]>
Commit | Line | Data |
---|---|---|
6fb5fa3c | 1 | /* RTL dead store elimination. |
8d9254fc | 2 | Copyright (C) 2005-2020 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" |
d7111da8 | 50 | #include "rtl-iter.h" |
6c5ed3f1 | 51 | #include "cfgcleanup.h" |
6783fdb7 | 52 | #include "calls.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. */ | |
6c1dae73 | 223 | class store_info |
6fb5fa3c | 224 | { |
6c1dae73 | 225 | public: |
6fb5fa3c DB |
226 | |
227 | /* False means this is a clobber. */ | |
228 | bool is_set; | |
229 | ||
8dd5516b JJ |
230 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
231 | bool is_large; | |
232 | ||
6fb5fa3c DB |
233 | /* The id of the mem group of the base address. If rtx_varies_p is |
234 | true, this is -1. Otherwise, it is the index into the group | |
235 | table. */ | |
236 | int group_id; | |
b8698a0f | 237 | |
6fb5fa3c DB |
238 | /* This is the cselib value. */ |
239 | cselib_val *cse_base; | |
240 | ||
241 | /* This canonized mem. */ | |
242 | rtx mem; | |
243 | ||
6216f94e | 244 | /* Canonized MEM address for use by canon_true_dependence. */ |
6fb5fa3c DB |
245 | rtx mem_addr; |
246 | ||
71d6a386 | 247 | /* The offset of the first byte associated with the operation. */ |
02ce5d90 | 248 | poly_int64 offset; |
71d6a386 RS |
249 | |
250 | /* The number of bytes covered by the operation. This is always exact | |
251 | and known (rather than -1). */ | |
02ce5d90 | 252 | poly_int64 width; |
8dd5516b JJ |
253 | |
254 | union | |
255 | { | |
256 | /* A bitmask as wide as the number of bytes in the word that | |
257 | contains a 1 if the byte may be needed. The store is unused if | |
258 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
259 | unsigned HOST_WIDE_INT small_bitmask; | |
260 | ||
261 | struct | |
262 | { | |
02ce5d90 RS |
263 | /* A bitmap with one bit per byte, or null if the number of |
264 | bytes isn't known at compile time. A cleared bit means | |
265 | the position is needed. Used if IS_LARGE is true. */ | |
dc491a25 | 266 | bitmap bmap; |
6fb5fa3c | 267 | |
02ce5d90 RS |
268 | /* When BITMAP is nonnull, this counts the number of set bits |
269 | (i.e. unneeded bytes) in the bitmap. If it is equal to | |
270 | WIDTH, the whole store is unused. | |
271 | ||
272 | When BITMAP is null: | |
273 | - the store is definitely not needed when COUNT == 1 | |
274 | - all the store is needed when COUNT == 0 and RHS is nonnull | |
275 | - otherwise we don't know which parts of the store are needed. */ | |
8dd5516b JJ |
276 | int count; |
277 | } large; | |
278 | } positions_needed; | |
6fb5fa3c DB |
279 | |
280 | /* The next store info for this insn. */ | |
99b1c316 | 281 | class store_info *next; |
6fb5fa3c DB |
282 | |
283 | /* The right hand side of the store. This is used if there is a | |
284 | subsequent reload of the mems address somewhere later in the | |
285 | basic block. */ | |
8dd5516b JJ |
286 | rtx rhs; |
287 | ||
288 | /* If rhs is or holds a constant, this contains that constant, | |
289 | otherwise NULL. */ | |
290 | rtx const_rhs; | |
291 | ||
292 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
293 | insn stored. These aren't eliminated early, because doing that | |
294 | might prevent the earlier larger store to be eliminated. */ | |
da6603c6 | 295 | struct insn_info_type *redundant_reason; |
6fb5fa3c DB |
296 | }; |
297 | ||
4fe663b0 L |
298 | /* Return a bitmask with the first N low bits set. */ |
299 | ||
300 | static unsigned HOST_WIDE_INT | |
301 | lowpart_bitmask (int n) | |
302 | { | |
dd4786fe | 303 | unsigned HOST_WIDE_INT mask = HOST_WIDE_INT_M1U; |
4fe663b0 L |
304 | return mask >> (HOST_BITS_PER_WIDE_INT - n); |
305 | } | |
306 | ||
fcb87c50 | 307 | static object_allocator<store_info> cse_store_info_pool ("cse_store_info_pool"); |
da6603c6 | 308 | |
fcb87c50 | 309 | static object_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool"); |
6fb5fa3c DB |
310 | |
311 | /* This structure holds information about a load. These are only | |
312 | built for rtx bases. */ | |
6c1dae73 | 313 | class read_info_type |
6fb5fa3c | 314 | { |
6c1dae73 | 315 | public: |
6fb5fa3c DB |
316 | /* The id of the mem group of the base address. */ |
317 | int group_id; | |
318 | ||
71d6a386 | 319 | /* The offset of the first byte associated with the operation. */ |
02ce5d90 | 320 | poly_int64 offset; |
71d6a386 RS |
321 | |
322 | /* The number of bytes covered by the operation, or -1 if not known. */ | |
02ce5d90 | 323 | poly_int64 width; |
6fb5fa3c DB |
324 | |
325 | /* The mem being read. */ | |
326 | rtx mem; | |
327 | ||
328 | /* The next read_info for this insn. */ | |
99b1c316 | 329 | class read_info_type *next; |
6fb5fa3c | 330 | }; |
99b1c316 | 331 | typedef class read_info_type *read_info_t; |
6fb5fa3c | 332 | |
fcb87c50 | 333 | static object_allocator<read_info_type> read_info_type_pool ("read_info_pool"); |
6fb5fa3c DB |
334 | |
335 | /* One of these records is created for each insn. */ | |
336 | ||
da6603c6 | 337 | struct insn_info_type |
6fb5fa3c DB |
338 | { |
339 | /* Set true if the insn contains a store but the insn itself cannot | |
340 | be deleted. This is set if the insn is a parallel and there is | |
341 | more than one non dead output or if the insn is in some way | |
342 | volatile. */ | |
343 | bool cannot_delete; | |
344 | ||
345 | /* This field is only used by the global algorithm. It is set true | |
346 | if the insn contains any read of mem except for a (1). This is | |
347 | also set if the insn is a call or has a clobber mem. If the insn | |
348 | contains a wild read, the use_rec will be null. */ | |
349 | bool wild_read; | |
350 | ||
d26c7090 ER |
351 | /* This is true only for CALL instructions which could potentially read |
352 | any non-frame memory location. This field is used by the global | |
353 | algorithm. */ | |
354 | bool non_frame_wild_read; | |
355 | ||
50f0f366 EB |
356 | /* This field is only used for the processing of const functions. |
357 | These functions cannot read memory, but they can read the stack | |
64520bdc EB |
358 | because that is where they may get their parms. We need to be |
359 | this conservative because, like the store motion pass, we don't | |
360 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
361 | Moreover, we need to distinguish two cases: | |
362 | 1. Before reload (register elimination), the stores related to | |
363 | outgoing arguments are stack pointer based and thus deemed | |
364 | of non-constant base in this pass. This requires special | |
365 | handling but also means that the frame pointer based stores | |
366 | need not be killed upon encountering a const function call. | |
367 | 2. After reload, the stores related to outgoing arguments can be | |
368 | either stack pointer or hard frame pointer based. This means | |
369 | that we have no other choice than also killing all the frame | |
370 | pointer based stores upon encountering a const function call. | |
57534689 JJ |
371 | This field is set after reload for const function calls and before |
372 | reload for const tail function calls on targets where arg pointer | |
373 | is the frame pointer. Having this set is less severe than a wild | |
374 | read, it just means that all the frame related stores are killed | |
375 | rather than all the stores. */ | |
64520bdc EB |
376 | bool frame_read; |
377 | ||
378 | /* This field is only used for the processing of const functions. | |
379 | It is set if the insn may contain a stack pointer based store. */ | |
50f0f366 | 380 | bool stack_pointer_based; |
6fb5fa3c DB |
381 | |
382 | /* This is true if any of the sets within the store contains a | |
383 | cselib base. Such stores can only be deleted by the local | |
384 | algorithm. */ | |
385 | bool contains_cselib_groups; | |
386 | ||
387 | /* The insn. */ | |
dd60a84c | 388 | rtx_insn *insn; |
6fb5fa3c DB |
389 | |
390 | /* The list of mem sets or mem clobbers that are contained in this | |
391 | insn. If the insn is deletable, it contains only one mem set. | |
392 | But it could also contain clobbers. Insns that contain more than | |
393 | one mem set are not deletable, but each of those mems are here in | |
6ed3da00 | 394 | order to provide info to delete other insns. */ |
fd6daec9 | 395 | store_info *store_rec; |
6fb5fa3c DB |
396 | |
397 | /* The linked list of mem uses in this insn. Only the reads from | |
398 | rtx bases are listed here. The reads to cselib bases are | |
399 | completely processed during the first scan and so are never | |
400 | created. */ | |
401 | read_info_t read_rec; | |
402 | ||
9e582b1d JR |
403 | /* The live fixed registers. We assume only fixed registers can |
404 | cause trouble by being clobbered from an expanded pattern; | |
405 | storing only the live fixed registers (rather than all registers) | |
406 | means less memory needs to be allocated / copied for the individual | |
407 | stores. */ | |
408 | regset fixed_regs_live; | |
409 | ||
6fb5fa3c | 410 | /* The prev insn in the basic block. */ |
da6603c6 | 411 | struct insn_info_type * prev_insn; |
6fb5fa3c DB |
412 | |
413 | /* The linked list of insns that are in consideration for removal in | |
073a8998 | 414 | the forwards pass through the basic block. This pointer may be |
6fb5fa3c | 415 | trash as it is not cleared when a wild read occurs. The only |
fa10beec | 416 | time it is guaranteed to be correct is when the traversal starts |
6fb5fa3c | 417 | at active_local_stores. */ |
da6603c6 | 418 | struct insn_info_type * next_local_store; |
6fb5fa3c | 419 | }; |
da6603c6 | 420 | typedef struct insn_info_type *insn_info_t; |
6fb5fa3c | 421 | |
fcb87c50 | 422 | static object_allocator<insn_info_type> insn_info_type_pool ("insn_info_pool"); |
6fb5fa3c DB |
423 | |
424 | /* The linked list of stores that are under consideration in this | |
b8698a0f | 425 | basic block. */ |
6fb5fa3c | 426 | static insn_info_t active_local_stores; |
dabd47e7 | 427 | static int active_local_stores_len; |
6fb5fa3c | 428 | |
da6603c6 | 429 | struct dse_bb_info_type |
6fb5fa3c | 430 | { |
6fb5fa3c DB |
431 | /* Pointer to the insn info for the last insn in the block. These |
432 | are linked so this is how all of the insns are reached. During | |
433 | scanning this is the current insn being scanned. */ | |
434 | insn_info_t last_insn; | |
435 | ||
436 | /* The info for the global dataflow problem. */ | |
437 | ||
438 | ||
439 | /* This is set if the transfer function should and in the wild_read | |
440 | bitmap before applying the kill and gen sets. That vector knocks | |
441 | out most of the bits in the bitmap and thus speeds up the | |
442 | operations. */ | |
443 | bool apply_wild_read; | |
444 | ||
02b47899 KZ |
445 | /* The following 4 bitvectors hold information about which positions |
446 | of which stores are live or dead. They are indexed by | |
447 | get_bitmap_index. */ | |
448 | ||
6fb5fa3c DB |
449 | /* The set of store positions that exist in this block before a wild read. */ |
450 | bitmap gen; | |
b8698a0f | 451 | |
6fb5fa3c DB |
452 | /* The set of load positions that exist in this block above the |
453 | same position of a store. */ | |
454 | bitmap kill; | |
455 | ||
456 | /* The set of stores that reach the top of the block without being | |
457 | killed by a read. | |
458 | ||
459 | Do not represent the in if it is all ones. Note that this is | |
460 | what the bitvector should logically be initialized to for a set | |
461 | intersection problem. However, like the kill set, this is too | |
462 | expensive. So initially, the in set will only be created for the | |
463 | exit block and any block that contains a wild read. */ | |
464 | bitmap in; | |
465 | ||
466 | /* The set of stores that reach the bottom of the block from it's | |
467 | successors. | |
468 | ||
469 | Do not represent the in if it is all ones. Note that this is | |
470 | what the bitvector should logically be initialized to for a set | |
471 | intersection problem. However, like the kill and in set, this is | |
472 | too expensive. So what is done is that the confluence operator | |
473 | just initializes the vector from one of the out sets of the | |
474 | successors of the block. */ | |
475 | bitmap out; | |
02b47899 KZ |
476 | |
477 | /* The following bitvector is indexed by the reg number. It | |
478 | contains the set of regs that are live at the current instruction | |
479 | being processed. While it contains info for all of the | |
9e582b1d JR |
480 | registers, only the hard registers are actually examined. It is used |
481 | to assure that shift and/or add sequences that are inserted do not | |
073a8998 | 482 | accidentally clobber live hard regs. */ |
02b47899 | 483 | bitmap regs_live; |
6fb5fa3c DB |
484 | }; |
485 | ||
da6603c6 | 486 | typedef struct dse_bb_info_type *bb_info_t; |
fb0b2914 ML |
487 | |
488 | static object_allocator<dse_bb_info_type> dse_bb_info_type_pool | |
fcb87c50 | 489 | ("bb_info_pool"); |
6fb5fa3c DB |
490 | |
491 | /* Table to hold all bb_infos. */ | |
492 | static bb_info_t *bb_table; | |
493 | ||
494 | /* There is a group_info for each rtx base that is used to reference | |
495 | memory. There are also not many of the rtx bases because they are | |
496 | very limited in scope. */ | |
497 | ||
b8698a0f | 498 | struct group_info |
6fb5fa3c DB |
499 | { |
500 | /* The actual base of the address. */ | |
501 | rtx rtx_base; | |
502 | ||
503 | /* The sequential id of the base. This allows us to have a | |
504 | canonical ordering of these that is not based on addresses. */ | |
505 | int id; | |
506 | ||
8f5929e1 JJ |
507 | /* True if there are any positions that are to be processed |
508 | globally. */ | |
509 | bool process_globally; | |
510 | ||
511 | /* True if the base of this group is either the frame_pointer or | |
512 | hard_frame_pointer. */ | |
513 | bool frame_related; | |
514 | ||
d32f725a OH |
515 | /* A mem wrapped around the base pointer for the group in order to do |
516 | read dependency. It must be given BLKmode in order to encompass all | |
517 | the possible offsets from the base. */ | |
6fb5fa3c | 518 | rtx base_mem; |
b8698a0f | 519 | |
6216f94e JJ |
520 | /* Canonized version of base_mem's address. */ |
521 | rtx canon_base_addr; | |
6fb5fa3c DB |
522 | |
523 | /* These two sets of two bitmaps are used to keep track of how many | |
6ed3da00 | 524 | stores are actually referencing that position from this base. We |
6fb5fa3c | 525 | only do this for rtx bases as this will be used to assign |
6ed3da00 | 526 | positions in the bitmaps for the global problem. Bit N is set in |
6fb5fa3c DB |
527 | store1 on the first store for offset N. Bit N is set in store2 |
528 | for the second store to offset N. This is all we need since we | |
529 | only care about offsets that have two or more stores for them. | |
530 | ||
531 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
532 | for 0 and greater offsets. | |
533 | ||
534 | There is one special case here, for stores into the stack frame, | |
535 | we will or store1 into store2 before deciding which stores look | |
536 | at globally. This is because stores to the stack frame that have | |
537 | no other reads before the end of the function can also be | |
538 | deleted. */ | |
539 | bitmap store1_n, store1_p, store2_n, store2_p; | |
540 | ||
d26c7090 ER |
541 | /* These bitmaps keep track of offsets in this group escape this function. |
542 | An offset escapes if it corresponds to a named variable whose | |
543 | addressable flag is set. */ | |
544 | bitmap escaped_n, escaped_p; | |
545 | ||
6ed3da00 | 546 | /* The positions in this bitmap have the same assignments as the in, |
6fb5fa3c | 547 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6ed3da00 | 548 | the positions that are occupied by stores for this group. */ |
6fb5fa3c DB |
549 | bitmap group_kill; |
550 | ||
6fb5fa3c | 551 | /* The offset_map is used to map the offsets from this base into |
6ed3da00 | 552 | positions in the global bitmaps. It is only created after all of |
6fb5fa3c DB |
553 | the all of stores have been scanned and we know which ones we |
554 | care about. */ | |
b8698a0f L |
555 | int *offset_map_n, *offset_map_p; |
556 | int offset_map_size_n, offset_map_size_p; | |
6fb5fa3c | 557 | }; |
da6603c6 | 558 | |
fcb87c50 | 559 | static object_allocator<group_info> group_info_pool ("rtx_group_info_pool"); |
6fb5fa3c | 560 | |
6fb5fa3c DB |
561 | /* Index into the rtx_group_vec. */ |
562 | static int rtx_group_next_id; | |
563 | ||
6fb5fa3c | 564 | |
fd6daec9 | 565 | static vec<group_info *> rtx_group_vec; |
6fb5fa3c DB |
566 | |
567 | ||
568 | /* This structure holds the set of changes that are being deferred | |
569 | when removing read operation. See replace_read. */ | |
b8698a0f | 570 | struct deferred_change |
6fb5fa3c DB |
571 | { |
572 | ||
573 | /* The mem that is being replaced. */ | |
574 | rtx *loc; | |
575 | ||
576 | /* The reg it is being replaced with. */ | |
577 | rtx reg; | |
578 | ||
579 | struct deferred_change *next; | |
580 | }; | |
581 | ||
fb0b2914 | 582 | static object_allocator<deferred_change> deferred_change_pool |
fcb87c50 | 583 | ("deferred_change_pool"); |
6fb5fa3c | 584 | |
fd6daec9 | 585 | static deferred_change *deferred_change_list = NULL; |
6fb5fa3c | 586 | |
e3b5732b | 587 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
9dd9bf80 | 588 | this for vararg functions because they play games with the frame. */ |
6fb5fa3c DB |
589 | static bool stores_off_frame_dead_at_return; |
590 | ||
591 | /* Counter for stats. */ | |
b8698a0f L |
592 | static int globally_deleted; |
593 | static int locally_deleted; | |
b8698a0f | 594 | |
6fb5fa3c DB |
595 | static bitmap all_blocks; |
596 | ||
d26c7090 ER |
597 | /* Locations that are killed by calls in the global phase. */ |
598 | static bitmap kill_on_calls; | |
599 | ||
6fb5fa3c DB |
600 | /* The number of bits used in the global bitmaps. */ |
601 | static unsigned int current_position; | |
71d6a386 RS |
602 | |
603 | /* Print offset range [OFFSET, OFFSET + WIDTH) to FILE. */ | |
604 | ||
605 | static void | |
606 | print_range (FILE *file, poly_int64 offset, poly_int64 width) | |
607 | { | |
608 | fprintf (file, "["); | |
609 | print_dec (offset, file, SIGNED); | |
610 | fprintf (file, ".."); | |
611 | print_dec (offset + width, file, SIGNED); | |
612 | fprintf (file, ")"); | |
613 | } | |
6fb5fa3c DB |
614 | \f |
615 | /*---------------------------------------------------------------------------- | |
616 | Zeroth step. | |
617 | ||
b8698a0f | 618 | Initialization. |
6fb5fa3c DB |
619 | ----------------------------------------------------------------------------*/ |
620 | ||
6fb5fa3c | 621 | |
6fb5fa3c DB |
622 | /* Hashtable callbacks for maintaining the "bases" field of |
623 | store_group_info, given that the addresses are function invariants. */ | |
624 | ||
8d67ee55 | 625 | struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
703c8606 | 626 | { |
67f58944 TS |
627 | static inline hashval_t hash (const group_info *); |
628 | static inline bool equal (const group_info *, const group_info *); | |
703c8606 LC |
629 | }; |
630 | ||
631 | inline bool | |
67f58944 TS |
632 | invariant_group_base_hasher::equal (const group_info *gi1, |
633 | const group_info *gi2) | |
6fb5fa3c | 634 | { |
6fb5fa3c DB |
635 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
636 | } | |
637 | ||
703c8606 | 638 | inline hashval_t |
67f58944 | 639 | invariant_group_base_hasher::hash (const group_info *gi) |
6fb5fa3c | 640 | { |
6fb5fa3c DB |
641 | int do_not_record; |
642 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
643 | } | |
644 | ||
703c8606 | 645 | /* Tables of group_info structures, hashed by base value. */ |
c203e8a7 | 646 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
703c8606 | 647 | |
6fb5fa3c DB |
648 | |
649 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
650 | ||
fd6daec9 | 651 | static group_info * |
6fb5fa3c DB |
652 | get_group_info (rtx base) |
653 | { | |
b8698a0f | 654 | struct group_info tmp_gi; |
fd6daec9 | 655 | group_info *gi; |
703c8606 | 656 | group_info **slot; |
6fb5fa3c | 657 | |
ac6929b5 RB |
658 | gcc_assert (base != NULL_RTX); |
659 | ||
660 | /* Find the store_base_info structure for BASE, creating a new one | |
661 | if necessary. */ | |
662 | tmp_gi.rtx_base = base; | |
663 | slot = rtx_group_table->find_slot (&tmp_gi, INSERT); | |
664 | gi = *slot; | |
6fb5fa3c DB |
665 | |
666 | if (gi == NULL) | |
667 | { | |
fb0b2914 | 668 | *slot = gi = group_info_pool.allocate (); |
6fb5fa3c DB |
669 | gi->rtx_base = base; |
670 | gi->id = rtx_group_next_id++; | |
d32f725a | 671 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
6216f94e | 672 | gi->canon_base_addr = canon_rtx (base); |
3f9b14ff SB |
673 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
674 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
675 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
676 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
677 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
678 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
679 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 680 | gi->process_globally = false; |
b8698a0f | 681 | gi->frame_related = |
6fb5fa3c DB |
682 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
683 | gi->offset_map_size_n = 0; | |
684 | gi->offset_map_size_p = 0; | |
685 | gi->offset_map_n = NULL; | |
686 | gi->offset_map_p = NULL; | |
9771b263 | 687 | rtx_group_vec.safe_push (gi); |
6fb5fa3c DB |
688 | } |
689 | ||
690 | return gi; | |
691 | } | |
692 | ||
693 | ||
694 | /* Initialization of data structures. */ | |
695 | ||
696 | static void | |
697 | dse_step0 (void) | |
698 | { | |
699 | locally_deleted = 0; | |
700 | globally_deleted = 0; | |
6fb5fa3c | 701 | |
3f9b14ff SB |
702 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
703 | gcc_obstack_init (&dse_obstack); | |
704 | ||
705 | scratch = BITMAP_ALLOC (®_obstack); | |
706 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 707 | |
6fb5fa3c | 708 | |
c203e8a7 | 709 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
6fb5fa3c | 710 | |
8b1c6fd7 | 711 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
6fb5fa3c DB |
712 | rtx_group_next_id = 0; |
713 | ||
e3b5732b | 714 | stores_off_frame_dead_at_return = !cfun->stdarg; |
6fb5fa3c DB |
715 | |
716 | init_alias_analysis (); | |
6fb5fa3c DB |
717 | } |
718 | ||
719 | ||
720 | \f | |
721 | /*---------------------------------------------------------------------------- | |
722 | First step. | |
723 | ||
724 | Scan all of the insns. Any random ordering of the blocks is fine. | |
0d52bcc1 | 725 | Each block is scanned in forward order to accommodate cselib which |
6fb5fa3c DB |
726 | is used to remove stores with non-constant bases. |
727 | ----------------------------------------------------------------------------*/ | |
728 | ||
729 | /* Delete all of the store_info recs from INSN_INFO. */ | |
730 | ||
b8698a0f | 731 | static void |
6fb5fa3c DB |
732 | free_store_info (insn_info_t insn_info) |
733 | { | |
fd6daec9 TS |
734 | store_info *cur = insn_info->store_rec; |
735 | while (cur) | |
6fb5fa3c | 736 | { |
fd6daec9 TS |
737 | store_info *next = cur->next; |
738 | if (cur->is_large) | |
739 | BITMAP_FREE (cur->positions_needed.large.bmap); | |
740 | if (cur->cse_base) | |
741 | cse_store_info_pool.remove (cur); | |
6fb5fa3c | 742 | else |
fd6daec9 TS |
743 | rtx_store_info_pool.remove (cur); |
744 | cur = next; | |
6fb5fa3c DB |
745 | } |
746 | ||
747 | insn_info->cannot_delete = true; | |
748 | insn_info->contains_cselib_groups = false; | |
749 | insn_info->store_rec = NULL; | |
750 | } | |
751 | ||
a79683d5 | 752 | struct note_add_store_info |
9e582b1d | 753 | { |
dc01c3d1 | 754 | rtx_insn *first, *current; |
9e582b1d JR |
755 | regset fixed_regs_live; |
756 | bool failure; | |
a79683d5 | 757 | }; |
9e582b1d JR |
758 | |
759 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
760 | Check if a register is clobbered which is live afterwards. */ | |
761 | ||
762 | static void | |
763 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
764 | { | |
dc01c3d1 | 765 | rtx_insn *insn; |
9e582b1d | 766 | note_add_store_info *info = (note_add_store_info *) data; |
9e582b1d JR |
767 | |
768 | if (!REG_P (loc)) | |
769 | return; | |
770 | ||
771 | /* If this register is referenced by the current or an earlier insn, | |
772 | that's OK. E.g. this applies to the register that is being incremented | |
773 | with this addition. */ | |
774 | for (insn = info->first; | |
775 | insn != NEXT_INSN (info->current); | |
776 | insn = NEXT_INSN (insn)) | |
777 | if (reg_referenced_p (loc, PATTERN (insn))) | |
778 | return; | |
779 | ||
780 | /* If we come here, we have a clobber of a register that's only OK | |
781 | if that register is not live. If we don't have liveness information | |
782 | available, fail now. */ | |
783 | if (!info->fixed_regs_live) | |
784 | { | |
53d1bae9 | 785 | info->failure = true; |
9e582b1d JR |
786 | return; |
787 | } | |
788 | /* Now check if this is a live fixed register. */ | |
53d1bae9 RS |
789 | unsigned int end_regno = END_REGNO (loc); |
790 | for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) | |
791 | if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) | |
792 | info->failure = true; | |
9e582b1d JR |
793 | } |
794 | ||
4deef538 AO |
795 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
796 | SRC + SRCOFF before insn ARG. */ | |
6fb5fa3c DB |
797 | |
798 | static int | |
4deef538 AO |
799 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
800 | rtx op ATTRIBUTE_UNUSED, | |
801 | rtx dest, rtx src, rtx srcoff, void *arg) | |
6fb5fa3c | 802 | { |
9e582b1d | 803 | insn_info_t insn_info = (insn_info_t) arg; |
dc01c3d1 | 804 | rtx_insn *insn = insn_info->insn, *new_insn, *cur; |
9e582b1d | 805 | note_add_store_info info; |
b8698a0f | 806 | |
4deef538 AO |
807 | /* We can reuse all operands without copying, because we are about |
808 | to delete the insn that contained it. */ | |
9e582b1d | 809 | if (srcoff) |
ed079c4b JJ |
810 | { |
811 | start_sequence (); | |
812 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
813 | new_insn = get_insns (); | |
814 | end_sequence (); | |
815 | } | |
9e582b1d | 816 | else |
1476d1bd | 817 | new_insn = gen_move_insn (dest, src); |
9e582b1d JR |
818 | info.first = new_insn; |
819 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
820 | info.failure = false; | |
821 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
822 | { | |
823 | info.current = cur; | |
e8448ba5 | 824 | note_stores (cur, note_add_store, &info); |
9e582b1d | 825 | } |
6fb5fa3c | 826 | |
9e582b1d JR |
827 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
828 | return it immediately, communicating the failure to its caller. */ | |
829 | if (info.failure) | |
830 | return 1; | |
831 | ||
832 | emit_insn_before (new_insn, insn); | |
6fb5fa3c | 833 | |
8d8e205b | 834 | return 0; |
6fb5fa3c DB |
835 | } |
836 | ||
9e582b1d JR |
837 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
838 | is there, is split into a separate insn. | |
839 | Return true on success (or if there was nothing to do), false on failure. */ | |
6fb5fa3c | 840 | |
9e582b1d JR |
841 | static bool |
842 | check_for_inc_dec_1 (insn_info_t insn_info) | |
6fb5fa3c | 843 | { |
dd60a84c | 844 | rtx_insn *insn = insn_info->insn; |
6fb5fa3c DB |
845 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
846 | if (note) | |
8d8e205b RS |
847 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
848 | insn_info) == 0; | |
2c0fa3ec JJ |
849 | |
850 | /* Punt on stack pushes, those don't have REG_INC notes and we are | |
851 | unprepared to deal with distribution of REG_ARGS_SIZE notes etc. */ | |
852 | subrtx_iterator::array_type array; | |
853 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) | |
854 | { | |
855 | const_rtx x = *iter; | |
856 | if (GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) | |
857 | return false; | |
858 | } | |
859 | ||
9e582b1d | 860 | return true; |
6fb5fa3c DB |
861 | } |
862 | ||
863 | ||
9e582b1d JR |
864 | /* Entry point for postreload. If you work on reload_cse, or you need this |
865 | anywhere else, consider if you can provide register liveness information | |
866 | and add a parameter to this function so that it can be passed down in | |
867 | insn_info.fixed_regs_live. */ | |
868 | bool | |
dd60a84c | 869 | check_for_inc_dec (rtx_insn *insn) |
9e582b1d | 870 | { |
da6603c6 | 871 | insn_info_type insn_info; |
9e582b1d JR |
872 | rtx note; |
873 | ||
874 | insn_info.insn = insn; | |
875 | insn_info.fixed_regs_live = NULL; | |
876 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
877 | if (note) | |
8d8e205b RS |
878 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
879 | &insn_info) == 0; | |
2c0fa3ec JJ |
880 | |
881 | /* Punt on stack pushes, those don't have REG_INC notes and we are | |
882 | unprepared to deal with distribution of REG_ARGS_SIZE notes etc. */ | |
883 | subrtx_iterator::array_type array; | |
884 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) | |
885 | { | |
886 | const_rtx x = *iter; | |
887 | if (GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) | |
888 | return false; | |
889 | } | |
890 | ||
9e582b1d JR |
891 | return true; |
892 | } | |
893 | ||
b8698a0f | 894 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
6fb5fa3c DB |
895 | |
896 | static void | |
897 | delete_dead_store_insn (insn_info_t insn_info) | |
898 | { | |
899 | read_info_t read_info; | |
900 | ||
901 | if (!dbg_cnt (dse)) | |
902 | return; | |
903 | ||
9e582b1d JR |
904 | if (!check_for_inc_dec_1 (insn_info)) |
905 | return; | |
456610d3 | 906 | if (dump_file && (dump_flags & TDF_DETAILS)) |
ac6929b5 RB |
907 | fprintf (dump_file, "Locally deleting insn %d\n", |
908 | INSN_UID (insn_info->insn)); | |
6fb5fa3c DB |
909 | |
910 | free_store_info (insn_info); | |
911 | read_info = insn_info->read_rec; | |
b8698a0f | 912 | |
6fb5fa3c DB |
913 | while (read_info) |
914 | { | |
915 | read_info_t next = read_info->next; | |
fb0b2914 | 916 | read_info_type_pool.remove (read_info); |
6fb5fa3c DB |
917 | read_info = next; |
918 | } | |
919 | insn_info->read_rec = NULL; | |
920 | ||
921 | delete_insn (insn_info->insn); | |
922 | locally_deleted++; | |
923 | insn_info->insn = NULL; | |
924 | ||
925 | insn_info->wild_read = false; | |
926 | } | |
927 | ||
88d8330d EB |
928 | /* Return whether DECL, a local variable, can possibly escape the current |
929 | function scope. */ | |
930 | ||
931 | static bool | |
932 | local_variable_can_escape (tree decl) | |
933 | { | |
934 | if (TREE_ADDRESSABLE (decl)) | |
935 | return true; | |
936 | ||
937 | /* If this is a partitioned variable, we need to consider all the variables | |
938 | in the partition. This is necessary because a store into one of them can | |
939 | be replaced with a store into another and this may not change the outcome | |
940 | of the escape analysis. */ | |
941 | if (cfun->gimple_df->decls_to_pointers != NULL) | |
942 | { | |
39c8aaa4 | 943 | tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); |
88d8330d | 944 | if (namep) |
39c8aaa4 | 945 | return TREE_ADDRESSABLE (*namep); |
88d8330d EB |
946 | } |
947 | ||
948 | return false; | |
949 | } | |
950 | ||
951 | /* Return whether EXPR can possibly escape the current function scope. */ | |
952 | ||
d26c7090 ER |
953 | static bool |
954 | can_escape (tree expr) | |
955 | { | |
956 | tree base; | |
957 | if (!expr) | |
958 | return true; | |
959 | base = get_base_address (expr); | |
960 | if (DECL_P (base) | |
88d8330d | 961 | && !may_be_aliased (base) |
8813a647 | 962 | && !(VAR_P (base) |
88d8330d EB |
963 | && !DECL_EXTERNAL (base) |
964 | && !TREE_STATIC (base) | |
965 | && local_variable_can_escape (base))) | |
d26c7090 ER |
966 | return false; |
967 | return true; | |
968 | } | |
6fb5fa3c DB |
969 | |
970 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
971 | OFFSET and WIDTH. */ | |
972 | ||
973 | static void | |
02ce5d90 | 974 | set_usage_bits (group_info *group, poly_int64 offset, poly_int64 width, |
d26c7090 | 975 | tree expr) |
6fb5fa3c | 976 | { |
02ce5d90 RS |
977 | /* Non-constant offsets and widths act as global kills, so there's no point |
978 | trying to use them to derive global DSE candidates. */ | |
979 | HOST_WIDE_INT i, const_offset, const_width; | |
d26c7090 | 980 | bool expr_escapes = can_escape (expr); |
02ce5d90 RS |
981 | if (offset.is_constant (&const_offset) |
982 | && width.is_constant (&const_width) | |
983 | && const_offset > -MAX_OFFSET | |
984 | && const_offset + const_width < MAX_OFFSET) | |
985 | for (i = const_offset; i < const_offset + const_width; ++i) | |
6fb5fa3c DB |
986 | { |
987 | bitmap store1; | |
988 | bitmap store2; | |
d26c7090 | 989 | bitmap escaped; |
6fb5fa3c DB |
990 | int ai; |
991 | if (i < 0) | |
992 | { | |
993 | store1 = group->store1_n; | |
994 | store2 = group->store2_n; | |
d26c7090 | 995 | escaped = group->escaped_n; |
6fb5fa3c DB |
996 | ai = -i; |
997 | } | |
998 | else | |
999 | { | |
1000 | store1 = group->store1_p; | |
1001 | store2 = group->store2_p; | |
d26c7090 | 1002 | escaped = group->escaped_p; |
6fb5fa3c DB |
1003 | ai = i; |
1004 | } | |
b8698a0f | 1005 | |
fcaa4ca4 | 1006 | if (!bitmap_set_bit (store1, ai)) |
6fb5fa3c | 1007 | bitmap_set_bit (store2, ai); |
b8698a0f | 1008 | else |
6fb5fa3c | 1009 | { |
6fb5fa3c DB |
1010 | if (i < 0) |
1011 | { | |
1012 | if (group->offset_map_size_n < ai) | |
1013 | group->offset_map_size_n = ai; | |
1014 | } | |
1015 | else | |
1016 | { | |
1017 | if (group->offset_map_size_p < ai) | |
1018 | group->offset_map_size_p = ai; | |
1019 | } | |
1020 | } | |
d26c7090 ER |
1021 | if (expr_escapes) |
1022 | bitmap_set_bit (escaped, ai); | |
6fb5fa3c DB |
1023 | } |
1024 | } | |
1025 | ||
d26c7090 ER |
1026 | static void |
1027 | reset_active_stores (void) | |
1028 | { | |
1029 | active_local_stores = NULL; | |
1030 | active_local_stores_len = 0; | |
1031 | } | |
6fb5fa3c | 1032 | |
d26c7090 | 1033 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
6fb5fa3c DB |
1034 | |
1035 | static void | |
d26c7090 | 1036 | free_read_records (bb_info_t bb_info) |
6fb5fa3c DB |
1037 | { |
1038 | insn_info_t insn_info = bb_info->last_insn; | |
1039 | read_info_t *ptr = &insn_info->read_rec; | |
6fb5fa3c DB |
1040 | while (*ptr) |
1041 | { | |
1042 | read_info_t next = (*ptr)->next; | |
ac6929b5 RB |
1043 | read_info_type_pool.remove (*ptr); |
1044 | *ptr = next; | |
6fb5fa3c | 1045 | } |
d26c7090 ER |
1046 | } |
1047 | ||
1048 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1049 | ||
1050 | static void | |
1051 | add_wild_read (bb_info_t bb_info) | |
1052 | { | |
1053 | insn_info_t insn_info = bb_info->last_insn; | |
6fb5fa3c | 1054 | insn_info->wild_read = true; |
d26c7090 ER |
1055 | free_read_records (bb_info); |
1056 | reset_active_stores (); | |
6fb5fa3c DB |
1057 | } |
1058 | ||
d26c7090 ER |
1059 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1060 | non-frame locations. */ | |
1061 | ||
1062 | static void | |
1063 | add_non_frame_wild_read (bb_info_t bb_info) | |
1064 | { | |
1065 | insn_info_t insn_info = bb_info->last_insn; | |
1066 | insn_info->non_frame_wild_read = true; | |
1067 | free_read_records (bb_info); | |
1068 | reset_active_stores (); | |
1069 | } | |
6fb5fa3c | 1070 | |
50f0f366 EB |
1071 | /* Return true if X is a constant or one of the registers that behave |
1072 | as a constant over the life of a function. This is equivalent to | |
1073 | !rtx_varies_p for memory addresses. */ | |
6fb5fa3c DB |
1074 | |
1075 | static bool | |
1076 | const_or_frame_p (rtx x) | |
1077 | { | |
d8116890 KZ |
1078 | if (CONSTANT_P (x)) |
1079 | return true; | |
1080 | ||
1081 | if (GET_CODE (x) == REG) | |
6fb5fa3c | 1082 | { |
6fb5fa3c DB |
1083 | /* Note that we have to test for the actual rtx used for the frame |
1084 | and arg pointers and not just the register number in case we have | |
1085 | eliminated the frame and/or arg pointer and are using it | |
1086 | for pseudos. */ | |
1087 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1088 | /* The arg pointer varies if it is not a fixed register. */ | |
1089 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1090 | || x == pic_offset_table_rtx) | |
1091 | return true; | |
1092 | return false; | |
6fb5fa3c | 1093 | } |
da6603c6 | 1094 | |
d8116890 | 1095 | return false; |
6fb5fa3c DB |
1096 | } |
1097 | ||
b8698a0f L |
1098 | /* Take all reasonable action to put the address of MEM into the form |
1099 | that we can do analysis on. | |
6fb5fa3c DB |
1100 | |
1101 | The gold standard is to get the address into the form: address + | |
1102 | OFFSET where address is something that rtx_varies_p considers a | |
1103 | constant. When we can get the address in this form, we can do | |
1104 | global analysis on it. Note that for constant bases, address is | |
1105 | not actually returned, only the group_id. The address can be | |
1106 | obtained from that. | |
1107 | ||
1108 | If that fails, we try cselib to get a value we can at least use | |
b8698a0f L |
1109 | locally. If that fails we return false. |
1110 | ||
6fb5fa3c DB |
1111 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1112 | group for non_varying bases. | |
1113 | ||
1114 | FOR_READ is true if this is a mem read and false if not. */ | |
1115 | ||
1116 | static bool | |
1117 | canon_address (rtx mem, | |
6fb5fa3c | 1118 | int *group_id, |
02ce5d90 | 1119 | poly_int64 *offset, |
6fb5fa3c DB |
1120 | cselib_val **base) |
1121 | { | |
ef4bddc2 | 1122 | machine_mode address_mode = get_address_mode (mem); |
6fb5fa3c DB |
1123 | rtx mem_address = XEXP (mem, 0); |
1124 | rtx expanded_address, address; | |
403c7520 JJ |
1125 | int expanded; |
1126 | ||
4deef538 | 1127 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
6fb5fa3c | 1128 | |
456610d3 | 1129 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1130 | { |
1131 | fprintf (dump_file, " mem: "); | |
1132 | print_inline_rtx (dump_file, mem_address, 0); | |
1133 | fprintf (dump_file, "\n"); | |
1134 | } | |
1135 | ||
403c7520 JJ |
1136 | /* First see if just canon_rtx (mem_address) is const or frame, |
1137 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1138 | address = NULL_RTX; | |
1139 | for (expanded = 0; expanded < 2; expanded++) | |
1140 | { | |
1141 | if (expanded) | |
1142 | { | |
1143 | /* Use cselib to replace all of the reg references with the full | |
b8698a0f | 1144 | expression. This will take care of the case where we have |
6fb5fa3c | 1145 | |
403c7520 JJ |
1146 | r_x = base + offset; |
1147 | val = *r_x; | |
b8698a0f L |
1148 | |
1149 | by making it into | |
6fb5fa3c | 1150 | |
403c7520 | 1151 | val = *(base + offset); */ |
6fb5fa3c | 1152 | |
403c7520 JJ |
1153 | expanded_address = cselib_expand_value_rtx (mem_address, |
1154 | scratch, 5); | |
6fb5fa3c | 1155 | |
403c7520 JJ |
1156 | /* If this fails, just go with the address from first |
1157 | iteration. */ | |
1158 | if (!expanded_address) | |
1159 | break; | |
1160 | } | |
1161 | else | |
1162 | expanded_address = mem_address; | |
6fb5fa3c | 1163 | |
403c7520 JJ |
1164 | /* Split the address into canonical BASE + OFFSET terms. */ |
1165 | address = canon_rtx (expanded_address); | |
6fb5fa3c | 1166 | |
403c7520 | 1167 | *offset = 0; |
6fb5fa3c | 1168 | |
456610d3 | 1169 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1170 | { |
1171 | if (expanded) | |
1172 | { | |
1173 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1174 | print_inline_rtx (dump_file, expanded_address, 0); | |
1175 | fprintf (dump_file, "\n"); | |
1176 | } | |
6fb5fa3c | 1177 | |
403c7520 JJ |
1178 | fprintf (dump_file, "\n after canon_rtx address: "); |
1179 | print_inline_rtx (dump_file, address, 0); | |
1180 | fprintf (dump_file, "\n"); | |
1181 | } | |
6fb5fa3c | 1182 | |
403c7520 JJ |
1183 | if (GET_CODE (address) == CONST) |
1184 | address = XEXP (address, 0); | |
6fb5fa3c | 1185 | |
02ce5d90 | 1186 | address = strip_offset_and_add (address, offset); |
6fb5fa3c | 1187 | |
09e881c9 BE |
1188 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1189 | && const_or_frame_p (address)) | |
6fb5fa3c | 1190 | { |
fd6daec9 | 1191 | group_info *group = get_group_info (address); |
403c7520 | 1192 | |
456610d3 | 1193 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
1194 | { |
1195 | fprintf (dump_file, " gid=%d offset=", group->id); | |
1196 | print_dec (*offset, dump_file); | |
1197 | fprintf (dump_file, "\n"); | |
1198 | } | |
403c7520 JJ |
1199 | *base = NULL; |
1200 | *group_id = group->id; | |
1201 | return true; | |
6fb5fa3c | 1202 | } |
403c7520 JJ |
1203 | } |
1204 | ||
4deef538 | 1205 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
403c7520 JJ |
1206 | *group_id = -1; |
1207 | ||
1208 | if (*base == NULL) | |
1209 | { | |
456610d3 | 1210 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1211 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1212 | return false; | |
6fb5fa3c | 1213 | } |
456610d3 | 1214 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
1215 | { |
1216 | fprintf (dump_file, " varying cselib base=%u:%u offset = ", | |
1217 | (*base)->uid, (*base)->hash); | |
1218 | print_dec (*offset, dump_file); | |
1219 | fprintf (dump_file, "\n"); | |
1220 | } | |
6fb5fa3c DB |
1221 | return true; |
1222 | } | |
1223 | ||
1224 | ||
1225 | /* Clear the rhs field from the active_local_stores array. */ | |
1226 | ||
1227 | static void | |
1228 | clear_rhs_from_active_local_stores (void) | |
1229 | { | |
1230 | insn_info_t ptr = active_local_stores; | |
1231 | ||
1232 | while (ptr) | |
1233 | { | |
fd6daec9 | 1234 | store_info *store_info = ptr->store_rec; |
6fb5fa3c DB |
1235 | /* Skip the clobbers. */ |
1236 | while (!store_info->is_set) | |
1237 | store_info = store_info->next; | |
1238 | ||
1239 | store_info->rhs = NULL; | |
8dd5516b | 1240 | store_info->const_rhs = NULL; |
6fb5fa3c DB |
1241 | |
1242 | ptr = ptr->next_local_store; | |
1243 | } | |
1244 | } | |
1245 | ||
1246 | ||
8dd5516b JJ |
1247 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1248 | ||
1249 | static inline void | |
fd6daec9 | 1250 | set_position_unneeded (store_info *s_info, int pos) |
8dd5516b JJ |
1251 | { |
1252 | if (__builtin_expect (s_info->is_large, false)) | |
1253 | { | |
fcaa4ca4 NF |
1254 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1255 | s_info->positions_needed.large.count++; | |
8dd5516b JJ |
1256 | } |
1257 | else | |
1258 | s_info->positions_needed.small_bitmask | |
fecfbfa4 | 1259 | &= ~(HOST_WIDE_INT_1U << pos); |
8dd5516b JJ |
1260 | } |
1261 | ||
1262 | /* Mark the whole store S_INFO as unneeded. */ | |
1263 | ||
1264 | static inline void | |
fd6daec9 | 1265 | set_all_positions_unneeded (store_info *s_info) |
8dd5516b JJ |
1266 | { |
1267 | if (__builtin_expect (s_info->is_large, false)) | |
1268 | { | |
02ce5d90 RS |
1269 | HOST_WIDE_INT width; |
1270 | if (s_info->width.is_constant (&width)) | |
1271 | { | |
1272 | bitmap_set_range (s_info->positions_needed.large.bmap, 0, width); | |
1273 | s_info->positions_needed.large.count = width; | |
1274 | } | |
1275 | else | |
1276 | { | |
1277 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1278 | s_info->positions_needed.large.count = 1; | |
1279 | } | |
8dd5516b JJ |
1280 | } |
1281 | else | |
07e96250 | 1282 | s_info->positions_needed.small_bitmask = HOST_WIDE_INT_0U; |
8dd5516b JJ |
1283 | } |
1284 | ||
1285 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1286 | ||
1287 | static inline bool | |
fd6daec9 | 1288 | any_positions_needed_p (store_info *s_info) |
8dd5516b JJ |
1289 | { |
1290 | if (__builtin_expect (s_info->is_large, false)) | |
02ce5d90 RS |
1291 | { |
1292 | HOST_WIDE_INT width; | |
1293 | if (s_info->width.is_constant (&width)) | |
1294 | { | |
1295 | gcc_checking_assert (s_info->positions_needed.large.bmap); | |
1296 | return s_info->positions_needed.large.count < width; | |
1297 | } | |
1298 | else | |
1299 | { | |
1300 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1301 | return s_info->positions_needed.large.count == 0; | |
1302 | } | |
1303 | } | |
8dd5516b | 1304 | else |
07e96250 | 1305 | return (s_info->positions_needed.small_bitmask != HOST_WIDE_INT_0U); |
8dd5516b JJ |
1306 | } |
1307 | ||
1308 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
02ce5d90 | 1309 | store are known to be needed. */ |
8dd5516b JJ |
1310 | |
1311 | static inline bool | |
02ce5d90 RS |
1312 | all_positions_needed_p (store_info *s_info, poly_int64 start, |
1313 | poly_int64 width) | |
8dd5516b | 1314 | { |
02ce5d90 RS |
1315 | gcc_assert (s_info->rhs); |
1316 | if (!s_info->width.is_constant ()) | |
1317 | { | |
1318 | gcc_assert (s_info->is_large | |
1319 | && !s_info->positions_needed.large.bmap); | |
1320 | return s_info->positions_needed.large.count == 0; | |
1321 | } | |
1322 | ||
1323 | /* Otherwise, if START and WIDTH are non-constant, we're asking about | |
1324 | a non-constant region of a constant-sized store. We can't say for | |
1325 | sure that all positions are needed. */ | |
1326 | HOST_WIDE_INT const_start, const_width; | |
1327 | if (!start.is_constant (&const_start) | |
1328 | || !width.is_constant (&const_width)) | |
1329 | return false; | |
1330 | ||
8dd5516b JJ |
1331 | if (__builtin_expect (s_info->is_large, false)) |
1332 | { | |
02ce5d90 RS |
1333 | for (HOST_WIDE_INT i = const_start; i < const_start + const_width; ++i) |
1334 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, i)) | |
8dd5516b JJ |
1335 | return false; |
1336 | return true; | |
1337 | } | |
1338 | else | |
1339 | { | |
02ce5d90 RS |
1340 | unsigned HOST_WIDE_INT mask |
1341 | = lowpart_bitmask (const_width) << const_start; | |
8dd5516b JJ |
1342 | return (s_info->positions_needed.small_bitmask & mask) == mask; |
1343 | } | |
1344 | } | |
1345 | ||
1346 | ||
02ce5d90 RS |
1347 | static rtx get_stored_val (store_info *, machine_mode, poly_int64, |
1348 | poly_int64, basic_block, bool); | |
8dd5516b JJ |
1349 | |
1350 | ||
6fb5fa3c DB |
1351 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1352 | there is a candidate store, after adding it to the appropriate | |
1353 | local store group if so. */ | |
1354 | ||
1355 | static int | |
1356 | record_store (rtx body, bb_info_t bb_info) | |
1357 | { | |
6216f94e | 1358 | rtx mem, rhs, const_rhs, mem_addr; |
02ce5d90 RS |
1359 | poly_int64 offset = 0; |
1360 | poly_int64 width = 0; | |
6fb5fa3c | 1361 | insn_info_t insn_info = bb_info->last_insn; |
fd6daec9 | 1362 | store_info *store_info = NULL; |
6fb5fa3c DB |
1363 | int group_id; |
1364 | cselib_val *base = NULL; | |
8dd5516b | 1365 | insn_info_t ptr, last, redundant_reason; |
6fb5fa3c DB |
1366 | bool store_is_unused; |
1367 | ||
1368 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1369 | return 0; | |
1370 | ||
8dd5516b JJ |
1371 | mem = SET_DEST (body); |
1372 | ||
6fb5fa3c DB |
1373 | /* If this is not used, then this cannot be used to keep the insn |
1374 | from being deleted. On the other hand, it does provide something | |
1375 | that can be used to prove that another store is dead. */ | |
1376 | store_is_unused | |
8dd5516b | 1377 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
6fb5fa3c DB |
1378 | |
1379 | /* Check whether that value is a suitable memory location. */ | |
6fb5fa3c DB |
1380 | if (!MEM_P (mem)) |
1381 | { | |
1382 | /* If the set or clobber is unused, then it does not effect our | |
1383 | ability to get rid of the entire insn. */ | |
1384 | if (!store_is_unused) | |
1385 | insn_info->cannot_delete = true; | |
1386 | return 0; | |
1387 | } | |
1388 | ||
1389 | /* At this point we know mem is a mem. */ | |
1390 | if (GET_MODE (mem) == BLKmode) | |
1391 | { | |
d05d7551 | 1392 | HOST_WIDE_INT const_size; |
6fb5fa3c DB |
1393 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) |
1394 | { | |
456610d3 | 1395 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1396 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1397 | add_wild_read (bb_info); | |
1398 | insn_info->cannot_delete = true; | |
8dd5516b | 1399 | return 0; |
6fb5fa3c | 1400 | } |
8dd5516b JJ |
1401 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1402 | as memset (addr, 0, 36); */ | |
f5541398 | 1403 | else if (!MEM_SIZE_KNOWN_P (mem) |
d05d7551 RS |
1404 | || maybe_le (MEM_SIZE (mem), 0) |
1405 | /* This is a limit on the bitmap size, which is only relevant | |
1406 | for constant-sized MEMs. */ | |
1407 | || (MEM_SIZE (mem).is_constant (&const_size) | |
1408 | && const_size > MAX_OFFSET) | |
8dd5516b | 1409 | || GET_CODE (body) != SET |
8dd5516b | 1410 | || !CONST_INT_P (SET_SRC (body))) |
6fb5fa3c | 1411 | { |
8dd5516b JJ |
1412 | if (!store_is_unused) |
1413 | { | |
1414 | /* If the set or clobber is unused, then it does not effect our | |
1415 | ability to get rid of the entire insn. */ | |
1416 | insn_info->cannot_delete = true; | |
1417 | clear_rhs_from_active_local_stores (); | |
1418 | } | |
1419 | return 0; | |
6fb5fa3c | 1420 | } |
6fb5fa3c DB |
1421 | } |
1422 | ||
1423 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1424 | if (MEM_VOLATILE_P (mem)) | |
8dd5516b | 1425 | insn_info->cannot_delete = true; |
6fb5fa3c | 1426 | |
ac6929b5 | 1427 | if (!canon_address (mem, &group_id, &offset, &base)) |
6fb5fa3c DB |
1428 | { |
1429 | clear_rhs_from_active_local_stores (); | |
1430 | return 0; | |
1431 | } | |
1432 | ||
8dd5516b | 1433 | if (GET_MODE (mem) == BLKmode) |
f5541398 | 1434 | width = MEM_SIZE (mem); |
8dd5516b | 1435 | else |
6f557e0e | 1436 | width = GET_MODE_SIZE (GET_MODE (mem)); |
6fb5fa3c | 1437 | |
02ce5d90 | 1438 | if (!endpoint_representable_p (offset, width)) |
213ffde2 JJ |
1439 | { |
1440 | clear_rhs_from_active_local_stores (); | |
1441 | return 0; | |
1442 | } | |
1443 | ||
cbe679a4 JJ |
1444 | if (known_eq (width, 0)) |
1445 | return 0; | |
1446 | ||
ac6929b5 | 1447 | if (group_id >= 0) |
6fb5fa3c DB |
1448 | { |
1449 | /* In the restrictive case where the base is a constant or the | |
1450 | frame pointer we can do global analysis. */ | |
b8698a0f | 1451 | |
fd6daec9 | 1452 | group_info *group |
9771b263 | 1453 | = rtx_group_vec[group_id]; |
d26c7090 | 1454 | tree expr = MEM_EXPR (mem); |
b8698a0f | 1455 | |
da6603c6 | 1456 | store_info = rtx_store_info_pool.allocate (); |
d26c7090 | 1457 | set_usage_bits (group, offset, width, expr); |
6fb5fa3c | 1458 | |
456610d3 | 1459 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1460 | { |
1461 | fprintf (dump_file, " processing const base store gid=%d", | |
1462 | group_id); | |
1463 | print_range (dump_file, offset, width); | |
1464 | fprintf (dump_file, "\n"); | |
1465 | } | |
6fb5fa3c DB |
1466 | } |
1467 | else | |
1468 | { | |
9e412ca3 | 1469 | if (may_be_sp_based_p (XEXP (mem, 0))) |
50f0f366 | 1470 | insn_info->stack_pointer_based = true; |
6fb5fa3c | 1471 | insn_info->contains_cselib_groups = true; |
50f0f366 | 1472 | |
da6603c6 | 1473 | store_info = cse_store_info_pool.allocate (); |
6fb5fa3c DB |
1474 | group_id = -1; |
1475 | ||
456610d3 | 1476 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1477 | { |
1478 | fprintf (dump_file, " processing cselib store "); | |
1479 | print_range (dump_file, offset, width); | |
1480 | fprintf (dump_file, "\n"); | |
1481 | } | |
6fb5fa3c DB |
1482 | } |
1483 | ||
8dd5516b JJ |
1484 | const_rhs = rhs = NULL_RTX; |
1485 | if (GET_CODE (body) == SET | |
1486 | /* No place to keep the value after ra. */ | |
1487 | && !reload_completed | |
1488 | && (REG_P (SET_SRC (body)) | |
1489 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1490 | || CONSTANT_P (SET_SRC (body))) | |
1491 | && !MEM_VOLATILE_P (mem) | |
1492 | /* Sometimes the store and reload is used for truncation and | |
1493 | rounding. */ | |
1494 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1495 | { | |
1496 | rhs = SET_SRC (body); | |
1497 | if (CONSTANT_P (rhs)) | |
1498 | const_rhs = rhs; | |
1499 | else if (body == PATTERN (insn_info->insn)) | |
1500 | { | |
1501 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1502 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1503 | const_rhs = XEXP (tem, 0); | |
1504 | } | |
1505 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1506 | { | |
1507 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1508 | ||
1509 | if (tem && CONSTANT_P (tem)) | |
1510 | const_rhs = tem; | |
1511 | } | |
1512 | } | |
1513 | ||
6fb5fa3c DB |
1514 | /* Check to see if this stores causes some other stores to be |
1515 | dead. */ | |
1516 | ptr = active_local_stores; | |
1517 | last = NULL; | |
8dd5516b | 1518 | redundant_reason = NULL; |
6216f94e | 1519 | mem = canon_rtx (mem); |
ac6929b5 RB |
1520 | |
1521 | if (group_id < 0) | |
1522 | mem_addr = base->val_rtx; | |
6216f94e JJ |
1523 | else |
1524 | { | |
ac6929b5 RB |
1525 | group_info *group = rtx_group_vec[group_id]; |
1526 | mem_addr = group->canon_base_addr; | |
6216f94e | 1527 | } |
02ce5d90 | 1528 | if (maybe_ne (offset, 0)) |
ac6929b5 | 1529 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6fb5fa3c DB |
1530 | |
1531 | while (ptr) | |
1532 | { | |
1533 | insn_info_t next = ptr->next_local_store; | |
99b1c316 | 1534 | class store_info *s_info = ptr->store_rec; |
60564289 | 1535 | bool del = true; |
6fb5fa3c DB |
1536 | |
1537 | /* Skip the clobbers. We delete the active insn if this insn | |
6ed3da00 | 1538 | shadows the set. To have been put on the active list, it |
6fb5fa3c DB |
1539 | has exactly on set. */ |
1540 | while (!s_info->is_set) | |
1541 | s_info = s_info->next; | |
1542 | ||
ac6929b5 | 1543 | if (s_info->group_id == group_id && s_info->cse_base == base) |
6fb5fa3c DB |
1544 | { |
1545 | HOST_WIDE_INT i; | |
456610d3 | 1546 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1547 | { |
1548 | fprintf (dump_file, " trying store in insn=%d gid=%d", | |
1549 | INSN_UID (ptr->insn), s_info->group_id); | |
1550 | print_range (dump_file, s_info->offset, s_info->width); | |
1551 | fprintf (dump_file, "\n"); | |
1552 | } | |
8dd5516b JJ |
1553 | |
1554 | /* Even if PTR won't be eliminated as unneeded, if both | |
1555 | PTR and this insn store the same constant value, we might | |
1556 | eliminate this insn instead. */ | |
1557 | if (s_info->const_rhs | |
1558 | && const_rhs | |
71d6a386 RS |
1559 | && known_subrange_p (offset, width, |
1560 | s_info->offset, s_info->width) | |
1561 | && all_positions_needed_p (s_info, offset - s_info->offset, | |
d7e1f499 RB |
1562 | width) |
1563 | /* We can only remove the later store if the earlier aliases | |
1564 | at least all accesses the later one. */ | |
3d6fd7ce RB |
1565 | && ((MEM_ALIAS_SET (mem) == MEM_ALIAS_SET (s_info->mem) |
1566 | || alias_set_subset_of (MEM_ALIAS_SET (mem), | |
1567 | MEM_ALIAS_SET (s_info->mem))) | |
1568 | && (!MEM_EXPR (s_info->mem) | |
1569 | || refs_same_for_tbaa_p (MEM_EXPR (s_info->mem), | |
1570 | MEM_EXPR (mem))))) | |
8dd5516b JJ |
1571 | { |
1572 | if (GET_MODE (mem) == BLKmode) | |
1573 | { | |
1574 | if (GET_MODE (s_info->mem) == BLKmode | |
1575 | && s_info->const_rhs == const_rhs) | |
1576 | redundant_reason = ptr; | |
1577 | } | |
1578 | else if (s_info->const_rhs == const0_rtx | |
1579 | && const_rhs == const0_rtx) | |
1580 | redundant_reason = ptr; | |
1581 | else | |
1582 | { | |
1583 | rtx val; | |
1584 | start_sequence (); | |
71d6a386 | 1585 | val = get_stored_val (s_info, GET_MODE (mem), offset, width, |
8dd5516b JJ |
1586 | BLOCK_FOR_INSN (insn_info->insn), |
1587 | true); | |
1588 | if (get_insns () != NULL) | |
1589 | val = NULL_RTX; | |
1590 | end_sequence (); | |
1591 | if (val && rtx_equal_p (val, const_rhs)) | |
1592 | redundant_reason = ptr; | |
1593 | } | |
1594 | } | |
1595 | ||
02ce5d90 | 1596 | HOST_WIDE_INT begin_unneeded, const_s_width, const_width; |
71d6a386 RS |
1597 | if (known_subrange_p (s_info->offset, s_info->width, offset, width)) |
1598 | /* The new store touches every byte that S_INFO does. */ | |
1599 | set_all_positions_unneeded (s_info); | |
02ce5d90 RS |
1600 | else if ((offset - s_info->offset).is_constant (&begin_unneeded) |
1601 | && s_info->width.is_constant (&const_s_width) | |
1602 | && width.is_constant (&const_width)) | |
71d6a386 | 1603 | { |
02ce5d90 | 1604 | HOST_WIDE_INT end_unneeded = begin_unneeded + const_width; |
71d6a386 | 1605 | begin_unneeded = MAX (begin_unneeded, 0); |
02ce5d90 | 1606 | end_unneeded = MIN (end_unneeded, const_s_width); |
71d6a386 RS |
1607 | for (i = begin_unneeded; i < end_unneeded; ++i) |
1608 | set_position_unneeded (s_info, i); | |
1609 | } | |
02ce5d90 RS |
1610 | else |
1611 | { | |
1612 | /* We don't know which parts of S_INFO are needed and | |
1613 | which aren't, so invalidate the RHS. */ | |
1614 | s_info->rhs = NULL; | |
1615 | s_info->const_rhs = NULL; | |
1616 | } | |
6fb5fa3c DB |
1617 | } |
1618 | else if (s_info->rhs) | |
1619 | /* Need to see if it is possible for this store to overwrite | |
1620 | the value of store_info. If it is, set the rhs to NULL to | |
1621 | keep it from being used to remove a load. */ | |
1622 | { | |
43b9f499 RB |
1623 | if (canon_output_dependence (s_info->mem, true, |
1624 | mem, GET_MODE (mem), | |
1625 | mem_addr)) | |
8dd5516b JJ |
1626 | { |
1627 | s_info->rhs = NULL; | |
1628 | s_info->const_rhs = NULL; | |
1629 | } | |
6fb5fa3c | 1630 | } |
6216f94e | 1631 | |
6fb5fa3c DB |
1632 | /* An insn can be deleted if every position of every one of |
1633 | its s_infos is zero. */ | |
1b6fa860 | 1634 | if (any_positions_needed_p (s_info)) |
60564289 | 1635 | del = false; |
8dd5516b | 1636 | |
60564289 | 1637 | if (del) |
6fb5fa3c DB |
1638 | { |
1639 | insn_info_t insn_to_delete = ptr; | |
b8698a0f | 1640 | |
dabd47e7 | 1641 | active_local_stores_len--; |
6fb5fa3c DB |
1642 | if (last) |
1643 | last->next_local_store = ptr->next_local_store; | |
1644 | else | |
1645 | active_local_stores = ptr->next_local_store; | |
b8698a0f | 1646 | |
1b6fa860 JJ |
1647 | if (!insn_to_delete->cannot_delete) |
1648 | delete_dead_store_insn (insn_to_delete); | |
6fb5fa3c DB |
1649 | } |
1650 | else | |
1651 | last = ptr; | |
b8698a0f | 1652 | |
6fb5fa3c DB |
1653 | ptr = next; |
1654 | } | |
b8698a0f | 1655 | |
6fb5fa3c DB |
1656 | /* Finish filling in the store_info. */ |
1657 | store_info->next = insn_info->store_rec; | |
1658 | insn_info->store_rec = store_info; | |
6216f94e | 1659 | store_info->mem = mem; |
6216f94e | 1660 | store_info->mem_addr = mem_addr; |
6fb5fa3c | 1661 | store_info->cse_base = base; |
02ce5d90 RS |
1662 | HOST_WIDE_INT const_width; |
1663 | if (!width.is_constant (&const_width)) | |
1664 | { | |
1665 | store_info->is_large = true; | |
1666 | store_info->positions_needed.large.count = 0; | |
1667 | store_info->positions_needed.large.bmap = NULL; | |
1668 | } | |
1669 | else if (const_width > HOST_BITS_PER_WIDE_INT) | |
8dd5516b JJ |
1670 | { |
1671 | store_info->is_large = true; | |
1672 | store_info->positions_needed.large.count = 0; | |
3f9b14ff | 1673 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
8dd5516b JJ |
1674 | } |
1675 | else | |
1676 | { | |
1677 | store_info->is_large = false; | |
02ce5d90 RS |
1678 | store_info->positions_needed.small_bitmask |
1679 | = lowpart_bitmask (const_width); | |
8dd5516b | 1680 | } |
6fb5fa3c | 1681 | store_info->group_id = group_id; |
71d6a386 RS |
1682 | store_info->offset = offset; |
1683 | store_info->width = width; | |
6fb5fa3c | 1684 | store_info->is_set = GET_CODE (body) == SET; |
8dd5516b JJ |
1685 | store_info->rhs = rhs; |
1686 | store_info->const_rhs = const_rhs; | |
1687 | store_info->redundant_reason = redundant_reason; | |
6fb5fa3c | 1688 | |
6fb5fa3c DB |
1689 | /* If this is a clobber, we return 0. We will only be able to |
1690 | delete this insn if there is only one store USED store, but we | |
1691 | can use the clobber to delete other stores earlier. */ | |
1692 | return store_info->is_set ? 1 : 0; | |
1693 | } | |
1694 | ||
1695 | ||
1696 | static void | |
1697 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1698 | { | |
b8698a0f | 1699 | fprintf (dump_file, "%s insn=%d %s\n", start, |
6fb5fa3c DB |
1700 | INSN_UID (insn_info->insn), |
1701 | insn_info->store_rec ? "has store" : "naked"); | |
1702 | } | |
1703 | ||
1704 | ||
8660aaae EC |
1705 | /* If the modes are different and the value's source and target do not |
1706 | line up, we need to extract the value from lower part of the rhs of | |
1707 | the store, shift it, and then put it into a form that can be shoved | |
1708 | into the read_insn. This function generates a right SHIFT of a | |
1709 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1710 | shift sequence is returned or NULL if we failed to find a | |
1711 | shift. */ | |
1712 | ||
1713 | static rtx | |
02ce5d90 | 1714 | find_shift_sequence (poly_int64 access_size, |
fd6daec9 | 1715 | store_info *store_info, |
ef4bddc2 | 1716 | machine_mode read_mode, |
02ce5d90 | 1717 | poly_int64 shift, bool speed, bool require_cst) |
8660aaae | 1718 | { |
ef4bddc2 | 1719 | machine_mode store_mode = GET_MODE (store_info->mem); |
f67f4dff | 1720 | scalar_int_mode new_mode; |
18b526e8 | 1721 | rtx read_reg = NULL; |
8660aaae EC |
1722 | |
1723 | /* Some machines like the x86 have shift insns for each size of | |
1724 | operand. Other machines like the ppc or the ia-64 may only have | |
1725 | shift insns that shift values within 32 or 64 bit registers. | |
1726 | This loop tries to find the smallest shift insn that will right | |
1727 | justify the value we want to read but is available in one insn on | |
1728 | the machine. */ | |
1729 | ||
f67f4dff RS |
1730 | opt_scalar_int_mode new_mode_iter; |
1731 | FOR_EACH_MODE_FROM (new_mode_iter, | |
1732 | smallest_int_mode_for_size (access_size * BITS_PER_UNIT)) | |
8660aaae | 1733 | { |
dc01c3d1 DM |
1734 | rtx target, new_reg, new_lhs; |
1735 | rtx_insn *shift_seq, *insn; | |
d898d29b | 1736 | int cost; |
348eea5f | 1737 | |
f67f4dff | 1738 | new_mode = new_mode_iter.require (); |
c94843d2 RS |
1739 | if (GET_MODE_BITSIZE (new_mode) > BITS_PER_WORD) |
1740 | break; | |
1741 | ||
72a2609f JJ |
1742 | /* If a constant was stored into memory, try to simplify it here, |
1743 | otherwise the cost of the shift might preclude this optimization | |
1744 | e.g. at -Os, even when no actual shift will be needed. */ | |
8dd5516b | 1745 | if (store_info->const_rhs) |
72a2609f | 1746 | { |
91914e56 | 1747 | poly_uint64 byte = subreg_lowpart_offset (new_mode, store_mode); |
8dd5516b JJ |
1748 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1749 | store_mode, byte); | |
72a2609f JJ |
1750 | if (ret && CONSTANT_P (ret)) |
1751 | { | |
abd3c800 | 1752 | rtx shift_rtx = gen_int_shift_amount (new_mode, shift); |
72a2609f | 1753 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, |
abd3c800 | 1754 | ret, shift_rtx); |
72a2609f JJ |
1755 | if (ret && CONSTANT_P (ret)) |
1756 | { | |
1757 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1758 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1759 | if (ret && CONSTANT_P (ret) | |
e548c9df AM |
1760 | && (set_src_cost (ret, read_mode, speed) |
1761 | <= COSTS_N_INSNS (1))) | |
72a2609f JJ |
1762 | return ret; |
1763 | } | |
1764 | } | |
1765 | } | |
1766 | ||
8dd5516b JJ |
1767 | if (require_cst) |
1768 | return NULL_RTX; | |
1769 | ||
18b526e8 RS |
1770 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1771 | requires a real instruction. */ | |
73a699ae | 1772 | if (maybe_lt (GET_MODE_SIZE (new_mode), GET_MODE_SIZE (store_mode)) |
d0edd768 | 1773 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
348eea5f RS |
1774 | continue; |
1775 | ||
18b526e8 RS |
1776 | /* Also try a wider mode if the necessary punning is either not |
1777 | desirable or not possible. */ | |
1778 | if (!CONSTANT_P (store_info->rhs) | |
99e1629f | 1779 | && !targetm.modes_tieable_p (new_mode, store_mode)) |
18b526e8 | 1780 | continue; |
18b526e8 | 1781 | |
348eea5f | 1782 | new_reg = gen_reg_rtx (new_mode); |
8660aaae EC |
1783 | |
1784 | start_sequence (); | |
1785 | ||
1786 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1787 | of one dsp where the cost of these two was not the same. But | |
1788 | this really is a rare case anyway. */ | |
1789 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
abd3c800 RS |
1790 | gen_int_shift_amount (new_mode, shift), |
1791 | new_reg, 1, OPTAB_DIRECT); | |
8660aaae | 1792 | |
c6f3019a RS |
1793 | shift_seq = get_insns (); |
1794 | end_sequence (); | |
8660aaae | 1795 | |
c6f3019a RS |
1796 | if (target != new_reg || shift_seq == NULL) |
1797 | continue; | |
1798 | ||
1799 | cost = 0; | |
1800 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1801 | if (INSN_P (insn)) | |
d09e78e8 | 1802 | cost += insn_cost (insn, speed); |
c6f3019a RS |
1803 | |
1804 | /* The computation up to here is essentially independent | |
1805 | of the arguments and could be precomputed. It may | |
1806 | not be worth doing so. We could precompute if | |
1807 | worthwhile or at least cache the results. The result | |
06acf7d0 RS |
1808 | technically depends on both SHIFT and ACCESS_SIZE, |
1809 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
c6f3019a RS |
1810 | |
1811 | if (cost > COSTS_N_INSNS (1)) | |
1812 | continue; | |
1813 | ||
d898d29b JJ |
1814 | new_lhs = extract_low_bits (new_mode, store_mode, |
1815 | copy_rtx (store_info->rhs)); | |
1816 | if (new_lhs == NULL_RTX) | |
1817 | continue; | |
1818 | ||
c6f3019a RS |
1819 | /* We found an acceptable shift. Generate a move to |
1820 | take the value from the store and put it into the | |
1821 | shift pseudo, then shift it, then generate another | |
1822 | move to put in into the target of the read. */ | |
18b526e8 | 1823 | emit_move_insn (new_reg, new_lhs); |
c6f3019a | 1824 | emit_insn (shift_seq); |
18b526e8 | 1825 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
c6f3019a | 1826 | break; |
8660aaae EC |
1827 | } |
1828 | ||
18b526e8 | 1829 | return read_reg; |
8660aaae EC |
1830 | } |
1831 | ||
1832 | ||
02b47899 KZ |
1833 | /* Call back for note_stores to find the hard regs set or clobbered by |
1834 | insn. Data is a bitmap of the hardregs set so far. */ | |
1835 | ||
1836 | static void | |
1837 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1838 | { | |
1839 | bitmap regs_set = (bitmap) data; | |
1840 | ||
1841 | if (REG_P (x) | |
f773c2bd | 1842 | && HARD_REGISTER_P (x)) |
dc8afb70 | 1843 | bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); |
02b47899 KZ |
1844 | } |
1845 | ||
8dd5516b | 1846 | /* Helper function for replace_read and record_store. |
71d6a386 RS |
1847 | Attempt to return a value of mode READ_MODE stored in STORE_INFO, |
1848 | consisting of READ_WIDTH bytes starting from READ_OFFSET. Return NULL | |
8dd5516b JJ |
1849 | if not successful. If REQUIRE_CST is true, return always constant. */ |
1850 | ||
1851 | static rtx | |
fd6daec9 | 1852 | get_stored_val (store_info *store_info, machine_mode read_mode, |
02ce5d90 | 1853 | poly_int64 read_offset, poly_int64 read_width, |
8dd5516b JJ |
1854 | basic_block bb, bool require_cst) |
1855 | { | |
ef4bddc2 | 1856 | machine_mode store_mode = GET_MODE (store_info->mem); |
02ce5d90 | 1857 | poly_int64 gap; |
8dd5516b JJ |
1858 | rtx read_reg; |
1859 | ||
1860 | /* To get here the read is within the boundaries of the write so | |
1861 | shift will never be negative. Start out with the shift being in | |
1862 | bytes. */ | |
1863 | if (store_mode == BLKmode) | |
71d6a386 | 1864 | gap = 0; |
8dd5516b | 1865 | else if (BYTES_BIG_ENDIAN) |
71d6a386 RS |
1866 | gap = ((store_info->offset + store_info->width) |
1867 | - (read_offset + read_width)); | |
8dd5516b | 1868 | else |
71d6a386 | 1869 | gap = read_offset - store_info->offset; |
8dd5516b | 1870 | |
e6575643 | 1871 | if (gap.is_constant () && maybe_ne (gap, 0)) |
71d6a386 | 1872 | { |
02ce5d90 RS |
1873 | poly_int64 shift = gap * BITS_PER_UNIT; |
1874 | poly_int64 access_size = GET_MODE_SIZE (read_mode) + gap; | |
71d6a386 RS |
1875 | read_reg = find_shift_sequence (access_size, store_info, read_mode, |
1876 | shift, optimize_bb_for_speed_p (bb), | |
1877 | require_cst); | |
1878 | } | |
8dd5516b JJ |
1879 | else if (store_mode == BLKmode) |
1880 | { | |
1881 | /* The store is a memset (addr, const_val, const_size). */ | |
1882 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
304b9962 RS |
1883 | scalar_int_mode int_store_mode; |
1884 | if (!int_mode_for_mode (read_mode).exists (&int_store_mode)) | |
8dd5516b JJ |
1885 | read_reg = NULL_RTX; |
1886 | else if (store_info->rhs == const0_rtx) | |
304b9962 RS |
1887 | read_reg = extract_low_bits (read_mode, int_store_mode, const0_rtx); |
1888 | else if (GET_MODE_BITSIZE (int_store_mode) > HOST_BITS_PER_WIDE_INT | |
8dd5516b JJ |
1889 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) |
1890 | read_reg = NULL_RTX; | |
1891 | else | |
1892 | { | |
1893 | unsigned HOST_WIDE_INT c | |
1894 | = INTVAL (store_info->rhs) | |
fecfbfa4 | 1895 | & ((HOST_WIDE_INT_1 << BITS_PER_UNIT) - 1); |
8dd5516b JJ |
1896 | int shift = BITS_PER_UNIT; |
1897 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1898 | { | |
1899 | c |= (c << shift); | |
1900 | shift <<= 1; | |
1901 | } | |
304b9962 RS |
1902 | read_reg = gen_int_mode (c, int_store_mode); |
1903 | read_reg = extract_low_bits (read_mode, int_store_mode, read_reg); | |
8dd5516b JJ |
1904 | } |
1905 | } | |
1906 | else if (store_info->const_rhs | |
1907 | && (require_cst | |
1908 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1909 | read_reg = extract_low_bits (read_mode, store_mode, | |
1910 | copy_rtx (store_info->const_rhs)); | |
1911 | else | |
1912 | read_reg = extract_low_bits (read_mode, store_mode, | |
1913 | copy_rtx (store_info->rhs)); | |
1914 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1915 | read_reg = NULL_RTX; | |
1916 | return read_reg; | |
1917 | } | |
02b47899 | 1918 | |
6fb5fa3c DB |
1919 | /* Take a sequence of: |
1920 | A <- r1 | |
1921 | ... | |
1922 | ... <- A | |
1923 | ||
b8698a0f | 1924 | and change it into |
6fb5fa3c DB |
1925 | r2 <- r1 |
1926 | A <- r1 | |
1927 | ... | |
1928 | ... <- r2 | |
1929 | ||
8660aaae EC |
1930 | or |
1931 | ||
1932 | r3 <- extract (r1) | |
1933 | r3 <- r3 >> shift | |
1934 | r2 <- extract (r3) | |
1935 | ... <- r2 | |
1936 | ||
1937 | or | |
1938 | ||
1939 | r2 <- extract (r1) | |
1940 | ... <- r2 | |
1941 | ||
1942 | Depending on the alignment and the mode of the store and | |
1943 | subsequent load. | |
1944 | ||
1945 | ||
1946 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
6fb5fa3c DB |
1947 | and READ_INSN are for the read. Return true if the replacement |
1948 | went ok. */ | |
1949 | ||
1950 | static bool | |
fd6daec9 | 1951 | replace_read (store_info *store_info, insn_info_t store_insn, |
8dd5516b JJ |
1952 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
1953 | bitmap regs_live) | |
6fb5fa3c | 1954 | { |
ef4bddc2 RS |
1955 | machine_mode store_mode = GET_MODE (store_info->mem); |
1956 | machine_mode read_mode = GET_MODE (read_info->mem); | |
dc01c3d1 DM |
1957 | rtx_insn *insns, *this_insn; |
1958 | rtx read_reg; | |
8dd5516b | 1959 | basic_block bb; |
8660aaae | 1960 | |
6fb5fa3c DB |
1961 | if (!dbg_cnt (dse)) |
1962 | return false; | |
1963 | ||
18b526e8 RS |
1964 | /* Create a sequence of instructions to set up the read register. |
1965 | This sequence goes immediately before the store and its result | |
1966 | is read by the load. | |
1967 | ||
1968 | We need to keep this in perspective. We are replacing a read | |
8660aaae EC |
1969 | with a sequence of insns, but the read will almost certainly be |
1970 | in cache, so it is not going to be an expensive one. Thus, we | |
1971 | are not willing to do a multi insn shift or worse a subroutine | |
1972 | call to get rid of the read. */ | |
456610d3 | 1973 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
1974 | fprintf (dump_file, "trying to replace %smode load in insn %d" |
1975 | " from %smode store in insn %d\n", | |
1976 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
1977 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
1978 | start_sequence (); | |
8dd5516b JJ |
1979 | bb = BLOCK_FOR_INSN (read_insn->insn); |
1980 | read_reg = get_stored_val (store_info, | |
71d6a386 | 1981 | read_mode, read_info->offset, read_info->width, |
8dd5516b | 1982 | bb, false); |
18b526e8 | 1983 | if (read_reg == NULL_RTX) |
8660aaae | 1984 | { |
18b526e8 | 1985 | end_sequence (); |
456610d3 | 1986 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
1987 | fprintf (dump_file, " -- could not extract bits of stored value\n"); |
1988 | return false; | |
8660aaae | 1989 | } |
18b526e8 RS |
1990 | /* Force the value into a new register so that it won't be clobbered |
1991 | between the store and the load. */ | |
1992 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
1993 | insns = get_insns (); | |
1994 | end_sequence (); | |
8660aaae | 1995 | |
02b47899 KZ |
1996 | if (insns != NULL_RTX) |
1997 | { | |
1998 | /* Now we have to scan the set of new instructions to see if the | |
1999 | sequence contains and sets of hardregs that happened to be | |
2000 | live at this point. For instance, this can happen if one of | |
2001 | the insns sets the CC and the CC happened to be live at that | |
2002 | point. This does occasionally happen, see PR 37922. */ | |
3f9b14ff | 2003 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
02b47899 | 2004 | |
2c0fa3ec JJ |
2005 | for (this_insn = insns; |
2006 | this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
2007 | { | |
2008 | if (insn_invalid_p (this_insn, false)) | |
2009 | { | |
2010 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2011 | { | |
2012 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2013 | print_simple_rtl (dump_file, read_reg); | |
2014 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2015 | } | |
2016 | BITMAP_FREE (regs_set); | |
2017 | return false; | |
2018 | } | |
2019 | note_stores (this_insn, look_for_hardregs, regs_set); | |
2020 | } | |
b8698a0f | 2021 | |
02b47899 KZ |
2022 | bitmap_and_into (regs_set, regs_live); |
2023 | if (!bitmap_empty_p (regs_set)) | |
2024 | { | |
456610d3 | 2025 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02b47899 | 2026 | { |
2c0fa3ec JJ |
2027 | fprintf (dump_file, "abandoning replacement because sequence " |
2028 | "clobbers live hardregs:"); | |
02b47899 KZ |
2029 | df_print_regset (dump_file, regs_set); |
2030 | } | |
b8698a0f | 2031 | |
02b47899 KZ |
2032 | BITMAP_FREE (regs_set); |
2033 | return false; | |
2034 | } | |
2035 | BITMAP_FREE (regs_set); | |
2036 | } | |
2037 | ||
2c0fa3ec JJ |
2038 | subrtx_iterator::array_type array; |
2039 | FOR_EACH_SUBRTX (iter, array, *loc, NONCONST) | |
2040 | { | |
2041 | const_rtx x = *iter; | |
2042 | if (GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) | |
2043 | { | |
2044 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2045 | fprintf (dump_file, " -- replacing the MEM failed due to address " | |
2046 | "side-effects\n"); | |
2047 | return false; | |
2048 | } | |
2049 | } | |
2050 | ||
8660aaae | 2051 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
6fb5fa3c | 2052 | { |
fd6daec9 | 2053 | deferred_change *change = deferred_change_pool.allocate (); |
b8698a0f | 2054 | |
8660aaae EC |
2055 | /* Insert this right before the store insn where it will be safe |
2056 | from later insns that might change it before the read. */ | |
2057 | emit_insn_before (insns, store_insn->insn); | |
b8698a0f | 2058 | |
8660aaae EC |
2059 | /* And now for the kludge part: cselib croaks if you just |
2060 | return at this point. There are two reasons for this: | |
b8698a0f | 2061 | |
8660aaae EC |
2062 | 1) Cselib has an idea of how many pseudos there are and |
2063 | that does not include the new ones we just added. | |
b8698a0f | 2064 | |
8660aaae EC |
2065 | 2) Cselib does not know about the move insn we added |
2066 | above the store_info, and there is no way to tell it | |
2067 | about it, because it has "moved on". | |
b8698a0f | 2068 | |
8660aaae EC |
2069 | Problem (1) is fixable with a certain amount of engineering. |
2070 | Problem (2) is requires starting the bb from scratch. This | |
2071 | could be expensive. | |
b8698a0f | 2072 | |
8660aaae EC |
2073 | So we are just going to have to lie. The move/extraction |
2074 | insns are not really an issue, cselib did not see them. But | |
2075 | the use of the new pseudo read_insn is a real problem because | |
2076 | cselib has not scanned this insn. The way that we solve this | |
2077 | problem is that we are just going to put the mem back for now | |
2078 | and when we are finished with the block, we undo this. We | |
2079 | keep a table of mems to get rid of. At the end of the basic | |
2080 | block we can put them back. */ | |
b8698a0f | 2081 | |
8660aaae | 2082 | *loc = read_info->mem; |
da6603c6 ML |
2083 | change->next = deferred_change_list; |
2084 | deferred_change_list = change; | |
2085 | change->loc = loc; | |
2086 | change->reg = read_reg; | |
b8698a0f | 2087 | |
8660aaae EC |
2088 | /* Get rid of the read_info, from the point of view of the |
2089 | rest of dse, play like this read never happened. */ | |
2090 | read_insn->read_rec = read_info->next; | |
fb0b2914 | 2091 | read_info_type_pool.remove (read_info); |
456610d3 | 2092 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2093 | { |
2094 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2095 | print_simple_rtl (dump_file, read_reg); | |
2096 | fprintf (dump_file, "\n"); | |
2097 | } | |
8660aaae | 2098 | return true; |
6fb5fa3c | 2099 | } |
b8698a0f | 2100 | else |
6fb5fa3c | 2101 | { |
456610d3 | 2102 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2103 | { |
2104 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2105 | print_simple_rtl (dump_file, read_reg); | |
2106 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2107 | } | |
6fb5fa3c DB |
2108 | return false; |
2109 | } | |
2110 | } | |
2111 | ||
d7111da8 RS |
2112 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2113 | be active. */ | |
6fb5fa3c | 2114 | |
d7111da8 RS |
2115 | static void |
2116 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info) | |
6fb5fa3c | 2117 | { |
6216f94e | 2118 | rtx mem = *loc, mem_addr; |
6fb5fa3c | 2119 | insn_info_t insn_info; |
02ce5d90 RS |
2120 | poly_int64 offset = 0; |
2121 | poly_int64 width = 0; | |
b8698a0f | 2122 | cselib_val *base = NULL; |
6fb5fa3c DB |
2123 | int group_id; |
2124 | read_info_t read_info; | |
2125 | ||
6fb5fa3c DB |
2126 | insn_info = bb_info->last_insn; |
2127 | ||
2128 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
a49a975f | 2129 | || MEM_VOLATILE_P (mem)) |
6fb5fa3c | 2130 | { |
a49a975f JJ |
2131 | if (crtl->stack_protect_guard |
2132 | && (MEM_EXPR (mem) == crtl->stack_protect_guard | |
2133 | || (crtl->stack_protect_guard_decl | |
2134 | && MEM_EXPR (mem) == crtl->stack_protect_guard_decl)) | |
2135 | && MEM_VOLATILE_P (mem)) | |
2136 | { | |
2137 | /* This is either the stack protector canary on the stack, | |
2138 | which ought to be written by a MEM_VOLATILE_P store and | |
2139 | thus shouldn't be deleted and is read at the very end of | |
2140 | function, but shouldn't conflict with any other store. | |
2141 | Or it is __stack_chk_guard variable or TLS or whatever else | |
2142 | MEM holding the canary value, which really shouldn't be | |
2143 | ever modified in -fstack-protector* protected functions, | |
2144 | otherwise the prologue store wouldn't match the epilogue | |
2145 | check. */ | |
2146 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2147 | fprintf (dump_file, " stack protector canary read ignored.\n"); | |
2148 | insn_info->cannot_delete = true; | |
2149 | return; | |
2150 | } | |
2151 | ||
456610d3 | 2152 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2153 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
2154 | add_wild_read (bb_info); | |
2155 | insn_info->cannot_delete = true; | |
d7111da8 | 2156 | return; |
6fb5fa3c DB |
2157 | } |
2158 | ||
2159 | /* If it is reading readonly mem, then there can be no conflict with | |
2160 | another write. */ | |
2161 | if (MEM_READONLY_P (mem)) | |
d7111da8 | 2162 | return; |
6fb5fa3c | 2163 | |
ac6929b5 | 2164 | if (!canon_address (mem, &group_id, &offset, &base)) |
6fb5fa3c | 2165 | { |
456610d3 | 2166 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2167 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
2168 | add_wild_read (bb_info); | |
d7111da8 | 2169 | return; |
6fb5fa3c DB |
2170 | } |
2171 | ||
2172 | if (GET_MODE (mem) == BLKmode) | |
2173 | width = -1; | |
2174 | else | |
2175 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2176 | ||
02ce5d90 | 2177 | if (!endpoint_representable_p (offset, known_eq (width, -1) ? 1 : width)) |
938f9248 | 2178 | { |
213ffde2 JJ |
2179 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2180 | fprintf (dump_file, " adding wild read, due to overflow.\n"); | |
2181 | add_wild_read (bb_info); | |
938f9248 ML |
2182 | return; |
2183 | } | |
2184 | ||
fb0b2914 | 2185 | read_info = read_info_type_pool.allocate (); |
6fb5fa3c DB |
2186 | read_info->group_id = group_id; |
2187 | read_info->mem = mem; | |
71d6a386 RS |
2188 | read_info->offset = offset; |
2189 | read_info->width = width; | |
6fb5fa3c DB |
2190 | read_info->next = insn_info->read_rec; |
2191 | insn_info->read_rec = read_info; | |
ac6929b5 RB |
2192 | if (group_id < 0) |
2193 | mem_addr = base->val_rtx; | |
6216f94e JJ |
2194 | else |
2195 | { | |
ac6929b5 RB |
2196 | group_info *group = rtx_group_vec[group_id]; |
2197 | mem_addr = group->canon_base_addr; | |
6216f94e | 2198 | } |
02ce5d90 | 2199 | if (maybe_ne (offset, 0)) |
ac6929b5 | 2200 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6fb5fa3c | 2201 | |
ac6929b5 | 2202 | if (group_id >= 0) |
6fb5fa3c DB |
2203 | { |
2204 | /* This is the restricted case where the base is a constant or | |
2205 | the frame pointer and offset is a constant. */ | |
2206 | insn_info_t i_ptr = active_local_stores; | |
2207 | insn_info_t last = NULL; | |
b8698a0f | 2208 | |
456610d3 | 2209 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2210 | { |
02ce5d90 | 2211 | if (!known_size_p (width)) |
6fb5fa3c DB |
2212 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", |
2213 | group_id); | |
2214 | else | |
71d6a386 RS |
2215 | { |
2216 | fprintf (dump_file, " processing const load gid=%d", group_id); | |
2217 | print_range (dump_file, offset, width); | |
2218 | fprintf (dump_file, "\n"); | |
2219 | } | |
6fb5fa3c DB |
2220 | } |
2221 | ||
2222 | while (i_ptr) | |
2223 | { | |
2224 | bool remove = false; | |
fd6daec9 | 2225 | store_info *store_info = i_ptr->store_rec; |
b8698a0f | 2226 | |
6fb5fa3c DB |
2227 | /* Skip the clobbers. */ |
2228 | while (!store_info->is_set) | |
2229 | store_info = store_info->next; | |
b8698a0f | 2230 | |
6fb5fa3c DB |
2231 | /* There are three cases here. */ |
2232 | if (store_info->group_id < 0) | |
2233 | /* We have a cselib store followed by a read from a | |
2234 | const base. */ | |
b8698a0f L |
2235 | remove |
2236 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2237 | GET_MODE (store_info->mem), |
2238 | store_info->mem_addr, | |
53d9622b | 2239 | mem, mem_addr); |
b8698a0f | 2240 | |
6fb5fa3c DB |
2241 | else if (group_id == store_info->group_id) |
2242 | { | |
2243 | /* This is a block mode load. We may get lucky and | |
2244 | canon_true_dependence may save the day. */ | |
02ce5d90 | 2245 | if (!known_size_p (width)) |
b8698a0f L |
2246 | remove |
2247 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2248 | GET_MODE (store_info->mem), |
2249 | store_info->mem_addr, | |
53d9622b | 2250 | mem, mem_addr); |
b8698a0f | 2251 | |
6fb5fa3c DB |
2252 | /* If this read is just reading back something that we just |
2253 | stored, rewrite the read. */ | |
b8698a0f | 2254 | else |
6fb5fa3c DB |
2255 | { |
2256 | if (store_info->rhs | |
71d6a386 RS |
2257 | && known_subrange_p (offset, width, store_info->offset, |
2258 | store_info->width) | |
8dd5516b | 2259 | && all_positions_needed_p (store_info, |
71d6a386 | 2260 | offset - store_info->offset, |
8dd5516b JJ |
2261 | width) |
2262 | && replace_read (store_info, i_ptr, read_info, | |
2263 | insn_info, loc, bb_info->regs_live)) | |
d7111da8 | 2264 | return; |
8dd5516b | 2265 | |
6fb5fa3c | 2266 | /* The bases are the same, just see if the offsets |
71d6a386 RS |
2267 | could overlap. */ |
2268 | if (ranges_maybe_overlap_p (offset, width, | |
2269 | store_info->offset, | |
2270 | store_info->width)) | |
6fb5fa3c DB |
2271 | remove = true; |
2272 | } | |
2273 | } | |
b8698a0f L |
2274 | |
2275 | /* else | |
6fb5fa3c DB |
2276 | The else case that is missing here is that the |
2277 | bases are constant but different. There is nothing | |
2278 | to do here because there is no overlap. */ | |
b8698a0f | 2279 | |
6fb5fa3c DB |
2280 | if (remove) |
2281 | { | |
456610d3 | 2282 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2283 | dump_insn_info ("removing from active", i_ptr); |
2284 | ||
dabd47e7 | 2285 | active_local_stores_len--; |
6fb5fa3c DB |
2286 | if (last) |
2287 | last->next_local_store = i_ptr->next_local_store; | |
2288 | else | |
2289 | active_local_stores = i_ptr->next_local_store; | |
2290 | } | |
2291 | else | |
2292 | last = i_ptr; | |
2293 | i_ptr = i_ptr->next_local_store; | |
2294 | } | |
2295 | } | |
b8698a0f | 2296 | else |
6fb5fa3c DB |
2297 | { |
2298 | insn_info_t i_ptr = active_local_stores; | |
2299 | insn_info_t last = NULL; | |
456610d3 | 2300 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2301 | { |
2302 | fprintf (dump_file, " processing cselib load mem:"); | |
2303 | print_inline_rtx (dump_file, mem, 0); | |
2304 | fprintf (dump_file, "\n"); | |
2305 | } | |
2306 | ||
2307 | while (i_ptr) | |
2308 | { | |
2309 | bool remove = false; | |
fd6daec9 | 2310 | store_info *store_info = i_ptr->store_rec; |
b8698a0f | 2311 | |
456610d3 | 2312 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2313 | fprintf (dump_file, " processing cselib load against insn %d\n", |
2314 | INSN_UID (i_ptr->insn)); | |
2315 | ||
2316 | /* Skip the clobbers. */ | |
2317 | while (!store_info->is_set) | |
2318 | store_info = store_info->next; | |
2319 | ||
2320 | /* If this read is just reading back something that we just | |
2321 | stored, rewrite the read. */ | |
2322 | if (store_info->rhs | |
2323 | && store_info->group_id == -1 | |
2324 | && store_info->cse_base == base | |
71d6a386 RS |
2325 | && known_subrange_p (offset, width, store_info->offset, |
2326 | store_info->width) | |
8dd5516b | 2327 | && all_positions_needed_p (store_info, |
71d6a386 | 2328 | offset - store_info->offset, width) |
8dd5516b JJ |
2329 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, |
2330 | bb_info->regs_live)) | |
d7111da8 | 2331 | return; |
6fb5fa3c | 2332 | |
ac6929b5 RB |
2333 | remove = canon_true_dependence (store_info->mem, |
2334 | GET_MODE (store_info->mem), | |
2335 | store_info->mem_addr, | |
2336 | mem, mem_addr); | |
b8698a0f | 2337 | |
6fb5fa3c DB |
2338 | if (remove) |
2339 | { | |
456610d3 | 2340 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2341 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2342 | |
dabd47e7 | 2343 | active_local_stores_len--; |
6fb5fa3c DB |
2344 | if (last) |
2345 | last->next_local_store = i_ptr->next_local_store; | |
2346 | else | |
2347 | active_local_stores = i_ptr->next_local_store; | |
2348 | } | |
2349 | else | |
2350 | last = i_ptr; | |
2351 | i_ptr = i_ptr->next_local_store; | |
2352 | } | |
2353 | } | |
6fb5fa3c DB |
2354 | } |
2355 | ||
d7111da8 | 2356 | /* A note_uses callback in which DATA points the INSN_INFO for |
6fb5fa3c DB |
2357 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2358 | true for any part of *LOC. */ | |
2359 | ||
2360 | static void | |
2361 | check_mem_read_use (rtx *loc, void *data) | |
2362 | { | |
d7111da8 RS |
2363 | subrtx_ptr_iterator::array_type array; |
2364 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
2365 | { | |
2366 | rtx *loc = *iter; | |
2367 | if (MEM_P (*loc)) | |
2368 | check_mem_read_rtx (loc, (bb_info_t) data); | |
2369 | } | |
6fb5fa3c DB |
2370 | } |
2371 | ||
8dd5516b JJ |
2372 | |
2373 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2374 | So far it only handles arguments passed in registers. */ | |
2375 | ||
2376 | static bool | |
2377 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2378 | { | |
d5cc9181 JR |
2379 | CUMULATIVE_ARGS args_so_far_v; |
2380 | cumulative_args_t args_so_far; | |
8dd5516b JJ |
2381 | tree arg; |
2382 | int idx; | |
2383 | ||
d5cc9181 JR |
2384 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2385 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
8dd5516b JJ |
2386 | |
2387 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2388 | for (idx = 0; | |
2389 | arg != void_list_node && idx < nargs; | |
2390 | arg = TREE_CHAIN (arg), idx++) | |
2391 | { | |
b4206259 | 2392 | scalar_int_mode mode; |
3c07301f | 2393 | rtx reg, link, tmp; |
b4206259 RS |
2394 | |
2395 | if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg)), &mode)) | |
2396 | return false; | |
2397 | ||
6783fdb7 RS |
2398 | function_arg_info arg (mode, /*named=*/true); |
2399 | reg = targetm.calls.function_arg (args_so_far, arg); | |
b4206259 | 2400 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode) |
8dd5516b JJ |
2401 | return false; |
2402 | ||
2403 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2404 | link; | |
2405 | link = XEXP (link, 1)) | |
2406 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2407 | { | |
b4206259 | 2408 | scalar_int_mode arg_mode; |
8dd5516b JJ |
2409 | args[idx] = XEXP (XEXP (link, 0), 0); |
2410 | if (REG_P (args[idx]) | |
2411 | && REGNO (args[idx]) == REGNO (reg) | |
2412 | && (GET_MODE (args[idx]) == mode | |
b4206259 RS |
2413 | || (is_int_mode (GET_MODE (args[idx]), &arg_mode) |
2414 | && (GET_MODE_SIZE (arg_mode) <= UNITS_PER_WORD) | |
2415 | && (GET_MODE_SIZE (arg_mode) > GET_MODE_SIZE (mode))))) | |
8dd5516b JJ |
2416 | break; |
2417 | } | |
2418 | if (!link) | |
2419 | return false; | |
2420 | ||
2421 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2422 | if (GET_MODE (args[idx]) != mode) | |
2423 | { | |
2424 | if (!tmp || !CONST_INT_P (tmp)) | |
2425 | return false; | |
6d26322f | 2426 | tmp = gen_int_mode (INTVAL (tmp), mode); |
8dd5516b JJ |
2427 | } |
2428 | if (tmp) | |
2429 | args[idx] = tmp; | |
2430 | ||
6930c98c | 2431 | targetm.calls.function_arg_advance (args_so_far, arg); |
8dd5516b JJ |
2432 | } |
2433 | if (arg != void_list_node || idx != nargs) | |
2434 | return false; | |
2435 | return true; | |
2436 | } | |
2437 | ||
9e582b1d JR |
2438 | /* Return a bitmap of the fixed registers contained in IN. */ |
2439 | ||
2440 | static bitmap | |
2441 | copy_fixed_regs (const_bitmap in) | |
2442 | { | |
2443 | bitmap ret; | |
2444 | ||
2445 | ret = ALLOC_REG_SET (NULL); | |
0b0310e9 | 2446 | bitmap_and (ret, in, bitmap_view<HARD_REG_SET> (fixed_reg_set)); |
9e582b1d JR |
2447 | return ret; |
2448 | } | |
8dd5516b | 2449 | |
6fb5fa3c DB |
2450 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2451 | if some part of it is not a candidate store and assigns to a | |
2452 | non-register target. */ | |
2453 | ||
2454 | static void | |
fc2d7303 | 2455 | scan_insn (bb_info_t bb_info, rtx_insn *insn, int max_active_local_stores) |
6fb5fa3c DB |
2456 | { |
2457 | rtx body; | |
fb0b2914 | 2458 | insn_info_type *insn_info = insn_info_type_pool.allocate (); |
6fb5fa3c | 2459 | int mems_found = 0; |
da6603c6 | 2460 | memset (insn_info, 0, sizeof (struct insn_info_type)); |
6fb5fa3c | 2461 | |
456610d3 | 2462 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2463 | fprintf (dump_file, "\n**scanning insn=%d\n", |
2464 | INSN_UID (insn)); | |
2465 | ||
2466 | insn_info->prev_insn = bb_info->last_insn; | |
2467 | insn_info->insn = insn; | |
2468 | bb_info->last_insn = insn_info; | |
b8698a0f | 2469 | |
b5b8b0ac AO |
2470 | if (DEBUG_INSN_P (insn)) |
2471 | { | |
2472 | insn_info->cannot_delete = true; | |
2473 | return; | |
2474 | } | |
6fb5fa3c | 2475 | |
6fb5fa3c DB |
2476 | /* Look at all of the uses in the insn. */ |
2477 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2478 | ||
2479 | if (CALL_P (insn)) | |
2480 | { | |
8dd5516b | 2481 | bool const_call; |
ee516de9 | 2482 | rtx call, sym; |
8dd5516b JJ |
2483 | tree memset_call = NULL_TREE; |
2484 | ||
6fb5fa3c | 2485 | insn_info->cannot_delete = true; |
50f0f366 | 2486 | |
6fb5fa3c | 2487 | /* Const functions cannot do anything bad i.e. read memory, |
50f0f366 | 2488 | however, they can read their parameters which may have |
8dd5516b JJ |
2489 | been pushed onto the stack. |
2490 | memset and bzero don't read memory either. */ | |
2491 | const_call = RTL_CONST_CALL_P (insn); | |
ee516de9 EB |
2492 | if (!const_call |
2493 | && (call = get_call_rtx_from (insn)) | |
2494 | && (sym = XEXP (XEXP (call, 0), 0)) | |
2495 | && GET_CODE (sym) == SYMBOL_REF | |
2496 | && SYMBOL_REF_DECL (sym) | |
2497 | && TREE_CODE (SYMBOL_REF_DECL (sym)) == FUNCTION_DECL | |
3d78e008 | 2498 | && fndecl_built_in_p (SYMBOL_REF_DECL (sym), BUILT_IN_MEMSET)) |
ee516de9 EB |
2499 | memset_call = SYMBOL_REF_DECL (sym); |
2500 | ||
8dd5516b | 2501 | if (const_call || memset_call) |
6fb5fa3c DB |
2502 | { |
2503 | insn_info_t i_ptr = active_local_stores; | |
2504 | insn_info_t last = NULL; | |
2505 | ||
456610d3 | 2506 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2507 | fprintf (dump_file, "%s call %d\n", |
2508 | const_call ? "const" : "memset", INSN_UID (insn)); | |
6fb5fa3c | 2509 | |
64520bdc | 2510 | /* See the head comment of the frame_read field. */ |
57534689 JJ |
2511 | if (reload_completed |
2512 | /* Tail calls are storing their arguments using | |
2513 | arg pointer. If it is a frame pointer on the target, | |
2514 | even before reload we need to kill frame pointer based | |
2515 | stores. */ | |
2516 | || (SIBLING_CALL_P (insn) | |
2517 | && HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
64520bdc EB |
2518 | insn_info->frame_read = true; |
2519 | ||
2520 | /* Loop over the active stores and remove those which are | |
2521 | killed by the const function call. */ | |
6fb5fa3c DB |
2522 | while (i_ptr) |
2523 | { | |
64520bdc EB |
2524 | bool remove_store = false; |
2525 | ||
2526 | /* The stack pointer based stores are always killed. */ | |
50f0f366 | 2527 | if (i_ptr->stack_pointer_based) |
64520bdc EB |
2528 | remove_store = true; |
2529 | ||
2530 | /* If the frame is read, the frame related stores are killed. */ | |
2531 | else if (insn_info->frame_read) | |
2532 | { | |
fd6daec9 | 2533 | store_info *store_info = i_ptr->store_rec; |
64520bdc EB |
2534 | |
2535 | /* Skip the clobbers. */ | |
2536 | while (!store_info->is_set) | |
2537 | store_info = store_info->next; | |
2538 | ||
2539 | if (store_info->group_id >= 0 | |
9771b263 | 2540 | && rtx_group_vec[store_info->group_id]->frame_related) |
64520bdc EB |
2541 | remove_store = true; |
2542 | } | |
2543 | ||
2544 | if (remove_store) | |
6fb5fa3c | 2545 | { |
456610d3 | 2546 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2547 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2548 | |
dabd47e7 | 2549 | active_local_stores_len--; |
6fb5fa3c DB |
2550 | if (last) |
2551 | last->next_local_store = i_ptr->next_local_store; | |
2552 | else | |
2553 | active_local_stores = i_ptr->next_local_store; | |
2554 | } | |
2555 | else | |
2556 | last = i_ptr; | |
64520bdc | 2557 | |
6fb5fa3c DB |
2558 | i_ptr = i_ptr->next_local_store; |
2559 | } | |
8dd5516b JJ |
2560 | |
2561 | if (memset_call) | |
2562 | { | |
2563 | rtx args[3]; | |
2564 | if (get_call_args (insn, memset_call, args, 3) | |
2565 | && CONST_INT_P (args[1]) | |
2566 | && CONST_INT_P (args[2]) | |
2567 | && INTVAL (args[2]) > 0) | |
2568 | { | |
2569 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
f5541398 | 2570 | set_mem_size (mem, INTVAL (args[2])); |
f7df4a84 | 2571 | body = gen_rtx_SET (mem, args[1]); |
8dd5516b | 2572 | mems_found += record_store (body, bb_info); |
456610d3 | 2573 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2574 | fprintf (dump_file, "handling memset as BLKmode store\n"); |
2575 | if (mems_found == 1) | |
2576 | { | |
fc2d7303 | 2577 | if (active_local_stores_len++ >= max_active_local_stores) |
dabd47e7 JJ |
2578 | { |
2579 | active_local_stores_len = 1; | |
2580 | active_local_stores = NULL; | |
2581 | } | |
9e582b1d JR |
2582 | insn_info->fixed_regs_live |
2583 | = copy_fixed_regs (bb_info->regs_live); | |
8dd5516b JJ |
2584 | insn_info->next_local_store = active_local_stores; |
2585 | active_local_stores = insn_info; | |
2586 | } | |
2587 | } | |
a2ad54a0 JJ |
2588 | else |
2589 | clear_rhs_from_active_local_stores (); | |
8dd5516b | 2590 | } |
6fb5fa3c | 2591 | } |
7ee98586 EB |
2592 | else if (SIBLING_CALL_P (insn) |
2593 | && (reload_completed || HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
57534689 JJ |
2594 | /* Arguments for a sibling call that are pushed to memory are passed |
2595 | using the incoming argument pointer of the current function. After | |
7ee98586 EB |
2596 | reload that might be (and likely is) frame pointer based. And, if |
2597 | it is a frame pointer on the target, even before reload we need to | |
2598 | kill frame pointer based stores. */ | |
57534689 | 2599 | add_wild_read (bb_info); |
50f0f366 | 2600 | else |
d26c7090 ER |
2601 | /* Every other call, including pure functions, may read any memory |
2602 | that is not relative to the frame. */ | |
2603 | add_non_frame_wild_read (bb_info); | |
50f0f366 | 2604 | |
6fb5fa3c DB |
2605 | return; |
2606 | } | |
2607 | ||
2608 | /* Assuming that there are sets in these insns, we cannot delete | |
2609 | them. */ | |
2610 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
0a64eeca | 2611 | || volatile_refs_p (PATTERN (insn)) |
2da02156 | 2612 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
6fb5fa3c DB |
2613 | || (RTX_FRAME_RELATED_P (insn)) |
2614 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2615 | insn_info->cannot_delete = true; | |
b8698a0f | 2616 | |
6fb5fa3c DB |
2617 | body = PATTERN (insn); |
2618 | if (GET_CODE (body) == PARALLEL) | |
2619 | { | |
2620 | int i; | |
2621 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2622 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2623 | } | |
2624 | else | |
2625 | mems_found += record_store (body, bb_info); | |
2626 | ||
456610d3 | 2627 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2628 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
6fb5fa3c DB |
2629 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2630 | ||
8dd5516b JJ |
2631 | /* If we found some sets of mems, add it into the active_local_stores so |
2632 | that it can be locally deleted if found dead or used for | |
2633 | replace_read and redundant constant store elimination. Otherwise mark | |
2634 | it as cannot delete. This simplifies the processing later. */ | |
2635 | if (mems_found == 1) | |
6fb5fa3c | 2636 | { |
fc2d7303 | 2637 | if (active_local_stores_len++ >= max_active_local_stores) |
dabd47e7 JJ |
2638 | { |
2639 | active_local_stores_len = 1; | |
2640 | active_local_stores = NULL; | |
2641 | } | |
9e582b1d | 2642 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
6fb5fa3c DB |
2643 | insn_info->next_local_store = active_local_stores; |
2644 | active_local_stores = insn_info; | |
2645 | } | |
2646 | else | |
2647 | insn_info->cannot_delete = true; | |
2648 | } | |
2649 | ||
2650 | ||
2651 | /* Remove BASE from the set of active_local_stores. This is a | |
2652 | callback from cselib that is used to get rid of the stores in | |
2653 | active_local_stores. */ | |
2654 | ||
2655 | static void | |
2656 | remove_useless_values (cselib_val *base) | |
2657 | { | |
2658 | insn_info_t insn_info = active_local_stores; | |
2659 | insn_info_t last = NULL; | |
2660 | ||
2661 | while (insn_info) | |
2662 | { | |
fd6daec9 | 2663 | store_info *store_info = insn_info->store_rec; |
60564289 | 2664 | bool del = false; |
6fb5fa3c DB |
2665 | |
2666 | /* If ANY of the store_infos match the cselib group that is | |
67914693 | 2667 | being deleted, then the insn cannot be deleted. */ |
6fb5fa3c DB |
2668 | while (store_info) |
2669 | { | |
b8698a0f | 2670 | if ((store_info->group_id == -1) |
6fb5fa3c DB |
2671 | && (store_info->cse_base == base)) |
2672 | { | |
60564289 | 2673 | del = true; |
6fb5fa3c DB |
2674 | break; |
2675 | } | |
2676 | store_info = store_info->next; | |
2677 | } | |
2678 | ||
60564289 | 2679 | if (del) |
6fb5fa3c | 2680 | { |
dabd47e7 | 2681 | active_local_stores_len--; |
6fb5fa3c DB |
2682 | if (last) |
2683 | last->next_local_store = insn_info->next_local_store; | |
2684 | else | |
2685 | active_local_stores = insn_info->next_local_store; | |
2686 | free_store_info (insn_info); | |
2687 | } | |
2688 | else | |
2689 | last = insn_info; | |
b8698a0f | 2690 | |
6fb5fa3c DB |
2691 | insn_info = insn_info->next_local_store; |
2692 | } | |
2693 | } | |
2694 | ||
2695 | ||
2696 | /* Do all of step 1. */ | |
2697 | ||
2698 | static void | |
2699 | dse_step1 (void) | |
2700 | { | |
2701 | basic_block bb; | |
3f9b14ff | 2702 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
b8698a0f | 2703 | |
457eeaae | 2704 | cselib_init (0); |
6fb5fa3c DB |
2705 | all_blocks = BITMAP_ALLOC (NULL); |
2706 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2707 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2708 | ||
fc2d7303 RB |
2709 | /* For -O1 reduce the maximum number of active local stores for RTL DSE |
2710 | since this can consume huge amounts of memory (PR89115). */ | |
028d4092 | 2711 | int max_active_local_stores = param_max_dse_active_local_stores; |
fc2d7303 RB |
2712 | if (optimize < 2) |
2713 | max_active_local_stores /= 10; | |
2714 | ||
04a90bec | 2715 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
2716 | { |
2717 | insn_info_t ptr; | |
fb0b2914 | 2718 | bb_info_t bb_info = dse_bb_info_type_pool.allocate (); |
6fb5fa3c | 2719 | |
da6603c6 | 2720 | memset (bb_info, 0, sizeof (dse_bb_info_type)); |
6fb5fa3c | 2721 | bitmap_set_bit (all_blocks, bb->index); |
02b47899 KZ |
2722 | bb_info->regs_live = regs_live; |
2723 | ||
2724 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2725 | df_simulate_initialize_forwards (bb, regs_live); | |
6fb5fa3c DB |
2726 | |
2727 | bb_table[bb->index] = bb_info; | |
2728 | cselib_discard_hook = remove_useless_values; | |
2729 | ||
2730 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2731 | { | |
dd60a84c | 2732 | rtx_insn *insn; |
6fb5fa3c | 2733 | |
6fb5fa3c | 2734 | active_local_stores = NULL; |
dabd47e7 | 2735 | active_local_stores_len = 0; |
6fb5fa3c | 2736 | cselib_clear_table (); |
b8698a0f | 2737 | |
6fb5fa3c DB |
2738 | /* Scan the insns. */ |
2739 | FOR_BB_INSNS (bb, insn) | |
2740 | { | |
2741 | if (INSN_P (insn)) | |
fc2d7303 | 2742 | scan_insn (bb_info, insn, max_active_local_stores); |
6fb5fa3c | 2743 | cselib_process_insn (insn); |
02b47899 KZ |
2744 | if (INSN_P (insn)) |
2745 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
6fb5fa3c | 2746 | } |
b8698a0f | 2747 | |
6fb5fa3c DB |
2748 | /* This is something of a hack, because the global algorithm |
2749 | is supposed to take care of the case where stores go dead | |
2750 | at the end of the function. However, the global | |
2751 | algorithm must take a more conservative view of block | |
2752 | mode reads than the local alg does. So to get the case | |
2753 | where you have a store to the frame followed by a non | |
0d52bcc1 | 2754 | overlapping block more read, we look at the active local |
6fb5fa3c DB |
2755 | stores at the end of the function and delete all of the |
2756 | frame and spill based ones. */ | |
2757 | if (stores_off_frame_dead_at_return | |
2758 | && (EDGE_COUNT (bb->succs) == 0 | |
2759 | || (single_succ_p (bb) | |
fefa31b5 | 2760 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
e3b5732b | 2761 | && ! crtl->calls_eh_return))) |
6fb5fa3c DB |
2762 | { |
2763 | insn_info_t i_ptr = active_local_stores; | |
2764 | while (i_ptr) | |
2765 | { | |
fd6daec9 | 2766 | store_info *store_info = i_ptr->store_rec; |
6fb5fa3c DB |
2767 | |
2768 | /* Skip the clobbers. */ | |
2769 | while (!store_info->is_set) | |
2770 | store_info = store_info->next; | |
ac6929b5 RB |
2771 | if (store_info->group_id >= 0) |
2772 | { | |
2773 | group_info *group = rtx_group_vec[store_info->group_id]; | |
2774 | if (group->frame_related && !i_ptr->cannot_delete) | |
2775 | delete_dead_store_insn (i_ptr); | |
2776 | } | |
6fb5fa3c DB |
2777 | |
2778 | i_ptr = i_ptr->next_local_store; | |
2779 | } | |
2780 | } | |
2781 | ||
2782 | /* Get rid of the loads that were discovered in | |
2783 | replace_read. Cselib is finished with this block. */ | |
2784 | while (deferred_change_list) | |
2785 | { | |
fd6daec9 | 2786 | deferred_change *next = deferred_change_list->next; |
6fb5fa3c DB |
2787 | |
2788 | /* There is no reason to validate this change. That was | |
2789 | done earlier. */ | |
2790 | *deferred_change_list->loc = deferred_change_list->reg; | |
fb0b2914 | 2791 | deferred_change_pool.remove (deferred_change_list); |
6fb5fa3c DB |
2792 | deferred_change_list = next; |
2793 | } | |
2794 | ||
2795 | /* Get rid of all of the cselib based store_infos in this | |
2796 | block and mark the containing insns as not being | |
2797 | deletable. */ | |
2798 | ptr = bb_info->last_insn; | |
2799 | while (ptr) | |
2800 | { | |
2801 | if (ptr->contains_cselib_groups) | |
8dd5516b | 2802 | { |
fd6daec9 | 2803 | store_info *s_info = ptr->store_rec; |
8dd5516b JJ |
2804 | while (s_info && !s_info->is_set) |
2805 | s_info = s_info->next; | |
2806 | if (s_info | |
2807 | && s_info->redundant_reason | |
2808 | && s_info->redundant_reason->insn | |
2809 | && !ptr->cannot_delete) | |
2810 | { | |
456610d3 | 2811 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2812 | fprintf (dump_file, "Locally deleting insn %d " |
2813 | "because insn %d stores the " | |
2814 | "same value and couldn't be " | |
2815 | "eliminated\n", | |
2816 | INSN_UID (ptr->insn), | |
2817 | INSN_UID (s_info->redundant_reason->insn)); | |
2818 | delete_dead_store_insn (ptr); | |
2819 | } | |
8dd5516b JJ |
2820 | free_store_info (ptr); |
2821 | } | |
2822 | else | |
2823 | { | |
fd6daec9 | 2824 | store_info *s_info; |
8dd5516b JJ |
2825 | |
2826 | /* Free at least positions_needed bitmaps. */ | |
2827 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2828 | if (s_info->is_large) | |
2829 | { | |
dc491a25 | 2830 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
8dd5516b JJ |
2831 | s_info->is_large = false; |
2832 | } | |
2833 | } | |
6fb5fa3c DB |
2834 | ptr = ptr->prev_insn; |
2835 | } | |
2836 | ||
da6603c6 | 2837 | cse_store_info_pool.release (); |
6fb5fa3c | 2838 | } |
02b47899 | 2839 | bb_info->regs_live = NULL; |
6fb5fa3c DB |
2840 | } |
2841 | ||
02b47899 | 2842 | BITMAP_FREE (regs_live); |
6fb5fa3c | 2843 | cselib_finish (); |
c203e8a7 | 2844 | rtx_group_table->empty (); |
6fb5fa3c DB |
2845 | } |
2846 | ||
2847 | \f | |
2848 | /*---------------------------------------------------------------------------- | |
2849 | Second step. | |
2850 | ||
2851 | Assign each byte position in the stores that we are going to | |
2852 | analyze globally to a position in the bitmaps. Returns true if | |
6ed3da00 | 2853 | there are any bit positions assigned. |
6fb5fa3c DB |
2854 | ----------------------------------------------------------------------------*/ |
2855 | ||
2856 | static void | |
2857 | dse_step2_init (void) | |
2858 | { | |
2859 | unsigned int i; | |
fd6daec9 | 2860 | group_info *group; |
6fb5fa3c | 2861 | |
9771b263 | 2862 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2863 | { |
2864 | /* For all non stack related bases, we only consider a store to | |
2865 | be deletable if there are two or more stores for that | |
2866 | position. This is because it takes one store to make the | |
2867 | other store redundant. However, for the stores that are | |
2868 | stack related, we consider them if there is only one store | |
2869 | for the position. We do this because the stack related | |
2870 | stores can be deleted if their is no read between them and | |
2871 | the end of the function. | |
b8698a0f | 2872 | |
6fb5fa3c DB |
2873 | To make this work in the current framework, we take the stack |
2874 | related bases add all of the bits from store1 into store2. | |
2875 | This has the effect of making the eligible even if there is | |
2876 | only one store. */ | |
2877 | ||
2878 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2879 | { | |
2880 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2881 | bitmap_ior_into (group->store2_p, group->store1_p); | |
456610d3 | 2882 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2883 | fprintf (dump_file, "group %d is frame related ", i); |
6fb5fa3c DB |
2884 | } |
2885 | ||
2886 | group->offset_map_size_n++; | |
3f9b14ff SB |
2887 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2888 | group->offset_map_size_n); | |
6fb5fa3c | 2889 | group->offset_map_size_p++; |
3f9b14ff SB |
2890 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2891 | group->offset_map_size_p); | |
6fb5fa3c | 2892 | group->process_globally = false; |
456610d3 | 2893 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2894 | { |
b8698a0f | 2895 | fprintf (dump_file, "group %d(%d+%d): ", i, |
6fb5fa3c DB |
2896 | (int)bitmap_count_bits (group->store2_n), |
2897 | (int)bitmap_count_bits (group->store2_p)); | |
2898 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2899 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2900 | } | |
2901 | } | |
2902 | } | |
2903 | ||
2904 | ||
ac6929b5 | 2905 | /* Init the offset tables. */ |
6fb5fa3c DB |
2906 | |
2907 | static bool | |
ac6929b5 | 2908 | dse_step2 (void) |
6fb5fa3c DB |
2909 | { |
2910 | unsigned int i; | |
fd6daec9 | 2911 | group_info *group; |
6fb5fa3c DB |
2912 | /* Position 0 is unused because 0 is used in the maps to mean |
2913 | unused. */ | |
2914 | current_position = 1; | |
9771b263 | 2915 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2916 | { |
2917 | bitmap_iterator bi; | |
2918 | unsigned int j; | |
2919 | ||
c3284718 RS |
2920 | memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
2921 | memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
6fb5fa3c DB |
2922 | bitmap_clear (group->group_kill); |
2923 | ||
2924 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2925 | { | |
2926 | bitmap_set_bit (group->group_kill, current_position); | |
d26c7090 ER |
2927 | if (bitmap_bit_p (group->escaped_n, j)) |
2928 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2929 | group->offset_map_n[j] = current_position++; |
2930 | group->process_globally = true; | |
2931 | } | |
2932 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2933 | { | |
b8698a0f | 2934 | bitmap_set_bit (group->group_kill, current_position); |
d26c7090 ER |
2935 | if (bitmap_bit_p (group->escaped_p, j)) |
2936 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2937 | group->offset_map_p[j] = current_position++; |
2938 | group->process_globally = true; | |
2939 | } | |
2940 | } | |
2941 | return current_position != 1; | |
2942 | } | |
2943 | ||
2944 | ||
6fb5fa3c DB |
2945 | \f |
2946 | /*---------------------------------------------------------------------------- | |
2947 | Third step. | |
b8698a0f | 2948 | |
6fb5fa3c DB |
2949 | Build the bit vectors for the transfer functions. |
2950 | ----------------------------------------------------------------------------*/ | |
2951 | ||
2952 | ||
6fb5fa3c DB |
2953 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
2954 | there, return 0. */ | |
2955 | ||
2956 | static int | |
fd6daec9 | 2957 | get_bitmap_index (group_info *group_info, HOST_WIDE_INT offset) |
6fb5fa3c DB |
2958 | { |
2959 | if (offset < 0) | |
2960 | { | |
2961 | HOST_WIDE_INT offset_p = -offset; | |
2962 | if (offset_p >= group_info->offset_map_size_n) | |
2963 | return 0; | |
2964 | return group_info->offset_map_n[offset_p]; | |
2965 | } | |
2966 | else | |
2967 | { | |
2968 | if (offset >= group_info->offset_map_size_p) | |
2969 | return 0; | |
2970 | return group_info->offset_map_p[offset]; | |
2971 | } | |
2972 | } | |
2973 | ||
2974 | ||
2975 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
2976 | may be NULL. */ | |
2977 | ||
b8698a0f | 2978 | static void |
ac6929b5 | 2979 | scan_stores (store_info *store_info, bitmap gen, bitmap kill) |
6fb5fa3c DB |
2980 | { |
2981 | while (store_info) | |
2982 | { | |
02ce5d90 | 2983 | HOST_WIDE_INT i, offset, width; |
fd6daec9 | 2984 | group_info *group_info |
9771b263 | 2985 | = rtx_group_vec[store_info->group_id]; |
02ce5d90 RS |
2986 | /* We can (conservatively) ignore stores whose bounds aren't known; |
2987 | they simply don't generate new global dse opportunities. */ | |
2988 | if (group_info->process_globally | |
2989 | && store_info->offset.is_constant (&offset) | |
2990 | && store_info->width.is_constant (&width)) | |
71d6a386 | 2991 | { |
02ce5d90 RS |
2992 | HOST_WIDE_INT end = offset + width; |
2993 | for (i = offset; i < end; i++) | |
71d6a386 RS |
2994 | { |
2995 | int index = get_bitmap_index (group_info, i); | |
2996 | if (index != 0) | |
2997 | { | |
2998 | bitmap_set_bit (gen, index); | |
2999 | if (kill) | |
3000 | bitmap_clear_bit (kill, index); | |
3001 | } | |
3002 | } | |
3003 | } | |
6fb5fa3c DB |
3004 | store_info = store_info->next; |
3005 | } | |
3006 | } | |
3007 | ||
3008 | ||
6fb5fa3c DB |
3009 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL |
3010 | may be NULL. */ | |
3011 | ||
3012 | static void | |
ac6929b5 | 3013 | scan_reads (insn_info_t insn_info, bitmap gen, bitmap kill) |
6fb5fa3c DB |
3014 | { |
3015 | read_info_t read_info = insn_info->read_rec; | |
3016 | int i; | |
fd6daec9 | 3017 | group_info *group; |
6fb5fa3c | 3018 | |
64520bdc EB |
3019 | /* If this insn reads the frame, kill all the frame related stores. */ |
3020 | if (insn_info->frame_read) | |
3021 | { | |
9771b263 | 3022 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
64520bdc EB |
3023 | if (group->process_globally && group->frame_related) |
3024 | { | |
3025 | if (kill) | |
3026 | bitmap_ior_into (kill, group->group_kill); | |
b8698a0f | 3027 | bitmap_and_compl_into (gen, group->group_kill); |
64520bdc EB |
3028 | } |
3029 | } | |
d26c7090 ER |
3030 | if (insn_info->non_frame_wild_read) |
3031 | { | |
3032 | /* Kill all non-frame related stores. Kill all stores of variables that | |
3033 | escape. */ | |
3034 | if (kill) | |
3035 | bitmap_ior_into (kill, kill_on_calls); | |
3036 | bitmap_and_compl_into (gen, kill_on_calls); | |
9771b263 | 3037 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
d26c7090 ER |
3038 | if (group->process_globally && !group->frame_related) |
3039 | { | |
3040 | if (kill) | |
3041 | bitmap_ior_into (kill, group->group_kill); | |
3042 | bitmap_and_compl_into (gen, group->group_kill); | |
3043 | } | |
3044 | } | |
6fb5fa3c DB |
3045 | while (read_info) |
3046 | { | |
9771b263 | 3047 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3048 | { |
3049 | if (group->process_globally) | |
3050 | { | |
3051 | if (i == read_info->group_id) | |
3052 | { | |
02ce5d90 RS |
3053 | HOST_WIDE_INT offset, width; |
3054 | /* Reads with non-constant size kill all DSE opportunities | |
3055 | in the group. */ | |
3056 | if (!read_info->offset.is_constant (&offset) | |
3057 | || !read_info->width.is_constant (&width) | |
3058 | || !known_size_p (width)) | |
6fb5fa3c | 3059 | { |
71d6a386 | 3060 | /* Handle block mode reads. */ |
6fb5fa3c DB |
3061 | if (kill) |
3062 | bitmap_ior_into (kill, group->group_kill); | |
3063 | bitmap_and_compl_into (gen, group->group_kill); | |
3064 | } | |
3065 | else | |
3066 | { | |
3067 | /* The groups are the same, just process the | |
3068 | offsets. */ | |
3069 | HOST_WIDE_INT j; | |
02ce5d90 RS |
3070 | HOST_WIDE_INT end = offset + width; |
3071 | for (j = offset; j < end; j++) | |
6fb5fa3c DB |
3072 | { |
3073 | int index = get_bitmap_index (group, j); | |
3074 | if (index != 0) | |
3075 | { | |
3076 | if (kill) | |
3077 | bitmap_set_bit (kill, index); | |
3078 | bitmap_clear_bit (gen, index); | |
3079 | } | |
3080 | } | |
3081 | } | |
3082 | } | |
3083 | else | |
3084 | { | |
3085 | /* The groups are different, if the alias sets | |
3086 | conflict, clear the entire group. We only need | |
3087 | to apply this test if the read_info is a cselib | |
3088 | read. Anything with a constant base cannot alias | |
3089 | something else with a different constant | |
3090 | base. */ | |
3091 | if ((read_info->group_id < 0) | |
b8698a0f | 3092 | && canon_true_dependence (group->base_mem, |
d32f725a | 3093 | GET_MODE (group->base_mem), |
6216f94e | 3094 | group->canon_base_addr, |
53d9622b | 3095 | read_info->mem, NULL_RTX)) |
6fb5fa3c DB |
3096 | { |
3097 | if (kill) | |
3098 | bitmap_ior_into (kill, group->group_kill); | |
3099 | bitmap_and_compl_into (gen, group->group_kill); | |
3100 | } | |
3101 | } | |
3102 | } | |
3103 | } | |
b8698a0f | 3104 | |
6fb5fa3c DB |
3105 | read_info = read_info->next; |
3106 | } | |
3107 | } | |
3108 | ||
6fb5fa3c DB |
3109 | |
3110 | /* Return the insn in BB_INFO before the first wild read or if there | |
3111 | are no wild reads in the block, return the last insn. */ | |
3112 | ||
3113 | static insn_info_t | |
3114 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3115 | { | |
3116 | insn_info_t insn_info = bb_info->last_insn; | |
3117 | insn_info_t last_wild_read = NULL; | |
3118 | ||
3119 | while (insn_info) | |
3120 | { | |
3121 | if (insn_info->wild_read) | |
3122 | { | |
3123 | last_wild_read = insn_info->prev_insn; | |
3124 | /* Block starts with wild read. */ | |
3125 | if (!last_wild_read) | |
3126 | return NULL; | |
3127 | } | |
3128 | ||
3129 | insn_info = insn_info->prev_insn; | |
3130 | } | |
3131 | ||
3132 | if (last_wild_read) | |
3133 | return last_wild_read; | |
3134 | else | |
3135 | return bb_info->last_insn; | |
3136 | } | |
3137 | ||
3138 | ||
3139 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3140 | the block in order to build the gen and kill sets for the block. | |
3141 | We start at ptr which may be the last insn in the block or may be | |
3142 | the first insn with a wild read. In the latter case we are able to | |
3143 | skip the rest of the block because it just does not matter: | |
3144 | anything that happens is hidden by the wild read. */ | |
3145 | ||
3146 | static void | |
ac6929b5 | 3147 | dse_step3_scan (basic_block bb) |
6fb5fa3c DB |
3148 | { |
3149 | bb_info_t bb_info = bb_table[bb->index]; | |
3150 | insn_info_t insn_info; | |
3151 | ||
ac6929b5 | 3152 | insn_info = find_insn_before_first_wild_read (bb_info); |
b8698a0f | 3153 | |
6fb5fa3c DB |
3154 | /* In the spill case or in the no_spill case if there is no wild |
3155 | read in the block, we will need a kill set. */ | |
3156 | if (insn_info == bb_info->last_insn) | |
3157 | { | |
3158 | if (bb_info->kill) | |
3159 | bitmap_clear (bb_info->kill); | |
3160 | else | |
3f9b14ff | 3161 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c | 3162 | } |
b8698a0f | 3163 | else |
6fb5fa3c DB |
3164 | if (bb_info->kill) |
3165 | BITMAP_FREE (bb_info->kill); | |
3166 | ||
3167 | while (insn_info) | |
3168 | { | |
3169 | /* There may have been code deleted by the dce pass run before | |
3170 | this phase. */ | |
3171 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3172 | { | |
ac6929b5 RB |
3173 | scan_stores (insn_info->store_rec, bb_info->gen, bb_info->kill); |
3174 | scan_reads (insn_info, bb_info->gen, bb_info->kill); | |
b8698a0f | 3175 | } |
6fb5fa3c DB |
3176 | |
3177 | insn_info = insn_info->prev_insn; | |
3178 | } | |
3179 | } | |
3180 | ||
3181 | ||
3182 | /* Set the gen set of the exit block, and also any block with no | |
3183 | successors that does not have a wild read. */ | |
3184 | ||
3185 | static void | |
3186 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3187 | { | |
3188 | /* The gen set is all 0's for the exit block except for the | |
3189 | frame_pointer_group. */ | |
b8698a0f | 3190 | |
6fb5fa3c DB |
3191 | if (stores_off_frame_dead_at_return) |
3192 | { | |
3193 | unsigned int i; | |
fd6daec9 | 3194 | group_info *group; |
b8698a0f | 3195 | |
9771b263 | 3196 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3197 | { |
3198 | if (group->process_globally && group->frame_related) | |
3199 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3200 | } | |
3201 | } | |
3202 | } | |
3203 | ||
3204 | ||
3205 | /* Find all of the blocks that are not backwards reachable from the | |
3206 | exit block or any block with no successors (BB). These are the | |
3207 | infinite loops or infinite self loops. These blocks will still | |
3208 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3209 | ||
3210 | static void | |
3211 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3212 | { | |
3213 | edge e; | |
3214 | edge_iterator ei; | |
3215 | ||
d7c028c0 | 3216 | if (bitmap_bit_p (unreachable_blocks, bb->index)) |
6fb5fa3c | 3217 | { |
d7c028c0 | 3218 | bitmap_clear_bit (unreachable_blocks, bb->index); |
6fb5fa3c | 3219 | FOR_EACH_EDGE (e, ei, bb->preds) |
b8698a0f | 3220 | { |
6fb5fa3c | 3221 | mark_reachable_blocks (unreachable_blocks, e->src); |
b8698a0f | 3222 | } |
6fb5fa3c DB |
3223 | } |
3224 | } | |
3225 | ||
3226 | /* Build the transfer functions for the function. */ | |
3227 | ||
3228 | static void | |
ac6929b5 | 3229 | dse_step3 () |
6fb5fa3c DB |
3230 | { |
3231 | basic_block bb; | |
6fb5fa3c DB |
3232 | sbitmap_iterator sbi; |
3233 | bitmap all_ones = NULL; | |
3234 | unsigned int i; | |
b8698a0f | 3235 | |
7ba9e72d | 3236 | auto_sbitmap unreachable_blocks (last_basic_block_for_fn (cfun)); |
f61e445a | 3237 | bitmap_ones (unreachable_blocks); |
6fb5fa3c | 3238 | |
04a90bec | 3239 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3240 | { |
3241 | bb_info_t bb_info = bb_table[bb->index]; | |
3242 | if (bb_info->gen) | |
3243 | bitmap_clear (bb_info->gen); | |
3244 | else | |
3f9b14ff | 3245 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3246 | |
3247 | if (bb->index == ENTRY_BLOCK) | |
3248 | ; | |
3249 | else if (bb->index == EXIT_BLOCK) | |
3250 | dse_step3_exit_block_scan (bb_info); | |
3251 | else | |
ac6929b5 | 3252 | dse_step3_scan (bb); |
6fb5fa3c DB |
3253 | if (EDGE_COUNT (bb->succs) == 0) |
3254 | mark_reachable_blocks (unreachable_blocks, bb); | |
3255 | ||
3256 | /* If this is the second time dataflow is run, delete the old | |
3257 | sets. */ | |
3258 | if (bb_info->in) | |
3259 | BITMAP_FREE (bb_info->in); | |
3260 | if (bb_info->out) | |
3261 | BITMAP_FREE (bb_info->out); | |
3262 | } | |
3263 | ||
3264 | /* For any block in an infinite loop, we must initialize the out set | |
3265 | to all ones. This could be expensive, but almost never occurs in | |
3266 | practice. However, it is common in regression tests. */ | |
d4ac4ce2 | 3267 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
6fb5fa3c DB |
3268 | { |
3269 | if (bitmap_bit_p (all_blocks, i)) | |
3270 | { | |
3271 | bb_info_t bb_info = bb_table[i]; | |
3272 | if (!all_ones) | |
3273 | { | |
3274 | unsigned int j; | |
fd6daec9 | 3275 | group_info *group; |
6fb5fa3c | 3276 | |
3f9b14ff | 3277 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
9771b263 | 3278 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
6fb5fa3c DB |
3279 | bitmap_ior_into (all_ones, group->group_kill); |
3280 | } | |
3281 | if (!bb_info->out) | |
3282 | { | |
3f9b14ff | 3283 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3284 | bitmap_copy (bb_info->out, all_ones); |
3285 | } | |
3286 | } | |
3287 | } | |
3288 | ||
3289 | if (all_ones) | |
3290 | BITMAP_FREE (all_ones); | |
6fb5fa3c DB |
3291 | } |
3292 | ||
3293 | ||
3294 | \f | |
3295 | /*---------------------------------------------------------------------------- | |
3296 | Fourth step. | |
3297 | ||
3298 | Solve the bitvector equations. | |
3299 | ----------------------------------------------------------------------------*/ | |
3300 | ||
3301 | ||
3302 | /* Confluence function for blocks with no successors. Create an out | |
3303 | set from the gen set of the exit block. This block logically has | |
3304 | the exit block as a successor. */ | |
3305 | ||
3306 | ||
3307 | ||
3308 | static void | |
3309 | dse_confluence_0 (basic_block bb) | |
3310 | { | |
3311 | bb_info_t bb_info = bb_table[bb->index]; | |
3312 | ||
3313 | if (bb->index == EXIT_BLOCK) | |
3314 | return; | |
3315 | ||
3316 | if (!bb_info->out) | |
3317 | { | |
3f9b14ff | 3318 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3319 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3320 | } | |
3321 | } | |
3322 | ||
3323 | /* Propagate the information from the in set of the dest of E to the | |
3324 | out set of the src of E. If the various in or out sets are not | |
3325 | there, that means they are all ones. */ | |
3326 | ||
1a0f3fa1 | 3327 | static bool |
6fb5fa3c DB |
3328 | dse_confluence_n (edge e) |
3329 | { | |
3330 | bb_info_t src_info = bb_table[e->src->index]; | |
3331 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3332 | ||
3333 | if (dest_info->in) | |
3334 | { | |
3335 | if (src_info->out) | |
3336 | bitmap_and_into (src_info->out, dest_info->in); | |
3337 | else | |
3338 | { | |
3f9b14ff | 3339 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3340 | bitmap_copy (src_info->out, dest_info->in); |
3341 | } | |
3342 | } | |
1a0f3fa1 | 3343 | return true; |
6fb5fa3c DB |
3344 | } |
3345 | ||
3346 | ||
3347 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
b8698a0f | 3348 | block. There are three cases: |
6fb5fa3c DB |
3349 | |
3350 | 1) The block has no kill set. In this case the kill set is all | |
3351 | ones. It does not matter what the out set of the block is, none of | |
3352 | the info can reach the top. The only thing that reaches the top is | |
3353 | the gen set and we just copy the set. | |
3354 | ||
3355 | 2) There is a kill set but no out set and bb has successors. In | |
3356 | this case we just return. Eventually an out set will be created and | |
3357 | it is better to wait than to create a set of ones. | |
3358 | ||
3359 | 3) There is both a kill and out set. We apply the obvious transfer | |
3360 | function. | |
3361 | */ | |
3362 | ||
3363 | static bool | |
3364 | dse_transfer_function (int bb_index) | |
3365 | { | |
3366 | bb_info_t bb_info = bb_table[bb_index]; | |
3367 | ||
3368 | if (bb_info->kill) | |
3369 | { | |
3370 | if (bb_info->out) | |
3371 | { | |
3372 | /* Case 3 above. */ | |
3373 | if (bb_info->in) | |
b8698a0f | 3374 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3375 | bb_info->out, bb_info->kill); |
3376 | else | |
3377 | { | |
3f9b14ff | 3378 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
b8698a0f | 3379 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3380 | bb_info->out, bb_info->kill); |
3381 | return true; | |
3382 | } | |
3383 | } | |
3384 | else | |
3385 | /* Case 2 above. */ | |
3386 | return false; | |
3387 | } | |
3388 | else | |
3389 | { | |
3390 | /* Case 1 above. If there is already an in set, nothing | |
3391 | happens. */ | |
3392 | if (bb_info->in) | |
3393 | return false; | |
3394 | else | |
3395 | { | |
3f9b14ff | 3396 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3397 | bitmap_copy (bb_info->in, bb_info->gen); |
3398 | return true; | |
3399 | } | |
3400 | } | |
3401 | } | |
3402 | ||
3403 | /* Solve the dataflow equations. */ | |
3404 | ||
3405 | static void | |
3406 | dse_step4 (void) | |
3407 | { | |
b8698a0f L |
3408 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3409 | dse_confluence_n, dse_transfer_function, | |
3410 | all_blocks, df_get_postorder (DF_BACKWARD), | |
6fb5fa3c | 3411 | df_get_n_blocks (DF_BACKWARD)); |
456610d3 | 3412 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3413 | { |
3414 | basic_block bb; | |
3415 | ||
3416 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
04a90bec | 3417 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3418 | { |
3419 | bb_info_t bb_info = bb_table[bb->index]; | |
3420 | ||
3421 | df_print_bb_index (bb, dump_file); | |
3422 | if (bb_info->in) | |
3423 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3424 | else | |
3425 | fprintf (dump_file, " in: *MISSING*\n"); | |
3426 | if (bb_info->gen) | |
3427 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3428 | else | |
3429 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3430 | if (bb_info->kill) | |
3431 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3432 | else | |
3433 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3434 | if (bb_info->out) | |
3435 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3436 | else | |
3437 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3438 | } | |
3439 | } | |
3440 | } | |
3441 | ||
3442 | ||
3443 | \f | |
3444 | /*---------------------------------------------------------------------------- | |
3445 | Fifth step. | |
3446 | ||
0d52bcc1 | 3447 | Delete the stores that can only be deleted using the global information. |
6fb5fa3c DB |
3448 | ----------------------------------------------------------------------------*/ |
3449 | ||
3450 | ||
3451 | static void | |
ac6929b5 | 3452 | dse_step5 (void) |
6fb5fa3c DB |
3453 | { |
3454 | basic_block bb; | |
11cd3bed | 3455 | FOR_EACH_BB_FN (bb, cfun) |
6fb5fa3c DB |
3456 | { |
3457 | bb_info_t bb_info = bb_table[bb->index]; | |
3458 | insn_info_t insn_info = bb_info->last_insn; | |
3459 | bitmap v = bb_info->out; | |
3460 | ||
3461 | while (insn_info) | |
3462 | { | |
3463 | bool deleted = false; | |
3464 | if (dump_file && insn_info->insn) | |
3465 | { | |
3466 | fprintf (dump_file, "starting to process insn %d\n", | |
3467 | INSN_UID (insn_info->insn)); | |
3468 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3469 | } | |
3470 | ||
3471 | /* There may have been code deleted by the dce pass run before | |
3472 | this phase. */ | |
b8698a0f | 3473 | if (insn_info->insn |
6fb5fa3c DB |
3474 | && INSN_P (insn_info->insn) |
3475 | && (!insn_info->cannot_delete) | |
3476 | && (!bitmap_empty_p (v))) | |
3477 | { | |
fd6daec9 | 3478 | store_info *store_info = insn_info->store_rec; |
6fb5fa3c DB |
3479 | |
3480 | /* Try to delete the current insn. */ | |
3481 | deleted = true; | |
b8698a0f | 3482 | |
6fb5fa3c DB |
3483 | /* Skip the clobbers. */ |
3484 | while (!store_info->is_set) | |
3485 | store_info = store_info->next; | |
3486 | ||
02ce5d90 | 3487 | HOST_WIDE_INT i, offset, width; |
ac6929b5 RB |
3488 | group_info *group_info = rtx_group_vec[store_info->group_id]; |
3489 | ||
02ce5d90 RS |
3490 | if (!store_info->offset.is_constant (&offset) |
3491 | || !store_info->width.is_constant (&width)) | |
3492 | deleted = false; | |
3493 | else | |
6fb5fa3c | 3494 | { |
02ce5d90 RS |
3495 | HOST_WIDE_INT end = offset + width; |
3496 | for (i = offset; i < end; i++) | |
6fb5fa3c | 3497 | { |
02ce5d90 RS |
3498 | int index = get_bitmap_index (group_info, i); |
3499 | ||
456610d3 | 3500 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
3501 | fprintf (dump_file, "i = %d, index = %d\n", |
3502 | (int) i, index); | |
3503 | if (index == 0 || !bitmap_bit_p (v, index)) | |
3504 | { | |
3505 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3506 | fprintf (dump_file, "failing at i = %d\n", | |
3507 | (int) i); | |
3508 | deleted = false; | |
3509 | break; | |
3510 | } | |
6fb5fa3c DB |
3511 | } |
3512 | } | |
3513 | if (deleted) | |
3514 | { | |
9e582b1d JR |
3515 | if (dbg_cnt (dse) |
3516 | && check_for_inc_dec_1 (insn_info)) | |
6fb5fa3c | 3517 | { |
6fb5fa3c DB |
3518 | delete_insn (insn_info->insn); |
3519 | insn_info->insn = NULL; | |
3520 | globally_deleted++; | |
3521 | } | |
3522 | } | |
3523 | } | |
3524 | /* We do want to process the local info if the insn was | |
6ed3da00 | 3525 | deleted. For instance, if the insn did a wild read, we |
6fb5fa3c | 3526 | no longer need to trash the info. */ |
b8698a0f | 3527 | if (insn_info->insn |
6fb5fa3c DB |
3528 | && INSN_P (insn_info->insn) |
3529 | && (!deleted)) | |
3530 | { | |
ac6929b5 | 3531 | scan_stores (insn_info->store_rec, v, NULL); |
6fb5fa3c DB |
3532 | if (insn_info->wild_read) |
3533 | { | |
456610d3 | 3534 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3535 | fprintf (dump_file, "wild read\n"); |
3536 | bitmap_clear (v); | |
3537 | } | |
d26c7090 | 3538 | else if (insn_info->read_rec |
a8404b9a JJ |
3539 | || insn_info->non_frame_wild_read |
3540 | || insn_info->frame_read) | |
6fb5fa3c | 3541 | { |
a8404b9a JJ |
3542 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3543 | { | |
3544 | if (!insn_info->non_frame_wild_read | |
3545 | && !insn_info->frame_read) | |
3546 | fprintf (dump_file, "regular read\n"); | |
3547 | if (insn_info->non_frame_wild_read) | |
3548 | fprintf (dump_file, "non-frame wild read\n"); | |
3549 | if (insn_info->frame_read) | |
3550 | fprintf (dump_file, "frame read\n"); | |
3551 | } | |
ac6929b5 | 3552 | scan_reads (insn_info, v, NULL); |
6fb5fa3c DB |
3553 | } |
3554 | } | |
b8698a0f | 3555 | |
6fb5fa3c DB |
3556 | insn_info = insn_info->prev_insn; |
3557 | } | |
3558 | } | |
3559 | } | |
3560 | ||
3561 | ||
6fb5fa3c DB |
3562 | \f |
3563 | /*---------------------------------------------------------------------------- | |
3564 | Sixth step. | |
3565 | ||
8dd5516b JJ |
3566 | Delete stores made redundant by earlier stores (which store the same |
3567 | value) that couldn't be eliminated. | |
3568 | ----------------------------------------------------------------------------*/ | |
3569 | ||
3570 | static void | |
3571 | dse_step6 (void) | |
3572 | { | |
3573 | basic_block bb; | |
3574 | ||
04a90bec | 3575 | FOR_ALL_BB_FN (bb, cfun) |
8dd5516b JJ |
3576 | { |
3577 | bb_info_t bb_info = bb_table[bb->index]; | |
3578 | insn_info_t insn_info = bb_info->last_insn; | |
3579 | ||
3580 | while (insn_info) | |
3581 | { | |
3582 | /* There may have been code deleted by the dce pass run before | |
3583 | this phase. */ | |
3584 | if (insn_info->insn | |
3585 | && INSN_P (insn_info->insn) | |
3586 | && !insn_info->cannot_delete) | |
3587 | { | |
fd6daec9 | 3588 | store_info *s_info = insn_info->store_rec; |
8dd5516b JJ |
3589 | |
3590 | while (s_info && !s_info->is_set) | |
3591 | s_info = s_info->next; | |
3592 | if (s_info | |
3593 | && s_info->redundant_reason | |
3594 | && s_info->redundant_reason->insn | |
3595 | && INSN_P (s_info->redundant_reason->insn)) | |
3596 | { | |
eb92d49a | 3597 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
456610d3 | 3598 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
3599 | fprintf (dump_file, "Locally deleting insn %d " |
3600 | "because insn %d stores the " | |
3601 | "same value and couldn't be " | |
3602 | "eliminated\n", | |
3603 | INSN_UID (insn_info->insn), | |
3604 | INSN_UID (rinsn)); | |
3605 | delete_dead_store_insn (insn_info); | |
3606 | } | |
3607 | } | |
3608 | insn_info = insn_info->prev_insn; | |
3609 | } | |
3610 | } | |
3611 | } | |
3612 | \f | |
3613 | /*---------------------------------------------------------------------------- | |
3614 | Seventh step. | |
3615 | ||
b8698a0f | 3616 | Destroy everything left standing. |
6fb5fa3c DB |
3617 | ----------------------------------------------------------------------------*/ |
3618 | ||
b8698a0f | 3619 | static void |
3f9b14ff | 3620 | dse_step7 (void) |
6fb5fa3c | 3621 | { |
3f9b14ff SB |
3622 | bitmap_obstack_release (&dse_bitmap_obstack); |
3623 | obstack_free (&dse_obstack, NULL); | |
370f38e8 | 3624 | |
6fb5fa3c DB |
3625 | end_alias_analysis (); |
3626 | free (bb_table); | |
c203e8a7 TS |
3627 | delete rtx_group_table; |
3628 | rtx_group_table = NULL; | |
9771b263 | 3629 | rtx_group_vec.release (); |
6fb5fa3c DB |
3630 | BITMAP_FREE (all_blocks); |
3631 | BITMAP_FREE (scratch); | |
3632 | ||
da6603c6 | 3633 | rtx_store_info_pool.release (); |
fb0b2914 ML |
3634 | read_info_type_pool.release (); |
3635 | insn_info_type_pool.release (); | |
3636 | dse_bb_info_type_pool.release (); | |
3637 | group_info_pool.release (); | |
3638 | deferred_change_pool.release (); | |
6fb5fa3c DB |
3639 | } |
3640 | ||
3641 | ||
6fb5fa3c DB |
3642 | /* ------------------------------------------------------------------------- |
3643 | DSE | |
3644 | ------------------------------------------------------------------------- */ | |
3645 | ||
3646 | /* Callback for running pass_rtl_dse. */ | |
3647 | ||
3648 | static unsigned int | |
3649 | rest_of_handle_dse (void) | |
3650 | { | |
6fb5fa3c DB |
3651 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3652 | ||
02b47899 KZ |
3653 | /* Need the notes since we must track live hardregs in the forwards |
3654 | direction. */ | |
3655 | df_note_add_problem (); | |
3656 | df_analyze (); | |
3657 | ||
6fb5fa3c DB |
3658 | dse_step0 (); |
3659 | dse_step1 (); | |
3660 | dse_step2_init (); | |
ac6929b5 | 3661 | if (dse_step2 ()) |
6fb5fa3c DB |
3662 | { |
3663 | df_set_flags (DF_LR_RUN_DCE); | |
3664 | df_analyze (); | |
456610d3 | 3665 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 3666 | fprintf (dump_file, "doing global processing\n"); |
ac6929b5 | 3667 | dse_step3 (); |
6fb5fa3c | 3668 | dse_step4 (); |
ac6929b5 | 3669 | dse_step5 (); |
6fb5fa3c DB |
3670 | } |
3671 | ||
8dd5516b | 3672 | dse_step6 (); |
3f9b14ff | 3673 | dse_step7 (); |
6fb5fa3c DB |
3674 | |
3675 | if (dump_file) | |
ac6929b5 RB |
3676 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d\n", |
3677 | locally_deleted, globally_deleted); | |
6c5ed3f1 JJ |
3678 | |
3679 | /* DSE can eliminate potentially-trapping MEMs. | |
3680 | Remove any EH edges associated with them. */ | |
3681 | if ((locally_deleted || globally_deleted) | |
3682 | && cfun->can_throw_non_call_exceptions | |
3683 | && purge_all_dead_edges ()) | |
7f5ddfcb JJ |
3684 | { |
3685 | free_dominance_info (CDI_DOMINATORS); | |
3686 | cleanup_cfg (0); | |
3687 | } | |
6c5ed3f1 | 3688 | |
6fb5fa3c DB |
3689 | return 0; |
3690 | } | |
3691 | ||
17795822 TS |
3692 | namespace { |
3693 | ||
3694 | const pass_data pass_data_rtl_dse1 = | |
27a4cd48 DM |
3695 | { |
3696 | RTL_PASS, /* type */ | |
3697 | "dse1", /* name */ | |
3698 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3699 | TV_DSE1, /* tv_id */ |
3700 | 0, /* properties_required */ | |
3701 | 0, /* properties_provided */ | |
3702 | 0, /* properties_destroyed */ | |
3703 | 0, /* todo_flags_start */ | |
3bea341f | 3704 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c DB |
3705 | }; |
3706 | ||
17795822 | 3707 | class pass_rtl_dse1 : public rtl_opt_pass |
27a4cd48 DM |
3708 | { |
3709 | public: | |
c3284718 RS |
3710 | pass_rtl_dse1 (gcc::context *ctxt) |
3711 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
27a4cd48 DM |
3712 | {} |
3713 | ||
3714 | /* opt_pass methods: */ | |
1a3d085c TS |
3715 | virtual bool gate (function *) |
3716 | { | |
3717 | return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
3718 | } | |
3719 | ||
be55bfe6 | 3720 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3721 | |
3722 | }; // class pass_rtl_dse1 | |
3723 | ||
17795822 TS |
3724 | } // anon namespace |
3725 | ||
27a4cd48 DM |
3726 | rtl_opt_pass * |
3727 | make_pass_rtl_dse1 (gcc::context *ctxt) | |
3728 | { | |
3729 | return new pass_rtl_dse1 (ctxt); | |
3730 | } | |
3731 | ||
17795822 TS |
3732 | namespace { |
3733 | ||
3734 | const pass_data pass_data_rtl_dse2 = | |
27a4cd48 DM |
3735 | { |
3736 | RTL_PASS, /* type */ | |
3737 | "dse2", /* name */ | |
3738 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3739 | TV_DSE2, /* tv_id */ |
3740 | 0, /* properties_required */ | |
3741 | 0, /* properties_provided */ | |
3742 | 0, /* properties_destroyed */ | |
3743 | 0, /* todo_flags_start */ | |
3bea341f | 3744 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c | 3745 | }; |
27a4cd48 | 3746 | |
17795822 | 3747 | class pass_rtl_dse2 : public rtl_opt_pass |
27a4cd48 DM |
3748 | { |
3749 | public: | |
c3284718 RS |
3750 | pass_rtl_dse2 (gcc::context *ctxt) |
3751 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
27a4cd48 DM |
3752 | {} |
3753 | ||
3754 | /* opt_pass methods: */ | |
1a3d085c TS |
3755 | virtual bool gate (function *) |
3756 | { | |
3757 | return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
3758 | } | |
3759 | ||
be55bfe6 | 3760 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3761 | |
3762 | }; // class pass_rtl_dse2 | |
3763 | ||
17795822 TS |
3764 | } // anon namespace |
3765 | ||
27a4cd48 DM |
3766 | rtl_opt_pass * |
3767 | make_pass_rtl_dse2 (gcc::context *ctxt) | |
3768 | { | |
3769 | return new pass_rtl_dse2 (ctxt); | |
3770 | } |