]>
Commit | Line | Data |
---|---|---|
3072d30e | 1 | /* RTL dead store elimination. |
fbd26352 | 2 | Copyright (C) 2005-2019 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" | |
9ef16211 | 28 | #include "backend.h" |
7c29e30e | 29 | #include "target.h" |
30 | #include "rtl.h" | |
9ef16211 | 31 | #include "tree.h" |
32 | #include "gimple.h" | |
7c29e30e | 33 | #include "predict.h" |
9ef16211 | 34 | #include "df.h" |
ad7b10a2 | 35 | #include "memmodel.h" |
7c29e30e | 36 | #include "tm_p.h" |
37 | #include "gimple-ssa.h" | |
38 | #include "expmed.h" | |
39 | #include "optabs.h" | |
7c29e30e | 40 | #include "emit-rtl.h" |
41 | #include "recog.h" | |
b20a8bb4 | 42 | #include "alias.h" |
9ed99284 | 43 | #include "stor-layout.h" |
94ea8568 | 44 | #include "cfgrtl.h" |
3072d30e | 45 | #include "cselib.h" |
3072d30e | 46 | #include "tree-pass.h" |
d53441c8 | 47 | #include "explow.h" |
3072d30e | 48 | #include "expr.h" |
3072d30e | 49 | #include "dbgcnt.h" |
1242bee6 | 50 | #include "params.h" |
ec1203cd | 51 | #include "rtl-iter.h" |
1f91a12d | 52 | #include "cfgcleanup.h" |
3072d30e | 53 | |
54 | /* This file contains three techniques for performing Dead Store | |
48e1416a | 55 | Elimination (dse). |
3072d30e | 56 | |
57 | * The first technique performs dse locally on any base address. It | |
58 | is based on the cselib which is a local value numbering technique. | |
59 | This technique is local to a basic block but deals with a fairly | |
60 | general addresses. | |
48e1416a | 61 | |
3072d30e | 62 | * The second technique performs dse globally but is restricted to |
63 | base addresses that are either constant or are relative to the | |
64 | frame_pointer. | |
65 | ||
66 | * The third technique, (which is only done after register allocation) | |
47ae02b7 | 67 | processes the spill slots. This differs from the second |
3072d30e | 68 | technique because it takes advantage of the fact that spilling is |
69 | completely free from the effects of aliasing. | |
70 | ||
71 | Logically, dse is a backwards dataflow problem. A store can be | |
72 | deleted if it if cannot be reached in the backward direction by any | |
73 | use of the value being stored. However, the local technique uses a | |
74 | forwards scan of the basic block because cselib requires that the | |
75 | block be processed in that order. | |
76 | ||
77 | The pass is logically broken into 7 steps: | |
78 | ||
79 | 0) Initialization. | |
80 | ||
81 | 1) The local algorithm, as well as scanning the insns for the two | |
82 | global algorithms. | |
83 | ||
84 | 2) Analysis to see if the global algs are necessary. In the case | |
85 | of stores base on a constant address, there must be at least two | |
86 | stores to that address, to make it possible to delete some of the | |
87 | stores. In the case of stores off of the frame or spill related | |
88 | stores, only one store to an address is necessary because those | |
89 | stores die at the end of the function. | |
90 | ||
48e1416a | 91 | 3) Set up the global dataflow equations based on processing the |
3072d30e | 92 | info parsed in the first step. |
93 | ||
94 | 4) Solve the dataflow equations. | |
95 | ||
96 | 5) Delete the insns that the global analysis has indicated are | |
97 | unnecessary. | |
98 | ||
9d75589a | 99 | 6) Delete insns that store the same value as preceding store |
aa140b76 | 100 | where the earlier store couldn't be eliminated. |
101 | ||
102 | 7) Cleanup. | |
3072d30e | 103 | |
104 | This step uses cselib and canon_rtx to build the largest expression | |
105 | possible for each address. This pass is a forwards pass through | |
106 | each basic block. From the point of view of the global technique, | |
107 | the first pass could examine a block in either direction. The | |
bef304b8 | 108 | forwards ordering is to accommodate cselib. |
3072d30e | 109 | |
2d0fd66d | 110 | We make a simplifying assumption: addresses fall into four broad |
3072d30e | 111 | categories: |
112 | ||
113 | 1) base has rtx_varies_p == false, offset is constant. | |
114 | 2) base has rtx_varies_p == false, offset variable. | |
115 | 3) base has rtx_varies_p == true, offset constant. | |
116 | 4) base has rtx_varies_p == true, offset variable. | |
117 | ||
118 | The local passes are able to process all 4 kinds of addresses. The | |
2d0fd66d | 119 | global pass only handles 1). |
3072d30e | 120 | |
121 | The global problem is formulated as follows: | |
122 | ||
123 | A store, S1, to address A, where A is not relative to the stack | |
124 | frame, can be eliminated if all paths from S1 to the end of the | |
2d0fd66d | 125 | function contain another store to A before a read to A. |
3072d30e | 126 | |
127 | If the address A is relative to the stack frame, a store S2 to A | |
2d0fd66d | 128 | can be eliminated if there are no paths from S2 that reach the |
3072d30e | 129 | end of the function that read A before another store to A. In |
2d0fd66d | 130 | this case S2 can be deleted if there are paths from S2 to the |
3072d30e | 131 | end of the function that have no reads or writes to A. This |
132 | second case allows stores to the stack frame to be deleted that | |
133 | would otherwise die when the function returns. This cannot be | |
134 | done if stores_off_frame_dead_at_return is not true. See the doc | |
135 | for that variable for when this variable is false. | |
136 | ||
137 | The global problem is formulated as a backwards set union | |
138 | dataflow problem where the stores are the gens and reads are the | |
139 | kills. Set union problems are rare and require some special | |
140 | handling given our representation of bitmaps. A straightforward | |
2d0fd66d | 141 | implementation requires a lot of bitmaps filled with 1s. |
3072d30e | 142 | These are expensive and cumbersome in our bitmap formulation so |
143 | care has been taken to avoid large vectors filled with 1s. See | |
144 | the comments in bb_info and in the dataflow confluence functions | |
48e1416a | 145 | for details. |
3072d30e | 146 | |
147 | There are two places for further enhancements to this algorithm: | |
48e1416a | 148 | |
3072d30e | 149 | 1) The original dse which was embedded in a pass called flow also |
150 | did local address forwarding. For example in | |
151 | ||
152 | A <- r100 | |
153 | ... <- A | |
154 | ||
155 | flow would replace the right hand side of the second insn with a | |
6dfdc153 | 156 | reference to r100. Most of the information is available to add this |
3072d30e | 157 | to this pass. It has not done it because it is a lot of work in |
158 | the case that either r100 is assigned to between the first and | |
159 | second insn and/or the second insn is a load of part of the value | |
160 | stored by the first insn. | |
161 | ||
162 | insn 5 in gcc.c-torture/compile/990203-1.c simple case. | |
163 | insn 15 in gcc.c-torture/execute/20001017-2.c simple case. | |
164 | insn 25 in gcc.c-torture/execute/20001026-1.c simple case. | |
165 | insn 44 in gcc.c-torture/execute/20010910-1.c simple case. | |
166 | ||
167 | 2) The cleaning up of spill code is quite profitable. It currently | |
168 | depends on reading tea leaves and chicken entrails left by reload. | |
169 | This pass depends on reload creating a singleton alias set for each | |
170 | spill slot and telling the next dse pass which of these alias sets | |
171 | are the singletons. Rather than analyze the addresses of the | |
172 | spills, dse's spill processing just does analysis of the loads and | |
173 | stores that use those alias sets. There are three cases where this | |
174 | falls short: | |
175 | ||
176 | a) Reload sometimes creates the slot for one mode of access, and | |
177 | then inserts loads and/or stores for a smaller mode. In this | |
178 | case, the current code just punts on the slot. The proper thing | |
179 | to do is to back out and use one bit vector position for each | |
180 | byte of the entity associated with the slot. This depends on | |
181 | KNOWING that reload always generates the accesses for each of the | |
182 | bytes in some canonical (read that easy to understand several | |
183 | passes after reload happens) way. | |
184 | ||
185 | b) Reload sometimes decides that spill slot it allocated was not | |
186 | large enough for the mode and goes back and allocates more slots | |
187 | with the same mode and alias set. The backout in this case is a | |
188 | little more graceful than (a). In this case the slot is unmarked | |
189 | as being a spill slot and if final address comes out to be based | |
48e1416a | 190 | off the frame pointer, the global algorithm handles this slot. |
3072d30e | 191 | |
192 | c) For any pass that may prespill, there is currently no | |
193 | mechanism to tell the dse pass that the slot being used has the | |
194 | special properties that reload uses. It may be that all that is | |
bef304b8 | 195 | required is to have those passes make the same calls that reload |
3072d30e | 196 | does, assuming that the alias sets can be manipulated in the same |
197 | way. */ | |
198 | ||
199 | /* There are limits to the size of constant offsets we model for the | |
200 | global problem. There are certainly test cases, that exceed this | |
201 | limit, however, it is unlikely that there are important programs | |
202 | that really have constant offsets this size. */ | |
203 | #define MAX_OFFSET (64 * 1024) | |
204 | ||
4fb07d00 | 205 | /* Obstack for the DSE dataflow bitmaps. We don't want to put these |
206 | on the default obstack because these bitmaps can grow quite large | |
207 | (~2GB for the small (!) test case of PR54146) and we'll hold on to | |
208 | all that memory until the end of the compiler run. | |
209 | As a bonus, delete_tree_live_info can destroy all the bitmaps by just | |
210 | releasing the whole obstack. */ | |
211 | static bitmap_obstack dse_bitmap_obstack; | |
212 | ||
213 | /* Obstack for other data. As for above: Kinda nice to be able to | |
214 | throw it all away at the end in one big sweep. */ | |
215 | static struct obstack dse_obstack; | |
216 | ||
217 | /* Scratch bitmap for cselib's cselib_expand_value_rtx. */ | |
3072d30e | 218 | static bitmap scratch = NULL; |
4fb07d00 | 219 | |
55c5ac9f | 220 | struct insn_info_type; |
3072d30e | 221 | |
222 | /* This structure holds information about a candidate store. */ | |
48e1416a | 223 | struct store_info |
3072d30e | 224 | { |
225 | ||
226 | /* False means this is a clobber. */ | |
227 | bool is_set; | |
228 | ||
aa140b76 | 229 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
230 | bool is_large; | |
231 | ||
3072d30e | 232 | /* The id of the mem group of the base address. If rtx_varies_p is |
233 | true, this is -1. Otherwise, it is the index into the group | |
234 | table. */ | |
235 | int group_id; | |
48e1416a | 236 | |
3072d30e | 237 | /* This is the cselib value. */ |
238 | cselib_val *cse_base; | |
239 | ||
240 | /* This canonized mem. */ | |
241 | rtx mem; | |
242 | ||
82d2c88b | 243 | /* Canonized MEM address for use by canon_true_dependence. */ |
3072d30e | 244 | rtx mem_addr; |
245 | ||
57bca8dc | 246 | /* The offset of the first byte associated with the operation. */ |
370ad826 | 247 | poly_int64 offset; |
57bca8dc | 248 | |
249 | /* The number of bytes covered by the operation. This is always exact | |
250 | and known (rather than -1). */ | |
370ad826 | 251 | poly_int64 width; |
aa140b76 | 252 | |
253 | union | |
254 | { | |
255 | /* A bitmask as wide as the number of bytes in the word that | |
256 | contains a 1 if the byte may be needed. The store is unused if | |
257 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
258 | unsigned HOST_WIDE_INT small_bitmask; | |
259 | ||
260 | struct | |
261 | { | |
370ad826 | 262 | /* A bitmap with one bit per byte, or null if the number of |
263 | bytes isn't known at compile time. A cleared bit means | |
264 | the position is needed. Used if IS_LARGE is true. */ | |
843bd2fa | 265 | bitmap bmap; |
3072d30e | 266 | |
370ad826 | 267 | /* When BITMAP is nonnull, this counts the number of set bits |
268 | (i.e. unneeded bytes) in the bitmap. If it is equal to | |
269 | WIDTH, the whole store is unused. | |
270 | ||
271 | When BITMAP is null: | |
272 | - the store is definitely not needed when COUNT == 1 | |
273 | - all the store is needed when COUNT == 0 and RHS is nonnull | |
274 | - otherwise we don't know which parts of the store are needed. */ | |
aa140b76 | 275 | int count; |
276 | } large; | |
277 | } positions_needed; | |
3072d30e | 278 | |
279 | /* The next store info for this insn. */ | |
280 | struct store_info *next; | |
281 | ||
282 | /* The right hand side of the store. This is used if there is a | |
283 | subsequent reload of the mems address somewhere later in the | |
284 | basic block. */ | |
aa140b76 | 285 | rtx rhs; |
286 | ||
287 | /* If rhs is or holds a constant, this contains that constant, | |
288 | otherwise NULL. */ | |
289 | rtx const_rhs; | |
290 | ||
291 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
292 | insn stored. These aren't eliminated early, because doing that | |
293 | might prevent the earlier larger store to be eliminated. */ | |
55c5ac9f | 294 | struct insn_info_type *redundant_reason; |
3072d30e | 295 | }; |
296 | ||
4e43e20a | 297 | /* Return a bitmask with the first N low bits set. */ |
298 | ||
299 | static unsigned HOST_WIDE_INT | |
300 | lowpart_bitmask (int n) | |
301 | { | |
7097b942 | 302 | unsigned HOST_WIDE_INT mask = HOST_WIDE_INT_M1U; |
4e43e20a | 303 | return mask >> (HOST_BITS_PER_WIDE_INT - n); |
304 | } | |
305 | ||
1dc6c44d | 306 | static object_allocator<store_info> cse_store_info_pool ("cse_store_info_pool"); |
55c5ac9f | 307 | |
1dc6c44d | 308 | static object_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool"); |
3072d30e | 309 | |
310 | /* This structure holds information about a load. These are only | |
311 | built for rtx bases. */ | |
55c5ac9f | 312 | struct read_info_type |
3072d30e | 313 | { |
314 | /* The id of the mem group of the base address. */ | |
315 | int group_id; | |
316 | ||
57bca8dc | 317 | /* The offset of the first byte associated with the operation. */ |
370ad826 | 318 | poly_int64 offset; |
57bca8dc | 319 | |
320 | /* The number of bytes covered by the operation, or -1 if not known. */ | |
370ad826 | 321 | poly_int64 width; |
3072d30e | 322 | |
323 | /* The mem being read. */ | |
324 | rtx mem; | |
325 | ||
326 | /* The next read_info for this insn. */ | |
55c5ac9f | 327 | struct read_info_type *next; |
3072d30e | 328 | }; |
55c5ac9f | 329 | typedef struct read_info_type *read_info_t; |
3072d30e | 330 | |
1dc6c44d | 331 | static object_allocator<read_info_type> read_info_type_pool ("read_info_pool"); |
3072d30e | 332 | |
333 | /* One of these records is created for each insn. */ | |
334 | ||
55c5ac9f | 335 | struct insn_info_type |
3072d30e | 336 | { |
337 | /* Set true if the insn contains a store but the insn itself cannot | |
338 | be deleted. This is set if the insn is a parallel and there is | |
339 | more than one non dead output or if the insn is in some way | |
340 | volatile. */ | |
341 | bool cannot_delete; | |
342 | ||
343 | /* This field is only used by the global algorithm. It is set true | |
344 | if the insn contains any read of mem except for a (1). This is | |
345 | also set if the insn is a call or has a clobber mem. If the insn | |
346 | contains a wild read, the use_rec will be null. */ | |
347 | bool wild_read; | |
348 | ||
b4a708fb | 349 | /* This is true only for CALL instructions which could potentially read |
350 | any non-frame memory location. This field is used by the global | |
351 | algorithm. */ | |
352 | bool non_frame_wild_read; | |
353 | ||
17e1318c | 354 | /* This field is only used for the processing of const functions. |
355 | These functions cannot read memory, but they can read the stack | |
16bf64db | 356 | because that is where they may get their parms. We need to be |
357 | this conservative because, like the store motion pass, we don't | |
358 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
359 | Moreover, we need to distinguish two cases: | |
360 | 1. Before reload (register elimination), the stores related to | |
361 | outgoing arguments are stack pointer based and thus deemed | |
362 | of non-constant base in this pass. This requires special | |
363 | handling but also means that the frame pointer based stores | |
364 | need not be killed upon encountering a const function call. | |
365 | 2. After reload, the stores related to outgoing arguments can be | |
366 | either stack pointer or hard frame pointer based. This means | |
367 | that we have no other choice than also killing all the frame | |
368 | pointer based stores upon encountering a const function call. | |
17853422 | 369 | This field is set after reload for const function calls and before |
370 | reload for const tail function calls on targets where arg pointer | |
371 | is the frame pointer. Having this set is less severe than a wild | |
372 | read, it just means that all the frame related stores are killed | |
373 | rather than all the stores. */ | |
16bf64db | 374 | bool frame_read; |
375 | ||
376 | /* This field is only used for the processing of const functions. | |
377 | It is set if the insn may contain a stack pointer based store. */ | |
17e1318c | 378 | bool stack_pointer_based; |
3072d30e | 379 | |
380 | /* This is true if any of the sets within the store contains a | |
381 | cselib base. Such stores can only be deleted by the local | |
382 | algorithm. */ | |
383 | bool contains_cselib_groups; | |
384 | ||
385 | /* The insn. */ | |
ebabb7a3 | 386 | rtx_insn *insn; |
3072d30e | 387 | |
388 | /* The list of mem sets or mem clobbers that are contained in this | |
389 | insn. If the insn is deletable, it contains only one mem set. | |
390 | But it could also contain clobbers. Insns that contain more than | |
391 | one mem set are not deletable, but each of those mems are here in | |
6dfdc153 | 392 | order to provide info to delete other insns. */ |
c2eab334 | 393 | store_info *store_rec; |
3072d30e | 394 | |
395 | /* The linked list of mem uses in this insn. Only the reads from | |
396 | rtx bases are listed here. The reads to cselib bases are | |
397 | completely processed during the first scan and so are never | |
398 | created. */ | |
399 | read_info_t read_rec; | |
400 | ||
5a9ecd4a | 401 | /* The live fixed registers. We assume only fixed registers can |
402 | cause trouble by being clobbered from an expanded pattern; | |
403 | storing only the live fixed registers (rather than all registers) | |
404 | means less memory needs to be allocated / copied for the individual | |
405 | stores. */ | |
406 | regset fixed_regs_live; | |
407 | ||
3072d30e | 408 | /* The prev insn in the basic block. */ |
55c5ac9f | 409 | struct insn_info_type * prev_insn; |
3072d30e | 410 | |
411 | /* The linked list of insns that are in consideration for removal in | |
9d75589a | 412 | the forwards pass through the basic block. This pointer may be |
3072d30e | 413 | trash as it is not cleared when a wild read occurs. The only |
f0b5f617 | 414 | time it is guaranteed to be correct is when the traversal starts |
3072d30e | 415 | at active_local_stores. */ |
55c5ac9f | 416 | struct insn_info_type * next_local_store; |
3072d30e | 417 | }; |
55c5ac9f | 418 | typedef struct insn_info_type *insn_info_t; |
3072d30e | 419 | |
1dc6c44d | 420 | static object_allocator<insn_info_type> insn_info_type_pool ("insn_info_pool"); |
3072d30e | 421 | |
422 | /* The linked list of stores that are under consideration in this | |
48e1416a | 423 | basic block. */ |
3072d30e | 424 | static insn_info_t active_local_stores; |
1242bee6 | 425 | static int active_local_stores_len; |
3072d30e | 426 | |
55c5ac9f | 427 | struct dse_bb_info_type |
3072d30e | 428 | { |
3072d30e | 429 | /* Pointer to the insn info for the last insn in the block. These |
430 | are linked so this is how all of the insns are reached. During | |
431 | scanning this is the current insn being scanned. */ | |
432 | insn_info_t last_insn; | |
433 | ||
434 | /* The info for the global dataflow problem. */ | |
435 | ||
436 | ||
437 | /* This is set if the transfer function should and in the wild_read | |
438 | bitmap before applying the kill and gen sets. That vector knocks | |
439 | out most of the bits in the bitmap and thus speeds up the | |
440 | operations. */ | |
441 | bool apply_wild_read; | |
442 | ||
a1b0a968 | 443 | /* The following 4 bitvectors hold information about which positions |
444 | of which stores are live or dead. They are indexed by | |
445 | get_bitmap_index. */ | |
446 | ||
3072d30e | 447 | /* The set of store positions that exist in this block before a wild read. */ |
448 | bitmap gen; | |
48e1416a | 449 | |
3072d30e | 450 | /* The set of load positions that exist in this block above the |
451 | same position of a store. */ | |
452 | bitmap kill; | |
453 | ||
454 | /* The set of stores that reach the top of the block without being | |
455 | killed by a read. | |
456 | ||
457 | Do not represent the in if it is all ones. Note that this is | |
458 | what the bitvector should logically be initialized to for a set | |
459 | intersection problem. However, like the kill set, this is too | |
460 | expensive. So initially, the in set will only be created for the | |
461 | exit block and any block that contains a wild read. */ | |
462 | bitmap in; | |
463 | ||
464 | /* The set of stores that reach the bottom of the block from it's | |
465 | successors. | |
466 | ||
467 | Do not represent the in if it is all ones. Note that this is | |
468 | what the bitvector should logically be initialized to for a set | |
469 | intersection problem. However, like the kill and in set, this is | |
470 | too expensive. So what is done is that the confluence operator | |
471 | just initializes the vector from one of the out sets of the | |
472 | successors of the block. */ | |
473 | bitmap out; | |
a1b0a968 | 474 | |
475 | /* The following bitvector is indexed by the reg number. It | |
476 | contains the set of regs that are live at the current instruction | |
477 | being processed. While it contains info for all of the | |
5a9ecd4a | 478 | registers, only the hard registers are actually examined. It is used |
479 | to assure that shift and/or add sequences that are inserted do not | |
9d75589a | 480 | accidentally clobber live hard regs. */ |
a1b0a968 | 481 | bitmap regs_live; |
3072d30e | 482 | }; |
483 | ||
55c5ac9f | 484 | typedef struct dse_bb_info_type *bb_info_t; |
e16712b1 | 485 | |
486 | static object_allocator<dse_bb_info_type> dse_bb_info_type_pool | |
1dc6c44d | 487 | ("bb_info_pool"); |
3072d30e | 488 | |
489 | /* Table to hold all bb_infos. */ | |
490 | static bb_info_t *bb_table; | |
491 | ||
492 | /* There is a group_info for each rtx base that is used to reference | |
493 | memory. There are also not many of the rtx bases because they are | |
494 | very limited in scope. */ | |
495 | ||
48e1416a | 496 | struct group_info |
3072d30e | 497 | { |
498 | /* The actual base of the address. */ | |
499 | rtx rtx_base; | |
500 | ||
501 | /* The sequential id of the base. This allows us to have a | |
502 | canonical ordering of these that is not based on addresses. */ | |
503 | int id; | |
504 | ||
0ac758f7 | 505 | /* True if there are any positions that are to be processed |
506 | globally. */ | |
507 | bool process_globally; | |
508 | ||
509 | /* True if the base of this group is either the frame_pointer or | |
510 | hard_frame_pointer. */ | |
511 | bool frame_related; | |
512 | ||
ec410bf1 | 513 | /* A mem wrapped around the base pointer for the group in order to do |
514 | read dependency. It must be given BLKmode in order to encompass all | |
515 | the possible offsets from the base. */ | |
3072d30e | 516 | rtx base_mem; |
48e1416a | 517 | |
82d2c88b | 518 | /* Canonized version of base_mem's address. */ |
519 | rtx canon_base_addr; | |
3072d30e | 520 | |
521 | /* These two sets of two bitmaps are used to keep track of how many | |
6dfdc153 | 522 | stores are actually referencing that position from this base. We |
3072d30e | 523 | only do this for rtx bases as this will be used to assign |
6dfdc153 | 524 | positions in the bitmaps for the global problem. Bit N is set in |
3072d30e | 525 | store1 on the first store for offset N. Bit N is set in store2 |
526 | for the second store to offset N. This is all we need since we | |
527 | only care about offsets that have two or more stores for them. | |
528 | ||
529 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
530 | for 0 and greater offsets. | |
531 | ||
532 | There is one special case here, for stores into the stack frame, | |
533 | we will or store1 into store2 before deciding which stores look | |
534 | at globally. This is because stores to the stack frame that have | |
535 | no other reads before the end of the function can also be | |
536 | deleted. */ | |
537 | bitmap store1_n, store1_p, store2_n, store2_p; | |
538 | ||
b4a708fb | 539 | /* These bitmaps keep track of offsets in this group escape this function. |
540 | An offset escapes if it corresponds to a named variable whose | |
541 | addressable flag is set. */ | |
542 | bitmap escaped_n, escaped_p; | |
543 | ||
6dfdc153 | 544 | /* The positions in this bitmap have the same assignments as the in, |
3072d30e | 545 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6dfdc153 | 546 | the positions that are occupied by stores for this group. */ |
3072d30e | 547 | bitmap group_kill; |
548 | ||
3072d30e | 549 | /* The offset_map is used to map the offsets from this base into |
6dfdc153 | 550 | positions in the global bitmaps. It is only created after all of |
3072d30e | 551 | the all of stores have been scanned and we know which ones we |
552 | care about. */ | |
48e1416a | 553 | int *offset_map_n, *offset_map_p; |
554 | int offset_map_size_n, offset_map_size_p; | |
3072d30e | 555 | }; |
55c5ac9f | 556 | |
1dc6c44d | 557 | static object_allocator<group_info> group_info_pool ("rtx_group_info_pool"); |
3072d30e | 558 | |
3072d30e | 559 | /* Index into the rtx_group_vec. */ |
560 | static int rtx_group_next_id; | |
561 | ||
3072d30e | 562 | |
c2eab334 | 563 | static vec<group_info *> rtx_group_vec; |
3072d30e | 564 | |
565 | ||
566 | /* This structure holds the set of changes that are being deferred | |
567 | when removing read operation. See replace_read. */ | |
48e1416a | 568 | struct deferred_change |
3072d30e | 569 | { |
570 | ||
571 | /* The mem that is being replaced. */ | |
572 | rtx *loc; | |
573 | ||
574 | /* The reg it is being replaced with. */ | |
575 | rtx reg; | |
576 | ||
577 | struct deferred_change *next; | |
578 | }; | |
579 | ||
e16712b1 | 580 | static object_allocator<deferred_change> deferred_change_pool |
1dc6c44d | 581 | ("deferred_change_pool"); |
3072d30e | 582 | |
c2eab334 | 583 | static deferred_change *deferred_change_list = NULL; |
3072d30e | 584 | |
18d50ae6 | 585 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
ff3ae375 | 586 | this for vararg functions because they play games with the frame. */ |
3072d30e | 587 | static bool stores_off_frame_dead_at_return; |
588 | ||
589 | /* Counter for stats. */ | |
48e1416a | 590 | static int globally_deleted; |
591 | static int locally_deleted; | |
48e1416a | 592 | |
3072d30e | 593 | static bitmap all_blocks; |
594 | ||
b4a708fb | 595 | /* Locations that are killed by calls in the global phase. */ |
596 | static bitmap kill_on_calls; | |
597 | ||
3072d30e | 598 | /* The number of bits used in the global bitmaps. */ |
599 | static unsigned int current_position; | |
57bca8dc | 600 | |
601 | /* Print offset range [OFFSET, OFFSET + WIDTH) to FILE. */ | |
602 | ||
603 | static void | |
604 | print_range (FILE *file, poly_int64 offset, poly_int64 width) | |
605 | { | |
606 | fprintf (file, "["); | |
607 | print_dec (offset, file, SIGNED); | |
608 | fprintf (file, ".."); | |
609 | print_dec (offset + width, file, SIGNED); | |
610 | fprintf (file, ")"); | |
611 | } | |
3072d30e | 612 | \f |
613 | /*---------------------------------------------------------------------------- | |
614 | Zeroth step. | |
615 | ||
48e1416a | 616 | Initialization. |
3072d30e | 617 | ----------------------------------------------------------------------------*/ |
618 | ||
3072d30e | 619 | |
3072d30e | 620 | /* Hashtable callbacks for maintaining the "bases" field of |
621 | store_group_info, given that the addresses are function invariants. */ | |
622 | ||
770ff93b | 623 | struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
d1455aa3 | 624 | { |
9969c043 | 625 | static inline hashval_t hash (const group_info *); |
626 | static inline bool equal (const group_info *, const group_info *); | |
d1455aa3 | 627 | }; |
628 | ||
629 | inline bool | |
9969c043 | 630 | invariant_group_base_hasher::equal (const group_info *gi1, |
631 | const group_info *gi2) | |
3072d30e | 632 | { |
3072d30e | 633 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
634 | } | |
635 | ||
d1455aa3 | 636 | inline hashval_t |
9969c043 | 637 | invariant_group_base_hasher::hash (const group_info *gi) |
3072d30e | 638 | { |
3072d30e | 639 | int do_not_record; |
640 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
641 | } | |
642 | ||
d1455aa3 | 643 | /* Tables of group_info structures, hashed by base value. */ |
c1f445d2 | 644 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
d1455aa3 | 645 | |
3072d30e | 646 | |
647 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
648 | ||
c2eab334 | 649 | static group_info * |
3072d30e | 650 | get_group_info (rtx base) |
651 | { | |
48e1416a | 652 | struct group_info tmp_gi; |
c2eab334 | 653 | group_info *gi; |
d1455aa3 | 654 | group_info **slot; |
3072d30e | 655 | |
5a3f5d69 | 656 | gcc_assert (base != NULL_RTX); |
657 | ||
658 | /* Find the store_base_info structure for BASE, creating a new one | |
659 | if necessary. */ | |
660 | tmp_gi.rtx_base = base; | |
661 | slot = rtx_group_table->find_slot (&tmp_gi, INSERT); | |
662 | gi = *slot; | |
3072d30e | 663 | |
664 | if (gi == NULL) | |
665 | { | |
e16712b1 | 666 | *slot = gi = group_info_pool.allocate (); |
3072d30e | 667 | gi->rtx_base = base; |
668 | gi->id = rtx_group_next_id++; | |
ec410bf1 | 669 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
82d2c88b | 670 | gi->canon_base_addr = canon_rtx (base); |
4fb07d00 | 671 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
672 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
673 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
674 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
675 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
676 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
677 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
3072d30e | 678 | gi->process_globally = false; |
48e1416a | 679 | gi->frame_related = |
3072d30e | 680 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
681 | gi->offset_map_size_n = 0; | |
682 | gi->offset_map_size_p = 0; | |
683 | gi->offset_map_n = NULL; | |
684 | gi->offset_map_p = NULL; | |
f1f41a6c | 685 | rtx_group_vec.safe_push (gi); |
3072d30e | 686 | } |
687 | ||
688 | return gi; | |
689 | } | |
690 | ||
691 | ||
692 | /* Initialization of data structures. */ | |
693 | ||
694 | static void | |
695 | dse_step0 (void) | |
696 | { | |
697 | locally_deleted = 0; | |
698 | globally_deleted = 0; | |
3072d30e | 699 | |
4fb07d00 | 700 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
701 | gcc_obstack_init (&dse_obstack); | |
702 | ||
703 | scratch = BITMAP_ALLOC (®_obstack); | |
704 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
3072d30e | 705 | |
3072d30e | 706 | |
c1f445d2 | 707 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
3072d30e | 708 | |
fe672ac0 | 709 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
3072d30e | 710 | rtx_group_next_id = 0; |
711 | ||
18d50ae6 | 712 | stores_off_frame_dead_at_return = !cfun->stdarg; |
3072d30e | 713 | |
714 | init_alias_analysis (); | |
3072d30e | 715 | } |
716 | ||
717 | ||
718 | \f | |
719 | /*---------------------------------------------------------------------------- | |
720 | First step. | |
721 | ||
722 | Scan all of the insns. Any random ordering of the blocks is fine. | |
bef304b8 | 723 | Each block is scanned in forward order to accommodate cselib which |
3072d30e | 724 | is used to remove stores with non-constant bases. |
725 | ----------------------------------------------------------------------------*/ | |
726 | ||
727 | /* Delete all of the store_info recs from INSN_INFO. */ | |
728 | ||
48e1416a | 729 | static void |
3072d30e | 730 | free_store_info (insn_info_t insn_info) |
731 | { | |
c2eab334 | 732 | store_info *cur = insn_info->store_rec; |
733 | while (cur) | |
3072d30e | 734 | { |
c2eab334 | 735 | store_info *next = cur->next; |
736 | if (cur->is_large) | |
737 | BITMAP_FREE (cur->positions_needed.large.bmap); | |
738 | if (cur->cse_base) | |
739 | cse_store_info_pool.remove (cur); | |
3072d30e | 740 | else |
c2eab334 | 741 | rtx_store_info_pool.remove (cur); |
742 | cur = next; | |
3072d30e | 743 | } |
744 | ||
745 | insn_info->cannot_delete = true; | |
746 | insn_info->contains_cselib_groups = false; | |
747 | insn_info->store_rec = NULL; | |
748 | } | |
749 | ||
6dc50383 | 750 | struct note_add_store_info |
5a9ecd4a | 751 | { |
4cd001d5 | 752 | rtx_insn *first, *current; |
5a9ecd4a | 753 | regset fixed_regs_live; |
754 | bool failure; | |
6dc50383 | 755 | }; |
5a9ecd4a | 756 | |
757 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
758 | Check if a register is clobbered which is live afterwards. */ | |
759 | ||
760 | static void | |
761 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
762 | { | |
4cd001d5 | 763 | rtx_insn *insn; |
5a9ecd4a | 764 | note_add_store_info *info = (note_add_store_info *) data; |
5a9ecd4a | 765 | |
766 | if (!REG_P (loc)) | |
767 | return; | |
768 | ||
769 | /* If this register is referenced by the current or an earlier insn, | |
770 | that's OK. E.g. this applies to the register that is being incremented | |
771 | with this addition. */ | |
772 | for (insn = info->first; | |
773 | insn != NEXT_INSN (info->current); | |
774 | insn = NEXT_INSN (insn)) | |
775 | if (reg_referenced_p (loc, PATTERN (insn))) | |
776 | return; | |
777 | ||
778 | /* If we come here, we have a clobber of a register that's only OK | |
779 | if that register is not live. If we don't have liveness information | |
780 | available, fail now. */ | |
781 | if (!info->fixed_regs_live) | |
782 | { | |
6a298741 | 783 | info->failure = true; |
5a9ecd4a | 784 | return; |
785 | } | |
786 | /* Now check if this is a live fixed register. */ | |
6a298741 | 787 | unsigned int end_regno = END_REGNO (loc); |
788 | for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) | |
789 | if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) | |
790 | info->failure = true; | |
5a9ecd4a | 791 | } |
792 | ||
1f864115 | 793 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
794 | SRC + SRCOFF before insn ARG. */ | |
3072d30e | 795 | |
796 | static int | |
1f864115 | 797 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
798 | rtx op ATTRIBUTE_UNUSED, | |
799 | rtx dest, rtx src, rtx srcoff, void *arg) | |
3072d30e | 800 | { |
5a9ecd4a | 801 | insn_info_t insn_info = (insn_info_t) arg; |
4cd001d5 | 802 | rtx_insn *insn = insn_info->insn, *new_insn, *cur; |
5a9ecd4a | 803 | note_add_store_info info; |
48e1416a | 804 | |
1f864115 | 805 | /* We can reuse all operands without copying, because we are about |
806 | to delete the insn that contained it. */ | |
5a9ecd4a | 807 | if (srcoff) |
a280a5da | 808 | { |
809 | start_sequence (); | |
810 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
811 | new_insn = get_insns (); | |
812 | end_sequence (); | |
813 | } | |
5a9ecd4a | 814 | else |
f9a00e9e | 815 | new_insn = gen_move_insn (dest, src); |
5a9ecd4a | 816 | info.first = new_insn; |
817 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
818 | info.failure = false; | |
819 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
820 | { | |
821 | info.current = cur; | |
822 | note_stores (PATTERN (cur), note_add_store, &info); | |
823 | } | |
3072d30e | 824 | |
5a9ecd4a | 825 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
826 | return it immediately, communicating the failure to its caller. */ | |
827 | if (info.failure) | |
828 | return 1; | |
829 | ||
830 | emit_insn_before (new_insn, insn); | |
3072d30e | 831 | |
623ad592 | 832 | return 0; |
3072d30e | 833 | } |
834 | ||
5a9ecd4a | 835 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
836 | is there, is split into a separate insn. | |
837 | Return true on success (or if there was nothing to do), false on failure. */ | |
3072d30e | 838 | |
5a9ecd4a | 839 | static bool |
840 | check_for_inc_dec_1 (insn_info_t insn_info) | |
3072d30e | 841 | { |
ebabb7a3 | 842 | rtx_insn *insn = insn_info->insn; |
3072d30e | 843 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
844 | if (note) | |
623ad592 | 845 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
846 | insn_info) == 0; | |
5a9ecd4a | 847 | return true; |
3072d30e | 848 | } |
849 | ||
850 | ||
5a9ecd4a | 851 | /* Entry point for postreload. If you work on reload_cse, or you need this |
852 | anywhere else, consider if you can provide register liveness information | |
853 | and add a parameter to this function so that it can be passed down in | |
854 | insn_info.fixed_regs_live. */ | |
855 | bool | |
ebabb7a3 | 856 | check_for_inc_dec (rtx_insn *insn) |
5a9ecd4a | 857 | { |
55c5ac9f | 858 | insn_info_type insn_info; |
5a9ecd4a | 859 | rtx note; |
860 | ||
861 | insn_info.insn = insn; | |
862 | insn_info.fixed_regs_live = NULL; | |
863 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
864 | if (note) | |
623ad592 | 865 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
866 | &insn_info) == 0; | |
5a9ecd4a | 867 | return true; |
868 | } | |
869 | ||
48e1416a | 870 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
3072d30e | 871 | |
872 | static void | |
873 | delete_dead_store_insn (insn_info_t insn_info) | |
874 | { | |
875 | read_info_t read_info; | |
876 | ||
877 | if (!dbg_cnt (dse)) | |
878 | return; | |
879 | ||
5a9ecd4a | 880 | if (!check_for_inc_dec_1 (insn_info)) |
881 | return; | |
1ca59310 | 882 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5a3f5d69 | 883 | fprintf (dump_file, "Locally deleting insn %d\n", |
884 | INSN_UID (insn_info->insn)); | |
3072d30e | 885 | |
886 | free_store_info (insn_info); | |
887 | read_info = insn_info->read_rec; | |
48e1416a | 888 | |
3072d30e | 889 | while (read_info) |
890 | { | |
891 | read_info_t next = read_info->next; | |
e16712b1 | 892 | read_info_type_pool.remove (read_info); |
3072d30e | 893 | read_info = next; |
894 | } | |
895 | insn_info->read_rec = NULL; | |
896 | ||
897 | delete_insn (insn_info->insn); | |
898 | locally_deleted++; | |
899 | insn_info->insn = NULL; | |
900 | ||
901 | insn_info->wild_read = false; | |
902 | } | |
903 | ||
f7b5f694 | 904 | /* Return whether DECL, a local variable, can possibly escape the current |
905 | function scope. */ | |
906 | ||
907 | static bool | |
908 | local_variable_can_escape (tree decl) | |
909 | { | |
910 | if (TREE_ADDRESSABLE (decl)) | |
911 | return true; | |
912 | ||
913 | /* If this is a partitioned variable, we need to consider all the variables | |
914 | in the partition. This is necessary because a store into one of them can | |
915 | be replaced with a store into another and this may not change the outcome | |
916 | of the escape analysis. */ | |
917 | if (cfun->gimple_df->decls_to_pointers != NULL) | |
918 | { | |
5f8841a5 | 919 | tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); |
f7b5f694 | 920 | if (namep) |
5f8841a5 | 921 | return TREE_ADDRESSABLE (*namep); |
f7b5f694 | 922 | } |
923 | ||
924 | return false; | |
925 | } | |
926 | ||
927 | /* Return whether EXPR can possibly escape the current function scope. */ | |
928 | ||
b4a708fb | 929 | static bool |
930 | can_escape (tree expr) | |
931 | { | |
932 | tree base; | |
933 | if (!expr) | |
934 | return true; | |
935 | base = get_base_address (expr); | |
936 | if (DECL_P (base) | |
f7b5f694 | 937 | && !may_be_aliased (base) |
53e9c5c4 | 938 | && !(VAR_P (base) |
f7b5f694 | 939 | && !DECL_EXTERNAL (base) |
940 | && !TREE_STATIC (base) | |
941 | && local_variable_can_escape (base))) | |
b4a708fb | 942 | return false; |
943 | return true; | |
944 | } | |
3072d30e | 945 | |
946 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
947 | OFFSET and WIDTH. */ | |
948 | ||
949 | static void | |
370ad826 | 950 | set_usage_bits (group_info *group, poly_int64 offset, poly_int64 width, |
b4a708fb | 951 | tree expr) |
3072d30e | 952 | { |
370ad826 | 953 | /* Non-constant offsets and widths act as global kills, so there's no point |
954 | trying to use them to derive global DSE candidates. */ | |
955 | HOST_WIDE_INT i, const_offset, const_width; | |
b4a708fb | 956 | bool expr_escapes = can_escape (expr); |
370ad826 | 957 | if (offset.is_constant (&const_offset) |
958 | && width.is_constant (&const_width) | |
959 | && const_offset > -MAX_OFFSET | |
960 | && const_offset + const_width < MAX_OFFSET) | |
961 | for (i = const_offset; i < const_offset + const_width; ++i) | |
3072d30e | 962 | { |
963 | bitmap store1; | |
964 | bitmap store2; | |
b4a708fb | 965 | bitmap escaped; |
3072d30e | 966 | int ai; |
967 | if (i < 0) | |
968 | { | |
969 | store1 = group->store1_n; | |
970 | store2 = group->store2_n; | |
b4a708fb | 971 | escaped = group->escaped_n; |
3072d30e | 972 | ai = -i; |
973 | } | |
974 | else | |
975 | { | |
976 | store1 = group->store1_p; | |
977 | store2 = group->store2_p; | |
b4a708fb | 978 | escaped = group->escaped_p; |
3072d30e | 979 | ai = i; |
980 | } | |
48e1416a | 981 | |
6ef9bbe0 | 982 | if (!bitmap_set_bit (store1, ai)) |
3072d30e | 983 | bitmap_set_bit (store2, ai); |
48e1416a | 984 | else |
3072d30e | 985 | { |
3072d30e | 986 | if (i < 0) |
987 | { | |
988 | if (group->offset_map_size_n < ai) | |
989 | group->offset_map_size_n = ai; | |
990 | } | |
991 | else | |
992 | { | |
993 | if (group->offset_map_size_p < ai) | |
994 | group->offset_map_size_p = ai; | |
995 | } | |
996 | } | |
b4a708fb | 997 | if (expr_escapes) |
998 | bitmap_set_bit (escaped, ai); | |
3072d30e | 999 | } |
1000 | } | |
1001 | ||
b4a708fb | 1002 | static void |
1003 | reset_active_stores (void) | |
1004 | { | |
1005 | active_local_stores = NULL; | |
1006 | active_local_stores_len = 0; | |
1007 | } | |
3072d30e | 1008 | |
b4a708fb | 1009 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
3072d30e | 1010 | |
1011 | static void | |
b4a708fb | 1012 | free_read_records (bb_info_t bb_info) |
3072d30e | 1013 | { |
1014 | insn_info_t insn_info = bb_info->last_insn; | |
1015 | read_info_t *ptr = &insn_info->read_rec; | |
3072d30e | 1016 | while (*ptr) |
1017 | { | |
1018 | read_info_t next = (*ptr)->next; | |
5a3f5d69 | 1019 | read_info_type_pool.remove (*ptr); |
1020 | *ptr = next; | |
3072d30e | 1021 | } |
b4a708fb | 1022 | } |
1023 | ||
1024 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1025 | ||
1026 | static void | |
1027 | add_wild_read (bb_info_t bb_info) | |
1028 | { | |
1029 | insn_info_t insn_info = bb_info->last_insn; | |
3072d30e | 1030 | insn_info->wild_read = true; |
b4a708fb | 1031 | free_read_records (bb_info); |
1032 | reset_active_stores (); | |
3072d30e | 1033 | } |
1034 | ||
b4a708fb | 1035 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1036 | non-frame locations. */ | |
1037 | ||
1038 | static void | |
1039 | add_non_frame_wild_read (bb_info_t bb_info) | |
1040 | { | |
1041 | insn_info_t insn_info = bb_info->last_insn; | |
1042 | insn_info->non_frame_wild_read = true; | |
1043 | free_read_records (bb_info); | |
1044 | reset_active_stores (); | |
1045 | } | |
3072d30e | 1046 | |
17e1318c | 1047 | /* Return true if X is a constant or one of the registers that behave |
1048 | as a constant over the life of a function. This is equivalent to | |
1049 | !rtx_varies_p for memory addresses. */ | |
3072d30e | 1050 | |
1051 | static bool | |
1052 | const_or_frame_p (rtx x) | |
1053 | { | |
0349edce | 1054 | if (CONSTANT_P (x)) |
1055 | return true; | |
1056 | ||
1057 | if (GET_CODE (x) == REG) | |
3072d30e | 1058 | { |
3072d30e | 1059 | /* Note that we have to test for the actual rtx used for the frame |
1060 | and arg pointers and not just the register number in case we have | |
1061 | eliminated the frame and/or arg pointer and are using it | |
1062 | for pseudos. */ | |
1063 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1064 | /* The arg pointer varies if it is not a fixed register. */ | |
1065 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1066 | || x == pic_offset_table_rtx) | |
1067 | return true; | |
1068 | return false; | |
3072d30e | 1069 | } |
55c5ac9f | 1070 | |
0349edce | 1071 | return false; |
3072d30e | 1072 | } |
1073 | ||
48e1416a | 1074 | /* Take all reasonable action to put the address of MEM into the form |
1075 | that we can do analysis on. | |
3072d30e | 1076 | |
1077 | The gold standard is to get the address into the form: address + | |
1078 | OFFSET where address is something that rtx_varies_p considers a | |
1079 | constant. When we can get the address in this form, we can do | |
1080 | global analysis on it. Note that for constant bases, address is | |
1081 | not actually returned, only the group_id. The address can be | |
1082 | obtained from that. | |
1083 | ||
1084 | If that fails, we try cselib to get a value we can at least use | |
48e1416a | 1085 | locally. If that fails we return false. |
1086 | ||
3072d30e | 1087 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1088 | group for non_varying bases. | |
1089 | ||
1090 | FOR_READ is true if this is a mem read and false if not. */ | |
1091 | ||
1092 | static bool | |
1093 | canon_address (rtx mem, | |
3072d30e | 1094 | int *group_id, |
370ad826 | 1095 | poly_int64 *offset, |
3072d30e | 1096 | cselib_val **base) |
1097 | { | |
3754d046 | 1098 | machine_mode address_mode = get_address_mode (mem); |
3072d30e | 1099 | rtx mem_address = XEXP (mem, 0); |
1100 | rtx expanded_address, address; | |
627540ce | 1101 | int expanded; |
1102 | ||
1f864115 | 1103 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
3072d30e | 1104 | |
1ca59310 | 1105 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1106 | { |
1107 | fprintf (dump_file, " mem: "); | |
1108 | print_inline_rtx (dump_file, mem_address, 0); | |
1109 | fprintf (dump_file, "\n"); | |
1110 | } | |
1111 | ||
627540ce | 1112 | /* First see if just canon_rtx (mem_address) is const or frame, |
1113 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1114 | address = NULL_RTX; | |
1115 | for (expanded = 0; expanded < 2; expanded++) | |
1116 | { | |
1117 | if (expanded) | |
1118 | { | |
1119 | /* Use cselib to replace all of the reg references with the full | |
48e1416a | 1120 | expression. This will take care of the case where we have |
3072d30e | 1121 | |
627540ce | 1122 | r_x = base + offset; |
1123 | val = *r_x; | |
48e1416a | 1124 | |
1125 | by making it into | |
3072d30e | 1126 | |
627540ce | 1127 | val = *(base + offset); */ |
3072d30e | 1128 | |
627540ce | 1129 | expanded_address = cselib_expand_value_rtx (mem_address, |
1130 | scratch, 5); | |
3072d30e | 1131 | |
627540ce | 1132 | /* If this fails, just go with the address from first |
1133 | iteration. */ | |
1134 | if (!expanded_address) | |
1135 | break; | |
1136 | } | |
1137 | else | |
1138 | expanded_address = mem_address; | |
3072d30e | 1139 | |
627540ce | 1140 | /* Split the address into canonical BASE + OFFSET terms. */ |
1141 | address = canon_rtx (expanded_address); | |
3072d30e | 1142 | |
627540ce | 1143 | *offset = 0; |
3072d30e | 1144 | |
1ca59310 | 1145 | if (dump_file && (dump_flags & TDF_DETAILS)) |
627540ce | 1146 | { |
1147 | if (expanded) | |
1148 | { | |
1149 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1150 | print_inline_rtx (dump_file, expanded_address, 0); | |
1151 | fprintf (dump_file, "\n"); | |
1152 | } | |
3072d30e | 1153 | |
627540ce | 1154 | fprintf (dump_file, "\n after canon_rtx address: "); |
1155 | print_inline_rtx (dump_file, address, 0); | |
1156 | fprintf (dump_file, "\n"); | |
1157 | } | |
3072d30e | 1158 | |
627540ce | 1159 | if (GET_CODE (address) == CONST) |
1160 | address = XEXP (address, 0); | |
3072d30e | 1161 | |
370ad826 | 1162 | address = strip_offset_and_add (address, offset); |
3072d30e | 1163 | |
bd1a81f7 | 1164 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1165 | && const_or_frame_p (address)) | |
3072d30e | 1166 | { |
c2eab334 | 1167 | group_info *group = get_group_info (address); |
627540ce | 1168 | |
1ca59310 | 1169 | if (dump_file && (dump_flags & TDF_DETAILS)) |
370ad826 | 1170 | { |
1171 | fprintf (dump_file, " gid=%d offset=", group->id); | |
1172 | print_dec (*offset, dump_file); | |
1173 | fprintf (dump_file, "\n"); | |
1174 | } | |
627540ce | 1175 | *base = NULL; |
1176 | *group_id = group->id; | |
1177 | return true; | |
3072d30e | 1178 | } |
627540ce | 1179 | } |
1180 | ||
1f864115 | 1181 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
627540ce | 1182 | *group_id = -1; |
1183 | ||
1184 | if (*base == NULL) | |
1185 | { | |
1ca59310 | 1186 | if (dump_file && (dump_flags & TDF_DETAILS)) |
627540ce | 1187 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1188 | return false; | |
3072d30e | 1189 | } |
1ca59310 | 1190 | if (dump_file && (dump_flags & TDF_DETAILS)) |
370ad826 | 1191 | { |
1192 | fprintf (dump_file, " varying cselib base=%u:%u offset = ", | |
1193 | (*base)->uid, (*base)->hash); | |
1194 | print_dec (*offset, dump_file); | |
1195 | fprintf (dump_file, "\n"); | |
1196 | } | |
3072d30e | 1197 | return true; |
1198 | } | |
1199 | ||
1200 | ||
1201 | /* Clear the rhs field from the active_local_stores array. */ | |
1202 | ||
1203 | static void | |
1204 | clear_rhs_from_active_local_stores (void) | |
1205 | { | |
1206 | insn_info_t ptr = active_local_stores; | |
1207 | ||
1208 | while (ptr) | |
1209 | { | |
c2eab334 | 1210 | store_info *store_info = ptr->store_rec; |
3072d30e | 1211 | /* Skip the clobbers. */ |
1212 | while (!store_info->is_set) | |
1213 | store_info = store_info->next; | |
1214 | ||
1215 | store_info->rhs = NULL; | |
aa140b76 | 1216 | store_info->const_rhs = NULL; |
3072d30e | 1217 | |
1218 | ptr = ptr->next_local_store; | |
1219 | } | |
1220 | } | |
1221 | ||
1222 | ||
aa140b76 | 1223 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1224 | ||
1225 | static inline void | |
c2eab334 | 1226 | set_position_unneeded (store_info *s_info, int pos) |
aa140b76 | 1227 | { |
1228 | if (__builtin_expect (s_info->is_large, false)) | |
1229 | { | |
6ef9bbe0 | 1230 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1231 | s_info->positions_needed.large.count++; | |
aa140b76 | 1232 | } |
1233 | else | |
1234 | s_info->positions_needed.small_bitmask | |
edc19fd0 | 1235 | &= ~(HOST_WIDE_INT_1U << pos); |
aa140b76 | 1236 | } |
1237 | ||
1238 | /* Mark the whole store S_INFO as unneeded. */ | |
1239 | ||
1240 | static inline void | |
c2eab334 | 1241 | set_all_positions_unneeded (store_info *s_info) |
aa140b76 | 1242 | { |
1243 | if (__builtin_expect (s_info->is_large, false)) | |
1244 | { | |
370ad826 | 1245 | HOST_WIDE_INT width; |
1246 | if (s_info->width.is_constant (&width)) | |
1247 | { | |
1248 | bitmap_set_range (s_info->positions_needed.large.bmap, 0, width); | |
1249 | s_info->positions_needed.large.count = width; | |
1250 | } | |
1251 | else | |
1252 | { | |
1253 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1254 | s_info->positions_needed.large.count = 1; | |
1255 | } | |
aa140b76 | 1256 | } |
1257 | else | |
457daa4e | 1258 | s_info->positions_needed.small_bitmask = HOST_WIDE_INT_0U; |
aa140b76 | 1259 | } |
1260 | ||
1261 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1262 | ||
1263 | static inline bool | |
c2eab334 | 1264 | any_positions_needed_p (store_info *s_info) |
aa140b76 | 1265 | { |
1266 | if (__builtin_expect (s_info->is_large, false)) | |
370ad826 | 1267 | { |
1268 | HOST_WIDE_INT width; | |
1269 | if (s_info->width.is_constant (&width)) | |
1270 | { | |
1271 | gcc_checking_assert (s_info->positions_needed.large.bmap); | |
1272 | return s_info->positions_needed.large.count < width; | |
1273 | } | |
1274 | else | |
1275 | { | |
1276 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1277 | return s_info->positions_needed.large.count == 0; | |
1278 | } | |
1279 | } | |
aa140b76 | 1280 | else |
457daa4e | 1281 | return (s_info->positions_needed.small_bitmask != HOST_WIDE_INT_0U); |
aa140b76 | 1282 | } |
1283 | ||
1284 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
370ad826 | 1285 | store are known to be needed. */ |
aa140b76 | 1286 | |
1287 | static inline bool | |
370ad826 | 1288 | all_positions_needed_p (store_info *s_info, poly_int64 start, |
1289 | poly_int64 width) | |
aa140b76 | 1290 | { |
370ad826 | 1291 | gcc_assert (s_info->rhs); |
1292 | if (!s_info->width.is_constant ()) | |
1293 | { | |
1294 | gcc_assert (s_info->is_large | |
1295 | && !s_info->positions_needed.large.bmap); | |
1296 | return s_info->positions_needed.large.count == 0; | |
1297 | } | |
1298 | ||
1299 | /* Otherwise, if START and WIDTH are non-constant, we're asking about | |
1300 | a non-constant region of a constant-sized store. We can't say for | |
1301 | sure that all positions are needed. */ | |
1302 | HOST_WIDE_INT const_start, const_width; | |
1303 | if (!start.is_constant (&const_start) | |
1304 | || !width.is_constant (&const_width)) | |
1305 | return false; | |
1306 | ||
aa140b76 | 1307 | if (__builtin_expect (s_info->is_large, false)) |
1308 | { | |
370ad826 | 1309 | for (HOST_WIDE_INT i = const_start; i < const_start + const_width; ++i) |
1310 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, i)) | |
aa140b76 | 1311 | return false; |
1312 | return true; | |
1313 | } | |
1314 | else | |
1315 | { | |
370ad826 | 1316 | unsigned HOST_WIDE_INT mask |
1317 | = lowpart_bitmask (const_width) << const_start; | |
aa140b76 | 1318 | return (s_info->positions_needed.small_bitmask & mask) == mask; |
1319 | } | |
1320 | } | |
1321 | ||
1322 | ||
370ad826 | 1323 | static rtx get_stored_val (store_info *, machine_mode, poly_int64, |
1324 | poly_int64, basic_block, bool); | |
aa140b76 | 1325 | |
1326 | ||
3072d30e | 1327 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1328 | there is a candidate store, after adding it to the appropriate | |
1329 | local store group if so. */ | |
1330 | ||
1331 | static int | |
1332 | record_store (rtx body, bb_info_t bb_info) | |
1333 | { | |
82d2c88b | 1334 | rtx mem, rhs, const_rhs, mem_addr; |
370ad826 | 1335 | poly_int64 offset = 0; |
1336 | poly_int64 width = 0; | |
3072d30e | 1337 | insn_info_t insn_info = bb_info->last_insn; |
c2eab334 | 1338 | store_info *store_info = NULL; |
3072d30e | 1339 | int group_id; |
1340 | cselib_val *base = NULL; | |
aa140b76 | 1341 | insn_info_t ptr, last, redundant_reason; |
3072d30e | 1342 | bool store_is_unused; |
1343 | ||
1344 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1345 | return 0; | |
1346 | ||
aa140b76 | 1347 | mem = SET_DEST (body); |
1348 | ||
3072d30e | 1349 | /* If this is not used, then this cannot be used to keep the insn |
1350 | from being deleted. On the other hand, it does provide something | |
1351 | that can be used to prove that another store is dead. */ | |
1352 | store_is_unused | |
aa140b76 | 1353 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
3072d30e | 1354 | |
1355 | /* Check whether that value is a suitable memory location. */ | |
3072d30e | 1356 | if (!MEM_P (mem)) |
1357 | { | |
1358 | /* If the set or clobber is unused, then it does not effect our | |
1359 | ability to get rid of the entire insn. */ | |
1360 | if (!store_is_unused) | |
1361 | insn_info->cannot_delete = true; | |
1362 | return 0; | |
1363 | } | |
1364 | ||
1365 | /* At this point we know mem is a mem. */ | |
1366 | if (GET_MODE (mem) == BLKmode) | |
1367 | { | |
711f137f | 1368 | HOST_WIDE_INT const_size; |
3072d30e | 1369 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) |
1370 | { | |
1ca59310 | 1371 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1372 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1373 | add_wild_read (bb_info); | |
1374 | insn_info->cannot_delete = true; | |
aa140b76 | 1375 | return 0; |
3072d30e | 1376 | } |
aa140b76 | 1377 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1378 | as memset (addr, 0, 36); */ | |
5b2a69fa | 1379 | else if (!MEM_SIZE_KNOWN_P (mem) |
711f137f | 1380 | || maybe_le (MEM_SIZE (mem), 0) |
1381 | /* This is a limit on the bitmap size, which is only relevant | |
1382 | for constant-sized MEMs. */ | |
1383 | || (MEM_SIZE (mem).is_constant (&const_size) | |
1384 | && const_size > MAX_OFFSET) | |
aa140b76 | 1385 | || GET_CODE (body) != SET |
aa140b76 | 1386 | || !CONST_INT_P (SET_SRC (body))) |
3072d30e | 1387 | { |
aa140b76 | 1388 | if (!store_is_unused) |
1389 | { | |
1390 | /* If the set or clobber is unused, then it does not effect our | |
1391 | ability to get rid of the entire insn. */ | |
1392 | insn_info->cannot_delete = true; | |
1393 | clear_rhs_from_active_local_stores (); | |
1394 | } | |
1395 | return 0; | |
3072d30e | 1396 | } |
3072d30e | 1397 | } |
1398 | ||
1399 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1400 | if (MEM_VOLATILE_P (mem)) | |
aa140b76 | 1401 | insn_info->cannot_delete = true; |
3072d30e | 1402 | |
5a3f5d69 | 1403 | if (!canon_address (mem, &group_id, &offset, &base)) |
3072d30e | 1404 | { |
1405 | clear_rhs_from_active_local_stores (); | |
1406 | return 0; | |
1407 | } | |
1408 | ||
aa140b76 | 1409 | if (GET_MODE (mem) == BLKmode) |
5b2a69fa | 1410 | width = MEM_SIZE (mem); |
aa140b76 | 1411 | else |
e4209874 | 1412 | width = GET_MODE_SIZE (GET_MODE (mem)); |
3072d30e | 1413 | |
370ad826 | 1414 | if (!endpoint_representable_p (offset, width)) |
8f4e0f7b | 1415 | { |
1416 | clear_rhs_from_active_local_stores (); | |
1417 | return 0; | |
1418 | } | |
1419 | ||
269483bb | 1420 | if (known_eq (width, 0)) |
1421 | return 0; | |
1422 | ||
5a3f5d69 | 1423 | if (group_id >= 0) |
3072d30e | 1424 | { |
1425 | /* In the restrictive case where the base is a constant or the | |
1426 | frame pointer we can do global analysis. */ | |
48e1416a | 1427 | |
c2eab334 | 1428 | group_info *group |
f1f41a6c | 1429 | = rtx_group_vec[group_id]; |
b4a708fb | 1430 | tree expr = MEM_EXPR (mem); |
48e1416a | 1431 | |
55c5ac9f | 1432 | store_info = rtx_store_info_pool.allocate (); |
b4a708fb | 1433 | set_usage_bits (group, offset, width, expr); |
3072d30e | 1434 | |
1ca59310 | 1435 | if (dump_file && (dump_flags & TDF_DETAILS)) |
57bca8dc | 1436 | { |
1437 | fprintf (dump_file, " processing const base store gid=%d", | |
1438 | group_id); | |
1439 | print_range (dump_file, offset, width); | |
1440 | fprintf (dump_file, "\n"); | |
1441 | } | |
3072d30e | 1442 | } |
1443 | else | |
1444 | { | |
86e87ef6 | 1445 | if (may_be_sp_based_p (XEXP (mem, 0))) |
17e1318c | 1446 | insn_info->stack_pointer_based = true; |
3072d30e | 1447 | insn_info->contains_cselib_groups = true; |
17e1318c | 1448 | |
55c5ac9f | 1449 | store_info = cse_store_info_pool.allocate (); |
3072d30e | 1450 | group_id = -1; |
1451 | ||
1ca59310 | 1452 | if (dump_file && (dump_flags & TDF_DETAILS)) |
57bca8dc | 1453 | { |
1454 | fprintf (dump_file, " processing cselib store "); | |
1455 | print_range (dump_file, offset, width); | |
1456 | fprintf (dump_file, "\n"); | |
1457 | } | |
3072d30e | 1458 | } |
1459 | ||
aa140b76 | 1460 | const_rhs = rhs = NULL_RTX; |
1461 | if (GET_CODE (body) == SET | |
1462 | /* No place to keep the value after ra. */ | |
1463 | && !reload_completed | |
1464 | && (REG_P (SET_SRC (body)) | |
1465 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1466 | || CONSTANT_P (SET_SRC (body))) | |
1467 | && !MEM_VOLATILE_P (mem) | |
1468 | /* Sometimes the store and reload is used for truncation and | |
1469 | rounding. */ | |
1470 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1471 | { | |
1472 | rhs = SET_SRC (body); | |
1473 | if (CONSTANT_P (rhs)) | |
1474 | const_rhs = rhs; | |
1475 | else if (body == PATTERN (insn_info->insn)) | |
1476 | { | |
1477 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1478 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1479 | const_rhs = XEXP (tem, 0); | |
1480 | } | |
1481 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1482 | { | |
1483 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1484 | ||
1485 | if (tem && CONSTANT_P (tem)) | |
1486 | const_rhs = tem; | |
1487 | } | |
1488 | } | |
1489 | ||
3072d30e | 1490 | /* Check to see if this stores causes some other stores to be |
1491 | dead. */ | |
1492 | ptr = active_local_stores; | |
1493 | last = NULL; | |
aa140b76 | 1494 | redundant_reason = NULL; |
82d2c88b | 1495 | mem = canon_rtx (mem); |
5a3f5d69 | 1496 | |
1497 | if (group_id < 0) | |
1498 | mem_addr = base->val_rtx; | |
82d2c88b | 1499 | else |
1500 | { | |
5a3f5d69 | 1501 | group_info *group = rtx_group_vec[group_id]; |
1502 | mem_addr = group->canon_base_addr; | |
82d2c88b | 1503 | } |
370ad826 | 1504 | if (maybe_ne (offset, 0)) |
5a3f5d69 | 1505 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
3072d30e | 1506 | |
1507 | while (ptr) | |
1508 | { | |
1509 | insn_info_t next = ptr->next_local_store; | |
c2eab334 | 1510 | struct store_info *s_info = ptr->store_rec; |
9ce37fa7 | 1511 | bool del = true; |
3072d30e | 1512 | |
1513 | /* Skip the clobbers. We delete the active insn if this insn | |
6dfdc153 | 1514 | shadows the set. To have been put on the active list, it |
3072d30e | 1515 | has exactly on set. */ |
1516 | while (!s_info->is_set) | |
1517 | s_info = s_info->next; | |
1518 | ||
5a3f5d69 | 1519 | if (s_info->group_id == group_id && s_info->cse_base == base) |
3072d30e | 1520 | { |
1521 | HOST_WIDE_INT i; | |
1ca59310 | 1522 | if (dump_file && (dump_flags & TDF_DETAILS)) |
57bca8dc | 1523 | { |
1524 | fprintf (dump_file, " trying store in insn=%d gid=%d", | |
1525 | INSN_UID (ptr->insn), s_info->group_id); | |
1526 | print_range (dump_file, s_info->offset, s_info->width); | |
1527 | fprintf (dump_file, "\n"); | |
1528 | } | |
aa140b76 | 1529 | |
1530 | /* Even if PTR won't be eliminated as unneeded, if both | |
1531 | PTR and this insn store the same constant value, we might | |
1532 | eliminate this insn instead. */ | |
1533 | if (s_info->const_rhs | |
1534 | && const_rhs | |
57bca8dc | 1535 | && known_subrange_p (offset, width, |
1536 | s_info->offset, s_info->width) | |
1537 | && all_positions_needed_p (s_info, offset - s_info->offset, | |
be292237 | 1538 | width) |
1539 | /* We can only remove the later store if the earlier aliases | |
1540 | at least all accesses the later one. */ | |
1541 | && (MEM_ALIAS_SET (mem) == MEM_ALIAS_SET (s_info->mem) | |
1542 | || alias_set_subset_of (MEM_ALIAS_SET (mem), | |
1543 | MEM_ALIAS_SET (s_info->mem)))) | |
aa140b76 | 1544 | { |
1545 | if (GET_MODE (mem) == BLKmode) | |
1546 | { | |
1547 | if (GET_MODE (s_info->mem) == BLKmode | |
1548 | && s_info->const_rhs == const_rhs) | |
1549 | redundant_reason = ptr; | |
1550 | } | |
1551 | else if (s_info->const_rhs == const0_rtx | |
1552 | && const_rhs == const0_rtx) | |
1553 | redundant_reason = ptr; | |
1554 | else | |
1555 | { | |
1556 | rtx val; | |
1557 | start_sequence (); | |
57bca8dc | 1558 | val = get_stored_val (s_info, GET_MODE (mem), offset, width, |
aa140b76 | 1559 | BLOCK_FOR_INSN (insn_info->insn), |
1560 | true); | |
1561 | if (get_insns () != NULL) | |
1562 | val = NULL_RTX; | |
1563 | end_sequence (); | |
1564 | if (val && rtx_equal_p (val, const_rhs)) | |
1565 | redundant_reason = ptr; | |
1566 | } | |
1567 | } | |
1568 | ||
370ad826 | 1569 | HOST_WIDE_INT begin_unneeded, const_s_width, const_width; |
57bca8dc | 1570 | if (known_subrange_p (s_info->offset, s_info->width, offset, width)) |
1571 | /* The new store touches every byte that S_INFO does. */ | |
1572 | set_all_positions_unneeded (s_info); | |
370ad826 | 1573 | else if ((offset - s_info->offset).is_constant (&begin_unneeded) |
1574 | && s_info->width.is_constant (&const_s_width) | |
1575 | && width.is_constant (&const_width)) | |
57bca8dc | 1576 | { |
370ad826 | 1577 | HOST_WIDE_INT end_unneeded = begin_unneeded + const_width; |
57bca8dc | 1578 | begin_unneeded = MAX (begin_unneeded, 0); |
370ad826 | 1579 | end_unneeded = MIN (end_unneeded, const_s_width); |
57bca8dc | 1580 | for (i = begin_unneeded; i < end_unneeded; ++i) |
1581 | set_position_unneeded (s_info, i); | |
1582 | } | |
370ad826 | 1583 | else |
1584 | { | |
1585 | /* We don't know which parts of S_INFO are needed and | |
1586 | which aren't, so invalidate the RHS. */ | |
1587 | s_info->rhs = NULL; | |
1588 | s_info->const_rhs = NULL; | |
1589 | } | |
3072d30e | 1590 | } |
1591 | else if (s_info->rhs) | |
1592 | /* Need to see if it is possible for this store to overwrite | |
1593 | the value of store_info. If it is, set the rhs to NULL to | |
1594 | keep it from being used to remove a load. */ | |
1595 | { | |
68ab33a6 | 1596 | if (canon_output_dependence (s_info->mem, true, |
1597 | mem, GET_MODE (mem), | |
1598 | mem_addr)) | |
aa140b76 | 1599 | { |
1600 | s_info->rhs = NULL; | |
1601 | s_info->const_rhs = NULL; | |
1602 | } | |
3072d30e | 1603 | } |
82d2c88b | 1604 | |
3072d30e | 1605 | /* An insn can be deleted if every position of every one of |
1606 | its s_infos is zero. */ | |
c75be2fe | 1607 | if (any_positions_needed_p (s_info)) |
9ce37fa7 | 1608 | del = false; |
aa140b76 | 1609 | |
9ce37fa7 | 1610 | if (del) |
3072d30e | 1611 | { |
1612 | insn_info_t insn_to_delete = ptr; | |
48e1416a | 1613 | |
1242bee6 | 1614 | active_local_stores_len--; |
3072d30e | 1615 | if (last) |
1616 | last->next_local_store = ptr->next_local_store; | |
1617 | else | |
1618 | active_local_stores = ptr->next_local_store; | |
48e1416a | 1619 | |
c75be2fe | 1620 | if (!insn_to_delete->cannot_delete) |
1621 | delete_dead_store_insn (insn_to_delete); | |
3072d30e | 1622 | } |
1623 | else | |
1624 | last = ptr; | |
48e1416a | 1625 | |
3072d30e | 1626 | ptr = next; |
1627 | } | |
48e1416a | 1628 | |
3072d30e | 1629 | /* Finish filling in the store_info. */ |
1630 | store_info->next = insn_info->store_rec; | |
1631 | insn_info->store_rec = store_info; | |
82d2c88b | 1632 | store_info->mem = mem; |
82d2c88b | 1633 | store_info->mem_addr = mem_addr; |
3072d30e | 1634 | store_info->cse_base = base; |
370ad826 | 1635 | HOST_WIDE_INT const_width; |
1636 | if (!width.is_constant (&const_width)) | |
1637 | { | |
1638 | store_info->is_large = true; | |
1639 | store_info->positions_needed.large.count = 0; | |
1640 | store_info->positions_needed.large.bmap = NULL; | |
1641 | } | |
1642 | else if (const_width > HOST_BITS_PER_WIDE_INT) | |
aa140b76 | 1643 | { |
1644 | store_info->is_large = true; | |
1645 | store_info->positions_needed.large.count = 0; | |
4fb07d00 | 1646 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
aa140b76 | 1647 | } |
1648 | else | |
1649 | { | |
1650 | store_info->is_large = false; | |
370ad826 | 1651 | store_info->positions_needed.small_bitmask |
1652 | = lowpart_bitmask (const_width); | |
aa140b76 | 1653 | } |
3072d30e | 1654 | store_info->group_id = group_id; |
57bca8dc | 1655 | store_info->offset = offset; |
1656 | store_info->width = width; | |
3072d30e | 1657 | store_info->is_set = GET_CODE (body) == SET; |
aa140b76 | 1658 | store_info->rhs = rhs; |
1659 | store_info->const_rhs = const_rhs; | |
1660 | store_info->redundant_reason = redundant_reason; | |
3072d30e | 1661 | |
3072d30e | 1662 | /* If this is a clobber, we return 0. We will only be able to |
1663 | delete this insn if there is only one store USED store, but we | |
1664 | can use the clobber to delete other stores earlier. */ | |
1665 | return store_info->is_set ? 1 : 0; | |
1666 | } | |
1667 | ||
1668 | ||
1669 | static void | |
1670 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1671 | { | |
48e1416a | 1672 | fprintf (dump_file, "%s insn=%d %s\n", start, |
3072d30e | 1673 | INSN_UID (insn_info->insn), |
1674 | insn_info->store_rec ? "has store" : "naked"); | |
1675 | } | |
1676 | ||
1677 | ||
5c9051a4 | 1678 | /* If the modes are different and the value's source and target do not |
1679 | line up, we need to extract the value from lower part of the rhs of | |
1680 | the store, shift it, and then put it into a form that can be shoved | |
1681 | into the read_insn. This function generates a right SHIFT of a | |
1682 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1683 | shift sequence is returned or NULL if we failed to find a | |
1684 | shift. */ | |
1685 | ||
1686 | static rtx | |
370ad826 | 1687 | find_shift_sequence (poly_int64 access_size, |
c2eab334 | 1688 | store_info *store_info, |
3754d046 | 1689 | machine_mode read_mode, |
370ad826 | 1690 | poly_int64 shift, bool speed, bool require_cst) |
5c9051a4 | 1691 | { |
3754d046 | 1692 | machine_mode store_mode = GET_MODE (store_info->mem); |
1a5d4b27 | 1693 | scalar_int_mode new_mode; |
10d4de0e | 1694 | rtx read_reg = NULL; |
5c9051a4 | 1695 | |
1696 | /* Some machines like the x86 have shift insns for each size of | |
1697 | operand. Other machines like the ppc or the ia-64 may only have | |
1698 | shift insns that shift values within 32 or 64 bit registers. | |
1699 | This loop tries to find the smallest shift insn that will right | |
1700 | justify the value we want to read but is available in one insn on | |
1701 | the machine. */ | |
1702 | ||
1a5d4b27 | 1703 | opt_scalar_int_mode new_mode_iter; |
1704 | FOR_EACH_MODE_FROM (new_mode_iter, | |
1705 | smallest_int_mode_for_size (access_size * BITS_PER_UNIT)) | |
5c9051a4 | 1706 | { |
4cd001d5 | 1707 | rtx target, new_reg, new_lhs; |
1708 | rtx_insn *shift_seq, *insn; | |
171557e8 | 1709 | int cost; |
af97461e | 1710 | |
1a5d4b27 | 1711 | new_mode = new_mode_iter.require (); |
19a4dce4 | 1712 | if (GET_MODE_BITSIZE (new_mode) > BITS_PER_WORD) |
1713 | break; | |
1714 | ||
4ed4afb9 | 1715 | /* If a constant was stored into memory, try to simplify it here, |
1716 | otherwise the cost of the shift might preclude this optimization | |
1717 | e.g. at -Os, even when no actual shift will be needed. */ | |
aa140b76 | 1718 | if (store_info->const_rhs) |
4ed4afb9 | 1719 | { |
9edf7ea8 | 1720 | poly_uint64 byte = subreg_lowpart_offset (new_mode, store_mode); |
aa140b76 | 1721 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1722 | store_mode, byte); | |
4ed4afb9 | 1723 | if (ret && CONSTANT_P (ret)) |
1724 | { | |
bd39703a | 1725 | rtx shift_rtx = gen_int_shift_amount (new_mode, shift); |
4ed4afb9 | 1726 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, |
bd39703a | 1727 | ret, shift_rtx); |
4ed4afb9 | 1728 | if (ret && CONSTANT_P (ret)) |
1729 | { | |
1730 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1731 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1732 | if (ret && CONSTANT_P (ret) | |
5ae4887d | 1733 | && (set_src_cost (ret, read_mode, speed) |
1734 | <= COSTS_N_INSNS (1))) | |
4ed4afb9 | 1735 | return ret; |
1736 | } | |
1737 | } | |
1738 | } | |
1739 | ||
aa140b76 | 1740 | if (require_cst) |
1741 | return NULL_RTX; | |
1742 | ||
10d4de0e | 1743 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1744 | requires a real instruction. */ | |
eafbcd13 | 1745 | if (maybe_lt (GET_MODE_SIZE (new_mode), GET_MODE_SIZE (store_mode)) |
396f2130 | 1746 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
af97461e | 1747 | continue; |
1748 | ||
10d4de0e | 1749 | /* Also try a wider mode if the necessary punning is either not |
1750 | desirable or not possible. */ | |
1751 | if (!CONSTANT_P (store_info->rhs) | |
5f6dcf1a | 1752 | && !targetm.modes_tieable_p (new_mode, store_mode)) |
10d4de0e | 1753 | continue; |
10d4de0e | 1754 | |
af97461e | 1755 | new_reg = gen_reg_rtx (new_mode); |
5c9051a4 | 1756 | |
1757 | start_sequence (); | |
1758 | ||
1759 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1760 | of one dsp where the cost of these two was not the same. But | |
1761 | this really is a rare case anyway. */ | |
1762 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
bd39703a | 1763 | gen_int_shift_amount (new_mode, shift), |
1764 | new_reg, 1, OPTAB_DIRECT); | |
5c9051a4 | 1765 | |
597d5470 | 1766 | shift_seq = get_insns (); |
1767 | end_sequence (); | |
5c9051a4 | 1768 | |
597d5470 | 1769 | if (target != new_reg || shift_seq == NULL) |
1770 | continue; | |
1771 | ||
1772 | cost = 0; | |
1773 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1774 | if (INSN_P (insn)) | |
3bdb5a4d | 1775 | cost += insn_cost (insn, speed); |
597d5470 | 1776 | |
1777 | /* The computation up to here is essentially independent | |
1778 | of the arguments and could be precomputed. It may | |
1779 | not be worth doing so. We could precompute if | |
1780 | worthwhile or at least cache the results. The result | |
5133fd21 | 1781 | technically depends on both SHIFT and ACCESS_SIZE, |
1782 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
597d5470 | 1783 | |
1784 | if (cost > COSTS_N_INSNS (1)) | |
1785 | continue; | |
1786 | ||
171557e8 | 1787 | new_lhs = extract_low_bits (new_mode, store_mode, |
1788 | copy_rtx (store_info->rhs)); | |
1789 | if (new_lhs == NULL_RTX) | |
1790 | continue; | |
1791 | ||
597d5470 | 1792 | /* We found an acceptable shift. Generate a move to |
1793 | take the value from the store and put it into the | |
1794 | shift pseudo, then shift it, then generate another | |
1795 | move to put in into the target of the read. */ | |
10d4de0e | 1796 | emit_move_insn (new_reg, new_lhs); |
597d5470 | 1797 | emit_insn (shift_seq); |
10d4de0e | 1798 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
597d5470 | 1799 | break; |
5c9051a4 | 1800 | } |
1801 | ||
10d4de0e | 1802 | return read_reg; |
5c9051a4 | 1803 | } |
1804 | ||
1805 | ||
a1b0a968 | 1806 | /* Call back for note_stores to find the hard regs set or clobbered by |
1807 | insn. Data is a bitmap of the hardregs set so far. */ | |
1808 | ||
1809 | static void | |
1810 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1811 | { | |
1812 | bitmap regs_set = (bitmap) data; | |
1813 | ||
1814 | if (REG_P (x) | |
771d4616 | 1815 | && HARD_REGISTER_P (x)) |
0933f1d9 | 1816 | bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); |
a1b0a968 | 1817 | } |
1818 | ||
aa140b76 | 1819 | /* Helper function for replace_read and record_store. |
57bca8dc | 1820 | Attempt to return a value of mode READ_MODE stored in STORE_INFO, |
1821 | consisting of READ_WIDTH bytes starting from READ_OFFSET. Return NULL | |
aa140b76 | 1822 | if not successful. If REQUIRE_CST is true, return always constant. */ |
1823 | ||
1824 | static rtx | |
c2eab334 | 1825 | get_stored_val (store_info *store_info, machine_mode read_mode, |
370ad826 | 1826 | poly_int64 read_offset, poly_int64 read_width, |
aa140b76 | 1827 | basic_block bb, bool require_cst) |
1828 | { | |
3754d046 | 1829 | machine_mode store_mode = GET_MODE (store_info->mem); |
370ad826 | 1830 | poly_int64 gap; |
aa140b76 | 1831 | rtx read_reg; |
1832 | ||
1833 | /* To get here the read is within the boundaries of the write so | |
1834 | shift will never be negative. Start out with the shift being in | |
1835 | bytes. */ | |
1836 | if (store_mode == BLKmode) | |
57bca8dc | 1837 | gap = 0; |
aa140b76 | 1838 | else if (BYTES_BIG_ENDIAN) |
57bca8dc | 1839 | gap = ((store_info->offset + store_info->width) |
1840 | - (read_offset + read_width)); | |
aa140b76 | 1841 | else |
57bca8dc | 1842 | gap = read_offset - store_info->offset; |
aa140b76 | 1843 | |
ba0d8d7a | 1844 | if (gap.is_constant () && maybe_ne (gap, 0)) |
57bca8dc | 1845 | { |
370ad826 | 1846 | poly_int64 shift = gap * BITS_PER_UNIT; |
1847 | poly_int64 access_size = GET_MODE_SIZE (read_mode) + gap; | |
57bca8dc | 1848 | read_reg = find_shift_sequence (access_size, store_info, read_mode, |
1849 | shift, optimize_bb_for_speed_p (bb), | |
1850 | require_cst); | |
1851 | } | |
aa140b76 | 1852 | else if (store_mode == BLKmode) |
1853 | { | |
1854 | /* The store is a memset (addr, const_val, const_size). */ | |
1855 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
2cf1bb25 | 1856 | scalar_int_mode int_store_mode; |
1857 | if (!int_mode_for_mode (read_mode).exists (&int_store_mode)) | |
aa140b76 | 1858 | read_reg = NULL_RTX; |
1859 | else if (store_info->rhs == const0_rtx) | |
2cf1bb25 | 1860 | read_reg = extract_low_bits (read_mode, int_store_mode, const0_rtx); |
1861 | else if (GET_MODE_BITSIZE (int_store_mode) > HOST_BITS_PER_WIDE_INT | |
aa140b76 | 1862 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) |
1863 | read_reg = NULL_RTX; | |
1864 | else | |
1865 | { | |
1866 | unsigned HOST_WIDE_INT c | |
1867 | = INTVAL (store_info->rhs) | |
edc19fd0 | 1868 | & ((HOST_WIDE_INT_1 << BITS_PER_UNIT) - 1); |
aa140b76 | 1869 | int shift = BITS_PER_UNIT; |
1870 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1871 | { | |
1872 | c |= (c << shift); | |
1873 | shift <<= 1; | |
1874 | } | |
2cf1bb25 | 1875 | read_reg = gen_int_mode (c, int_store_mode); |
1876 | read_reg = extract_low_bits (read_mode, int_store_mode, read_reg); | |
aa140b76 | 1877 | } |
1878 | } | |
1879 | else if (store_info->const_rhs | |
1880 | && (require_cst | |
1881 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1882 | read_reg = extract_low_bits (read_mode, store_mode, | |
1883 | copy_rtx (store_info->const_rhs)); | |
1884 | else | |
1885 | read_reg = extract_low_bits (read_mode, store_mode, | |
1886 | copy_rtx (store_info->rhs)); | |
1887 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1888 | read_reg = NULL_RTX; | |
1889 | return read_reg; | |
1890 | } | |
a1b0a968 | 1891 | |
3072d30e | 1892 | /* Take a sequence of: |
1893 | A <- r1 | |
1894 | ... | |
1895 | ... <- A | |
1896 | ||
48e1416a | 1897 | and change it into |
3072d30e | 1898 | r2 <- r1 |
1899 | A <- r1 | |
1900 | ... | |
1901 | ... <- r2 | |
1902 | ||
5c9051a4 | 1903 | or |
1904 | ||
1905 | r3 <- extract (r1) | |
1906 | r3 <- r3 >> shift | |
1907 | r2 <- extract (r3) | |
1908 | ... <- r2 | |
1909 | ||
1910 | or | |
1911 | ||
1912 | r2 <- extract (r1) | |
1913 | ... <- r2 | |
1914 | ||
1915 | Depending on the alignment and the mode of the store and | |
1916 | subsequent load. | |
1917 | ||
1918 | ||
1919 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
3072d30e | 1920 | and READ_INSN are for the read. Return true if the replacement |
1921 | went ok. */ | |
1922 | ||
1923 | static bool | |
c2eab334 | 1924 | replace_read (store_info *store_info, insn_info_t store_insn, |
aa140b76 | 1925 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
1926 | bitmap regs_live) | |
3072d30e | 1927 | { |
3754d046 | 1928 | machine_mode store_mode = GET_MODE (store_info->mem); |
1929 | machine_mode read_mode = GET_MODE (read_info->mem); | |
4cd001d5 | 1930 | rtx_insn *insns, *this_insn; |
1931 | rtx read_reg; | |
aa140b76 | 1932 | basic_block bb; |
5c9051a4 | 1933 | |
3072d30e | 1934 | if (!dbg_cnt (dse)) |
1935 | return false; | |
1936 | ||
10d4de0e | 1937 | /* Create a sequence of instructions to set up the read register. |
1938 | This sequence goes immediately before the store and its result | |
1939 | is read by the load. | |
1940 | ||
1941 | We need to keep this in perspective. We are replacing a read | |
5c9051a4 | 1942 | with a sequence of insns, but the read will almost certainly be |
1943 | in cache, so it is not going to be an expensive one. Thus, we | |
1944 | are not willing to do a multi insn shift or worse a subroutine | |
1945 | call to get rid of the read. */ | |
1ca59310 | 1946 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 1947 | fprintf (dump_file, "trying to replace %smode load in insn %d" |
1948 | " from %smode store in insn %d\n", | |
1949 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
1950 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
1951 | start_sequence (); | |
aa140b76 | 1952 | bb = BLOCK_FOR_INSN (read_insn->insn); |
1953 | read_reg = get_stored_val (store_info, | |
57bca8dc | 1954 | read_mode, read_info->offset, read_info->width, |
aa140b76 | 1955 | bb, false); |
10d4de0e | 1956 | if (read_reg == NULL_RTX) |
5c9051a4 | 1957 | { |
10d4de0e | 1958 | end_sequence (); |
1ca59310 | 1959 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 1960 | fprintf (dump_file, " -- could not extract bits of stored value\n"); |
1961 | return false; | |
5c9051a4 | 1962 | } |
10d4de0e | 1963 | /* Force the value into a new register so that it won't be clobbered |
1964 | between the store and the load. */ | |
1965 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
1966 | insns = get_insns (); | |
1967 | end_sequence (); | |
5c9051a4 | 1968 | |
a1b0a968 | 1969 | if (insns != NULL_RTX) |
1970 | { | |
1971 | /* Now we have to scan the set of new instructions to see if the | |
1972 | sequence contains and sets of hardregs that happened to be | |
1973 | live at this point. For instance, this can happen if one of | |
1974 | the insns sets the CC and the CC happened to be live at that | |
1975 | point. This does occasionally happen, see PR 37922. */ | |
4fb07d00 | 1976 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
a1b0a968 | 1977 | |
1978 | for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
1979 | note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
48e1416a | 1980 | |
a1b0a968 | 1981 | bitmap_and_into (regs_set, regs_live); |
1982 | if (!bitmap_empty_p (regs_set)) | |
1983 | { | |
1ca59310 | 1984 | if (dump_file && (dump_flags & TDF_DETAILS)) |
a1b0a968 | 1985 | { |
48e1416a | 1986 | fprintf (dump_file, |
a1b0a968 | 1987 | "abandoning replacement because sequence clobbers live hardregs:"); |
1988 | df_print_regset (dump_file, regs_set); | |
1989 | } | |
48e1416a | 1990 | |
a1b0a968 | 1991 | BITMAP_FREE (regs_set); |
1992 | return false; | |
1993 | } | |
1994 | BITMAP_FREE (regs_set); | |
1995 | } | |
1996 | ||
5c9051a4 | 1997 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
3072d30e | 1998 | { |
c2eab334 | 1999 | deferred_change *change = deferred_change_pool.allocate (); |
48e1416a | 2000 | |
5c9051a4 | 2001 | /* Insert this right before the store insn where it will be safe |
2002 | from later insns that might change it before the read. */ | |
2003 | emit_insn_before (insns, store_insn->insn); | |
48e1416a | 2004 | |
5c9051a4 | 2005 | /* And now for the kludge part: cselib croaks if you just |
2006 | return at this point. There are two reasons for this: | |
48e1416a | 2007 | |
5c9051a4 | 2008 | 1) Cselib has an idea of how many pseudos there are and |
2009 | that does not include the new ones we just added. | |
48e1416a | 2010 | |
5c9051a4 | 2011 | 2) Cselib does not know about the move insn we added |
2012 | above the store_info, and there is no way to tell it | |
2013 | about it, because it has "moved on". | |
48e1416a | 2014 | |
5c9051a4 | 2015 | Problem (1) is fixable with a certain amount of engineering. |
2016 | Problem (2) is requires starting the bb from scratch. This | |
2017 | could be expensive. | |
48e1416a | 2018 | |
5c9051a4 | 2019 | So we are just going to have to lie. The move/extraction |
2020 | insns are not really an issue, cselib did not see them. But | |
2021 | the use of the new pseudo read_insn is a real problem because | |
2022 | cselib has not scanned this insn. The way that we solve this | |
2023 | problem is that we are just going to put the mem back for now | |
2024 | and when we are finished with the block, we undo this. We | |
2025 | keep a table of mems to get rid of. At the end of the basic | |
2026 | block we can put them back. */ | |
48e1416a | 2027 | |
5c9051a4 | 2028 | *loc = read_info->mem; |
55c5ac9f | 2029 | change->next = deferred_change_list; |
2030 | deferred_change_list = change; | |
2031 | change->loc = loc; | |
2032 | change->reg = read_reg; | |
48e1416a | 2033 | |
5c9051a4 | 2034 | /* Get rid of the read_info, from the point of view of the |
2035 | rest of dse, play like this read never happened. */ | |
2036 | read_insn->read_rec = read_info->next; | |
e16712b1 | 2037 | read_info_type_pool.remove (read_info); |
1ca59310 | 2038 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 2039 | { |
2040 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2041 | print_simple_rtl (dump_file, read_reg); | |
2042 | fprintf (dump_file, "\n"); | |
2043 | } | |
5c9051a4 | 2044 | return true; |
3072d30e | 2045 | } |
48e1416a | 2046 | else |
3072d30e | 2047 | { |
1ca59310 | 2048 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 2049 | { |
2050 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2051 | print_simple_rtl (dump_file, read_reg); | |
2052 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2053 | } | |
3072d30e | 2054 | return false; |
2055 | } | |
2056 | } | |
2057 | ||
ec1203cd | 2058 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2059 | be active. */ | |
3072d30e | 2060 | |
ec1203cd | 2061 | static void |
2062 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info) | |
3072d30e | 2063 | { |
82d2c88b | 2064 | rtx mem = *loc, mem_addr; |
3072d30e | 2065 | insn_info_t insn_info; |
370ad826 | 2066 | poly_int64 offset = 0; |
2067 | poly_int64 width = 0; | |
48e1416a | 2068 | cselib_val *base = NULL; |
3072d30e | 2069 | int group_id; |
2070 | read_info_t read_info; | |
2071 | ||
3072d30e | 2072 | insn_info = bb_info->last_insn; |
2073 | ||
2074 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
c7a7ba46 | 2075 | || MEM_VOLATILE_P (mem)) |
3072d30e | 2076 | { |
c7a7ba46 | 2077 | if (crtl->stack_protect_guard |
2078 | && (MEM_EXPR (mem) == crtl->stack_protect_guard | |
2079 | || (crtl->stack_protect_guard_decl | |
2080 | && MEM_EXPR (mem) == crtl->stack_protect_guard_decl)) | |
2081 | && MEM_VOLATILE_P (mem)) | |
2082 | { | |
2083 | /* This is either the stack protector canary on the stack, | |
2084 | which ought to be written by a MEM_VOLATILE_P store and | |
2085 | thus shouldn't be deleted and is read at the very end of | |
2086 | function, but shouldn't conflict with any other store. | |
2087 | Or it is __stack_chk_guard variable or TLS or whatever else | |
2088 | MEM holding the canary value, which really shouldn't be | |
2089 | ever modified in -fstack-protector* protected functions, | |
2090 | otherwise the prologue store wouldn't match the epilogue | |
2091 | check. */ | |
2092 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2093 | fprintf (dump_file, " stack protector canary read ignored.\n"); | |
2094 | insn_info->cannot_delete = true; | |
2095 | return; | |
2096 | } | |
2097 | ||
1ca59310 | 2098 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2099 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
2100 | add_wild_read (bb_info); | |
2101 | insn_info->cannot_delete = true; | |
ec1203cd | 2102 | return; |
3072d30e | 2103 | } |
2104 | ||
2105 | /* If it is reading readonly mem, then there can be no conflict with | |
2106 | another write. */ | |
2107 | if (MEM_READONLY_P (mem)) | |
ec1203cd | 2108 | return; |
3072d30e | 2109 | |
5a3f5d69 | 2110 | if (!canon_address (mem, &group_id, &offset, &base)) |
3072d30e | 2111 | { |
1ca59310 | 2112 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2113 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
2114 | add_wild_read (bb_info); | |
ec1203cd | 2115 | return; |
3072d30e | 2116 | } |
2117 | ||
2118 | if (GET_MODE (mem) == BLKmode) | |
2119 | width = -1; | |
2120 | else | |
2121 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2122 | ||
370ad826 | 2123 | if (!endpoint_representable_p (offset, known_eq (width, -1) ? 1 : width)) |
89a03a50 | 2124 | { |
8f4e0f7b | 2125 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2126 | fprintf (dump_file, " adding wild read, due to overflow.\n"); | |
2127 | add_wild_read (bb_info); | |
89a03a50 | 2128 | return; |
2129 | } | |
2130 | ||
e16712b1 | 2131 | read_info = read_info_type_pool.allocate (); |
3072d30e | 2132 | read_info->group_id = group_id; |
2133 | read_info->mem = mem; | |
57bca8dc | 2134 | read_info->offset = offset; |
2135 | read_info->width = width; | |
3072d30e | 2136 | read_info->next = insn_info->read_rec; |
2137 | insn_info->read_rec = read_info; | |
5a3f5d69 | 2138 | if (group_id < 0) |
2139 | mem_addr = base->val_rtx; | |
82d2c88b | 2140 | else |
2141 | { | |
5a3f5d69 | 2142 | group_info *group = rtx_group_vec[group_id]; |
2143 | mem_addr = group->canon_base_addr; | |
82d2c88b | 2144 | } |
370ad826 | 2145 | if (maybe_ne (offset, 0)) |
5a3f5d69 | 2146 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
3072d30e | 2147 | |
5a3f5d69 | 2148 | if (group_id >= 0) |
3072d30e | 2149 | { |
2150 | /* This is the restricted case where the base is a constant or | |
2151 | the frame pointer and offset is a constant. */ | |
2152 | insn_info_t i_ptr = active_local_stores; | |
2153 | insn_info_t last = NULL; | |
48e1416a | 2154 | |
1ca59310 | 2155 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2156 | { |
370ad826 | 2157 | if (!known_size_p (width)) |
3072d30e | 2158 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", |
2159 | group_id); | |
2160 | else | |
57bca8dc | 2161 | { |
2162 | fprintf (dump_file, " processing const load gid=%d", group_id); | |
2163 | print_range (dump_file, offset, width); | |
2164 | fprintf (dump_file, "\n"); | |
2165 | } | |
3072d30e | 2166 | } |
2167 | ||
2168 | while (i_ptr) | |
2169 | { | |
2170 | bool remove = false; | |
c2eab334 | 2171 | store_info *store_info = i_ptr->store_rec; |
48e1416a | 2172 | |
3072d30e | 2173 | /* Skip the clobbers. */ |
2174 | while (!store_info->is_set) | |
2175 | store_info = store_info->next; | |
48e1416a | 2176 | |
3072d30e | 2177 | /* There are three cases here. */ |
2178 | if (store_info->group_id < 0) | |
2179 | /* We have a cselib store followed by a read from a | |
2180 | const base. */ | |
48e1416a | 2181 | remove |
2182 | = canon_true_dependence (store_info->mem, | |
3072d30e | 2183 | GET_MODE (store_info->mem), |
2184 | store_info->mem_addr, | |
376a287d | 2185 | mem, mem_addr); |
48e1416a | 2186 | |
3072d30e | 2187 | else if (group_id == store_info->group_id) |
2188 | { | |
2189 | /* This is a block mode load. We may get lucky and | |
2190 | canon_true_dependence may save the day. */ | |
370ad826 | 2191 | if (!known_size_p (width)) |
48e1416a | 2192 | remove |
2193 | = canon_true_dependence (store_info->mem, | |
3072d30e | 2194 | GET_MODE (store_info->mem), |
2195 | store_info->mem_addr, | |
376a287d | 2196 | mem, mem_addr); |
48e1416a | 2197 | |
3072d30e | 2198 | /* If this read is just reading back something that we just |
2199 | stored, rewrite the read. */ | |
48e1416a | 2200 | else |
3072d30e | 2201 | { |
2202 | if (store_info->rhs | |
57bca8dc | 2203 | && known_subrange_p (offset, width, store_info->offset, |
2204 | store_info->width) | |
aa140b76 | 2205 | && all_positions_needed_p (store_info, |
57bca8dc | 2206 | offset - store_info->offset, |
aa140b76 | 2207 | width) |
2208 | && replace_read (store_info, i_ptr, read_info, | |
2209 | insn_info, loc, bb_info->regs_live)) | |
ec1203cd | 2210 | return; |
aa140b76 | 2211 | |
3072d30e | 2212 | /* The bases are the same, just see if the offsets |
57bca8dc | 2213 | could overlap. */ |
2214 | if (ranges_maybe_overlap_p (offset, width, | |
2215 | store_info->offset, | |
2216 | store_info->width)) | |
3072d30e | 2217 | remove = true; |
2218 | } | |
2219 | } | |
48e1416a | 2220 | |
2221 | /* else | |
3072d30e | 2222 | The else case that is missing here is that the |
2223 | bases are constant but different. There is nothing | |
2224 | to do here because there is no overlap. */ | |
48e1416a | 2225 | |
3072d30e | 2226 | if (remove) |
2227 | { | |
1ca59310 | 2228 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2229 | dump_insn_info ("removing from active", i_ptr); |
2230 | ||
1242bee6 | 2231 | active_local_stores_len--; |
3072d30e | 2232 | if (last) |
2233 | last->next_local_store = i_ptr->next_local_store; | |
2234 | else | |
2235 | active_local_stores = i_ptr->next_local_store; | |
2236 | } | |
2237 | else | |
2238 | last = i_ptr; | |
2239 | i_ptr = i_ptr->next_local_store; | |
2240 | } | |
2241 | } | |
48e1416a | 2242 | else |
3072d30e | 2243 | { |
2244 | insn_info_t i_ptr = active_local_stores; | |
2245 | insn_info_t last = NULL; | |
1ca59310 | 2246 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2247 | { |
2248 | fprintf (dump_file, " processing cselib load mem:"); | |
2249 | print_inline_rtx (dump_file, mem, 0); | |
2250 | fprintf (dump_file, "\n"); | |
2251 | } | |
2252 | ||
2253 | while (i_ptr) | |
2254 | { | |
2255 | bool remove = false; | |
c2eab334 | 2256 | store_info *store_info = i_ptr->store_rec; |
48e1416a | 2257 | |
1ca59310 | 2258 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2259 | fprintf (dump_file, " processing cselib load against insn %d\n", |
2260 | INSN_UID (i_ptr->insn)); | |
2261 | ||
2262 | /* Skip the clobbers. */ | |
2263 | while (!store_info->is_set) | |
2264 | store_info = store_info->next; | |
2265 | ||
2266 | /* If this read is just reading back something that we just | |
2267 | stored, rewrite the read. */ | |
2268 | if (store_info->rhs | |
2269 | && store_info->group_id == -1 | |
2270 | && store_info->cse_base == base | |
57bca8dc | 2271 | && known_subrange_p (offset, width, store_info->offset, |
2272 | store_info->width) | |
aa140b76 | 2273 | && all_positions_needed_p (store_info, |
57bca8dc | 2274 | offset - store_info->offset, width) |
aa140b76 | 2275 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, |
2276 | bb_info->regs_live)) | |
ec1203cd | 2277 | return; |
3072d30e | 2278 | |
5a3f5d69 | 2279 | remove = canon_true_dependence (store_info->mem, |
2280 | GET_MODE (store_info->mem), | |
2281 | store_info->mem_addr, | |
2282 | mem, mem_addr); | |
48e1416a | 2283 | |
3072d30e | 2284 | if (remove) |
2285 | { | |
1ca59310 | 2286 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2287 | dump_insn_info ("removing from active", i_ptr); |
48e1416a | 2288 | |
1242bee6 | 2289 | active_local_stores_len--; |
3072d30e | 2290 | if (last) |
2291 | last->next_local_store = i_ptr->next_local_store; | |
2292 | else | |
2293 | active_local_stores = i_ptr->next_local_store; | |
2294 | } | |
2295 | else | |
2296 | last = i_ptr; | |
2297 | i_ptr = i_ptr->next_local_store; | |
2298 | } | |
2299 | } | |
3072d30e | 2300 | } |
2301 | ||
ec1203cd | 2302 | /* A note_uses callback in which DATA points the INSN_INFO for |
3072d30e | 2303 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2304 | true for any part of *LOC. */ | |
2305 | ||
2306 | static void | |
2307 | check_mem_read_use (rtx *loc, void *data) | |
2308 | { | |
ec1203cd | 2309 | subrtx_ptr_iterator::array_type array; |
2310 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
2311 | { | |
2312 | rtx *loc = *iter; | |
2313 | if (MEM_P (*loc)) | |
2314 | check_mem_read_rtx (loc, (bb_info_t) data); | |
2315 | } | |
3072d30e | 2316 | } |
2317 | ||
aa140b76 | 2318 | |
2319 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2320 | So far it only handles arguments passed in registers. */ | |
2321 | ||
2322 | static bool | |
2323 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2324 | { | |
39cba157 | 2325 | CUMULATIVE_ARGS args_so_far_v; |
2326 | cumulative_args_t args_so_far; | |
aa140b76 | 2327 | tree arg; |
2328 | int idx; | |
2329 | ||
39cba157 | 2330 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2331 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
aa140b76 | 2332 | |
2333 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2334 | for (idx = 0; | |
2335 | arg != void_list_node && idx < nargs; | |
2336 | arg = TREE_CHAIN (arg), idx++) | |
2337 | { | |
8b449599 | 2338 | scalar_int_mode mode; |
f387af4f | 2339 | rtx reg, link, tmp; |
8b449599 | 2340 | |
2341 | if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg)), &mode)) | |
2342 | return false; | |
2343 | ||
39cba157 | 2344 | reg = targetm.calls.function_arg (args_so_far, mode, NULL_TREE, true); |
8b449599 | 2345 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode) |
aa140b76 | 2346 | return false; |
2347 | ||
2348 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2349 | link; | |
2350 | link = XEXP (link, 1)) | |
2351 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2352 | { | |
8b449599 | 2353 | scalar_int_mode arg_mode; |
aa140b76 | 2354 | args[idx] = XEXP (XEXP (link, 0), 0); |
2355 | if (REG_P (args[idx]) | |
2356 | && REGNO (args[idx]) == REGNO (reg) | |
2357 | && (GET_MODE (args[idx]) == mode | |
8b449599 | 2358 | || (is_int_mode (GET_MODE (args[idx]), &arg_mode) |
2359 | && (GET_MODE_SIZE (arg_mode) <= UNITS_PER_WORD) | |
2360 | && (GET_MODE_SIZE (arg_mode) > GET_MODE_SIZE (mode))))) | |
aa140b76 | 2361 | break; |
2362 | } | |
2363 | if (!link) | |
2364 | return false; | |
2365 | ||
2366 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2367 | if (GET_MODE (args[idx]) != mode) | |
2368 | { | |
2369 | if (!tmp || !CONST_INT_P (tmp)) | |
2370 | return false; | |
f62058c3 | 2371 | tmp = gen_int_mode (INTVAL (tmp), mode); |
aa140b76 | 2372 | } |
2373 | if (tmp) | |
2374 | args[idx] = tmp; | |
2375 | ||
39cba157 | 2376 | targetm.calls.function_arg_advance (args_so_far, mode, NULL_TREE, true); |
aa140b76 | 2377 | } |
2378 | if (arg != void_list_node || idx != nargs) | |
2379 | return false; | |
2380 | return true; | |
2381 | } | |
2382 | ||
5a9ecd4a | 2383 | /* Return a bitmap of the fixed registers contained in IN. */ |
2384 | ||
2385 | static bitmap | |
2386 | copy_fixed_regs (const_bitmap in) | |
2387 | { | |
2388 | bitmap ret; | |
2389 | ||
2390 | ret = ALLOC_REG_SET (NULL); | |
2391 | bitmap_and (ret, in, fixed_reg_set_regset); | |
2392 | return ret; | |
2393 | } | |
aa140b76 | 2394 | |
3072d30e | 2395 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2396 | if some part of it is not a candidate store and assigns to a | |
2397 | non-register target. */ | |
2398 | ||
2399 | static void | |
ebabb7a3 | 2400 | scan_insn (bb_info_t bb_info, rtx_insn *insn) |
3072d30e | 2401 | { |
2402 | rtx body; | |
e16712b1 | 2403 | insn_info_type *insn_info = insn_info_type_pool.allocate (); |
3072d30e | 2404 | int mems_found = 0; |
55c5ac9f | 2405 | memset (insn_info, 0, sizeof (struct insn_info_type)); |
3072d30e | 2406 | |
1ca59310 | 2407 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2408 | fprintf (dump_file, "\n**scanning insn=%d\n", |
2409 | INSN_UID (insn)); | |
2410 | ||
2411 | insn_info->prev_insn = bb_info->last_insn; | |
2412 | insn_info->insn = insn; | |
2413 | bb_info->last_insn = insn_info; | |
48e1416a | 2414 | |
9845d120 | 2415 | if (DEBUG_INSN_P (insn)) |
2416 | { | |
2417 | insn_info->cannot_delete = true; | |
2418 | return; | |
2419 | } | |
3072d30e | 2420 | |
3072d30e | 2421 | /* Look at all of the uses in the insn. */ |
2422 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2423 | ||
2424 | if (CALL_P (insn)) | |
2425 | { | |
aa140b76 | 2426 | bool const_call; |
61ffc71a | 2427 | rtx call, sym; |
aa140b76 | 2428 | tree memset_call = NULL_TREE; |
2429 | ||
3072d30e | 2430 | insn_info->cannot_delete = true; |
17e1318c | 2431 | |
3072d30e | 2432 | /* Const functions cannot do anything bad i.e. read memory, |
17e1318c | 2433 | however, they can read their parameters which may have |
aa140b76 | 2434 | been pushed onto the stack. |
2435 | memset and bzero don't read memory either. */ | |
2436 | const_call = RTL_CONST_CALL_P (insn); | |
61ffc71a | 2437 | if (!const_call |
2438 | && (call = get_call_rtx_from (insn)) | |
2439 | && (sym = XEXP (XEXP (call, 0), 0)) | |
2440 | && GET_CODE (sym) == SYMBOL_REF | |
2441 | && SYMBOL_REF_DECL (sym) | |
2442 | && TREE_CODE (SYMBOL_REF_DECL (sym)) == FUNCTION_DECL | |
a0e9bfbb | 2443 | && fndecl_built_in_p (SYMBOL_REF_DECL (sym), BUILT_IN_MEMSET)) |
61ffc71a | 2444 | memset_call = SYMBOL_REF_DECL (sym); |
2445 | ||
aa140b76 | 2446 | if (const_call || memset_call) |
3072d30e | 2447 | { |
2448 | insn_info_t i_ptr = active_local_stores; | |
2449 | insn_info_t last = NULL; | |
2450 | ||
1ca59310 | 2451 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 2452 | fprintf (dump_file, "%s call %d\n", |
2453 | const_call ? "const" : "memset", INSN_UID (insn)); | |
3072d30e | 2454 | |
16bf64db | 2455 | /* See the head comment of the frame_read field. */ |
17853422 | 2456 | if (reload_completed |
2457 | /* Tail calls are storing their arguments using | |
2458 | arg pointer. If it is a frame pointer on the target, | |
2459 | even before reload we need to kill frame pointer based | |
2460 | stores. */ | |
2461 | || (SIBLING_CALL_P (insn) | |
2462 | && HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
16bf64db | 2463 | insn_info->frame_read = true; |
2464 | ||
2465 | /* Loop over the active stores and remove those which are | |
2466 | killed by the const function call. */ | |
3072d30e | 2467 | while (i_ptr) |
2468 | { | |
16bf64db | 2469 | bool remove_store = false; |
2470 | ||
2471 | /* The stack pointer based stores are always killed. */ | |
17e1318c | 2472 | if (i_ptr->stack_pointer_based) |
16bf64db | 2473 | remove_store = true; |
2474 | ||
2475 | /* If the frame is read, the frame related stores are killed. */ | |
2476 | else if (insn_info->frame_read) | |
2477 | { | |
c2eab334 | 2478 | store_info *store_info = i_ptr->store_rec; |
16bf64db | 2479 | |
2480 | /* Skip the clobbers. */ | |
2481 | while (!store_info->is_set) | |
2482 | store_info = store_info->next; | |
2483 | ||
2484 | if (store_info->group_id >= 0 | |
f1f41a6c | 2485 | && rtx_group_vec[store_info->group_id]->frame_related) |
16bf64db | 2486 | remove_store = true; |
2487 | } | |
2488 | ||
2489 | if (remove_store) | |
3072d30e | 2490 | { |
1ca59310 | 2491 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2492 | dump_insn_info ("removing from active", i_ptr); |
48e1416a | 2493 | |
1242bee6 | 2494 | active_local_stores_len--; |
3072d30e | 2495 | if (last) |
2496 | last->next_local_store = i_ptr->next_local_store; | |
2497 | else | |
2498 | active_local_stores = i_ptr->next_local_store; | |
2499 | } | |
2500 | else | |
2501 | last = i_ptr; | |
16bf64db | 2502 | |
3072d30e | 2503 | i_ptr = i_ptr->next_local_store; |
2504 | } | |
aa140b76 | 2505 | |
2506 | if (memset_call) | |
2507 | { | |
2508 | rtx args[3]; | |
2509 | if (get_call_args (insn, memset_call, args, 3) | |
2510 | && CONST_INT_P (args[1]) | |
2511 | && CONST_INT_P (args[2]) | |
2512 | && INTVAL (args[2]) > 0) | |
2513 | { | |
2514 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
5b2a69fa | 2515 | set_mem_size (mem, INTVAL (args[2])); |
d1f9b275 | 2516 | body = gen_rtx_SET (mem, args[1]); |
aa140b76 | 2517 | mems_found += record_store (body, bb_info); |
1ca59310 | 2518 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 2519 | fprintf (dump_file, "handling memset as BLKmode store\n"); |
2520 | if (mems_found == 1) | |
2521 | { | |
1242bee6 | 2522 | if (active_local_stores_len++ |
2523 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2524 | { | |
2525 | active_local_stores_len = 1; | |
2526 | active_local_stores = NULL; | |
2527 | } | |
5a9ecd4a | 2528 | insn_info->fixed_regs_live |
2529 | = copy_fixed_regs (bb_info->regs_live); | |
aa140b76 | 2530 | insn_info->next_local_store = active_local_stores; |
2531 | active_local_stores = insn_info; | |
2532 | } | |
2533 | } | |
9441fa4e | 2534 | else |
2535 | clear_rhs_from_active_local_stores (); | |
aa140b76 | 2536 | } |
3072d30e | 2537 | } |
17853422 | 2538 | else if (SIBLING_CALL_P (insn) && reload_completed) |
2539 | /* Arguments for a sibling call that are pushed to memory are passed | |
2540 | using the incoming argument pointer of the current function. After | |
2541 | reload that might be (and likely is) frame pointer based. */ | |
2542 | add_wild_read (bb_info); | |
17e1318c | 2543 | else |
b4a708fb | 2544 | /* Every other call, including pure functions, may read any memory |
2545 | that is not relative to the frame. */ | |
2546 | add_non_frame_wild_read (bb_info); | |
17e1318c | 2547 | |
3072d30e | 2548 | return; |
2549 | } | |
2550 | ||
2551 | /* Assuming that there are sets in these insns, we cannot delete | |
2552 | them. */ | |
2553 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
4aafe72f | 2554 | || volatile_refs_p (PATTERN (insn)) |
bc0dfc8d | 2555 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
3072d30e | 2556 | || (RTX_FRAME_RELATED_P (insn)) |
2557 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2558 | insn_info->cannot_delete = true; | |
48e1416a | 2559 | |
3072d30e | 2560 | body = PATTERN (insn); |
2561 | if (GET_CODE (body) == PARALLEL) | |
2562 | { | |
2563 | int i; | |
2564 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2565 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2566 | } | |
2567 | else | |
2568 | mems_found += record_store (body, bb_info); | |
2569 | ||
1ca59310 | 2570 | if (dump_file && (dump_flags & TDF_DETAILS)) |
48e1416a | 2571 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
3072d30e | 2572 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2573 | ||
aa140b76 | 2574 | /* If we found some sets of mems, add it into the active_local_stores so |
2575 | that it can be locally deleted if found dead or used for | |
2576 | replace_read and redundant constant store elimination. Otherwise mark | |
2577 | it as cannot delete. This simplifies the processing later. */ | |
2578 | if (mems_found == 1) | |
3072d30e | 2579 | { |
1242bee6 | 2580 | if (active_local_stores_len++ |
2581 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2582 | { | |
2583 | active_local_stores_len = 1; | |
2584 | active_local_stores = NULL; | |
2585 | } | |
5a9ecd4a | 2586 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
3072d30e | 2587 | insn_info->next_local_store = active_local_stores; |
2588 | active_local_stores = insn_info; | |
2589 | } | |
2590 | else | |
2591 | insn_info->cannot_delete = true; | |
2592 | } | |
2593 | ||
2594 | ||
2595 | /* Remove BASE from the set of active_local_stores. This is a | |
2596 | callback from cselib that is used to get rid of the stores in | |
2597 | active_local_stores. */ | |
2598 | ||
2599 | static void | |
2600 | remove_useless_values (cselib_val *base) | |
2601 | { | |
2602 | insn_info_t insn_info = active_local_stores; | |
2603 | insn_info_t last = NULL; | |
2604 | ||
2605 | while (insn_info) | |
2606 | { | |
c2eab334 | 2607 | store_info *store_info = insn_info->store_rec; |
9ce37fa7 | 2608 | bool del = false; |
3072d30e | 2609 | |
2610 | /* If ANY of the store_infos match the cselib group that is | |
f4d3c071 | 2611 | being deleted, then the insn cannot be deleted. */ |
3072d30e | 2612 | while (store_info) |
2613 | { | |
48e1416a | 2614 | if ((store_info->group_id == -1) |
3072d30e | 2615 | && (store_info->cse_base == base)) |
2616 | { | |
9ce37fa7 | 2617 | del = true; |
3072d30e | 2618 | break; |
2619 | } | |
2620 | store_info = store_info->next; | |
2621 | } | |
2622 | ||
9ce37fa7 | 2623 | if (del) |
3072d30e | 2624 | { |
1242bee6 | 2625 | active_local_stores_len--; |
3072d30e | 2626 | if (last) |
2627 | last->next_local_store = insn_info->next_local_store; | |
2628 | else | |
2629 | active_local_stores = insn_info->next_local_store; | |
2630 | free_store_info (insn_info); | |
2631 | } | |
2632 | else | |
2633 | last = insn_info; | |
48e1416a | 2634 | |
3072d30e | 2635 | insn_info = insn_info->next_local_store; |
2636 | } | |
2637 | } | |
2638 | ||
2639 | ||
2640 | /* Do all of step 1. */ | |
2641 | ||
2642 | static void | |
2643 | dse_step1 (void) | |
2644 | { | |
2645 | basic_block bb; | |
4fb07d00 | 2646 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
48e1416a | 2647 | |
35af0188 | 2648 | cselib_init (0); |
3072d30e | 2649 | all_blocks = BITMAP_ALLOC (NULL); |
2650 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2651 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2652 | ||
ed7d889a | 2653 | FOR_ALL_BB_FN (bb, cfun) |
3072d30e | 2654 | { |
2655 | insn_info_t ptr; | |
e16712b1 | 2656 | bb_info_t bb_info = dse_bb_info_type_pool.allocate (); |
3072d30e | 2657 | |
55c5ac9f | 2658 | memset (bb_info, 0, sizeof (dse_bb_info_type)); |
3072d30e | 2659 | bitmap_set_bit (all_blocks, bb->index); |
a1b0a968 | 2660 | bb_info->regs_live = regs_live; |
2661 | ||
2662 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2663 | df_simulate_initialize_forwards (bb, regs_live); | |
3072d30e | 2664 | |
2665 | bb_table[bb->index] = bb_info; | |
2666 | cselib_discard_hook = remove_useless_values; | |
2667 | ||
2668 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2669 | { | |
ebabb7a3 | 2670 | rtx_insn *insn; |
3072d30e | 2671 | |
3072d30e | 2672 | active_local_stores = NULL; |
1242bee6 | 2673 | active_local_stores_len = 0; |
3072d30e | 2674 | cselib_clear_table (); |
48e1416a | 2675 | |
3072d30e | 2676 | /* Scan the insns. */ |
2677 | FOR_BB_INSNS (bb, insn) | |
2678 | { | |
2679 | if (INSN_P (insn)) | |
2680 | scan_insn (bb_info, insn); | |
2681 | cselib_process_insn (insn); | |
a1b0a968 | 2682 | if (INSN_P (insn)) |
2683 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
3072d30e | 2684 | } |
48e1416a | 2685 | |
3072d30e | 2686 | /* This is something of a hack, because the global algorithm |
2687 | is supposed to take care of the case where stores go dead | |
2688 | at the end of the function. However, the global | |
2689 | algorithm must take a more conservative view of block | |
2690 | mode reads than the local alg does. So to get the case | |
2691 | where you have a store to the frame followed by a non | |
bef304b8 | 2692 | overlapping block more read, we look at the active local |
3072d30e | 2693 | stores at the end of the function and delete all of the |
2694 | frame and spill based ones. */ | |
2695 | if (stores_off_frame_dead_at_return | |
2696 | && (EDGE_COUNT (bb->succs) == 0 | |
2697 | || (single_succ_p (bb) | |
34154e27 | 2698 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
18d50ae6 | 2699 | && ! crtl->calls_eh_return))) |
3072d30e | 2700 | { |
2701 | insn_info_t i_ptr = active_local_stores; | |
2702 | while (i_ptr) | |
2703 | { | |
c2eab334 | 2704 | store_info *store_info = i_ptr->store_rec; |
3072d30e | 2705 | |
2706 | /* Skip the clobbers. */ | |
2707 | while (!store_info->is_set) | |
2708 | store_info = store_info->next; | |
5a3f5d69 | 2709 | if (store_info->group_id >= 0) |
2710 | { | |
2711 | group_info *group = rtx_group_vec[store_info->group_id]; | |
2712 | if (group->frame_related && !i_ptr->cannot_delete) | |
2713 | delete_dead_store_insn (i_ptr); | |
2714 | } | |
3072d30e | 2715 | |
2716 | i_ptr = i_ptr->next_local_store; | |
2717 | } | |
2718 | } | |
2719 | ||
2720 | /* Get rid of the loads that were discovered in | |
2721 | replace_read. Cselib is finished with this block. */ | |
2722 | while (deferred_change_list) | |
2723 | { | |
c2eab334 | 2724 | deferred_change *next = deferred_change_list->next; |
3072d30e | 2725 | |
2726 | /* There is no reason to validate this change. That was | |
2727 | done earlier. */ | |
2728 | *deferred_change_list->loc = deferred_change_list->reg; | |
e16712b1 | 2729 | deferred_change_pool.remove (deferred_change_list); |
3072d30e | 2730 | deferred_change_list = next; |
2731 | } | |
2732 | ||
2733 | /* Get rid of all of the cselib based store_infos in this | |
2734 | block and mark the containing insns as not being | |
2735 | deletable. */ | |
2736 | ptr = bb_info->last_insn; | |
2737 | while (ptr) | |
2738 | { | |
2739 | if (ptr->contains_cselib_groups) | |
aa140b76 | 2740 | { |
c2eab334 | 2741 | store_info *s_info = ptr->store_rec; |
aa140b76 | 2742 | while (s_info && !s_info->is_set) |
2743 | s_info = s_info->next; | |
2744 | if (s_info | |
2745 | && s_info->redundant_reason | |
2746 | && s_info->redundant_reason->insn | |
2747 | && !ptr->cannot_delete) | |
2748 | { | |
1ca59310 | 2749 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 2750 | fprintf (dump_file, "Locally deleting insn %d " |
2751 | "because insn %d stores the " | |
2752 | "same value and couldn't be " | |
2753 | "eliminated\n", | |
2754 | INSN_UID (ptr->insn), | |
2755 | INSN_UID (s_info->redundant_reason->insn)); | |
2756 | delete_dead_store_insn (ptr); | |
2757 | } | |
aa140b76 | 2758 | free_store_info (ptr); |
2759 | } | |
2760 | else | |
2761 | { | |
c2eab334 | 2762 | store_info *s_info; |
aa140b76 | 2763 | |
2764 | /* Free at least positions_needed bitmaps. */ | |
2765 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2766 | if (s_info->is_large) | |
2767 | { | |
843bd2fa | 2768 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
aa140b76 | 2769 | s_info->is_large = false; |
2770 | } | |
2771 | } | |
3072d30e | 2772 | ptr = ptr->prev_insn; |
2773 | } | |
2774 | ||
55c5ac9f | 2775 | cse_store_info_pool.release (); |
3072d30e | 2776 | } |
a1b0a968 | 2777 | bb_info->regs_live = NULL; |
3072d30e | 2778 | } |
2779 | ||
a1b0a968 | 2780 | BITMAP_FREE (regs_live); |
3072d30e | 2781 | cselib_finish (); |
c1f445d2 | 2782 | rtx_group_table->empty (); |
3072d30e | 2783 | } |
2784 | ||
2785 | \f | |
2786 | /*---------------------------------------------------------------------------- | |
2787 | Second step. | |
2788 | ||
2789 | Assign each byte position in the stores that we are going to | |
2790 | analyze globally to a position in the bitmaps. Returns true if | |
6dfdc153 | 2791 | there are any bit positions assigned. |
3072d30e | 2792 | ----------------------------------------------------------------------------*/ |
2793 | ||
2794 | static void | |
2795 | dse_step2_init (void) | |
2796 | { | |
2797 | unsigned int i; | |
c2eab334 | 2798 | group_info *group; |
3072d30e | 2799 | |
f1f41a6c | 2800 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 2801 | { |
2802 | /* For all non stack related bases, we only consider a store to | |
2803 | be deletable if there are two or more stores for that | |
2804 | position. This is because it takes one store to make the | |
2805 | other store redundant. However, for the stores that are | |
2806 | stack related, we consider them if there is only one store | |
2807 | for the position. We do this because the stack related | |
2808 | stores can be deleted if their is no read between them and | |
2809 | the end of the function. | |
48e1416a | 2810 | |
3072d30e | 2811 | To make this work in the current framework, we take the stack |
2812 | related bases add all of the bits from store1 into store2. | |
2813 | This has the effect of making the eligible even if there is | |
2814 | only one store. */ | |
2815 | ||
2816 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2817 | { | |
2818 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2819 | bitmap_ior_into (group->store2_p, group->store1_p); | |
1ca59310 | 2820 | if (dump_file && (dump_flags & TDF_DETAILS)) |
48e1416a | 2821 | fprintf (dump_file, "group %d is frame related ", i); |
3072d30e | 2822 | } |
2823 | ||
2824 | group->offset_map_size_n++; | |
4fb07d00 | 2825 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2826 | group->offset_map_size_n); | |
3072d30e | 2827 | group->offset_map_size_p++; |
4fb07d00 | 2828 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2829 | group->offset_map_size_p); | |
3072d30e | 2830 | group->process_globally = false; |
1ca59310 | 2831 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2832 | { |
48e1416a | 2833 | fprintf (dump_file, "group %d(%d+%d): ", i, |
3072d30e | 2834 | (int)bitmap_count_bits (group->store2_n), |
2835 | (int)bitmap_count_bits (group->store2_p)); | |
2836 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2837 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2838 | } | |
2839 | } | |
2840 | } | |
2841 | ||
2842 | ||
5a3f5d69 | 2843 | /* Init the offset tables. */ |
3072d30e | 2844 | |
2845 | static bool | |
5a3f5d69 | 2846 | dse_step2 (void) |
3072d30e | 2847 | { |
2848 | unsigned int i; | |
c2eab334 | 2849 | group_info *group; |
3072d30e | 2850 | /* Position 0 is unused because 0 is used in the maps to mean |
2851 | unused. */ | |
2852 | current_position = 1; | |
f1f41a6c | 2853 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 2854 | { |
2855 | bitmap_iterator bi; | |
2856 | unsigned int j; | |
2857 | ||
9af5ce0c | 2858 | memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
2859 | memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
3072d30e | 2860 | bitmap_clear (group->group_kill); |
2861 | ||
2862 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2863 | { | |
2864 | bitmap_set_bit (group->group_kill, current_position); | |
b4a708fb | 2865 | if (bitmap_bit_p (group->escaped_n, j)) |
2866 | bitmap_set_bit (kill_on_calls, current_position); | |
3072d30e | 2867 | group->offset_map_n[j] = current_position++; |
2868 | group->process_globally = true; | |
2869 | } | |
2870 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2871 | { | |
48e1416a | 2872 | bitmap_set_bit (group->group_kill, current_position); |
b4a708fb | 2873 | if (bitmap_bit_p (group->escaped_p, j)) |
2874 | bitmap_set_bit (kill_on_calls, current_position); | |
3072d30e | 2875 | group->offset_map_p[j] = current_position++; |
2876 | group->process_globally = true; | |
2877 | } | |
2878 | } | |
2879 | return current_position != 1; | |
2880 | } | |
2881 | ||
2882 | ||
3072d30e | 2883 | \f |
2884 | /*---------------------------------------------------------------------------- | |
2885 | Third step. | |
48e1416a | 2886 | |
3072d30e | 2887 | Build the bit vectors for the transfer functions. |
2888 | ----------------------------------------------------------------------------*/ | |
2889 | ||
2890 | ||
3072d30e | 2891 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
2892 | there, return 0. */ | |
2893 | ||
2894 | static int | |
c2eab334 | 2895 | get_bitmap_index (group_info *group_info, HOST_WIDE_INT offset) |
3072d30e | 2896 | { |
2897 | if (offset < 0) | |
2898 | { | |
2899 | HOST_WIDE_INT offset_p = -offset; | |
2900 | if (offset_p >= group_info->offset_map_size_n) | |
2901 | return 0; | |
2902 | return group_info->offset_map_n[offset_p]; | |
2903 | } | |
2904 | else | |
2905 | { | |
2906 | if (offset >= group_info->offset_map_size_p) | |
2907 | return 0; | |
2908 | return group_info->offset_map_p[offset]; | |
2909 | } | |
2910 | } | |
2911 | ||
2912 | ||
2913 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
2914 | may be NULL. */ | |
2915 | ||
48e1416a | 2916 | static void |
5a3f5d69 | 2917 | scan_stores (store_info *store_info, bitmap gen, bitmap kill) |
3072d30e | 2918 | { |
2919 | while (store_info) | |
2920 | { | |
370ad826 | 2921 | HOST_WIDE_INT i, offset, width; |
c2eab334 | 2922 | group_info *group_info |
f1f41a6c | 2923 | = rtx_group_vec[store_info->group_id]; |
370ad826 | 2924 | /* We can (conservatively) ignore stores whose bounds aren't known; |
2925 | they simply don't generate new global dse opportunities. */ | |
2926 | if (group_info->process_globally | |
2927 | && store_info->offset.is_constant (&offset) | |
2928 | && store_info->width.is_constant (&width)) | |
57bca8dc | 2929 | { |
370ad826 | 2930 | HOST_WIDE_INT end = offset + width; |
2931 | for (i = offset; i < end; i++) | |
57bca8dc | 2932 | { |
2933 | int index = get_bitmap_index (group_info, i); | |
2934 | if (index != 0) | |
2935 | { | |
2936 | bitmap_set_bit (gen, index); | |
2937 | if (kill) | |
2938 | bitmap_clear_bit (kill, index); | |
2939 | } | |
2940 | } | |
2941 | } | |
3072d30e | 2942 | store_info = store_info->next; |
2943 | } | |
2944 | } | |
2945 | ||
2946 | ||
3072d30e | 2947 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL |
2948 | may be NULL. */ | |
2949 | ||
2950 | static void | |
5a3f5d69 | 2951 | scan_reads (insn_info_t insn_info, bitmap gen, bitmap kill) |
3072d30e | 2952 | { |
2953 | read_info_t read_info = insn_info->read_rec; | |
2954 | int i; | |
c2eab334 | 2955 | group_info *group; |
3072d30e | 2956 | |
16bf64db | 2957 | /* If this insn reads the frame, kill all the frame related stores. */ |
2958 | if (insn_info->frame_read) | |
2959 | { | |
f1f41a6c | 2960 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
16bf64db | 2961 | if (group->process_globally && group->frame_related) |
2962 | { | |
2963 | if (kill) | |
2964 | bitmap_ior_into (kill, group->group_kill); | |
48e1416a | 2965 | bitmap_and_compl_into (gen, group->group_kill); |
16bf64db | 2966 | } |
2967 | } | |
b4a708fb | 2968 | if (insn_info->non_frame_wild_read) |
2969 | { | |
2970 | /* Kill all non-frame related stores. Kill all stores of variables that | |
2971 | escape. */ | |
2972 | if (kill) | |
2973 | bitmap_ior_into (kill, kill_on_calls); | |
2974 | bitmap_and_compl_into (gen, kill_on_calls); | |
f1f41a6c | 2975 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
b4a708fb | 2976 | if (group->process_globally && !group->frame_related) |
2977 | { | |
2978 | if (kill) | |
2979 | bitmap_ior_into (kill, group->group_kill); | |
2980 | bitmap_and_compl_into (gen, group->group_kill); | |
2981 | } | |
2982 | } | |
3072d30e | 2983 | while (read_info) |
2984 | { | |
f1f41a6c | 2985 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 2986 | { |
2987 | if (group->process_globally) | |
2988 | { | |
2989 | if (i == read_info->group_id) | |
2990 | { | |
370ad826 | 2991 | HOST_WIDE_INT offset, width; |
2992 | /* Reads with non-constant size kill all DSE opportunities | |
2993 | in the group. */ | |
2994 | if (!read_info->offset.is_constant (&offset) | |
2995 | || !read_info->width.is_constant (&width) | |
2996 | || !known_size_p (width)) | |
3072d30e | 2997 | { |
57bca8dc | 2998 | /* Handle block mode reads. */ |
3072d30e | 2999 | if (kill) |
3000 | bitmap_ior_into (kill, group->group_kill); | |
3001 | bitmap_and_compl_into (gen, group->group_kill); | |
3002 | } | |
3003 | else | |
3004 | { | |
3005 | /* The groups are the same, just process the | |
3006 | offsets. */ | |
3007 | HOST_WIDE_INT j; | |
370ad826 | 3008 | HOST_WIDE_INT end = offset + width; |
3009 | for (j = offset; j < end; j++) | |
3072d30e | 3010 | { |
3011 | int index = get_bitmap_index (group, j); | |
3012 | if (index != 0) | |
3013 | { | |
3014 | if (kill) | |
3015 | bitmap_set_bit (kill, index); | |
3016 | bitmap_clear_bit (gen, index); | |
3017 | } | |
3018 | } | |
3019 | } | |
3020 | } | |
3021 | else | |
3022 | { | |
3023 | /* The groups are different, if the alias sets | |
3024 | conflict, clear the entire group. We only need | |
3025 | to apply this test if the read_info is a cselib | |
3026 | read. Anything with a constant base cannot alias | |
3027 | something else with a different constant | |
3028 | base. */ | |
3029 | if ((read_info->group_id < 0) | |
48e1416a | 3030 | && canon_true_dependence (group->base_mem, |
ec410bf1 | 3031 | GET_MODE (group->base_mem), |
82d2c88b | 3032 | group->canon_base_addr, |
376a287d | 3033 | read_info->mem, NULL_RTX)) |
3072d30e | 3034 | { |
3035 | if (kill) | |
3036 | bitmap_ior_into (kill, group->group_kill); | |
3037 | bitmap_and_compl_into (gen, group->group_kill); | |
3038 | } | |
3039 | } | |
3040 | } | |
3041 | } | |
48e1416a | 3042 | |
3072d30e | 3043 | read_info = read_info->next; |
3044 | } | |
3045 | } | |
3046 | ||
3072d30e | 3047 | |
3048 | /* Return the insn in BB_INFO before the first wild read or if there | |
3049 | are no wild reads in the block, return the last insn. */ | |
3050 | ||
3051 | static insn_info_t | |
3052 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3053 | { | |
3054 | insn_info_t insn_info = bb_info->last_insn; | |
3055 | insn_info_t last_wild_read = NULL; | |
3056 | ||
3057 | while (insn_info) | |
3058 | { | |
3059 | if (insn_info->wild_read) | |
3060 | { | |
3061 | last_wild_read = insn_info->prev_insn; | |
3062 | /* Block starts with wild read. */ | |
3063 | if (!last_wild_read) | |
3064 | return NULL; | |
3065 | } | |
3066 | ||
3067 | insn_info = insn_info->prev_insn; | |
3068 | } | |
3069 | ||
3070 | if (last_wild_read) | |
3071 | return last_wild_read; | |
3072 | else | |
3073 | return bb_info->last_insn; | |
3074 | } | |
3075 | ||
3076 | ||
3077 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3078 | the block in order to build the gen and kill sets for the block. | |
3079 | We start at ptr which may be the last insn in the block or may be | |
3080 | the first insn with a wild read. In the latter case we are able to | |
3081 | skip the rest of the block because it just does not matter: | |
3082 | anything that happens is hidden by the wild read. */ | |
3083 | ||
3084 | static void | |
5a3f5d69 | 3085 | dse_step3_scan (basic_block bb) |
3072d30e | 3086 | { |
3087 | bb_info_t bb_info = bb_table[bb->index]; | |
3088 | insn_info_t insn_info; | |
3089 | ||
5a3f5d69 | 3090 | insn_info = find_insn_before_first_wild_read (bb_info); |
48e1416a | 3091 | |
3072d30e | 3092 | /* In the spill case or in the no_spill case if there is no wild |
3093 | read in the block, we will need a kill set. */ | |
3094 | if (insn_info == bb_info->last_insn) | |
3095 | { | |
3096 | if (bb_info->kill) | |
3097 | bitmap_clear (bb_info->kill); | |
3098 | else | |
4fb07d00 | 3099 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3100 | } |
48e1416a | 3101 | else |
3072d30e | 3102 | if (bb_info->kill) |
3103 | BITMAP_FREE (bb_info->kill); | |
3104 | ||
3105 | while (insn_info) | |
3106 | { | |
3107 | /* There may have been code deleted by the dce pass run before | |
3108 | this phase. */ | |
3109 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3110 | { | |
5a3f5d69 | 3111 | scan_stores (insn_info->store_rec, bb_info->gen, bb_info->kill); |
3112 | scan_reads (insn_info, bb_info->gen, bb_info->kill); | |
48e1416a | 3113 | } |
3072d30e | 3114 | |
3115 | insn_info = insn_info->prev_insn; | |
3116 | } | |
3117 | } | |
3118 | ||
3119 | ||
3120 | /* Set the gen set of the exit block, and also any block with no | |
3121 | successors that does not have a wild read. */ | |
3122 | ||
3123 | static void | |
3124 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3125 | { | |
3126 | /* The gen set is all 0's for the exit block except for the | |
3127 | frame_pointer_group. */ | |
48e1416a | 3128 | |
3072d30e | 3129 | if (stores_off_frame_dead_at_return) |
3130 | { | |
3131 | unsigned int i; | |
c2eab334 | 3132 | group_info *group; |
48e1416a | 3133 | |
f1f41a6c | 3134 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 3135 | { |
3136 | if (group->process_globally && group->frame_related) | |
3137 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3138 | } | |
3139 | } | |
3140 | } | |
3141 | ||
3142 | ||
3143 | /* Find all of the blocks that are not backwards reachable from the | |
3144 | exit block or any block with no successors (BB). These are the | |
3145 | infinite loops or infinite self loops. These blocks will still | |
3146 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3147 | ||
3148 | static void | |
3149 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3150 | { | |
3151 | edge e; | |
3152 | edge_iterator ei; | |
3153 | ||
08b7917c | 3154 | if (bitmap_bit_p (unreachable_blocks, bb->index)) |
3072d30e | 3155 | { |
08b7917c | 3156 | bitmap_clear_bit (unreachable_blocks, bb->index); |
3072d30e | 3157 | FOR_EACH_EDGE (e, ei, bb->preds) |
48e1416a | 3158 | { |
3072d30e | 3159 | mark_reachable_blocks (unreachable_blocks, e->src); |
48e1416a | 3160 | } |
3072d30e | 3161 | } |
3162 | } | |
3163 | ||
3164 | /* Build the transfer functions for the function. */ | |
3165 | ||
3166 | static void | |
5a3f5d69 | 3167 | dse_step3 () |
3072d30e | 3168 | { |
3169 | basic_block bb; | |
3072d30e | 3170 | sbitmap_iterator sbi; |
3171 | bitmap all_ones = NULL; | |
3172 | unsigned int i; | |
48e1416a | 3173 | |
3c6549f8 | 3174 | auto_sbitmap unreachable_blocks (last_basic_block_for_fn (cfun)); |
53c5d9d4 | 3175 | bitmap_ones (unreachable_blocks); |
3072d30e | 3176 | |
ed7d889a | 3177 | FOR_ALL_BB_FN (bb, cfun) |
3072d30e | 3178 | { |
3179 | bb_info_t bb_info = bb_table[bb->index]; | |
3180 | if (bb_info->gen) | |
3181 | bitmap_clear (bb_info->gen); | |
3182 | else | |
4fb07d00 | 3183 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3184 | |
3185 | if (bb->index == ENTRY_BLOCK) | |
3186 | ; | |
3187 | else if (bb->index == EXIT_BLOCK) | |
3188 | dse_step3_exit_block_scan (bb_info); | |
3189 | else | |
5a3f5d69 | 3190 | dse_step3_scan (bb); |
3072d30e | 3191 | if (EDGE_COUNT (bb->succs) == 0) |
3192 | mark_reachable_blocks (unreachable_blocks, bb); | |
3193 | ||
3194 | /* If this is the second time dataflow is run, delete the old | |
3195 | sets. */ | |
3196 | if (bb_info->in) | |
3197 | BITMAP_FREE (bb_info->in); | |
3198 | if (bb_info->out) | |
3199 | BITMAP_FREE (bb_info->out); | |
3200 | } | |
3201 | ||
3202 | /* For any block in an infinite loop, we must initialize the out set | |
3203 | to all ones. This could be expensive, but almost never occurs in | |
3204 | practice. However, it is common in regression tests. */ | |
0d211963 | 3205 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
3072d30e | 3206 | { |
3207 | if (bitmap_bit_p (all_blocks, i)) | |
3208 | { | |
3209 | bb_info_t bb_info = bb_table[i]; | |
3210 | if (!all_ones) | |
3211 | { | |
3212 | unsigned int j; | |
c2eab334 | 3213 | group_info *group; |
3072d30e | 3214 | |
4fb07d00 | 3215 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
f1f41a6c | 3216 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
3072d30e | 3217 | bitmap_ior_into (all_ones, group->group_kill); |
3218 | } | |
3219 | if (!bb_info->out) | |
3220 | { | |
4fb07d00 | 3221 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3222 | bitmap_copy (bb_info->out, all_ones); |
3223 | } | |
3224 | } | |
3225 | } | |
3226 | ||
3227 | if (all_ones) | |
3228 | BITMAP_FREE (all_ones); | |
3072d30e | 3229 | } |
3230 | ||
3231 | ||
3232 | \f | |
3233 | /*---------------------------------------------------------------------------- | |
3234 | Fourth step. | |
3235 | ||
3236 | Solve the bitvector equations. | |
3237 | ----------------------------------------------------------------------------*/ | |
3238 | ||
3239 | ||
3240 | /* Confluence function for blocks with no successors. Create an out | |
3241 | set from the gen set of the exit block. This block logically has | |
3242 | the exit block as a successor. */ | |
3243 | ||
3244 | ||
3245 | ||
3246 | static void | |
3247 | dse_confluence_0 (basic_block bb) | |
3248 | { | |
3249 | bb_info_t bb_info = bb_table[bb->index]; | |
3250 | ||
3251 | if (bb->index == EXIT_BLOCK) | |
3252 | return; | |
3253 | ||
3254 | if (!bb_info->out) | |
3255 | { | |
4fb07d00 | 3256 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3257 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3258 | } | |
3259 | } | |
3260 | ||
3261 | /* Propagate the information from the in set of the dest of E to the | |
3262 | out set of the src of E. If the various in or out sets are not | |
3263 | there, that means they are all ones. */ | |
3264 | ||
a703ca31 | 3265 | static bool |
3072d30e | 3266 | dse_confluence_n (edge e) |
3267 | { | |
3268 | bb_info_t src_info = bb_table[e->src->index]; | |
3269 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3270 | ||
3271 | if (dest_info->in) | |
3272 | { | |
3273 | if (src_info->out) | |
3274 | bitmap_and_into (src_info->out, dest_info->in); | |
3275 | else | |
3276 | { | |
4fb07d00 | 3277 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3278 | bitmap_copy (src_info->out, dest_info->in); |
3279 | } | |
3280 | } | |
a703ca31 | 3281 | return true; |
3072d30e | 3282 | } |
3283 | ||
3284 | ||
3285 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
48e1416a | 3286 | block. There are three cases: |
3072d30e | 3287 | |
3288 | 1) The block has no kill set. In this case the kill set is all | |
3289 | ones. It does not matter what the out set of the block is, none of | |
3290 | the info can reach the top. The only thing that reaches the top is | |
3291 | the gen set and we just copy the set. | |
3292 | ||
3293 | 2) There is a kill set but no out set and bb has successors. In | |
3294 | this case we just return. Eventually an out set will be created and | |
3295 | it is better to wait than to create a set of ones. | |
3296 | ||
3297 | 3) There is both a kill and out set. We apply the obvious transfer | |
3298 | function. | |
3299 | */ | |
3300 | ||
3301 | static bool | |
3302 | dse_transfer_function (int bb_index) | |
3303 | { | |
3304 | bb_info_t bb_info = bb_table[bb_index]; | |
3305 | ||
3306 | if (bb_info->kill) | |
3307 | { | |
3308 | if (bb_info->out) | |
3309 | { | |
3310 | /* Case 3 above. */ | |
3311 | if (bb_info->in) | |
48e1416a | 3312 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
3072d30e | 3313 | bb_info->out, bb_info->kill); |
3314 | else | |
3315 | { | |
4fb07d00 | 3316 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
48e1416a | 3317 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
3072d30e | 3318 | bb_info->out, bb_info->kill); |
3319 | return true; | |
3320 | } | |
3321 | } | |
3322 | else | |
3323 | /* Case 2 above. */ | |
3324 | return false; | |
3325 | } | |
3326 | else | |
3327 | { | |
3328 | /* Case 1 above. If there is already an in set, nothing | |
3329 | happens. */ | |
3330 | if (bb_info->in) | |
3331 | return false; | |
3332 | else | |
3333 | { | |
4fb07d00 | 3334 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3335 | bitmap_copy (bb_info->in, bb_info->gen); |
3336 | return true; | |
3337 | } | |
3338 | } | |
3339 | } | |
3340 | ||
3341 | /* Solve the dataflow equations. */ | |
3342 | ||
3343 | static void | |
3344 | dse_step4 (void) | |
3345 | { | |
48e1416a | 3346 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3347 | dse_confluence_n, dse_transfer_function, | |
3348 | all_blocks, df_get_postorder (DF_BACKWARD), | |
3072d30e | 3349 | df_get_n_blocks (DF_BACKWARD)); |
1ca59310 | 3350 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 3351 | { |
3352 | basic_block bb; | |
3353 | ||
3354 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
ed7d889a | 3355 | FOR_ALL_BB_FN (bb, cfun) |
3072d30e | 3356 | { |
3357 | bb_info_t bb_info = bb_table[bb->index]; | |
3358 | ||
3359 | df_print_bb_index (bb, dump_file); | |
3360 | if (bb_info->in) | |
3361 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3362 | else | |
3363 | fprintf (dump_file, " in: *MISSING*\n"); | |
3364 | if (bb_info->gen) | |
3365 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3366 | else | |
3367 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3368 | if (bb_info->kill) | |
3369 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3370 | else | |
3371 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3372 | if (bb_info->out) | |
3373 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3374 | else | |
3375 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3376 | } | |
3377 | } | |
3378 | } | |
3379 | ||
3380 | ||
3381 | \f | |
3382 | /*---------------------------------------------------------------------------- | |
3383 | Fifth step. | |
3384 | ||
bef304b8 | 3385 | Delete the stores that can only be deleted using the global information. |
3072d30e | 3386 | ----------------------------------------------------------------------------*/ |
3387 | ||
3388 | ||
3389 | static void | |
5a3f5d69 | 3390 | dse_step5 (void) |
3072d30e | 3391 | { |
3392 | basic_block bb; | |
fc00614f | 3393 | FOR_EACH_BB_FN (bb, cfun) |
3072d30e | 3394 | { |
3395 | bb_info_t bb_info = bb_table[bb->index]; | |
3396 | insn_info_t insn_info = bb_info->last_insn; | |
3397 | bitmap v = bb_info->out; | |
3398 | ||
3399 | while (insn_info) | |
3400 | { | |
3401 | bool deleted = false; | |
3402 | if (dump_file && insn_info->insn) | |
3403 | { | |
3404 | fprintf (dump_file, "starting to process insn %d\n", | |
3405 | INSN_UID (insn_info->insn)); | |
3406 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3407 | } | |
3408 | ||
3409 | /* There may have been code deleted by the dce pass run before | |
3410 | this phase. */ | |
48e1416a | 3411 | if (insn_info->insn |
3072d30e | 3412 | && INSN_P (insn_info->insn) |
3413 | && (!insn_info->cannot_delete) | |
3414 | && (!bitmap_empty_p (v))) | |
3415 | { | |
c2eab334 | 3416 | store_info *store_info = insn_info->store_rec; |
3072d30e | 3417 | |
3418 | /* Try to delete the current insn. */ | |
3419 | deleted = true; | |
48e1416a | 3420 | |
3072d30e | 3421 | /* Skip the clobbers. */ |
3422 | while (!store_info->is_set) | |
3423 | store_info = store_info->next; | |
3424 | ||
370ad826 | 3425 | HOST_WIDE_INT i, offset, width; |
5a3f5d69 | 3426 | group_info *group_info = rtx_group_vec[store_info->group_id]; |
3427 | ||
370ad826 | 3428 | if (!store_info->offset.is_constant (&offset) |
3429 | || !store_info->width.is_constant (&width)) | |
3430 | deleted = false; | |
3431 | else | |
3072d30e | 3432 | { |
370ad826 | 3433 | HOST_WIDE_INT end = offset + width; |
3434 | for (i = offset; i < end; i++) | |
3072d30e | 3435 | { |
370ad826 | 3436 | int index = get_bitmap_index (group_info, i); |
3437 | ||
1ca59310 | 3438 | if (dump_file && (dump_flags & TDF_DETAILS)) |
370ad826 | 3439 | fprintf (dump_file, "i = %d, index = %d\n", |
3440 | (int) i, index); | |
3441 | if (index == 0 || !bitmap_bit_p (v, index)) | |
3442 | { | |
3443 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3444 | fprintf (dump_file, "failing at i = %d\n", | |
3445 | (int) i); | |
3446 | deleted = false; | |
3447 | break; | |
3448 | } | |
3072d30e | 3449 | } |
3450 | } | |
3451 | if (deleted) | |
3452 | { | |
5a9ecd4a | 3453 | if (dbg_cnt (dse) |
3454 | && check_for_inc_dec_1 (insn_info)) | |
3072d30e | 3455 | { |
3072d30e | 3456 | delete_insn (insn_info->insn); |
3457 | insn_info->insn = NULL; | |
3458 | globally_deleted++; | |
3459 | } | |
3460 | } | |
3461 | } | |
3462 | /* We do want to process the local info if the insn was | |
6dfdc153 | 3463 | deleted. For instance, if the insn did a wild read, we |
3072d30e | 3464 | no longer need to trash the info. */ |
48e1416a | 3465 | if (insn_info->insn |
3072d30e | 3466 | && INSN_P (insn_info->insn) |
3467 | && (!deleted)) | |
3468 | { | |
5a3f5d69 | 3469 | scan_stores (insn_info->store_rec, v, NULL); |
3072d30e | 3470 | if (insn_info->wild_read) |
3471 | { | |
1ca59310 | 3472 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 3473 | fprintf (dump_file, "wild read\n"); |
3474 | bitmap_clear (v); | |
3475 | } | |
b4a708fb | 3476 | else if (insn_info->read_rec |
3d35f723 | 3477 | || insn_info->non_frame_wild_read |
3478 | || insn_info->frame_read) | |
3072d30e | 3479 | { |
3d35f723 | 3480 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3481 | { | |
3482 | if (!insn_info->non_frame_wild_read | |
3483 | && !insn_info->frame_read) | |
3484 | fprintf (dump_file, "regular read\n"); | |
3485 | if (insn_info->non_frame_wild_read) | |
3486 | fprintf (dump_file, "non-frame wild read\n"); | |
3487 | if (insn_info->frame_read) | |
3488 | fprintf (dump_file, "frame read\n"); | |
3489 | } | |
5a3f5d69 | 3490 | scan_reads (insn_info, v, NULL); |
3072d30e | 3491 | } |
3492 | } | |
48e1416a | 3493 | |
3072d30e | 3494 | insn_info = insn_info->prev_insn; |
3495 | } | |
3496 | } | |
3497 | } | |
3498 | ||
3499 | ||
3072d30e | 3500 | \f |
3501 | /*---------------------------------------------------------------------------- | |
3502 | Sixth step. | |
3503 | ||
aa140b76 | 3504 | Delete stores made redundant by earlier stores (which store the same |
3505 | value) that couldn't be eliminated. | |
3506 | ----------------------------------------------------------------------------*/ | |
3507 | ||
3508 | static void | |
3509 | dse_step6 (void) | |
3510 | { | |
3511 | basic_block bb; | |
3512 | ||
ed7d889a | 3513 | FOR_ALL_BB_FN (bb, cfun) |
aa140b76 | 3514 | { |
3515 | bb_info_t bb_info = bb_table[bb->index]; | |
3516 | insn_info_t insn_info = bb_info->last_insn; | |
3517 | ||
3518 | while (insn_info) | |
3519 | { | |
3520 | /* There may have been code deleted by the dce pass run before | |
3521 | this phase. */ | |
3522 | if (insn_info->insn | |
3523 | && INSN_P (insn_info->insn) | |
3524 | && !insn_info->cannot_delete) | |
3525 | { | |
c2eab334 | 3526 | store_info *s_info = insn_info->store_rec; |
aa140b76 | 3527 | |
3528 | while (s_info && !s_info->is_set) | |
3529 | s_info = s_info->next; | |
3530 | if (s_info | |
3531 | && s_info->redundant_reason | |
3532 | && s_info->redundant_reason->insn | |
3533 | && INSN_P (s_info->redundant_reason->insn)) | |
3534 | { | |
cccc26f6 | 3535 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
1ca59310 | 3536 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 3537 | fprintf (dump_file, "Locally deleting insn %d " |
3538 | "because insn %d stores the " | |
3539 | "same value and couldn't be " | |
3540 | "eliminated\n", | |
3541 | INSN_UID (insn_info->insn), | |
3542 | INSN_UID (rinsn)); | |
3543 | delete_dead_store_insn (insn_info); | |
3544 | } | |
3545 | } | |
3546 | insn_info = insn_info->prev_insn; | |
3547 | } | |
3548 | } | |
3549 | } | |
3550 | \f | |
3551 | /*---------------------------------------------------------------------------- | |
3552 | Seventh step. | |
3553 | ||
48e1416a | 3554 | Destroy everything left standing. |
3072d30e | 3555 | ----------------------------------------------------------------------------*/ |
3556 | ||
48e1416a | 3557 | static void |
4fb07d00 | 3558 | dse_step7 (void) |
3072d30e | 3559 | { |
4fb07d00 | 3560 | bitmap_obstack_release (&dse_bitmap_obstack); |
3561 | obstack_free (&dse_obstack, NULL); | |
ce299759 | 3562 | |
3072d30e | 3563 | end_alias_analysis (); |
3564 | free (bb_table); | |
c1f445d2 | 3565 | delete rtx_group_table; |
3566 | rtx_group_table = NULL; | |
f1f41a6c | 3567 | rtx_group_vec.release (); |
3072d30e | 3568 | BITMAP_FREE (all_blocks); |
3569 | BITMAP_FREE (scratch); | |
3570 | ||
55c5ac9f | 3571 | rtx_store_info_pool.release (); |
e16712b1 | 3572 | read_info_type_pool.release (); |
3573 | insn_info_type_pool.release (); | |
3574 | dse_bb_info_type_pool.release (); | |
3575 | group_info_pool.release (); | |
3576 | deferred_change_pool.release (); | |
3072d30e | 3577 | } |
3578 | ||
3579 | ||
3072d30e | 3580 | /* ------------------------------------------------------------------------- |
3581 | DSE | |
3582 | ------------------------------------------------------------------------- */ | |
3583 | ||
3584 | /* Callback for running pass_rtl_dse. */ | |
3585 | ||
3586 | static unsigned int | |
3587 | rest_of_handle_dse (void) | |
3588 | { | |
3072d30e | 3589 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3590 | ||
a1b0a968 | 3591 | /* Need the notes since we must track live hardregs in the forwards |
3592 | direction. */ | |
3593 | df_note_add_problem (); | |
3594 | df_analyze (); | |
3595 | ||
3072d30e | 3596 | dse_step0 (); |
3597 | dse_step1 (); | |
3598 | dse_step2_init (); | |
5a3f5d69 | 3599 | if (dse_step2 ()) |
3072d30e | 3600 | { |
3601 | df_set_flags (DF_LR_RUN_DCE); | |
3602 | df_analyze (); | |
1ca59310 | 3603 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 3604 | fprintf (dump_file, "doing global processing\n"); |
5a3f5d69 | 3605 | dse_step3 (); |
3072d30e | 3606 | dse_step4 (); |
5a3f5d69 | 3607 | dse_step5 (); |
3072d30e | 3608 | } |
3609 | ||
aa140b76 | 3610 | dse_step6 (); |
4fb07d00 | 3611 | dse_step7 (); |
3072d30e | 3612 | |
3613 | if (dump_file) | |
5a3f5d69 | 3614 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d\n", |
3615 | locally_deleted, globally_deleted); | |
1f91a12d | 3616 | |
3617 | /* DSE can eliminate potentially-trapping MEMs. | |
3618 | Remove any EH edges associated with them. */ | |
3619 | if ((locally_deleted || globally_deleted) | |
3620 | && cfun->can_throw_non_call_exceptions | |
3621 | && purge_all_dead_edges ()) | |
3622 | cleanup_cfg (0); | |
3623 | ||
3072d30e | 3624 | return 0; |
3625 | } | |
3626 | ||
7620bc82 | 3627 | namespace { |
3628 | ||
3629 | const pass_data pass_data_rtl_dse1 = | |
cbe8bda8 | 3630 | { |
3631 | RTL_PASS, /* type */ | |
3632 | "dse1", /* name */ | |
3633 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 3634 | TV_DSE1, /* tv_id */ |
3635 | 0, /* properties_required */ | |
3636 | 0, /* properties_provided */ | |
3637 | 0, /* properties_destroyed */ | |
3638 | 0, /* todo_flags_start */ | |
8b88439e | 3639 | TODO_df_finish, /* todo_flags_finish */ |
3072d30e | 3640 | }; |
3641 | ||
7620bc82 | 3642 | class pass_rtl_dse1 : public rtl_opt_pass |
cbe8bda8 | 3643 | { |
3644 | public: | |
9af5ce0c | 3645 | pass_rtl_dse1 (gcc::context *ctxt) |
3646 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
cbe8bda8 | 3647 | {} |
3648 | ||
3649 | /* opt_pass methods: */ | |
31315c24 | 3650 | virtual bool gate (function *) |
3651 | { | |
3652 | return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
3653 | } | |
3654 | ||
65b0537f | 3655 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
cbe8bda8 | 3656 | |
3657 | }; // class pass_rtl_dse1 | |
3658 | ||
7620bc82 | 3659 | } // anon namespace |
3660 | ||
cbe8bda8 | 3661 | rtl_opt_pass * |
3662 | make_pass_rtl_dse1 (gcc::context *ctxt) | |
3663 | { | |
3664 | return new pass_rtl_dse1 (ctxt); | |
3665 | } | |
3666 | ||
7620bc82 | 3667 | namespace { |
3668 | ||
3669 | const pass_data pass_data_rtl_dse2 = | |
cbe8bda8 | 3670 | { |
3671 | RTL_PASS, /* type */ | |
3672 | "dse2", /* name */ | |
3673 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 3674 | TV_DSE2, /* tv_id */ |
3675 | 0, /* properties_required */ | |
3676 | 0, /* properties_provided */ | |
3677 | 0, /* properties_destroyed */ | |
3678 | 0, /* todo_flags_start */ | |
8b88439e | 3679 | TODO_df_finish, /* todo_flags_finish */ |
3072d30e | 3680 | }; |
cbe8bda8 | 3681 | |
7620bc82 | 3682 | class pass_rtl_dse2 : public rtl_opt_pass |
cbe8bda8 | 3683 | { |
3684 | public: | |
9af5ce0c | 3685 | pass_rtl_dse2 (gcc::context *ctxt) |
3686 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
cbe8bda8 | 3687 | {} |
3688 | ||
3689 | /* opt_pass methods: */ | |
31315c24 | 3690 | virtual bool gate (function *) |
3691 | { | |
3692 | return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
3693 | } | |
3694 | ||
65b0537f | 3695 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
cbe8bda8 | 3696 | |
3697 | }; // class pass_rtl_dse2 | |
3698 | ||
7620bc82 | 3699 | } // anon namespace |
3700 | ||
cbe8bda8 | 3701 | rtl_opt_pass * |
3702 | make_pass_rtl_dse2 (gcc::context *ctxt) | |
3703 | { | |
3704 | return new pass_rtl_dse2 (ctxt); | |
3705 | } |