]>
Commit | Line | Data |
---|---|---|
6fb5fa3c | 1 | /* RTL dead store elimination. |
85ec4feb | 2 | Copyright (C) 2005-2018 Free Software Foundation, Inc. |
6fb5fa3c DB |
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 | |
9dcd6f09 | 11 | Software Foundation; either version 3, or (at your option) any later |
6fb5fa3c DB |
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 | |
9dcd6f09 NC |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ | |
6fb5fa3c DB |
22 | |
23 | #undef BASELINE | |
24 | ||
25 | #include "config.h" | |
26 | #include "system.h" | |
27 | #include "coretypes.h" | |
c7131fb2 | 28 | #include "backend.h" |
957060b5 AM |
29 | #include "target.h" |
30 | #include "rtl.h" | |
c7131fb2 AM |
31 | #include "tree.h" |
32 | #include "gimple.h" | |
957060b5 | 33 | #include "predict.h" |
c7131fb2 | 34 | #include "df.h" |
4d0cdd0c | 35 | #include "memmodel.h" |
957060b5 AM |
36 | #include "tm_p.h" |
37 | #include "gimple-ssa.h" | |
38 | #include "expmed.h" | |
39 | #include "optabs.h" | |
957060b5 AM |
40 | #include "emit-rtl.h" |
41 | #include "recog.h" | |
40e23961 | 42 | #include "alias.h" |
d8a2d370 | 43 | #include "stor-layout.h" |
60393bbc | 44 | #include "cfgrtl.h" |
6fb5fa3c | 45 | #include "cselib.h" |
6fb5fa3c | 46 | #include "tree-pass.h" |
36566b39 | 47 | #include "explow.h" |
6fb5fa3c | 48 | #include "expr.h" |
6fb5fa3c | 49 | #include "dbgcnt.h" |
dabd47e7 | 50 | #include "params.h" |
d7111da8 | 51 | #include "rtl-iter.h" |
6c5ed3f1 | 52 | #include "cfgcleanup.h" |
6fb5fa3c DB |
53 | |
54 | /* This file contains three techniques for performing Dead Store | |
b8698a0f | 55 | Elimination (dse). |
6fb5fa3c DB |
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. | |
b8698a0f | 61 | |
6fb5fa3c DB |
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) | |
026c3cfd | 67 | processes the spill slots. This differs from the second |
6fb5fa3c DB |
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 | ||
b8698a0f | 91 | 3) Set up the global dataflow equations based on processing the |
6fb5fa3c DB |
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 | ||
073a8998 | 99 | 6) Delete insns that store the same value as preceding store |
8dd5516b JJ |
100 | where the earlier store couldn't be eliminated. |
101 | ||
102 | 7) Cleanup. | |
6fb5fa3c DB |
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 | |
0d52bcc1 | 108 | forwards ordering is to accommodate cselib. |
6fb5fa3c | 109 | |
5ef0b50d | 110 | We make a simplifying assumption: addresses fall into four broad |
6fb5fa3c DB |
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 | |
5ef0b50d | 119 | global pass only handles 1). |
6fb5fa3c DB |
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 | |
5ef0b50d | 125 | function contain another store to A before a read to A. |
6fb5fa3c DB |
126 | |
127 | If the address A is relative to the stack frame, a store S2 to A | |
5ef0b50d | 128 | can be eliminated if there are no paths from S2 that reach the |
6fb5fa3c | 129 | end of the function that read A before another store to A. In |
5ef0b50d | 130 | this case S2 can be deleted if there are paths from S2 to the |
6fb5fa3c DB |
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 | |
5ef0b50d | 141 | implementation requires a lot of bitmaps filled with 1s. |
6fb5fa3c DB |
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 | |
b8698a0f | 145 | for details. |
6fb5fa3c DB |
146 | |
147 | There are two places for further enhancements to this algorithm: | |
b8698a0f | 148 | |
6fb5fa3c DB |
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 | |
6ed3da00 | 156 | reference to r100. Most of the information is available to add this |
6fb5fa3c DB |
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 | |
b8698a0f | 190 | off the frame pointer, the global algorithm handles this slot. |
6fb5fa3c DB |
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 | |
0d52bcc1 | 195 | required is to have those passes make the same calls that reload |
6fb5fa3c DB |
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 | ||
3f9b14ff SB |
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. */ | |
6fb5fa3c | 218 | static bitmap scratch = NULL; |
3f9b14ff | 219 | |
da6603c6 | 220 | struct insn_info_type; |
6fb5fa3c DB |
221 | |
222 | /* This structure holds information about a candidate store. */ | |
b8698a0f | 223 | struct store_info |
6fb5fa3c DB |
224 | { |
225 | ||
226 | /* False means this is a clobber. */ | |
227 | bool is_set; | |
228 | ||
8dd5516b JJ |
229 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
230 | bool is_large; | |
231 | ||
6fb5fa3c DB |
232 | /* The id of the mem group of the base address. If rtx_varies_p is |
233 | true, this is -1. Otherwise, it is the index into the group | |
234 | table. */ | |
235 | int group_id; | |
b8698a0f | 236 | |
6fb5fa3c DB |
237 | /* This is the cselib value. */ |
238 | cselib_val *cse_base; | |
239 | ||
240 | /* This canonized mem. */ | |
241 | rtx mem; | |
242 | ||
6216f94e | 243 | /* Canonized MEM address for use by canon_true_dependence. */ |
6fb5fa3c DB |
244 | rtx mem_addr; |
245 | ||
71d6a386 | 246 | /* The offset of the first byte associated with the operation. */ |
02ce5d90 | 247 | poly_int64 offset; |
71d6a386 RS |
248 | |
249 | /* The number of bytes covered by the operation. This is always exact | |
250 | and known (rather than -1). */ | |
02ce5d90 | 251 | poly_int64 width; |
8dd5516b JJ |
252 | |
253 | union | |
254 | { | |
255 | /* A bitmask as wide as the number of bytes in the word that | |
256 | contains a 1 if the byte may be needed. The store is unused if | |
257 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
258 | unsigned HOST_WIDE_INT small_bitmask; | |
259 | ||
260 | struct | |
261 | { | |
02ce5d90 RS |
262 | /* A bitmap with one bit per byte, or null if the number of |
263 | bytes isn't known at compile time. A cleared bit means | |
264 | the position is needed. Used if IS_LARGE is true. */ | |
dc491a25 | 265 | bitmap bmap; |
6fb5fa3c | 266 | |
02ce5d90 RS |
267 | /* When BITMAP is nonnull, this counts the number of set bits |
268 | (i.e. unneeded bytes) in the bitmap. If it is equal to | |
269 | WIDTH, the whole store is unused. | |
270 | ||
271 | When BITMAP is null: | |
272 | - the store is definitely not needed when COUNT == 1 | |
273 | - all the store is needed when COUNT == 0 and RHS is nonnull | |
274 | - otherwise we don't know which parts of the store are needed. */ | |
8dd5516b JJ |
275 | int count; |
276 | } large; | |
277 | } positions_needed; | |
6fb5fa3c DB |
278 | |
279 | /* The next store info for this insn. */ | |
280 | struct store_info *next; | |
281 | ||
282 | /* The right hand side of the store. This is used if there is a | |
283 | subsequent reload of the mems address somewhere later in the | |
284 | basic block. */ | |
8dd5516b JJ |
285 | rtx rhs; |
286 | ||
287 | /* If rhs is or holds a constant, this contains that constant, | |
288 | otherwise NULL. */ | |
289 | rtx const_rhs; | |
290 | ||
291 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
292 | insn stored. These aren't eliminated early, because doing that | |
293 | might prevent the earlier larger store to be eliminated. */ | |
da6603c6 | 294 | struct insn_info_type *redundant_reason; |
6fb5fa3c DB |
295 | }; |
296 | ||
4fe663b0 L |
297 | /* Return a bitmask with the first N low bits set. */ |
298 | ||
299 | static unsigned HOST_WIDE_INT | |
300 | lowpart_bitmask (int n) | |
301 | { | |
dd4786fe | 302 | unsigned HOST_WIDE_INT mask = HOST_WIDE_INT_M1U; |
4fe663b0 L |
303 | return mask >> (HOST_BITS_PER_WIDE_INT - n); |
304 | } | |
305 | ||
fcb87c50 | 306 | static object_allocator<store_info> cse_store_info_pool ("cse_store_info_pool"); |
da6603c6 | 307 | |
fcb87c50 | 308 | static object_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool"); |
6fb5fa3c DB |
309 | |
310 | /* This structure holds information about a load. These are only | |
311 | built for rtx bases. */ | |
da6603c6 | 312 | struct read_info_type |
6fb5fa3c DB |
313 | { |
314 | /* The id of the mem group of the base address. */ | |
315 | int group_id; | |
316 | ||
71d6a386 | 317 | /* The offset of the first byte associated with the operation. */ |
02ce5d90 | 318 | poly_int64 offset; |
71d6a386 RS |
319 | |
320 | /* The number of bytes covered by the operation, or -1 if not known. */ | |
02ce5d90 | 321 | poly_int64 width; |
6fb5fa3c DB |
322 | |
323 | /* The mem being read. */ | |
324 | rtx mem; | |
325 | ||
326 | /* The next read_info for this insn. */ | |
da6603c6 | 327 | struct read_info_type *next; |
6fb5fa3c | 328 | }; |
da6603c6 | 329 | typedef struct read_info_type *read_info_t; |
6fb5fa3c | 330 | |
fcb87c50 | 331 | static object_allocator<read_info_type> read_info_type_pool ("read_info_pool"); |
6fb5fa3c DB |
332 | |
333 | /* One of these records is created for each insn. */ | |
334 | ||
da6603c6 | 335 | struct insn_info_type |
6fb5fa3c DB |
336 | { |
337 | /* Set true if the insn contains a store but the insn itself cannot | |
338 | be deleted. This is set if the insn is a parallel and there is | |
339 | more than one non dead output or if the insn is in some way | |
340 | volatile. */ | |
341 | bool cannot_delete; | |
342 | ||
343 | /* This field is only used by the global algorithm. It is set true | |
344 | if the insn contains any read of mem except for a (1). This is | |
345 | also set if the insn is a call or has a clobber mem. If the insn | |
346 | contains a wild read, the use_rec will be null. */ | |
347 | bool wild_read; | |
348 | ||
d26c7090 ER |
349 | /* This is true only for CALL instructions which could potentially read |
350 | any non-frame memory location. This field is used by the global | |
351 | algorithm. */ | |
352 | bool non_frame_wild_read; | |
353 | ||
50f0f366 EB |
354 | /* This field is only used for the processing of const functions. |
355 | These functions cannot read memory, but they can read the stack | |
64520bdc EB |
356 | because that is where they may get their parms. We need to be |
357 | this conservative because, like the store motion pass, we don't | |
358 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
359 | Moreover, we need to distinguish two cases: | |
360 | 1. Before reload (register elimination), the stores related to | |
361 | outgoing arguments are stack pointer based and thus deemed | |
362 | of non-constant base in this pass. This requires special | |
363 | handling but also means that the frame pointer based stores | |
364 | need not be killed upon encountering a const function call. | |
365 | 2. After reload, the stores related to outgoing arguments can be | |
366 | either stack pointer or hard frame pointer based. This means | |
367 | that we have no other choice than also killing all the frame | |
368 | pointer based stores upon encountering a const function call. | |
57534689 JJ |
369 | This field is set after reload for const function calls and before |
370 | reload for const tail function calls on targets where arg pointer | |
371 | is the frame pointer. Having this set is less severe than a wild | |
372 | read, it just means that all the frame related stores are killed | |
373 | rather than all the stores. */ | |
64520bdc EB |
374 | bool frame_read; |
375 | ||
376 | /* This field is only used for the processing of const functions. | |
377 | It is set if the insn may contain a stack pointer based store. */ | |
50f0f366 | 378 | bool stack_pointer_based; |
6fb5fa3c DB |
379 | |
380 | /* This is true if any of the sets within the store contains a | |
381 | cselib base. Such stores can only be deleted by the local | |
382 | algorithm. */ | |
383 | bool contains_cselib_groups; | |
384 | ||
385 | /* The insn. */ | |
dd60a84c | 386 | rtx_insn *insn; |
6fb5fa3c DB |
387 | |
388 | /* The list of mem sets or mem clobbers that are contained in this | |
389 | insn. If the insn is deletable, it contains only one mem set. | |
390 | But it could also contain clobbers. Insns that contain more than | |
391 | one mem set are not deletable, but each of those mems are here in | |
6ed3da00 | 392 | order to provide info to delete other insns. */ |
fd6daec9 | 393 | store_info *store_rec; |
6fb5fa3c DB |
394 | |
395 | /* The linked list of mem uses in this insn. Only the reads from | |
396 | rtx bases are listed here. The reads to cselib bases are | |
397 | completely processed during the first scan and so are never | |
398 | created. */ | |
399 | read_info_t read_rec; | |
400 | ||
9e582b1d JR |
401 | /* The live fixed registers. We assume only fixed registers can |
402 | cause trouble by being clobbered from an expanded pattern; | |
403 | storing only the live fixed registers (rather than all registers) | |
404 | means less memory needs to be allocated / copied for the individual | |
405 | stores. */ | |
406 | regset fixed_regs_live; | |
407 | ||
6fb5fa3c | 408 | /* The prev insn in the basic block. */ |
da6603c6 | 409 | struct insn_info_type * prev_insn; |
6fb5fa3c DB |
410 | |
411 | /* The linked list of insns that are in consideration for removal in | |
073a8998 | 412 | the forwards pass through the basic block. This pointer may be |
6fb5fa3c | 413 | trash as it is not cleared when a wild read occurs. The only |
fa10beec | 414 | time it is guaranteed to be correct is when the traversal starts |
6fb5fa3c | 415 | at active_local_stores. */ |
da6603c6 | 416 | struct insn_info_type * next_local_store; |
6fb5fa3c | 417 | }; |
da6603c6 | 418 | typedef struct insn_info_type *insn_info_t; |
6fb5fa3c | 419 | |
fcb87c50 | 420 | static object_allocator<insn_info_type> insn_info_type_pool ("insn_info_pool"); |
6fb5fa3c DB |
421 | |
422 | /* The linked list of stores that are under consideration in this | |
b8698a0f | 423 | basic block. */ |
6fb5fa3c | 424 | static insn_info_t active_local_stores; |
dabd47e7 | 425 | static int active_local_stores_len; |
6fb5fa3c | 426 | |
da6603c6 | 427 | struct dse_bb_info_type |
6fb5fa3c | 428 | { |
6fb5fa3c DB |
429 | /* Pointer to the insn info for the last insn in the block. These |
430 | are linked so this is how all of the insns are reached. During | |
431 | scanning this is the current insn being scanned. */ | |
432 | insn_info_t last_insn; | |
433 | ||
434 | /* The info for the global dataflow problem. */ | |
435 | ||
436 | ||
437 | /* This is set if the transfer function should and in the wild_read | |
438 | bitmap before applying the kill and gen sets. That vector knocks | |
439 | out most of the bits in the bitmap and thus speeds up the | |
440 | operations. */ | |
441 | bool apply_wild_read; | |
442 | ||
02b47899 KZ |
443 | /* The following 4 bitvectors hold information about which positions |
444 | of which stores are live or dead. They are indexed by | |
445 | get_bitmap_index. */ | |
446 | ||
6fb5fa3c DB |
447 | /* The set of store positions that exist in this block before a wild read. */ |
448 | bitmap gen; | |
b8698a0f | 449 | |
6fb5fa3c DB |
450 | /* The set of load positions that exist in this block above the |
451 | same position of a store. */ | |
452 | bitmap kill; | |
453 | ||
454 | /* The set of stores that reach the top of the block without being | |
455 | killed by a read. | |
456 | ||
457 | Do not represent the in if it is all ones. Note that this is | |
458 | what the bitvector should logically be initialized to for a set | |
459 | intersection problem. However, like the kill set, this is too | |
460 | expensive. So initially, the in set will only be created for the | |
461 | exit block and any block that contains a wild read. */ | |
462 | bitmap in; | |
463 | ||
464 | /* The set of stores that reach the bottom of the block from it's | |
465 | successors. | |
466 | ||
467 | Do not represent the in if it is all ones. Note that this is | |
468 | what the bitvector should logically be initialized to for a set | |
469 | intersection problem. However, like the kill and in set, this is | |
470 | too expensive. So what is done is that the confluence operator | |
471 | just initializes the vector from one of the out sets of the | |
472 | successors of the block. */ | |
473 | bitmap out; | |
02b47899 KZ |
474 | |
475 | /* The following bitvector is indexed by the reg number. It | |
476 | contains the set of regs that are live at the current instruction | |
477 | being processed. While it contains info for all of the | |
9e582b1d JR |
478 | registers, only the hard registers are actually examined. It is used |
479 | to assure that shift and/or add sequences that are inserted do not | |
073a8998 | 480 | accidentally clobber live hard regs. */ |
02b47899 | 481 | bitmap regs_live; |
6fb5fa3c DB |
482 | }; |
483 | ||
da6603c6 | 484 | typedef struct dse_bb_info_type *bb_info_t; |
fb0b2914 ML |
485 | |
486 | static object_allocator<dse_bb_info_type> dse_bb_info_type_pool | |
fcb87c50 | 487 | ("bb_info_pool"); |
6fb5fa3c DB |
488 | |
489 | /* Table to hold all bb_infos. */ | |
490 | static bb_info_t *bb_table; | |
491 | ||
492 | /* There is a group_info for each rtx base that is used to reference | |
493 | memory. There are also not many of the rtx bases because they are | |
494 | very limited in scope. */ | |
495 | ||
b8698a0f | 496 | struct group_info |
6fb5fa3c DB |
497 | { |
498 | /* The actual base of the address. */ | |
499 | rtx rtx_base; | |
500 | ||
501 | /* The sequential id of the base. This allows us to have a | |
502 | canonical ordering of these that is not based on addresses. */ | |
503 | int id; | |
504 | ||
8f5929e1 JJ |
505 | /* True if there are any positions that are to be processed |
506 | globally. */ | |
507 | bool process_globally; | |
508 | ||
509 | /* True if the base of this group is either the frame_pointer or | |
510 | hard_frame_pointer. */ | |
511 | bool frame_related; | |
512 | ||
d32f725a OH |
513 | /* A mem wrapped around the base pointer for the group in order to do |
514 | read dependency. It must be given BLKmode in order to encompass all | |
515 | the possible offsets from the base. */ | |
6fb5fa3c | 516 | rtx base_mem; |
b8698a0f | 517 | |
6216f94e JJ |
518 | /* Canonized version of base_mem's address. */ |
519 | rtx canon_base_addr; | |
6fb5fa3c DB |
520 | |
521 | /* These two sets of two bitmaps are used to keep track of how many | |
6ed3da00 | 522 | stores are actually referencing that position from this base. We |
6fb5fa3c | 523 | only do this for rtx bases as this will be used to assign |
6ed3da00 | 524 | positions in the bitmaps for the global problem. Bit N is set in |
6fb5fa3c DB |
525 | store1 on the first store for offset N. Bit N is set in store2 |
526 | for the second store to offset N. This is all we need since we | |
527 | only care about offsets that have two or more stores for them. | |
528 | ||
529 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
530 | for 0 and greater offsets. | |
531 | ||
532 | There is one special case here, for stores into the stack frame, | |
533 | we will or store1 into store2 before deciding which stores look | |
534 | at globally. This is because stores to the stack frame that have | |
535 | no other reads before the end of the function can also be | |
536 | deleted. */ | |
537 | bitmap store1_n, store1_p, store2_n, store2_p; | |
538 | ||
d26c7090 ER |
539 | /* These bitmaps keep track of offsets in this group escape this function. |
540 | An offset escapes if it corresponds to a named variable whose | |
541 | addressable flag is set. */ | |
542 | bitmap escaped_n, escaped_p; | |
543 | ||
6ed3da00 | 544 | /* The positions in this bitmap have the same assignments as the in, |
6fb5fa3c | 545 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6ed3da00 | 546 | the positions that are occupied by stores for this group. */ |
6fb5fa3c DB |
547 | bitmap group_kill; |
548 | ||
6fb5fa3c | 549 | /* The offset_map is used to map the offsets from this base into |
6ed3da00 | 550 | positions in the global bitmaps. It is only created after all of |
6fb5fa3c DB |
551 | the all of stores have been scanned and we know which ones we |
552 | care about. */ | |
b8698a0f L |
553 | int *offset_map_n, *offset_map_p; |
554 | int offset_map_size_n, offset_map_size_p; | |
6fb5fa3c | 555 | }; |
da6603c6 | 556 | |
fcb87c50 | 557 | static object_allocator<group_info> group_info_pool ("rtx_group_info_pool"); |
6fb5fa3c | 558 | |
6fb5fa3c DB |
559 | /* Index into the rtx_group_vec. */ |
560 | static int rtx_group_next_id; | |
561 | ||
6fb5fa3c | 562 | |
fd6daec9 | 563 | static vec<group_info *> rtx_group_vec; |
6fb5fa3c DB |
564 | |
565 | ||
566 | /* This structure holds the set of changes that are being deferred | |
567 | when removing read operation. See replace_read. */ | |
b8698a0f | 568 | struct deferred_change |
6fb5fa3c DB |
569 | { |
570 | ||
571 | /* The mem that is being replaced. */ | |
572 | rtx *loc; | |
573 | ||
574 | /* The reg it is being replaced with. */ | |
575 | rtx reg; | |
576 | ||
577 | struct deferred_change *next; | |
578 | }; | |
579 | ||
fb0b2914 | 580 | static object_allocator<deferred_change> deferred_change_pool |
fcb87c50 | 581 | ("deferred_change_pool"); |
6fb5fa3c | 582 | |
fd6daec9 | 583 | static deferred_change *deferred_change_list = NULL; |
6fb5fa3c | 584 | |
e3b5732b | 585 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
9dd9bf80 | 586 | this for vararg functions because they play games with the frame. */ |
6fb5fa3c DB |
587 | static bool stores_off_frame_dead_at_return; |
588 | ||
589 | /* Counter for stats. */ | |
b8698a0f L |
590 | static int globally_deleted; |
591 | static int locally_deleted; | |
b8698a0f | 592 | |
6fb5fa3c DB |
593 | static bitmap all_blocks; |
594 | ||
d26c7090 ER |
595 | /* Locations that are killed by calls in the global phase. */ |
596 | static bitmap kill_on_calls; | |
597 | ||
6fb5fa3c DB |
598 | /* The number of bits used in the global bitmaps. */ |
599 | static unsigned int current_position; | |
71d6a386 RS |
600 | |
601 | /* Print offset range [OFFSET, OFFSET + WIDTH) to FILE. */ | |
602 | ||
603 | static void | |
604 | print_range (FILE *file, poly_int64 offset, poly_int64 width) | |
605 | { | |
606 | fprintf (file, "["); | |
607 | print_dec (offset, file, SIGNED); | |
608 | fprintf (file, ".."); | |
609 | print_dec (offset + width, file, SIGNED); | |
610 | fprintf (file, ")"); | |
611 | } | |
6fb5fa3c DB |
612 | \f |
613 | /*---------------------------------------------------------------------------- | |
614 | Zeroth step. | |
615 | ||
b8698a0f | 616 | Initialization. |
6fb5fa3c DB |
617 | ----------------------------------------------------------------------------*/ |
618 | ||
6fb5fa3c | 619 | |
6fb5fa3c DB |
620 | /* Hashtable callbacks for maintaining the "bases" field of |
621 | store_group_info, given that the addresses are function invariants. */ | |
622 | ||
8d67ee55 | 623 | struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
703c8606 | 624 | { |
67f58944 TS |
625 | static inline hashval_t hash (const group_info *); |
626 | static inline bool equal (const group_info *, const group_info *); | |
703c8606 LC |
627 | }; |
628 | ||
629 | inline bool | |
67f58944 TS |
630 | invariant_group_base_hasher::equal (const group_info *gi1, |
631 | const group_info *gi2) | |
6fb5fa3c | 632 | { |
6fb5fa3c DB |
633 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
634 | } | |
635 | ||
703c8606 | 636 | inline hashval_t |
67f58944 | 637 | invariant_group_base_hasher::hash (const group_info *gi) |
6fb5fa3c | 638 | { |
6fb5fa3c DB |
639 | int do_not_record; |
640 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
641 | } | |
642 | ||
703c8606 | 643 | /* Tables of group_info structures, hashed by base value. */ |
c203e8a7 | 644 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
703c8606 | 645 | |
6fb5fa3c DB |
646 | |
647 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
648 | ||
fd6daec9 | 649 | static group_info * |
6fb5fa3c DB |
650 | get_group_info (rtx base) |
651 | { | |
b8698a0f | 652 | struct group_info tmp_gi; |
fd6daec9 | 653 | group_info *gi; |
703c8606 | 654 | group_info **slot; |
6fb5fa3c | 655 | |
ac6929b5 RB |
656 | gcc_assert (base != NULL_RTX); |
657 | ||
658 | /* Find the store_base_info structure for BASE, creating a new one | |
659 | if necessary. */ | |
660 | tmp_gi.rtx_base = base; | |
661 | slot = rtx_group_table->find_slot (&tmp_gi, INSERT); | |
662 | gi = *slot; | |
6fb5fa3c DB |
663 | |
664 | if (gi == NULL) | |
665 | { | |
fb0b2914 | 666 | *slot = gi = group_info_pool.allocate (); |
6fb5fa3c DB |
667 | gi->rtx_base = base; |
668 | gi->id = rtx_group_next_id++; | |
d32f725a | 669 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
6216f94e | 670 | gi->canon_base_addr = canon_rtx (base); |
3f9b14ff SB |
671 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
672 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
673 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
674 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
675 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
676 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
677 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 678 | gi->process_globally = false; |
b8698a0f | 679 | gi->frame_related = |
6fb5fa3c DB |
680 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
681 | gi->offset_map_size_n = 0; | |
682 | gi->offset_map_size_p = 0; | |
683 | gi->offset_map_n = NULL; | |
684 | gi->offset_map_p = NULL; | |
9771b263 | 685 | rtx_group_vec.safe_push (gi); |
6fb5fa3c DB |
686 | } |
687 | ||
688 | return gi; | |
689 | } | |
690 | ||
691 | ||
692 | /* Initialization of data structures. */ | |
693 | ||
694 | static void | |
695 | dse_step0 (void) | |
696 | { | |
697 | locally_deleted = 0; | |
698 | globally_deleted = 0; | |
6fb5fa3c | 699 | |
3f9b14ff SB |
700 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
701 | gcc_obstack_init (&dse_obstack); | |
702 | ||
703 | scratch = BITMAP_ALLOC (®_obstack); | |
704 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 705 | |
6fb5fa3c | 706 | |
c203e8a7 | 707 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
6fb5fa3c | 708 | |
8b1c6fd7 | 709 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
6fb5fa3c DB |
710 | rtx_group_next_id = 0; |
711 | ||
e3b5732b | 712 | stores_off_frame_dead_at_return = !cfun->stdarg; |
6fb5fa3c DB |
713 | |
714 | init_alias_analysis (); | |
6fb5fa3c DB |
715 | } |
716 | ||
717 | ||
718 | \f | |
719 | /*---------------------------------------------------------------------------- | |
720 | First step. | |
721 | ||
722 | Scan all of the insns. Any random ordering of the blocks is fine. | |
0d52bcc1 | 723 | Each block is scanned in forward order to accommodate cselib which |
6fb5fa3c DB |
724 | is used to remove stores with non-constant bases. |
725 | ----------------------------------------------------------------------------*/ | |
726 | ||
727 | /* Delete all of the store_info recs from INSN_INFO. */ | |
728 | ||
b8698a0f | 729 | static void |
6fb5fa3c DB |
730 | free_store_info (insn_info_t insn_info) |
731 | { | |
fd6daec9 TS |
732 | store_info *cur = insn_info->store_rec; |
733 | while (cur) | |
6fb5fa3c | 734 | { |
fd6daec9 TS |
735 | store_info *next = cur->next; |
736 | if (cur->is_large) | |
737 | BITMAP_FREE (cur->positions_needed.large.bmap); | |
738 | if (cur->cse_base) | |
739 | cse_store_info_pool.remove (cur); | |
6fb5fa3c | 740 | else |
fd6daec9 TS |
741 | rtx_store_info_pool.remove (cur); |
742 | cur = next; | |
6fb5fa3c DB |
743 | } |
744 | ||
745 | insn_info->cannot_delete = true; | |
746 | insn_info->contains_cselib_groups = false; | |
747 | insn_info->store_rec = NULL; | |
748 | } | |
749 | ||
a79683d5 | 750 | struct note_add_store_info |
9e582b1d | 751 | { |
dc01c3d1 | 752 | rtx_insn *first, *current; |
9e582b1d JR |
753 | regset fixed_regs_live; |
754 | bool failure; | |
a79683d5 | 755 | }; |
9e582b1d JR |
756 | |
757 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
758 | Check if a register is clobbered which is live afterwards. */ | |
759 | ||
760 | static void | |
761 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
762 | { | |
dc01c3d1 | 763 | rtx_insn *insn; |
9e582b1d | 764 | note_add_store_info *info = (note_add_store_info *) data; |
9e582b1d JR |
765 | |
766 | if (!REG_P (loc)) | |
767 | return; | |
768 | ||
769 | /* If this register is referenced by the current or an earlier insn, | |
770 | that's OK. E.g. this applies to the register that is being incremented | |
771 | with this addition. */ | |
772 | for (insn = info->first; | |
773 | insn != NEXT_INSN (info->current); | |
774 | insn = NEXT_INSN (insn)) | |
775 | if (reg_referenced_p (loc, PATTERN (insn))) | |
776 | return; | |
777 | ||
778 | /* If we come here, we have a clobber of a register that's only OK | |
779 | if that register is not live. If we don't have liveness information | |
780 | available, fail now. */ | |
781 | if (!info->fixed_regs_live) | |
782 | { | |
53d1bae9 | 783 | info->failure = true; |
9e582b1d JR |
784 | return; |
785 | } | |
786 | /* Now check if this is a live fixed register. */ | |
53d1bae9 RS |
787 | unsigned int end_regno = END_REGNO (loc); |
788 | for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) | |
789 | if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) | |
790 | info->failure = true; | |
9e582b1d JR |
791 | } |
792 | ||
4deef538 AO |
793 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
794 | SRC + SRCOFF before insn ARG. */ | |
6fb5fa3c DB |
795 | |
796 | static int | |
4deef538 AO |
797 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
798 | rtx op ATTRIBUTE_UNUSED, | |
799 | rtx dest, rtx src, rtx srcoff, void *arg) | |
6fb5fa3c | 800 | { |
9e582b1d | 801 | insn_info_t insn_info = (insn_info_t) arg; |
dc01c3d1 | 802 | rtx_insn *insn = insn_info->insn, *new_insn, *cur; |
9e582b1d | 803 | note_add_store_info info; |
b8698a0f | 804 | |
4deef538 AO |
805 | /* We can reuse all operands without copying, because we are about |
806 | to delete the insn that contained it. */ | |
9e582b1d | 807 | if (srcoff) |
ed079c4b JJ |
808 | { |
809 | start_sequence (); | |
810 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
811 | new_insn = get_insns (); | |
812 | end_sequence (); | |
813 | } | |
9e582b1d | 814 | else |
1476d1bd | 815 | new_insn = gen_move_insn (dest, src); |
9e582b1d JR |
816 | info.first = new_insn; |
817 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
818 | info.failure = false; | |
819 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
820 | { | |
821 | info.current = cur; | |
822 | note_stores (PATTERN (cur), note_add_store, &info); | |
823 | } | |
6fb5fa3c | 824 | |
9e582b1d JR |
825 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
826 | return it immediately, communicating the failure to its caller. */ | |
827 | if (info.failure) | |
828 | return 1; | |
829 | ||
830 | emit_insn_before (new_insn, insn); | |
6fb5fa3c | 831 | |
8d8e205b | 832 | return 0; |
6fb5fa3c DB |
833 | } |
834 | ||
9e582b1d JR |
835 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
836 | is there, is split into a separate insn. | |
837 | Return true on success (or if there was nothing to do), false on failure. */ | |
6fb5fa3c | 838 | |
9e582b1d JR |
839 | static bool |
840 | check_for_inc_dec_1 (insn_info_t insn_info) | |
6fb5fa3c | 841 | { |
dd60a84c | 842 | rtx_insn *insn = insn_info->insn; |
6fb5fa3c DB |
843 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
844 | if (note) | |
8d8e205b RS |
845 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
846 | insn_info) == 0; | |
9e582b1d | 847 | return true; |
6fb5fa3c DB |
848 | } |
849 | ||
850 | ||
9e582b1d JR |
851 | /* Entry point for postreload. If you work on reload_cse, or you need this |
852 | anywhere else, consider if you can provide register liveness information | |
853 | and add a parameter to this function so that it can be passed down in | |
854 | insn_info.fixed_regs_live. */ | |
855 | bool | |
dd60a84c | 856 | check_for_inc_dec (rtx_insn *insn) |
9e582b1d | 857 | { |
da6603c6 | 858 | insn_info_type insn_info; |
9e582b1d JR |
859 | rtx note; |
860 | ||
861 | insn_info.insn = insn; | |
862 | insn_info.fixed_regs_live = NULL; | |
863 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
864 | if (note) | |
8d8e205b RS |
865 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
866 | &insn_info) == 0; | |
9e582b1d JR |
867 | return true; |
868 | } | |
869 | ||
b8698a0f | 870 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
6fb5fa3c DB |
871 | |
872 | static void | |
873 | delete_dead_store_insn (insn_info_t insn_info) | |
874 | { | |
875 | read_info_t read_info; | |
876 | ||
877 | if (!dbg_cnt (dse)) | |
878 | return; | |
879 | ||
9e582b1d JR |
880 | if (!check_for_inc_dec_1 (insn_info)) |
881 | return; | |
456610d3 | 882 | if (dump_file && (dump_flags & TDF_DETAILS)) |
ac6929b5 RB |
883 | fprintf (dump_file, "Locally deleting insn %d\n", |
884 | INSN_UID (insn_info->insn)); | |
6fb5fa3c DB |
885 | |
886 | free_store_info (insn_info); | |
887 | read_info = insn_info->read_rec; | |
b8698a0f | 888 | |
6fb5fa3c DB |
889 | while (read_info) |
890 | { | |
891 | read_info_t next = read_info->next; | |
fb0b2914 | 892 | read_info_type_pool.remove (read_info); |
6fb5fa3c DB |
893 | read_info = next; |
894 | } | |
895 | insn_info->read_rec = NULL; | |
896 | ||
897 | delete_insn (insn_info->insn); | |
898 | locally_deleted++; | |
899 | insn_info->insn = NULL; | |
900 | ||
901 | insn_info->wild_read = false; | |
902 | } | |
903 | ||
88d8330d EB |
904 | /* Return whether DECL, a local variable, can possibly escape the current |
905 | function scope. */ | |
906 | ||
907 | static bool | |
908 | local_variable_can_escape (tree decl) | |
909 | { | |
910 | if (TREE_ADDRESSABLE (decl)) | |
911 | return true; | |
912 | ||
913 | /* If this is a partitioned variable, we need to consider all the variables | |
914 | in the partition. This is necessary because a store into one of them can | |
915 | be replaced with a store into another and this may not change the outcome | |
916 | of the escape analysis. */ | |
917 | if (cfun->gimple_df->decls_to_pointers != NULL) | |
918 | { | |
39c8aaa4 | 919 | tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); |
88d8330d | 920 | if (namep) |
39c8aaa4 | 921 | return TREE_ADDRESSABLE (*namep); |
88d8330d EB |
922 | } |
923 | ||
924 | return false; | |
925 | } | |
926 | ||
927 | /* Return whether EXPR can possibly escape the current function scope. */ | |
928 | ||
d26c7090 ER |
929 | static bool |
930 | can_escape (tree expr) | |
931 | { | |
932 | tree base; | |
933 | if (!expr) | |
934 | return true; | |
935 | base = get_base_address (expr); | |
936 | if (DECL_P (base) | |
88d8330d | 937 | && !may_be_aliased (base) |
8813a647 | 938 | && !(VAR_P (base) |
88d8330d EB |
939 | && !DECL_EXTERNAL (base) |
940 | && !TREE_STATIC (base) | |
941 | && local_variable_can_escape (base))) | |
d26c7090 ER |
942 | return false; |
943 | return true; | |
944 | } | |
6fb5fa3c DB |
945 | |
946 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
947 | OFFSET and WIDTH. */ | |
948 | ||
949 | static void | |
02ce5d90 | 950 | set_usage_bits (group_info *group, poly_int64 offset, poly_int64 width, |
d26c7090 | 951 | tree expr) |
6fb5fa3c | 952 | { |
02ce5d90 RS |
953 | /* Non-constant offsets and widths act as global kills, so there's no point |
954 | trying to use them to derive global DSE candidates. */ | |
955 | HOST_WIDE_INT i, const_offset, const_width; | |
d26c7090 | 956 | bool expr_escapes = can_escape (expr); |
02ce5d90 RS |
957 | if (offset.is_constant (&const_offset) |
958 | && width.is_constant (&const_width) | |
959 | && const_offset > -MAX_OFFSET | |
960 | && const_offset + const_width < MAX_OFFSET) | |
961 | for (i = const_offset; i < const_offset + const_width; ++i) | |
6fb5fa3c DB |
962 | { |
963 | bitmap store1; | |
964 | bitmap store2; | |
d26c7090 | 965 | bitmap escaped; |
6fb5fa3c DB |
966 | int ai; |
967 | if (i < 0) | |
968 | { | |
969 | store1 = group->store1_n; | |
970 | store2 = group->store2_n; | |
d26c7090 | 971 | escaped = group->escaped_n; |
6fb5fa3c DB |
972 | ai = -i; |
973 | } | |
974 | else | |
975 | { | |
976 | store1 = group->store1_p; | |
977 | store2 = group->store2_p; | |
d26c7090 | 978 | escaped = group->escaped_p; |
6fb5fa3c DB |
979 | ai = i; |
980 | } | |
b8698a0f | 981 | |
fcaa4ca4 | 982 | if (!bitmap_set_bit (store1, ai)) |
6fb5fa3c | 983 | bitmap_set_bit (store2, ai); |
b8698a0f | 984 | else |
6fb5fa3c | 985 | { |
6fb5fa3c DB |
986 | if (i < 0) |
987 | { | |
988 | if (group->offset_map_size_n < ai) | |
989 | group->offset_map_size_n = ai; | |
990 | } | |
991 | else | |
992 | { | |
993 | if (group->offset_map_size_p < ai) | |
994 | group->offset_map_size_p = ai; | |
995 | } | |
996 | } | |
d26c7090 ER |
997 | if (expr_escapes) |
998 | bitmap_set_bit (escaped, ai); | |
6fb5fa3c DB |
999 | } |
1000 | } | |
1001 | ||
d26c7090 ER |
1002 | static void |
1003 | reset_active_stores (void) | |
1004 | { | |
1005 | active_local_stores = NULL; | |
1006 | active_local_stores_len = 0; | |
1007 | } | |
6fb5fa3c | 1008 | |
d26c7090 | 1009 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
6fb5fa3c DB |
1010 | |
1011 | static void | |
d26c7090 | 1012 | free_read_records (bb_info_t bb_info) |
6fb5fa3c DB |
1013 | { |
1014 | insn_info_t insn_info = bb_info->last_insn; | |
1015 | read_info_t *ptr = &insn_info->read_rec; | |
6fb5fa3c DB |
1016 | while (*ptr) |
1017 | { | |
1018 | read_info_t next = (*ptr)->next; | |
ac6929b5 RB |
1019 | read_info_type_pool.remove (*ptr); |
1020 | *ptr = next; | |
6fb5fa3c | 1021 | } |
d26c7090 ER |
1022 | } |
1023 | ||
1024 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1025 | ||
1026 | static void | |
1027 | add_wild_read (bb_info_t bb_info) | |
1028 | { | |
1029 | insn_info_t insn_info = bb_info->last_insn; | |
6fb5fa3c | 1030 | insn_info->wild_read = true; |
d26c7090 ER |
1031 | free_read_records (bb_info); |
1032 | reset_active_stores (); | |
6fb5fa3c DB |
1033 | } |
1034 | ||
d26c7090 ER |
1035 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1036 | non-frame locations. */ | |
1037 | ||
1038 | static void | |
1039 | add_non_frame_wild_read (bb_info_t bb_info) | |
1040 | { | |
1041 | insn_info_t insn_info = bb_info->last_insn; | |
1042 | insn_info->non_frame_wild_read = true; | |
1043 | free_read_records (bb_info); | |
1044 | reset_active_stores (); | |
1045 | } | |
6fb5fa3c | 1046 | |
50f0f366 EB |
1047 | /* Return true if X is a constant or one of the registers that behave |
1048 | as a constant over the life of a function. This is equivalent to | |
1049 | !rtx_varies_p for memory addresses. */ | |
6fb5fa3c DB |
1050 | |
1051 | static bool | |
1052 | const_or_frame_p (rtx x) | |
1053 | { | |
d8116890 KZ |
1054 | if (CONSTANT_P (x)) |
1055 | return true; | |
1056 | ||
1057 | if (GET_CODE (x) == REG) | |
6fb5fa3c | 1058 | { |
6fb5fa3c DB |
1059 | /* Note that we have to test for the actual rtx used for the frame |
1060 | and arg pointers and not just the register number in case we have | |
1061 | eliminated the frame and/or arg pointer and are using it | |
1062 | for pseudos. */ | |
1063 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1064 | /* The arg pointer varies if it is not a fixed register. */ | |
1065 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1066 | || x == pic_offset_table_rtx) | |
1067 | return true; | |
1068 | return false; | |
6fb5fa3c | 1069 | } |
da6603c6 | 1070 | |
d8116890 | 1071 | return false; |
6fb5fa3c DB |
1072 | } |
1073 | ||
b8698a0f L |
1074 | /* Take all reasonable action to put the address of MEM into the form |
1075 | that we can do analysis on. | |
6fb5fa3c DB |
1076 | |
1077 | The gold standard is to get the address into the form: address + | |
1078 | OFFSET where address is something that rtx_varies_p considers a | |
1079 | constant. When we can get the address in this form, we can do | |
1080 | global analysis on it. Note that for constant bases, address is | |
1081 | not actually returned, only the group_id. The address can be | |
1082 | obtained from that. | |
1083 | ||
1084 | If that fails, we try cselib to get a value we can at least use | |
b8698a0f L |
1085 | locally. If that fails we return false. |
1086 | ||
6fb5fa3c DB |
1087 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1088 | group for non_varying bases. | |
1089 | ||
1090 | FOR_READ is true if this is a mem read and false if not. */ | |
1091 | ||
1092 | static bool | |
1093 | canon_address (rtx mem, | |
6fb5fa3c | 1094 | int *group_id, |
02ce5d90 | 1095 | poly_int64 *offset, |
6fb5fa3c DB |
1096 | cselib_val **base) |
1097 | { | |
ef4bddc2 | 1098 | machine_mode address_mode = get_address_mode (mem); |
6fb5fa3c DB |
1099 | rtx mem_address = XEXP (mem, 0); |
1100 | rtx expanded_address, address; | |
403c7520 JJ |
1101 | int expanded; |
1102 | ||
4deef538 | 1103 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
6fb5fa3c | 1104 | |
456610d3 | 1105 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1106 | { |
1107 | fprintf (dump_file, " mem: "); | |
1108 | print_inline_rtx (dump_file, mem_address, 0); | |
1109 | fprintf (dump_file, "\n"); | |
1110 | } | |
1111 | ||
403c7520 JJ |
1112 | /* First see if just canon_rtx (mem_address) is const or frame, |
1113 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1114 | address = NULL_RTX; | |
1115 | for (expanded = 0; expanded < 2; expanded++) | |
1116 | { | |
1117 | if (expanded) | |
1118 | { | |
1119 | /* Use cselib to replace all of the reg references with the full | |
b8698a0f | 1120 | expression. This will take care of the case where we have |
6fb5fa3c | 1121 | |
403c7520 JJ |
1122 | r_x = base + offset; |
1123 | val = *r_x; | |
b8698a0f L |
1124 | |
1125 | by making it into | |
6fb5fa3c | 1126 | |
403c7520 | 1127 | val = *(base + offset); */ |
6fb5fa3c | 1128 | |
403c7520 JJ |
1129 | expanded_address = cselib_expand_value_rtx (mem_address, |
1130 | scratch, 5); | |
6fb5fa3c | 1131 | |
403c7520 JJ |
1132 | /* If this fails, just go with the address from first |
1133 | iteration. */ | |
1134 | if (!expanded_address) | |
1135 | break; | |
1136 | } | |
1137 | else | |
1138 | expanded_address = mem_address; | |
6fb5fa3c | 1139 | |
403c7520 JJ |
1140 | /* Split the address into canonical BASE + OFFSET terms. */ |
1141 | address = canon_rtx (expanded_address); | |
6fb5fa3c | 1142 | |
403c7520 | 1143 | *offset = 0; |
6fb5fa3c | 1144 | |
456610d3 | 1145 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1146 | { |
1147 | if (expanded) | |
1148 | { | |
1149 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1150 | print_inline_rtx (dump_file, expanded_address, 0); | |
1151 | fprintf (dump_file, "\n"); | |
1152 | } | |
6fb5fa3c | 1153 | |
403c7520 JJ |
1154 | fprintf (dump_file, "\n after canon_rtx address: "); |
1155 | print_inline_rtx (dump_file, address, 0); | |
1156 | fprintf (dump_file, "\n"); | |
1157 | } | |
6fb5fa3c | 1158 | |
403c7520 JJ |
1159 | if (GET_CODE (address) == CONST) |
1160 | address = XEXP (address, 0); | |
6fb5fa3c | 1161 | |
02ce5d90 | 1162 | address = strip_offset_and_add (address, offset); |
6fb5fa3c | 1163 | |
09e881c9 BE |
1164 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1165 | && const_or_frame_p (address)) | |
6fb5fa3c | 1166 | { |
fd6daec9 | 1167 | group_info *group = get_group_info (address); |
403c7520 | 1168 | |
456610d3 | 1169 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
1170 | { |
1171 | fprintf (dump_file, " gid=%d offset=", group->id); | |
1172 | print_dec (*offset, dump_file); | |
1173 | fprintf (dump_file, "\n"); | |
1174 | } | |
403c7520 JJ |
1175 | *base = NULL; |
1176 | *group_id = group->id; | |
1177 | return true; | |
6fb5fa3c | 1178 | } |
403c7520 JJ |
1179 | } |
1180 | ||
4deef538 | 1181 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
403c7520 JJ |
1182 | *group_id = -1; |
1183 | ||
1184 | if (*base == NULL) | |
1185 | { | |
456610d3 | 1186 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1187 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1188 | return false; | |
6fb5fa3c | 1189 | } |
456610d3 | 1190 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
1191 | { |
1192 | fprintf (dump_file, " varying cselib base=%u:%u offset = ", | |
1193 | (*base)->uid, (*base)->hash); | |
1194 | print_dec (*offset, dump_file); | |
1195 | fprintf (dump_file, "\n"); | |
1196 | } | |
6fb5fa3c DB |
1197 | return true; |
1198 | } | |
1199 | ||
1200 | ||
1201 | /* Clear the rhs field from the active_local_stores array. */ | |
1202 | ||
1203 | static void | |
1204 | clear_rhs_from_active_local_stores (void) | |
1205 | { | |
1206 | insn_info_t ptr = active_local_stores; | |
1207 | ||
1208 | while (ptr) | |
1209 | { | |
fd6daec9 | 1210 | store_info *store_info = ptr->store_rec; |
6fb5fa3c DB |
1211 | /* Skip the clobbers. */ |
1212 | while (!store_info->is_set) | |
1213 | store_info = store_info->next; | |
1214 | ||
1215 | store_info->rhs = NULL; | |
8dd5516b | 1216 | store_info->const_rhs = NULL; |
6fb5fa3c DB |
1217 | |
1218 | ptr = ptr->next_local_store; | |
1219 | } | |
1220 | } | |
1221 | ||
1222 | ||
8dd5516b JJ |
1223 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1224 | ||
1225 | static inline void | |
fd6daec9 | 1226 | set_position_unneeded (store_info *s_info, int pos) |
8dd5516b JJ |
1227 | { |
1228 | if (__builtin_expect (s_info->is_large, false)) | |
1229 | { | |
fcaa4ca4 NF |
1230 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1231 | s_info->positions_needed.large.count++; | |
8dd5516b JJ |
1232 | } |
1233 | else | |
1234 | s_info->positions_needed.small_bitmask | |
fecfbfa4 | 1235 | &= ~(HOST_WIDE_INT_1U << pos); |
8dd5516b JJ |
1236 | } |
1237 | ||
1238 | /* Mark the whole store S_INFO as unneeded. */ | |
1239 | ||
1240 | static inline void | |
fd6daec9 | 1241 | set_all_positions_unneeded (store_info *s_info) |
8dd5516b JJ |
1242 | { |
1243 | if (__builtin_expect (s_info->is_large, false)) | |
1244 | { | |
02ce5d90 RS |
1245 | HOST_WIDE_INT width; |
1246 | if (s_info->width.is_constant (&width)) | |
1247 | { | |
1248 | bitmap_set_range (s_info->positions_needed.large.bmap, 0, width); | |
1249 | s_info->positions_needed.large.count = width; | |
1250 | } | |
1251 | else | |
1252 | { | |
1253 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1254 | s_info->positions_needed.large.count = 1; | |
1255 | } | |
8dd5516b JJ |
1256 | } |
1257 | else | |
07e96250 | 1258 | s_info->positions_needed.small_bitmask = HOST_WIDE_INT_0U; |
8dd5516b JJ |
1259 | } |
1260 | ||
1261 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1262 | ||
1263 | static inline bool | |
fd6daec9 | 1264 | any_positions_needed_p (store_info *s_info) |
8dd5516b JJ |
1265 | { |
1266 | if (__builtin_expect (s_info->is_large, false)) | |
02ce5d90 RS |
1267 | { |
1268 | HOST_WIDE_INT width; | |
1269 | if (s_info->width.is_constant (&width)) | |
1270 | { | |
1271 | gcc_checking_assert (s_info->positions_needed.large.bmap); | |
1272 | return s_info->positions_needed.large.count < width; | |
1273 | } | |
1274 | else | |
1275 | { | |
1276 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1277 | return s_info->positions_needed.large.count == 0; | |
1278 | } | |
1279 | } | |
8dd5516b | 1280 | else |
07e96250 | 1281 | return (s_info->positions_needed.small_bitmask != HOST_WIDE_INT_0U); |
8dd5516b JJ |
1282 | } |
1283 | ||
1284 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
02ce5d90 | 1285 | store are known to be needed. */ |
8dd5516b JJ |
1286 | |
1287 | static inline bool | |
02ce5d90 RS |
1288 | all_positions_needed_p (store_info *s_info, poly_int64 start, |
1289 | poly_int64 width) | |
8dd5516b | 1290 | { |
02ce5d90 RS |
1291 | gcc_assert (s_info->rhs); |
1292 | if (!s_info->width.is_constant ()) | |
1293 | { | |
1294 | gcc_assert (s_info->is_large | |
1295 | && !s_info->positions_needed.large.bmap); | |
1296 | return s_info->positions_needed.large.count == 0; | |
1297 | } | |
1298 | ||
1299 | /* Otherwise, if START and WIDTH are non-constant, we're asking about | |
1300 | a non-constant region of a constant-sized store. We can't say for | |
1301 | sure that all positions are needed. */ | |
1302 | HOST_WIDE_INT const_start, const_width; | |
1303 | if (!start.is_constant (&const_start) | |
1304 | || !width.is_constant (&const_width)) | |
1305 | return false; | |
1306 | ||
8dd5516b JJ |
1307 | if (__builtin_expect (s_info->is_large, false)) |
1308 | { | |
02ce5d90 RS |
1309 | for (HOST_WIDE_INT i = const_start; i < const_start + const_width; ++i) |
1310 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, i)) | |
8dd5516b JJ |
1311 | return false; |
1312 | return true; | |
1313 | } | |
1314 | else | |
1315 | { | |
02ce5d90 RS |
1316 | unsigned HOST_WIDE_INT mask |
1317 | = lowpart_bitmask (const_width) << const_start; | |
8dd5516b JJ |
1318 | return (s_info->positions_needed.small_bitmask & mask) == mask; |
1319 | } | |
1320 | } | |
1321 | ||
1322 | ||
02ce5d90 RS |
1323 | static rtx get_stored_val (store_info *, machine_mode, poly_int64, |
1324 | poly_int64, basic_block, bool); | |
8dd5516b JJ |
1325 | |
1326 | ||
6fb5fa3c DB |
1327 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1328 | there is a candidate store, after adding it to the appropriate | |
1329 | local store group if so. */ | |
1330 | ||
1331 | static int | |
1332 | record_store (rtx body, bb_info_t bb_info) | |
1333 | { | |
6216f94e | 1334 | rtx mem, rhs, const_rhs, mem_addr; |
02ce5d90 RS |
1335 | poly_int64 offset = 0; |
1336 | poly_int64 width = 0; | |
6fb5fa3c | 1337 | insn_info_t insn_info = bb_info->last_insn; |
fd6daec9 | 1338 | store_info *store_info = NULL; |
6fb5fa3c DB |
1339 | int group_id; |
1340 | cselib_val *base = NULL; | |
8dd5516b | 1341 | insn_info_t ptr, last, redundant_reason; |
6fb5fa3c DB |
1342 | bool store_is_unused; |
1343 | ||
1344 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1345 | return 0; | |
1346 | ||
8dd5516b JJ |
1347 | mem = SET_DEST (body); |
1348 | ||
6fb5fa3c DB |
1349 | /* If this is not used, then this cannot be used to keep the insn |
1350 | from being deleted. On the other hand, it does provide something | |
1351 | that can be used to prove that another store is dead. */ | |
1352 | store_is_unused | |
8dd5516b | 1353 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
6fb5fa3c DB |
1354 | |
1355 | /* Check whether that value is a suitable memory location. */ | |
6fb5fa3c DB |
1356 | if (!MEM_P (mem)) |
1357 | { | |
1358 | /* If the set or clobber is unused, then it does not effect our | |
1359 | ability to get rid of the entire insn. */ | |
1360 | if (!store_is_unused) | |
1361 | insn_info->cannot_delete = true; | |
1362 | return 0; | |
1363 | } | |
1364 | ||
1365 | /* At this point we know mem is a mem. */ | |
1366 | if (GET_MODE (mem) == BLKmode) | |
1367 | { | |
d05d7551 | 1368 | HOST_WIDE_INT const_size; |
6fb5fa3c DB |
1369 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) |
1370 | { | |
456610d3 | 1371 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1372 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1373 | add_wild_read (bb_info); | |
1374 | insn_info->cannot_delete = true; | |
8dd5516b | 1375 | return 0; |
6fb5fa3c | 1376 | } |
8dd5516b JJ |
1377 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1378 | as memset (addr, 0, 36); */ | |
f5541398 | 1379 | else if (!MEM_SIZE_KNOWN_P (mem) |
d05d7551 RS |
1380 | || maybe_le (MEM_SIZE (mem), 0) |
1381 | /* This is a limit on the bitmap size, which is only relevant | |
1382 | for constant-sized MEMs. */ | |
1383 | || (MEM_SIZE (mem).is_constant (&const_size) | |
1384 | && const_size > MAX_OFFSET) | |
8dd5516b | 1385 | || GET_CODE (body) != SET |
8dd5516b | 1386 | || !CONST_INT_P (SET_SRC (body))) |
6fb5fa3c | 1387 | { |
8dd5516b JJ |
1388 | if (!store_is_unused) |
1389 | { | |
1390 | /* If the set or clobber is unused, then it does not effect our | |
1391 | ability to get rid of the entire insn. */ | |
1392 | insn_info->cannot_delete = true; | |
1393 | clear_rhs_from_active_local_stores (); | |
1394 | } | |
1395 | return 0; | |
6fb5fa3c | 1396 | } |
6fb5fa3c DB |
1397 | } |
1398 | ||
1399 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1400 | if (MEM_VOLATILE_P (mem)) | |
8dd5516b | 1401 | insn_info->cannot_delete = true; |
6fb5fa3c | 1402 | |
ac6929b5 | 1403 | if (!canon_address (mem, &group_id, &offset, &base)) |
6fb5fa3c DB |
1404 | { |
1405 | clear_rhs_from_active_local_stores (); | |
1406 | return 0; | |
1407 | } | |
1408 | ||
8dd5516b | 1409 | if (GET_MODE (mem) == BLKmode) |
f5541398 | 1410 | width = MEM_SIZE (mem); |
8dd5516b | 1411 | else |
6f557e0e | 1412 | width = GET_MODE_SIZE (GET_MODE (mem)); |
6fb5fa3c | 1413 | |
02ce5d90 | 1414 | if (!endpoint_representable_p (offset, width)) |
213ffde2 JJ |
1415 | { |
1416 | clear_rhs_from_active_local_stores (); | |
1417 | return 0; | |
1418 | } | |
1419 | ||
ac6929b5 | 1420 | if (group_id >= 0) |
6fb5fa3c DB |
1421 | { |
1422 | /* In the restrictive case where the base is a constant or the | |
1423 | frame pointer we can do global analysis. */ | |
b8698a0f | 1424 | |
fd6daec9 | 1425 | group_info *group |
9771b263 | 1426 | = rtx_group_vec[group_id]; |
d26c7090 | 1427 | tree expr = MEM_EXPR (mem); |
b8698a0f | 1428 | |
da6603c6 | 1429 | store_info = rtx_store_info_pool.allocate (); |
d26c7090 | 1430 | set_usage_bits (group, offset, width, expr); |
6fb5fa3c | 1431 | |
456610d3 | 1432 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1433 | { |
1434 | fprintf (dump_file, " processing const base store gid=%d", | |
1435 | group_id); | |
1436 | print_range (dump_file, offset, width); | |
1437 | fprintf (dump_file, "\n"); | |
1438 | } | |
6fb5fa3c DB |
1439 | } |
1440 | else | |
1441 | { | |
9e412ca3 | 1442 | if (may_be_sp_based_p (XEXP (mem, 0))) |
50f0f366 | 1443 | insn_info->stack_pointer_based = true; |
6fb5fa3c | 1444 | insn_info->contains_cselib_groups = true; |
50f0f366 | 1445 | |
da6603c6 | 1446 | store_info = cse_store_info_pool.allocate (); |
6fb5fa3c DB |
1447 | group_id = -1; |
1448 | ||
456610d3 | 1449 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1450 | { |
1451 | fprintf (dump_file, " processing cselib store "); | |
1452 | print_range (dump_file, offset, width); | |
1453 | fprintf (dump_file, "\n"); | |
1454 | } | |
6fb5fa3c DB |
1455 | } |
1456 | ||
8dd5516b JJ |
1457 | const_rhs = rhs = NULL_RTX; |
1458 | if (GET_CODE (body) == SET | |
1459 | /* No place to keep the value after ra. */ | |
1460 | && !reload_completed | |
1461 | && (REG_P (SET_SRC (body)) | |
1462 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1463 | || CONSTANT_P (SET_SRC (body))) | |
1464 | && !MEM_VOLATILE_P (mem) | |
1465 | /* Sometimes the store and reload is used for truncation and | |
1466 | rounding. */ | |
1467 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1468 | { | |
1469 | rhs = SET_SRC (body); | |
1470 | if (CONSTANT_P (rhs)) | |
1471 | const_rhs = rhs; | |
1472 | else if (body == PATTERN (insn_info->insn)) | |
1473 | { | |
1474 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1475 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1476 | const_rhs = XEXP (tem, 0); | |
1477 | } | |
1478 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1479 | { | |
1480 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1481 | ||
1482 | if (tem && CONSTANT_P (tem)) | |
1483 | const_rhs = tem; | |
1484 | } | |
1485 | } | |
1486 | ||
6fb5fa3c DB |
1487 | /* Check to see if this stores causes some other stores to be |
1488 | dead. */ | |
1489 | ptr = active_local_stores; | |
1490 | last = NULL; | |
8dd5516b | 1491 | redundant_reason = NULL; |
6216f94e | 1492 | mem = canon_rtx (mem); |
ac6929b5 RB |
1493 | |
1494 | if (group_id < 0) | |
1495 | mem_addr = base->val_rtx; | |
6216f94e JJ |
1496 | else |
1497 | { | |
ac6929b5 RB |
1498 | group_info *group = rtx_group_vec[group_id]; |
1499 | mem_addr = group->canon_base_addr; | |
6216f94e | 1500 | } |
02ce5d90 | 1501 | if (maybe_ne (offset, 0)) |
ac6929b5 | 1502 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6fb5fa3c DB |
1503 | |
1504 | while (ptr) | |
1505 | { | |
1506 | insn_info_t next = ptr->next_local_store; | |
fd6daec9 | 1507 | struct store_info *s_info = ptr->store_rec; |
60564289 | 1508 | bool del = true; |
6fb5fa3c DB |
1509 | |
1510 | /* Skip the clobbers. We delete the active insn if this insn | |
6ed3da00 | 1511 | shadows the set. To have been put on the active list, it |
6fb5fa3c DB |
1512 | has exactly on set. */ |
1513 | while (!s_info->is_set) | |
1514 | s_info = s_info->next; | |
1515 | ||
ac6929b5 | 1516 | if (s_info->group_id == group_id && s_info->cse_base == base) |
6fb5fa3c DB |
1517 | { |
1518 | HOST_WIDE_INT i; | |
456610d3 | 1519 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1520 | { |
1521 | fprintf (dump_file, " trying store in insn=%d gid=%d", | |
1522 | INSN_UID (ptr->insn), s_info->group_id); | |
1523 | print_range (dump_file, s_info->offset, s_info->width); | |
1524 | fprintf (dump_file, "\n"); | |
1525 | } | |
8dd5516b JJ |
1526 | |
1527 | /* Even if PTR won't be eliminated as unneeded, if both | |
1528 | PTR and this insn store the same constant value, we might | |
1529 | eliminate this insn instead. */ | |
1530 | if (s_info->const_rhs | |
1531 | && const_rhs | |
71d6a386 RS |
1532 | && known_subrange_p (offset, width, |
1533 | s_info->offset, s_info->width) | |
1534 | && all_positions_needed_p (s_info, offset - s_info->offset, | |
d7e1f499 RB |
1535 | width) |
1536 | /* We can only remove the later store if the earlier aliases | |
1537 | at least all accesses the later one. */ | |
1538 | && (MEM_ALIAS_SET (mem) == MEM_ALIAS_SET (s_info->mem) | |
1539 | || alias_set_subset_of (MEM_ALIAS_SET (mem), | |
1540 | MEM_ALIAS_SET (s_info->mem)))) | |
8dd5516b JJ |
1541 | { |
1542 | if (GET_MODE (mem) == BLKmode) | |
1543 | { | |
1544 | if (GET_MODE (s_info->mem) == BLKmode | |
1545 | && s_info->const_rhs == const_rhs) | |
1546 | redundant_reason = ptr; | |
1547 | } | |
1548 | else if (s_info->const_rhs == const0_rtx | |
1549 | && const_rhs == const0_rtx) | |
1550 | redundant_reason = ptr; | |
1551 | else | |
1552 | { | |
1553 | rtx val; | |
1554 | start_sequence (); | |
71d6a386 | 1555 | val = get_stored_val (s_info, GET_MODE (mem), offset, width, |
8dd5516b JJ |
1556 | BLOCK_FOR_INSN (insn_info->insn), |
1557 | true); | |
1558 | if (get_insns () != NULL) | |
1559 | val = NULL_RTX; | |
1560 | end_sequence (); | |
1561 | if (val && rtx_equal_p (val, const_rhs)) | |
1562 | redundant_reason = ptr; | |
1563 | } | |
1564 | } | |
1565 | ||
02ce5d90 | 1566 | HOST_WIDE_INT begin_unneeded, const_s_width, const_width; |
71d6a386 RS |
1567 | if (known_subrange_p (s_info->offset, s_info->width, offset, width)) |
1568 | /* The new store touches every byte that S_INFO does. */ | |
1569 | set_all_positions_unneeded (s_info); | |
02ce5d90 RS |
1570 | else if ((offset - s_info->offset).is_constant (&begin_unneeded) |
1571 | && s_info->width.is_constant (&const_s_width) | |
1572 | && width.is_constant (&const_width)) | |
71d6a386 | 1573 | { |
02ce5d90 | 1574 | HOST_WIDE_INT end_unneeded = begin_unneeded + const_width; |
71d6a386 | 1575 | begin_unneeded = MAX (begin_unneeded, 0); |
02ce5d90 | 1576 | end_unneeded = MIN (end_unneeded, const_s_width); |
71d6a386 RS |
1577 | for (i = begin_unneeded; i < end_unneeded; ++i) |
1578 | set_position_unneeded (s_info, i); | |
1579 | } | |
02ce5d90 RS |
1580 | else |
1581 | { | |
1582 | /* We don't know which parts of S_INFO are needed and | |
1583 | which aren't, so invalidate the RHS. */ | |
1584 | s_info->rhs = NULL; | |
1585 | s_info->const_rhs = NULL; | |
1586 | } | |
6fb5fa3c DB |
1587 | } |
1588 | else if (s_info->rhs) | |
1589 | /* Need to see if it is possible for this store to overwrite | |
1590 | the value of store_info. If it is, set the rhs to NULL to | |
1591 | keep it from being used to remove a load. */ | |
1592 | { | |
43b9f499 RB |
1593 | if (canon_output_dependence (s_info->mem, true, |
1594 | mem, GET_MODE (mem), | |
1595 | mem_addr)) | |
8dd5516b JJ |
1596 | { |
1597 | s_info->rhs = NULL; | |
1598 | s_info->const_rhs = NULL; | |
1599 | } | |
6fb5fa3c | 1600 | } |
6216f94e | 1601 | |
6fb5fa3c DB |
1602 | /* An insn can be deleted if every position of every one of |
1603 | its s_infos is zero. */ | |
1b6fa860 | 1604 | if (any_positions_needed_p (s_info)) |
60564289 | 1605 | del = false; |
8dd5516b | 1606 | |
60564289 | 1607 | if (del) |
6fb5fa3c DB |
1608 | { |
1609 | insn_info_t insn_to_delete = ptr; | |
b8698a0f | 1610 | |
dabd47e7 | 1611 | active_local_stores_len--; |
6fb5fa3c DB |
1612 | if (last) |
1613 | last->next_local_store = ptr->next_local_store; | |
1614 | else | |
1615 | active_local_stores = ptr->next_local_store; | |
b8698a0f | 1616 | |
1b6fa860 JJ |
1617 | if (!insn_to_delete->cannot_delete) |
1618 | delete_dead_store_insn (insn_to_delete); | |
6fb5fa3c DB |
1619 | } |
1620 | else | |
1621 | last = ptr; | |
b8698a0f | 1622 | |
6fb5fa3c DB |
1623 | ptr = next; |
1624 | } | |
b8698a0f | 1625 | |
6fb5fa3c DB |
1626 | /* Finish filling in the store_info. */ |
1627 | store_info->next = insn_info->store_rec; | |
1628 | insn_info->store_rec = store_info; | |
6216f94e | 1629 | store_info->mem = mem; |
6216f94e | 1630 | store_info->mem_addr = mem_addr; |
6fb5fa3c | 1631 | store_info->cse_base = base; |
02ce5d90 RS |
1632 | HOST_WIDE_INT const_width; |
1633 | if (!width.is_constant (&const_width)) | |
1634 | { | |
1635 | store_info->is_large = true; | |
1636 | store_info->positions_needed.large.count = 0; | |
1637 | store_info->positions_needed.large.bmap = NULL; | |
1638 | } | |
1639 | else if (const_width > HOST_BITS_PER_WIDE_INT) | |
8dd5516b JJ |
1640 | { |
1641 | store_info->is_large = true; | |
1642 | store_info->positions_needed.large.count = 0; | |
3f9b14ff | 1643 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
8dd5516b JJ |
1644 | } |
1645 | else | |
1646 | { | |
1647 | store_info->is_large = false; | |
02ce5d90 RS |
1648 | store_info->positions_needed.small_bitmask |
1649 | = lowpart_bitmask (const_width); | |
8dd5516b | 1650 | } |
6fb5fa3c | 1651 | store_info->group_id = group_id; |
71d6a386 RS |
1652 | store_info->offset = offset; |
1653 | store_info->width = width; | |
6fb5fa3c | 1654 | store_info->is_set = GET_CODE (body) == SET; |
8dd5516b JJ |
1655 | store_info->rhs = rhs; |
1656 | store_info->const_rhs = const_rhs; | |
1657 | store_info->redundant_reason = redundant_reason; | |
6fb5fa3c | 1658 | |
6fb5fa3c DB |
1659 | /* If this is a clobber, we return 0. We will only be able to |
1660 | delete this insn if there is only one store USED store, but we | |
1661 | can use the clobber to delete other stores earlier. */ | |
1662 | return store_info->is_set ? 1 : 0; | |
1663 | } | |
1664 | ||
1665 | ||
1666 | static void | |
1667 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1668 | { | |
b8698a0f | 1669 | fprintf (dump_file, "%s insn=%d %s\n", start, |
6fb5fa3c DB |
1670 | INSN_UID (insn_info->insn), |
1671 | insn_info->store_rec ? "has store" : "naked"); | |
1672 | } | |
1673 | ||
1674 | ||
8660aaae EC |
1675 | /* If the modes are different and the value's source and target do not |
1676 | line up, we need to extract the value from lower part of the rhs of | |
1677 | the store, shift it, and then put it into a form that can be shoved | |
1678 | into the read_insn. This function generates a right SHIFT of a | |
1679 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1680 | shift sequence is returned or NULL if we failed to find a | |
1681 | shift. */ | |
1682 | ||
1683 | static rtx | |
02ce5d90 | 1684 | find_shift_sequence (poly_int64 access_size, |
fd6daec9 | 1685 | store_info *store_info, |
ef4bddc2 | 1686 | machine_mode read_mode, |
02ce5d90 | 1687 | poly_int64 shift, bool speed, bool require_cst) |
8660aaae | 1688 | { |
ef4bddc2 | 1689 | machine_mode store_mode = GET_MODE (store_info->mem); |
f67f4dff | 1690 | scalar_int_mode new_mode; |
18b526e8 | 1691 | rtx read_reg = NULL; |
8660aaae EC |
1692 | |
1693 | /* Some machines like the x86 have shift insns for each size of | |
1694 | operand. Other machines like the ppc or the ia-64 may only have | |
1695 | shift insns that shift values within 32 or 64 bit registers. | |
1696 | This loop tries to find the smallest shift insn that will right | |
1697 | justify the value we want to read but is available in one insn on | |
1698 | the machine. */ | |
1699 | ||
f67f4dff RS |
1700 | opt_scalar_int_mode new_mode_iter; |
1701 | FOR_EACH_MODE_FROM (new_mode_iter, | |
1702 | smallest_int_mode_for_size (access_size * BITS_PER_UNIT)) | |
8660aaae | 1703 | { |
dc01c3d1 DM |
1704 | rtx target, new_reg, new_lhs; |
1705 | rtx_insn *shift_seq, *insn; | |
d898d29b | 1706 | int cost; |
348eea5f | 1707 | |
f67f4dff | 1708 | new_mode = new_mode_iter.require (); |
c94843d2 RS |
1709 | if (GET_MODE_BITSIZE (new_mode) > BITS_PER_WORD) |
1710 | break; | |
1711 | ||
72a2609f JJ |
1712 | /* If a constant was stored into memory, try to simplify it here, |
1713 | otherwise the cost of the shift might preclude this optimization | |
1714 | e.g. at -Os, even when no actual shift will be needed. */ | |
8dd5516b | 1715 | if (store_info->const_rhs) |
72a2609f | 1716 | { |
91914e56 | 1717 | poly_uint64 byte = subreg_lowpart_offset (new_mode, store_mode); |
8dd5516b JJ |
1718 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1719 | store_mode, byte); | |
72a2609f JJ |
1720 | if (ret && CONSTANT_P (ret)) |
1721 | { | |
abd3c800 | 1722 | rtx shift_rtx = gen_int_shift_amount (new_mode, shift); |
72a2609f | 1723 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, |
abd3c800 | 1724 | ret, shift_rtx); |
72a2609f JJ |
1725 | if (ret && CONSTANT_P (ret)) |
1726 | { | |
1727 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1728 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1729 | if (ret && CONSTANT_P (ret) | |
e548c9df AM |
1730 | && (set_src_cost (ret, read_mode, speed) |
1731 | <= COSTS_N_INSNS (1))) | |
72a2609f JJ |
1732 | return ret; |
1733 | } | |
1734 | } | |
1735 | } | |
1736 | ||
8dd5516b JJ |
1737 | if (require_cst) |
1738 | return NULL_RTX; | |
1739 | ||
18b526e8 RS |
1740 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1741 | requires a real instruction. */ | |
73a699ae | 1742 | if (maybe_lt (GET_MODE_SIZE (new_mode), GET_MODE_SIZE (store_mode)) |
d0edd768 | 1743 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
348eea5f RS |
1744 | continue; |
1745 | ||
18b526e8 RS |
1746 | /* Also try a wider mode if the necessary punning is either not |
1747 | desirable or not possible. */ | |
1748 | if (!CONSTANT_P (store_info->rhs) | |
99e1629f | 1749 | && !targetm.modes_tieable_p (new_mode, store_mode)) |
18b526e8 | 1750 | continue; |
18b526e8 | 1751 | |
348eea5f | 1752 | new_reg = gen_reg_rtx (new_mode); |
8660aaae EC |
1753 | |
1754 | start_sequence (); | |
1755 | ||
1756 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1757 | of one dsp where the cost of these two was not the same. But | |
1758 | this really is a rare case anyway. */ | |
1759 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
abd3c800 RS |
1760 | gen_int_shift_amount (new_mode, shift), |
1761 | new_reg, 1, OPTAB_DIRECT); | |
8660aaae | 1762 | |
c6f3019a RS |
1763 | shift_seq = get_insns (); |
1764 | end_sequence (); | |
8660aaae | 1765 | |
c6f3019a RS |
1766 | if (target != new_reg || shift_seq == NULL) |
1767 | continue; | |
1768 | ||
1769 | cost = 0; | |
1770 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1771 | if (INSN_P (insn)) | |
d09e78e8 | 1772 | cost += insn_cost (insn, speed); |
c6f3019a RS |
1773 | |
1774 | /* The computation up to here is essentially independent | |
1775 | of the arguments and could be precomputed. It may | |
1776 | not be worth doing so. We could precompute if | |
1777 | worthwhile or at least cache the results. The result | |
06acf7d0 RS |
1778 | technically depends on both SHIFT and ACCESS_SIZE, |
1779 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
c6f3019a RS |
1780 | |
1781 | if (cost > COSTS_N_INSNS (1)) | |
1782 | continue; | |
1783 | ||
d898d29b JJ |
1784 | new_lhs = extract_low_bits (new_mode, store_mode, |
1785 | copy_rtx (store_info->rhs)); | |
1786 | if (new_lhs == NULL_RTX) | |
1787 | continue; | |
1788 | ||
c6f3019a RS |
1789 | /* We found an acceptable shift. Generate a move to |
1790 | take the value from the store and put it into the | |
1791 | shift pseudo, then shift it, then generate another | |
1792 | move to put in into the target of the read. */ | |
18b526e8 | 1793 | emit_move_insn (new_reg, new_lhs); |
c6f3019a | 1794 | emit_insn (shift_seq); |
18b526e8 | 1795 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
c6f3019a | 1796 | break; |
8660aaae EC |
1797 | } |
1798 | ||
18b526e8 | 1799 | return read_reg; |
8660aaae EC |
1800 | } |
1801 | ||
1802 | ||
02b47899 KZ |
1803 | /* Call back for note_stores to find the hard regs set or clobbered by |
1804 | insn. Data is a bitmap of the hardregs set so far. */ | |
1805 | ||
1806 | static void | |
1807 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1808 | { | |
1809 | bitmap regs_set = (bitmap) data; | |
1810 | ||
1811 | if (REG_P (x) | |
f773c2bd | 1812 | && HARD_REGISTER_P (x)) |
dc8afb70 | 1813 | bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); |
02b47899 KZ |
1814 | } |
1815 | ||
8dd5516b | 1816 | /* Helper function for replace_read and record_store. |
71d6a386 RS |
1817 | Attempt to return a value of mode READ_MODE stored in STORE_INFO, |
1818 | consisting of READ_WIDTH bytes starting from READ_OFFSET. Return NULL | |
8dd5516b JJ |
1819 | if not successful. If REQUIRE_CST is true, return always constant. */ |
1820 | ||
1821 | static rtx | |
fd6daec9 | 1822 | get_stored_val (store_info *store_info, machine_mode read_mode, |
02ce5d90 | 1823 | poly_int64 read_offset, poly_int64 read_width, |
8dd5516b JJ |
1824 | basic_block bb, bool require_cst) |
1825 | { | |
ef4bddc2 | 1826 | machine_mode store_mode = GET_MODE (store_info->mem); |
02ce5d90 | 1827 | poly_int64 gap; |
8dd5516b JJ |
1828 | rtx read_reg; |
1829 | ||
1830 | /* To get here the read is within the boundaries of the write so | |
1831 | shift will never be negative. Start out with the shift being in | |
1832 | bytes. */ | |
1833 | if (store_mode == BLKmode) | |
71d6a386 | 1834 | gap = 0; |
8dd5516b | 1835 | else if (BYTES_BIG_ENDIAN) |
71d6a386 RS |
1836 | gap = ((store_info->offset + store_info->width) |
1837 | - (read_offset + read_width)); | |
8dd5516b | 1838 | else |
71d6a386 | 1839 | gap = read_offset - store_info->offset; |
8dd5516b | 1840 | |
02ce5d90 | 1841 | if (maybe_ne (gap, 0)) |
71d6a386 | 1842 | { |
02ce5d90 RS |
1843 | poly_int64 shift = gap * BITS_PER_UNIT; |
1844 | poly_int64 access_size = GET_MODE_SIZE (read_mode) + gap; | |
71d6a386 RS |
1845 | read_reg = find_shift_sequence (access_size, store_info, read_mode, |
1846 | shift, optimize_bb_for_speed_p (bb), | |
1847 | require_cst); | |
1848 | } | |
8dd5516b JJ |
1849 | else if (store_mode == BLKmode) |
1850 | { | |
1851 | /* The store is a memset (addr, const_val, const_size). */ | |
1852 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
304b9962 RS |
1853 | scalar_int_mode int_store_mode; |
1854 | if (!int_mode_for_mode (read_mode).exists (&int_store_mode)) | |
8dd5516b JJ |
1855 | read_reg = NULL_RTX; |
1856 | else if (store_info->rhs == const0_rtx) | |
304b9962 RS |
1857 | read_reg = extract_low_bits (read_mode, int_store_mode, const0_rtx); |
1858 | else if (GET_MODE_BITSIZE (int_store_mode) > HOST_BITS_PER_WIDE_INT | |
8dd5516b JJ |
1859 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) |
1860 | read_reg = NULL_RTX; | |
1861 | else | |
1862 | { | |
1863 | unsigned HOST_WIDE_INT c | |
1864 | = INTVAL (store_info->rhs) | |
fecfbfa4 | 1865 | & ((HOST_WIDE_INT_1 << BITS_PER_UNIT) - 1); |
8dd5516b JJ |
1866 | int shift = BITS_PER_UNIT; |
1867 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1868 | { | |
1869 | c |= (c << shift); | |
1870 | shift <<= 1; | |
1871 | } | |
304b9962 RS |
1872 | read_reg = gen_int_mode (c, int_store_mode); |
1873 | read_reg = extract_low_bits (read_mode, int_store_mode, read_reg); | |
8dd5516b JJ |
1874 | } |
1875 | } | |
1876 | else if (store_info->const_rhs | |
1877 | && (require_cst | |
1878 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1879 | read_reg = extract_low_bits (read_mode, store_mode, | |
1880 | copy_rtx (store_info->const_rhs)); | |
1881 | else | |
1882 | read_reg = extract_low_bits (read_mode, store_mode, | |
1883 | copy_rtx (store_info->rhs)); | |
1884 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1885 | read_reg = NULL_RTX; | |
1886 | return read_reg; | |
1887 | } | |
02b47899 | 1888 | |
6fb5fa3c DB |
1889 | /* Take a sequence of: |
1890 | A <- r1 | |
1891 | ... | |
1892 | ... <- A | |
1893 | ||
b8698a0f | 1894 | and change it into |
6fb5fa3c DB |
1895 | r2 <- r1 |
1896 | A <- r1 | |
1897 | ... | |
1898 | ... <- r2 | |
1899 | ||
8660aaae EC |
1900 | or |
1901 | ||
1902 | r3 <- extract (r1) | |
1903 | r3 <- r3 >> shift | |
1904 | r2 <- extract (r3) | |
1905 | ... <- r2 | |
1906 | ||
1907 | or | |
1908 | ||
1909 | r2 <- extract (r1) | |
1910 | ... <- r2 | |
1911 | ||
1912 | Depending on the alignment and the mode of the store and | |
1913 | subsequent load. | |
1914 | ||
1915 | ||
1916 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
6fb5fa3c DB |
1917 | and READ_INSN are for the read. Return true if the replacement |
1918 | went ok. */ | |
1919 | ||
1920 | static bool | |
fd6daec9 | 1921 | replace_read (store_info *store_info, insn_info_t store_insn, |
8dd5516b JJ |
1922 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
1923 | bitmap regs_live) | |
6fb5fa3c | 1924 | { |
ef4bddc2 RS |
1925 | machine_mode store_mode = GET_MODE (store_info->mem); |
1926 | machine_mode read_mode = GET_MODE (read_info->mem); | |
dc01c3d1 DM |
1927 | rtx_insn *insns, *this_insn; |
1928 | rtx read_reg; | |
8dd5516b | 1929 | basic_block bb; |
8660aaae | 1930 | |
6fb5fa3c DB |
1931 | if (!dbg_cnt (dse)) |
1932 | return false; | |
1933 | ||
18b526e8 RS |
1934 | /* Create a sequence of instructions to set up the read register. |
1935 | This sequence goes immediately before the store and its result | |
1936 | is read by the load. | |
1937 | ||
1938 | We need to keep this in perspective. We are replacing a read | |
8660aaae EC |
1939 | with a sequence of insns, but the read will almost certainly be |
1940 | in cache, so it is not going to be an expensive one. Thus, we | |
1941 | are not willing to do a multi insn shift or worse a subroutine | |
1942 | call to get rid of the read. */ | |
456610d3 | 1943 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
1944 | fprintf (dump_file, "trying to replace %smode load in insn %d" |
1945 | " from %smode store in insn %d\n", | |
1946 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
1947 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
1948 | start_sequence (); | |
8dd5516b JJ |
1949 | bb = BLOCK_FOR_INSN (read_insn->insn); |
1950 | read_reg = get_stored_val (store_info, | |
71d6a386 | 1951 | read_mode, read_info->offset, read_info->width, |
8dd5516b | 1952 | bb, false); |
18b526e8 | 1953 | if (read_reg == NULL_RTX) |
8660aaae | 1954 | { |
18b526e8 | 1955 | end_sequence (); |
456610d3 | 1956 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
1957 | fprintf (dump_file, " -- could not extract bits of stored value\n"); |
1958 | return false; | |
8660aaae | 1959 | } |
18b526e8 RS |
1960 | /* Force the value into a new register so that it won't be clobbered |
1961 | between the store and the load. */ | |
1962 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
1963 | insns = get_insns (); | |
1964 | end_sequence (); | |
8660aaae | 1965 | |
02b47899 KZ |
1966 | if (insns != NULL_RTX) |
1967 | { | |
1968 | /* Now we have to scan the set of new instructions to see if the | |
1969 | sequence contains and sets of hardregs that happened to be | |
1970 | live at this point. For instance, this can happen if one of | |
1971 | the insns sets the CC and the CC happened to be live at that | |
1972 | point. This does occasionally happen, see PR 37922. */ | |
3f9b14ff | 1973 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
02b47899 KZ |
1974 | |
1975 | for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
1976 | note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
b8698a0f | 1977 | |
02b47899 KZ |
1978 | bitmap_and_into (regs_set, regs_live); |
1979 | if (!bitmap_empty_p (regs_set)) | |
1980 | { | |
456610d3 | 1981 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02b47899 | 1982 | { |
b8698a0f | 1983 | fprintf (dump_file, |
02b47899 KZ |
1984 | "abandoning replacement because sequence clobbers live hardregs:"); |
1985 | df_print_regset (dump_file, regs_set); | |
1986 | } | |
b8698a0f | 1987 | |
02b47899 KZ |
1988 | BITMAP_FREE (regs_set); |
1989 | return false; | |
1990 | } | |
1991 | BITMAP_FREE (regs_set); | |
1992 | } | |
1993 | ||
8660aaae | 1994 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
6fb5fa3c | 1995 | { |
fd6daec9 | 1996 | deferred_change *change = deferred_change_pool.allocate (); |
b8698a0f | 1997 | |
8660aaae EC |
1998 | /* Insert this right before the store insn where it will be safe |
1999 | from later insns that might change it before the read. */ | |
2000 | emit_insn_before (insns, store_insn->insn); | |
b8698a0f | 2001 | |
8660aaae EC |
2002 | /* And now for the kludge part: cselib croaks if you just |
2003 | return at this point. There are two reasons for this: | |
b8698a0f | 2004 | |
8660aaae EC |
2005 | 1) Cselib has an idea of how many pseudos there are and |
2006 | that does not include the new ones we just added. | |
b8698a0f | 2007 | |
8660aaae EC |
2008 | 2) Cselib does not know about the move insn we added |
2009 | above the store_info, and there is no way to tell it | |
2010 | about it, because it has "moved on". | |
b8698a0f | 2011 | |
8660aaae EC |
2012 | Problem (1) is fixable with a certain amount of engineering. |
2013 | Problem (2) is requires starting the bb from scratch. This | |
2014 | could be expensive. | |
b8698a0f | 2015 | |
8660aaae EC |
2016 | So we are just going to have to lie. The move/extraction |
2017 | insns are not really an issue, cselib did not see them. But | |
2018 | the use of the new pseudo read_insn is a real problem because | |
2019 | cselib has not scanned this insn. The way that we solve this | |
2020 | problem is that we are just going to put the mem back for now | |
2021 | and when we are finished with the block, we undo this. We | |
2022 | keep a table of mems to get rid of. At the end of the basic | |
2023 | block we can put them back. */ | |
b8698a0f | 2024 | |
8660aaae | 2025 | *loc = read_info->mem; |
da6603c6 ML |
2026 | change->next = deferred_change_list; |
2027 | deferred_change_list = change; | |
2028 | change->loc = loc; | |
2029 | change->reg = read_reg; | |
b8698a0f | 2030 | |
8660aaae EC |
2031 | /* Get rid of the read_info, from the point of view of the |
2032 | rest of dse, play like this read never happened. */ | |
2033 | read_insn->read_rec = read_info->next; | |
fb0b2914 | 2034 | read_info_type_pool.remove (read_info); |
456610d3 | 2035 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2036 | { |
2037 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2038 | print_simple_rtl (dump_file, read_reg); | |
2039 | fprintf (dump_file, "\n"); | |
2040 | } | |
8660aaae | 2041 | return true; |
6fb5fa3c | 2042 | } |
b8698a0f | 2043 | else |
6fb5fa3c | 2044 | { |
456610d3 | 2045 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2046 | { |
2047 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2048 | print_simple_rtl (dump_file, read_reg); | |
2049 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2050 | } | |
6fb5fa3c DB |
2051 | return false; |
2052 | } | |
2053 | } | |
2054 | ||
d7111da8 RS |
2055 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2056 | be active. */ | |
6fb5fa3c | 2057 | |
d7111da8 RS |
2058 | static void |
2059 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info) | |
6fb5fa3c | 2060 | { |
6216f94e | 2061 | rtx mem = *loc, mem_addr; |
6fb5fa3c | 2062 | insn_info_t insn_info; |
02ce5d90 RS |
2063 | poly_int64 offset = 0; |
2064 | poly_int64 width = 0; | |
b8698a0f | 2065 | cselib_val *base = NULL; |
6fb5fa3c DB |
2066 | int group_id; |
2067 | read_info_t read_info; | |
2068 | ||
6fb5fa3c DB |
2069 | insn_info = bb_info->last_insn; |
2070 | ||
2071 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2072 | || (MEM_VOLATILE_P (mem))) | |
2073 | { | |
456610d3 | 2074 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2075 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
2076 | add_wild_read (bb_info); | |
2077 | insn_info->cannot_delete = true; | |
d7111da8 | 2078 | return; |
6fb5fa3c DB |
2079 | } |
2080 | ||
2081 | /* If it is reading readonly mem, then there can be no conflict with | |
2082 | another write. */ | |
2083 | if (MEM_READONLY_P (mem)) | |
d7111da8 | 2084 | return; |
6fb5fa3c | 2085 | |
ac6929b5 | 2086 | if (!canon_address (mem, &group_id, &offset, &base)) |
6fb5fa3c | 2087 | { |
456610d3 | 2088 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2089 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
2090 | add_wild_read (bb_info); | |
d7111da8 | 2091 | return; |
6fb5fa3c DB |
2092 | } |
2093 | ||
2094 | if (GET_MODE (mem) == BLKmode) | |
2095 | width = -1; | |
2096 | else | |
2097 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2098 | ||
02ce5d90 | 2099 | if (!endpoint_representable_p (offset, known_eq (width, -1) ? 1 : width)) |
938f9248 | 2100 | { |
213ffde2 JJ |
2101 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2102 | fprintf (dump_file, " adding wild read, due to overflow.\n"); | |
2103 | add_wild_read (bb_info); | |
938f9248 ML |
2104 | return; |
2105 | } | |
2106 | ||
fb0b2914 | 2107 | read_info = read_info_type_pool.allocate (); |
6fb5fa3c DB |
2108 | read_info->group_id = group_id; |
2109 | read_info->mem = mem; | |
71d6a386 RS |
2110 | read_info->offset = offset; |
2111 | read_info->width = width; | |
6fb5fa3c DB |
2112 | read_info->next = insn_info->read_rec; |
2113 | insn_info->read_rec = read_info; | |
ac6929b5 RB |
2114 | if (group_id < 0) |
2115 | mem_addr = base->val_rtx; | |
6216f94e JJ |
2116 | else |
2117 | { | |
ac6929b5 RB |
2118 | group_info *group = rtx_group_vec[group_id]; |
2119 | mem_addr = group->canon_base_addr; | |
6216f94e | 2120 | } |
02ce5d90 | 2121 | if (maybe_ne (offset, 0)) |
ac6929b5 | 2122 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6fb5fa3c | 2123 | |
ac6929b5 | 2124 | if (group_id >= 0) |
6fb5fa3c DB |
2125 | { |
2126 | /* This is the restricted case where the base is a constant or | |
2127 | the frame pointer and offset is a constant. */ | |
2128 | insn_info_t i_ptr = active_local_stores; | |
2129 | insn_info_t last = NULL; | |
b8698a0f | 2130 | |
456610d3 | 2131 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2132 | { |
02ce5d90 | 2133 | if (!known_size_p (width)) |
6fb5fa3c DB |
2134 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", |
2135 | group_id); | |
2136 | else | |
71d6a386 RS |
2137 | { |
2138 | fprintf (dump_file, " processing const load gid=%d", group_id); | |
2139 | print_range (dump_file, offset, width); | |
2140 | fprintf (dump_file, "\n"); | |
2141 | } | |
6fb5fa3c DB |
2142 | } |
2143 | ||
2144 | while (i_ptr) | |
2145 | { | |
2146 | bool remove = false; | |
fd6daec9 | 2147 | store_info *store_info = i_ptr->store_rec; |
b8698a0f | 2148 | |
6fb5fa3c DB |
2149 | /* Skip the clobbers. */ |
2150 | while (!store_info->is_set) | |
2151 | store_info = store_info->next; | |
b8698a0f | 2152 | |
6fb5fa3c DB |
2153 | /* There are three cases here. */ |
2154 | if (store_info->group_id < 0) | |
2155 | /* We have a cselib store followed by a read from a | |
2156 | const base. */ | |
b8698a0f L |
2157 | remove |
2158 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2159 | GET_MODE (store_info->mem), |
2160 | store_info->mem_addr, | |
53d9622b | 2161 | mem, mem_addr); |
b8698a0f | 2162 | |
6fb5fa3c DB |
2163 | else if (group_id == store_info->group_id) |
2164 | { | |
2165 | /* This is a block mode load. We may get lucky and | |
2166 | canon_true_dependence may save the day. */ | |
02ce5d90 | 2167 | if (!known_size_p (width)) |
b8698a0f L |
2168 | remove |
2169 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2170 | GET_MODE (store_info->mem), |
2171 | store_info->mem_addr, | |
53d9622b | 2172 | mem, mem_addr); |
b8698a0f | 2173 | |
6fb5fa3c DB |
2174 | /* If this read is just reading back something that we just |
2175 | stored, rewrite the read. */ | |
b8698a0f | 2176 | else |
6fb5fa3c DB |
2177 | { |
2178 | if (store_info->rhs | |
71d6a386 RS |
2179 | && known_subrange_p (offset, width, store_info->offset, |
2180 | store_info->width) | |
8dd5516b | 2181 | && all_positions_needed_p (store_info, |
71d6a386 | 2182 | offset - store_info->offset, |
8dd5516b JJ |
2183 | width) |
2184 | && replace_read (store_info, i_ptr, read_info, | |
2185 | insn_info, loc, bb_info->regs_live)) | |
d7111da8 | 2186 | return; |
8dd5516b | 2187 | |
6fb5fa3c | 2188 | /* The bases are the same, just see if the offsets |
71d6a386 RS |
2189 | could overlap. */ |
2190 | if (ranges_maybe_overlap_p (offset, width, | |
2191 | store_info->offset, | |
2192 | store_info->width)) | |
6fb5fa3c DB |
2193 | remove = true; |
2194 | } | |
2195 | } | |
b8698a0f L |
2196 | |
2197 | /* else | |
6fb5fa3c DB |
2198 | The else case that is missing here is that the |
2199 | bases are constant but different. There is nothing | |
2200 | to do here because there is no overlap. */ | |
b8698a0f | 2201 | |
6fb5fa3c DB |
2202 | if (remove) |
2203 | { | |
456610d3 | 2204 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2205 | dump_insn_info ("removing from active", i_ptr); |
2206 | ||
dabd47e7 | 2207 | active_local_stores_len--; |
6fb5fa3c DB |
2208 | if (last) |
2209 | last->next_local_store = i_ptr->next_local_store; | |
2210 | else | |
2211 | active_local_stores = i_ptr->next_local_store; | |
2212 | } | |
2213 | else | |
2214 | last = i_ptr; | |
2215 | i_ptr = i_ptr->next_local_store; | |
2216 | } | |
2217 | } | |
b8698a0f | 2218 | else |
6fb5fa3c DB |
2219 | { |
2220 | insn_info_t i_ptr = active_local_stores; | |
2221 | insn_info_t last = NULL; | |
456610d3 | 2222 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2223 | { |
2224 | fprintf (dump_file, " processing cselib load mem:"); | |
2225 | print_inline_rtx (dump_file, mem, 0); | |
2226 | fprintf (dump_file, "\n"); | |
2227 | } | |
2228 | ||
2229 | while (i_ptr) | |
2230 | { | |
2231 | bool remove = false; | |
fd6daec9 | 2232 | store_info *store_info = i_ptr->store_rec; |
b8698a0f | 2233 | |
456610d3 | 2234 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2235 | fprintf (dump_file, " processing cselib load against insn %d\n", |
2236 | INSN_UID (i_ptr->insn)); | |
2237 | ||
2238 | /* Skip the clobbers. */ | |
2239 | while (!store_info->is_set) | |
2240 | store_info = store_info->next; | |
2241 | ||
2242 | /* If this read is just reading back something that we just | |
2243 | stored, rewrite the read. */ | |
2244 | if (store_info->rhs | |
2245 | && store_info->group_id == -1 | |
2246 | && store_info->cse_base == base | |
71d6a386 RS |
2247 | && known_subrange_p (offset, width, store_info->offset, |
2248 | store_info->width) | |
8dd5516b | 2249 | && all_positions_needed_p (store_info, |
71d6a386 | 2250 | offset - store_info->offset, width) |
8dd5516b JJ |
2251 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, |
2252 | bb_info->regs_live)) | |
d7111da8 | 2253 | return; |
6fb5fa3c | 2254 | |
ac6929b5 RB |
2255 | remove = canon_true_dependence (store_info->mem, |
2256 | GET_MODE (store_info->mem), | |
2257 | store_info->mem_addr, | |
2258 | mem, mem_addr); | |
b8698a0f | 2259 | |
6fb5fa3c DB |
2260 | if (remove) |
2261 | { | |
456610d3 | 2262 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2263 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2264 | |
dabd47e7 | 2265 | active_local_stores_len--; |
6fb5fa3c DB |
2266 | if (last) |
2267 | last->next_local_store = i_ptr->next_local_store; | |
2268 | else | |
2269 | active_local_stores = i_ptr->next_local_store; | |
2270 | } | |
2271 | else | |
2272 | last = i_ptr; | |
2273 | i_ptr = i_ptr->next_local_store; | |
2274 | } | |
2275 | } | |
6fb5fa3c DB |
2276 | } |
2277 | ||
d7111da8 | 2278 | /* A note_uses callback in which DATA points the INSN_INFO for |
6fb5fa3c DB |
2279 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2280 | true for any part of *LOC. */ | |
2281 | ||
2282 | static void | |
2283 | check_mem_read_use (rtx *loc, void *data) | |
2284 | { | |
d7111da8 RS |
2285 | subrtx_ptr_iterator::array_type array; |
2286 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
2287 | { | |
2288 | rtx *loc = *iter; | |
2289 | if (MEM_P (*loc)) | |
2290 | check_mem_read_rtx (loc, (bb_info_t) data); | |
2291 | } | |
6fb5fa3c DB |
2292 | } |
2293 | ||
8dd5516b JJ |
2294 | |
2295 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2296 | So far it only handles arguments passed in registers. */ | |
2297 | ||
2298 | static bool | |
2299 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2300 | { | |
d5cc9181 JR |
2301 | CUMULATIVE_ARGS args_so_far_v; |
2302 | cumulative_args_t args_so_far; | |
8dd5516b JJ |
2303 | tree arg; |
2304 | int idx; | |
2305 | ||
d5cc9181 JR |
2306 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2307 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
8dd5516b JJ |
2308 | |
2309 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2310 | for (idx = 0; | |
2311 | arg != void_list_node && idx < nargs; | |
2312 | arg = TREE_CHAIN (arg), idx++) | |
2313 | { | |
b4206259 | 2314 | scalar_int_mode mode; |
3c07301f | 2315 | rtx reg, link, tmp; |
b4206259 RS |
2316 | |
2317 | if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg)), &mode)) | |
2318 | return false; | |
2319 | ||
d5cc9181 | 2320 | reg = targetm.calls.function_arg (args_so_far, mode, NULL_TREE, true); |
b4206259 | 2321 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode) |
8dd5516b JJ |
2322 | return false; |
2323 | ||
2324 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2325 | link; | |
2326 | link = XEXP (link, 1)) | |
2327 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2328 | { | |
b4206259 | 2329 | scalar_int_mode arg_mode; |
8dd5516b JJ |
2330 | args[idx] = XEXP (XEXP (link, 0), 0); |
2331 | if (REG_P (args[idx]) | |
2332 | && REGNO (args[idx]) == REGNO (reg) | |
2333 | && (GET_MODE (args[idx]) == mode | |
b4206259 RS |
2334 | || (is_int_mode (GET_MODE (args[idx]), &arg_mode) |
2335 | && (GET_MODE_SIZE (arg_mode) <= UNITS_PER_WORD) | |
2336 | && (GET_MODE_SIZE (arg_mode) > GET_MODE_SIZE (mode))))) | |
8dd5516b JJ |
2337 | break; |
2338 | } | |
2339 | if (!link) | |
2340 | return false; | |
2341 | ||
2342 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2343 | if (GET_MODE (args[idx]) != mode) | |
2344 | { | |
2345 | if (!tmp || !CONST_INT_P (tmp)) | |
2346 | return false; | |
6d26322f | 2347 | tmp = gen_int_mode (INTVAL (tmp), mode); |
8dd5516b JJ |
2348 | } |
2349 | if (tmp) | |
2350 | args[idx] = tmp; | |
2351 | ||
d5cc9181 | 2352 | targetm.calls.function_arg_advance (args_so_far, mode, NULL_TREE, true); |
8dd5516b JJ |
2353 | } |
2354 | if (arg != void_list_node || idx != nargs) | |
2355 | return false; | |
2356 | return true; | |
2357 | } | |
2358 | ||
9e582b1d JR |
2359 | /* Return a bitmap of the fixed registers contained in IN. */ |
2360 | ||
2361 | static bitmap | |
2362 | copy_fixed_regs (const_bitmap in) | |
2363 | { | |
2364 | bitmap ret; | |
2365 | ||
2366 | ret = ALLOC_REG_SET (NULL); | |
2367 | bitmap_and (ret, in, fixed_reg_set_regset); | |
2368 | return ret; | |
2369 | } | |
8dd5516b | 2370 | |
6fb5fa3c DB |
2371 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2372 | if some part of it is not a candidate store and assigns to a | |
2373 | non-register target. */ | |
2374 | ||
2375 | static void | |
dd60a84c | 2376 | scan_insn (bb_info_t bb_info, rtx_insn *insn) |
6fb5fa3c DB |
2377 | { |
2378 | rtx body; | |
fb0b2914 | 2379 | insn_info_type *insn_info = insn_info_type_pool.allocate (); |
6fb5fa3c | 2380 | int mems_found = 0; |
da6603c6 | 2381 | memset (insn_info, 0, sizeof (struct insn_info_type)); |
6fb5fa3c | 2382 | |
456610d3 | 2383 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2384 | fprintf (dump_file, "\n**scanning insn=%d\n", |
2385 | INSN_UID (insn)); | |
2386 | ||
2387 | insn_info->prev_insn = bb_info->last_insn; | |
2388 | insn_info->insn = insn; | |
2389 | bb_info->last_insn = insn_info; | |
b8698a0f | 2390 | |
b5b8b0ac AO |
2391 | if (DEBUG_INSN_P (insn)) |
2392 | { | |
2393 | insn_info->cannot_delete = true; | |
2394 | return; | |
2395 | } | |
6fb5fa3c | 2396 | |
6fb5fa3c DB |
2397 | /* Look at all of the uses in the insn. */ |
2398 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2399 | ||
2400 | if (CALL_P (insn)) | |
2401 | { | |
8dd5516b | 2402 | bool const_call; |
ee516de9 | 2403 | rtx call, sym; |
8dd5516b JJ |
2404 | tree memset_call = NULL_TREE; |
2405 | ||
6fb5fa3c | 2406 | insn_info->cannot_delete = true; |
50f0f366 | 2407 | |
6fb5fa3c | 2408 | /* Const functions cannot do anything bad i.e. read memory, |
50f0f366 | 2409 | however, they can read their parameters which may have |
8dd5516b JJ |
2410 | been pushed onto the stack. |
2411 | memset and bzero don't read memory either. */ | |
2412 | const_call = RTL_CONST_CALL_P (insn); | |
ee516de9 EB |
2413 | if (!const_call |
2414 | && (call = get_call_rtx_from (insn)) | |
2415 | && (sym = XEXP (XEXP (call, 0), 0)) | |
2416 | && GET_CODE (sym) == SYMBOL_REF | |
2417 | && SYMBOL_REF_DECL (sym) | |
2418 | && TREE_CODE (SYMBOL_REF_DECL (sym)) == FUNCTION_DECL | |
2419 | && DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (sym)) == BUILT_IN_NORMAL | |
2420 | && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (sym)) == BUILT_IN_MEMSET) | |
2421 | memset_call = SYMBOL_REF_DECL (sym); | |
2422 | ||
8dd5516b | 2423 | if (const_call || memset_call) |
6fb5fa3c DB |
2424 | { |
2425 | insn_info_t i_ptr = active_local_stores; | |
2426 | insn_info_t last = NULL; | |
2427 | ||
456610d3 | 2428 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2429 | fprintf (dump_file, "%s call %d\n", |
2430 | const_call ? "const" : "memset", INSN_UID (insn)); | |
6fb5fa3c | 2431 | |
64520bdc | 2432 | /* See the head comment of the frame_read field. */ |
57534689 JJ |
2433 | if (reload_completed |
2434 | /* Tail calls are storing their arguments using | |
2435 | arg pointer. If it is a frame pointer on the target, | |
2436 | even before reload we need to kill frame pointer based | |
2437 | stores. */ | |
2438 | || (SIBLING_CALL_P (insn) | |
2439 | && HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
64520bdc EB |
2440 | insn_info->frame_read = true; |
2441 | ||
2442 | /* Loop over the active stores and remove those which are | |
2443 | killed by the const function call. */ | |
6fb5fa3c DB |
2444 | while (i_ptr) |
2445 | { | |
64520bdc EB |
2446 | bool remove_store = false; |
2447 | ||
2448 | /* The stack pointer based stores are always killed. */ | |
50f0f366 | 2449 | if (i_ptr->stack_pointer_based) |
64520bdc EB |
2450 | remove_store = true; |
2451 | ||
2452 | /* If the frame is read, the frame related stores are killed. */ | |
2453 | else if (insn_info->frame_read) | |
2454 | { | |
fd6daec9 | 2455 | store_info *store_info = i_ptr->store_rec; |
64520bdc EB |
2456 | |
2457 | /* Skip the clobbers. */ | |
2458 | while (!store_info->is_set) | |
2459 | store_info = store_info->next; | |
2460 | ||
2461 | if (store_info->group_id >= 0 | |
9771b263 | 2462 | && rtx_group_vec[store_info->group_id]->frame_related) |
64520bdc EB |
2463 | remove_store = true; |
2464 | } | |
2465 | ||
2466 | if (remove_store) | |
6fb5fa3c | 2467 | { |
456610d3 | 2468 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2469 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2470 | |
dabd47e7 | 2471 | active_local_stores_len--; |
6fb5fa3c DB |
2472 | if (last) |
2473 | last->next_local_store = i_ptr->next_local_store; | |
2474 | else | |
2475 | active_local_stores = i_ptr->next_local_store; | |
2476 | } | |
2477 | else | |
2478 | last = i_ptr; | |
64520bdc | 2479 | |
6fb5fa3c DB |
2480 | i_ptr = i_ptr->next_local_store; |
2481 | } | |
8dd5516b JJ |
2482 | |
2483 | if (memset_call) | |
2484 | { | |
2485 | rtx args[3]; | |
2486 | if (get_call_args (insn, memset_call, args, 3) | |
2487 | && CONST_INT_P (args[1]) | |
2488 | && CONST_INT_P (args[2]) | |
2489 | && INTVAL (args[2]) > 0) | |
2490 | { | |
2491 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
f5541398 | 2492 | set_mem_size (mem, INTVAL (args[2])); |
f7df4a84 | 2493 | body = gen_rtx_SET (mem, args[1]); |
8dd5516b | 2494 | mems_found += record_store (body, bb_info); |
456610d3 | 2495 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2496 | fprintf (dump_file, "handling memset as BLKmode store\n"); |
2497 | if (mems_found == 1) | |
2498 | { | |
dabd47e7 JJ |
2499 | if (active_local_stores_len++ |
2500 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2501 | { | |
2502 | active_local_stores_len = 1; | |
2503 | active_local_stores = NULL; | |
2504 | } | |
9e582b1d JR |
2505 | insn_info->fixed_regs_live |
2506 | = copy_fixed_regs (bb_info->regs_live); | |
8dd5516b JJ |
2507 | insn_info->next_local_store = active_local_stores; |
2508 | active_local_stores = insn_info; | |
2509 | } | |
2510 | } | |
a2ad54a0 JJ |
2511 | else |
2512 | clear_rhs_from_active_local_stores (); | |
8dd5516b | 2513 | } |
6fb5fa3c | 2514 | } |
57534689 JJ |
2515 | else if (SIBLING_CALL_P (insn) && reload_completed) |
2516 | /* Arguments for a sibling call that are pushed to memory are passed | |
2517 | using the incoming argument pointer of the current function. After | |
2518 | reload that might be (and likely is) frame pointer based. */ | |
2519 | add_wild_read (bb_info); | |
50f0f366 | 2520 | else |
d26c7090 ER |
2521 | /* Every other call, including pure functions, may read any memory |
2522 | that is not relative to the frame. */ | |
2523 | add_non_frame_wild_read (bb_info); | |
50f0f366 | 2524 | |
6fb5fa3c DB |
2525 | return; |
2526 | } | |
2527 | ||
2528 | /* Assuming that there are sets in these insns, we cannot delete | |
2529 | them. */ | |
2530 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
0a64eeca | 2531 | || volatile_refs_p (PATTERN (insn)) |
2da02156 | 2532 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
6fb5fa3c DB |
2533 | || (RTX_FRAME_RELATED_P (insn)) |
2534 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2535 | insn_info->cannot_delete = true; | |
b8698a0f | 2536 | |
6fb5fa3c DB |
2537 | body = PATTERN (insn); |
2538 | if (GET_CODE (body) == PARALLEL) | |
2539 | { | |
2540 | int i; | |
2541 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2542 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2543 | } | |
2544 | else | |
2545 | mems_found += record_store (body, bb_info); | |
2546 | ||
456610d3 | 2547 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2548 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
6fb5fa3c DB |
2549 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2550 | ||
8dd5516b JJ |
2551 | /* If we found some sets of mems, add it into the active_local_stores so |
2552 | that it can be locally deleted if found dead or used for | |
2553 | replace_read and redundant constant store elimination. Otherwise mark | |
2554 | it as cannot delete. This simplifies the processing later. */ | |
2555 | if (mems_found == 1) | |
6fb5fa3c | 2556 | { |
dabd47e7 JJ |
2557 | if (active_local_stores_len++ |
2558 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2559 | { | |
2560 | active_local_stores_len = 1; | |
2561 | active_local_stores = NULL; | |
2562 | } | |
9e582b1d | 2563 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
6fb5fa3c DB |
2564 | insn_info->next_local_store = active_local_stores; |
2565 | active_local_stores = insn_info; | |
2566 | } | |
2567 | else | |
2568 | insn_info->cannot_delete = true; | |
2569 | } | |
2570 | ||
2571 | ||
2572 | /* Remove BASE from the set of active_local_stores. This is a | |
2573 | callback from cselib that is used to get rid of the stores in | |
2574 | active_local_stores. */ | |
2575 | ||
2576 | static void | |
2577 | remove_useless_values (cselib_val *base) | |
2578 | { | |
2579 | insn_info_t insn_info = active_local_stores; | |
2580 | insn_info_t last = NULL; | |
2581 | ||
2582 | while (insn_info) | |
2583 | { | |
fd6daec9 | 2584 | store_info *store_info = insn_info->store_rec; |
60564289 | 2585 | bool del = false; |
6fb5fa3c DB |
2586 | |
2587 | /* If ANY of the store_infos match the cselib group that is | |
2588 | being deleted, then the insn can not be deleted. */ | |
2589 | while (store_info) | |
2590 | { | |
b8698a0f | 2591 | if ((store_info->group_id == -1) |
6fb5fa3c DB |
2592 | && (store_info->cse_base == base)) |
2593 | { | |
60564289 | 2594 | del = true; |
6fb5fa3c DB |
2595 | break; |
2596 | } | |
2597 | store_info = store_info->next; | |
2598 | } | |
2599 | ||
60564289 | 2600 | if (del) |
6fb5fa3c | 2601 | { |
dabd47e7 | 2602 | active_local_stores_len--; |
6fb5fa3c DB |
2603 | if (last) |
2604 | last->next_local_store = insn_info->next_local_store; | |
2605 | else | |
2606 | active_local_stores = insn_info->next_local_store; | |
2607 | free_store_info (insn_info); | |
2608 | } | |
2609 | else | |
2610 | last = insn_info; | |
b8698a0f | 2611 | |
6fb5fa3c DB |
2612 | insn_info = insn_info->next_local_store; |
2613 | } | |
2614 | } | |
2615 | ||
2616 | ||
2617 | /* Do all of step 1. */ | |
2618 | ||
2619 | static void | |
2620 | dse_step1 (void) | |
2621 | { | |
2622 | basic_block bb; | |
3f9b14ff | 2623 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
b8698a0f | 2624 | |
457eeaae | 2625 | cselib_init (0); |
6fb5fa3c DB |
2626 | all_blocks = BITMAP_ALLOC (NULL); |
2627 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2628 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2629 | ||
04a90bec | 2630 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
2631 | { |
2632 | insn_info_t ptr; | |
fb0b2914 | 2633 | bb_info_t bb_info = dse_bb_info_type_pool.allocate (); |
6fb5fa3c | 2634 | |
da6603c6 | 2635 | memset (bb_info, 0, sizeof (dse_bb_info_type)); |
6fb5fa3c | 2636 | bitmap_set_bit (all_blocks, bb->index); |
02b47899 KZ |
2637 | bb_info->regs_live = regs_live; |
2638 | ||
2639 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2640 | df_simulate_initialize_forwards (bb, regs_live); | |
6fb5fa3c DB |
2641 | |
2642 | bb_table[bb->index] = bb_info; | |
2643 | cselib_discard_hook = remove_useless_values; | |
2644 | ||
2645 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2646 | { | |
dd60a84c | 2647 | rtx_insn *insn; |
6fb5fa3c | 2648 | |
6fb5fa3c | 2649 | active_local_stores = NULL; |
dabd47e7 | 2650 | active_local_stores_len = 0; |
6fb5fa3c | 2651 | cselib_clear_table (); |
b8698a0f | 2652 | |
6fb5fa3c DB |
2653 | /* Scan the insns. */ |
2654 | FOR_BB_INSNS (bb, insn) | |
2655 | { | |
2656 | if (INSN_P (insn)) | |
2657 | scan_insn (bb_info, insn); | |
2658 | cselib_process_insn (insn); | |
02b47899 KZ |
2659 | if (INSN_P (insn)) |
2660 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
6fb5fa3c | 2661 | } |
b8698a0f | 2662 | |
6fb5fa3c DB |
2663 | /* This is something of a hack, because the global algorithm |
2664 | is supposed to take care of the case where stores go dead | |
2665 | at the end of the function. However, the global | |
2666 | algorithm must take a more conservative view of block | |
2667 | mode reads than the local alg does. So to get the case | |
2668 | where you have a store to the frame followed by a non | |
0d52bcc1 | 2669 | overlapping block more read, we look at the active local |
6fb5fa3c DB |
2670 | stores at the end of the function and delete all of the |
2671 | frame and spill based ones. */ | |
2672 | if (stores_off_frame_dead_at_return | |
2673 | && (EDGE_COUNT (bb->succs) == 0 | |
2674 | || (single_succ_p (bb) | |
fefa31b5 | 2675 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
e3b5732b | 2676 | && ! crtl->calls_eh_return))) |
6fb5fa3c DB |
2677 | { |
2678 | insn_info_t i_ptr = active_local_stores; | |
2679 | while (i_ptr) | |
2680 | { | |
fd6daec9 | 2681 | store_info *store_info = i_ptr->store_rec; |
6fb5fa3c DB |
2682 | |
2683 | /* Skip the clobbers. */ | |
2684 | while (!store_info->is_set) | |
2685 | store_info = store_info->next; | |
ac6929b5 RB |
2686 | if (store_info->group_id >= 0) |
2687 | { | |
2688 | group_info *group = rtx_group_vec[store_info->group_id]; | |
2689 | if (group->frame_related && !i_ptr->cannot_delete) | |
2690 | delete_dead_store_insn (i_ptr); | |
2691 | } | |
6fb5fa3c DB |
2692 | |
2693 | i_ptr = i_ptr->next_local_store; | |
2694 | } | |
2695 | } | |
2696 | ||
2697 | /* Get rid of the loads that were discovered in | |
2698 | replace_read. Cselib is finished with this block. */ | |
2699 | while (deferred_change_list) | |
2700 | { | |
fd6daec9 | 2701 | deferred_change *next = deferred_change_list->next; |
6fb5fa3c DB |
2702 | |
2703 | /* There is no reason to validate this change. That was | |
2704 | done earlier. */ | |
2705 | *deferred_change_list->loc = deferred_change_list->reg; | |
fb0b2914 | 2706 | deferred_change_pool.remove (deferred_change_list); |
6fb5fa3c DB |
2707 | deferred_change_list = next; |
2708 | } | |
2709 | ||
2710 | /* Get rid of all of the cselib based store_infos in this | |
2711 | block and mark the containing insns as not being | |
2712 | deletable. */ | |
2713 | ptr = bb_info->last_insn; | |
2714 | while (ptr) | |
2715 | { | |
2716 | if (ptr->contains_cselib_groups) | |
8dd5516b | 2717 | { |
fd6daec9 | 2718 | store_info *s_info = ptr->store_rec; |
8dd5516b JJ |
2719 | while (s_info && !s_info->is_set) |
2720 | s_info = s_info->next; | |
2721 | if (s_info | |
2722 | && s_info->redundant_reason | |
2723 | && s_info->redundant_reason->insn | |
2724 | && !ptr->cannot_delete) | |
2725 | { | |
456610d3 | 2726 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2727 | fprintf (dump_file, "Locally deleting insn %d " |
2728 | "because insn %d stores the " | |
2729 | "same value and couldn't be " | |
2730 | "eliminated\n", | |
2731 | INSN_UID (ptr->insn), | |
2732 | INSN_UID (s_info->redundant_reason->insn)); | |
2733 | delete_dead_store_insn (ptr); | |
2734 | } | |
8dd5516b JJ |
2735 | free_store_info (ptr); |
2736 | } | |
2737 | else | |
2738 | { | |
fd6daec9 | 2739 | store_info *s_info; |
8dd5516b JJ |
2740 | |
2741 | /* Free at least positions_needed bitmaps. */ | |
2742 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2743 | if (s_info->is_large) | |
2744 | { | |
dc491a25 | 2745 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
8dd5516b JJ |
2746 | s_info->is_large = false; |
2747 | } | |
2748 | } | |
6fb5fa3c DB |
2749 | ptr = ptr->prev_insn; |
2750 | } | |
2751 | ||
da6603c6 | 2752 | cse_store_info_pool.release (); |
6fb5fa3c | 2753 | } |
02b47899 | 2754 | bb_info->regs_live = NULL; |
6fb5fa3c DB |
2755 | } |
2756 | ||
02b47899 | 2757 | BITMAP_FREE (regs_live); |
6fb5fa3c | 2758 | cselib_finish (); |
c203e8a7 | 2759 | rtx_group_table->empty (); |
6fb5fa3c DB |
2760 | } |
2761 | ||
2762 | \f | |
2763 | /*---------------------------------------------------------------------------- | |
2764 | Second step. | |
2765 | ||
2766 | Assign each byte position in the stores that we are going to | |
2767 | analyze globally to a position in the bitmaps. Returns true if | |
6ed3da00 | 2768 | there are any bit positions assigned. |
6fb5fa3c DB |
2769 | ----------------------------------------------------------------------------*/ |
2770 | ||
2771 | static void | |
2772 | dse_step2_init (void) | |
2773 | { | |
2774 | unsigned int i; | |
fd6daec9 | 2775 | group_info *group; |
6fb5fa3c | 2776 | |
9771b263 | 2777 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2778 | { |
2779 | /* For all non stack related bases, we only consider a store to | |
2780 | be deletable if there are two or more stores for that | |
2781 | position. This is because it takes one store to make the | |
2782 | other store redundant. However, for the stores that are | |
2783 | stack related, we consider them if there is only one store | |
2784 | for the position. We do this because the stack related | |
2785 | stores can be deleted if their is no read between them and | |
2786 | the end of the function. | |
b8698a0f | 2787 | |
6fb5fa3c DB |
2788 | To make this work in the current framework, we take the stack |
2789 | related bases add all of the bits from store1 into store2. | |
2790 | This has the effect of making the eligible even if there is | |
2791 | only one store. */ | |
2792 | ||
2793 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2794 | { | |
2795 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2796 | bitmap_ior_into (group->store2_p, group->store1_p); | |
456610d3 | 2797 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2798 | fprintf (dump_file, "group %d is frame related ", i); |
6fb5fa3c DB |
2799 | } |
2800 | ||
2801 | group->offset_map_size_n++; | |
3f9b14ff SB |
2802 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2803 | group->offset_map_size_n); | |
6fb5fa3c | 2804 | group->offset_map_size_p++; |
3f9b14ff SB |
2805 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2806 | group->offset_map_size_p); | |
6fb5fa3c | 2807 | group->process_globally = false; |
456610d3 | 2808 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2809 | { |
b8698a0f | 2810 | fprintf (dump_file, "group %d(%d+%d): ", i, |
6fb5fa3c DB |
2811 | (int)bitmap_count_bits (group->store2_n), |
2812 | (int)bitmap_count_bits (group->store2_p)); | |
2813 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2814 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2815 | } | |
2816 | } | |
2817 | } | |
2818 | ||
2819 | ||
ac6929b5 | 2820 | /* Init the offset tables. */ |
6fb5fa3c DB |
2821 | |
2822 | static bool | |
ac6929b5 | 2823 | dse_step2 (void) |
6fb5fa3c DB |
2824 | { |
2825 | unsigned int i; | |
fd6daec9 | 2826 | group_info *group; |
6fb5fa3c DB |
2827 | /* Position 0 is unused because 0 is used in the maps to mean |
2828 | unused. */ | |
2829 | current_position = 1; | |
9771b263 | 2830 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2831 | { |
2832 | bitmap_iterator bi; | |
2833 | unsigned int j; | |
2834 | ||
c3284718 RS |
2835 | memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
2836 | memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
6fb5fa3c DB |
2837 | bitmap_clear (group->group_kill); |
2838 | ||
2839 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2840 | { | |
2841 | bitmap_set_bit (group->group_kill, current_position); | |
d26c7090 ER |
2842 | if (bitmap_bit_p (group->escaped_n, j)) |
2843 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2844 | group->offset_map_n[j] = current_position++; |
2845 | group->process_globally = true; | |
2846 | } | |
2847 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2848 | { | |
b8698a0f | 2849 | bitmap_set_bit (group->group_kill, current_position); |
d26c7090 ER |
2850 | if (bitmap_bit_p (group->escaped_p, j)) |
2851 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2852 | group->offset_map_p[j] = current_position++; |
2853 | group->process_globally = true; | |
2854 | } | |
2855 | } | |
2856 | return current_position != 1; | |
2857 | } | |
2858 | ||
2859 | ||
6fb5fa3c DB |
2860 | \f |
2861 | /*---------------------------------------------------------------------------- | |
2862 | Third step. | |
b8698a0f | 2863 | |
6fb5fa3c DB |
2864 | Build the bit vectors for the transfer functions. |
2865 | ----------------------------------------------------------------------------*/ | |
2866 | ||
2867 | ||
6fb5fa3c DB |
2868 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
2869 | there, return 0. */ | |
2870 | ||
2871 | static int | |
fd6daec9 | 2872 | get_bitmap_index (group_info *group_info, HOST_WIDE_INT offset) |
6fb5fa3c DB |
2873 | { |
2874 | if (offset < 0) | |
2875 | { | |
2876 | HOST_WIDE_INT offset_p = -offset; | |
2877 | if (offset_p >= group_info->offset_map_size_n) | |
2878 | return 0; | |
2879 | return group_info->offset_map_n[offset_p]; | |
2880 | } | |
2881 | else | |
2882 | { | |
2883 | if (offset >= group_info->offset_map_size_p) | |
2884 | return 0; | |
2885 | return group_info->offset_map_p[offset]; | |
2886 | } | |
2887 | } | |
2888 | ||
2889 | ||
2890 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
2891 | may be NULL. */ | |
2892 | ||
b8698a0f | 2893 | static void |
ac6929b5 | 2894 | scan_stores (store_info *store_info, bitmap gen, bitmap kill) |
6fb5fa3c DB |
2895 | { |
2896 | while (store_info) | |
2897 | { | |
02ce5d90 | 2898 | HOST_WIDE_INT i, offset, width; |
fd6daec9 | 2899 | group_info *group_info |
9771b263 | 2900 | = rtx_group_vec[store_info->group_id]; |
02ce5d90 RS |
2901 | /* We can (conservatively) ignore stores whose bounds aren't known; |
2902 | they simply don't generate new global dse opportunities. */ | |
2903 | if (group_info->process_globally | |
2904 | && store_info->offset.is_constant (&offset) | |
2905 | && store_info->width.is_constant (&width)) | |
71d6a386 | 2906 | { |
02ce5d90 RS |
2907 | HOST_WIDE_INT end = offset + width; |
2908 | for (i = offset; i < end; i++) | |
71d6a386 RS |
2909 | { |
2910 | int index = get_bitmap_index (group_info, i); | |
2911 | if (index != 0) | |
2912 | { | |
2913 | bitmap_set_bit (gen, index); | |
2914 | if (kill) | |
2915 | bitmap_clear_bit (kill, index); | |
2916 | } | |
2917 | } | |
2918 | } | |
6fb5fa3c DB |
2919 | store_info = store_info->next; |
2920 | } | |
2921 | } | |
2922 | ||
2923 | ||
6fb5fa3c DB |
2924 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL |
2925 | may be NULL. */ | |
2926 | ||
2927 | static void | |
ac6929b5 | 2928 | scan_reads (insn_info_t insn_info, bitmap gen, bitmap kill) |
6fb5fa3c DB |
2929 | { |
2930 | read_info_t read_info = insn_info->read_rec; | |
2931 | int i; | |
fd6daec9 | 2932 | group_info *group; |
6fb5fa3c | 2933 | |
64520bdc EB |
2934 | /* If this insn reads the frame, kill all the frame related stores. */ |
2935 | if (insn_info->frame_read) | |
2936 | { | |
9771b263 | 2937 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
64520bdc EB |
2938 | if (group->process_globally && group->frame_related) |
2939 | { | |
2940 | if (kill) | |
2941 | bitmap_ior_into (kill, group->group_kill); | |
b8698a0f | 2942 | bitmap_and_compl_into (gen, group->group_kill); |
64520bdc EB |
2943 | } |
2944 | } | |
d26c7090 ER |
2945 | if (insn_info->non_frame_wild_read) |
2946 | { | |
2947 | /* Kill all non-frame related stores. Kill all stores of variables that | |
2948 | escape. */ | |
2949 | if (kill) | |
2950 | bitmap_ior_into (kill, kill_on_calls); | |
2951 | bitmap_and_compl_into (gen, kill_on_calls); | |
9771b263 | 2952 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
d26c7090 ER |
2953 | if (group->process_globally && !group->frame_related) |
2954 | { | |
2955 | if (kill) | |
2956 | bitmap_ior_into (kill, group->group_kill); | |
2957 | bitmap_and_compl_into (gen, group->group_kill); | |
2958 | } | |
2959 | } | |
6fb5fa3c DB |
2960 | while (read_info) |
2961 | { | |
9771b263 | 2962 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2963 | { |
2964 | if (group->process_globally) | |
2965 | { | |
2966 | if (i == read_info->group_id) | |
2967 | { | |
02ce5d90 RS |
2968 | HOST_WIDE_INT offset, width; |
2969 | /* Reads with non-constant size kill all DSE opportunities | |
2970 | in the group. */ | |
2971 | if (!read_info->offset.is_constant (&offset) | |
2972 | || !read_info->width.is_constant (&width) | |
2973 | || !known_size_p (width)) | |
6fb5fa3c | 2974 | { |
71d6a386 | 2975 | /* Handle block mode reads. */ |
6fb5fa3c DB |
2976 | if (kill) |
2977 | bitmap_ior_into (kill, group->group_kill); | |
2978 | bitmap_and_compl_into (gen, group->group_kill); | |
2979 | } | |
2980 | else | |
2981 | { | |
2982 | /* The groups are the same, just process the | |
2983 | offsets. */ | |
2984 | HOST_WIDE_INT j; | |
02ce5d90 RS |
2985 | HOST_WIDE_INT end = offset + width; |
2986 | for (j = offset; j < end; j++) | |
6fb5fa3c DB |
2987 | { |
2988 | int index = get_bitmap_index (group, j); | |
2989 | if (index != 0) | |
2990 | { | |
2991 | if (kill) | |
2992 | bitmap_set_bit (kill, index); | |
2993 | bitmap_clear_bit (gen, index); | |
2994 | } | |
2995 | } | |
2996 | } | |
2997 | } | |
2998 | else | |
2999 | { | |
3000 | /* The groups are different, if the alias sets | |
3001 | conflict, clear the entire group. We only need | |
3002 | to apply this test if the read_info is a cselib | |
3003 | read. Anything with a constant base cannot alias | |
3004 | something else with a different constant | |
3005 | base. */ | |
3006 | if ((read_info->group_id < 0) | |
b8698a0f | 3007 | && canon_true_dependence (group->base_mem, |
d32f725a | 3008 | GET_MODE (group->base_mem), |
6216f94e | 3009 | group->canon_base_addr, |
53d9622b | 3010 | read_info->mem, NULL_RTX)) |
6fb5fa3c DB |
3011 | { |
3012 | if (kill) | |
3013 | bitmap_ior_into (kill, group->group_kill); | |
3014 | bitmap_and_compl_into (gen, group->group_kill); | |
3015 | } | |
3016 | } | |
3017 | } | |
3018 | } | |
b8698a0f | 3019 | |
6fb5fa3c DB |
3020 | read_info = read_info->next; |
3021 | } | |
3022 | } | |
3023 | ||
6fb5fa3c DB |
3024 | |
3025 | /* Return the insn in BB_INFO before the first wild read or if there | |
3026 | are no wild reads in the block, return the last insn. */ | |
3027 | ||
3028 | static insn_info_t | |
3029 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3030 | { | |
3031 | insn_info_t insn_info = bb_info->last_insn; | |
3032 | insn_info_t last_wild_read = NULL; | |
3033 | ||
3034 | while (insn_info) | |
3035 | { | |
3036 | if (insn_info->wild_read) | |
3037 | { | |
3038 | last_wild_read = insn_info->prev_insn; | |
3039 | /* Block starts with wild read. */ | |
3040 | if (!last_wild_read) | |
3041 | return NULL; | |
3042 | } | |
3043 | ||
3044 | insn_info = insn_info->prev_insn; | |
3045 | } | |
3046 | ||
3047 | if (last_wild_read) | |
3048 | return last_wild_read; | |
3049 | else | |
3050 | return bb_info->last_insn; | |
3051 | } | |
3052 | ||
3053 | ||
3054 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3055 | the block in order to build the gen and kill sets for the block. | |
3056 | We start at ptr which may be the last insn in the block or may be | |
3057 | the first insn with a wild read. In the latter case we are able to | |
3058 | skip the rest of the block because it just does not matter: | |
3059 | anything that happens is hidden by the wild read. */ | |
3060 | ||
3061 | static void | |
ac6929b5 | 3062 | dse_step3_scan (basic_block bb) |
6fb5fa3c DB |
3063 | { |
3064 | bb_info_t bb_info = bb_table[bb->index]; | |
3065 | insn_info_t insn_info; | |
3066 | ||
ac6929b5 | 3067 | insn_info = find_insn_before_first_wild_read (bb_info); |
b8698a0f | 3068 | |
6fb5fa3c DB |
3069 | /* In the spill case or in the no_spill case if there is no wild |
3070 | read in the block, we will need a kill set. */ | |
3071 | if (insn_info == bb_info->last_insn) | |
3072 | { | |
3073 | if (bb_info->kill) | |
3074 | bitmap_clear (bb_info->kill); | |
3075 | else | |
3f9b14ff | 3076 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c | 3077 | } |
b8698a0f | 3078 | else |
6fb5fa3c DB |
3079 | if (bb_info->kill) |
3080 | BITMAP_FREE (bb_info->kill); | |
3081 | ||
3082 | while (insn_info) | |
3083 | { | |
3084 | /* There may have been code deleted by the dce pass run before | |
3085 | this phase. */ | |
3086 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3087 | { | |
ac6929b5 RB |
3088 | scan_stores (insn_info->store_rec, bb_info->gen, bb_info->kill); |
3089 | scan_reads (insn_info, bb_info->gen, bb_info->kill); | |
b8698a0f | 3090 | } |
6fb5fa3c DB |
3091 | |
3092 | insn_info = insn_info->prev_insn; | |
3093 | } | |
3094 | } | |
3095 | ||
3096 | ||
3097 | /* Set the gen set of the exit block, and also any block with no | |
3098 | successors that does not have a wild read. */ | |
3099 | ||
3100 | static void | |
3101 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3102 | { | |
3103 | /* The gen set is all 0's for the exit block except for the | |
3104 | frame_pointer_group. */ | |
b8698a0f | 3105 | |
6fb5fa3c DB |
3106 | if (stores_off_frame_dead_at_return) |
3107 | { | |
3108 | unsigned int i; | |
fd6daec9 | 3109 | group_info *group; |
b8698a0f | 3110 | |
9771b263 | 3111 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3112 | { |
3113 | if (group->process_globally && group->frame_related) | |
3114 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3115 | } | |
3116 | } | |
3117 | } | |
3118 | ||
3119 | ||
3120 | /* Find all of the blocks that are not backwards reachable from the | |
3121 | exit block or any block with no successors (BB). These are the | |
3122 | infinite loops or infinite self loops. These blocks will still | |
3123 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3124 | ||
3125 | static void | |
3126 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3127 | { | |
3128 | edge e; | |
3129 | edge_iterator ei; | |
3130 | ||
d7c028c0 | 3131 | if (bitmap_bit_p (unreachable_blocks, bb->index)) |
6fb5fa3c | 3132 | { |
d7c028c0 | 3133 | bitmap_clear_bit (unreachable_blocks, bb->index); |
6fb5fa3c | 3134 | FOR_EACH_EDGE (e, ei, bb->preds) |
b8698a0f | 3135 | { |
6fb5fa3c | 3136 | mark_reachable_blocks (unreachable_blocks, e->src); |
b8698a0f | 3137 | } |
6fb5fa3c DB |
3138 | } |
3139 | } | |
3140 | ||
3141 | /* Build the transfer functions for the function. */ | |
3142 | ||
3143 | static void | |
ac6929b5 | 3144 | dse_step3 () |
6fb5fa3c DB |
3145 | { |
3146 | basic_block bb; | |
6fb5fa3c DB |
3147 | sbitmap_iterator sbi; |
3148 | bitmap all_ones = NULL; | |
3149 | unsigned int i; | |
b8698a0f | 3150 | |
7ba9e72d | 3151 | auto_sbitmap unreachable_blocks (last_basic_block_for_fn (cfun)); |
f61e445a | 3152 | bitmap_ones (unreachable_blocks); |
6fb5fa3c | 3153 | |
04a90bec | 3154 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3155 | { |
3156 | bb_info_t bb_info = bb_table[bb->index]; | |
3157 | if (bb_info->gen) | |
3158 | bitmap_clear (bb_info->gen); | |
3159 | else | |
3f9b14ff | 3160 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3161 | |
3162 | if (bb->index == ENTRY_BLOCK) | |
3163 | ; | |
3164 | else if (bb->index == EXIT_BLOCK) | |
3165 | dse_step3_exit_block_scan (bb_info); | |
3166 | else | |
ac6929b5 | 3167 | dse_step3_scan (bb); |
6fb5fa3c DB |
3168 | if (EDGE_COUNT (bb->succs) == 0) |
3169 | mark_reachable_blocks (unreachable_blocks, bb); | |
3170 | ||
3171 | /* If this is the second time dataflow is run, delete the old | |
3172 | sets. */ | |
3173 | if (bb_info->in) | |
3174 | BITMAP_FREE (bb_info->in); | |
3175 | if (bb_info->out) | |
3176 | BITMAP_FREE (bb_info->out); | |
3177 | } | |
3178 | ||
3179 | /* For any block in an infinite loop, we must initialize the out set | |
3180 | to all ones. This could be expensive, but almost never occurs in | |
3181 | practice. However, it is common in regression tests. */ | |
d4ac4ce2 | 3182 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
6fb5fa3c DB |
3183 | { |
3184 | if (bitmap_bit_p (all_blocks, i)) | |
3185 | { | |
3186 | bb_info_t bb_info = bb_table[i]; | |
3187 | if (!all_ones) | |
3188 | { | |
3189 | unsigned int j; | |
fd6daec9 | 3190 | group_info *group; |
6fb5fa3c | 3191 | |
3f9b14ff | 3192 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
9771b263 | 3193 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
6fb5fa3c DB |
3194 | bitmap_ior_into (all_ones, group->group_kill); |
3195 | } | |
3196 | if (!bb_info->out) | |
3197 | { | |
3f9b14ff | 3198 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3199 | bitmap_copy (bb_info->out, all_ones); |
3200 | } | |
3201 | } | |
3202 | } | |
3203 | ||
3204 | if (all_ones) | |
3205 | BITMAP_FREE (all_ones); | |
6fb5fa3c DB |
3206 | } |
3207 | ||
3208 | ||
3209 | \f | |
3210 | /*---------------------------------------------------------------------------- | |
3211 | Fourth step. | |
3212 | ||
3213 | Solve the bitvector equations. | |
3214 | ----------------------------------------------------------------------------*/ | |
3215 | ||
3216 | ||
3217 | /* Confluence function for blocks with no successors. Create an out | |
3218 | set from the gen set of the exit block. This block logically has | |
3219 | the exit block as a successor. */ | |
3220 | ||
3221 | ||
3222 | ||
3223 | static void | |
3224 | dse_confluence_0 (basic_block bb) | |
3225 | { | |
3226 | bb_info_t bb_info = bb_table[bb->index]; | |
3227 | ||
3228 | if (bb->index == EXIT_BLOCK) | |
3229 | return; | |
3230 | ||
3231 | if (!bb_info->out) | |
3232 | { | |
3f9b14ff | 3233 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3234 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3235 | } | |
3236 | } | |
3237 | ||
3238 | /* Propagate the information from the in set of the dest of E to the | |
3239 | out set of the src of E. If the various in or out sets are not | |
3240 | there, that means they are all ones. */ | |
3241 | ||
1a0f3fa1 | 3242 | static bool |
6fb5fa3c DB |
3243 | dse_confluence_n (edge e) |
3244 | { | |
3245 | bb_info_t src_info = bb_table[e->src->index]; | |
3246 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3247 | ||
3248 | if (dest_info->in) | |
3249 | { | |
3250 | if (src_info->out) | |
3251 | bitmap_and_into (src_info->out, dest_info->in); | |
3252 | else | |
3253 | { | |
3f9b14ff | 3254 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3255 | bitmap_copy (src_info->out, dest_info->in); |
3256 | } | |
3257 | } | |
1a0f3fa1 | 3258 | return true; |
6fb5fa3c DB |
3259 | } |
3260 | ||
3261 | ||
3262 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
b8698a0f | 3263 | block. There are three cases: |
6fb5fa3c DB |
3264 | |
3265 | 1) The block has no kill set. In this case the kill set is all | |
3266 | ones. It does not matter what the out set of the block is, none of | |
3267 | the info can reach the top. The only thing that reaches the top is | |
3268 | the gen set and we just copy the set. | |
3269 | ||
3270 | 2) There is a kill set but no out set and bb has successors. In | |
3271 | this case we just return. Eventually an out set will be created and | |
3272 | it is better to wait than to create a set of ones. | |
3273 | ||
3274 | 3) There is both a kill and out set. We apply the obvious transfer | |
3275 | function. | |
3276 | */ | |
3277 | ||
3278 | static bool | |
3279 | dse_transfer_function (int bb_index) | |
3280 | { | |
3281 | bb_info_t bb_info = bb_table[bb_index]; | |
3282 | ||
3283 | if (bb_info->kill) | |
3284 | { | |
3285 | if (bb_info->out) | |
3286 | { | |
3287 | /* Case 3 above. */ | |
3288 | if (bb_info->in) | |
b8698a0f | 3289 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3290 | bb_info->out, bb_info->kill); |
3291 | else | |
3292 | { | |
3f9b14ff | 3293 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
b8698a0f | 3294 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3295 | bb_info->out, bb_info->kill); |
3296 | return true; | |
3297 | } | |
3298 | } | |
3299 | else | |
3300 | /* Case 2 above. */ | |
3301 | return false; | |
3302 | } | |
3303 | else | |
3304 | { | |
3305 | /* Case 1 above. If there is already an in set, nothing | |
3306 | happens. */ | |
3307 | if (bb_info->in) | |
3308 | return false; | |
3309 | else | |
3310 | { | |
3f9b14ff | 3311 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3312 | bitmap_copy (bb_info->in, bb_info->gen); |
3313 | return true; | |
3314 | } | |
3315 | } | |
3316 | } | |
3317 | ||
3318 | /* Solve the dataflow equations. */ | |
3319 | ||
3320 | static void | |
3321 | dse_step4 (void) | |
3322 | { | |
b8698a0f L |
3323 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3324 | dse_confluence_n, dse_transfer_function, | |
3325 | all_blocks, df_get_postorder (DF_BACKWARD), | |
6fb5fa3c | 3326 | df_get_n_blocks (DF_BACKWARD)); |
456610d3 | 3327 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3328 | { |
3329 | basic_block bb; | |
3330 | ||
3331 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
04a90bec | 3332 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3333 | { |
3334 | bb_info_t bb_info = bb_table[bb->index]; | |
3335 | ||
3336 | df_print_bb_index (bb, dump_file); | |
3337 | if (bb_info->in) | |
3338 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3339 | else | |
3340 | fprintf (dump_file, " in: *MISSING*\n"); | |
3341 | if (bb_info->gen) | |
3342 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3343 | else | |
3344 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3345 | if (bb_info->kill) | |
3346 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3347 | else | |
3348 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3349 | if (bb_info->out) | |
3350 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3351 | else | |
3352 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3353 | } | |
3354 | } | |
3355 | } | |
3356 | ||
3357 | ||
3358 | \f | |
3359 | /*---------------------------------------------------------------------------- | |
3360 | Fifth step. | |
3361 | ||
0d52bcc1 | 3362 | Delete the stores that can only be deleted using the global information. |
6fb5fa3c DB |
3363 | ----------------------------------------------------------------------------*/ |
3364 | ||
3365 | ||
3366 | static void | |
ac6929b5 | 3367 | dse_step5 (void) |
6fb5fa3c DB |
3368 | { |
3369 | basic_block bb; | |
11cd3bed | 3370 | FOR_EACH_BB_FN (bb, cfun) |
6fb5fa3c DB |
3371 | { |
3372 | bb_info_t bb_info = bb_table[bb->index]; | |
3373 | insn_info_t insn_info = bb_info->last_insn; | |
3374 | bitmap v = bb_info->out; | |
3375 | ||
3376 | while (insn_info) | |
3377 | { | |
3378 | bool deleted = false; | |
3379 | if (dump_file && insn_info->insn) | |
3380 | { | |
3381 | fprintf (dump_file, "starting to process insn %d\n", | |
3382 | INSN_UID (insn_info->insn)); | |
3383 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3384 | } | |
3385 | ||
3386 | /* There may have been code deleted by the dce pass run before | |
3387 | this phase. */ | |
b8698a0f | 3388 | if (insn_info->insn |
6fb5fa3c DB |
3389 | && INSN_P (insn_info->insn) |
3390 | && (!insn_info->cannot_delete) | |
3391 | && (!bitmap_empty_p (v))) | |
3392 | { | |
fd6daec9 | 3393 | store_info *store_info = insn_info->store_rec; |
6fb5fa3c DB |
3394 | |
3395 | /* Try to delete the current insn. */ | |
3396 | deleted = true; | |
b8698a0f | 3397 | |
6fb5fa3c DB |
3398 | /* Skip the clobbers. */ |
3399 | while (!store_info->is_set) | |
3400 | store_info = store_info->next; | |
3401 | ||
02ce5d90 | 3402 | HOST_WIDE_INT i, offset, width; |
ac6929b5 RB |
3403 | group_info *group_info = rtx_group_vec[store_info->group_id]; |
3404 | ||
02ce5d90 RS |
3405 | if (!store_info->offset.is_constant (&offset) |
3406 | || !store_info->width.is_constant (&width)) | |
3407 | deleted = false; | |
3408 | else | |
6fb5fa3c | 3409 | { |
02ce5d90 RS |
3410 | HOST_WIDE_INT end = offset + width; |
3411 | for (i = offset; i < end; i++) | |
6fb5fa3c | 3412 | { |
02ce5d90 RS |
3413 | int index = get_bitmap_index (group_info, i); |
3414 | ||
456610d3 | 3415 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
3416 | fprintf (dump_file, "i = %d, index = %d\n", |
3417 | (int) i, index); | |
3418 | if (index == 0 || !bitmap_bit_p (v, index)) | |
3419 | { | |
3420 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3421 | fprintf (dump_file, "failing at i = %d\n", | |
3422 | (int) i); | |
3423 | deleted = false; | |
3424 | break; | |
3425 | } | |
6fb5fa3c DB |
3426 | } |
3427 | } | |
3428 | if (deleted) | |
3429 | { | |
9e582b1d JR |
3430 | if (dbg_cnt (dse) |
3431 | && check_for_inc_dec_1 (insn_info)) | |
6fb5fa3c | 3432 | { |
6fb5fa3c DB |
3433 | delete_insn (insn_info->insn); |
3434 | insn_info->insn = NULL; | |
3435 | globally_deleted++; | |
3436 | } | |
3437 | } | |
3438 | } | |
3439 | /* We do want to process the local info if the insn was | |
6ed3da00 | 3440 | deleted. For instance, if the insn did a wild read, we |
6fb5fa3c | 3441 | no longer need to trash the info. */ |
b8698a0f | 3442 | if (insn_info->insn |
6fb5fa3c DB |
3443 | && INSN_P (insn_info->insn) |
3444 | && (!deleted)) | |
3445 | { | |
ac6929b5 | 3446 | scan_stores (insn_info->store_rec, v, NULL); |
6fb5fa3c DB |
3447 | if (insn_info->wild_read) |
3448 | { | |
456610d3 | 3449 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3450 | fprintf (dump_file, "wild read\n"); |
3451 | bitmap_clear (v); | |
3452 | } | |
d26c7090 | 3453 | else if (insn_info->read_rec |
a8404b9a JJ |
3454 | || insn_info->non_frame_wild_read |
3455 | || insn_info->frame_read) | |
6fb5fa3c | 3456 | { |
a8404b9a JJ |
3457 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3458 | { | |
3459 | if (!insn_info->non_frame_wild_read | |
3460 | && !insn_info->frame_read) | |
3461 | fprintf (dump_file, "regular read\n"); | |
3462 | if (insn_info->non_frame_wild_read) | |
3463 | fprintf (dump_file, "non-frame wild read\n"); | |
3464 | if (insn_info->frame_read) | |
3465 | fprintf (dump_file, "frame read\n"); | |
3466 | } | |
ac6929b5 | 3467 | scan_reads (insn_info, v, NULL); |
6fb5fa3c DB |
3468 | } |
3469 | } | |
b8698a0f | 3470 | |
6fb5fa3c DB |
3471 | insn_info = insn_info->prev_insn; |
3472 | } | |
3473 | } | |
3474 | } | |
3475 | ||
3476 | ||
6fb5fa3c DB |
3477 | \f |
3478 | /*---------------------------------------------------------------------------- | |
3479 | Sixth step. | |
3480 | ||
8dd5516b JJ |
3481 | Delete stores made redundant by earlier stores (which store the same |
3482 | value) that couldn't be eliminated. | |
3483 | ----------------------------------------------------------------------------*/ | |
3484 | ||
3485 | static void | |
3486 | dse_step6 (void) | |
3487 | { | |
3488 | basic_block bb; | |
3489 | ||
04a90bec | 3490 | FOR_ALL_BB_FN (bb, cfun) |
8dd5516b JJ |
3491 | { |
3492 | bb_info_t bb_info = bb_table[bb->index]; | |
3493 | insn_info_t insn_info = bb_info->last_insn; | |
3494 | ||
3495 | while (insn_info) | |
3496 | { | |
3497 | /* There may have been code deleted by the dce pass run before | |
3498 | this phase. */ | |
3499 | if (insn_info->insn | |
3500 | && INSN_P (insn_info->insn) | |
3501 | && !insn_info->cannot_delete) | |
3502 | { | |
fd6daec9 | 3503 | store_info *s_info = insn_info->store_rec; |
8dd5516b JJ |
3504 | |
3505 | while (s_info && !s_info->is_set) | |
3506 | s_info = s_info->next; | |
3507 | if (s_info | |
3508 | && s_info->redundant_reason | |
3509 | && s_info->redundant_reason->insn | |
3510 | && INSN_P (s_info->redundant_reason->insn)) | |
3511 | { | |
eb92d49a | 3512 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
456610d3 | 3513 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
3514 | fprintf (dump_file, "Locally deleting insn %d " |
3515 | "because insn %d stores the " | |
3516 | "same value and couldn't be " | |
3517 | "eliminated\n", | |
3518 | INSN_UID (insn_info->insn), | |
3519 | INSN_UID (rinsn)); | |
3520 | delete_dead_store_insn (insn_info); | |
3521 | } | |
3522 | } | |
3523 | insn_info = insn_info->prev_insn; | |
3524 | } | |
3525 | } | |
3526 | } | |
3527 | \f | |
3528 | /*---------------------------------------------------------------------------- | |
3529 | Seventh step. | |
3530 | ||
b8698a0f | 3531 | Destroy everything left standing. |
6fb5fa3c DB |
3532 | ----------------------------------------------------------------------------*/ |
3533 | ||
b8698a0f | 3534 | static void |
3f9b14ff | 3535 | dse_step7 (void) |
6fb5fa3c | 3536 | { |
3f9b14ff SB |
3537 | bitmap_obstack_release (&dse_bitmap_obstack); |
3538 | obstack_free (&dse_obstack, NULL); | |
370f38e8 | 3539 | |
6fb5fa3c DB |
3540 | end_alias_analysis (); |
3541 | free (bb_table); | |
c203e8a7 TS |
3542 | delete rtx_group_table; |
3543 | rtx_group_table = NULL; | |
9771b263 | 3544 | rtx_group_vec.release (); |
6fb5fa3c DB |
3545 | BITMAP_FREE (all_blocks); |
3546 | BITMAP_FREE (scratch); | |
3547 | ||
da6603c6 | 3548 | rtx_store_info_pool.release (); |
fb0b2914 ML |
3549 | read_info_type_pool.release (); |
3550 | insn_info_type_pool.release (); | |
3551 | dse_bb_info_type_pool.release (); | |
3552 | group_info_pool.release (); | |
3553 | deferred_change_pool.release (); | |
6fb5fa3c DB |
3554 | } |
3555 | ||
3556 | ||
6fb5fa3c DB |
3557 | /* ------------------------------------------------------------------------- |
3558 | DSE | |
3559 | ------------------------------------------------------------------------- */ | |
3560 | ||
3561 | /* Callback for running pass_rtl_dse. */ | |
3562 | ||
3563 | static unsigned int | |
3564 | rest_of_handle_dse (void) | |
3565 | { | |
6fb5fa3c DB |
3566 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3567 | ||
02b47899 KZ |
3568 | /* Need the notes since we must track live hardregs in the forwards |
3569 | direction. */ | |
3570 | df_note_add_problem (); | |
3571 | df_analyze (); | |
3572 | ||
6fb5fa3c DB |
3573 | dse_step0 (); |
3574 | dse_step1 (); | |
3575 | dse_step2_init (); | |
ac6929b5 | 3576 | if (dse_step2 ()) |
6fb5fa3c DB |
3577 | { |
3578 | df_set_flags (DF_LR_RUN_DCE); | |
3579 | df_analyze (); | |
456610d3 | 3580 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 3581 | fprintf (dump_file, "doing global processing\n"); |
ac6929b5 | 3582 | dse_step3 (); |
6fb5fa3c | 3583 | dse_step4 (); |
ac6929b5 | 3584 | dse_step5 (); |
6fb5fa3c DB |
3585 | } |
3586 | ||
8dd5516b | 3587 | dse_step6 (); |
3f9b14ff | 3588 | dse_step7 (); |
6fb5fa3c DB |
3589 | |
3590 | if (dump_file) | |
ac6929b5 RB |
3591 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d\n", |
3592 | locally_deleted, globally_deleted); | |
6c5ed3f1 JJ |
3593 | |
3594 | /* DSE can eliminate potentially-trapping MEMs. | |
3595 | Remove any EH edges associated with them. */ | |
3596 | if ((locally_deleted || globally_deleted) | |
3597 | && cfun->can_throw_non_call_exceptions | |
3598 | && purge_all_dead_edges ()) | |
3599 | cleanup_cfg (0); | |
3600 | ||
6fb5fa3c DB |
3601 | return 0; |
3602 | } | |
3603 | ||
17795822 TS |
3604 | namespace { |
3605 | ||
3606 | const pass_data pass_data_rtl_dse1 = | |
27a4cd48 DM |
3607 | { |
3608 | RTL_PASS, /* type */ | |
3609 | "dse1", /* name */ | |
3610 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3611 | TV_DSE1, /* tv_id */ |
3612 | 0, /* properties_required */ | |
3613 | 0, /* properties_provided */ | |
3614 | 0, /* properties_destroyed */ | |
3615 | 0, /* todo_flags_start */ | |
3bea341f | 3616 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c DB |
3617 | }; |
3618 | ||
17795822 | 3619 | class pass_rtl_dse1 : public rtl_opt_pass |
27a4cd48 DM |
3620 | { |
3621 | public: | |
c3284718 RS |
3622 | pass_rtl_dse1 (gcc::context *ctxt) |
3623 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
27a4cd48 DM |
3624 | {} |
3625 | ||
3626 | /* opt_pass methods: */ | |
1a3d085c TS |
3627 | virtual bool gate (function *) |
3628 | { | |
3629 | return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
3630 | } | |
3631 | ||
be55bfe6 | 3632 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3633 | |
3634 | }; // class pass_rtl_dse1 | |
3635 | ||
17795822 TS |
3636 | } // anon namespace |
3637 | ||
27a4cd48 DM |
3638 | rtl_opt_pass * |
3639 | make_pass_rtl_dse1 (gcc::context *ctxt) | |
3640 | { | |
3641 | return new pass_rtl_dse1 (ctxt); | |
3642 | } | |
3643 | ||
17795822 TS |
3644 | namespace { |
3645 | ||
3646 | const pass_data pass_data_rtl_dse2 = | |
27a4cd48 DM |
3647 | { |
3648 | RTL_PASS, /* type */ | |
3649 | "dse2", /* name */ | |
3650 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3651 | TV_DSE2, /* tv_id */ |
3652 | 0, /* properties_required */ | |
3653 | 0, /* properties_provided */ | |
3654 | 0, /* properties_destroyed */ | |
3655 | 0, /* todo_flags_start */ | |
3bea341f | 3656 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c | 3657 | }; |
27a4cd48 | 3658 | |
17795822 | 3659 | class pass_rtl_dse2 : public rtl_opt_pass |
27a4cd48 DM |
3660 | { |
3661 | public: | |
c3284718 RS |
3662 | pass_rtl_dse2 (gcc::context *ctxt) |
3663 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
27a4cd48 DM |
3664 | {} |
3665 | ||
3666 | /* opt_pass methods: */ | |
1a3d085c TS |
3667 | virtual bool gate (function *) |
3668 | { | |
3669 | return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
3670 | } | |
3671 | ||
be55bfe6 | 3672 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3673 | |
3674 | }; // class pass_rtl_dse2 | |
3675 | ||
17795822 TS |
3676 | } // anon namespace |
3677 | ||
27a4cd48 DM |
3678 | rtl_opt_pass * |
3679 | make_pass_rtl_dse2 (gcc::context *ctxt) | |
3680 | { | |
3681 | return new pass_rtl_dse2 (ctxt); | |
3682 | } |