 2651 | mems_found += record_store (body, bb_info); |
1ca59310 | 2652 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 2653 | fprintf (dump_file, "handling memset as BLKmode store\n"); |
2654 | if (mems_found == 1) | |
2655 | { | |
1242bee6 | 2656 | if (active_local_stores_len++ |
2657 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2658 | { | |
2659 | active_local_stores_len = 1; | |
2660 | active_local_stores = NULL; | |
2661 | } | |
5a9ecd4a | 2662 | insn_info->fixed_regs_live |
2663 | = copy_fixed_regs (bb_info->regs_live); | |
aa140b76 | 2664 | insn_info->next_local_store = active_local_stores; |
2665 | active_local_stores = insn_info; | |
2666 | } | |
2667 | } | |
2668 | } | |
3072d30e | 2669 | } |
17853422 | 2670 | else if (SIBLING_CALL_P (insn) && reload_completed) |
2671 | /* Arguments for a sibling call that are pushed to memory are passed | |
2672 | using the incoming argument pointer of the current function. After | |
2673 | reload that might be (and likely is) frame pointer based. */ | |
2674 | add_wild_read (bb_info); | |
17e1318c | 2675 | else |
b4a708fb | 2676 | /* Every other call, including pure functions, may read any memory |
2677 | that is not relative to the frame. */ | |
2678 | add_non_frame_wild_read (bb_info); | |
17e1318c | 2679 | |
3072d30e | 2680 | return; |
2681 | } | |
2682 | ||
2683 | /* Assuming that there are sets in these insns, we cannot delete | |
2684 | them. */ | |
2685 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
4aafe72f | 2686 | || volatile_refs_p (PATTERN (insn)) |
bc0dfc8d | 2687 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
3072d30e | 2688 | || (RTX_FRAME_RELATED_P (insn)) |
2689 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2690 | insn_info->cannot_delete = true; | |
48e1416a | 2691 | |
3072d30e | 2692 | body = PATTERN (insn); |
2693 | if (GET_CODE (body) == PARALLEL) | |
2694 | { | |
2695 | int i; | |
2696 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2697 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2698 | } | |
2699 | else | |
2700 | mems_found += record_store (body, bb_info); | |
2701 | ||
1ca59310 | 2702 | if (dump_file && (dump_flags & TDF_DETAILS)) |
48e1416a | 2703 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
3072d30e | 2704 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2705 | ||
aa140b76 | 2706 | /* If we found some sets of mems, add it into the active_local_stores so |
2707 | that it can be locally deleted if found dead or used for | |
2708 | replace_read and redundant constant store elimination. Otherwise mark | |
2709 | it as cannot delete. This simplifies the processing later. */ | |
2710 | if (mems_found == 1) | |
3072d30e | 2711 | { |
1242bee6 | 2712 | if (active_local_stores_len++ |
2713 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2714 | { | |
2715 | active_local_stores_len = 1; | |
2716 | active_local_stores = NULL; | |
2717 | } | |
5a9ecd4a | 2718 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
3072d30e | 2719 | insn_info->next_local_store = active_local_stores; |
2720 | active_local_stores = insn_info; | |
2721 | } | |
2722 | else | |
2723 | insn_info->cannot_delete = true; | |
2724 | } | |
2725 | ||
2726 | ||
2727 | /* Remove BASE from the set of active_local_stores. This is a | |
2728 | callback from cselib that is used to get rid of the stores in | |
2729 | active_local_stores. */ | |
2730 | ||
2731 | static void | |
2732 | remove_useless_values (cselib_val *base) | |
2733 | { | |
2734 | insn_info_t insn_info = active_local_stores; | |
2735 | insn_info_t last = NULL; | |
2736 | ||
2737 | while (insn_info) | |
2738 | { | |
2739 | store_info_t store_info = insn_info->store_rec; | |
9ce37fa7 | 2740 | bool del = false; |
3072d30e | 2741 | |
2742 | /* If ANY of the store_infos match the cselib group that is | |
2743 | being deleted, then the insn can not be deleted. */ | |
2744 | while (store_info) | |
2745 | { | |
48e1416a | 2746 | if ((store_info->group_id == -1) |
3072d30e | 2747 | && (store_info->cse_base == base)) |
2748 | { | |
9ce37fa7 | 2749 | del = true; |
3072d30e | 2750 | break; |
2751 | } | |
2752 | store_info = store_info->next; | |
2753 | } | |
2754 | ||
9ce37fa7 | 2755 | if (del) |
3072d30e | 2756 | { |
1242bee6 | 2757 | active_local_stores_len--; |
3072d30e | 2758 | if (last) |
2759 | last->next_local_store = insn_info->next_local_store; | |
2760 | else | |
2761 | active_local_stores = insn_info->next_local_store; | |
2762 | free_store_info (insn_info); | |
2763 | } | |
2764 | else | |
2765 | last = insn_info; | |
48e1416a | 2766 | |
3072d30e | 2767 | insn_info = insn_info->next_local_store; |
2768 | } | |
2769 | } | |
2770 | ||
2771 | ||
2772 | /* Do all of step 1. */ | |
2773 | ||
2774 | static void | |
2775 | dse_step1 (void) | |
2776 | { | |
2777 | basic_block bb; | |
4fb07d00 | 2778 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
48e1416a | 2779 | |
35af0188 | 2780 | cselib_init (0); |
3072d30e | 2781 | all_blocks = BITMAP_ALLOC (NULL); |
2782 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2783 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2784 | ||
ed7d889a | 2785 | FOR_ALL_BB_FN (bb, cfun) |
3072d30e | 2786 | { |
2787 | insn_info_t ptr; | |
55c5ac9f | 2788 | bb_info_t bb_info = new dse_bb_info_type; |
3072d30e | 2789 | |
55c5ac9f | 2790 | memset (bb_info, 0, sizeof (dse_bb_info_type)); |
3072d30e | 2791 | bitmap_set_bit (all_blocks, bb->index); |
a1b0a968 | 2792 | bb_info->regs_live = regs_live; |
2793 | ||
2794 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2795 | df_simulate_initialize_forwards (bb, regs_live); | |
3072d30e | 2796 | |
2797 | bb_table[bb->index] = bb_info; | |
2798 | cselib_discard_hook = remove_useless_values; | |
2799 | ||
2800 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2801 | { | |
ebabb7a3 | 2802 | rtx_insn *insn; |
3072d30e | 2803 | |
3072d30e | 2804 | active_local_stores = NULL; |
1242bee6 | 2805 | active_local_stores_len = 0; |
3072d30e | 2806 | cselib_clear_table (); |
48e1416a | 2807 | |
3072d30e | 2808 | /* Scan the insns. */ |
2809 | FOR_BB_INSNS (bb, insn) | |
2810 | { | |
2811 | if (INSN_P (insn)) | |
2812 | scan_insn (bb_info, insn); | |
2813 | cselib_process_insn (insn); | |
a1b0a968 | 2814 | if (INSN_P (insn)) |
2815 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
3072d30e | 2816 | } |
48e1416a | 2817 | |
3072d30e | 2818 | /* This is something of a hack, because the global algorithm |
2819 | is supposed to take care of the case where stores go dead | |
2820 | at the end of the function. However, the global | |
2821 | algorithm must take a more conservative view of block | |
2822 | mode reads than the local alg does. So to get the case | |
2823 | where you have a store to the frame followed by a non | |
bef304b8 | 2824 | overlapping block more read, we look at the active local |
3072d30e | 2825 | stores at the end of the function and delete all of the |
2826 | frame and spill based ones. */ | |
2827 | if (stores_off_frame_dead_at_return | |
2828 | && (EDGE_COUNT (bb->succs) == 0 | |
2829 | || (single_succ_p (bb) | |
34154e27 | 2830 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
18d50ae6 | 2831 | && ! crtl->calls_eh_return))) |
3072d30e | 2832 | { |
2833 | insn_info_t i_ptr = active_local_stores; | |
2834 | while (i_ptr) | |
2835 | { | |
2836 | store_info_t store_info = i_ptr->store_rec; | |
2837 | ||
2838 | /* Skip the clobbers. */ | |
2839 | while (!store_info->is_set) | |
2840 | store_info = store_info->next; | |
aa140b76 | 2841 | if (store_info->alias_set && !i_ptr->cannot_delete) |
3072d30e | 2842 | delete_dead_store_insn (i_ptr); |
48e1416a | 2843 | else |
3072d30e | 2844 | if (store_info->group_id >= 0) |
2845 | { | |
48e1416a | 2846 | group_info_t group |
f1f41a6c | 2847 | = rtx_group_vec[store_info->group_id]; |
aa140b76 | 2848 | if (group->frame_related && !i_ptr->cannot_delete) |
3072d30e | 2849 | delete_dead_store_insn (i_ptr); |
2850 | } | |
2851 | ||
2852 | i_ptr = i_ptr->next_local_store; | |
2853 | } | |
2854 | } | |
2855 | ||
2856 | /* Get rid of the loads that were discovered in | |
2857 | replace_read. Cselib is finished with this block. */ | |
2858 | while (deferred_change_list) | |
2859 | { | |
2860 | deferred_change_t next = deferred_change_list->next; | |
2861 | ||
2862 | /* There is no reason to validate this change. That was | |
2863 | done earlier. */ | |
2864 | *deferred_change_list->loc = deferred_change_list->reg; | |
55c5ac9f | 2865 | delete deferred_change_list; |
3072d30e | 2866 | deferred_change_list = next; |
2867 | } | |
2868 | ||
2869 | /* Get rid of all of the cselib based store_infos in this | |
2870 | block and mark the containing insns as not being | |
2871 | deletable. */ | |
2872 | ptr = bb_info->last_insn; | |
2873 | while (ptr) | |
2874 | { | |
2875 | if (ptr->contains_cselib_groups) | |
aa140b76 | 2876 | { |
2877 | store_info_t s_info = ptr->store_rec; | |
2878 | while (s_info && !s_info->is_set) | |
2879 | s_info = s_info->next; | |
2880 | if (s_info | |
2881 | && s_info->redundant_reason | |
2882 | && s_info->redundant_reason->insn | |
2883 | && !ptr->cannot_delete) | |
2884 | { | |
1ca59310 | 2885 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 2886 | fprintf (dump_file, "Locally deleting insn %d " |
2887 | "because insn %d stores the " | |
2888 | "same value and couldn't be " | |
2889 | "eliminated\n", | |
2890 | INSN_UID (ptr->insn), | |
2891 | INSN_UID (s_info->redundant_reason->insn)); | |
2892 | delete_dead_store_insn (ptr); | |
2893 | } | |
aa140b76 | 2894 | free_store_info (ptr); |
2895 | } | |
2896 | else | |
2897 | { | |
2898 | store_info_t s_info; | |
2899 | ||
2900 | /* Free at least positions_needed bitmaps. */ | |
2901 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2902 | if (s_info->is_large) | |
2903 | { | |
843bd2fa | 2904 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
aa140b76 | 2905 | s_info->is_large = false; |
2906 | } | |
2907 | } | |
3072d30e | 2908 | ptr = ptr->prev_insn; |
2909 | } | |
2910 | ||
55c5ac9f | 2911 | cse_store_info_pool.release (); |
3072d30e | 2912 | } |
a1b0a968 | 2913 | bb_info->regs_live = NULL; |
3072d30e | 2914 | } |
2915 | ||
a1b0a968 | 2916 | BITMAP_FREE (regs_live); |
3072d30e | 2917 | cselib_finish (); |
c1f445d2 | 2918 | rtx_group_table->empty (); |
3072d30e | 2919 | } |
2920 | ||
2921 | \f | |
2922 | /*---------------------------------------------------------------------------- | |
2923 | Second step. | |
2924 | ||
2925 | Assign each byte position in the stores that we are going to | |
2926 | analyze globally to a position in the bitmaps. Returns true if | |
6dfdc153 | 2927 | there are any bit positions assigned. |
3072d30e | 2928 | ----------------------------------------------------------------------------*/ |
2929 | ||
2930 | static void | |
2931 | dse_step2_init (void) | |
2932 | { | |
2933 | unsigned int i; | |
2934 | group_info_t group; | |
2935 | ||
f1f41a6c | 2936 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 2937 | { |
2938 | /* For all non stack related bases, we only consider a store to | |
2939 | be deletable if there are two or more stores for that | |
2940 | position. This is because it takes one store to make the | |
2941 | other store redundant. However, for the stores that are | |
2942 | stack related, we consider them if there is only one store | |
2943 | for the position. We do this because the stack related | |
2944 | stores can be deleted if their is no read between them and | |
2945 | the end of the function. | |
48e1416a | 2946 | |
3072d30e | 2947 | To make this work in the current framework, we take the stack |
2948 | related bases add all of the bits from store1 into store2. | |
2949 | This has the effect of making the eligible even if there is | |
2950 | only one store. */ | |
2951 | ||
2952 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2953 | { | |
2954 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2955 | bitmap_ior_into (group->store2_p, group->store1_p); | |
1ca59310 | 2956 | if (dump_file && (dump_flags & TDF_DETAILS)) |
48e1416a | 2957 | fprintf (dump_file, "group %d is frame related ", i); |
3072d30e | 2958 | } |
2959 | ||
2960 | group->offset_map_size_n++; | |
4fb07d00 | 2961 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2962 | group->offset_map_size_n); | |
3072d30e | 2963 | group->offset_map_size_p++; |
4fb07d00 | 2964 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2965 | group->offset_map_size_p); | |
3072d30e | 2966 | group->process_globally = false; |
1ca59310 | 2967 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2968 | { |
48e1416a | 2969 | fprintf (dump_file, "group %d(%d+%d): ", i, |
3072d30e | 2970 | (int)bitmap_count_bits (group->store2_n), |
2971 | (int)bitmap_count_bits (group->store2_p)); | |
2972 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2973 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2974 | } | |
2975 | } | |
2976 | } | |
2977 | ||
2978 | ||
2979 | /* Init the offset tables for the normal case. */ | |
2980 | ||
2981 | static bool | |
2982 | dse_step2_nospill (void) | |
2983 | { | |
2984 | unsigned int i; | |
2985 | group_info_t group; | |
2986 | /* Position 0 is unused because 0 is used in the maps to mean | |
2987 | unused. */ | |
2988 | current_position = 1; | |
f1f41a6c | 2989 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 2990 | { |
2991 | bitmap_iterator bi; | |
2992 | unsigned int j; | |
2993 | ||
2994 | if (group == clear_alias_group) | |
2995 | continue; | |
2996 | ||
9af5ce0c | 2997 | memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
2998 | memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
3072d30e | 2999 | bitmap_clear (group->group_kill); |
3000 | ||
3001 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
3002 | { | |
3003 | bitmap_set_bit (group->group_kill, current_position); | |
b4a708fb | 3004 | if (bitmap_bit_p (group->escaped_n, j)) |
3005 | bitmap_set_bit (kill_on_calls, current_position); | |
3072d30e | 3006 | group->offset_map_n[j] = current_position++; |
3007 | group->process_globally = true; | |
3008 | } | |
3009 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
3010 | { | |
48e1416a | 3011 | bitmap_set_bit (group->group_kill, current_position); |
b4a708fb | 3012 | if (bitmap_bit_p (group->escaped_p, j)) |
3013 | bitmap_set_bit (kill_on_calls, current_position); | |
3072d30e | 3014 | group->offset_map_p[j] = current_position++; |
3015 | group->process_globally = true; | |
3016 | } | |
3017 | } | |
3018 | return current_position != 1; | |
3019 | } | |
3020 | ||
3021 | ||
3072d30e | 3022 | \f |
3023 | /*---------------------------------------------------------------------------- | |
3024 | Third step. | |
48e1416a | 3025 | |
3072d30e | 3026 | Build the bit vectors for the transfer functions. |
3027 | ----------------------------------------------------------------------------*/ | |
3028 | ||
3029 | ||
3072d30e | 3030 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
3031 | there, return 0. */ | |
3032 | ||
3033 | static int | |
3034 | get_bitmap_index (group_info_t group_info, HOST_WIDE_INT offset) | |
3035 | { | |
3036 | if (offset < 0) | |
3037 | { | |
3038 | HOST_WIDE_INT offset_p = -offset; | |
3039 | if (offset_p >= group_info->offset_map_size_n) | |
3040 | return 0; | |
3041 | return group_info->offset_map_n[offset_p]; | |
3042 | } | |
3043 | else | |
3044 | { | |
3045 | if (offset >= group_info->offset_map_size_p) | |
3046 | return 0; | |
3047 | return group_info->offset_map_p[offset]; | |
3048 | } | |
3049 | } | |
3050 | ||
3051 | ||
3052 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
3053 | may be NULL. */ | |
3054 | ||
48e1416a | 3055 | static void |
3072d30e | 3056 | scan_stores_nospill (store_info_t store_info, bitmap gen, bitmap kill) |
3057 | { | |
3058 | while (store_info) | |
3059 | { | |
3060 | HOST_WIDE_INT i; | |
48e1416a | 3061 | group_info_t group_info |
f1f41a6c | 3062 | = rtx_group_vec[store_info->group_id]; |
3072d30e | 3063 | if (group_info->process_globally) |
3064 | for (i = store_info->begin; i < store_info->end; i++) | |
3065 | { | |
3066 | int index = get_bitmap_index (group_info, i); | |
3067 | if (index != 0) | |
3068 | { | |
3069 | bitmap_set_bit (gen, index); | |
3070 | if (kill) | |
3071 | bitmap_clear_bit (kill, index); | |
3072 | } | |
3073 | } | |
3074 | store_info = store_info->next; | |
3075 | } | |
3076 | } | |
3077 | ||
3078 | ||
3079 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
3080 | may be NULL. */ | |
3081 | ||
48e1416a | 3082 | static void |
3072d30e | 3083 | scan_stores_spill (store_info_t store_info, bitmap gen, bitmap kill) |
3084 | { | |
3085 | while (store_info) | |
3086 | { | |
3087 | if (store_info->alias_set) | |
3088 | { | |
48e1416a | 3089 | int index = get_bitmap_index (clear_alias_group, |
3072d30e | 3090 | store_info->alias_set); |
3091 | if (index != 0) | |
3092 | { | |
3093 | bitmap_set_bit (gen, index); | |
3094 | if (kill) | |
3095 | bitmap_clear_bit (kill, index); | |
3096 | } | |
3097 | } | |
3098 | store_info = store_info->next; | |
3099 | } | |
3100 | } | |
3101 | ||
3102 | ||
3103 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3104 | may be NULL. */ | |
3105 | ||
3106 | static void | |
3107 | scan_reads_nospill (insn_info_t insn_info, bitmap gen, bitmap kill) | |
3108 | { | |
3109 | read_info_t read_info = insn_info->read_rec; | |
3110 | int i; | |
3111 | group_info_t group; | |
3112 | ||
16bf64db | 3113 | /* If this insn reads the frame, kill all the frame related stores. */ |
3114 | if (insn_info->frame_read) | |
3115 | { | |
f1f41a6c | 3116 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
16bf64db | 3117 | if (group->process_globally && group->frame_related) |
3118 | { | |
3119 | if (kill) | |
3120 | bitmap_ior_into (kill, group->group_kill); | |
48e1416a | 3121 | bitmap_and_compl_into (gen, group->group_kill); |
16bf64db | 3122 | } |
3123 | } | |
b4a708fb | 3124 | if (insn_info->non_frame_wild_read) |
3125 | { | |
3126 | /* Kill all non-frame related stores. Kill all stores of variables that | |
3127 | escape. */ | |
3128 | if (kill) | |
3129 | bitmap_ior_into (kill, kill_on_calls); | |
3130 | bitmap_and_compl_into (gen, kill_on_calls); | |
f1f41a6c | 3131 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
b4a708fb | 3132 | if (group->process_globally && !group->frame_related) |
3133 | { | |
3134 | if (kill) | |
3135 | bitmap_ior_into (kill, group->group_kill); | |
3136 | bitmap_and_compl_into (gen, group->group_kill); | |
3137 | } | |
3138 | } | |
3072d30e | 3139 | while (read_info) |
3140 | { | |
f1f41a6c | 3141 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 3142 | { |
3143 | if (group->process_globally) | |
3144 | { | |
3145 | if (i == read_info->group_id) | |
3146 | { | |
3147 | if (read_info->begin > read_info->end) | |
3148 | { | |
3149 | /* Begin > end for block mode reads. */ | |
3150 | if (kill) | |
3151 | bitmap_ior_into (kill, group->group_kill); | |
3152 | bitmap_and_compl_into (gen, group->group_kill); | |
3153 | } | |
3154 | else | |
3155 | { | |
3156 | /* The groups are the same, just process the | |
3157 | offsets. */ | |
3158 | HOST_WIDE_INT j; | |
3159 | for (j = read_info->begin; j < read_info->end; j++) | |
3160 | { | |
3161 | int index = get_bitmap_index (group, j); | |
3162 | if (index != 0) | |
3163 | { | |
3164 | if (kill) | |
3165 | bitmap_set_bit (kill, index); | |
3166 | bitmap_clear_bit (gen, index); | |
3167 | } | |
3168 | } | |
3169 | } | |
3170 | } | |
3171 | else | |
3172 | { | |
3173 | /* The groups are different, if the alias sets | |
3174 | conflict, clear the entire group. We only need | |
3175 | to apply this test if the read_info is a cselib | |
3176 | read. Anything with a constant base cannot alias | |
3177 | something else with a different constant | |
3178 | base. */ | |
3179 | if ((read_info->group_id < 0) | |
48e1416a | 3180 | && canon_true_dependence (group->base_mem, |
ec410bf1 | 3181 | GET_MODE (group->base_mem), |
82d2c88b | 3182 | group->canon_base_addr, |
376a287d | 3183 | read_info->mem, NULL_RTX)) |
3072d30e | 3184 | { |
3185 | if (kill) | |
3186 | bitmap_ior_into (kill, group->group_kill); | |
3187 | bitmap_and_compl_into (gen, group->group_kill); | |
3188 | } | |
3189 | } | |
3190 | } | |
3191 | } | |
48e1416a | 3192 | |
3072d30e | 3193 | read_info = read_info->next; |
3194 | } | |
3195 | } | |
3196 | ||
3197 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3198 | may be NULL. */ | |
3199 | ||
3200 | static void | |
3201 | scan_reads_spill (read_info_t read_info, bitmap gen, bitmap kill) | |
3202 | { | |
3203 | while (read_info) | |
3204 | { | |
3205 | if (read_info->alias_set) | |
3206 | { | |
48e1416a | 3207 | int index = get_bitmap_index (clear_alias_group, |
3072d30e | 3208 | read_info->alias_set); |
3209 | if (index != 0) | |
3210 | { | |
3211 | if (kill) | |
3212 | bitmap_set_bit (kill, index); | |
3213 | bitmap_clear_bit (gen, index); | |
3214 | } | |
3215 | } | |
48e1416a | 3216 | |
3072d30e | 3217 | read_info = read_info->next; |
3218 | } | |
3219 | } | |
3220 | ||
3221 | ||
3222 | /* Return the insn in BB_INFO before the first wild read or if there | |
3223 | are no wild reads in the block, return the last insn. */ | |
3224 | ||
3225 | static insn_info_t | |
3226 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3227 | { | |
3228 | insn_info_t insn_info = bb_info->last_insn; | |
3229 | insn_info_t last_wild_read = NULL; | |
3230 | ||
3231 | while (insn_info) | |
3232 | { | |
3233 | if (insn_info->wild_read) | |
3234 | { | |
3235 | last_wild_read = insn_info->prev_insn; | |
3236 | /* Block starts with wild read. */ | |
3237 | if (!last_wild_read) | |
3238 | return NULL; | |
3239 | } | |
3240 | ||
3241 | insn_info = insn_info->prev_insn; | |
3242 | } | |
3243 | ||
3244 | if (last_wild_read) | |
3245 | return last_wild_read; | |
3246 | else | |
3247 | return bb_info->last_insn; | |
3248 | } | |
3249 | ||
3250 | ||
3251 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3252 | the block in order to build the gen and kill sets for the block. | |
3253 | We start at ptr which may be the last insn in the block or may be | |
3254 | the first insn with a wild read. In the latter case we are able to | |
3255 | skip the rest of the block because it just does not matter: | |
3256 | anything that happens is hidden by the wild read. */ | |
3257 | ||
3258 | static void | |
3259 | dse_step3_scan (bool for_spills, basic_block bb) | |
3260 | { | |
3261 | bb_info_t bb_info = bb_table[bb->index]; | |
3262 | insn_info_t insn_info; | |
3263 | ||
3264 | if (for_spills) | |
3265 | /* There are no wild reads in the spill case. */ | |
3266 | insn_info = bb_info->last_insn; | |
3267 | else | |
3268 | insn_info = find_insn_before_first_wild_read (bb_info); | |
48e1416a | 3269 | |
3072d30e | 3270 | /* In the spill case or in the no_spill case if there is no wild |
3271 | read in the block, we will need a kill set. */ | |
3272 | if (insn_info == bb_info->last_insn) | |
3273 | { | |
3274 | if (bb_info->kill) | |
3275 | bitmap_clear (bb_info->kill); | |
3276 | else | |
4fb07d00 | 3277 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3278 | } |
48e1416a | 3279 | else |
3072d30e | 3280 | if (bb_info->kill) |
3281 | BITMAP_FREE (bb_info->kill); | |
3282 | ||
3283 | while (insn_info) | |
3284 | { | |
3285 | /* There may have been code deleted by the dce pass run before | |
3286 | this phase. */ | |
3287 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3288 | { | |
48e1416a | 3289 | /* Process the read(s) last. */ |
3072d30e | 3290 | if (for_spills) |
3291 | { | |
3292 | scan_stores_spill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3293 | scan_reads_spill (insn_info->read_rec, bb_info->gen, bb_info->kill); | |
3294 | } | |
3295 | else | |
3296 | { | |
3297 | scan_stores_nospill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3298 | scan_reads_nospill (insn_info, bb_info->gen, bb_info->kill); | |
3299 | } | |
48e1416a | 3300 | } |
3072d30e | 3301 | |
3302 | insn_info = insn_info->prev_insn; | |
3303 | } | |
3304 | } | |
3305 | ||
3306 | ||
3307 | /* Set the gen set of the exit block, and also any block with no | |
3308 | successors that does not have a wild read. */ | |
3309 | ||
3310 | static void | |
3311 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3312 | { | |
3313 | /* The gen set is all 0's for the exit block except for the | |
3314 | frame_pointer_group. */ | |
48e1416a | 3315 | |
3072d30e | 3316 | if (stores_off_frame_dead_at_return) |
3317 | { | |
3318 | unsigned int i; | |
3319 | group_info_t group; | |
48e1416a | 3320 | |
f1f41a6c | 3321 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 3322 | { |
3323 | if (group->process_globally && group->frame_related) | |
3324 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3325 | } | |
3326 | } | |
3327 | } | |
3328 | ||
3329 | ||
3330 | /* Find all of the blocks that are not backwards reachable from the | |
3331 | exit block or any block with no successors (BB). These are the | |
3332 | infinite loops or infinite self loops. These blocks will still | |
3333 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3334 | ||
3335 | static void | |
3336 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3337 | { | |
3338 | edge e; | |
3339 | edge_iterator ei; | |
3340 | ||
08b7917c | 3341 | if (bitmap_bit_p (unreachable_blocks, bb->index)) |
3072d30e | 3342 | { |
08b7917c | 3343 | bitmap_clear_bit (unreachable_blocks, bb->index); |
3072d30e | 3344 | FOR_EACH_EDGE (e, ei, bb->preds) |
48e1416a | 3345 | { |
3072d30e | 3346 | mark_reachable_blocks (unreachable_blocks, e->src); |
48e1416a | 3347 | } |
3072d30e | 3348 | } |
3349 | } | |
3350 | ||
3351 | /* Build the transfer functions for the function. */ | |
3352 | ||
3353 | static void | |
3354 | dse_step3 (bool for_spills) | |
3355 | { | |
3356 | basic_block bb; | |
fe672ac0 | 3357 | sbitmap unreachable_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
3072d30e | 3358 | sbitmap_iterator sbi; |
3359 | bitmap all_ones = NULL; | |
3360 | unsigned int i; | |
48e1416a | 3361 | |
53c5d9d4 | 3362 | bitmap_ones (unreachable_blocks); |
3072d30e | 3363 | |
ed7d889a | 3364 | FOR_ALL_BB_FN (bb, cfun) |
3072d30e | 3365 | { |
3366 | bb_info_t bb_info = bb_table[bb->index]; | |
3367 | if (bb_info->gen) | |
3368 | bitmap_clear (bb_info->gen); | |
3369 | else | |
4fb07d00 | 3370 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3371 | |
3372 | if (bb->index == ENTRY_BLOCK) | |
3373 | ; | |
3374 | else if (bb->index == EXIT_BLOCK) | |
3375 | dse_step3_exit_block_scan (bb_info); | |
3376 | else | |
3377 | dse_step3_scan (for_spills, bb); | |
3378 | if (EDGE_COUNT (bb->succs) == 0) | |
3379 | mark_reachable_blocks (unreachable_blocks, bb); | |
3380 | ||
3381 | /* If this is the second time dataflow is run, delete the old | |
3382 | sets. */ | |
3383 | if (bb_info->in) | |
3384 | BITMAP_FREE (bb_info->in); | |
3385 | if (bb_info->out) | |
3386 | BITMAP_FREE (bb_info->out); | |
3387 | } | |
3388 | ||
3389 | /* For any block in an infinite loop, we must initialize the out set | |
3390 | to all ones. This could be expensive, but almost never occurs in | |
3391 | practice. However, it is common in regression tests. */ | |
0d211963 | 3392 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
3072d30e | 3393 | { |
3394 | if (bitmap_bit_p (all_blocks, i)) | |
3395 | { | |
3396 | bb_info_t bb_info = bb_table[i]; | |
3397 | if (!all_ones) | |
3398 | { | |
3399 | unsigned int j; | |
3400 | group_info_t group; | |
3401 | ||
4fb07d00 | 3402 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
f1f41a6c | 3403 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
3072d30e | 3404 | bitmap_ior_into (all_ones, group->group_kill); |
3405 | } | |
3406 | if (!bb_info->out) | |
3407 | { | |
4fb07d00 | 3408 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3409 | bitmap_copy (bb_info->out, all_ones); |
3410 | } | |
3411 | } | |
3412 | } | |
3413 | ||
3414 | if (all_ones) | |
3415 | BITMAP_FREE (all_ones); | |
3416 | sbitmap_free (unreachable_blocks); | |
3417 | } | |
3418 | ||
3419 | ||
3420 | \f | |
3421 | /*---------------------------------------------------------------------------- | |
3422 | Fourth step. | |
3423 | ||
3424 | Solve the bitvector equations. | |
3425 | ----------------------------------------------------------------------------*/ | |
3426 | ||
3427 | ||
3428 | /* Confluence function for blocks with no successors. Create an out | |
3429 | set from the gen set of the exit block. This block logically has | |
3430 | the exit block as a successor. */ | |
3431 | ||
3432 | ||
3433 | ||
3434 | static void | |
3435 | dse_confluence_0 (basic_block bb) | |
3436 | { | |
3437 | bb_info_t bb_info = bb_table[bb->index]; | |
3438 | ||
3439 | if (bb->index == EXIT_BLOCK) | |
3440 | return; | |
3441 | ||
3442 | if (!bb_info->out) | |
3443 | { | |
4fb07d00 | 3444 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3445 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3446 | } | |
3447 | } | |
3448 | ||
3449 | /* Propagate the information from the in set of the dest of E to the | |
3450 | out set of the src of E. If the various in or out sets are not | |
3451 | there, that means they are all ones. */ | |
3452 | ||
a703ca31 | 3453 | static bool |
3072d30e | 3454 | dse_confluence_n (edge e) |
3455 | { | |
3456 | bb_info_t src_info = bb_table[e->src->index]; | |
3457 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3458 | ||
3459 | if (dest_info->in) | |
3460 | { | |
3461 | if (src_info->out) | |
3462 | bitmap_and_into (src_info->out, dest_info->in); | |
3463 | else | |
3464 | { | |
4fb07d00 | 3465 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3466 | bitmap_copy (src_info->out, dest_info->in); |
3467 | } | |
3468 | } | |
a703ca31 | 3469 | return true; |
3072d30e | 3470 | } |
3471 | ||
3472 | ||
3473 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
48e1416a | 3474 | block. There are three cases: |
3072d30e | 3475 | |
3476 | 1) The block has no kill set. In this case the kill set is all | |
3477 | ones. It does not matter what the out set of the block is, none of | |
3478 | the info can reach the top. The only thing that reaches the top is | |
3479 | the gen set and we just copy the set. | |
3480 | ||
3481 | 2) There is a kill set but no out set and bb has successors. In | |
3482 | this case we just return. Eventually an out set will be created and | |
3483 | it is better to wait than to create a set of ones. | |
3484 | ||
3485 | 3) There is both a kill and out set. We apply the obvious transfer | |
3486 | function. | |
3487 | */ | |
3488 | ||
3489 | static bool | |
3490 | dse_transfer_function (int bb_index) | |
3491 | { | |
3492 | bb_info_t bb_info = bb_table[bb_index]; | |
3493 | ||
3494 | if (bb_info->kill) | |
3495 | { | |
3496 | if (bb_info->out) | |
3497 | { | |
3498 | /* Case 3 above. */ | |
3499 | if (bb_info->in) | |
48e1416a | 3500 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
3072d30e | 3501 | bb_info->out, bb_info->kill); |
3502 | else | |
3503 | { | |
4fb07d00 | 3504 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
48e1416a | 3505 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
3072d30e | 3506 | bb_info->out, bb_info->kill); |
3507 | return true; | |
3508 | } | |
3509 | } | |
3510 | else | |
3511 | /* Case 2 above. */ | |
3512 | return false; | |
3513 | } | |
3514 | else | |
3515 | { | |
3516 | /* Case 1 above. If there is already an in set, nothing | |
3517 | happens. */ | |
3518 | if (bb_info->in) | |
3519 | return false; | |
3520 | else | |
3521 | { | |
4fb07d00 | 3522 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3523 | bitmap_copy (bb_info->in, bb_info->gen); |
3524 | return true; | |
3525 | } | |
3526 | } | |
3527 | } | |
3528 | ||
3529 | /* Solve the dataflow equations. */ | |
3530 | ||
3531 | static void | |
3532 | dse_step4 (void) | |
3533 | { | |
48e1416a | 3534 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3535 | dse_confluence_n, dse_transfer_function, | |
3536 | all_blocks, df_get_postorder (DF_BACKWARD), | |
3072d30e | 3537 | df_get_n_blocks (DF_BACKWARD)); |
1ca59310 | 3538 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 3539 | { |
3540 | basic_block bb; | |
3541 | ||
3542 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
ed7d889a | 3543 | FOR_ALL_BB_FN (bb, cfun) |
3072d30e | 3544 | { |
3545 | bb_info_t bb_info = bb_table[bb->index]; | |
3546 | ||
3547 | df_print_bb_index (bb, dump_file); | |
3548 | if (bb_info->in) | |
3549 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3550 | else | |
3551 | fprintf (dump_file, " in: *MISSING*\n"); | |
3552 | if (bb_info->gen) | |
3553 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3554 | else | |
3555 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3556 | if (bb_info->kill) | |
3557 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3558 | else | |
3559 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3560 | if (bb_info->out) | |
3561 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3562 | else | |
3563 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3564 | } | |
3565 | } | |
3566 | } | |
3567 | ||
3568 | ||
3569 | \f | |
3570 | /*---------------------------------------------------------------------------- | |
3571 | Fifth step. | |
3572 | ||
bef304b8 | 3573 | Delete the stores that can only be deleted using the global information. |
3072d30e | 3574 | ----------------------------------------------------------------------------*/ |
3575 | ||
3576 | ||
3577 | static void | |
3578 | dse_step5_nospill (void) | |
3579 | { | |
3580 | basic_block bb; | |
fc00614f | 3581 | FOR_EACH_BB_FN (bb, cfun) |
3072d30e | 3582 | { |
3583 | bb_info_t bb_info = bb_table[bb->index]; | |
3584 | insn_info_t insn_info = bb_info->last_insn; | |
3585 | bitmap v = bb_info->out; | |
3586 | ||
3587 | while (insn_info) | |
3588 | { | |
3589 | bool deleted = false; | |
3590 | if (dump_file && insn_info->insn) | |
3591 | { | |
3592 | fprintf (dump_file, "starting to process insn %d\n", | |
3593 | INSN_UID (insn_info->insn)); | |
3594 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3595 | } | |
3596 | ||
3597 | /* There may have been code deleted by the dce pass run before | |
3598 | this phase. */ | |
48e1416a | 3599 | if (insn_info->insn |
3072d30e | 3600 | && INSN_P (insn_info->insn) |
3601 | && (!insn_info->cannot_delete) | |
3602 | && (!bitmap_empty_p (v))) | |
3603 | { | |
3604 | store_info_t store_info = insn_info->store_rec; | |
3605 | ||
3606 | /* Try to delete the current insn. */ | |
3607 | deleted = true; | |
48e1416a | 3608 | |
3072d30e | 3609 | /* Skip the clobbers. */ |
3610 | while (!store_info->is_set) | |
3611 | store_info = store_info->next; | |
3612 | ||
3613 | if (store_info->alias_set) | |
3614 | deleted = false; | |
3615 | else | |
3616 | { | |
3617 | HOST_WIDE_INT i; | |
48e1416a | 3618 | group_info_t group_info |
f1f41a6c | 3619 | = rtx_group_vec[store_info->group_id]; |
48e1416a | 3620 | |
3072d30e | 3621 | for (i = store_info->begin; i < store_info->end; i++) |
3622 | { | |
3623 | int index = get_bitmap_index (group_info, i); | |
48e1416a | 3624 | |
1ca59310 | 3625 | if (dump_file && (dump_flags & TDF_DETAILS)) |
48e1416a | 3626 | fprintf (dump_file, "i = %d, index = %d\n", (int)i, index); |
3072d30e | 3627 | if (index == 0 || !bitmap_bit_p (v, index)) |
3628 | { | |
1ca59310 | 3629 | if (dump_file && (dump_flags & TDF_DETAILS)) |
48e1416a | 3630 | fprintf (dump_file, "failing at i = %d\n", (int)i); |
3072d30e | 3631 | deleted = false; |
3632 | break; | |
3633 | } | |
3634 | } | |
3635 | } | |
3636 | if (deleted) | |
3637 | { | |
5a9ecd4a | 3638 | if (dbg_cnt (dse) |
3639 | && check_for_inc_dec_1 (insn_info)) | |
3072d30e | 3640 | { |
3072d30e | 3641 | delete_insn (insn_info->insn); |
3642 | insn_info->insn = NULL; | |
3643 | globally_deleted++; | |
3644 | } | |
3645 | } | |
3646 | } | |
3647 | /* We do want to process the local info if the insn was | |
6dfdc153 | 3648 | deleted. For instance, if the insn did a wild read, we |
3072d30e | 3649 | no longer need to trash the info. */ |
48e1416a | 3650 | if (insn_info->insn |
3072d30e | 3651 | && INSN_P (insn_info->insn) |
3652 | && (!deleted)) | |
3653 | { | |
3654 | scan_stores_nospill (insn_info->store_rec, v, NULL); | |
3655 | if (insn_info->wild_read) | |
3656 | { | |
1ca59310 | 3657 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 3658 | fprintf (dump_file, "wild read\n"); |
3659 | bitmap_clear (v); | |
3660 | } | |
b4a708fb | 3661 | else if (insn_info->read_rec |
3662 | || insn_info->non_frame_wild_read) | |
3072d30e | 3663 | { |
b4a708fb | 3664 | if (dump_file && !insn_info->non_frame_wild_read) |
3072d30e | 3665 | fprintf (dump_file, "regular read\n"); |
1ca59310 | 3666 | else if (dump_file && (dump_flags & TDF_DETAILS)) |
b4a708fb | 3667 | fprintf (dump_file, "non-frame wild read\n"); |
3072d30e | 3668 | scan_reads_nospill (insn_info, v, NULL); |
3669 | } | |
3670 | } | |
48e1416a | 3671 | |
3072d30e | 3672 | insn_info = insn_info->prev_insn; |
3673 | } | |
3674 | } | |
3675 | } | |
3676 | ||
3677 | ||
3072d30e | 3678 | \f |
3679 | /*---------------------------------------------------------------------------- | |
3680 | Sixth step. | |
3681 | ||
aa140b76 | 3682 | Delete stores made redundant by earlier stores (which store the same |
3683 | value) that couldn't be eliminated. | |
3684 | ----------------------------------------------------------------------------*/ | |
3685 | ||
3686 | static void | |
3687 | dse_step6 (void) | |
3688 | { | |
3689 | basic_block bb; | |
3690 | ||
ed7d889a | 3691 | FOR_ALL_BB_FN (bb, cfun) |
aa140b76 | 3692 | { |
3693 | bb_info_t bb_info = bb_table[bb->index]; | |
3694 | insn_info_t insn_info = bb_info->last_insn; | |
3695 | ||
3696 | while (insn_info) | |
3697 | { | |
3698 | /* There may have been code deleted by the dce pass run before | |
3699 | this phase. */ | |
3700 | if (insn_info->insn | |
3701 | && INSN_P (insn_info->insn) | |
3702 | && !insn_info->cannot_delete) | |
3703 | { | |
3704 | store_info_t s_info = insn_info->store_rec; | |
3705 | ||
3706 | while (s_info && !s_info->is_set) | |
3707 | s_info = s_info->next; | |
3708 | if (s_info | |
3709 | && s_info->redundant_reason | |
3710 | && s_info->redundant_reason->insn | |
3711 | && INSN_P (s_info->redundant_reason->insn)) | |
3712 | { | |
cccc26f6 | 3713 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
1ca59310 | 3714 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 3715 | fprintf (dump_file, "Locally deleting insn %d " |
3716 | "because insn %d stores the " | |
3717 | "same value and couldn't be " | |
3718 | "eliminated\n", | |
3719 | INSN_UID (insn_info->insn), | |
3720 | INSN_UID (rinsn)); | |
3721 | delete_dead_store_insn (insn_info); | |
3722 | } | |
3723 | } | |
3724 | insn_info = insn_info->prev_insn; | |
3725 | } | |
3726 | } | |
3727 | } | |
3728 | \f | |
3729 | /*---------------------------------------------------------------------------- | |
3730 | Seventh step. | |
3731 | ||
48e1416a | 3732 | Destroy everything left standing. |
3072d30e | 3733 | ----------------------------------------------------------------------------*/ |
3734 | ||
48e1416a | 3735 | static void |
4fb07d00 | 3736 | dse_step7 (void) |
3072d30e | 3737 | { |
4fb07d00 | 3738 | bitmap_obstack_release (&dse_bitmap_obstack); |
3739 | obstack_free (&dse_obstack, NULL); | |
ce299759 | 3740 | |
3072d30e | 3741 | end_alias_analysis (); |
3742 | free (bb_table); | |
c1f445d2 | 3743 | delete rtx_group_table; |
3744 | rtx_group_table = NULL; | |
f1f41a6c | 3745 | rtx_group_vec.release (); |
3072d30e | 3746 | BITMAP_FREE (all_blocks); |
3747 | BITMAP_FREE (scratch); | |
3748 | ||
55c5ac9f | 3749 | rtx_store_info_pool.release (); |
3750 | read_info_type::pool.release (); | |
3751 | insn_info_type::pool.release (); | |
3752 | dse_bb_info_type::pool.release (); | |
3753 | group_info::pool.release (); | |
3754 | deferred_change::pool.release (); | |
3072d30e | 3755 | } |
3756 | ||
3757 | ||
3072d30e | 3758 | /* ------------------------------------------------------------------------- |
3759 | DSE | |
3760 | ------------------------------------------------------------------------- */ | |
3761 | ||
3762 | /* Callback for running pass_rtl_dse. */ | |
3763 | ||
3764 | static unsigned int | |
3765 | rest_of_handle_dse (void) | |
3766 | { | |
3072d30e | 3767 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3768 | ||
a1b0a968 | 3769 | /* Need the notes since we must track live hardregs in the forwards |
3770 | direction. */ | |
3771 | df_note_add_problem (); | |
3772 | df_analyze (); | |
3773 | ||
3072d30e | 3774 | dse_step0 (); |
3775 | dse_step1 (); | |
3776 | dse_step2_init (); | |
3777 | if (dse_step2_nospill ()) | |
3778 | { | |
3779 | df_set_flags (DF_LR_RUN_DCE); | |
3780 | df_analyze (); | |
1ca59310 | 3781 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 3782 | fprintf (dump_file, "doing global processing\n"); |
3783 | dse_step3 (false); | |
3784 | dse_step4 (); | |
3785 | dse_step5_nospill (); | |
3786 | } | |
3787 | ||
aa140b76 | 3788 | dse_step6 (); |
4fb07d00 | 3789 | dse_step7 (); |
3072d30e | 3790 | |
3791 | if (dump_file) | |
3792 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d, spill deletions = %d\n", | |
3793 | locally_deleted, globally_deleted, spill_deleted); | |
1f91a12d | 3794 | |
3795 | /* DSE can eliminate potentially-trapping MEMs. | |
3796 | Remove any EH edges associated with them. */ | |
3797 | if ((locally_deleted || globally_deleted) | |
3798 | && cfun->can_throw_non_call_exceptions | |
3799 | && purge_all_dead_edges ()) | |
3800 | cleanup_cfg (0); | |
3801 | ||
3072d30e | 3802 | return 0; |
3803 | } | |
3804 | ||
cbe8bda8 | 3805 | namespace { |
3806 | ||
3807 | const pass_data pass_data_rtl_dse1 = | |
3808 | { | |
3809 | RTL_PASS, /* type */ | |
3810 | "dse1", /* name */ | |
3811 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 3812 | TV_DSE1, /* tv_id */ |
3813 | 0, /* properties_required */ | |
3814 | 0, /* properties_provided */ | |
3815 | 0, /* properties_destroyed */ | |
3816 | 0, /* todo_flags_start */ | |
8b88439e | 3817 | TODO_df_finish, /* todo_flags_finish */ |
3072d30e | 3818 | }; |
3819 | ||
cbe8bda8 | 3820 | class pass_rtl_dse1 : public rtl_opt_pass |
3821 | { | |
3822 | public: | |
9af5ce0c | 3823 | pass_rtl_dse1 (gcc::context *ctxt) |
3824 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
cbe8bda8 | 3825 | {} |
3826 | ||
3827 | /* opt_pass methods: */ | |
31315c24 | 3828 | virtual bool gate (function *) |
3829 | { | |
3830 | return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
3831 | } | |
3832 | ||
65b0537f | 3833 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
cbe8bda8 | 3834 | |
3835 | }; // class pass_rtl_dse1 | |
3836 | ||
3837 | } // anon namespace | |
3838 | ||
3839 | rtl_opt_pass * | |
3840 | make_pass_rtl_dse1 (gcc::context *ctxt) | |
3841 | { | |
3842 | return new pass_rtl_dse1 (ctxt); | |
3843 | } | |
3844 | ||
3845 | namespace { | |
3846 | ||
3847 | const pass_data pass_data_rtl_dse2 = | |
3848 | { | |
3849 | RTL_PASS, /* type */ | |
3850 | "dse2", /* name */ | |
3851 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 3852 | TV_DSE2, /* tv_id */ |
3853 | 0, /* properties_required */ | |
3854 | 0, /* properties_provided */ | |
3855 | 0, /* properties_destroyed */ | |
3856 | 0, /* todo_flags_start */ | |
8b88439e | 3857 | TODO_df_finish, /* todo_flags_finish */ |
3072d30e | 3858 | }; |
cbe8bda8 | 3859 | |
3860 | class pass_rtl_dse2 : public rtl_opt_pass | |
3861 | { | |
3862 | public: | |
9af5ce0c | 3863 | pass_rtl_dse2 (gcc::context *ctxt) |
3864 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
cbe8bda8 | 3865 | {} |
3866 | ||
3867 | /* opt_pass methods: */ | |
31315c24 | 3868 | virtual bool gate (function *) |
3869 | { | |
3870 | return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
3871 | } | |
3872 | ||
65b0537f | 3873 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
cbe8bda8 | 3874 | |
3875 | }; // class pass_rtl_dse2 | |
3876 | ||
3877 | } // anon namespace | |
3878 | ||
3879 | rtl_opt_pass * | |
3880 | make_pass_rtl_dse2 (gcc::context *ctxt) | |
3881 | { | |
3882 | return new pass_rtl_dse2 (ctxt); | |
3883 | } |