]>
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. */ | |
251317e4 | 223 | class store_info |
3072d30e | 224 | { |
251317e4 | 225 | public: |
3072d30e | 226 | |
227 | /* False means this is a clobber. */ | |
228 | bool is_set; | |
229 | ||
aa140b76 | 230 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
231 | bool is_large; | |
232 | ||
3072d30e | 233 | /* The id of the mem group of the base address. If rtx_varies_p is |
234 | true, this is -1. Otherwise, it is the index into the group | |
235 | table. */ | |
236 | int group_id; | |
48e1416a | 237 | |
3072d30e | 238 | /* This is the cselib value. */ |
239 | cselib_val *cse_base; | |
240 | ||
241 | /* This canonized mem. */ | |
242 | rtx mem; | |
243 | ||
82d2c88b | 244 | /* Canonized MEM address for use by canon_true_dependence. */ |
3072d30e | 245 | rtx mem_addr; |
246 | ||
57bca8dc | 247 | /* The offset of the first byte associated with the operation. */ |
370ad826 | 248 | poly_int64 offset; |
57bca8dc | 249 | |
250 | /* The number of bytes covered by the operation. This is always exact | |
251 | and known (rather than -1). */ | |
370ad826 | 252 | poly_int64 width; |
aa140b76 | 253 | |
254 | union | |
255 | { | |
256 | /* A bitmask as wide as the number of bytes in the word that | |
257 | contains a 1 if the byte may be needed. The store is unused if | |
258 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
259 | unsigned HOST_WIDE_INT small_bitmask; | |
260 | ||
261 | struct | |
262 | { | |
370ad826 | 263 | /* A bitmap with one bit per byte, or null if the number of |
264 | bytes isn't known at compile time. A cleared bit means | |
265 | the position is needed. Used if IS_LARGE is true. */ | |
843bd2fa | 266 | bitmap bmap; |
3072d30e | 267 | |
370ad826 | 268 | /* When BITMAP is nonnull, this counts the number of set bits |
269 | (i.e. unneeded bytes) in the bitmap. If it is equal to | |
270 | WIDTH, the whole store is unused. | |
271 | ||
272 | When BITMAP is null: | |
273 | - the store is definitely not needed when COUNT == 1 | |
274 | - all the store is needed when COUNT == 0 and RHS is nonnull | |
275 | - otherwise we don't know which parts of the store are needed. */ | |
aa140b76 | 276 | int count; |
277 | } large; | |
278 | } positions_needed; | |
3072d30e | 279 | |
280 | /* The next store info for this insn. */ | |
2e966e2a | 281 | class store_info *next; |
3072d30e | 282 | |
283 | /* The right hand side of the store. This is used if there is a | |
284 | subsequent reload of the mems address somewhere later in the | |
285 | basic block. */ | |
aa140b76 | 286 | rtx rhs; |
287 | ||
288 | /* If rhs is or holds a constant, this contains that constant, | |
289 | otherwise NULL. */ | |
290 | rtx const_rhs; | |
291 | ||
292 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
293 | insn stored. These aren't eliminated early, because doing that | |
294 | might prevent the earlier larger store to be eliminated. */ | |
55c5ac9f | 295 | struct insn_info_type *redundant_reason; |
3072d30e | 296 | }; |
297 | ||
4e43e20a | 298 | /* Return a bitmask with the first N low bits set. */ |
299 | ||
300 | static unsigned HOST_WIDE_INT | |
301 | lowpart_bitmask (int n) | |
302 | { | |
7097b942 | 303 | unsigned HOST_WIDE_INT mask = HOST_WIDE_INT_M1U; |
4e43e20a | 304 | return mask >> (HOST_BITS_PER_WIDE_INT - n); |
305 | } | |
306 | ||
1dc6c44d | 307 | static object_allocator<store_info> cse_store_info_pool ("cse_store_info_pool"); |
55c5ac9f | 308 | |
1dc6c44d | 309 | static object_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool"); |
3072d30e | 310 | |
311 | /* This structure holds information about a load. These are only | |
312 | built for rtx bases. */ | |
251317e4 | 313 | class read_info_type |
3072d30e | 314 | { |
251317e4 | 315 | public: |
3072d30e | 316 | /* The id of the mem group of the base address. */ |
317 | int group_id; | |
318 | ||
57bca8dc | 319 | /* The offset of the first byte associated with the operation. */ |
370ad826 | 320 | poly_int64 offset; |
57bca8dc | 321 | |
322 | /* The number of bytes covered by the operation, or -1 if not known. */ | |
370ad826 | 323 | poly_int64 width; |
3072d30e | 324 | |
325 | /* The mem being read. */ | |
326 | rtx mem; | |
327 | ||
328 | /* The next read_info for this insn. */ | |
2e966e2a | 329 | class read_info_type *next; |
3072d30e | 330 | }; |
2e966e2a | 331 | typedef class read_info_type *read_info_t; |
3072d30e | 332 | |
1dc6c44d | 333 | static object_allocator<read_info_type> read_info_type_pool ("read_info_pool"); |
3072d30e | 334 | |
335 | /* One of these records is created for each insn. */ | |
336 | ||
55c5ac9f | 337 | struct insn_info_type |
3072d30e | 338 | { |
339 | /* Set true if the insn contains a store but the insn itself cannot | |
340 | be deleted. This is set if the insn is a parallel and there is | |
341 | more than one non dead output or if the insn is in some way | |
342 | volatile. */ | |
343 | bool cannot_delete; | |
344 | ||
345 | /* This field is only used by the global algorithm. It is set true | |
346 | if the insn contains any read of mem except for a (1). This is | |
347 | also set if the insn is a call or has a clobber mem. If the insn | |
348 | contains a wild read, the use_rec will be null. */ | |
349 | bool wild_read; | |
350 | ||
b4a708fb | 351 | /* This is true only for CALL instructions which could potentially read |
352 | any non-frame memory location. This field is used by the global | |
353 | algorithm. */ | |
354 | bool non_frame_wild_read; | |
355 | ||
17e1318c | 356 | /* This field is only used for the processing of const functions. |
357 | These functions cannot read memory, but they can read the stack | |
16bf64db | 358 | because that is where they may get their parms. We need to be |
359 | this conservative because, like the store motion pass, we don't | |
360 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
361 | Moreover, we need to distinguish two cases: | |
362 | 1. Before reload (register elimination), the stores related to | |
363 | outgoing arguments are stack pointer based and thus deemed | |
364 | of non-constant base in this pass. This requires special | |
365 | handling but also means that the frame pointer based stores | |
366 | need not be killed upon encountering a const function call. | |
367 | 2. After reload, the stores related to outgoing arguments can be | |
368 | either stack pointer or hard frame pointer based. This means | |
369 | that we have no other choice than also killing all the frame | |
370 | pointer based stores upon encountering a const function call. | |
17853422 | 371 | This field is set after reload for const function calls and before |
372 | reload for const tail function calls on targets where arg pointer | |
373 | is the frame pointer. Having this set is less severe than a wild | |
374 | read, it just means that all the frame related stores are killed | |
375 | rather than all the stores. */ | |
16bf64db | 376 | bool frame_read; |
377 | ||
378 | /* This field is only used for the processing of const functions. | |
379 | It is set if the insn may contain a stack pointer based store. */ | |
17e1318c | 380 | bool stack_pointer_based; |
3072d30e | 381 | |
382 | /* This is true if any of the sets within the store contains a | |
383 | cselib base. Such stores can only be deleted by the local | |
384 | algorithm. */ | |
385 | bool contains_cselib_groups; | |
386 | ||
387 | /* The insn. */ | |
ebabb7a3 | 388 | rtx_insn *insn; |
3072d30e | 389 | |
390 | /* The list of mem sets or mem clobbers that are contained in this | |
391 | insn. If the insn is deletable, it contains only one mem set. | |
392 | But it could also contain clobbers. Insns that contain more than | |
393 | one mem set are not deletable, but each of those mems are here in | |
6dfdc153 | 394 | order to provide info to delete other insns. */ |
c2eab334 | 395 | store_info *store_rec; |
3072d30e | 396 | |
397 | /* The linked list of mem uses in this insn. Only the reads from | |
398 | rtx bases are listed here. The reads to cselib bases are | |
399 | completely processed during the first scan and so are never | |
400 | created. */ | |
401 | read_info_t read_rec; | |
402 | ||
5a9ecd4a | 403 | /* The live fixed registers. We assume only fixed registers can |
404 | cause trouble by being clobbered from an expanded pattern; | |
405 | storing only the live fixed registers (rather than all registers) | |
406 | means less memory needs to be allocated / copied for the individual | |
407 | stores. */ | |
408 | regset fixed_regs_live; | |
409 | ||
3072d30e | 410 | /* The prev insn in the basic block. */ |
55c5ac9f | 411 | struct insn_info_type * prev_insn; |
3072d30e | 412 | |
413 | /* The linked list of insns that are in consideration for removal in | |
9d75589a | 414 | the forwards pass through the basic block. This pointer may be |
3072d30e | 415 | trash as it is not cleared when a wild read occurs. The only |
f0b5f617 | 416 | time it is guaranteed to be correct is when the traversal starts |
3072d30e | 417 | at active_local_stores. */ |
55c5ac9f | 418 | struct insn_info_type * next_local_store; |
3072d30e | 419 | }; |
55c5ac9f | 420 | typedef struct insn_info_type *insn_info_t; |
3072d30e | 421 | |
1dc6c44d | 422 | static object_allocator<insn_info_type> insn_info_type_pool ("insn_info_pool"); |
3072d30e | 423 | |
424 | /* The linked list of stores that are under consideration in this | |
48e1416a | 425 | basic block. */ |
3072d30e | 426 | static insn_info_t active_local_stores; |
1242bee6 | 427 | static int active_local_stores_len; |
3072d30e | 428 | |
55c5ac9f | 429 | struct dse_bb_info_type |
3072d30e | 430 | { |
3072d30e | 431 | /* Pointer to the insn info for the last insn in the block. These |
432 | are linked so this is how all of the insns are reached. During | |
433 | scanning this is the current insn being scanned. */ | |
434 | insn_info_t last_insn; | |
435 | ||
436 | /* The info for the global dataflow problem. */ | |
437 | ||
438 | ||
439 | /* This is set if the transfer function should and in the wild_read | |
440 | bitmap before applying the kill and gen sets. That vector knocks | |
441 | out most of the bits in the bitmap and thus speeds up the | |
442 | operations. */ | |
443 | bool apply_wild_read; | |
444 | ||
a1b0a968 | 445 | /* The following 4 bitvectors hold information about which positions |
446 | of which stores are live or dead. They are indexed by | |
447 | get_bitmap_index. */ | |
448 | ||
3072d30e | 449 | /* The set of store positions that exist in this block before a wild read. */ |
450 | bitmap gen; | |
48e1416a | 451 | |
3072d30e | 452 | /* The set of load positions that exist in this block above the |
453 | same position of a store. */ | |
454 | bitmap kill; | |
455 | ||
456 | /* The set of stores that reach the top of the block without being | |
457 | killed by a read. | |
458 | ||
459 | Do not represent the in if it is all ones. Note that this is | |
460 | what the bitvector should logically be initialized to for a set | |
461 | intersection problem. However, like the kill set, this is too | |
462 | expensive. So initially, the in set will only be created for the | |
463 | exit block and any block that contains a wild read. */ | |
464 | bitmap in; | |
465 | ||
466 | /* The set of stores that reach the bottom of the block from it's | |
467 | successors. | |
468 | ||
469 | Do not represent the in if it is all ones. Note that this is | |
470 | what the bitvector should logically be initialized to for a set | |
471 | intersection problem. However, like the kill and in set, this is | |
472 | too expensive. So what is done is that the confluence operator | |
473 | just initializes the vector from one of the out sets of the | |
474 | successors of the block. */ | |
475 | bitmap out; | |
a1b0a968 | 476 | |
477 | /* The following bitvector is indexed by the reg number. It | |
478 | contains the set of regs that are live at the current instruction | |
479 | being processed. While it contains info for all of the | |
5a9ecd4a | 480 | registers, only the hard registers are actually examined. It is used |
481 | to assure that shift and/or add sequences that are inserted do not | |
9d75589a | 482 | accidentally clobber live hard regs. */ |
a1b0a968 | 483 | bitmap regs_live; |
3072d30e | 484 | }; |
485 | ||
55c5ac9f | 486 | typedef struct dse_bb_info_type *bb_info_t; |
e16712b1 | 487 | |
488 | static object_allocator<dse_bb_info_type> dse_bb_info_type_pool | |
1dc6c44d | 489 | ("bb_info_pool"); |
3072d30e | 490 | |
491 | /* Table to hold all bb_infos. */ | |
492 | static bb_info_t *bb_table; | |
493 | ||
494 | /* There is a group_info for each rtx base that is used to reference | |
495 | memory. There are also not many of the rtx bases because they are | |
496 | very limited in scope. */ | |
497 | ||
48e1416a | 498 | struct group_info |
3072d30e | 499 | { |
500 | /* The actual base of the address. */ | |
501 | rtx rtx_base; | |
502 | ||
503 | /* The sequential id of the base. This allows us to have a | |
504 | canonical ordering of these that is not based on addresses. */ | |
505 | int id; | |
506 | ||
0ac758f7 | 507 | /* True if there are any positions that are to be processed |
508 | globally. */ | |
509 | bool process_globally; | |
510 | ||
511 | /* True if the base of this group is either the frame_pointer or | |
512 | hard_frame_pointer. */ | |
513 | bool frame_related; | |
514 | ||
ec410bf1 | 515 | /* A mem wrapped around the base pointer for the group in order to do |
516 | read dependency. It must be given BLKmode in order to encompass all | |
517 | the possible offsets from the base. */ | |
3072d30e | 518 | rtx base_mem; |
48e1416a | 519 | |
82d2c88b | 520 | /* Canonized version of base_mem's address. */ |
521 | rtx canon_base_addr; | |
3072d30e | 522 | |
523 | /* These two sets of two bitmaps are used to keep track of how many | |
6dfdc153 | 524 | stores are actually referencing that position from this base. We |
3072d30e | 525 | only do this for rtx bases as this will be used to assign |
6dfdc153 | 526 | positions in the bitmaps for the global problem. Bit N is set in |
3072d30e | 527 | store1 on the first store for offset N. Bit N is set in store2 |
528 | for the second store to offset N. This is all we need since we | |
529 | only care about offsets that have two or more stores for them. | |
530 | ||
531 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
532 | for 0 and greater offsets. | |
533 | ||
534 | There is one special case here, for stores into the stack frame, | |
535 | we will or store1 into store2 before deciding which stores look | |
536 | at globally. This is because stores to the stack frame that have | |
537 | no other reads before the end of the function can also be | |
538 | deleted. */ | |
539 | bitmap store1_n, store1_p, store2_n, store2_p; | |
540 | ||
b4a708fb | 541 | /* These bitmaps keep track of offsets in this group escape this function. |
542 | An offset escapes if it corresponds to a named variable whose | |
543 | addressable flag is set. */ | |
544 | bitmap escaped_n, escaped_p; | |
545 | ||
6dfdc153 | 546 | /* The positions in this bitmap have the same assignments as the in, |
3072d30e | 547 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6dfdc153 | 548 | the positions that are occupied by stores for this group. */ |
3072d30e | 549 | bitmap group_kill; |
550 | ||
3072d30e | 551 | /* The offset_map is used to map the offsets from this base into |
6dfdc153 | 552 | positions in the global bitmaps. It is only created after all of |
3072d30e | 553 | the all of stores have been scanned and we know which ones we |
554 | care about. */ | |
48e1416a | 555 | int *offset_map_n, *offset_map_p; |
556 | int offset_map_size_n, offset_map_size_p; | |
3072d30e | 557 | }; |
55c5ac9f | 558 | |
1dc6c44d | 559 | static object_allocator<group_info> group_info_pool ("rtx_group_info_pool"); |
3072d30e | 560 | |
3072d30e | 561 | /* Index into the rtx_group_vec. */ |
562 | static int rtx_group_next_id; | |
563 | ||
3072d30e | 564 | |
c2eab334 | 565 | static vec<group_info *> rtx_group_vec; |
3072d30e | 566 | |
567 | ||
568 | /* This structure holds the set of changes that are being deferred | |
569 | when removing read operation. See replace_read. */ | |
48e1416a | 570 | struct deferred_change |
3072d30e | 571 | { |
572 | ||
573 | /* The mem that is being replaced. */ | |
574 | rtx *loc; | |
575 | ||
576 | /* The reg it is being replaced with. */ | |
577 | rtx reg; | |
578 | ||
579 | struct deferred_change *next; | |
580 | }; | |
581 | ||
e16712b1 | 582 | static object_allocator<deferred_change> deferred_change_pool |
1dc6c44d | 583 | ("deferred_change_pool"); |
3072d30e | 584 | |
c2eab334 | 585 | static deferred_change *deferred_change_list = NULL; |
3072d30e | 586 | |
18d50ae6 | 587 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
ff3ae375 | 588 | this for vararg functions because they play games with the frame. */ |
3072d30e | 589 | static bool stores_off_frame_dead_at_return; |
590 | ||
591 | /* Counter for stats. */ | |
48e1416a | 592 | static int globally_deleted; |
593 | static int locally_deleted; | |
48e1416a | 594 | |
3072d30e | 595 | static bitmap all_blocks; |
596 | ||
b4a708fb | 597 | /* Locations that are killed by calls in the global phase. */ |
598 | static bitmap kill_on_calls; | |
599 | ||
3072d30e | 600 | /* The number of bits used in the global bitmaps. */ |
601 | static unsigned int current_position; | |
57bca8dc | 602 | |
603 | /* Print offset range [OFFSET, OFFSET + WIDTH) to FILE. */ | |
604 | ||
605 | static void | |
606 | print_range (FILE *file, poly_int64 offset, poly_int64 width) | |
607 | { | |
608 | fprintf (file, "["); | |
609 | print_dec (offset, file, SIGNED); | |
610 | fprintf (file, ".."); | |
611 | print_dec (offset + width, file, SIGNED); | |
612 | fprintf (file, ")"); | |
613 | } | |
3072d30e | 614 | \f |
615 | /*---------------------------------------------------------------------------- | |
616 | Zeroth step. | |
617 | ||
48e1416a | 618 | Initialization. |
3072d30e | 619 | ----------------------------------------------------------------------------*/ |
620 | ||
3072d30e | 621 | |
3072d30e | 622 | /* Hashtable callbacks for maintaining the "bases" field of |
623 | store_group_info, given that the addresses are function invariants. */ | |
624 | ||
770ff93b | 625 | struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
d1455aa3 | 626 | { |
9969c043 | 627 | static inline hashval_t hash (const group_info *); |
628 | static inline bool equal (const group_info *, const group_info *); | |
d1455aa3 | 629 | }; |
630 | ||
631 | inline bool | |
9969c043 | 632 | invariant_group_base_hasher::equal (const group_info *gi1, |
633 | const group_info *gi2) | |
3072d30e | 634 | { |
3072d30e | 635 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
636 | } | |
637 | ||
d1455aa3 | 638 | inline hashval_t |
9969c043 | 639 | invariant_group_base_hasher::hash (const group_info *gi) |
3072d30e | 640 | { |
3072d30e | 641 | int do_not_record; |
642 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
643 | } | |
644 | ||
d1455aa3 | 645 | /* Tables of group_info structures, hashed by base value. */ |
c1f445d2 | 646 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
d1455aa3 | 647 | |
3072d30e | 648 | |
649 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
650 | ||
c2eab334 | 651 | static group_info * |
3072d30e | 652 | get_group_info (rtx base) |
653 | { | |
48e1416a | 654 | struct group_info tmp_gi; |
c2eab334 | 655 | group_info *gi; |
d1455aa3 | 656 | group_info **slot; |
3072d30e | 657 | |
5a3f5d69 | 658 | gcc_assert (base != NULL_RTX); |
659 | ||
660 | /* Find the store_base_info structure for BASE, creating a new one | |
661 | if necessary. */ | |
662 | tmp_gi.rtx_base = base; | |
663 | slot = rtx_group_table->find_slot (&tmp_gi, INSERT); | |
664 | gi = *slot; | |
3072d30e | 665 | |
666 | if (gi == NULL) | |
667 | { | |
e16712b1 | 668 | *slot = gi = group_info_pool.allocate (); |
3072d30e | 669 | gi->rtx_base = base; |
670 | gi->id = rtx_group_next_id++; | |
ec410bf1 | 671 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
82d2c88b | 672 | gi->canon_base_addr = canon_rtx (base); |
4fb07d00 | 673 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
674 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
675 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
676 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
677 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
678 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
679 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
3072d30e | 680 | gi->process_globally = false; |
48e1416a | 681 | gi->frame_related = |
3072d30e | 682 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
683 | gi->offset_map_size_n = 0; | |
684 | gi->offset_map_size_p = 0; | |
685 | gi->offset_map_n = NULL; | |
686 | gi->offset_map_p = NULL; | |
f1f41a6c | 687 | rtx_group_vec.safe_push (gi); |
3072d30e | 688 | } |
689 | ||
690 | return gi; | |
691 | } | |
692 | ||
693 | ||
694 | /* Initialization of data structures. */ | |
695 | ||
696 | static void | |
697 | dse_step0 (void) | |
698 | { | |
699 | locally_deleted = 0; | |
700 | globally_deleted = 0; | |
3072d30e | 701 | |
4fb07d00 | 702 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
703 | gcc_obstack_init (&dse_obstack); | |
704 | ||
705 | scratch = BITMAP_ALLOC (®_obstack); | |
706 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
3072d30e | 707 | |
3072d30e | 708 | |
c1f445d2 | 709 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
3072d30e | 710 | |
fe672ac0 | 711 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
3072d30e | 712 | rtx_group_next_id = 0; |
713 | ||
18d50ae6 | 714 | stores_off_frame_dead_at_return = !cfun->stdarg; |
3072d30e | 715 | |
716 | init_alias_analysis (); | |
3072d30e | 717 | } |
718 | ||
719 | ||
720 | \f | |
721 | /*---------------------------------------------------------------------------- | |
722 | First step. | |
723 | ||
724 | Scan all of the insns. Any random ordering of the blocks is fine. | |
bef304b8 | 725 | Each block is scanned in forward order to accommodate cselib which |
3072d30e | 726 | is used to remove stores with non-constant bases. |
727 | ----------------------------------------------------------------------------*/ | |
728 | ||
729 | /* Delete all of the store_info recs from INSN_INFO. */ | |
730 | ||
48e1416a | 731 | static void |
3072d30e | 732 | free_store_info (insn_info_t insn_info) |
733 | { | |
c2eab334 | 734 | store_info *cur = insn_info->store_rec; |
735 | while (cur) | |
3072d30e | 736 | { |
c2eab334 | 737 | store_info *next = cur->next; |
738 | if (cur->is_large) | |
739 | BITMAP_FREE (cur->positions_needed.large.bmap); | |
740 | if (cur->cse_base) | |
741 | cse_store_info_pool.remove (cur); | |
3072d30e | 742 | else |
c2eab334 | 743 | rtx_store_info_pool.remove (cur); |
744 | cur = next; | |
3072d30e | 745 | } |
746 | ||
747 | insn_info->cannot_delete = true; | |
748 | insn_info->contains_cselib_groups = false; | |
749 | insn_info->store_rec = NULL; | |
750 | } | |
751 | ||
6dc50383 | 752 | struct note_add_store_info |
5a9ecd4a | 753 | { |
4cd001d5 | 754 | rtx_insn *first, *current; |
5a9ecd4a | 755 | regset fixed_regs_live; |
756 | bool failure; | |
6dc50383 | 757 | }; |
5a9ecd4a | 758 | |
759 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
760 | Check if a register is clobbered which is live afterwards. */ | |
761 | ||
762 | static void | |
763 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
764 | { | |
4cd001d5 | 765 | rtx_insn *insn; |
5a9ecd4a | 766 | note_add_store_info *info = (note_add_store_info *) data; |
5a9ecd4a | 767 | |
768 | if (!REG_P (loc)) | |
769 | return; | |
770 | ||
771 | /* If this register is referenced by the current or an earlier insn, | |
772 | that's OK. E.g. this applies to the register that is being incremented | |
773 | with this addition. */ | |
774 | for (insn = info->first; | |
775 | insn != NEXT_INSN (info->current); | |
776 | insn = NEXT_INSN (insn)) | |
777 | if (reg_referenced_p (loc, PATTERN (insn))) | |
778 | return; | |
779 | ||
780 | /* If we come here, we have a clobber of a register that's only OK | |
781 | if that register is not live. If we don't have liveness information | |
782 | available, fail now. */ | |
783 | if (!info->fixed_regs_live) | |
784 | { | |
6a298741 | 785 | info->failure = true; |
5a9ecd4a | 786 | return; |
787 | } | |
788 | /* Now check if this is a live fixed register. */ | |
6a298741 | 789 | unsigned int end_regno = END_REGNO (loc); |
790 | for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) | |
791 | if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) | |
792 | info->failure = true; | |
5a9ecd4a | 793 | } |
794 | ||
1f864115 | 795 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
796 | SRC + SRCOFF before insn ARG. */ | |
3072d30e | 797 | |
798 | static int | |
1f864115 | 799 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
800 | rtx op ATTRIBUTE_UNUSED, | |
801 | rtx dest, rtx src, rtx srcoff, void *arg) | |
3072d30e | 802 | { |
5a9ecd4a | 803 | insn_info_t insn_info = (insn_info_t) arg; |
4cd001d5 | 804 | rtx_insn *insn = insn_info->insn, *new_insn, *cur; |
5a9ecd4a | 805 | note_add_store_info info; |
48e1416a | 806 | |
1f864115 | 807 | /* We can reuse all operands without copying, because we are about |
808 | to delete the insn that contained it. */ | |
5a9ecd4a | 809 | if (srcoff) |
a280a5da | 810 | { |
811 | start_sequence (); | |
812 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
813 | new_insn = get_insns (); | |
814 | end_sequence (); | |
815 | } | |
5a9ecd4a | 816 | else |
f9a00e9e | 817 | new_insn = gen_move_insn (dest, src); |
5a9ecd4a | 818 | info.first = new_insn; |
819 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
820 | info.failure = false; | |
821 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
822 | { | |
823 | info.current = cur; | |
824 | note_stores (PATTERN (cur), note_add_store, &info); | |
825 | } | |
3072d30e | 826 | |
5a9ecd4a | 827 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
828 | return it immediately, communicating the failure to its caller. */ | |
829 | if (info.failure) | |
830 | return 1; | |
831 | ||
832 | emit_insn_before (new_insn, insn); | |
3072d30e | 833 | |
623ad592 | 834 | return 0; |
3072d30e | 835 | } |
836 | ||
5a9ecd4a | 837 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
838 | is there, is split into a separate insn. | |
839 | Return true on success (or if there was nothing to do), false on failure. */ | |
3072d30e | 840 | |
5a9ecd4a | 841 | static bool |
842 | check_for_inc_dec_1 (insn_info_t insn_info) | |
3072d30e | 843 | { |
ebabb7a3 | 844 | rtx_insn *insn = insn_info->insn; |
3072d30e | 845 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
846 | if (note) | |
623ad592 | 847 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
848 | insn_info) == 0; | |
5a9ecd4a | 849 | return true; |
3072d30e | 850 | } |
851 | ||
852 | ||
5a9ecd4a | 853 | /* Entry point for postreload. If you work on reload_cse, or you need this |
854 | anywhere else, consider if you can provide register liveness information | |
855 | and add a parameter to this function so that it can be passed down in | |
856 | insn_info.fixed_regs_live. */ | |
857 | bool | |
ebabb7a3 | 858 | check_for_inc_dec (rtx_insn *insn) |
5a9ecd4a | 859 | { |
55c5ac9f | 860 | insn_info_type insn_info; |
5a9ecd4a | 861 | rtx note; |
862 | ||
863 | insn_info.insn = insn; | |
864 | insn_info.fixed_regs_live = NULL; | |
865 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
866 | if (note) | |
623ad592 | 867 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
868 | &insn_info) == 0; | |
5a9ecd4a | 869 | return true; |
870 | } | |
871 | ||
48e1416a | 872 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
3072d30e | 873 | |
874 | static void | |
875 | delete_dead_store_insn (insn_info_t insn_info) | |
876 | { | |
877 | read_info_t read_info; | |
878 | ||
879 | if (!dbg_cnt (dse)) | |
880 | return; | |
881 | ||
5a9ecd4a | 882 | if (!check_for_inc_dec_1 (insn_info)) |
883 | return; | |
1ca59310 | 884 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5a3f5d69 | 885 | fprintf (dump_file, "Locally deleting insn %d\n", |
886 | INSN_UID (insn_info->insn)); | |
3072d30e | 887 | |
888 | free_store_info (insn_info); | |
889 | read_info = insn_info->read_rec; | |
48e1416a | 890 | |
3072d30e | 891 | while (read_info) |
892 | { | |
893 | read_info_t next = read_info->next; | |
e16712b1 | 894 | read_info_type_pool.remove (read_info); |
3072d30e | 895 | read_info = next; |
896 | } | |
897 | insn_info->read_rec = NULL; | |
898 | ||
899 | delete_insn (insn_info->insn); | |
900 | locally_deleted++; | |
901 | insn_info->insn = NULL; | |
902 | ||
903 | insn_info->wild_read = false; | |
904 | } | |
905 | ||
f7b5f694 | 906 | /* Return whether DECL, a local variable, can possibly escape the current |
907 | function scope. */ | |
908 | ||
909 | static bool | |
910 | local_variable_can_escape (tree decl) | |
911 | { | |
912 | if (TREE_ADDRESSABLE (decl)) | |
913 | return true; | |
914 | ||
915 | /* If this is a partitioned variable, we need to consider all the variables | |
916 | in the partition. This is necessary because a store into one of them can | |
917 | be replaced with a store into another and this may not change the outcome | |
918 | of the escape analysis. */ | |
919 | if (cfun->gimple_df->decls_to_pointers != NULL) | |
920 | { | |
5f8841a5 | 921 | tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); |
f7b5f694 | 922 | if (namep) |
5f8841a5 | 923 | return TREE_ADDRESSABLE (*namep); |
f7b5f694 | 924 | } |
925 | ||
926 | return false; | |
927 | } | |
928 | ||
929 | /* Return whether EXPR can possibly escape the current function scope. */ | |
930 | ||
b4a708fb | 931 | static bool |
932 | can_escape (tree expr) | |
933 | { | |
934 | tree base; | |
935 | if (!expr) | |
936 | return true; | |
937 | base = get_base_address (expr); | |
938 | if (DECL_P (base) | |
f7b5f694 | 939 | && !may_be_aliased (base) |
53e9c5c4 | 940 | && !(VAR_P (base) |
f7b5f694 | 941 | && !DECL_EXTERNAL (base) |
942 | && !TREE_STATIC (base) | |
943 | && local_variable_can_escape (base))) | |
b4a708fb | 944 | return false; |
945 | return true; | |
946 | } | |
3072d30e | 947 | |
948 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
949 | OFFSET and WIDTH. */ | |
950 | ||
951 | static void | |
370ad826 | 952 | set_usage_bits (group_info *group, poly_int64 offset, poly_int64 width, |
b4a708fb | 953 | tree expr) |
3072d30e | 954 | { |
370ad826 | 955 | /* Non-constant offsets and widths act as global kills, so there's no point |
956 | trying to use them to derive global DSE candidates. */ | |
957 | HOST_WIDE_INT i, const_offset, const_width; | |
b4a708fb | 958 | bool expr_escapes = can_escape (expr); |
370ad826 | 959 | if (offset.is_constant (&const_offset) |
960 | && width.is_constant (&const_width) | |
961 | && const_offset > -MAX_OFFSET | |
962 | && const_offset + const_width < MAX_OFFSET) | |
963 | for (i = const_offset; i < const_offset + const_width; ++i) | |
3072d30e | 964 | { |
965 | bitmap store1; | |
966 | bitmap store2; | |
b4a708fb | 967 | bitmap escaped; |
3072d30e | 968 | int ai; |
969 | if (i < 0) | |
970 | { | |
971 | store1 = group->store1_n; | |
972 | store2 = group->store2_n; | |
b4a708fb | 973 | escaped = group->escaped_n; |
3072d30e | 974 | ai = -i; |
975 | } | |
976 | else | |
977 | { | |
978 | store1 = group->store1_p; | |
979 | store2 = group->store2_p; | |
b4a708fb | 980 | escaped = group->escaped_p; |
3072d30e | 981 | ai = i; |
982 | } | |
48e1416a | 983 | |
6ef9bbe0 | 984 | if (!bitmap_set_bit (store1, ai)) |
3072d30e | 985 | bitmap_set_bit (store2, ai); |
48e1416a | 986 | else |
3072d30e | 987 | { |
3072d30e | 988 | if (i < 0) |
989 | { | |
990 | if (group->offset_map_size_n < ai) | |
991 | group->offset_map_size_n = ai; | |
992 | } | |
993 | else | |
994 | { | |
995 | if (group->offset_map_size_p < ai) | |
996 | group->offset_map_size_p = ai; | |
997 | } | |
998 | } | |
b4a708fb | 999 | if (expr_escapes) |
1000 | bitmap_set_bit (escaped, ai); | |
3072d30e | 1001 | } |
1002 | } | |
1003 | ||
b4a708fb | 1004 | static void |
1005 | reset_active_stores (void) | |
1006 | { | |
1007 | active_local_stores = NULL; | |
1008 | active_local_stores_len = 0; | |
1009 | } | |
3072d30e | 1010 | |
b4a708fb | 1011 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
3072d30e | 1012 | |
1013 | static void | |
b4a708fb | 1014 | free_read_records (bb_info_t bb_info) |
3072d30e | 1015 | { |
1016 | insn_info_t insn_info = bb_info->last_insn; | |
1017 | read_info_t *ptr = &insn_info->read_rec; | |
3072d30e | 1018 | while (*ptr) |
1019 | { | |
1020 | read_info_t next = (*ptr)->next; | |
5a3f5d69 | 1021 | read_info_type_pool.remove (*ptr); |
1022 | *ptr = next; | |
3072d30e | 1023 | } |
b4a708fb | 1024 | } |
1025 | ||
1026 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1027 | ||
1028 | static void | |
1029 | add_wild_read (bb_info_t bb_info) | |
1030 | { | |
1031 | insn_info_t insn_info = bb_info->last_insn; | |
3072d30e | 1032 | insn_info->wild_read = true; |
b4a708fb | 1033 | free_read_records (bb_info); |
1034 | reset_active_stores (); | |
3072d30e | 1035 | } |
1036 | ||
b4a708fb | 1037 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1038 | non-frame locations. */ | |
1039 | ||
1040 | static void | |
1041 | add_non_frame_wild_read (bb_info_t bb_info) | |
1042 | { | |
1043 | insn_info_t insn_info = bb_info->last_insn; | |
1044 | insn_info->non_frame_wild_read = true; | |
1045 | free_read_records (bb_info); | |
1046 | reset_active_stores (); | |
1047 | } | |
3072d30e | 1048 | |
17e1318c | 1049 | /* Return true if X is a constant or one of the registers that behave |
1050 | as a constant over the life of a function. This is equivalent to | |
1051 | !rtx_varies_p for memory addresses. */ | |
3072d30e | 1052 | |
1053 | static bool | |
1054 | const_or_frame_p (rtx x) | |
1055 | { | |
0349edce | 1056 | if (CONSTANT_P (x)) |
1057 | return true; | |
1058 | ||
1059 | if (GET_CODE (x) == REG) | |
3072d30e | 1060 | { |
3072d30e | 1061 | /* Note that we have to test for the actual rtx used for the frame |
1062 | and arg pointers and not just the register number in case we have | |
1063 | eliminated the frame and/or arg pointer and are using it | |
1064 | for pseudos. */ | |
1065 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1066 | /* The arg pointer varies if it is not a fixed register. */ | |
1067 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1068 | || x == pic_offset_table_rtx) | |
1069 | return true; | |
1070 | return false; | |
3072d30e | 1071 | } |
55c5ac9f | 1072 | |
0349edce | 1073 | return false; |
3072d30e | 1074 | } |
1075 | ||
48e1416a | 1076 | /* Take all reasonable action to put the address of MEM into the form |
1077 | that we can do analysis on. | |
3072d30e | 1078 | |
1079 | The gold standard is to get the address into the form: address + | |
1080 | OFFSET where address is something that rtx_varies_p considers a | |
1081 | constant. When we can get the address in this form, we can do | |
1082 | global analysis on it. Note that for constant bases, address is | |
1083 | not actually returned, only the group_id. The address can be | |
1084 | obtained from that. | |
1085 | ||
1086 | If that fails, we try cselib to get a value we can at least use | |
48e1416a | 1087 | locally. If that fails we return false. |
1088 | ||
3072d30e | 1089 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1090 | group for non_varying bases. | |
1091 | ||
1092 | FOR_READ is true if this is a mem read and false if not. */ | |
1093 | ||
1094 | static bool | |
1095 | canon_address (rtx mem, | |
3072d30e | 1096 | int *group_id, |
370ad826 | 1097 | poly_int64 *offset, |
3072d30e | 1098 | cselib_val **base) |
1099 | { | |
3754d046 | 1100 | machine_mode address_mode = get_address_mode (mem); |
3072d30e | 1101 | rtx mem_address = XEXP (mem, 0); |
1102 | rtx expanded_address, address; | |
627540ce | 1103 | int expanded; |
1104 | ||
1f864115 | 1105 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
3072d30e | 1106 | |
1ca59310 | 1107 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1108 | { |
1109 | fprintf (dump_file, " mem: "); | |
1110 | print_inline_rtx (dump_file, mem_address, 0); | |
1111 | fprintf (dump_file, "\n"); | |
1112 | } | |
1113 | ||
627540ce | 1114 | /* First see if just canon_rtx (mem_address) is const or frame, |
1115 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1116 | address = NULL_RTX; | |
1117 | for (expanded = 0; expanded < 2; expanded++) | |
1118 | { | |
1119 | if (expanded) | |
1120 | { | |
1121 | /* Use cselib to replace all of the reg references with the full | |
48e1416a | 1122 | expression. This will take care of the case where we have |
3072d30e | 1123 | |
627540ce | 1124 | r_x = base + offset; |
1125 | val = *r_x; | |
48e1416a | 1126 | |
1127 | by making it into | |
3072d30e | 1128 | |
627540ce | 1129 | val = *(base + offset); */ |
3072d30e | 1130 | |
627540ce | 1131 | expanded_address = cselib_expand_value_rtx (mem_address, |
1132 | scratch, 5); | |
3072d30e | 1133 | |
627540ce | 1134 | /* If this fails, just go with the address from first |
1135 | iteration. */ | |
1136 | if (!expanded_address) | |
1137 | break; | |
1138 | } | |
1139 | else | |
1140 | expanded_address = mem_address; | |
3072d30e | 1141 | |
627540ce | 1142 | /* Split the address into canonical BASE + OFFSET terms. */ |
1143 | address = canon_rtx (expanded_address); | |
3072d30e | 1144 | |
627540ce | 1145 | *offset = 0; |
3072d30e | 1146 | |
1ca59310 | 1147 | if (dump_file && (dump_flags & TDF_DETAILS)) |
627540ce | 1148 | { |
1149 | if (expanded) | |
1150 | { | |
1151 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1152 | print_inline_rtx (dump_file, expanded_address, 0); | |
1153 | fprintf (dump_file, "\n"); | |
1154 | } | |
3072d30e | 1155 | |
627540ce | 1156 | fprintf (dump_file, "\n after canon_rtx address: "); |
1157 | print_inline_rtx (dump_file, address, 0); | |
1158 | fprintf (dump_file, "\n"); | |
1159 | } | |
3072d30e | 1160 | |
627540ce | 1161 | if (GET_CODE (address) == CONST) |
1162 | address = XEXP (address, 0); | |
3072d30e | 1163 | |
370ad826 | 1164 | address = strip_offset_and_add (address, offset); |
3072d30e | 1165 | |
bd1a81f7 | 1166 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1167 | && const_or_frame_p (address)) | |
3072d30e | 1168 | { |
c2eab334 | 1169 | group_info *group = get_group_info (address); |
627540ce | 1170 | |
1ca59310 | 1171 | if (dump_file && (dump_flags & TDF_DETAILS)) |
370ad826 | 1172 | { |
1173 | fprintf (dump_file, " gid=%d offset=", group->id); | |
1174 | print_dec (*offset, dump_file); | |
1175 | fprintf (dump_file, "\n"); | |
1176 | } | |
627540ce | 1177 | *base = NULL; |
1178 | *group_id = group->id; | |
1179 | return true; | |
3072d30e | 1180 | } |
627540ce | 1181 | } |
1182 | ||
1f864115 | 1183 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
627540ce | 1184 | *group_id = -1; |
1185 | ||
1186 | if (*base == NULL) | |
1187 | { | |
1ca59310 | 1188 | if (dump_file && (dump_flags & TDF_DETAILS)) |
627540ce | 1189 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1190 | return false; | |
3072d30e | 1191 | } |
1ca59310 | 1192 | if (dump_file && (dump_flags & TDF_DETAILS)) |
370ad826 | 1193 | { |
1194 | fprintf (dump_file, " varying cselib base=%u:%u offset = ", | |
1195 | (*base)->uid, (*base)->hash); | |
1196 | print_dec (*offset, dump_file); | |
1197 | fprintf (dump_file, "\n"); | |
1198 | } | |
3072d30e | 1199 | return true; |
1200 | } | |
1201 | ||
1202 | ||
1203 | /* Clear the rhs field from the active_local_stores array. */ | |
1204 | ||
1205 | static void | |
1206 | clear_rhs_from_active_local_stores (void) | |
1207 | { | |
1208 | insn_info_t ptr = active_local_stores; | |
1209 | ||
1210 | while (ptr) | |
1211 | { | |
c2eab334 | 1212 | store_info *store_info = ptr->store_rec; |
3072d30e | 1213 | /* Skip the clobbers. */ |
1214 | while (!store_info->is_set) | |
1215 | store_info = store_info->next; | |
1216 | ||
1217 | store_info->rhs = NULL; | |
aa140b76 | 1218 | store_info->const_rhs = NULL; |
3072d30e | 1219 | |
1220 | ptr = ptr->next_local_store; | |
1221 | } | |
1222 | } | |
1223 | ||
1224 | ||
aa140b76 | 1225 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1226 | ||
1227 | static inline void | |
c2eab334 | 1228 | set_position_unneeded (store_info *s_info, int pos) |
aa140b76 | 1229 | { |
1230 | if (__builtin_expect (s_info->is_large, false)) | |
1231 | { | |
6ef9bbe0 | 1232 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1233 | s_info->positions_needed.large.count++; | |
aa140b76 | 1234 | } |
1235 | else | |
1236 | s_info->positions_needed.small_bitmask | |
edc19fd0 | 1237 | &= ~(HOST_WIDE_INT_1U << pos); |
aa140b76 | 1238 | } |
1239 | ||
1240 | /* Mark the whole store S_INFO as unneeded. */ | |
1241 | ||
1242 | static inline void | |
c2eab334 | 1243 | set_all_positions_unneeded (store_info *s_info) |
aa140b76 | 1244 | { |
1245 | if (__builtin_expect (s_info->is_large, false)) | |
1246 | { | |
370ad826 | 1247 | HOST_WIDE_INT width; |
1248 | if (s_info->width.is_constant (&width)) | |
1249 | { | |
1250 | bitmap_set_range (s_info->positions_needed.large.bmap, 0, width); | |
1251 | s_info->positions_needed.large.count = width; | |
1252 | } | |
1253 | else | |
1254 | { | |
1255 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1256 | s_info->positions_needed.large.count = 1; | |
1257 | } | |
aa140b76 | 1258 | } |
1259 | else | |
457daa4e | 1260 | s_info->positions_needed.small_bitmask = HOST_WIDE_INT_0U; |
aa140b76 | 1261 | } |
1262 | ||
1263 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1264 | ||
1265 | static inline bool | |
c2eab334 | 1266 | any_positions_needed_p (store_info *s_info) |
aa140b76 | 1267 | { |
1268 | if (__builtin_expect (s_info->is_large, false)) | |
370ad826 | 1269 | { |
1270 | HOST_WIDE_INT width; | |
1271 | if (s_info->width.is_constant (&width)) | |
1272 | { | |
1273 | gcc_checking_assert (s_info->positions_needed.large.bmap); | |
1274 | return s_info->positions_needed.large.count < width; | |
1275 | } | |
1276 | else | |
1277 | { | |
1278 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1279 | return s_info->positions_needed.large.count == 0; | |
1280 | } | |
1281 | } | |
aa140b76 | 1282 | else |
457daa4e | 1283 | return (s_info->positions_needed.small_bitmask != HOST_WIDE_INT_0U); |
aa140b76 | 1284 | } |
1285 | ||
1286 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
370ad826 | 1287 | store are known to be needed. */ |
aa140b76 | 1288 | |
1289 | static inline bool | |
370ad826 | 1290 | all_positions_needed_p (store_info *s_info, poly_int64 start, |
1291 | poly_int64 width) | |
aa140b76 | 1292 | { |
370ad826 | 1293 | gcc_assert (s_info->rhs); |
1294 | if (!s_info->width.is_constant ()) | |
1295 | { | |
1296 | gcc_assert (s_info->is_large | |
1297 | && !s_info->positions_needed.large.bmap); | |
1298 | return s_info->positions_needed.large.count == 0; | |
1299 | } | |
1300 | ||
1301 | /* Otherwise, if START and WIDTH are non-constant, we're asking about | |
1302 | a non-constant region of a constant-sized store. We can't say for | |
1303 | sure that all positions are needed. */ | |
1304 | HOST_WIDE_INT const_start, const_width; | |
1305 | if (!start.is_constant (&const_start) | |
1306 | || !width.is_constant (&const_width)) | |
1307 | return false; | |
1308 | ||
aa140b76 | 1309 | if (__builtin_expect (s_info->is_large, false)) |
1310 | { | |
370ad826 | 1311 | for (HOST_WIDE_INT i = const_start; i < const_start + const_width; ++i) |
1312 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, i)) | |
aa140b76 | 1313 | return false; |
1314 | return true; | |
1315 | } | |
1316 | else | |
1317 | { | |
370ad826 | 1318 | unsigned HOST_WIDE_INT mask |
1319 | = lowpart_bitmask (const_width) << const_start; | |
aa140b76 | 1320 | return (s_info->positions_needed.small_bitmask & mask) == mask; |
1321 | } | |
1322 | } | |
1323 | ||
1324 | ||
370ad826 | 1325 | static rtx get_stored_val (store_info *, machine_mode, poly_int64, |
1326 | poly_int64, basic_block, bool); | |
aa140b76 | 1327 | |
1328 | ||
3072d30e | 1329 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1330 | there is a candidate store, after adding it to the appropriate | |
1331 | local store group if so. */ | |
1332 | ||
1333 | static int | |
1334 | record_store (rtx body, bb_info_t bb_info) | |
1335 | { | |
82d2c88b | 1336 | rtx mem, rhs, const_rhs, mem_addr; |
370ad826 | 1337 | poly_int64 offset = 0; |
1338 | poly_int64 width = 0; | |
3072d30e | 1339 | insn_info_t insn_info = bb_info->last_insn; |
c2eab334 | 1340 | store_info *store_info = NULL; |
3072d30e | 1341 | int group_id; |
1342 | cselib_val *base = NULL; | |
aa140b76 | 1343 | insn_info_t ptr, last, redundant_reason; |
3072d30e | 1344 | bool store_is_unused; |
1345 | ||
1346 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1347 | return 0; | |
1348 | ||
aa140b76 | 1349 | mem = SET_DEST (body); |
1350 | ||
3072d30e | 1351 | /* If this is not used, then this cannot be used to keep the insn |
1352 | from being deleted. On the other hand, it does provide something | |
1353 | that can be used to prove that another store is dead. */ | |
1354 | store_is_unused | |
aa140b76 | 1355 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
3072d30e | 1356 | |
1357 | /* Check whether that value is a suitable memory location. */ | |
3072d30e | 1358 | if (!MEM_P (mem)) |
1359 | { | |
1360 | /* If the set or clobber is unused, then it does not effect our | |
1361 | ability to get rid of the entire insn. */ | |
1362 | if (!store_is_unused) | |
1363 | insn_info->cannot_delete = true; | |
1364 | return 0; | |
1365 | } | |
1366 | ||
1367 | /* At this point we know mem is a mem. */ | |
1368 | if (GET_MODE (mem) == BLKmode) | |
1369 | { | |
711f137f | 1370 | HOST_WIDE_INT const_size; |
3072d30e | 1371 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) |
1372 | { | |
1ca59310 | 1373 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 1374 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1375 | add_wild_read (bb_info); | |
1376 | insn_info->cannot_delete = true; | |
aa140b76 | 1377 | return 0; |
3072d30e | 1378 | } |
aa140b76 | 1379 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1380 | as memset (addr, 0, 36); */ | |
5b2a69fa | 1381 | else if (!MEM_SIZE_KNOWN_P (mem) |
711f137f | 1382 | || maybe_le (MEM_SIZE (mem), 0) |
1383 | /* This is a limit on the bitmap size, which is only relevant | |
1384 | for constant-sized MEMs. */ | |
1385 | || (MEM_SIZE (mem).is_constant (&const_size) | |
1386 | && const_size > MAX_OFFSET) | |
aa140b76 | 1387 | || GET_CODE (body) != SET |
aa140b76 | 1388 | || !CONST_INT_P (SET_SRC (body))) |
3072d30e | 1389 | { |
aa140b76 | 1390 | if (!store_is_unused) |
1391 | { | |
1392 | /* If the set or clobber is unused, then it does not effect our | |
1393 | ability to get rid of the entire insn. */ | |
1394 | insn_info->cannot_delete = true; | |
1395 | clear_rhs_from_active_local_stores (); | |
1396 | } | |
1397 | return 0; | |
3072d30e | 1398 | } |
3072d30e | 1399 | } |
1400 | ||
1401 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1402 | if (MEM_VOLATILE_P (mem)) | |
aa140b76 | 1403 | insn_info->cannot_delete = true; |
3072d30e | 1404 | |
5a3f5d69 | 1405 | if (!canon_address (mem, &group_id, &offset, &base)) |
3072d30e | 1406 | { |
1407 | clear_rhs_from_active_local_stores (); | |
1408 | return 0; | |
1409 | } | |
1410 | ||
aa140b76 | 1411 | if (GET_MODE (mem) == BLKmode) |
5b2a69fa | 1412 | width = MEM_SIZE (mem); |
aa140b76 | 1413 | else |
e4209874 | 1414 | width = GET_MODE_SIZE (GET_MODE (mem)); |
3072d30e | 1415 | |
370ad826 | 1416 | if (!endpoint_representable_p (offset, width)) |
8f4e0f7b | 1417 | { |
1418 | clear_rhs_from_active_local_stores (); | |
1419 | return 0; | |
1420 | } | |
1421 | ||
269483bb | 1422 | if (known_eq (width, 0)) |
1423 | return 0; | |
1424 | ||
5a3f5d69 | 1425 | if (group_id >= 0) |
3072d30e | 1426 | { |
1427 | /* In the restrictive case where the base is a constant or the | |
1428 | frame pointer we can do global analysis. */ | |
48e1416a | 1429 | |
c2eab334 | 1430 | group_info *group |
f1f41a6c | 1431 | = rtx_group_vec[group_id]; |
b4a708fb | 1432 | tree expr = MEM_EXPR (mem); |
48e1416a | 1433 | |
55c5ac9f | 1434 | store_info = rtx_store_info_pool.allocate (); |
b4a708fb | 1435 | set_usage_bits (group, offset, width, expr); |
3072d30e | 1436 | |
1ca59310 | 1437 | if (dump_file && (dump_flags & TDF_DETAILS)) |
57bca8dc | 1438 | { |
1439 | fprintf (dump_file, " processing const base store gid=%d", | |
1440 | group_id); | |
1441 | print_range (dump_file, offset, width); | |
1442 | fprintf (dump_file, "\n"); | |
1443 | } | |
3072d30e | 1444 | } |
1445 | else | |
1446 | { | |
86e87ef6 | 1447 | if (may_be_sp_based_p (XEXP (mem, 0))) |
17e1318c | 1448 | insn_info->stack_pointer_based = true; |
3072d30e | 1449 | insn_info->contains_cselib_groups = true; |
17e1318c | 1450 | |
55c5ac9f | 1451 | store_info = cse_store_info_pool.allocate (); |
3072d30e | 1452 | group_id = -1; |
1453 | ||
1ca59310 | 1454 | if (dump_file && (dump_flags & TDF_DETAILS)) |
57bca8dc | 1455 | { |
1456 | fprintf (dump_file, " processing cselib store "); | |
1457 | print_range (dump_file, offset, width); | |
1458 | fprintf (dump_file, "\n"); | |
1459 | } | |
3072d30e | 1460 | } |
1461 | ||
aa140b76 | 1462 | const_rhs = rhs = NULL_RTX; |
1463 | if (GET_CODE (body) == SET | |
1464 | /* No place to keep the value after ra. */ | |
1465 | && !reload_completed | |
1466 | && (REG_P (SET_SRC (body)) | |
1467 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1468 | || CONSTANT_P (SET_SRC (body))) | |
1469 | && !MEM_VOLATILE_P (mem) | |
1470 | /* Sometimes the store and reload is used for truncation and | |
1471 | rounding. */ | |
1472 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1473 | { | |
1474 | rhs = SET_SRC (body); | |
1475 | if (CONSTANT_P (rhs)) | |
1476 | const_rhs = rhs; | |
1477 | else if (body == PATTERN (insn_info->insn)) | |
1478 | { | |
1479 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1480 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1481 | const_rhs = XEXP (tem, 0); | |
1482 | } | |
1483 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1484 | { | |
1485 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1486 | ||
1487 | if (tem && CONSTANT_P (tem)) | |
1488 | const_rhs = tem; | |
1489 | } | |
1490 | } | |
1491 | ||
3072d30e | 1492 | /* Check to see if this stores causes some other stores to be |
1493 | dead. */ | |
1494 | ptr = active_local_stores; | |
1495 | last = NULL; | |
aa140b76 | 1496 | redundant_reason = NULL; |
82d2c88b | 1497 | mem = canon_rtx (mem); |
5a3f5d69 | 1498 | |
1499 | if (group_id < 0) | |
1500 | mem_addr = base->val_rtx; | |
82d2c88b | 1501 | else |
1502 | { | |
5a3f5d69 | 1503 | group_info *group = rtx_group_vec[group_id]; |
1504 | mem_addr = group->canon_base_addr; | |
82d2c88b | 1505 | } |
370ad826 | 1506 | if (maybe_ne (offset, 0)) |
5a3f5d69 | 1507 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
3072d30e | 1508 | |
1509 | while (ptr) | |
1510 | { | |
1511 | insn_info_t next = ptr->next_local_store; | |
2e966e2a | 1512 | class store_info *s_info = ptr->store_rec; |
9ce37fa7 | 1513 | bool del = true; |
3072d30e | 1514 | |
1515 | /* Skip the clobbers. We delete the active insn if this insn | |
6dfdc153 | 1516 | shadows the set. To have been put on the active list, it |
3072d30e | 1517 | has exactly on set. */ |
1518 | while (!s_info->is_set) | |
1519 | s_info = s_info->next; | |
1520 | ||
5a3f5d69 | 1521 | if (s_info->group_id == group_id && s_info->cse_base == base) |
3072d30e | 1522 | { |
1523 | HOST_WIDE_INT i; | |
1ca59310 | 1524 | if (dump_file && (dump_flags & TDF_DETAILS)) |
57bca8dc | 1525 | { |
1526 | fprintf (dump_file, " trying store in insn=%d gid=%d", | |
1527 | INSN_UID (ptr->insn), s_info->group_id); | |
1528 | print_range (dump_file, s_info->offset, s_info->width); | |
1529 | fprintf (dump_file, "\n"); | |
1530 | } | |
aa140b76 | 1531 | |
1532 | /* Even if PTR won't be eliminated as unneeded, if both | |
1533 | PTR and this insn store the same constant value, we might | |
1534 | eliminate this insn instead. */ | |
1535 | if (s_info->const_rhs | |
1536 | && const_rhs | |
57bca8dc | 1537 | && known_subrange_p (offset, width, |
1538 | s_info->offset, s_info->width) | |
1539 | && all_positions_needed_p (s_info, offset - s_info->offset, | |
be292237 | 1540 | width) |
1541 | /* We can only remove the later store if the earlier aliases | |
1542 | at least all accesses the later one. */ | |
1543 | && (MEM_ALIAS_SET (mem) == MEM_ALIAS_SET (s_info->mem) | |
1544 | || alias_set_subset_of (MEM_ALIAS_SET (mem), | |
1545 | MEM_ALIAS_SET (s_info->mem)))) | |
aa140b76 | 1546 | { |
1547 | if (GET_MODE (mem) == BLKmode) | |
1548 | { | |
1549 | if (GET_MODE (s_info->mem) == BLKmode | |
1550 | && s_info->const_rhs == const_rhs) | |
1551 | redundant_reason = ptr; | |
1552 | } | |
1553 | else if (s_info->const_rhs == const0_rtx | |
1554 | && const_rhs == const0_rtx) | |
1555 | redundant_reason = ptr; | |
1556 | else | |
1557 | { | |
1558 | rtx val; | |
1559 | start_sequence (); | |
57bca8dc | 1560 | val = get_stored_val (s_info, GET_MODE (mem), offset, width, |
aa140b76 | 1561 | BLOCK_FOR_INSN (insn_info->insn), |
1562 | true); | |
1563 | if (get_insns () != NULL) | |
1564 | val = NULL_RTX; | |
1565 | end_sequence (); | |
1566 | if (val && rtx_equal_p (val, const_rhs)) | |
1567 | redundant_reason = ptr; | |
1568 | } | |
1569 | } | |
1570 | ||
370ad826 | 1571 | HOST_WIDE_INT begin_unneeded, const_s_width, const_width; |
57bca8dc | 1572 | if (known_subrange_p (s_info->offset, s_info->width, offset, width)) |
1573 | /* The new store touches every byte that S_INFO does. */ | |
1574 | set_all_positions_unneeded (s_info); | |
370ad826 | 1575 | else if ((offset - s_info->offset).is_constant (&begin_unneeded) |
1576 | && s_info->width.is_constant (&const_s_width) | |
1577 | && width.is_constant (&const_width)) | |
57bca8dc | 1578 | { |
370ad826 | 1579 | HOST_WIDE_INT end_unneeded = begin_unneeded + const_width; |
57bca8dc | 1580 | begin_unneeded = MAX (begin_unneeded, 0); |
370ad826 | 1581 | end_unneeded = MIN (end_unneeded, const_s_width); |
57bca8dc | 1582 | for (i = begin_unneeded; i < end_unneeded; ++i) |
1583 | set_position_unneeded (s_info, i); | |
1584 | } | |
370ad826 | 1585 | else |
1586 | { | |
1587 | /* We don't know which parts of S_INFO are needed and | |
1588 | which aren't, so invalidate the RHS. */ | |
1589 | s_info->rhs = NULL; | |
1590 | s_info->const_rhs = NULL; | |
1591 | } | |
3072d30e | 1592 | } |
1593 | else if (s_info->rhs) | |
1594 | /* Need to see if it is possible for this store to overwrite | |
1595 | the value of store_info. If it is, set the rhs to NULL to | |
1596 | keep it from being used to remove a load. */ | |
1597 | { | |
68ab33a6 | 1598 | if (canon_output_dependence (s_info->mem, true, |
1599 | mem, GET_MODE (mem), | |
1600 | mem_addr)) | |
aa140b76 | 1601 | { |
1602 | s_info->rhs = NULL; | |
1603 | s_info->const_rhs = NULL; | |
1604 | } | |
3072d30e | 1605 | } |
82d2c88b | 1606 | |
3072d30e | 1607 | /* An insn can be deleted if every position of every one of |
1608 | its s_infos is zero. */ | |
c75be2fe | 1609 | if (any_positions_needed_p (s_info)) |
9ce37fa7 | 1610 | del = false; |
aa140b76 | 1611 | |
9ce37fa7 | 1612 | if (del) |
3072d30e | 1613 | { |
1614 | insn_info_t insn_to_delete = ptr; | |
48e1416a | 1615 | |
1242bee6 | 1616 | active_local_stores_len--; |
3072d30e | 1617 | if (last) |
1618 | last->next_local_store = ptr->next_local_store; | |
1619 | else | |
1620 | active_local_stores = ptr->next_local_store; | |
48e1416a | 1621 | |
c75be2fe | 1622 | if (!insn_to_delete->cannot_delete) |
1623 | delete_dead_store_insn (insn_to_delete); | |
3072d30e | 1624 | } |
1625 | else | |
1626 | last = ptr; | |
48e1416a | 1627 | |
3072d30e | 1628 | ptr = next; |
1629 | } | |
48e1416a | 1630 | |
3072d30e | 1631 | /* Finish filling in the store_info. */ |
1632 | store_info->next = insn_info->store_rec; | |
1633 | insn_info->store_rec = store_info; | |
82d2c88b | 1634 | store_info->mem = mem; |
82d2c88b | 1635 | store_info->mem_addr = mem_addr; |
3072d30e | 1636 | store_info->cse_base = base; |
370ad826 | 1637 | HOST_WIDE_INT const_width; |
1638 | if (!width.is_constant (&const_width)) | |
1639 | { | |
1640 | store_info->is_large = true; | |
1641 | store_info->positions_needed.large.count = 0; | |
1642 | store_info->positions_needed.large.bmap = NULL; | |
1643 | } | |
1644 | else if (const_width > HOST_BITS_PER_WIDE_INT) | |
aa140b76 | 1645 | { |
1646 | store_info->is_large = true; | |
1647 | store_info->positions_needed.large.count = 0; | |
4fb07d00 | 1648 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
aa140b76 | 1649 | } |
1650 | else | |
1651 | { | |
1652 | store_info->is_large = false; | |
370ad826 | 1653 | store_info->positions_needed.small_bitmask |
1654 | = lowpart_bitmask (const_width); | |
aa140b76 | 1655 | } |
3072d30e | 1656 | store_info->group_id = group_id; |
57bca8dc | 1657 | store_info->offset = offset; |
1658 | store_info->width = width; | |
3072d30e | 1659 | store_info->is_set = GET_CODE (body) == SET; |
aa140b76 | 1660 | store_info->rhs = rhs; |
1661 | store_info->const_rhs = const_rhs; | |
1662 | store_info->redundant_reason = redundant_reason; | |
3072d30e | 1663 | |
3072d30e | 1664 | /* If this is a clobber, we return 0. We will only be able to |
1665 | delete this insn if there is only one store USED store, but we | |
1666 | can use the clobber to delete other stores earlier. */ | |
1667 | return store_info->is_set ? 1 : 0; | |
1668 | } | |
1669 | ||
1670 | ||
1671 | static void | |
1672 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1673 | { | |
48e1416a | 1674 | fprintf (dump_file, "%s insn=%d %s\n", start, |
3072d30e | 1675 | INSN_UID (insn_info->insn), |
1676 | insn_info->store_rec ? "has store" : "naked"); | |
1677 | } | |
1678 | ||
1679 | ||
5c9051a4 | 1680 | /* If the modes are different and the value's source and target do not |
1681 | line up, we need to extract the value from lower part of the rhs of | |
1682 | the store, shift it, and then put it into a form that can be shoved | |
1683 | into the read_insn. This function generates a right SHIFT of a | |
1684 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1685 | shift sequence is returned or NULL if we failed to find a | |
1686 | shift. */ | |
1687 | ||
1688 | static rtx | |
370ad826 | 1689 | find_shift_sequence (poly_int64 access_size, |
c2eab334 | 1690 | store_info *store_info, |
3754d046 | 1691 | machine_mode read_mode, |
370ad826 | 1692 | poly_int64 shift, bool speed, bool require_cst) |
5c9051a4 | 1693 | { |
3754d046 | 1694 | machine_mode store_mode = GET_MODE (store_info->mem); |
1a5d4b27 | 1695 | scalar_int_mode new_mode; |
10d4de0e | 1696 | rtx read_reg = NULL; |
5c9051a4 | 1697 | |
1698 | /* Some machines like the x86 have shift insns for each size of | |
1699 | operand. Other machines like the ppc or the ia-64 may only have | |
1700 | shift insns that shift values within 32 or 64 bit registers. | |
1701 | This loop tries to find the smallest shift insn that will right | |
1702 | justify the value we want to read but is available in one insn on | |
1703 | the machine. */ | |
1704 | ||
1a5d4b27 | 1705 | opt_scalar_int_mode new_mode_iter; |
1706 | FOR_EACH_MODE_FROM (new_mode_iter, | |
1707 | smallest_int_mode_for_size (access_size * BITS_PER_UNIT)) | |
5c9051a4 | 1708 | { |
4cd001d5 | 1709 | rtx target, new_reg, new_lhs; |
1710 | rtx_insn *shift_seq, *insn; | |
171557e8 | 1711 | int cost; |
af97461e | 1712 | |
1a5d4b27 | 1713 | new_mode = new_mode_iter.require (); |
19a4dce4 | 1714 | if (GET_MODE_BITSIZE (new_mode) > BITS_PER_WORD) |
1715 | break; | |
1716 | ||
4ed4afb9 | 1717 | /* If a constant was stored into memory, try to simplify it here, |
1718 | otherwise the cost of the shift might preclude this optimization | |
1719 | e.g. at -Os, even when no actual shift will be needed. */ | |
aa140b76 | 1720 | if (store_info->const_rhs) |
4ed4afb9 | 1721 | { |
9edf7ea8 | 1722 | poly_uint64 byte = subreg_lowpart_offset (new_mode, store_mode); |
aa140b76 | 1723 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1724 | store_mode, byte); | |
4ed4afb9 | 1725 | if (ret && CONSTANT_P (ret)) |
1726 | { | |
bd39703a | 1727 | rtx shift_rtx = gen_int_shift_amount (new_mode, shift); |
4ed4afb9 | 1728 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, |
bd39703a | 1729 | ret, shift_rtx); |
4ed4afb9 | 1730 | if (ret && CONSTANT_P (ret)) |
1731 | { | |
1732 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1733 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1734 | if (ret && CONSTANT_P (ret) | |
5ae4887d | 1735 | && (set_src_cost (ret, read_mode, speed) |
1736 | <= COSTS_N_INSNS (1))) | |
4ed4afb9 | 1737 | return ret; |
1738 | } | |
1739 | } | |
1740 | } | |
1741 | ||
aa140b76 | 1742 | if (require_cst) |
1743 | return NULL_RTX; | |
1744 | ||
10d4de0e | 1745 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1746 | requires a real instruction. */ | |
eafbcd13 | 1747 | if (maybe_lt (GET_MODE_SIZE (new_mode), GET_MODE_SIZE (store_mode)) |
396f2130 | 1748 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
af97461e | 1749 | continue; |
1750 | ||
10d4de0e | 1751 | /* Also try a wider mode if the necessary punning is either not |
1752 | desirable or not possible. */ | |
1753 | if (!CONSTANT_P (store_info->rhs) | |
5f6dcf1a | 1754 | && !targetm.modes_tieable_p (new_mode, store_mode)) |
10d4de0e | 1755 | continue; |
10d4de0e | 1756 | |
af97461e | 1757 | new_reg = gen_reg_rtx (new_mode); |
5c9051a4 | 1758 | |
1759 | start_sequence (); | |
1760 | ||
1761 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1762 | of one dsp where the cost of these two was not the same. But | |
1763 | this really is a rare case anyway. */ | |
1764 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
bd39703a | 1765 | gen_int_shift_amount (new_mode, shift), |
1766 | new_reg, 1, OPTAB_DIRECT); | |
5c9051a4 | 1767 | |
597d5470 | 1768 | shift_seq = get_insns (); |
1769 | end_sequence (); | |
5c9051a4 | 1770 | |
597d5470 | 1771 | if (target != new_reg || shift_seq == NULL) |
1772 | continue; | |
1773 | ||
1774 | cost = 0; | |
1775 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1776 | if (INSN_P (insn)) | |
3bdb5a4d | 1777 | cost += insn_cost (insn, speed); |
597d5470 | 1778 | |
1779 | /* The computation up to here is essentially independent | |
1780 | of the arguments and could be precomputed. It may | |
1781 | not be worth doing so. We could precompute if | |
1782 | worthwhile or at least cache the results. The result | |
5133fd21 | 1783 | technically depends on both SHIFT and ACCESS_SIZE, |
1784 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
597d5470 | 1785 | |
1786 | if (cost > COSTS_N_INSNS (1)) | |
1787 | continue; | |
1788 | ||
171557e8 | 1789 | new_lhs = extract_low_bits (new_mode, store_mode, |
1790 | copy_rtx (store_info->rhs)); | |
1791 | if (new_lhs == NULL_RTX) | |
1792 | continue; | |
1793 | ||
597d5470 | 1794 | /* We found an acceptable shift. Generate a move to |
1795 | take the value from the store and put it into the | |
1796 | shift pseudo, then shift it, then generate another | |
1797 | move to put in into the target of the read. */ | |
10d4de0e | 1798 | emit_move_insn (new_reg, new_lhs); |
597d5470 | 1799 | emit_insn (shift_seq); |
10d4de0e | 1800 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
597d5470 | 1801 | break; |
5c9051a4 | 1802 | } |
1803 | ||
10d4de0e | 1804 | return read_reg; |
5c9051a4 | 1805 | } |
1806 | ||
1807 | ||
a1b0a968 | 1808 | /* Call back for note_stores to find the hard regs set or clobbered by |
1809 | insn. Data is a bitmap of the hardregs set so far. */ | |
1810 | ||
1811 | static void | |
1812 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1813 | { | |
1814 | bitmap regs_set = (bitmap) data; | |
1815 | ||
1816 | if (REG_P (x) | |
771d4616 | 1817 | && HARD_REGISTER_P (x)) |
0933f1d9 | 1818 | bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); |
a1b0a968 | 1819 | } |
1820 | ||
aa140b76 | 1821 | /* Helper function for replace_read and record_store. |
57bca8dc | 1822 | Attempt to return a value of mode READ_MODE stored in STORE_INFO, |
1823 | consisting of READ_WIDTH bytes starting from READ_OFFSET. Return NULL | |
aa140b76 | 1824 | if not successful. If REQUIRE_CST is true, return always constant. */ |
1825 | ||
1826 | static rtx | |
c2eab334 | 1827 | get_stored_val (store_info *store_info, machine_mode read_mode, |
370ad826 | 1828 | poly_int64 read_offset, poly_int64 read_width, |
aa140b76 | 1829 | basic_block bb, bool require_cst) |
1830 | { | |
3754d046 | 1831 | machine_mode store_mode = GET_MODE (store_info->mem); |
370ad826 | 1832 | poly_int64 gap; |
aa140b76 | 1833 | rtx read_reg; |
1834 | ||
1835 | /* To get here the read is within the boundaries of the write so | |
1836 | shift will never be negative. Start out with the shift being in | |
1837 | bytes. */ | |
1838 | if (store_mode == BLKmode) | |
57bca8dc | 1839 | gap = 0; |
aa140b76 | 1840 | else if (BYTES_BIG_ENDIAN) |
57bca8dc | 1841 | gap = ((store_info->offset + store_info->width) |
1842 | - (read_offset + read_width)); | |
aa140b76 | 1843 | else |
57bca8dc | 1844 | gap = read_offset - store_info->offset; |
aa140b76 | 1845 | |
ba0d8d7a | 1846 | if (gap.is_constant () && maybe_ne (gap, 0)) |
57bca8dc | 1847 | { |
370ad826 | 1848 | poly_int64 shift = gap * BITS_PER_UNIT; |
1849 | poly_int64 access_size = GET_MODE_SIZE (read_mode) + gap; | |
57bca8dc | 1850 | read_reg = find_shift_sequence (access_size, store_info, read_mode, |
1851 | shift, optimize_bb_for_speed_p (bb), | |
1852 | require_cst); | |
1853 | } | |
aa140b76 | 1854 | else if (store_mode == BLKmode) |
1855 | { | |
1856 | /* The store is a memset (addr, const_val, const_size). */ | |
1857 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
2cf1bb25 | 1858 | scalar_int_mode int_store_mode; |
1859 | if (!int_mode_for_mode (read_mode).exists (&int_store_mode)) | |
aa140b76 | 1860 | read_reg = NULL_RTX; |
1861 | else if (store_info->rhs == const0_rtx) | |
2cf1bb25 | 1862 | read_reg = extract_low_bits (read_mode, int_store_mode, const0_rtx); |
1863 | else if (GET_MODE_BITSIZE (int_store_mode) > HOST_BITS_PER_WIDE_INT | |
aa140b76 | 1864 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) |
1865 | read_reg = NULL_RTX; | |
1866 | else | |
1867 | { | |
1868 | unsigned HOST_WIDE_INT c | |
1869 | = INTVAL (store_info->rhs) | |
edc19fd0 | 1870 | & ((HOST_WIDE_INT_1 << BITS_PER_UNIT) - 1); |
aa140b76 | 1871 | int shift = BITS_PER_UNIT; |
1872 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1873 | { | |
1874 | c |= (c << shift); | |
1875 | shift <<= 1; | |
1876 | } | |
2cf1bb25 | 1877 | read_reg = gen_int_mode (c, int_store_mode); |
1878 | read_reg = extract_low_bits (read_mode, int_store_mode, read_reg); | |
aa140b76 | 1879 | } |
1880 | } | |
1881 | else if (store_info->const_rhs | |
1882 | && (require_cst | |
1883 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1884 | read_reg = extract_low_bits (read_mode, store_mode, | |
1885 | copy_rtx (store_info->const_rhs)); | |
1886 | else | |
1887 | read_reg = extract_low_bits (read_mode, store_mode, | |
1888 | copy_rtx (store_info->rhs)); | |
1889 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1890 | read_reg = NULL_RTX; | |
1891 | return read_reg; | |
1892 | } | |
a1b0a968 | 1893 | |
3072d30e | 1894 | /* Take a sequence of: |
1895 | A <- r1 | |
1896 | ... | |
1897 | ... <- A | |
1898 | ||
48e1416a | 1899 | and change it into |
3072d30e | 1900 | r2 <- r1 |
1901 | A <- r1 | |
1902 | ... | |
1903 | ... <- r2 | |
1904 | ||
5c9051a4 | 1905 | or |
1906 | ||
1907 | r3 <- extract (r1) | |
1908 | r3 <- r3 >> shift | |
1909 | r2 <- extract (r3) | |
1910 | ... <- r2 | |
1911 | ||
1912 | or | |
1913 | ||
1914 | r2 <- extract (r1) | |
1915 | ... <- r2 | |
1916 | ||
1917 | Depending on the alignment and the mode of the store and | |
1918 | subsequent load. | |
1919 | ||
1920 | ||
1921 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
3072d30e | 1922 | and READ_INSN are for the read. Return true if the replacement |
1923 | went ok. */ | |
1924 | ||
1925 | static bool | |
c2eab334 | 1926 | replace_read (store_info *store_info, insn_info_t store_insn, |
aa140b76 | 1927 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
1928 | bitmap regs_live) | |
3072d30e | 1929 | { |
3754d046 | 1930 | machine_mode store_mode = GET_MODE (store_info->mem); |
1931 | machine_mode read_mode = GET_MODE (read_info->mem); | |
4cd001d5 | 1932 | rtx_insn *insns, *this_insn; |
1933 | rtx read_reg; | |
aa140b76 | 1934 | basic_block bb; |
5c9051a4 | 1935 | |
3072d30e | 1936 | if (!dbg_cnt (dse)) |
1937 | return false; | |
1938 | ||
10d4de0e | 1939 | /* Create a sequence of instructions to set up the read register. |
1940 | This sequence goes immediately before the store and its result | |
1941 | is read by the load. | |
1942 | ||
1943 | We need to keep this in perspective. We are replacing a read | |
5c9051a4 | 1944 | with a sequence of insns, but the read will almost certainly be |
1945 | in cache, so it is not going to be an expensive one. Thus, we | |
1946 | are not willing to do a multi insn shift or worse a subroutine | |
1947 | call to get rid of the read. */ | |
1ca59310 | 1948 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 1949 | fprintf (dump_file, "trying to replace %smode load in insn %d" |
1950 | " from %smode store in insn %d\n", | |
1951 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
1952 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
1953 | start_sequence (); | |
aa140b76 | 1954 | bb = BLOCK_FOR_INSN (read_insn->insn); |
1955 | read_reg = get_stored_val (store_info, | |
57bca8dc | 1956 | read_mode, read_info->offset, read_info->width, |
aa140b76 | 1957 | bb, false); |
10d4de0e | 1958 | if (read_reg == NULL_RTX) |
5c9051a4 | 1959 | { |
10d4de0e | 1960 | end_sequence (); |
1ca59310 | 1961 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 1962 | fprintf (dump_file, " -- could not extract bits of stored value\n"); |
1963 | return false; | |
5c9051a4 | 1964 | } |
10d4de0e | 1965 | /* Force the value into a new register so that it won't be clobbered |
1966 | between the store and the load. */ | |
1967 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
1968 | insns = get_insns (); | |
1969 | end_sequence (); | |
5c9051a4 | 1970 | |
a1b0a968 | 1971 | if (insns != NULL_RTX) |
1972 | { | |
1973 | /* Now we have to scan the set of new instructions to see if the | |
1974 | sequence contains and sets of hardregs that happened to be | |
1975 | live at this point. For instance, this can happen if one of | |
1976 | the insns sets the CC and the CC happened to be live at that | |
1977 | point. This does occasionally happen, see PR 37922. */ | |
4fb07d00 | 1978 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
a1b0a968 | 1979 | |
1980 | for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
1981 | note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
48e1416a | 1982 | |
a1b0a968 | 1983 | bitmap_and_into (regs_set, regs_live); |
1984 | if (!bitmap_empty_p (regs_set)) | |
1985 | { | |
1ca59310 | 1986 | if (dump_file && (dump_flags & TDF_DETAILS)) |
a1b0a968 | 1987 | { |
48e1416a | 1988 | fprintf (dump_file, |
a1b0a968 | 1989 | "abandoning replacement because sequence clobbers live hardregs:"); |
1990 | df_print_regset (dump_file, regs_set); | |
1991 | } | |
48e1416a | 1992 | |
a1b0a968 | 1993 | BITMAP_FREE (regs_set); |
1994 | return false; | |
1995 | } | |
1996 | BITMAP_FREE (regs_set); | |
1997 | } | |
1998 | ||
5c9051a4 | 1999 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
3072d30e | 2000 | { |
c2eab334 | 2001 | deferred_change *change = deferred_change_pool.allocate (); |
48e1416a | 2002 | |
5c9051a4 | 2003 | /* Insert this right before the store insn where it will be safe |
2004 | from later insns that might change it before the read. */ | |
2005 | emit_insn_before (insns, store_insn->insn); | |
48e1416a | 2006 | |
5c9051a4 | 2007 | /* And now for the kludge part: cselib croaks if you just |
2008 | return at this point. There are two reasons for this: | |
48e1416a | 2009 | |
5c9051a4 | 2010 | 1) Cselib has an idea of how many pseudos there are and |
2011 | that does not include the new ones we just added. | |
48e1416a | 2012 | |
5c9051a4 | 2013 | 2) Cselib does not know about the move insn we added |
2014 | above the store_info, and there is no way to tell it | |
2015 | about it, because it has "moved on". | |
48e1416a | 2016 | |
5c9051a4 | 2017 | Problem (1) is fixable with a certain amount of engineering. |
2018 | Problem (2) is requires starting the bb from scratch. This | |
2019 | could be expensive. | |
48e1416a | 2020 | |
5c9051a4 | 2021 | So we are just going to have to lie. The move/extraction |
2022 | insns are not really an issue, cselib did not see them. But | |
2023 | the use of the new pseudo read_insn is a real problem because | |
2024 | cselib has not scanned this insn. The way that we solve this | |
2025 | problem is that we are just going to put the mem back for now | |
2026 | and when we are finished with the block, we undo this. We | |
2027 | keep a table of mems to get rid of. At the end of the basic | |
2028 | block we can put them back. */ | |
48e1416a | 2029 | |
5c9051a4 | 2030 | *loc = read_info->mem; |
55c5ac9f | 2031 | change->next = deferred_change_list; |
2032 | deferred_change_list = change; | |
2033 | change->loc = loc; | |
2034 | change->reg = read_reg; | |
48e1416a | 2035 | |
5c9051a4 | 2036 | /* Get rid of the read_info, from the point of view of the |
2037 | rest of dse, play like this read never happened. */ | |
2038 | read_insn->read_rec = read_info->next; | |
e16712b1 | 2039 | read_info_type_pool.remove (read_info); |
1ca59310 | 2040 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 2041 | { |
2042 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2043 | print_simple_rtl (dump_file, read_reg); | |
2044 | fprintf (dump_file, "\n"); | |
2045 | } | |
5c9051a4 | 2046 | return true; |
3072d30e | 2047 | } |
48e1416a | 2048 | else |
3072d30e | 2049 | { |
1ca59310 | 2050 | if (dump_file && (dump_flags & TDF_DETAILS)) |
10d4de0e | 2051 | { |
2052 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2053 | print_simple_rtl (dump_file, read_reg); | |
2054 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2055 | } | |
3072d30e | 2056 | return false; |
2057 | } | |
2058 | } | |
2059 | ||
ec1203cd | 2060 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2061 | be active. */ | |
3072d30e | 2062 | |
ec1203cd | 2063 | static void |
2064 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info) | |
3072d30e | 2065 | { |
82d2c88b | 2066 | rtx mem = *loc, mem_addr; |
3072d30e | 2067 | insn_info_t insn_info; |
370ad826 | 2068 | poly_int64 offset = 0; |
2069 | poly_int64 width = 0; | |
48e1416a | 2070 | cselib_val *base = NULL; |
3072d30e | 2071 | int group_id; |
2072 | read_info_t read_info; | |
2073 | ||
3072d30e | 2074 | insn_info = bb_info->last_insn; |
2075 | ||
2076 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
c7a7ba46 | 2077 | || MEM_VOLATILE_P (mem)) |
3072d30e | 2078 | { |
c7a7ba46 | 2079 | if (crtl->stack_protect_guard |
2080 | && (MEM_EXPR (mem) == crtl->stack_protect_guard | |
2081 | || (crtl->stack_protect_guard_decl | |
2082 | && MEM_EXPR (mem) == crtl->stack_protect_guard_decl)) | |
2083 | && MEM_VOLATILE_P (mem)) | |
2084 | { | |
2085 | /* This is either the stack protector canary on the stack, | |
2086 | which ought to be written by a MEM_VOLATILE_P store and | |
2087 | thus shouldn't be deleted and is read at the very end of | |
2088 | function, but shouldn't conflict with any other store. | |
2089 | Or it is __stack_chk_guard variable or TLS or whatever else | |
2090 | MEM holding the canary value, which really shouldn't be | |
2091 | ever modified in -fstack-protector* protected functions, | |
2092 | otherwise the prologue store wouldn't match the epilogue | |
2093 | check. */ | |
2094 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2095 | fprintf (dump_file, " stack protector canary read ignored.\n"); | |
2096 | insn_info->cannot_delete = true; | |
2097 | return; | |
2098 | } | |
2099 | ||
1ca59310 | 2100 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2101 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
2102 | add_wild_read (bb_info); | |
2103 | insn_info->cannot_delete = true; | |
ec1203cd | 2104 | return; |
3072d30e | 2105 | } |
2106 | ||
2107 | /* If it is reading readonly mem, then there can be no conflict with | |
2108 | another write. */ | |
2109 | if (MEM_READONLY_P (mem)) | |
ec1203cd | 2110 | return; |
3072d30e | 2111 | |
5a3f5d69 | 2112 | if (!canon_address (mem, &group_id, &offset, &base)) |
3072d30e | 2113 | { |
1ca59310 | 2114 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2115 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
2116 | add_wild_read (bb_info); | |
ec1203cd | 2117 | return; |
3072d30e | 2118 | } |
2119 | ||
2120 | if (GET_MODE (mem) == BLKmode) | |
2121 | width = -1; | |
2122 | else | |
2123 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2124 | ||
370ad826 | 2125 | if (!endpoint_representable_p (offset, known_eq (width, -1) ? 1 : width)) |
89a03a50 | 2126 | { |
8f4e0f7b | 2127 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2128 | fprintf (dump_file, " adding wild read, due to overflow.\n"); | |
2129 | add_wild_read (bb_info); | |
89a03a50 | 2130 | return; |
2131 | } | |
2132 | ||
e16712b1 | 2133 | read_info = read_info_type_pool.allocate (); |
3072d30e | 2134 | read_info->group_id = group_id; |
2135 | read_info->mem = mem; | |
57bca8dc | 2136 | read_info->offset = offset; |
2137 | read_info->width = width; | |
3072d30e | 2138 | read_info->next = insn_info->read_rec; |
2139 | insn_info->read_rec = read_info; | |
5a3f5d69 | 2140 | if (group_id < 0) |
2141 | mem_addr = base->val_rtx; | |
82d2c88b | 2142 | else |
2143 | { | |
5a3f5d69 | 2144 | group_info *group = rtx_group_vec[group_id]; |
2145 | mem_addr = group->canon_base_addr; | |
82d2c88b | 2146 | } |
370ad826 | 2147 | if (maybe_ne (offset, 0)) |
5a3f5d69 | 2148 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
3072d30e | 2149 | |
5a3f5d69 | 2150 | if (group_id >= 0) |
3072d30e | 2151 | { |
2152 | /* This is the restricted case where the base is a constant or | |
2153 | the frame pointer and offset is a constant. */ | |
2154 | insn_info_t i_ptr = active_local_stores; | |
2155 | insn_info_t last = NULL; | |
48e1416a | 2156 | |
1ca59310 | 2157 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2158 | { |
370ad826 | 2159 | if (!known_size_p (width)) |
3072d30e | 2160 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", |
2161 | group_id); | |
2162 | else | |
57bca8dc | 2163 | { |
2164 | fprintf (dump_file, " processing const load gid=%d", group_id); | |
2165 | print_range (dump_file, offset, width); | |
2166 | fprintf (dump_file, "\n"); | |
2167 | } | |
3072d30e | 2168 | } |
2169 | ||
2170 | while (i_ptr) | |
2171 | { | |
2172 | bool remove = false; | |
c2eab334 | 2173 | store_info *store_info = i_ptr->store_rec; |
48e1416a | 2174 | |
3072d30e | 2175 | /* Skip the clobbers. */ |
2176 | while (!store_info->is_set) | |
2177 | store_info = store_info->next; | |
48e1416a | 2178 | |
3072d30e | 2179 | /* There are three cases here. */ |
2180 | if (store_info->group_id < 0) | |
2181 | /* We have a cselib store followed by a read from a | |
2182 | const base. */ | |
48e1416a | 2183 | remove |
2184 | = canon_true_dependence (store_info->mem, | |
3072d30e | 2185 | GET_MODE (store_info->mem), |
2186 | store_info->mem_addr, | |
376a287d | 2187 | mem, mem_addr); |
48e1416a | 2188 | |
3072d30e | 2189 | else if (group_id == store_info->group_id) |
2190 | { | |
2191 | /* This is a block mode load. We may get lucky and | |
2192 | canon_true_dependence may save the day. */ | |
370ad826 | 2193 | if (!known_size_p (width)) |
48e1416a | 2194 | remove |
2195 | = canon_true_dependence (store_info->mem, | |
3072d30e | 2196 | GET_MODE (store_info->mem), |
2197 | store_info->mem_addr, | |
376a287d | 2198 | mem, mem_addr); |
48e1416a | 2199 | |
3072d30e | 2200 | /* If this read is just reading back something that we just |
2201 | stored, rewrite the read. */ | |
48e1416a | 2202 | else |
3072d30e | 2203 | { |
2204 | if (store_info->rhs | |
57bca8dc | 2205 | && known_subrange_p (offset, width, store_info->offset, |
2206 | store_info->width) | |
aa140b76 | 2207 | && all_positions_needed_p (store_info, |
57bca8dc | 2208 | offset - store_info->offset, |
aa140b76 | 2209 | width) |
2210 | && replace_read (store_info, i_ptr, read_info, | |
2211 | insn_info, loc, bb_info->regs_live)) | |
ec1203cd | 2212 | return; |
aa140b76 | 2213 | |
3072d30e | 2214 | /* The bases are the same, just see if the offsets |
57bca8dc | 2215 | could overlap. */ |
2216 | if (ranges_maybe_overlap_p (offset, width, | |
2217 | store_info->offset, | |
2218 | store_info->width)) | |
3072d30e | 2219 | remove = true; |
2220 | } | |
2221 | } | |
48e1416a | 2222 | |
2223 | /* else | |
3072d30e | 2224 | The else case that is missing here is that the |
2225 | bases are constant but different. There is nothing | |
2226 | to do here because there is no overlap. */ | |
48e1416a | 2227 | |
3072d30e | 2228 | if (remove) |
2229 | { | |
1ca59310 | 2230 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2231 | dump_insn_info ("removing from active", i_ptr); |
2232 | ||
1242bee6 | 2233 | active_local_stores_len--; |
3072d30e | 2234 | if (last) |
2235 | last->next_local_store = i_ptr->next_local_store; | |
2236 | else | |
2237 | active_local_stores = i_ptr->next_local_store; | |
2238 | } | |
2239 | else | |
2240 | last = i_ptr; | |
2241 | i_ptr = i_ptr->next_local_store; | |
2242 | } | |
2243 | } | |
48e1416a | 2244 | else |
3072d30e | 2245 | { |
2246 | insn_info_t i_ptr = active_local_stores; | |
2247 | insn_info_t last = NULL; | |
1ca59310 | 2248 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2249 | { |
2250 | fprintf (dump_file, " processing cselib load mem:"); | |
2251 | print_inline_rtx (dump_file, mem, 0); | |
2252 | fprintf (dump_file, "\n"); | |
2253 | } | |
2254 | ||
2255 | while (i_ptr) | |
2256 | { | |
2257 | bool remove = false; | |
c2eab334 | 2258 | store_info *store_info = i_ptr->store_rec; |
48e1416a | 2259 | |
1ca59310 | 2260 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2261 | fprintf (dump_file, " processing cselib load against insn %d\n", |
2262 | INSN_UID (i_ptr->insn)); | |
2263 | ||
2264 | /* Skip the clobbers. */ | |
2265 | while (!store_info->is_set) | |
2266 | store_info = store_info->next; | |
2267 | ||
2268 | /* If this read is just reading back something that we just | |
2269 | stored, rewrite the read. */ | |
2270 | if (store_info->rhs | |
2271 | && store_info->group_id == -1 | |
2272 | && store_info->cse_base == base | |
57bca8dc | 2273 | && known_subrange_p (offset, width, store_info->offset, |
2274 | store_info->width) | |
aa140b76 | 2275 | && all_positions_needed_p (store_info, |
57bca8dc | 2276 | offset - store_info->offset, width) |
aa140b76 | 2277 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, |
2278 | bb_info->regs_live)) | |
ec1203cd | 2279 | return; |
3072d30e | 2280 | |
5a3f5d69 | 2281 | remove = canon_true_dependence (store_info->mem, |
2282 | GET_MODE (store_info->mem), | |
2283 | store_info->mem_addr, | |
2284 | mem, mem_addr); | |
48e1416a | 2285 | |
3072d30e | 2286 | if (remove) |
2287 | { | |
1ca59310 | 2288 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2289 | dump_insn_info ("removing from active", i_ptr); |
48e1416a | 2290 | |
1242bee6 | 2291 | active_local_stores_len--; |
3072d30e | 2292 | if (last) |
2293 | last->next_local_store = i_ptr->next_local_store; | |
2294 | else | |
2295 | active_local_stores = i_ptr->next_local_store; | |
2296 | } | |
2297 | else | |
2298 | last = i_ptr; | |
2299 | i_ptr = i_ptr->next_local_store; | |
2300 | } | |
2301 | } | |
3072d30e | 2302 | } |
2303 | ||
ec1203cd | 2304 | /* A note_uses callback in which DATA points the INSN_INFO for |
3072d30e | 2305 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2306 | true for any part of *LOC. */ | |
2307 | ||
2308 | static void | |
2309 | check_mem_read_use (rtx *loc, void *data) | |
2310 | { | |
ec1203cd | 2311 | subrtx_ptr_iterator::array_type array; |
2312 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
2313 | { | |
2314 | rtx *loc = *iter; | |
2315 | if (MEM_P (*loc)) | |
2316 | check_mem_read_rtx (loc, (bb_info_t) data); | |
2317 | } | |
3072d30e | 2318 | } |
2319 | ||
aa140b76 | 2320 | |
2321 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2322 | So far it only handles arguments passed in registers. */ | |
2323 | ||
2324 | static bool | |
2325 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2326 | { | |
39cba157 | 2327 | CUMULATIVE_ARGS args_so_far_v; |
2328 | cumulative_args_t args_so_far; | |
aa140b76 | 2329 | tree arg; |
2330 | int idx; | |
2331 | ||
39cba157 | 2332 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2333 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
aa140b76 | 2334 | |
2335 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2336 | for (idx = 0; | |
2337 | arg != void_list_node && idx < nargs; | |
2338 | arg = TREE_CHAIN (arg), idx++) | |
2339 | { | |
8b449599 | 2340 | scalar_int_mode mode; |
f387af4f | 2341 | rtx reg, link, tmp; |
8b449599 | 2342 | |
2343 | if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg)), &mode)) | |
2344 | return false; | |
2345 | ||
39cba157 | 2346 | reg = targetm.calls.function_arg (args_so_far, mode, NULL_TREE, true); |
8b449599 | 2347 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode) |
aa140b76 | 2348 | return false; |
2349 | ||
2350 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2351 | link; | |
2352 | link = XEXP (link, 1)) | |
2353 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2354 | { | |
8b449599 | 2355 | scalar_int_mode arg_mode; |
aa140b76 | 2356 | args[idx] = XEXP (XEXP (link, 0), 0); |
2357 | if (REG_P (args[idx]) | |
2358 | && REGNO (args[idx]) == REGNO (reg) | |
2359 | && (GET_MODE (args[idx]) == mode | |
8b449599 | 2360 | || (is_int_mode (GET_MODE (args[idx]), &arg_mode) |
2361 | && (GET_MODE_SIZE (arg_mode) <= UNITS_PER_WORD) | |
2362 | && (GET_MODE_SIZE (arg_mode) > GET_MODE_SIZE (mode))))) | |
aa140b76 | 2363 | break; |
2364 | } | |
2365 | if (!link) | |
2366 | return false; | |
2367 | ||
2368 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2369 | if (GET_MODE (args[idx]) != mode) | |
2370 | { | |
2371 | if (!tmp || !CONST_INT_P (tmp)) | |
2372 | return false; | |
f62058c3 | 2373 | tmp = gen_int_mode (INTVAL (tmp), mode); |
aa140b76 | 2374 | } |
2375 | if (tmp) | |
2376 | args[idx] = tmp; | |
2377 | ||
39cba157 | 2378 | targetm.calls.function_arg_advance (args_so_far, mode, NULL_TREE, true); |
aa140b76 | 2379 | } |
2380 | if (arg != void_list_node || idx != nargs) | |
2381 | return false; | |
2382 | return true; | |
2383 | } | |
2384 | ||
5a9ecd4a | 2385 | /* Return a bitmap of the fixed registers contained in IN. */ |
2386 | ||
2387 | static bitmap | |
2388 | copy_fixed_regs (const_bitmap in) | |
2389 | { | |
2390 | bitmap ret; | |
2391 | ||
2392 | ret = ALLOC_REG_SET (NULL); | |
2393 | bitmap_and (ret, in, fixed_reg_set_regset); | |
2394 | return ret; | |
2395 | } | |
aa140b76 | 2396 | |
3072d30e | 2397 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2398 | if some part of it is not a candidate store and assigns to a | |
2399 | non-register target. */ | |
2400 | ||
2401 | static void | |
ebabb7a3 | 2402 | scan_insn (bb_info_t bb_info, rtx_insn *insn) |
3072d30e | 2403 | { |
2404 | rtx body; | |
e16712b1 | 2405 | insn_info_type *insn_info = insn_info_type_pool.allocate (); |
3072d30e | 2406 | int mems_found = 0; |
55c5ac9f | 2407 | memset (insn_info, 0, sizeof (struct insn_info_type)); |
3072d30e | 2408 | |
1ca59310 | 2409 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2410 | fprintf (dump_file, "\n**scanning insn=%d\n", |
2411 | INSN_UID (insn)); | |
2412 | ||
2413 | insn_info->prev_insn = bb_info->last_insn; | |
2414 | insn_info->insn = insn; | |
2415 | bb_info->last_insn = insn_info; | |
48e1416a | 2416 | |
9845d120 | 2417 | if (DEBUG_INSN_P (insn)) |
2418 | { | |
2419 | insn_info->cannot_delete = true; | |
2420 | return; | |
2421 | } | |
3072d30e | 2422 | |
3072d30e | 2423 | /* Look at all of the uses in the insn. */ |
2424 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2425 | ||
2426 | if (CALL_P (insn)) | |
2427 | { | |
aa140b76 | 2428 | bool const_call; |
61ffc71a | 2429 | rtx call, sym; |
aa140b76 | 2430 | tree memset_call = NULL_TREE; |
2431 | ||
3072d30e | 2432 | insn_info->cannot_delete = true; |
17e1318c | 2433 | |
3072d30e | 2434 | /* Const functions cannot do anything bad i.e. read memory, |
17e1318c | 2435 | however, they can read their parameters which may have |
aa140b76 | 2436 | been pushed onto the stack. |
2437 | memset and bzero don't read memory either. */ | |
2438 | const_call = RTL_CONST_CALL_P (insn); | |
61ffc71a | 2439 | if (!const_call |
2440 | && (call = get_call_rtx_from (insn)) | |
2441 | && (sym = XEXP (XEXP (call, 0), 0)) | |
2442 | && GET_CODE (sym) == SYMBOL_REF | |
2443 | && SYMBOL_REF_DECL (sym) | |
2444 | && TREE_CODE (SYMBOL_REF_DECL (sym)) == FUNCTION_DECL | |
a0e9bfbb | 2445 | && fndecl_built_in_p (SYMBOL_REF_DECL (sym), BUILT_IN_MEMSET)) |
61ffc71a | 2446 | memset_call = SYMBOL_REF_DECL (sym); |
2447 | ||
aa140b76 | 2448 | if (const_call || memset_call) |
3072d30e | 2449 | { |
2450 | insn_info_t i_ptr = active_local_stores; | |
2451 | insn_info_t last = NULL; | |
2452 | ||
1ca59310 | 2453 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 2454 | fprintf (dump_file, "%s call %d\n", |
2455 | const_call ? "const" : "memset", INSN_UID (insn)); | |
3072d30e | 2456 | |
16bf64db | 2457 | /* See the head comment of the frame_read field. */ |
17853422 | 2458 | if (reload_completed |
2459 | /* Tail calls are storing their arguments using | |
2460 | arg pointer. If it is a frame pointer on the target, | |
2461 | even before reload we need to kill frame pointer based | |
2462 | stores. */ | |
2463 | || (SIBLING_CALL_P (insn) | |
2464 | && HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
16bf64db | 2465 | insn_info->frame_read = true; |
2466 | ||
2467 | /* Loop over the active stores and remove those which are | |
2468 | killed by the const function call. */ | |
3072d30e | 2469 | while (i_ptr) |
2470 | { | |
16bf64db | 2471 | bool remove_store = false; |
2472 | ||
2473 | /* The stack pointer based stores are always killed. */ | |
17e1318c | 2474 | if (i_ptr->stack_pointer_based) |
16bf64db | 2475 | remove_store = true; |
2476 | ||
2477 | /* If the frame is read, the frame related stores are killed. */ | |
2478 | else if (insn_info->frame_read) | |
2479 | { | |
c2eab334 | 2480 | store_info *store_info = i_ptr->store_rec; |
16bf64db | 2481 | |
2482 | /* Skip the clobbers. */ | |
2483 | while (!store_info->is_set) | |
2484 | store_info = store_info->next; | |
2485 | ||
2486 | if (store_info->group_id >= 0 | |
f1f41a6c | 2487 | && rtx_group_vec[store_info->group_id]->frame_related) |
16bf64db | 2488 | remove_store = true; |
2489 | } | |
2490 | ||
2491 | if (remove_store) | |
3072d30e | 2492 | { |
1ca59310 | 2493 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2494 | dump_insn_info ("removing from active", i_ptr); |
48e1416a | 2495 | |
1242bee6 | 2496 | active_local_stores_len--; |
3072d30e | 2497 | if (last) |
2498 | last->next_local_store = i_ptr->next_local_store; | |
2499 | else | |
2500 | active_local_stores = i_ptr->next_local_store; | |
2501 | } | |
2502 | else | |
2503 | last = i_ptr; | |
16bf64db | 2504 | |
3072d30e | 2505 | i_ptr = i_ptr->next_local_store; |
2506 | } | |
aa140b76 | 2507 | |
2508 | if (memset_call) | |
2509 | { | |
2510 | rtx args[3]; | |
2511 | if (get_call_args (insn, memset_call, args, 3) | |
2512 | && CONST_INT_P (args[1]) | |
2513 | && CONST_INT_P (args[2]) | |
2514 | && INTVAL (args[2]) > 0) | |
2515 | { | |
2516 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
5b2a69fa | 2517 | set_mem_size (mem, INTVAL (args[2])); |
d1f9b275 | 2518 | body = gen_rtx_SET (mem, args[1]); |
aa140b76 | 2519 | mems_found += record_store (body, bb_info); |
1ca59310 | 2520 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 2521 | fprintf (dump_file, "handling memset as BLKmode store\n"); |
2522 | if (mems_found == 1) | |
2523 | { | |
1242bee6 | 2524 | if (active_local_stores_len++ |
2525 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2526 | { | |
2527 | active_local_stores_len = 1; | |
2528 | active_local_stores = NULL; | |
2529 | } | |
5a9ecd4a | 2530 | insn_info->fixed_regs_live |
2531 | = copy_fixed_regs (bb_info->regs_live); | |
aa140b76 | 2532 | insn_info->next_local_store = active_local_stores; |
2533 | active_local_stores = insn_info; | |
2534 | } | |
2535 | } | |
9441fa4e | 2536 | else |
2537 | clear_rhs_from_active_local_stores (); | |
aa140b76 | 2538 | } |
3072d30e | 2539 | } |
17853422 | 2540 | else if (SIBLING_CALL_P (insn) && reload_completed) |
2541 | /* Arguments for a sibling call that are pushed to memory are passed | |
2542 | using the incoming argument pointer of the current function. After | |
2543 | reload that might be (and likely is) frame pointer based. */ | |
2544 | add_wild_read (bb_info); | |
17e1318c | 2545 | else |
b4a708fb | 2546 | /* Every other call, including pure functions, may read any memory |
2547 | that is not relative to the frame. */ | |
2548 | add_non_frame_wild_read (bb_info); | |
17e1318c | 2549 | |
3072d30e | 2550 | return; |
2551 | } | |
2552 | ||
2553 | /* Assuming that there are sets in these insns, we cannot delete | |
2554 | them. */ | |
2555 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
4aafe72f | 2556 | || volatile_refs_p (PATTERN (insn)) |
bc0dfc8d | 2557 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
3072d30e | 2558 | || (RTX_FRAME_RELATED_P (insn)) |
2559 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2560 | insn_info->cannot_delete = true; | |
48e1416a | 2561 | |
3072d30e | 2562 | body = PATTERN (insn); |
2563 | if (GET_CODE (body) == PARALLEL) | |
2564 | { | |
2565 | int i; | |
2566 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2567 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2568 | } | |
2569 | else | |
2570 | mems_found += record_store (body, bb_info); | |
2571 | ||
1ca59310 | 2572 | if (dump_file && (dump_flags & TDF_DETAILS)) |
48e1416a | 2573 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
3072d30e | 2574 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2575 | ||
aa140b76 | 2576 | /* If we found some sets of mems, add it into the active_local_stores so |
2577 | that it can be locally deleted if found dead or used for | |
2578 | replace_read and redundant constant store elimination. Otherwise mark | |
2579 | it as cannot delete. This simplifies the processing later. */ | |
2580 | if (mems_found == 1) | |
3072d30e | 2581 | { |
1242bee6 | 2582 | if (active_local_stores_len++ |
2583 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2584 | { | |
2585 | active_local_stores_len = 1; | |
2586 | active_local_stores = NULL; | |
2587 | } | |
5a9ecd4a | 2588 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
3072d30e | 2589 | insn_info->next_local_store = active_local_stores; |
2590 | active_local_stores = insn_info; | |
2591 | } | |
2592 | else | |
2593 | insn_info->cannot_delete = true; | |
2594 | } | |
2595 | ||
2596 | ||
2597 | /* Remove BASE from the set of active_local_stores. This is a | |
2598 | callback from cselib that is used to get rid of the stores in | |
2599 | active_local_stores. */ | |
2600 | ||
2601 | static void | |
2602 | remove_useless_values (cselib_val *base) | |
2603 | { | |
2604 | insn_info_t insn_info = active_local_stores; | |
2605 | insn_info_t last = NULL; | |
2606 | ||
2607 | while (insn_info) | |
2608 | { | |
c2eab334 | 2609 | store_info *store_info = insn_info->store_rec; |
9ce37fa7 | 2610 | bool del = false; |
3072d30e | 2611 | |
2612 | /* If ANY of the store_infos match the cselib group that is | |
f4d3c071 | 2613 | being deleted, then the insn cannot be deleted. */ |
3072d30e | 2614 | while (store_info) |
2615 | { | |
48e1416a | 2616 | if ((store_info->group_id == -1) |
3072d30e | 2617 | && (store_info->cse_base == base)) |
2618 | { | |
9ce37fa7 | 2619 | del = true; |
3072d30e | 2620 | break; |
2621 | } | |
2622 | store_info = store_info->next; | |
2623 | } | |
2624 | ||
9ce37fa7 | 2625 | if (del) |
3072d30e | 2626 | { |
1242bee6 | 2627 | active_local_stores_len--; |
3072d30e | 2628 | if (last) |
2629 | last->next_local_store = insn_info->next_local_store; | |
2630 | else | |
2631 | active_local_stores = insn_info->next_local_store; | |
2632 | free_store_info (insn_info); | |
2633 | } | |
2634 | else | |
2635 | last = insn_info; | |
48e1416a | 2636 | |
3072d30e | 2637 | insn_info = insn_info->next_local_store; |
2638 | } | |
2639 | } | |
2640 | ||
2641 | ||
2642 | /* Do all of step 1. */ | |
2643 | ||
2644 | static void | |
2645 | dse_step1 (void) | |
2646 | { | |
2647 | basic_block bb; | |
4fb07d00 | 2648 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
48e1416a | 2649 | |
35af0188 | 2650 | cselib_init (0); |
3072d30e | 2651 | all_blocks = BITMAP_ALLOC (NULL); |
2652 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2653 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2654 | ||
ed7d889a | 2655 | FOR_ALL_BB_FN (bb, cfun) |
3072d30e | 2656 | { |
2657 | insn_info_t ptr; | |
e16712b1 | 2658 | bb_info_t bb_info = dse_bb_info_type_pool.allocate (); |
3072d30e | 2659 | |
55c5ac9f | 2660 | memset (bb_info, 0, sizeof (dse_bb_info_type)); |
3072d30e | 2661 | bitmap_set_bit (all_blocks, bb->index); |
a1b0a968 | 2662 | bb_info->regs_live = regs_live; |
2663 | ||
2664 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2665 | df_simulate_initialize_forwards (bb, regs_live); | |
3072d30e | 2666 | |
2667 | bb_table[bb->index] = bb_info; | |
2668 | cselib_discard_hook = remove_useless_values; | |
2669 | ||
2670 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2671 | { | |
ebabb7a3 | 2672 | rtx_insn *insn; |
3072d30e | 2673 | |
3072d30e | 2674 | active_local_stores = NULL; |
1242bee6 | 2675 | active_local_stores_len = 0; |
3072d30e | 2676 | cselib_clear_table (); |
48e1416a | 2677 | |
3072d30e | 2678 | /* Scan the insns. */ |
2679 | FOR_BB_INSNS (bb, insn) | |
2680 | { | |
2681 | if (INSN_P (insn)) | |
2682 | scan_insn (bb_info, insn); | |
2683 | cselib_process_insn (insn); | |
a1b0a968 | 2684 | if (INSN_P (insn)) |
2685 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
3072d30e | 2686 | } |
48e1416a | 2687 | |
3072d30e | 2688 | /* This is something of a hack, because the global algorithm |
2689 | is supposed to take care of the case where stores go dead | |
2690 | at the end of the function. However, the global | |
2691 | algorithm must take a more conservative view of block | |
2692 | mode reads than the local alg does. So to get the case | |
2693 | where you have a store to the frame followed by a non | |
bef304b8 | 2694 | overlapping block more read, we look at the active local |
3072d30e | 2695 | stores at the end of the function and delete all of the |
2696 | frame and spill based ones. */ | |
2697 | if (stores_off_frame_dead_at_return | |
2698 | && (EDGE_COUNT (bb->succs) == 0 | |
2699 | || (single_succ_p (bb) | |
34154e27 | 2700 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
18d50ae6 | 2701 | && ! crtl->calls_eh_return))) |
3072d30e | 2702 | { |
2703 | insn_info_t i_ptr = active_local_stores; | |
2704 | while (i_ptr) | |
2705 | { | |
c2eab334 | 2706 | store_info *store_info = i_ptr->store_rec; |
3072d30e | 2707 | |
2708 | /* Skip the clobbers. */ | |
2709 | while (!store_info->is_set) | |
2710 | store_info = store_info->next; | |
5a3f5d69 | 2711 | if (store_info->group_id >= 0) |
2712 | { | |
2713 | group_info *group = rtx_group_vec[store_info->group_id]; | |
2714 | if (group->frame_related && !i_ptr->cannot_delete) | |
2715 | delete_dead_store_insn (i_ptr); | |
2716 | } | |
3072d30e | 2717 | |
2718 | i_ptr = i_ptr->next_local_store; | |
2719 | } | |
2720 | } | |
2721 | ||
2722 | /* Get rid of the loads that were discovered in | |
2723 | replace_read. Cselib is finished with this block. */ | |
2724 | while (deferred_change_list) | |
2725 | { | |
c2eab334 | 2726 | deferred_change *next = deferred_change_list->next; |
3072d30e | 2727 | |
2728 | /* There is no reason to validate this change. That was | |
2729 | done earlier. */ | |
2730 | *deferred_change_list->loc = deferred_change_list->reg; | |
e16712b1 | 2731 | deferred_change_pool.remove (deferred_change_list); |
3072d30e | 2732 | deferred_change_list = next; |
2733 | } | |
2734 | ||
2735 | /* Get rid of all of the cselib based store_infos in this | |
2736 | block and mark the containing insns as not being | |
2737 | deletable. */ | |
2738 | ptr = bb_info->last_insn; | |
2739 | while (ptr) | |
2740 | { | |
2741 | if (ptr->contains_cselib_groups) | |
aa140b76 | 2742 | { |
c2eab334 | 2743 | store_info *s_info = ptr->store_rec; |
aa140b76 | 2744 | while (s_info && !s_info->is_set) |
2745 | s_info = s_info->next; | |
2746 | if (s_info | |
2747 | && s_info->redundant_reason | |
2748 | && s_info->redundant_reason->insn | |
2749 | && !ptr->cannot_delete) | |
2750 | { | |
1ca59310 | 2751 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 2752 | fprintf (dump_file, "Locally deleting insn %d " |
2753 | "because insn %d stores the " | |
2754 | "same value and couldn't be " | |
2755 | "eliminated\n", | |
2756 | INSN_UID (ptr->insn), | |
2757 | INSN_UID (s_info->redundant_reason->insn)); | |
2758 | delete_dead_store_insn (ptr); | |
2759 | } | |
aa140b76 | 2760 | free_store_info (ptr); |
2761 | } | |
2762 | else | |
2763 | { | |
c2eab334 | 2764 | store_info *s_info; |
aa140b76 | 2765 | |
2766 | /* Free at least positions_needed bitmaps. */ | |
2767 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2768 | if (s_info->is_large) | |
2769 | { | |
843bd2fa | 2770 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
aa140b76 | 2771 | s_info->is_large = false; |
2772 | } | |
2773 | } | |
3072d30e | 2774 | ptr = ptr->prev_insn; |
2775 | } | |
2776 | ||
55c5ac9f | 2777 | cse_store_info_pool.release (); |
3072d30e | 2778 | } |
a1b0a968 | 2779 | bb_info->regs_live = NULL; |
3072d30e | 2780 | } |
2781 | ||
a1b0a968 | 2782 | BITMAP_FREE (regs_live); |
3072d30e | 2783 | cselib_finish (); |
c1f445d2 | 2784 | rtx_group_table->empty (); |
3072d30e | 2785 | } |
2786 | ||
2787 | \f | |
2788 | /*---------------------------------------------------------------------------- | |
2789 | Second step. | |
2790 | ||
2791 | Assign each byte position in the stores that we are going to | |
2792 | analyze globally to a position in the bitmaps. Returns true if | |
6dfdc153 | 2793 | there are any bit positions assigned. |
3072d30e | 2794 | ----------------------------------------------------------------------------*/ |
2795 | ||
2796 | static void | |
2797 | dse_step2_init (void) | |
2798 | { | |
2799 | unsigned int i; | |
c2eab334 | 2800 | group_info *group; |
3072d30e | 2801 | |
f1f41a6c | 2802 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 2803 | { |
2804 | /* For all non stack related bases, we only consider a store to | |
2805 | be deletable if there are two or more stores for that | |
2806 | position. This is because it takes one store to make the | |
2807 | other store redundant. However, for the stores that are | |
2808 | stack related, we consider them if there is only one store | |
2809 | for the position. We do this because the stack related | |
2810 | stores can be deleted if their is no read between them and | |
2811 | the end of the function. | |
48e1416a | 2812 | |
3072d30e | 2813 | To make this work in the current framework, we take the stack |
2814 | related bases add all of the bits from store1 into store2. | |
2815 | This has the effect of making the eligible even if there is | |
2816 | only one store. */ | |
2817 | ||
2818 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2819 | { | |
2820 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2821 | bitmap_ior_into (group->store2_p, group->store1_p); | |
1ca59310 | 2822 | if (dump_file && (dump_flags & TDF_DETAILS)) |
48e1416a | 2823 | fprintf (dump_file, "group %d is frame related ", i); |
3072d30e | 2824 | } |
2825 | ||
2826 | group->offset_map_size_n++; | |
4fb07d00 | 2827 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2828 | group->offset_map_size_n); | |
3072d30e | 2829 | group->offset_map_size_p++; |
4fb07d00 | 2830 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2831 | group->offset_map_size_p); | |
3072d30e | 2832 | group->process_globally = false; |
1ca59310 | 2833 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 2834 | { |
48e1416a | 2835 | fprintf (dump_file, "group %d(%d+%d): ", i, |
3072d30e | 2836 | (int)bitmap_count_bits (group->store2_n), |
2837 | (int)bitmap_count_bits (group->store2_p)); | |
2838 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2839 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2840 | } | |
2841 | } | |
2842 | } | |
2843 | ||
2844 | ||
5a3f5d69 | 2845 | /* Init the offset tables. */ |
3072d30e | 2846 | |
2847 | static bool | |
5a3f5d69 | 2848 | dse_step2 (void) |
3072d30e | 2849 | { |
2850 | unsigned int i; | |
c2eab334 | 2851 | group_info *group; |
3072d30e | 2852 | /* Position 0 is unused because 0 is used in the maps to mean |
2853 | unused. */ | |
2854 | current_position = 1; | |
f1f41a6c | 2855 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 2856 | { |
2857 | bitmap_iterator bi; | |
2858 | unsigned int j; | |
2859 | ||
9af5ce0c | 2860 | memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
2861 | memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
3072d30e | 2862 | bitmap_clear (group->group_kill); |
2863 | ||
2864 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2865 | { | |
2866 | bitmap_set_bit (group->group_kill, current_position); | |
b4a708fb | 2867 | if (bitmap_bit_p (group->escaped_n, j)) |
2868 | bitmap_set_bit (kill_on_calls, current_position); | |
3072d30e | 2869 | group->offset_map_n[j] = current_position++; |
2870 | group->process_globally = true; | |
2871 | } | |
2872 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2873 | { | |
48e1416a | 2874 | bitmap_set_bit (group->group_kill, current_position); |
b4a708fb | 2875 | if (bitmap_bit_p (group->escaped_p, j)) |
2876 | bitmap_set_bit (kill_on_calls, current_position); | |
3072d30e | 2877 | group->offset_map_p[j] = current_position++; |
2878 | group->process_globally = true; | |
2879 | } | |
2880 | } | |
2881 | return current_position != 1; | |
2882 | } | |
2883 | ||
2884 | ||
3072d30e | 2885 | \f |
2886 | /*---------------------------------------------------------------------------- | |
2887 | Third step. | |
48e1416a | 2888 | |
3072d30e | 2889 | Build the bit vectors for the transfer functions. |
2890 | ----------------------------------------------------------------------------*/ | |
2891 | ||
2892 | ||
3072d30e | 2893 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
2894 | there, return 0. */ | |
2895 | ||
2896 | static int | |
c2eab334 | 2897 | get_bitmap_index (group_info *group_info, HOST_WIDE_INT offset) |
3072d30e | 2898 | { |
2899 | if (offset < 0) | |
2900 | { | |
2901 | HOST_WIDE_INT offset_p = -offset; | |
2902 | if (offset_p >= group_info->offset_map_size_n) | |
2903 | return 0; | |
2904 | return group_info->offset_map_n[offset_p]; | |
2905 | } | |
2906 | else | |
2907 | { | |
2908 | if (offset >= group_info->offset_map_size_p) | |
2909 | return 0; | |
2910 | return group_info->offset_map_p[offset]; | |
2911 | } | |
2912 | } | |
2913 | ||
2914 | ||
2915 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
2916 | may be NULL. */ | |
2917 | ||
48e1416a | 2918 | static void |
5a3f5d69 | 2919 | scan_stores (store_info *store_info, bitmap gen, bitmap kill) |
3072d30e | 2920 | { |
2921 | while (store_info) | |
2922 | { | |
370ad826 | 2923 | HOST_WIDE_INT i, offset, width; |
c2eab334 | 2924 | group_info *group_info |
f1f41a6c | 2925 | = rtx_group_vec[store_info->group_id]; |
370ad826 | 2926 | /* We can (conservatively) ignore stores whose bounds aren't known; |
2927 | they simply don't generate new global dse opportunities. */ | |
2928 | if (group_info->process_globally | |
2929 | && store_info->offset.is_constant (&offset) | |
2930 | && store_info->width.is_constant (&width)) | |
57bca8dc | 2931 | { |
370ad826 | 2932 | HOST_WIDE_INT end = offset + width; |
2933 | for (i = offset; i < end; i++) | |
57bca8dc | 2934 | { |
2935 | int index = get_bitmap_index (group_info, i); | |
2936 | if (index != 0) | |
2937 | { | |
2938 | bitmap_set_bit (gen, index); | |
2939 | if (kill) | |
2940 | bitmap_clear_bit (kill, index); | |
2941 | } | |
2942 | } | |
2943 | } | |
3072d30e | 2944 | store_info = store_info->next; |
2945 | } | |
2946 | } | |
2947 | ||
2948 | ||
3072d30e | 2949 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL |
2950 | may be NULL. */ | |
2951 | ||
2952 | static void | |
5a3f5d69 | 2953 | scan_reads (insn_info_t insn_info, bitmap gen, bitmap kill) |
3072d30e | 2954 | { |
2955 | read_info_t read_info = insn_info->read_rec; | |
2956 | int i; | |
c2eab334 | 2957 | group_info *group; |
3072d30e | 2958 | |
16bf64db | 2959 | /* If this insn reads the frame, kill all the frame related stores. */ |
2960 | if (insn_info->frame_read) | |
2961 | { | |
f1f41a6c | 2962 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
16bf64db | 2963 | if (group->process_globally && group->frame_related) |
2964 | { | |
2965 | if (kill) | |
2966 | bitmap_ior_into (kill, group->group_kill); | |
48e1416a | 2967 | bitmap_and_compl_into (gen, group->group_kill); |
16bf64db | 2968 | } |
2969 | } | |
b4a708fb | 2970 | if (insn_info->non_frame_wild_read) |
2971 | { | |
2972 | /* Kill all non-frame related stores. Kill all stores of variables that | |
2973 | escape. */ | |
2974 | if (kill) | |
2975 | bitmap_ior_into (kill, kill_on_calls); | |
2976 | bitmap_and_compl_into (gen, kill_on_calls); | |
f1f41a6c | 2977 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
b4a708fb | 2978 | if (group->process_globally && !group->frame_related) |
2979 | { | |
2980 | if (kill) | |
2981 | bitmap_ior_into (kill, group->group_kill); | |
2982 | bitmap_and_compl_into (gen, group->group_kill); | |
2983 | } | |
2984 | } | |
3072d30e | 2985 | while (read_info) |
2986 | { | |
f1f41a6c | 2987 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 2988 | { |
2989 | if (group->process_globally) | |
2990 | { | |
2991 | if (i == read_info->group_id) | |
2992 | { | |
370ad826 | 2993 | HOST_WIDE_INT offset, width; |
2994 | /* Reads with non-constant size kill all DSE opportunities | |
2995 | in the group. */ | |
2996 | if (!read_info->offset.is_constant (&offset) | |
2997 | || !read_info->width.is_constant (&width) | |
2998 | || !known_size_p (width)) | |
3072d30e | 2999 | { |
57bca8dc | 3000 | /* Handle block mode reads. */ |
3072d30e | 3001 | if (kill) |
3002 | bitmap_ior_into (kill, group->group_kill); | |
3003 | bitmap_and_compl_into (gen, group->group_kill); | |
3004 | } | |
3005 | else | |
3006 | { | |
3007 | /* The groups are the same, just process the | |
3008 | offsets. */ | |
3009 | HOST_WIDE_INT j; | |
370ad826 | 3010 | HOST_WIDE_INT end = offset + width; |
3011 | for (j = offset; j < end; j++) | |
3072d30e | 3012 | { |
3013 | int index = get_bitmap_index (group, j); | |
3014 | if (index != 0) | |
3015 | { | |
3016 | if (kill) | |
3017 | bitmap_set_bit (kill, index); | |
3018 | bitmap_clear_bit (gen, index); | |
3019 | } | |
3020 | } | |
3021 | } | |
3022 | } | |
3023 | else | |
3024 | { | |
3025 | /* The groups are different, if the alias sets | |
3026 | conflict, clear the entire group. We only need | |
3027 | to apply this test if the read_info is a cselib | |
3028 | read. Anything with a constant base cannot alias | |
3029 | something else with a different constant | |
3030 | base. */ | |
3031 | if ((read_info->group_id < 0) | |
48e1416a | 3032 | && canon_true_dependence (group->base_mem, |
ec410bf1 | 3033 | GET_MODE (group->base_mem), |
82d2c88b | 3034 | group->canon_base_addr, |
376a287d | 3035 | read_info->mem, NULL_RTX)) |
3072d30e | 3036 | { |
3037 | if (kill) | |
3038 | bitmap_ior_into (kill, group->group_kill); | |
3039 | bitmap_and_compl_into (gen, group->group_kill); | |
3040 | } | |
3041 | } | |
3042 | } | |
3043 | } | |
48e1416a | 3044 | |
3072d30e | 3045 | read_info = read_info->next; |
3046 | } | |
3047 | } | |
3048 | ||
3072d30e | 3049 | |
3050 | /* Return the insn in BB_INFO before the first wild read or if there | |
3051 | are no wild reads in the block, return the last insn. */ | |
3052 | ||
3053 | static insn_info_t | |
3054 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3055 | { | |
3056 | insn_info_t insn_info = bb_info->last_insn; | |
3057 | insn_info_t last_wild_read = NULL; | |
3058 | ||
3059 | while (insn_info) | |
3060 | { | |
3061 | if (insn_info->wild_read) | |
3062 | { | |
3063 | last_wild_read = insn_info->prev_insn; | |
3064 | /* Block starts with wild read. */ | |
3065 | if (!last_wild_read) | |
3066 | return NULL; | |
3067 | } | |
3068 | ||
3069 | insn_info = insn_info->prev_insn; | |
3070 | } | |
3071 | ||
3072 | if (last_wild_read) | |
3073 | return last_wild_read; | |
3074 | else | |
3075 | return bb_info->last_insn; | |
3076 | } | |
3077 | ||
3078 | ||
3079 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3080 | the block in order to build the gen and kill sets for the block. | |
3081 | We start at ptr which may be the last insn in the block or may be | |
3082 | the first insn with a wild read. In the latter case we are able to | |
3083 | skip the rest of the block because it just does not matter: | |
3084 | anything that happens is hidden by the wild read. */ | |
3085 | ||
3086 | static void | |
5a3f5d69 | 3087 | dse_step3_scan (basic_block bb) |
3072d30e | 3088 | { |
3089 | bb_info_t bb_info = bb_table[bb->index]; | |
3090 | insn_info_t insn_info; | |
3091 | ||
5a3f5d69 | 3092 | insn_info = find_insn_before_first_wild_read (bb_info); |
48e1416a | 3093 | |
3072d30e | 3094 | /* In the spill case or in the no_spill case if there is no wild |
3095 | read in the block, we will need a kill set. */ | |
3096 | if (insn_info == bb_info->last_insn) | |
3097 | { | |
3098 | if (bb_info->kill) | |
3099 | bitmap_clear (bb_info->kill); | |
3100 | else | |
4fb07d00 | 3101 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3102 | } |
48e1416a | 3103 | else |
3072d30e | 3104 | if (bb_info->kill) |
3105 | BITMAP_FREE (bb_info->kill); | |
3106 | ||
3107 | while (insn_info) | |
3108 | { | |
3109 | /* There may have been code deleted by the dce pass run before | |
3110 | this phase. */ | |
3111 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3112 | { | |
5a3f5d69 | 3113 | scan_stores (insn_info->store_rec, bb_info->gen, bb_info->kill); |
3114 | scan_reads (insn_info, bb_info->gen, bb_info->kill); | |
48e1416a | 3115 | } |
3072d30e | 3116 | |
3117 | insn_info = insn_info->prev_insn; | |
3118 | } | |
3119 | } | |
3120 | ||
3121 | ||
3122 | /* Set the gen set of the exit block, and also any block with no | |
3123 | successors that does not have a wild read. */ | |
3124 | ||
3125 | static void | |
3126 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3127 | { | |
3128 | /* The gen set is all 0's for the exit block except for the | |
3129 | frame_pointer_group. */ | |
48e1416a | 3130 | |
3072d30e | 3131 | if (stores_off_frame_dead_at_return) |
3132 | { | |
3133 | unsigned int i; | |
c2eab334 | 3134 | group_info *group; |
48e1416a | 3135 | |
f1f41a6c | 3136 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3072d30e | 3137 | { |
3138 | if (group->process_globally && group->frame_related) | |
3139 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3140 | } | |
3141 | } | |
3142 | } | |
3143 | ||
3144 | ||
3145 | /* Find all of the blocks that are not backwards reachable from the | |
3146 | exit block or any block with no successors (BB). These are the | |
3147 | infinite loops or infinite self loops. These blocks will still | |
3148 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3149 | ||
3150 | static void | |
3151 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3152 | { | |
3153 | edge e; | |
3154 | edge_iterator ei; | |
3155 | ||
08b7917c | 3156 | if (bitmap_bit_p (unreachable_blocks, bb->index)) |
3072d30e | 3157 | { |
08b7917c | 3158 | bitmap_clear_bit (unreachable_blocks, bb->index); |
3072d30e | 3159 | FOR_EACH_EDGE (e, ei, bb->preds) |
48e1416a | 3160 | { |
3072d30e | 3161 | mark_reachable_blocks (unreachable_blocks, e->src); |
48e1416a | 3162 | } |
3072d30e | 3163 | } |
3164 | } | |
3165 | ||
3166 | /* Build the transfer functions for the function. */ | |
3167 | ||
3168 | static void | |
5a3f5d69 | 3169 | dse_step3 () |
3072d30e | 3170 | { |
3171 | basic_block bb; | |
3072d30e | 3172 | sbitmap_iterator sbi; |
3173 | bitmap all_ones = NULL; | |
3174 | unsigned int i; | |
48e1416a | 3175 | |
3c6549f8 | 3176 | auto_sbitmap unreachable_blocks (last_basic_block_for_fn (cfun)); |
53c5d9d4 | 3177 | bitmap_ones (unreachable_blocks); |
3072d30e | 3178 | |
ed7d889a | 3179 | FOR_ALL_BB_FN (bb, cfun) |
3072d30e | 3180 | { |
3181 | bb_info_t bb_info = bb_table[bb->index]; | |
3182 | if (bb_info->gen) | |
3183 | bitmap_clear (bb_info->gen); | |
3184 | else | |
4fb07d00 | 3185 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3186 | |
3187 | if (bb->index == ENTRY_BLOCK) | |
3188 | ; | |
3189 | else if (bb->index == EXIT_BLOCK) | |
3190 | dse_step3_exit_block_scan (bb_info); | |
3191 | else | |
5a3f5d69 | 3192 | dse_step3_scan (bb); |
3072d30e | 3193 | if (EDGE_COUNT (bb->succs) == 0) |
3194 | mark_reachable_blocks (unreachable_blocks, bb); | |
3195 | ||
3196 | /* If this is the second time dataflow is run, delete the old | |
3197 | sets. */ | |
3198 | if (bb_info->in) | |
3199 | BITMAP_FREE (bb_info->in); | |
3200 | if (bb_info->out) | |
3201 | BITMAP_FREE (bb_info->out); | |
3202 | } | |
3203 | ||
3204 | /* For any block in an infinite loop, we must initialize the out set | |
3205 | to all ones. This could be expensive, but almost never occurs in | |
3206 | practice. However, it is common in regression tests. */ | |
0d211963 | 3207 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
3072d30e | 3208 | { |
3209 | if (bitmap_bit_p (all_blocks, i)) | |
3210 | { | |
3211 | bb_info_t bb_info = bb_table[i]; | |
3212 | if (!all_ones) | |
3213 | { | |
3214 | unsigned int j; | |
c2eab334 | 3215 | group_info *group; |
3072d30e | 3216 | |
4fb07d00 | 3217 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
f1f41a6c | 3218 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
3072d30e | 3219 | bitmap_ior_into (all_ones, group->group_kill); |
3220 | } | |
3221 | if (!bb_info->out) | |
3222 | { | |
4fb07d00 | 3223 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3224 | bitmap_copy (bb_info->out, all_ones); |
3225 | } | |
3226 | } | |
3227 | } | |
3228 | ||
3229 | if (all_ones) | |
3230 | BITMAP_FREE (all_ones); | |
3072d30e | 3231 | } |
3232 | ||
3233 | ||
3234 | \f | |
3235 | /*---------------------------------------------------------------------------- | |
3236 | Fourth step. | |
3237 | ||
3238 | Solve the bitvector equations. | |
3239 | ----------------------------------------------------------------------------*/ | |
3240 | ||
3241 | ||
3242 | /* Confluence function for blocks with no successors. Create an out | |
3243 | set from the gen set of the exit block. This block logically has | |
3244 | the exit block as a successor. */ | |
3245 | ||
3246 | ||
3247 | ||
3248 | static void | |
3249 | dse_confluence_0 (basic_block bb) | |
3250 | { | |
3251 | bb_info_t bb_info = bb_table[bb->index]; | |
3252 | ||
3253 | if (bb->index == EXIT_BLOCK) | |
3254 | return; | |
3255 | ||
3256 | if (!bb_info->out) | |
3257 | { | |
4fb07d00 | 3258 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3259 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3260 | } | |
3261 | } | |
3262 | ||
3263 | /* Propagate the information from the in set of the dest of E to the | |
3264 | out set of the src of E. If the various in or out sets are not | |
3265 | there, that means they are all ones. */ | |
3266 | ||
a703ca31 | 3267 | static bool |
3072d30e | 3268 | dse_confluence_n (edge e) |
3269 | { | |
3270 | bb_info_t src_info = bb_table[e->src->index]; | |
3271 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3272 | ||
3273 | if (dest_info->in) | |
3274 | { | |
3275 | if (src_info->out) | |
3276 | bitmap_and_into (src_info->out, dest_info->in); | |
3277 | else | |
3278 | { | |
4fb07d00 | 3279 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3280 | bitmap_copy (src_info->out, dest_info->in); |
3281 | } | |
3282 | } | |
a703ca31 | 3283 | return true; |
3072d30e | 3284 | } |
3285 | ||
3286 | ||
3287 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
48e1416a | 3288 | block. There are three cases: |
3072d30e | 3289 | |
3290 | 1) The block has no kill set. In this case the kill set is all | |
3291 | ones. It does not matter what the out set of the block is, none of | |
3292 | the info can reach the top. The only thing that reaches the top is | |
3293 | the gen set and we just copy the set. | |
3294 | ||
3295 | 2) There is a kill set but no out set and bb has successors. In | |
3296 | this case we just return. Eventually an out set will be created and | |
3297 | it is better to wait than to create a set of ones. | |
3298 | ||
3299 | 3) There is both a kill and out set. We apply the obvious transfer | |
3300 | function. | |
3301 | */ | |
3302 | ||
3303 | static bool | |
3304 | dse_transfer_function (int bb_index) | |
3305 | { | |
3306 | bb_info_t bb_info = bb_table[bb_index]; | |
3307 | ||
3308 | if (bb_info->kill) | |
3309 | { | |
3310 | if (bb_info->out) | |
3311 | { | |
3312 | /* Case 3 above. */ | |
3313 | if (bb_info->in) | |
48e1416a | 3314 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
3072d30e | 3315 | bb_info->out, bb_info->kill); |
3316 | else | |
3317 | { | |
4fb07d00 | 3318 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
48e1416a | 3319 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
3072d30e | 3320 | bb_info->out, bb_info->kill); |
3321 | return true; | |
3322 | } | |
3323 | } | |
3324 | else | |
3325 | /* Case 2 above. */ | |
3326 | return false; | |
3327 | } | |
3328 | else | |
3329 | { | |
3330 | /* Case 1 above. If there is already an in set, nothing | |
3331 | happens. */ | |
3332 | if (bb_info->in) | |
3333 | return false; | |
3334 | else | |
3335 | { | |
4fb07d00 | 3336 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
3072d30e | 3337 | bitmap_copy (bb_info->in, bb_info->gen); |
3338 | return true; | |
3339 | } | |
3340 | } | |
3341 | } | |
3342 | ||
3343 | /* Solve the dataflow equations. */ | |
3344 | ||
3345 | static void | |
3346 | dse_step4 (void) | |
3347 | { | |
48e1416a | 3348 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3349 | dse_confluence_n, dse_transfer_function, | |
3350 | all_blocks, df_get_postorder (DF_BACKWARD), | |
3072d30e | 3351 | df_get_n_blocks (DF_BACKWARD)); |
1ca59310 | 3352 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 3353 | { |
3354 | basic_block bb; | |
3355 | ||
3356 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
ed7d889a | 3357 | FOR_ALL_BB_FN (bb, cfun) |
3072d30e | 3358 | { |
3359 | bb_info_t bb_info = bb_table[bb->index]; | |
3360 | ||
3361 | df_print_bb_index (bb, dump_file); | |
3362 | if (bb_info->in) | |
3363 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3364 | else | |
3365 | fprintf (dump_file, " in: *MISSING*\n"); | |
3366 | if (bb_info->gen) | |
3367 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3368 | else | |
3369 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3370 | if (bb_info->kill) | |
3371 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3372 | else | |
3373 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3374 | if (bb_info->out) | |
3375 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3376 | else | |
3377 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3378 | } | |
3379 | } | |
3380 | } | |
3381 | ||
3382 | ||
3383 | \f | |
3384 | /*---------------------------------------------------------------------------- | |
3385 | Fifth step. | |
3386 | ||
bef304b8 | 3387 | Delete the stores that can only be deleted using the global information. |
3072d30e | 3388 | ----------------------------------------------------------------------------*/ |
3389 | ||
3390 | ||
3391 | static void | |
5a3f5d69 | 3392 | dse_step5 (void) |
3072d30e | 3393 | { |
3394 | basic_block bb; | |
fc00614f | 3395 | FOR_EACH_BB_FN (bb, cfun) |
3072d30e | 3396 | { |
3397 | bb_info_t bb_info = bb_table[bb->index]; | |
3398 | insn_info_t insn_info = bb_info->last_insn; | |
3399 | bitmap v = bb_info->out; | |
3400 | ||
3401 | while (insn_info) | |
3402 | { | |
3403 | bool deleted = false; | |
3404 | if (dump_file && insn_info->insn) | |
3405 | { | |
3406 | fprintf (dump_file, "starting to process insn %d\n", | |
3407 | INSN_UID (insn_info->insn)); | |
3408 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3409 | } | |
3410 | ||
3411 | /* There may have been code deleted by the dce pass run before | |
3412 | this phase. */ | |
48e1416a | 3413 | if (insn_info->insn |
3072d30e | 3414 | && INSN_P (insn_info->insn) |
3415 | && (!insn_info->cannot_delete) | |
3416 | && (!bitmap_empty_p (v))) | |
3417 | { | |
c2eab334 | 3418 | store_info *store_info = insn_info->store_rec; |
3072d30e | 3419 | |
3420 | /* Try to delete the current insn. */ | |
3421 | deleted = true; | |
48e1416a | 3422 | |
3072d30e | 3423 | /* Skip the clobbers. */ |
3424 | while (!store_info->is_set) | |
3425 | store_info = store_info->next; | |
3426 | ||
370ad826 | 3427 | HOST_WIDE_INT i, offset, width; |
5a3f5d69 | 3428 | group_info *group_info = rtx_group_vec[store_info->group_id]; |
3429 | ||
370ad826 | 3430 | if (!store_info->offset.is_constant (&offset) |
3431 | || !store_info->width.is_constant (&width)) | |
3432 | deleted = false; | |
3433 | else | |
3072d30e | 3434 | { |
370ad826 | 3435 | HOST_WIDE_INT end = offset + width; |
3436 | for (i = offset; i < end; i++) | |
3072d30e | 3437 | { |
370ad826 | 3438 | int index = get_bitmap_index (group_info, i); |
3439 | ||
1ca59310 | 3440 | if (dump_file && (dump_flags & TDF_DETAILS)) |
370ad826 | 3441 | fprintf (dump_file, "i = %d, index = %d\n", |
3442 | (int) i, index); | |
3443 | if (index == 0 || !bitmap_bit_p (v, index)) | |
3444 | { | |
3445 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3446 | fprintf (dump_file, "failing at i = %d\n", | |
3447 | (int) i); | |
3448 | deleted = false; | |
3449 | break; | |
3450 | } | |
3072d30e | 3451 | } |
3452 | } | |
3453 | if (deleted) | |
3454 | { | |
5a9ecd4a | 3455 | if (dbg_cnt (dse) |
3456 | && check_for_inc_dec_1 (insn_info)) | |
3072d30e | 3457 | { |
3072d30e | 3458 | delete_insn (insn_info->insn); |
3459 | insn_info->insn = NULL; | |
3460 | globally_deleted++; | |
3461 | } | |
3462 | } | |
3463 | } | |
3464 | /* We do want to process the local info if the insn was | |
6dfdc153 | 3465 | deleted. For instance, if the insn did a wild read, we |
3072d30e | 3466 | no longer need to trash the info. */ |
48e1416a | 3467 | if (insn_info->insn |
3072d30e | 3468 | && INSN_P (insn_info->insn) |
3469 | && (!deleted)) | |
3470 | { | |
5a3f5d69 | 3471 | scan_stores (insn_info->store_rec, v, NULL); |
3072d30e | 3472 | if (insn_info->wild_read) |
3473 | { | |
1ca59310 | 3474 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 3475 | fprintf (dump_file, "wild read\n"); |
3476 | bitmap_clear (v); | |
3477 | } | |
b4a708fb | 3478 | else if (insn_info->read_rec |
3d35f723 | 3479 | || insn_info->non_frame_wild_read |
3480 | || insn_info->frame_read) | |
3072d30e | 3481 | { |
3d35f723 | 3482 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3483 | { | |
3484 | if (!insn_info->non_frame_wild_read | |
3485 | && !insn_info->frame_read) | |
3486 | fprintf (dump_file, "regular read\n"); | |
3487 | if (insn_info->non_frame_wild_read) | |
3488 | fprintf (dump_file, "non-frame wild read\n"); | |
3489 | if (insn_info->frame_read) | |
3490 | fprintf (dump_file, "frame read\n"); | |
3491 | } | |
5a3f5d69 | 3492 | scan_reads (insn_info, v, NULL); |
3072d30e | 3493 | } |
3494 | } | |
48e1416a | 3495 | |
3072d30e | 3496 | insn_info = insn_info->prev_insn; |
3497 | } | |
3498 | } | |
3499 | } | |
3500 | ||
3501 | ||
3072d30e | 3502 | \f |
3503 | /*---------------------------------------------------------------------------- | |
3504 | Sixth step. | |
3505 | ||
aa140b76 | 3506 | Delete stores made redundant by earlier stores (which store the same |
3507 | value) that couldn't be eliminated. | |
3508 | ----------------------------------------------------------------------------*/ | |
3509 | ||
3510 | static void | |
3511 | dse_step6 (void) | |
3512 | { | |
3513 | basic_block bb; | |
3514 | ||
ed7d889a | 3515 | FOR_ALL_BB_FN (bb, cfun) |
aa140b76 | 3516 | { |
3517 | bb_info_t bb_info = bb_table[bb->index]; | |
3518 | insn_info_t insn_info = bb_info->last_insn; | |
3519 | ||
3520 | while (insn_info) | |
3521 | { | |
3522 | /* There may have been code deleted by the dce pass run before | |
3523 | this phase. */ | |
3524 | if (insn_info->insn | |
3525 | && INSN_P (insn_info->insn) | |
3526 | && !insn_info->cannot_delete) | |
3527 | { | |
c2eab334 | 3528 | store_info *s_info = insn_info->store_rec; |
aa140b76 | 3529 | |
3530 | while (s_info && !s_info->is_set) | |
3531 | s_info = s_info->next; | |
3532 | if (s_info | |
3533 | && s_info->redundant_reason | |
3534 | && s_info->redundant_reason->insn | |
3535 | && INSN_P (s_info->redundant_reason->insn)) | |
3536 | { | |
cccc26f6 | 3537 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
1ca59310 | 3538 | if (dump_file && (dump_flags & TDF_DETAILS)) |
aa140b76 | 3539 | fprintf (dump_file, "Locally deleting insn %d " |
3540 | "because insn %d stores the " | |
3541 | "same value and couldn't be " | |
3542 | "eliminated\n", | |
3543 | INSN_UID (insn_info->insn), | |
3544 | INSN_UID (rinsn)); | |
3545 | delete_dead_store_insn (insn_info); | |
3546 | } | |
3547 | } | |
3548 | insn_info = insn_info->prev_insn; | |
3549 | } | |
3550 | } | |
3551 | } | |
3552 | \f | |
3553 | /*---------------------------------------------------------------------------- | |
3554 | Seventh step. | |
3555 | ||
48e1416a | 3556 | Destroy everything left standing. |
3072d30e | 3557 | ----------------------------------------------------------------------------*/ |
3558 | ||
48e1416a | 3559 | static void |
4fb07d00 | 3560 | dse_step7 (void) |
3072d30e | 3561 | { |
4fb07d00 | 3562 | bitmap_obstack_release (&dse_bitmap_obstack); |
3563 | obstack_free (&dse_obstack, NULL); | |
ce299759 | 3564 | |
3072d30e | 3565 | end_alias_analysis (); |
3566 | free (bb_table); | |
c1f445d2 | 3567 | delete rtx_group_table; |
3568 | rtx_group_table = NULL; | |
f1f41a6c | 3569 | rtx_group_vec.release (); |
3072d30e | 3570 | BITMAP_FREE (all_blocks); |
3571 | BITMAP_FREE (scratch); | |
3572 | ||
55c5ac9f | 3573 | rtx_store_info_pool.release (); |
e16712b1 | 3574 | read_info_type_pool.release (); |
3575 | insn_info_type_pool.release (); | |
3576 | dse_bb_info_type_pool.release (); | |
3577 | group_info_pool.release (); | |
3578 | deferred_change_pool.release (); | |
3072d30e | 3579 | } |
3580 | ||
3581 | ||
3072d30e | 3582 | /* ------------------------------------------------------------------------- |
3583 | DSE | |
3584 | ------------------------------------------------------------------------- */ | |
3585 | ||
3586 | /* Callback for running pass_rtl_dse. */ | |
3587 | ||
3588 | static unsigned int | |
3589 | rest_of_handle_dse (void) | |
3590 | { | |
3072d30e | 3591 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3592 | ||
a1b0a968 | 3593 | /* Need the notes since we must track live hardregs in the forwards |
3594 | direction. */ | |
3595 | df_note_add_problem (); | |
3596 | df_analyze (); | |
3597 | ||
3072d30e | 3598 | dse_step0 (); |
3599 | dse_step1 (); | |
3600 | dse_step2_init (); | |
5a3f5d69 | 3601 | if (dse_step2 ()) |
3072d30e | 3602 | { |
3603 | df_set_flags (DF_LR_RUN_DCE); | |
3604 | df_analyze (); | |
1ca59310 | 3605 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3072d30e | 3606 | fprintf (dump_file, "doing global processing\n"); |
5a3f5d69 | 3607 | dse_step3 (); |
3072d30e | 3608 | dse_step4 (); |
5a3f5d69 | 3609 | dse_step5 (); |
3072d30e | 3610 | } |
3611 | ||
aa140b76 | 3612 | dse_step6 (); |
4fb07d00 | 3613 | dse_step7 (); |
3072d30e | 3614 | |
3615 | if (dump_file) | |
5a3f5d69 | 3616 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d\n", |
3617 | locally_deleted, globally_deleted); | |
1f91a12d | 3618 | |
3619 | /* DSE can eliminate potentially-trapping MEMs. | |
3620 | Remove any EH edges associated with them. */ | |
3621 | if ((locally_deleted || globally_deleted) | |
3622 | && cfun->can_throw_non_call_exceptions | |
3623 | && purge_all_dead_edges ()) | |
150ceef9 | 3624 | { |
3625 | free_dominance_info (CDI_DOMINATORS); | |
3626 | cleanup_cfg (0); | |
3627 | } | |
1f91a12d | 3628 | |
3072d30e | 3629 | return 0; |
3630 | } | |
3631 | ||
7620bc82 | 3632 | namespace { |
3633 | ||
3634 | const pass_data pass_data_rtl_dse1 = | |
cbe8bda8 | 3635 | { |
3636 | RTL_PASS, /* type */ | |
3637 | "dse1", /* name */ | |
3638 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 3639 | TV_DSE1, /* tv_id */ |
3640 | 0, /* properties_required */ | |
3641 | 0, /* properties_provided */ | |
3642 | 0, /* properties_destroyed */ | |
3643 | 0, /* todo_flags_start */ | |
8b88439e | 3644 | TODO_df_finish, /* todo_flags_finish */ |
3072d30e | 3645 | }; |
3646 | ||
7620bc82 | 3647 | class pass_rtl_dse1 : public rtl_opt_pass |
cbe8bda8 | 3648 | { |
3649 | public: | |
9af5ce0c | 3650 | pass_rtl_dse1 (gcc::context *ctxt) |
3651 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
cbe8bda8 | 3652 | {} |
3653 | ||
3654 | /* opt_pass methods: */ | |
31315c24 | 3655 | virtual bool gate (function *) |
3656 | { | |
3657 | return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
3658 | } | |
3659 | ||
65b0537f | 3660 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
cbe8bda8 | 3661 | |
3662 | }; // class pass_rtl_dse1 | |
3663 | ||
7620bc82 | 3664 | } // anon namespace |
3665 | ||
cbe8bda8 | 3666 | rtl_opt_pass * |
3667 | make_pass_rtl_dse1 (gcc::context *ctxt) | |
3668 | { | |
3669 | return new pass_rtl_dse1 (ctxt); | |
3670 | } | |
3671 | ||
7620bc82 | 3672 | namespace { |
3673 | ||
3674 | const pass_data pass_data_rtl_dse2 = | |
cbe8bda8 | 3675 | { |
3676 | RTL_PASS, /* type */ | |
3677 | "dse2", /* name */ | |
3678 | OPTGROUP_NONE, /* optinfo_flags */ | |
cbe8bda8 | 3679 | TV_DSE2, /* tv_id */ |
3680 | 0, /* properties_required */ | |
3681 | 0, /* properties_provided */ | |
3682 | 0, /* properties_destroyed */ | |
3683 | 0, /* todo_flags_start */ | |
8b88439e | 3684 | TODO_df_finish, /* todo_flags_finish */ |
3072d30e | 3685 | }; |
cbe8bda8 | 3686 | |
7620bc82 | 3687 | class pass_rtl_dse2 : public rtl_opt_pass |
cbe8bda8 | 3688 | { |
3689 | public: | |
9af5ce0c | 3690 | pass_rtl_dse2 (gcc::context *ctxt) |
3691 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
cbe8bda8 | 3692 | {} |
3693 | ||
3694 | /* opt_pass methods: */ | |
31315c24 | 3695 | virtual bool gate (function *) |
3696 | { | |
3697 | return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
3698 | } | |
3699 | ||
65b0537f | 3700 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
cbe8bda8 | 3701 | |
3702 | }; // class pass_rtl_dse2 | |
3703 | ||
7620bc82 | 3704 | } // anon namespace |
3705 | ||
cbe8bda8 | 3706 | rtl_opt_pass * |
3707 | make_pass_rtl_dse2 (gcc::context *ctxt) | |
3708 | { | |
3709 | return new pass_rtl_dse2 (ctxt); | |
3710 | } |