]>
Commit | Line | Data |
---|---|---|
6fb5fa3c | 1 | /* RTL dead store elimination. |
a945c346 | 2 | Copyright (C) 2005-2024 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 | 253 | |
64ca6aa7 AP |
254 | /* The address space that the memory reference uses. */ |
255 | unsigned char addrspace; | |
256 | ||
8dd5516b JJ |
257 | union |
258 | { | |
259 | /* A bitmask as wide as the number of bytes in the word that | |
260 | contains a 1 if the byte may be needed. The store is unused if | |
261 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
262 | unsigned HOST_WIDE_INT small_bitmask; | |
263 | ||
264 | struct | |
265 | { | |
02ce5d90 RS |
266 | /* A bitmap with one bit per byte, or null if the number of |
267 | bytes isn't known at compile time. A cleared bit means | |
268 | the position is needed. Used if IS_LARGE is true. */ | |
dc491a25 | 269 | bitmap bmap; |
6fb5fa3c | 270 | |
02ce5d90 RS |
271 | /* When BITMAP is nonnull, this counts the number of set bits |
272 | (i.e. unneeded bytes) in the bitmap. If it is equal to | |
273 | WIDTH, the whole store is unused. | |
274 | ||
275 | When BITMAP is null: | |
276 | - the store is definitely not needed when COUNT == 1 | |
277 | - all the store is needed when COUNT == 0 and RHS is nonnull | |
278 | - otherwise we don't know which parts of the store are needed. */ | |
8dd5516b JJ |
279 | int count; |
280 | } large; | |
281 | } positions_needed; | |
6fb5fa3c DB |
282 | |
283 | /* The next store info for this insn. */ | |
99b1c316 | 284 | class store_info *next; |
6fb5fa3c DB |
285 | |
286 | /* The right hand side of the store. This is used if there is a | |
287 | subsequent reload of the mems address somewhere later in the | |
288 | basic block. */ | |
8dd5516b JJ |
289 | rtx rhs; |
290 | ||
291 | /* If rhs is or holds a constant, this contains that constant, | |
292 | otherwise NULL. */ | |
293 | rtx const_rhs; | |
294 | ||
295 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
296 | insn stored. These aren't eliminated early, because doing that | |
297 | might prevent the earlier larger store to be eliminated. */ | |
da6603c6 | 298 | struct insn_info_type *redundant_reason; |
6fb5fa3c DB |
299 | }; |
300 | ||
4fe663b0 L |
301 | /* Return a bitmask with the first N low bits set. */ |
302 | ||
303 | static unsigned HOST_WIDE_INT | |
304 | lowpart_bitmask (int n) | |
305 | { | |
dd4786fe | 306 | unsigned HOST_WIDE_INT mask = HOST_WIDE_INT_M1U; |
4fe663b0 L |
307 | return mask >> (HOST_BITS_PER_WIDE_INT - n); |
308 | } | |
309 | ||
fcb87c50 | 310 | static object_allocator<store_info> cse_store_info_pool ("cse_store_info_pool"); |
da6603c6 | 311 | |
fcb87c50 | 312 | static object_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool"); |
6fb5fa3c DB |
313 | |
314 | /* This structure holds information about a load. These are only | |
315 | built for rtx bases. */ | |
6c1dae73 | 316 | class read_info_type |
6fb5fa3c | 317 | { |
6c1dae73 | 318 | public: |
6fb5fa3c DB |
319 | /* The id of the mem group of the base address. */ |
320 | int group_id; | |
321 | ||
71d6a386 | 322 | /* The offset of the first byte associated with the operation. */ |
02ce5d90 | 323 | poly_int64 offset; |
71d6a386 RS |
324 | |
325 | /* The number of bytes covered by the operation, or -1 if not known. */ | |
02ce5d90 | 326 | poly_int64 width; |
6fb5fa3c DB |
327 | |
328 | /* The mem being read. */ | |
329 | rtx mem; | |
330 | ||
331 | /* The next read_info for this insn. */ | |
99b1c316 | 332 | class read_info_type *next; |
6fb5fa3c | 333 | }; |
99b1c316 | 334 | typedef class read_info_type *read_info_t; |
6fb5fa3c | 335 | |
fcb87c50 | 336 | static object_allocator<read_info_type> read_info_type_pool ("read_info_pool"); |
6fb5fa3c DB |
337 | |
338 | /* One of these records is created for each insn. */ | |
339 | ||
da6603c6 | 340 | struct insn_info_type |
6fb5fa3c DB |
341 | { |
342 | /* Set true if the insn contains a store but the insn itself cannot | |
343 | be deleted. This is set if the insn is a parallel and there is | |
344 | more than one non dead output or if the insn is in some way | |
345 | volatile. */ | |
346 | bool cannot_delete; | |
347 | ||
348 | /* This field is only used by the global algorithm. It is set true | |
349 | if the insn contains any read of mem except for a (1). This is | |
350 | also set if the insn is a call or has a clobber mem. If the insn | |
351 | contains a wild read, the use_rec will be null. */ | |
352 | bool wild_read; | |
353 | ||
d26c7090 ER |
354 | /* This is true only for CALL instructions which could potentially read |
355 | any non-frame memory location. This field is used by the global | |
356 | algorithm. */ | |
357 | bool non_frame_wild_read; | |
358 | ||
50f0f366 EB |
359 | /* This field is only used for the processing of const functions. |
360 | These functions cannot read memory, but they can read the stack | |
64520bdc EB |
361 | because that is where they may get their parms. We need to be |
362 | this conservative because, like the store motion pass, we don't | |
363 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
364 | Moreover, we need to distinguish two cases: | |
365 | 1. Before reload (register elimination), the stores related to | |
366 | outgoing arguments are stack pointer based and thus deemed | |
367 | of non-constant base in this pass. This requires special | |
368 | handling but also means that the frame pointer based stores | |
369 | need not be killed upon encountering a const function call. | |
370 | 2. After reload, the stores related to outgoing arguments can be | |
371 | either stack pointer or hard frame pointer based. This means | |
372 | that we have no other choice than also killing all the frame | |
373 | pointer based stores upon encountering a const function call. | |
57534689 JJ |
374 | This field is set after reload for const function calls and before |
375 | reload for const tail function calls on targets where arg pointer | |
376 | is the frame pointer. Having this set is less severe than a wild | |
377 | read, it just means that all the frame related stores are killed | |
378 | rather than all the stores. */ | |
64520bdc EB |
379 | bool frame_read; |
380 | ||
381 | /* This field is only used for the processing of const functions. | |
382 | It is set if the insn may contain a stack pointer based store. */ | |
50f0f366 | 383 | bool stack_pointer_based; |
6fb5fa3c DB |
384 | |
385 | /* This is true if any of the sets within the store contains a | |
386 | cselib base. Such stores can only be deleted by the local | |
387 | algorithm. */ | |
388 | bool contains_cselib_groups; | |
389 | ||
390 | /* The insn. */ | |
dd60a84c | 391 | rtx_insn *insn; |
6fb5fa3c DB |
392 | |
393 | /* The list of mem sets or mem clobbers that are contained in this | |
394 | insn. If the insn is deletable, it contains only one mem set. | |
395 | But it could also contain clobbers. Insns that contain more than | |
396 | one mem set are not deletable, but each of those mems are here in | |
6ed3da00 | 397 | order to provide info to delete other insns. */ |
fd6daec9 | 398 | store_info *store_rec; |
6fb5fa3c DB |
399 | |
400 | /* The linked list of mem uses in this insn. Only the reads from | |
401 | rtx bases are listed here. The reads to cselib bases are | |
402 | completely processed during the first scan and so are never | |
403 | created. */ | |
404 | read_info_t read_rec; | |
405 | ||
9e582b1d JR |
406 | /* The live fixed registers. We assume only fixed registers can |
407 | cause trouble by being clobbered from an expanded pattern; | |
408 | storing only the live fixed registers (rather than all registers) | |
409 | means less memory needs to be allocated / copied for the individual | |
410 | stores. */ | |
411 | regset fixed_regs_live; | |
412 | ||
6fb5fa3c | 413 | /* The prev insn in the basic block. */ |
da6603c6 | 414 | struct insn_info_type * prev_insn; |
6fb5fa3c DB |
415 | |
416 | /* The linked list of insns that are in consideration for removal in | |
073a8998 | 417 | the forwards pass through the basic block. This pointer may be |
6fb5fa3c | 418 | trash as it is not cleared when a wild read occurs. The only |
fa10beec | 419 | time it is guaranteed to be correct is when the traversal starts |
6fb5fa3c | 420 | at active_local_stores. */ |
da6603c6 | 421 | struct insn_info_type * next_local_store; |
6fb5fa3c | 422 | }; |
da6603c6 | 423 | typedef struct insn_info_type *insn_info_t; |
6fb5fa3c | 424 | |
fcb87c50 | 425 | static object_allocator<insn_info_type> insn_info_type_pool ("insn_info_pool"); |
6fb5fa3c DB |
426 | |
427 | /* The linked list of stores that are under consideration in this | |
b8698a0f | 428 | basic block. */ |
6fb5fa3c | 429 | static insn_info_t active_local_stores; |
dabd47e7 | 430 | static int active_local_stores_len; |
6fb5fa3c | 431 | |
da6603c6 | 432 | struct dse_bb_info_type |
6fb5fa3c | 433 | { |
6fb5fa3c DB |
434 | /* Pointer to the insn info for the last insn in the block. These |
435 | are linked so this is how all of the insns are reached. During | |
436 | scanning this is the current insn being scanned. */ | |
437 | insn_info_t last_insn; | |
438 | ||
439 | /* The info for the global dataflow problem. */ | |
440 | ||
441 | ||
442 | /* This is set if the transfer function should and in the wild_read | |
443 | bitmap before applying the kill and gen sets. That vector knocks | |
444 | out most of the bits in the bitmap and thus speeds up the | |
445 | operations. */ | |
446 | bool apply_wild_read; | |
447 | ||
02b47899 KZ |
448 | /* The following 4 bitvectors hold information about which positions |
449 | of which stores are live or dead. They are indexed by | |
450 | get_bitmap_index. */ | |
451 | ||
6fb5fa3c DB |
452 | /* The set of store positions that exist in this block before a wild read. */ |
453 | bitmap gen; | |
b8698a0f | 454 | |
6fb5fa3c DB |
455 | /* The set of load positions that exist in this block above the |
456 | same position of a store. */ | |
457 | bitmap kill; | |
458 | ||
459 | /* The set of stores that reach the top of the block without being | |
460 | killed by a read. | |
461 | ||
462 | Do not represent the in if it is all ones. Note that this is | |
463 | what the bitvector should logically be initialized to for a set | |
464 | intersection problem. However, like the kill set, this is too | |
465 | expensive. So initially, the in set will only be created for the | |
466 | exit block and any block that contains a wild read. */ | |
467 | bitmap in; | |
468 | ||
469 | /* The set of stores that reach the bottom of the block from it's | |
470 | successors. | |
471 | ||
472 | Do not represent the in if it is all ones. Note that this is | |
473 | what the bitvector should logically be initialized to for a set | |
474 | intersection problem. However, like the kill and in set, this is | |
475 | too expensive. So what is done is that the confluence operator | |
476 | just initializes the vector from one of the out sets of the | |
477 | successors of the block. */ | |
478 | bitmap out; | |
02b47899 KZ |
479 | |
480 | /* The following bitvector is indexed by the reg number. It | |
481 | contains the set of regs that are live at the current instruction | |
482 | being processed. While it contains info for all of the | |
9e582b1d JR |
483 | registers, only the hard registers are actually examined. It is used |
484 | to assure that shift and/or add sequences that are inserted do not | |
073a8998 | 485 | accidentally clobber live hard regs. */ |
02b47899 | 486 | bitmap regs_live; |
6fb5fa3c DB |
487 | }; |
488 | ||
da6603c6 | 489 | typedef struct dse_bb_info_type *bb_info_t; |
fb0b2914 ML |
490 | |
491 | static object_allocator<dse_bb_info_type> dse_bb_info_type_pool | |
fcb87c50 | 492 | ("bb_info_pool"); |
6fb5fa3c DB |
493 | |
494 | /* Table to hold all bb_infos. */ | |
495 | static bb_info_t *bb_table; | |
496 | ||
497 | /* There is a group_info for each rtx base that is used to reference | |
498 | memory. There are also not many of the rtx bases because they are | |
499 | very limited in scope. */ | |
500 | ||
b8698a0f | 501 | struct group_info |
6fb5fa3c DB |
502 | { |
503 | /* The actual base of the address. */ | |
504 | rtx rtx_base; | |
505 | ||
506 | /* The sequential id of the base. This allows us to have a | |
507 | canonical ordering of these that is not based on addresses. */ | |
508 | int id; | |
509 | ||
8f5929e1 JJ |
510 | /* True if there are any positions that are to be processed |
511 | globally. */ | |
512 | bool process_globally; | |
513 | ||
514 | /* True if the base of this group is either the frame_pointer or | |
515 | hard_frame_pointer. */ | |
516 | bool frame_related; | |
517 | ||
d32f725a OH |
518 | /* A mem wrapped around the base pointer for the group in order to do |
519 | read dependency. It must be given BLKmode in order to encompass all | |
520 | the possible offsets from the base. */ | |
6fb5fa3c | 521 | rtx base_mem; |
b8698a0f | 522 | |
6216f94e JJ |
523 | /* Canonized version of base_mem's address. */ |
524 | rtx canon_base_addr; | |
6fb5fa3c DB |
525 | |
526 | /* These two sets of two bitmaps are used to keep track of how many | |
6ed3da00 | 527 | stores are actually referencing that position from this base. We |
6fb5fa3c | 528 | only do this for rtx bases as this will be used to assign |
6ed3da00 | 529 | positions in the bitmaps for the global problem. Bit N is set in |
6fb5fa3c DB |
530 | store1 on the first store for offset N. Bit N is set in store2 |
531 | for the second store to offset N. This is all we need since we | |
532 | only care about offsets that have two or more stores for them. | |
533 | ||
534 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
535 | for 0 and greater offsets. | |
536 | ||
537 | There is one special case here, for stores into the stack frame, | |
538 | we will or store1 into store2 before deciding which stores look | |
539 | at globally. This is because stores to the stack frame that have | |
540 | no other reads before the end of the function can also be | |
541 | deleted. */ | |
542 | bitmap store1_n, store1_p, store2_n, store2_p; | |
543 | ||
d26c7090 ER |
544 | /* These bitmaps keep track of offsets in this group escape this function. |
545 | An offset escapes if it corresponds to a named variable whose | |
546 | addressable flag is set. */ | |
547 | bitmap escaped_n, escaped_p; | |
548 | ||
6ed3da00 | 549 | /* The positions in this bitmap have the same assignments as the in, |
6fb5fa3c | 550 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6ed3da00 | 551 | the positions that are occupied by stores for this group. */ |
6fb5fa3c DB |
552 | bitmap group_kill; |
553 | ||
6fb5fa3c | 554 | /* The offset_map is used to map the offsets from this base into |
6ed3da00 | 555 | positions in the global bitmaps. It is only created after all of |
6fb5fa3c DB |
556 | the all of stores have been scanned and we know which ones we |
557 | care about. */ | |
b8698a0f L |
558 | int *offset_map_n, *offset_map_p; |
559 | int offset_map_size_n, offset_map_size_p; | |
6fb5fa3c | 560 | }; |
da6603c6 | 561 | |
fcb87c50 | 562 | static object_allocator<group_info> group_info_pool ("rtx_group_info_pool"); |
6fb5fa3c | 563 | |
6fb5fa3c DB |
564 | /* Index into the rtx_group_vec. */ |
565 | static int rtx_group_next_id; | |
566 | ||
6fb5fa3c | 567 | |
fd6daec9 | 568 | static vec<group_info *> rtx_group_vec; |
6fb5fa3c DB |
569 | |
570 | ||
571 | /* This structure holds the set of changes that are being deferred | |
572 | when removing read operation. See replace_read. */ | |
b8698a0f | 573 | struct deferred_change |
6fb5fa3c DB |
574 | { |
575 | ||
576 | /* The mem that is being replaced. */ | |
577 | rtx *loc; | |
578 | ||
579 | /* The reg it is being replaced with. */ | |
580 | rtx reg; | |
581 | ||
582 | struct deferred_change *next; | |
583 | }; | |
584 | ||
fb0b2914 | 585 | static object_allocator<deferred_change> deferred_change_pool |
fcb87c50 | 586 | ("deferred_change_pool"); |
6fb5fa3c | 587 | |
fd6daec9 | 588 | static deferred_change *deferred_change_list = NULL; |
6fb5fa3c | 589 | |
e3b5732b | 590 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
9dd9bf80 | 591 | this for vararg functions because they play games with the frame. */ |
6fb5fa3c DB |
592 | static bool stores_off_frame_dead_at_return; |
593 | ||
594 | /* Counter for stats. */ | |
b8698a0f L |
595 | static int globally_deleted; |
596 | static int locally_deleted; | |
b8698a0f | 597 | |
6fb5fa3c DB |
598 | static bitmap all_blocks; |
599 | ||
d26c7090 ER |
600 | /* Locations that are killed by calls in the global phase. */ |
601 | static bitmap kill_on_calls; | |
602 | ||
6fb5fa3c DB |
603 | /* The number of bits used in the global bitmaps. */ |
604 | static unsigned int current_position; | |
71d6a386 RS |
605 | |
606 | /* Print offset range [OFFSET, OFFSET + WIDTH) to FILE. */ | |
607 | ||
608 | static void | |
609 | print_range (FILE *file, poly_int64 offset, poly_int64 width) | |
610 | { | |
611 | fprintf (file, "["); | |
612 | print_dec (offset, file, SIGNED); | |
613 | fprintf (file, ".."); | |
614 | print_dec (offset + width, file, SIGNED); | |
615 | fprintf (file, ")"); | |
616 | } | |
6fb5fa3c DB |
617 | \f |
618 | /*---------------------------------------------------------------------------- | |
619 | Zeroth step. | |
620 | ||
b8698a0f | 621 | Initialization. |
6fb5fa3c DB |
622 | ----------------------------------------------------------------------------*/ |
623 | ||
6fb5fa3c | 624 | |
6fb5fa3c DB |
625 | /* Hashtable callbacks for maintaining the "bases" field of |
626 | store_group_info, given that the addresses are function invariants. */ | |
627 | ||
8d67ee55 | 628 | struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
703c8606 | 629 | { |
67f58944 TS |
630 | static inline hashval_t hash (const group_info *); |
631 | static inline bool equal (const group_info *, const group_info *); | |
703c8606 LC |
632 | }; |
633 | ||
634 | inline bool | |
67f58944 TS |
635 | invariant_group_base_hasher::equal (const group_info *gi1, |
636 | const group_info *gi2) | |
6fb5fa3c | 637 | { |
6fb5fa3c DB |
638 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
639 | } | |
640 | ||
703c8606 | 641 | inline hashval_t |
67f58944 | 642 | invariant_group_base_hasher::hash (const group_info *gi) |
6fb5fa3c | 643 | { |
6fb5fa3c DB |
644 | int do_not_record; |
645 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
646 | } | |
647 | ||
703c8606 | 648 | /* Tables of group_info structures, hashed by base value. */ |
c203e8a7 | 649 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
703c8606 | 650 | |
6fb5fa3c DB |
651 | |
652 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
653 | ||
fd6daec9 | 654 | static group_info * |
6fb5fa3c DB |
655 | get_group_info (rtx base) |
656 | { | |
b8698a0f | 657 | struct group_info tmp_gi; |
fd6daec9 | 658 | group_info *gi; |
703c8606 | 659 | group_info **slot; |
6fb5fa3c | 660 | |
ac6929b5 RB |
661 | gcc_assert (base != NULL_RTX); |
662 | ||
663 | /* Find the store_base_info structure for BASE, creating a new one | |
664 | if necessary. */ | |
665 | tmp_gi.rtx_base = base; | |
666 | slot = rtx_group_table->find_slot (&tmp_gi, INSERT); | |
667 | gi = *slot; | |
6fb5fa3c DB |
668 | |
669 | if (gi == NULL) | |
670 | { | |
fb0b2914 | 671 | *slot = gi = group_info_pool.allocate (); |
6fb5fa3c DB |
672 | gi->rtx_base = base; |
673 | gi->id = rtx_group_next_id++; | |
d32f725a | 674 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
6216f94e | 675 | gi->canon_base_addr = canon_rtx (base); |
3f9b14ff SB |
676 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
677 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
678 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
679 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
680 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
681 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
682 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 683 | gi->process_globally = false; |
b8698a0f | 684 | gi->frame_related = |
47593832 JG |
685 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx) |
686 | || (base == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]); | |
6fb5fa3c DB |
687 | gi->offset_map_size_n = 0; |
688 | gi->offset_map_size_p = 0; | |
689 | gi->offset_map_n = NULL; | |
690 | gi->offset_map_p = NULL; | |
9771b263 | 691 | rtx_group_vec.safe_push (gi); |
6fb5fa3c DB |
692 | } |
693 | ||
694 | return gi; | |
695 | } | |
696 | ||
697 | ||
698 | /* Initialization of data structures. */ | |
699 | ||
700 | static void | |
701 | dse_step0 (void) | |
702 | { | |
703 | locally_deleted = 0; | |
704 | globally_deleted = 0; | |
6fb5fa3c | 705 | |
3f9b14ff SB |
706 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
707 | gcc_obstack_init (&dse_obstack); | |
708 | ||
709 | scratch = BITMAP_ALLOC (®_obstack); | |
710 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 711 | |
6fb5fa3c | 712 | |
c203e8a7 | 713 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
6fb5fa3c | 714 | |
8b1c6fd7 | 715 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
6fb5fa3c DB |
716 | rtx_group_next_id = 0; |
717 | ||
e3b5732b | 718 | stores_off_frame_dead_at_return = !cfun->stdarg; |
6fb5fa3c DB |
719 | |
720 | init_alias_analysis (); | |
6fb5fa3c DB |
721 | } |
722 | ||
723 | ||
724 | \f | |
725 | /*---------------------------------------------------------------------------- | |
726 | First step. | |
727 | ||
728 | Scan all of the insns. Any random ordering of the blocks is fine. | |
0d52bcc1 | 729 | Each block is scanned in forward order to accommodate cselib which |
6fb5fa3c DB |
730 | is used to remove stores with non-constant bases. |
731 | ----------------------------------------------------------------------------*/ | |
732 | ||
733 | /* Delete all of the store_info recs from INSN_INFO. */ | |
734 | ||
b8698a0f | 735 | static void |
6fb5fa3c DB |
736 | free_store_info (insn_info_t insn_info) |
737 | { | |
fd6daec9 TS |
738 | store_info *cur = insn_info->store_rec; |
739 | while (cur) | |
6fb5fa3c | 740 | { |
fd6daec9 TS |
741 | store_info *next = cur->next; |
742 | if (cur->is_large) | |
743 | BITMAP_FREE (cur->positions_needed.large.bmap); | |
744 | if (cur->cse_base) | |
745 | cse_store_info_pool.remove (cur); | |
6fb5fa3c | 746 | else |
fd6daec9 TS |
747 | rtx_store_info_pool.remove (cur); |
748 | cur = next; | |
6fb5fa3c DB |
749 | } |
750 | ||
751 | insn_info->cannot_delete = true; | |
752 | insn_info->contains_cselib_groups = false; | |
753 | insn_info->store_rec = NULL; | |
754 | } | |
755 | ||
a79683d5 | 756 | struct note_add_store_info |
9e582b1d | 757 | { |
dc01c3d1 | 758 | rtx_insn *first, *current; |
9e582b1d JR |
759 | regset fixed_regs_live; |
760 | bool failure; | |
a79683d5 | 761 | }; |
9e582b1d JR |
762 | |
763 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
764 | Check if a register is clobbered which is live afterwards. */ | |
765 | ||
766 | static void | |
767 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
768 | { | |
dc01c3d1 | 769 | rtx_insn *insn; |
9e582b1d | 770 | note_add_store_info *info = (note_add_store_info *) data; |
9e582b1d JR |
771 | |
772 | if (!REG_P (loc)) | |
773 | return; | |
774 | ||
775 | /* If this register is referenced by the current or an earlier insn, | |
776 | that's OK. E.g. this applies to the register that is being incremented | |
777 | with this addition. */ | |
778 | for (insn = info->first; | |
779 | insn != NEXT_INSN (info->current); | |
780 | insn = NEXT_INSN (insn)) | |
781 | if (reg_referenced_p (loc, PATTERN (insn))) | |
782 | return; | |
783 | ||
784 | /* If we come here, we have a clobber of a register that's only OK | |
785 | if that register is not live. If we don't have liveness information | |
786 | available, fail now. */ | |
787 | if (!info->fixed_regs_live) | |
788 | { | |
53d1bae9 | 789 | info->failure = true; |
9e582b1d JR |
790 | return; |
791 | } | |
792 | /* Now check if this is a live fixed register. */ | |
53d1bae9 RS |
793 | unsigned int end_regno = END_REGNO (loc); |
794 | for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) | |
795 | if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) | |
796 | info->failure = true; | |
9e582b1d JR |
797 | } |
798 | ||
4deef538 AO |
799 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
800 | SRC + SRCOFF before insn ARG. */ | |
6fb5fa3c DB |
801 | |
802 | static int | |
4deef538 AO |
803 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
804 | rtx op ATTRIBUTE_UNUSED, | |
805 | rtx dest, rtx src, rtx srcoff, void *arg) | |
6fb5fa3c | 806 | { |
9e582b1d | 807 | insn_info_t insn_info = (insn_info_t) arg; |
dc01c3d1 | 808 | rtx_insn *insn = insn_info->insn, *new_insn, *cur; |
9e582b1d | 809 | note_add_store_info info; |
b8698a0f | 810 | |
4deef538 AO |
811 | /* We can reuse all operands without copying, because we are about |
812 | to delete the insn that contained it. */ | |
9e582b1d | 813 | if (srcoff) |
ed079c4b JJ |
814 | { |
815 | start_sequence (); | |
816 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
817 | new_insn = get_insns (); | |
818 | end_sequence (); | |
819 | } | |
9e582b1d | 820 | else |
1476d1bd | 821 | new_insn = gen_move_insn (dest, src); |
9e582b1d JR |
822 | info.first = new_insn; |
823 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
824 | info.failure = false; | |
825 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
826 | { | |
827 | info.current = cur; | |
e8448ba5 | 828 | note_stores (cur, note_add_store, &info); |
9e582b1d | 829 | } |
6fb5fa3c | 830 | |
9e582b1d JR |
831 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
832 | return it immediately, communicating the failure to its caller. */ | |
833 | if (info.failure) | |
834 | return 1; | |
835 | ||
836 | emit_insn_before (new_insn, insn); | |
6fb5fa3c | 837 | |
8d8e205b | 838 | return 0; |
6fb5fa3c DB |
839 | } |
840 | ||
9e582b1d JR |
841 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
842 | is there, is split into a separate insn. | |
843 | Return true on success (or if there was nothing to do), false on failure. */ | |
6fb5fa3c | 844 | |
9e582b1d JR |
845 | static bool |
846 | check_for_inc_dec_1 (insn_info_t insn_info) | |
6fb5fa3c | 847 | { |
dd60a84c | 848 | rtx_insn *insn = insn_info->insn; |
6fb5fa3c DB |
849 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
850 | if (note) | |
8d8e205b RS |
851 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
852 | insn_info) == 0; | |
2c0fa3ec JJ |
853 | |
854 | /* Punt on stack pushes, those don't have REG_INC notes and we are | |
855 | unprepared to deal with distribution of REG_ARGS_SIZE notes etc. */ | |
856 | subrtx_iterator::array_type array; | |
857 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) | |
858 | { | |
859 | const_rtx x = *iter; | |
860 | if (GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) | |
861 | return false; | |
862 | } | |
863 | ||
9e582b1d | 864 | return true; |
6fb5fa3c DB |
865 | } |
866 | ||
867 | ||
9e582b1d JR |
868 | /* Entry point for postreload. If you work on reload_cse, or you need this |
869 | anywhere else, consider if you can provide register liveness information | |
870 | and add a parameter to this function so that it can be passed down in | |
871 | insn_info.fixed_regs_live. */ | |
872 | bool | |
dd60a84c | 873 | check_for_inc_dec (rtx_insn *insn) |
9e582b1d | 874 | { |
da6603c6 | 875 | insn_info_type insn_info; |
9e582b1d JR |
876 | rtx note; |
877 | ||
878 | insn_info.insn = insn; | |
879 | insn_info.fixed_regs_live = NULL; | |
880 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
881 | if (note) | |
8d8e205b RS |
882 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
883 | &insn_info) == 0; | |
2c0fa3ec JJ |
884 | |
885 | /* Punt on stack pushes, those don't have REG_INC notes and we are | |
886 | unprepared to deal with distribution of REG_ARGS_SIZE notes etc. */ | |
887 | subrtx_iterator::array_type array; | |
888 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) | |
889 | { | |
890 | const_rtx x = *iter; | |
891 | if (GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) | |
892 | return false; | |
893 | } | |
894 | ||
9e582b1d JR |
895 | return true; |
896 | } | |
897 | ||
b8698a0f | 898 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
6fb5fa3c DB |
899 | |
900 | static void | |
901 | delete_dead_store_insn (insn_info_t insn_info) | |
902 | { | |
903 | read_info_t read_info; | |
904 | ||
905 | if (!dbg_cnt (dse)) | |
906 | return; | |
907 | ||
9e582b1d JR |
908 | if (!check_for_inc_dec_1 (insn_info)) |
909 | return; | |
456610d3 | 910 | if (dump_file && (dump_flags & TDF_DETAILS)) |
ac6929b5 RB |
911 | fprintf (dump_file, "Locally deleting insn %d\n", |
912 | INSN_UID (insn_info->insn)); | |
6fb5fa3c DB |
913 | |
914 | free_store_info (insn_info); | |
915 | read_info = insn_info->read_rec; | |
b8698a0f | 916 | |
6fb5fa3c DB |
917 | while (read_info) |
918 | { | |
919 | read_info_t next = read_info->next; | |
fb0b2914 | 920 | read_info_type_pool.remove (read_info); |
6fb5fa3c DB |
921 | read_info = next; |
922 | } | |
923 | insn_info->read_rec = NULL; | |
924 | ||
925 | delete_insn (insn_info->insn); | |
926 | locally_deleted++; | |
927 | insn_info->insn = NULL; | |
928 | ||
929 | insn_info->wild_read = false; | |
930 | } | |
931 | ||
88d8330d EB |
932 | /* Return whether DECL, a local variable, can possibly escape the current |
933 | function scope. */ | |
934 | ||
935 | static bool | |
936 | local_variable_can_escape (tree decl) | |
937 | { | |
938 | if (TREE_ADDRESSABLE (decl)) | |
939 | return true; | |
940 | ||
941 | /* If this is a partitioned variable, we need to consider all the variables | |
942 | in the partition. This is necessary because a store into one of them can | |
943 | be replaced with a store into another and this may not change the outcome | |
944 | of the escape analysis. */ | |
945 | if (cfun->gimple_df->decls_to_pointers != NULL) | |
946 | { | |
39c8aaa4 | 947 | tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); |
88d8330d | 948 | if (namep) |
39c8aaa4 | 949 | return TREE_ADDRESSABLE (*namep); |
88d8330d EB |
950 | } |
951 | ||
952 | return false; | |
953 | } | |
954 | ||
955 | /* Return whether EXPR can possibly escape the current function scope. */ | |
956 | ||
d26c7090 ER |
957 | static bool |
958 | can_escape (tree expr) | |
959 | { | |
960 | tree base; | |
961 | if (!expr) | |
962 | return true; | |
963 | base = get_base_address (expr); | |
964 | if (DECL_P (base) | |
88d8330d | 965 | && !may_be_aliased (base) |
8813a647 | 966 | && !(VAR_P (base) |
88d8330d EB |
967 | && !DECL_EXTERNAL (base) |
968 | && !TREE_STATIC (base) | |
969 | && local_variable_can_escape (base))) | |
d26c7090 ER |
970 | return false; |
971 | return true; | |
972 | } | |
6fb5fa3c DB |
973 | |
974 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
975 | OFFSET and WIDTH. */ | |
976 | ||
977 | static void | |
02ce5d90 | 978 | set_usage_bits (group_info *group, poly_int64 offset, poly_int64 width, |
d26c7090 | 979 | tree expr) |
6fb5fa3c | 980 | { |
02ce5d90 RS |
981 | /* Non-constant offsets and widths act as global kills, so there's no point |
982 | trying to use them to derive global DSE candidates. */ | |
983 | HOST_WIDE_INT i, const_offset, const_width; | |
d26c7090 | 984 | bool expr_escapes = can_escape (expr); |
02ce5d90 RS |
985 | if (offset.is_constant (&const_offset) |
986 | && width.is_constant (&const_width) | |
987 | && const_offset > -MAX_OFFSET | |
988 | && const_offset + const_width < MAX_OFFSET) | |
989 | for (i = const_offset; i < const_offset + const_width; ++i) | |
6fb5fa3c DB |
990 | { |
991 | bitmap store1; | |
992 | bitmap store2; | |
d26c7090 | 993 | bitmap escaped; |
6fb5fa3c DB |
994 | int ai; |
995 | if (i < 0) | |
996 | { | |
997 | store1 = group->store1_n; | |
998 | store2 = group->store2_n; | |
d26c7090 | 999 | escaped = group->escaped_n; |
6fb5fa3c DB |
1000 | ai = -i; |
1001 | } | |
1002 | else | |
1003 | { | |
1004 | store1 = group->store1_p; | |
1005 | store2 = group->store2_p; | |
d26c7090 | 1006 | escaped = group->escaped_p; |
6fb5fa3c DB |
1007 | ai = i; |
1008 | } | |
b8698a0f | 1009 | |
fcaa4ca4 | 1010 | if (!bitmap_set_bit (store1, ai)) |
6fb5fa3c | 1011 | bitmap_set_bit (store2, ai); |
b8698a0f | 1012 | else |
6fb5fa3c | 1013 | { |
6fb5fa3c DB |
1014 | if (i < 0) |
1015 | { | |
1016 | if (group->offset_map_size_n < ai) | |
1017 | group->offset_map_size_n = ai; | |
1018 | } | |
1019 | else | |
1020 | { | |
1021 | if (group->offset_map_size_p < ai) | |
1022 | group->offset_map_size_p = ai; | |
1023 | } | |
1024 | } | |
d26c7090 ER |
1025 | if (expr_escapes) |
1026 | bitmap_set_bit (escaped, ai); | |
6fb5fa3c DB |
1027 | } |
1028 | } | |
1029 | ||
d26c7090 ER |
1030 | static void |
1031 | reset_active_stores (void) | |
1032 | { | |
1033 | active_local_stores = NULL; | |
1034 | active_local_stores_len = 0; | |
1035 | } | |
6fb5fa3c | 1036 | |
d26c7090 | 1037 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
6fb5fa3c DB |
1038 | |
1039 | static void | |
d26c7090 | 1040 | free_read_records (bb_info_t bb_info) |
6fb5fa3c DB |
1041 | { |
1042 | insn_info_t insn_info = bb_info->last_insn; | |
1043 | read_info_t *ptr = &insn_info->read_rec; | |
6fb5fa3c DB |
1044 | while (*ptr) |
1045 | { | |
1046 | read_info_t next = (*ptr)->next; | |
ac6929b5 RB |
1047 | read_info_type_pool.remove (*ptr); |
1048 | *ptr = next; | |
6fb5fa3c | 1049 | } |
d26c7090 ER |
1050 | } |
1051 | ||
1052 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1053 | ||
1054 | static void | |
1055 | add_wild_read (bb_info_t bb_info) | |
1056 | { | |
1057 | insn_info_t insn_info = bb_info->last_insn; | |
6fb5fa3c | 1058 | insn_info->wild_read = true; |
d26c7090 ER |
1059 | free_read_records (bb_info); |
1060 | reset_active_stores (); | |
6fb5fa3c DB |
1061 | } |
1062 | ||
d26c7090 ER |
1063 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1064 | non-frame locations. */ | |
1065 | ||
1066 | static void | |
1067 | add_non_frame_wild_read (bb_info_t bb_info) | |
1068 | { | |
1069 | insn_info_t insn_info = bb_info->last_insn; | |
1070 | insn_info->non_frame_wild_read = true; | |
1071 | free_read_records (bb_info); | |
1072 | reset_active_stores (); | |
1073 | } | |
6fb5fa3c | 1074 | |
50f0f366 EB |
1075 | /* Return true if X is a constant or one of the registers that behave |
1076 | as a constant over the life of a function. This is equivalent to | |
1077 | !rtx_varies_p for memory addresses. */ | |
6fb5fa3c DB |
1078 | |
1079 | static bool | |
1080 | const_or_frame_p (rtx x) | |
1081 | { | |
d8116890 KZ |
1082 | if (CONSTANT_P (x)) |
1083 | return true; | |
1084 | ||
1085 | if (GET_CODE (x) == REG) | |
6fb5fa3c | 1086 | { |
6fb5fa3c DB |
1087 | /* Note that we have to test for the actual rtx used for the frame |
1088 | and arg pointers and not just the register number in case we have | |
1089 | eliminated the frame and/or arg pointer and are using it | |
1090 | for pseudos. */ | |
1091 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1092 | /* The arg pointer varies if it is not a fixed register. */ | |
1093 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1094 | || x == pic_offset_table_rtx) | |
1095 | return true; | |
1096 | return false; | |
6fb5fa3c | 1097 | } |
da6603c6 | 1098 | |
d8116890 | 1099 | return false; |
6fb5fa3c DB |
1100 | } |
1101 | ||
b8698a0f L |
1102 | /* Take all reasonable action to put the address of MEM into the form |
1103 | that we can do analysis on. | |
6fb5fa3c DB |
1104 | |
1105 | The gold standard is to get the address into the form: address + | |
1106 | OFFSET where address is something that rtx_varies_p considers a | |
1107 | constant. When we can get the address in this form, we can do | |
1108 | global analysis on it. Note that for constant bases, address is | |
1109 | not actually returned, only the group_id. The address can be | |
1110 | obtained from that. | |
1111 | ||
1112 | If that fails, we try cselib to get a value we can at least use | |
b8698a0f L |
1113 | locally. If that fails we return false. |
1114 | ||
6fb5fa3c DB |
1115 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1116 | group for non_varying bases. | |
1117 | ||
1118 | FOR_READ is true if this is a mem read and false if not. */ | |
1119 | ||
1120 | static bool | |
1121 | canon_address (rtx mem, | |
6fb5fa3c | 1122 | int *group_id, |
02ce5d90 | 1123 | poly_int64 *offset, |
6fb5fa3c DB |
1124 | cselib_val **base) |
1125 | { | |
ef4bddc2 | 1126 | machine_mode address_mode = get_address_mode (mem); |
6fb5fa3c DB |
1127 | rtx mem_address = XEXP (mem, 0); |
1128 | rtx expanded_address, address; | |
403c7520 JJ |
1129 | int expanded; |
1130 | ||
4deef538 | 1131 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
6fb5fa3c | 1132 | |
456610d3 | 1133 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1134 | { |
1135 | fprintf (dump_file, " mem: "); | |
1136 | print_inline_rtx (dump_file, mem_address, 0); | |
1137 | fprintf (dump_file, "\n"); | |
1138 | } | |
1139 | ||
403c7520 JJ |
1140 | /* First see if just canon_rtx (mem_address) is const or frame, |
1141 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1142 | address = NULL_RTX; | |
1143 | for (expanded = 0; expanded < 2; expanded++) | |
1144 | { | |
1145 | if (expanded) | |
1146 | { | |
1147 | /* Use cselib to replace all of the reg references with the full | |
b8698a0f | 1148 | expression. This will take care of the case where we have |
6fb5fa3c | 1149 | |
403c7520 JJ |
1150 | r_x = base + offset; |
1151 | val = *r_x; | |
b8698a0f L |
1152 | |
1153 | by making it into | |
6fb5fa3c | 1154 | |
403c7520 | 1155 | val = *(base + offset); */ |
6fb5fa3c | 1156 | |
403c7520 JJ |
1157 | expanded_address = cselib_expand_value_rtx (mem_address, |
1158 | scratch, 5); | |
6fb5fa3c | 1159 | |
403c7520 JJ |
1160 | /* If this fails, just go with the address from first |
1161 | iteration. */ | |
1162 | if (!expanded_address) | |
1163 | break; | |
1164 | } | |
1165 | else | |
1166 | expanded_address = mem_address; | |
6fb5fa3c | 1167 | |
403c7520 JJ |
1168 | /* Split the address into canonical BASE + OFFSET terms. */ |
1169 | address = canon_rtx (expanded_address); | |
6fb5fa3c | 1170 | |
403c7520 | 1171 | *offset = 0; |
6fb5fa3c | 1172 | |
456610d3 | 1173 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1174 | { |
1175 | if (expanded) | |
1176 | { | |
1177 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1178 | print_inline_rtx (dump_file, expanded_address, 0); | |
1179 | fprintf (dump_file, "\n"); | |
1180 | } | |
6fb5fa3c | 1181 | |
403c7520 JJ |
1182 | fprintf (dump_file, "\n after canon_rtx address: "); |
1183 | print_inline_rtx (dump_file, address, 0); | |
1184 | fprintf (dump_file, "\n"); | |
1185 | } | |
6fb5fa3c | 1186 | |
403c7520 JJ |
1187 | if (GET_CODE (address) == CONST) |
1188 | address = XEXP (address, 0); | |
6fb5fa3c | 1189 | |
02ce5d90 | 1190 | address = strip_offset_and_add (address, offset); |
6fb5fa3c | 1191 | |
09e881c9 BE |
1192 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1193 | && const_or_frame_p (address)) | |
6fb5fa3c | 1194 | { |
fd6daec9 | 1195 | group_info *group = get_group_info (address); |
403c7520 | 1196 | |
456610d3 | 1197 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
1198 | { |
1199 | fprintf (dump_file, " gid=%d offset=", group->id); | |
1200 | print_dec (*offset, dump_file); | |
1201 | fprintf (dump_file, "\n"); | |
1202 | } | |
403c7520 JJ |
1203 | *base = NULL; |
1204 | *group_id = group->id; | |
1205 | return true; | |
6fb5fa3c | 1206 | } |
403c7520 JJ |
1207 | } |
1208 | ||
4deef538 | 1209 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
403c7520 JJ |
1210 | *group_id = -1; |
1211 | ||
1212 | if (*base == NULL) | |
1213 | { | |
456610d3 | 1214 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1215 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1216 | return false; | |
6fb5fa3c | 1217 | } |
456610d3 | 1218 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
1219 | { |
1220 | fprintf (dump_file, " varying cselib base=%u:%u offset = ", | |
1221 | (*base)->uid, (*base)->hash); | |
1222 | print_dec (*offset, dump_file); | |
1223 | fprintf (dump_file, "\n"); | |
1224 | } | |
6fb5fa3c DB |
1225 | return true; |
1226 | } | |
1227 | ||
1228 | ||
1229 | /* Clear the rhs field from the active_local_stores array. */ | |
1230 | ||
1231 | static void | |
1232 | clear_rhs_from_active_local_stores (void) | |
1233 | { | |
1234 | insn_info_t ptr = active_local_stores; | |
1235 | ||
1236 | while (ptr) | |
1237 | { | |
fd6daec9 | 1238 | store_info *store_info = ptr->store_rec; |
6fb5fa3c DB |
1239 | /* Skip the clobbers. */ |
1240 | while (!store_info->is_set) | |
1241 | store_info = store_info->next; | |
1242 | ||
1243 | store_info->rhs = NULL; | |
8dd5516b | 1244 | store_info->const_rhs = NULL; |
6fb5fa3c DB |
1245 | |
1246 | ptr = ptr->next_local_store; | |
1247 | } | |
1248 | } | |
1249 | ||
1250 | ||
8dd5516b JJ |
1251 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1252 | ||
1253 | static inline void | |
fd6daec9 | 1254 | set_position_unneeded (store_info *s_info, int pos) |
8dd5516b | 1255 | { |
22d9c880 | 1256 | if (UNLIKELY (s_info->is_large)) |
8dd5516b | 1257 | { |
fcaa4ca4 NF |
1258 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1259 | s_info->positions_needed.large.count++; | |
8dd5516b JJ |
1260 | } |
1261 | else | |
1262 | s_info->positions_needed.small_bitmask | |
fecfbfa4 | 1263 | &= ~(HOST_WIDE_INT_1U << pos); |
8dd5516b JJ |
1264 | } |
1265 | ||
1266 | /* Mark the whole store S_INFO as unneeded. */ | |
1267 | ||
1268 | static inline void | |
fd6daec9 | 1269 | set_all_positions_unneeded (store_info *s_info) |
8dd5516b | 1270 | { |
22d9c880 | 1271 | if (UNLIKELY (s_info->is_large)) |
8dd5516b | 1272 | { |
02ce5d90 RS |
1273 | HOST_WIDE_INT width; |
1274 | if (s_info->width.is_constant (&width)) | |
1275 | { | |
1276 | bitmap_set_range (s_info->positions_needed.large.bmap, 0, width); | |
1277 | s_info->positions_needed.large.count = width; | |
1278 | } | |
1279 | else | |
1280 | { | |
1281 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1282 | s_info->positions_needed.large.count = 1; | |
1283 | } | |
8dd5516b JJ |
1284 | } |
1285 | else | |
07e96250 | 1286 | s_info->positions_needed.small_bitmask = HOST_WIDE_INT_0U; |
8dd5516b JJ |
1287 | } |
1288 | ||
1289 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1290 | ||
1291 | static inline bool | |
fd6daec9 | 1292 | any_positions_needed_p (store_info *s_info) |
8dd5516b | 1293 | { |
22d9c880 | 1294 | if (UNLIKELY (s_info->is_large)) |
02ce5d90 RS |
1295 | { |
1296 | HOST_WIDE_INT width; | |
1297 | if (s_info->width.is_constant (&width)) | |
1298 | { | |
1299 | gcc_checking_assert (s_info->positions_needed.large.bmap); | |
1300 | return s_info->positions_needed.large.count < width; | |
1301 | } | |
1302 | else | |
1303 | { | |
1304 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1305 | return s_info->positions_needed.large.count == 0; | |
1306 | } | |
1307 | } | |
8dd5516b | 1308 | else |
07e96250 | 1309 | return (s_info->positions_needed.small_bitmask != HOST_WIDE_INT_0U); |
8dd5516b JJ |
1310 | } |
1311 | ||
1312 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
02ce5d90 | 1313 | store are known to be needed. */ |
8dd5516b JJ |
1314 | |
1315 | static inline bool | |
02ce5d90 RS |
1316 | all_positions_needed_p (store_info *s_info, poly_int64 start, |
1317 | poly_int64 width) | |
8dd5516b | 1318 | { |
02ce5d90 RS |
1319 | gcc_assert (s_info->rhs); |
1320 | if (!s_info->width.is_constant ()) | |
1321 | { | |
1322 | gcc_assert (s_info->is_large | |
1323 | && !s_info->positions_needed.large.bmap); | |
1324 | return s_info->positions_needed.large.count == 0; | |
1325 | } | |
1326 | ||
1327 | /* Otherwise, if START and WIDTH are non-constant, we're asking about | |
1328 | a non-constant region of a constant-sized store. We can't say for | |
1329 | sure that all positions are needed. */ | |
1330 | HOST_WIDE_INT const_start, const_width; | |
1331 | if (!start.is_constant (&const_start) | |
1332 | || !width.is_constant (&const_width)) | |
1333 | return false; | |
1334 | ||
22d9c880 | 1335 | if (UNLIKELY (s_info->is_large)) |
8dd5516b | 1336 | { |
02ce5d90 RS |
1337 | for (HOST_WIDE_INT i = const_start; i < const_start + const_width; ++i) |
1338 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, i)) | |
8dd5516b JJ |
1339 | return false; |
1340 | return true; | |
1341 | } | |
1342 | else | |
1343 | { | |
02ce5d90 RS |
1344 | unsigned HOST_WIDE_INT mask |
1345 | = lowpart_bitmask (const_width) << const_start; | |
8dd5516b JJ |
1346 | return (s_info->positions_needed.small_bitmask & mask) == mask; |
1347 | } | |
1348 | } | |
1349 | ||
1350 | ||
02ce5d90 RS |
1351 | static rtx get_stored_val (store_info *, machine_mode, poly_int64, |
1352 | poly_int64, basic_block, bool); | |
8dd5516b JJ |
1353 | |
1354 | ||
6fb5fa3c DB |
1355 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1356 | there is a candidate store, after adding it to the appropriate | |
1357 | local store group if so. */ | |
1358 | ||
1359 | static int | |
1360 | record_store (rtx body, bb_info_t bb_info) | |
1361 | { | |
6216f94e | 1362 | rtx mem, rhs, const_rhs, mem_addr; |
02ce5d90 RS |
1363 | poly_int64 offset = 0; |
1364 | poly_int64 width = 0; | |
6fb5fa3c | 1365 | insn_info_t insn_info = bb_info->last_insn; |
fd6daec9 | 1366 | store_info *store_info = NULL; |
6fb5fa3c DB |
1367 | int group_id; |
1368 | cselib_val *base = NULL; | |
8dd5516b | 1369 | insn_info_t ptr, last, redundant_reason; |
6fb5fa3c DB |
1370 | bool store_is_unused; |
1371 | ||
1372 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1373 | return 0; | |
1374 | ||
8dd5516b JJ |
1375 | mem = SET_DEST (body); |
1376 | ||
6fb5fa3c DB |
1377 | /* If this is not used, then this cannot be used to keep the insn |
1378 | from being deleted. On the other hand, it does provide something | |
1379 | that can be used to prove that another store is dead. */ | |
1380 | store_is_unused | |
8dd5516b | 1381 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
6fb5fa3c DB |
1382 | |
1383 | /* Check whether that value is a suitable memory location. */ | |
6fb5fa3c DB |
1384 | if (!MEM_P (mem)) |
1385 | { | |
1386 | /* If the set or clobber is unused, then it does not effect our | |
1387 | ability to get rid of the entire insn. */ | |
1388 | if (!store_is_unused) | |
1389 | insn_info->cannot_delete = true; | |
1390 | return 0; | |
1391 | } | |
1392 | ||
1393 | /* At this point we know mem is a mem. */ | |
1394 | if (GET_MODE (mem) == BLKmode) | |
1395 | { | |
d05d7551 | 1396 | HOST_WIDE_INT const_size; |
6fb5fa3c DB |
1397 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) |
1398 | { | |
456610d3 | 1399 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1400 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1401 | add_wild_read (bb_info); | |
1402 | insn_info->cannot_delete = true; | |
8dd5516b | 1403 | return 0; |
6fb5fa3c | 1404 | } |
8dd5516b JJ |
1405 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1406 | as memset (addr, 0, 36); */ | |
f5541398 | 1407 | else if (!MEM_SIZE_KNOWN_P (mem) |
d05d7551 RS |
1408 | || maybe_le (MEM_SIZE (mem), 0) |
1409 | /* This is a limit on the bitmap size, which is only relevant | |
1410 | for constant-sized MEMs. */ | |
1411 | || (MEM_SIZE (mem).is_constant (&const_size) | |
1412 | && const_size > MAX_OFFSET) | |
8dd5516b | 1413 | || GET_CODE (body) != SET |
8dd5516b | 1414 | || !CONST_INT_P (SET_SRC (body))) |
6fb5fa3c | 1415 | { |
8dd5516b JJ |
1416 | if (!store_is_unused) |
1417 | { | |
1418 | /* If the set or clobber is unused, then it does not effect our | |
1419 | ability to get rid of the entire insn. */ | |
1420 | insn_info->cannot_delete = true; | |
1421 | clear_rhs_from_active_local_stores (); | |
1422 | } | |
1423 | return 0; | |
6fb5fa3c | 1424 | } |
6fb5fa3c DB |
1425 | } |
1426 | ||
1427 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1428 | if (MEM_VOLATILE_P (mem)) | |
8dd5516b | 1429 | insn_info->cannot_delete = true; |
6fb5fa3c | 1430 | |
ac6929b5 | 1431 | if (!canon_address (mem, &group_id, &offset, &base)) |
6fb5fa3c DB |
1432 | { |
1433 | clear_rhs_from_active_local_stores (); | |
1434 | return 0; | |
1435 | } | |
1436 | ||
8dd5516b | 1437 | if (GET_MODE (mem) == BLKmode) |
f5541398 | 1438 | width = MEM_SIZE (mem); |
8dd5516b | 1439 | else |
6f557e0e | 1440 | width = GET_MODE_SIZE (GET_MODE (mem)); |
6fb5fa3c | 1441 | |
02ce5d90 | 1442 | if (!endpoint_representable_p (offset, width)) |
213ffde2 JJ |
1443 | { |
1444 | clear_rhs_from_active_local_stores (); | |
1445 | return 0; | |
1446 | } | |
1447 | ||
cbe679a4 JJ |
1448 | if (known_eq (width, 0)) |
1449 | return 0; | |
1450 | ||
ac6929b5 | 1451 | if (group_id >= 0) |
6fb5fa3c DB |
1452 | { |
1453 | /* In the restrictive case where the base is a constant or the | |
1454 | frame pointer we can do global analysis. */ | |
b8698a0f | 1455 | |
fd6daec9 | 1456 | group_info *group |
9771b263 | 1457 | = rtx_group_vec[group_id]; |
d26c7090 | 1458 | tree expr = MEM_EXPR (mem); |
b8698a0f | 1459 | |
da6603c6 | 1460 | store_info = rtx_store_info_pool.allocate (); |
d26c7090 | 1461 | set_usage_bits (group, offset, width, expr); |
6fb5fa3c | 1462 | |
456610d3 | 1463 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1464 | { |
1465 | fprintf (dump_file, " processing const base store gid=%d", | |
1466 | group_id); | |
1467 | print_range (dump_file, offset, width); | |
1468 | fprintf (dump_file, "\n"); | |
1469 | } | |
6fb5fa3c DB |
1470 | } |
1471 | else | |
1472 | { | |
9e412ca3 | 1473 | if (may_be_sp_based_p (XEXP (mem, 0))) |
50f0f366 | 1474 | insn_info->stack_pointer_based = true; |
6fb5fa3c | 1475 | insn_info->contains_cselib_groups = true; |
50f0f366 | 1476 | |
da6603c6 | 1477 | store_info = cse_store_info_pool.allocate (); |
6fb5fa3c DB |
1478 | group_id = -1; |
1479 | ||
456610d3 | 1480 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1481 | { |
1482 | fprintf (dump_file, " processing cselib store "); | |
1483 | print_range (dump_file, offset, width); | |
1484 | fprintf (dump_file, "\n"); | |
1485 | } | |
6fb5fa3c DB |
1486 | } |
1487 | ||
8dd5516b JJ |
1488 | const_rhs = rhs = NULL_RTX; |
1489 | if (GET_CODE (body) == SET | |
1490 | /* No place to keep the value after ra. */ | |
1491 | && !reload_completed | |
1492 | && (REG_P (SET_SRC (body)) | |
1493 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1494 | || CONSTANT_P (SET_SRC (body))) | |
1495 | && !MEM_VOLATILE_P (mem) | |
1496 | /* Sometimes the store and reload is used for truncation and | |
1497 | rounding. */ | |
1498 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1499 | { | |
1500 | rhs = SET_SRC (body); | |
1501 | if (CONSTANT_P (rhs)) | |
1502 | const_rhs = rhs; | |
1503 | else if (body == PATTERN (insn_info->insn)) | |
1504 | { | |
1505 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1506 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1507 | const_rhs = XEXP (tem, 0); | |
1508 | } | |
1509 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1510 | { | |
1511 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1512 | ||
1513 | if (tem && CONSTANT_P (tem)) | |
1514 | const_rhs = tem; | |
a743a727 L |
1515 | else |
1516 | { | |
1517 | /* If RHS is set only once to a constant, set CONST_RHS | |
1518 | to the constant. */ | |
1519 | rtx def_src = df_find_single_def_src (rhs); | |
1520 | if (def_src != nullptr && CONSTANT_P (def_src)) | |
1521 | const_rhs = def_src; | |
1522 | } | |
8dd5516b JJ |
1523 | } |
1524 | } | |
1525 | ||
6fb5fa3c DB |
1526 | /* Check to see if this stores causes some other stores to be |
1527 | dead. */ | |
1528 | ptr = active_local_stores; | |
1529 | last = NULL; | |
8dd5516b | 1530 | redundant_reason = NULL; |
64ca6aa7 | 1531 | unsigned char addrspace = MEM_ADDR_SPACE (mem); |
6216f94e | 1532 | mem = canon_rtx (mem); |
ac6929b5 RB |
1533 | |
1534 | if (group_id < 0) | |
1535 | mem_addr = base->val_rtx; | |
6216f94e JJ |
1536 | else |
1537 | { | |
ac6929b5 RB |
1538 | group_info *group = rtx_group_vec[group_id]; |
1539 | mem_addr = group->canon_base_addr; | |
6216f94e | 1540 | } |
02ce5d90 | 1541 | if (maybe_ne (offset, 0)) |
ac6929b5 | 1542 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6fb5fa3c DB |
1543 | |
1544 | while (ptr) | |
1545 | { | |
1546 | insn_info_t next = ptr->next_local_store; | |
99b1c316 | 1547 | class store_info *s_info = ptr->store_rec; |
60564289 | 1548 | bool del = true; |
6fb5fa3c DB |
1549 | |
1550 | /* Skip the clobbers. We delete the active insn if this insn | |
6ed3da00 | 1551 | shadows the set. To have been put on the active list, it |
6fb5fa3c DB |
1552 | has exactly on set. */ |
1553 | while (!s_info->is_set) | |
1554 | s_info = s_info->next; | |
1555 | ||
64ca6aa7 AP |
1556 | if (s_info->group_id == group_id |
1557 | && s_info->cse_base == base | |
1558 | && s_info->addrspace == addrspace) | |
6fb5fa3c DB |
1559 | { |
1560 | HOST_WIDE_INT i; | |
456610d3 | 1561 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1562 | { |
1563 | fprintf (dump_file, " trying store in insn=%d gid=%d", | |
1564 | INSN_UID (ptr->insn), s_info->group_id); | |
1565 | print_range (dump_file, s_info->offset, s_info->width); | |
1566 | fprintf (dump_file, "\n"); | |
1567 | } | |
8dd5516b JJ |
1568 | |
1569 | /* Even if PTR won't be eliminated as unneeded, if both | |
1570 | PTR and this insn store the same constant value, we might | |
1571 | eliminate this insn instead. */ | |
1572 | if (s_info->const_rhs | |
1573 | && const_rhs | |
71d6a386 RS |
1574 | && known_subrange_p (offset, width, |
1575 | s_info->offset, s_info->width) | |
1576 | && all_positions_needed_p (s_info, offset - s_info->offset, | |
d7e1f499 RB |
1577 | width) |
1578 | /* We can only remove the later store if the earlier aliases | |
1579 | at least all accesses the later one. */ | |
64ce76d9 | 1580 | && mems_same_for_tbaa_p (s_info->mem, mem)) |
8dd5516b JJ |
1581 | { |
1582 | if (GET_MODE (mem) == BLKmode) | |
1583 | { | |
1584 | if (GET_MODE (s_info->mem) == BLKmode | |
1585 | && s_info->const_rhs == const_rhs) | |
1586 | redundant_reason = ptr; | |
1587 | } | |
1588 | else if (s_info->const_rhs == const0_rtx | |
1589 | && const_rhs == const0_rtx) | |
1590 | redundant_reason = ptr; | |
1591 | else | |
1592 | { | |
1593 | rtx val; | |
1594 | start_sequence (); | |
71d6a386 | 1595 | val = get_stored_val (s_info, GET_MODE (mem), offset, width, |
8dd5516b JJ |
1596 | BLOCK_FOR_INSN (insn_info->insn), |
1597 | true); | |
1598 | if (get_insns () != NULL) | |
1599 | val = NULL_RTX; | |
1600 | end_sequence (); | |
1601 | if (val && rtx_equal_p (val, const_rhs)) | |
1602 | redundant_reason = ptr; | |
1603 | } | |
1604 | } | |
1605 | ||
02ce5d90 | 1606 | HOST_WIDE_INT begin_unneeded, const_s_width, const_width; |
71d6a386 RS |
1607 | if (known_subrange_p (s_info->offset, s_info->width, offset, width)) |
1608 | /* The new store touches every byte that S_INFO does. */ | |
1609 | set_all_positions_unneeded (s_info); | |
02ce5d90 RS |
1610 | else if ((offset - s_info->offset).is_constant (&begin_unneeded) |
1611 | && s_info->width.is_constant (&const_s_width) | |
1612 | && width.is_constant (&const_width)) | |
71d6a386 | 1613 | { |
02ce5d90 | 1614 | HOST_WIDE_INT end_unneeded = begin_unneeded + const_width; |
71d6a386 | 1615 | begin_unneeded = MAX (begin_unneeded, 0); |
02ce5d90 | 1616 | end_unneeded = MIN (end_unneeded, const_s_width); |
71d6a386 RS |
1617 | for (i = begin_unneeded; i < end_unneeded; ++i) |
1618 | set_position_unneeded (s_info, i); | |
1619 | } | |
02ce5d90 RS |
1620 | else |
1621 | { | |
1622 | /* We don't know which parts of S_INFO are needed and | |
1623 | which aren't, so invalidate the RHS. */ | |
1624 | s_info->rhs = NULL; | |
1625 | s_info->const_rhs = NULL; | |
1626 | } | |
6fb5fa3c DB |
1627 | } |
1628 | else if (s_info->rhs) | |
1629 | /* Need to see if it is possible for this store to overwrite | |
1630 | the value of store_info. If it is, set the rhs to NULL to | |
1631 | keep it from being used to remove a load. */ | |
1632 | { | |
43b9f499 RB |
1633 | if (canon_output_dependence (s_info->mem, true, |
1634 | mem, GET_MODE (mem), | |
1635 | mem_addr)) | |
8dd5516b JJ |
1636 | { |
1637 | s_info->rhs = NULL; | |
1638 | s_info->const_rhs = NULL; | |
1639 | } | |
6fb5fa3c | 1640 | } |
6216f94e | 1641 | |
6fb5fa3c DB |
1642 | /* An insn can be deleted if every position of every one of |
1643 | its s_infos is zero. */ | |
1b6fa860 | 1644 | if (any_positions_needed_p (s_info)) |
60564289 | 1645 | del = false; |
8dd5516b | 1646 | |
60564289 | 1647 | if (del) |
6fb5fa3c DB |
1648 | { |
1649 | insn_info_t insn_to_delete = ptr; | |
b8698a0f | 1650 | |
dabd47e7 | 1651 | active_local_stores_len--; |
6fb5fa3c DB |
1652 | if (last) |
1653 | last->next_local_store = ptr->next_local_store; | |
1654 | else | |
1655 | active_local_stores = ptr->next_local_store; | |
b8698a0f | 1656 | |
1b6fa860 JJ |
1657 | if (!insn_to_delete->cannot_delete) |
1658 | delete_dead_store_insn (insn_to_delete); | |
6fb5fa3c DB |
1659 | } |
1660 | else | |
1661 | last = ptr; | |
b8698a0f | 1662 | |
6fb5fa3c DB |
1663 | ptr = next; |
1664 | } | |
b8698a0f | 1665 | |
6fb5fa3c DB |
1666 | /* Finish filling in the store_info. */ |
1667 | store_info->next = insn_info->store_rec; | |
1668 | insn_info->store_rec = store_info; | |
6216f94e | 1669 | store_info->mem = mem; |
6216f94e | 1670 | store_info->mem_addr = mem_addr; |
6fb5fa3c | 1671 | store_info->cse_base = base; |
02ce5d90 RS |
1672 | HOST_WIDE_INT const_width; |
1673 | if (!width.is_constant (&const_width)) | |
1674 | { | |
1675 | store_info->is_large = true; | |
1676 | store_info->positions_needed.large.count = 0; | |
1677 | store_info->positions_needed.large.bmap = NULL; | |
1678 | } | |
1679 | else if (const_width > HOST_BITS_PER_WIDE_INT) | |
8dd5516b JJ |
1680 | { |
1681 | store_info->is_large = true; | |
1682 | store_info->positions_needed.large.count = 0; | |
3f9b14ff | 1683 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
8dd5516b JJ |
1684 | } |
1685 | else | |
1686 | { | |
1687 | store_info->is_large = false; | |
02ce5d90 RS |
1688 | store_info->positions_needed.small_bitmask |
1689 | = lowpart_bitmask (const_width); | |
8dd5516b | 1690 | } |
6fb5fa3c | 1691 | store_info->group_id = group_id; |
71d6a386 RS |
1692 | store_info->offset = offset; |
1693 | store_info->width = width; | |
6fb5fa3c | 1694 | store_info->is_set = GET_CODE (body) == SET; |
8dd5516b JJ |
1695 | store_info->rhs = rhs; |
1696 | store_info->const_rhs = const_rhs; | |
1697 | store_info->redundant_reason = redundant_reason; | |
64ca6aa7 | 1698 | store_info->addrspace = addrspace; |
6fb5fa3c | 1699 | |
6fb5fa3c DB |
1700 | /* If this is a clobber, we return 0. We will only be able to |
1701 | delete this insn if there is only one store USED store, but we | |
1702 | can use the clobber to delete other stores earlier. */ | |
1703 | return store_info->is_set ? 1 : 0; | |
1704 | } | |
1705 | ||
1706 | ||
1707 | static void | |
1708 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1709 | { | |
b8698a0f | 1710 | fprintf (dump_file, "%s insn=%d %s\n", start, |
6fb5fa3c DB |
1711 | INSN_UID (insn_info->insn), |
1712 | insn_info->store_rec ? "has store" : "naked"); | |
1713 | } | |
1714 | ||
1715 | ||
8660aaae EC |
1716 | /* If the modes are different and the value's source and target do not |
1717 | line up, we need to extract the value from lower part of the rhs of | |
1718 | the store, shift it, and then put it into a form that can be shoved | |
1719 | into the read_insn. This function generates a right SHIFT of a | |
1720 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1721 | shift sequence is returned or NULL if we failed to find a | |
1722 | shift. */ | |
1723 | ||
1724 | static rtx | |
02ce5d90 | 1725 | find_shift_sequence (poly_int64 access_size, |
fd6daec9 | 1726 | store_info *store_info, |
ef4bddc2 | 1727 | machine_mode read_mode, |
02ce5d90 | 1728 | poly_int64 shift, bool speed, bool require_cst) |
8660aaae | 1729 | { |
ef4bddc2 | 1730 | machine_mode store_mode = GET_MODE (store_info->mem); |
f67f4dff | 1731 | scalar_int_mode new_mode; |
18b526e8 | 1732 | rtx read_reg = NULL; |
8660aaae | 1733 | |
265d817b XL |
1734 | /* If a constant was stored into memory, try to simplify it here, |
1735 | otherwise the cost of the shift might preclude this optimization | |
1736 | e.g. at -Os, even when no actual shift will be needed. */ | |
073849da JZ |
1737 | if (store_info->const_rhs |
1738 | && known_le (access_size, GET_MODE_SIZE (MAX_MODE_INT))) | |
265d817b XL |
1739 | { |
1740 | auto new_mode = smallest_int_mode_for_size (access_size * BITS_PER_UNIT); | |
1741 | auto byte = subreg_lowpart_offset (new_mode, store_mode); | |
1742 | rtx ret | |
1743 | = simplify_subreg (new_mode, store_info->const_rhs, store_mode, byte); | |
1744 | if (ret && CONSTANT_P (ret)) | |
1745 | { | |
1746 | rtx shift_rtx = gen_int_shift_amount (new_mode, shift); | |
1747 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, ret, | |
1748 | shift_rtx); | |
1749 | if (ret && CONSTANT_P (ret)) | |
1750 | { | |
1751 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1752 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1753 | if (ret && CONSTANT_P (ret) | |
1754 | && (set_src_cost (ret, read_mode, speed) | |
1755 | <= COSTS_N_INSNS (1))) | |
1756 | return ret; | |
1757 | } | |
1758 | } | |
1759 | } | |
1760 | ||
1761 | if (require_cst) | |
1762 | return NULL_RTX; | |
1763 | ||
8660aaae EC |
1764 | /* Some machines like the x86 have shift insns for each size of |
1765 | operand. Other machines like the ppc or the ia-64 may only have | |
1766 | shift insns that shift values within 32 or 64 bit registers. | |
1767 | This loop tries to find the smallest shift insn that will right | |
1768 | justify the value we want to read but is available in one insn on | |
1769 | the machine. */ | |
1770 | ||
f67f4dff | 1771 | opt_scalar_int_mode new_mode_iter; |
f835e9f6 | 1772 | FOR_EACH_MODE_IN_CLASS (new_mode_iter, MODE_INT) |
8660aaae | 1773 | { |
dc01c3d1 DM |
1774 | rtx target, new_reg, new_lhs; |
1775 | rtx_insn *shift_seq, *insn; | |
d898d29b | 1776 | int cost; |
348eea5f | 1777 | |
f67f4dff | 1778 | new_mode = new_mode_iter.require (); |
c94843d2 RS |
1779 | if (GET_MODE_BITSIZE (new_mode) > BITS_PER_WORD) |
1780 | break; | |
f835e9f6 RS |
1781 | if (maybe_lt (GET_MODE_SIZE (new_mode), GET_MODE_SIZE (read_mode))) |
1782 | continue; | |
c94843d2 | 1783 | |
18b526e8 RS |
1784 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1785 | requires a real instruction. */ | |
73a699ae | 1786 | if (maybe_lt (GET_MODE_SIZE (new_mode), GET_MODE_SIZE (store_mode)) |
d0edd768 | 1787 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
348eea5f RS |
1788 | continue; |
1789 | ||
18b526e8 RS |
1790 | /* Also try a wider mode if the necessary punning is either not |
1791 | desirable or not possible. */ | |
1792 | if (!CONSTANT_P (store_info->rhs) | |
99e1629f | 1793 | && !targetm.modes_tieable_p (new_mode, store_mode)) |
18b526e8 | 1794 | continue; |
18b526e8 | 1795 | |
265d817b XL |
1796 | if (multiple_p (shift, GET_MODE_BITSIZE (new_mode)) |
1797 | && known_le (GET_MODE_SIZE (new_mode), GET_MODE_SIZE (store_mode))) | |
1798 | { | |
1799 | /* Try to implement the shift using a subreg. */ | |
1800 | poly_int64 offset | |
1801 | = subreg_offset_from_lsb (new_mode, store_mode, shift); | |
1802 | rtx rhs_subreg = simplify_gen_subreg (new_mode, store_info->rhs, | |
1803 | store_mode, offset); | |
1804 | if (rhs_subreg) | |
1805 | { | |
1806 | read_reg | |
1807 | = extract_low_bits (read_mode, new_mode, copy_rtx (rhs_subreg)); | |
1808 | break; | |
1809 | } | |
1810 | } | |
1811 | ||
1812 | if (maybe_lt (GET_MODE_SIZE (new_mode), access_size)) | |
1813 | continue; | |
1814 | ||
348eea5f | 1815 | new_reg = gen_reg_rtx (new_mode); |
8660aaae EC |
1816 | |
1817 | start_sequence (); | |
1818 | ||
1819 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1820 | of one dsp where the cost of these two was not the same. But | |
1821 | this really is a rare case anyway. */ | |
1822 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
abd3c800 RS |
1823 | gen_int_shift_amount (new_mode, shift), |
1824 | new_reg, 1, OPTAB_DIRECT); | |
8660aaae | 1825 | |
c6f3019a RS |
1826 | shift_seq = get_insns (); |
1827 | end_sequence (); | |
8660aaae | 1828 | |
c6f3019a RS |
1829 | if (target != new_reg || shift_seq == NULL) |
1830 | continue; | |
1831 | ||
1832 | cost = 0; | |
1833 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1834 | if (INSN_P (insn)) | |
d09e78e8 | 1835 | cost += insn_cost (insn, speed); |
c6f3019a RS |
1836 | |
1837 | /* The computation up to here is essentially independent | |
1838 | of the arguments and could be precomputed. It may | |
1839 | not be worth doing so. We could precompute if | |
1840 | worthwhile or at least cache the results. The result | |
06acf7d0 RS |
1841 | technically depends on both SHIFT and ACCESS_SIZE, |
1842 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
c6f3019a RS |
1843 | |
1844 | if (cost > COSTS_N_INSNS (1)) | |
1845 | continue; | |
1846 | ||
d898d29b JJ |
1847 | new_lhs = extract_low_bits (new_mode, store_mode, |
1848 | copy_rtx (store_info->rhs)); | |
1849 | if (new_lhs == NULL_RTX) | |
1850 | continue; | |
1851 | ||
c6f3019a RS |
1852 | /* We found an acceptable shift. Generate a move to |
1853 | take the value from the store and put it into the | |
1854 | shift pseudo, then shift it, then generate another | |
1855 | move to put in into the target of the read. */ | |
18b526e8 | 1856 | emit_move_insn (new_reg, new_lhs); |
c6f3019a | 1857 | emit_insn (shift_seq); |
18b526e8 | 1858 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
c6f3019a | 1859 | break; |
8660aaae EC |
1860 | } |
1861 | ||
18b526e8 | 1862 | return read_reg; |
8660aaae EC |
1863 | } |
1864 | ||
1865 | ||
02b47899 KZ |
1866 | /* Call back for note_stores to find the hard regs set or clobbered by |
1867 | insn. Data is a bitmap of the hardregs set so far. */ | |
1868 | ||
1869 | static void | |
1870 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1871 | { | |
1872 | bitmap regs_set = (bitmap) data; | |
1873 | ||
1874 | if (REG_P (x) | |
f773c2bd | 1875 | && HARD_REGISTER_P (x)) |
dc8afb70 | 1876 | bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); |
02b47899 KZ |
1877 | } |
1878 | ||
8dd5516b | 1879 | /* Helper function for replace_read and record_store. |
71d6a386 RS |
1880 | Attempt to return a value of mode READ_MODE stored in STORE_INFO, |
1881 | consisting of READ_WIDTH bytes starting from READ_OFFSET. Return NULL | |
8dd5516b JJ |
1882 | if not successful. If REQUIRE_CST is true, return always constant. */ |
1883 | ||
1884 | static rtx | |
fd6daec9 | 1885 | get_stored_val (store_info *store_info, machine_mode read_mode, |
02ce5d90 | 1886 | poly_int64 read_offset, poly_int64 read_width, |
8dd5516b JJ |
1887 | basic_block bb, bool require_cst) |
1888 | { | |
ef4bddc2 | 1889 | machine_mode store_mode = GET_MODE (store_info->mem); |
02ce5d90 | 1890 | poly_int64 gap; |
8dd5516b JJ |
1891 | rtx read_reg; |
1892 | ||
1893 | /* To get here the read is within the boundaries of the write so | |
1894 | shift will never be negative. Start out with the shift being in | |
1895 | bytes. */ | |
1896 | if (store_mode == BLKmode) | |
71d6a386 | 1897 | gap = 0; |
8dd5516b | 1898 | else if (BYTES_BIG_ENDIAN) |
71d6a386 RS |
1899 | gap = ((store_info->offset + store_info->width) |
1900 | - (read_offset + read_width)); | |
8dd5516b | 1901 | else |
71d6a386 | 1902 | gap = read_offset - store_info->offset; |
8dd5516b | 1903 | |
990769a3 | 1904 | if (maybe_ne (gap, 0)) |
71d6a386 | 1905 | { |
990769a3 PL |
1906 | if (!gap.is_constant ()) |
1907 | return NULL_RTX; | |
1908 | ||
02ce5d90 RS |
1909 | poly_int64 shift = gap * BITS_PER_UNIT; |
1910 | poly_int64 access_size = GET_MODE_SIZE (read_mode) + gap; | |
71d6a386 RS |
1911 | read_reg = find_shift_sequence (access_size, store_info, read_mode, |
1912 | shift, optimize_bb_for_speed_p (bb), | |
1913 | require_cst); | |
1914 | } | |
8dd5516b JJ |
1915 | else if (store_mode == BLKmode) |
1916 | { | |
1917 | /* The store is a memset (addr, const_val, const_size). */ | |
1918 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
304b9962 RS |
1919 | scalar_int_mode int_store_mode; |
1920 | if (!int_mode_for_mode (read_mode).exists (&int_store_mode)) | |
8dd5516b JJ |
1921 | read_reg = NULL_RTX; |
1922 | else if (store_info->rhs == const0_rtx) | |
304b9962 RS |
1923 | read_reg = extract_low_bits (read_mode, int_store_mode, const0_rtx); |
1924 | else if (GET_MODE_BITSIZE (int_store_mode) > HOST_BITS_PER_WIDE_INT | |
8dd5516b JJ |
1925 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) |
1926 | read_reg = NULL_RTX; | |
1927 | else | |
1928 | { | |
1929 | unsigned HOST_WIDE_INT c | |
1930 | = INTVAL (store_info->rhs) | |
fecfbfa4 | 1931 | & ((HOST_WIDE_INT_1 << BITS_PER_UNIT) - 1); |
8dd5516b JJ |
1932 | int shift = BITS_PER_UNIT; |
1933 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1934 | { | |
1935 | c |= (c << shift); | |
1936 | shift <<= 1; | |
1937 | } | |
304b9962 RS |
1938 | read_reg = gen_int_mode (c, int_store_mode); |
1939 | read_reg = extract_low_bits (read_mode, int_store_mode, read_reg); | |
8dd5516b JJ |
1940 | } |
1941 | } | |
1942 | else if (store_info->const_rhs | |
1943 | && (require_cst | |
1944 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1945 | read_reg = extract_low_bits (read_mode, store_mode, | |
1946 | copy_rtx (store_info->const_rhs)); | |
990769a3 PL |
1947 | else if (VECTOR_MODE_P (read_mode) && VECTOR_MODE_P (store_mode) |
1948 | && known_le (GET_MODE_BITSIZE (read_mode), GET_MODE_BITSIZE (store_mode)) | |
1949 | && targetm.modes_tieable_p (read_mode, store_mode)) | |
1950 | read_reg = gen_lowpart (read_mode, copy_rtx (store_info->rhs)); | |
8dd5516b JJ |
1951 | else |
1952 | read_reg = extract_low_bits (read_mode, store_mode, | |
1953 | copy_rtx (store_info->rhs)); | |
1954 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1955 | read_reg = NULL_RTX; | |
1956 | return read_reg; | |
1957 | } | |
02b47899 | 1958 | |
6fb5fa3c DB |
1959 | /* Take a sequence of: |
1960 | A <- r1 | |
1961 | ... | |
1962 | ... <- A | |
1963 | ||
b8698a0f | 1964 | and change it into |
6fb5fa3c DB |
1965 | r2 <- r1 |
1966 | A <- r1 | |
1967 | ... | |
1968 | ... <- r2 | |
1969 | ||
8660aaae EC |
1970 | or |
1971 | ||
1972 | r3 <- extract (r1) | |
1973 | r3 <- r3 >> shift | |
1974 | r2 <- extract (r3) | |
1975 | ... <- r2 | |
1976 | ||
1977 | or | |
1978 | ||
1979 | r2 <- extract (r1) | |
1980 | ... <- r2 | |
1981 | ||
1982 | Depending on the alignment and the mode of the store and | |
1983 | subsequent load. | |
1984 | ||
1985 | ||
1986 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
6fb5fa3c DB |
1987 | and READ_INSN are for the read. Return true if the replacement |
1988 | went ok. */ | |
1989 | ||
1990 | static bool | |
fd6daec9 | 1991 | replace_read (store_info *store_info, insn_info_t store_insn, |
9c747368 | 1992 | read_info_t read_info, insn_info_t read_insn, rtx *loc) |
6fb5fa3c | 1993 | { |
ef4bddc2 RS |
1994 | machine_mode store_mode = GET_MODE (store_info->mem); |
1995 | machine_mode read_mode = GET_MODE (read_info->mem); | |
dc01c3d1 DM |
1996 | rtx_insn *insns, *this_insn; |
1997 | rtx read_reg; | |
8dd5516b | 1998 | basic_block bb; |
8660aaae | 1999 | |
6fb5fa3c DB |
2000 | if (!dbg_cnt (dse)) |
2001 | return false; | |
2002 | ||
18b526e8 RS |
2003 | /* Create a sequence of instructions to set up the read register. |
2004 | This sequence goes immediately before the store and its result | |
2005 | is read by the load. | |
2006 | ||
2007 | We need to keep this in perspective. We are replacing a read | |
8660aaae EC |
2008 | with a sequence of insns, but the read will almost certainly be |
2009 | in cache, so it is not going to be an expensive one. Thus, we | |
2010 | are not willing to do a multi insn shift or worse a subroutine | |
2011 | call to get rid of the read. */ | |
456610d3 | 2012 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2013 | fprintf (dump_file, "trying to replace %smode load in insn %d" |
2014 | " from %smode store in insn %d\n", | |
2015 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
2016 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
2017 | start_sequence (); | |
8dd5516b JJ |
2018 | bb = BLOCK_FOR_INSN (read_insn->insn); |
2019 | read_reg = get_stored_val (store_info, | |
71d6a386 | 2020 | read_mode, read_info->offset, read_info->width, |
8dd5516b | 2021 | bb, false); |
18b526e8 | 2022 | if (read_reg == NULL_RTX) |
8660aaae | 2023 | { |
18b526e8 | 2024 | end_sequence (); |
456610d3 | 2025 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2026 | fprintf (dump_file, " -- could not extract bits of stored value\n"); |
2027 | return false; | |
8660aaae | 2028 | } |
18b526e8 RS |
2029 | /* Force the value into a new register so that it won't be clobbered |
2030 | between the store and the load. */ | |
76f44fbf JJ |
2031 | if (WORD_REGISTER_OPERATIONS |
2032 | && GET_CODE (read_reg) == SUBREG | |
2033 | && REG_P (SUBREG_REG (read_reg)) | |
2034 | && GET_MODE (SUBREG_REG (read_reg)) == word_mode) | |
2035 | { | |
2036 | /* For WORD_REGISTER_OPERATIONS with subreg of word_mode register | |
2037 | force SUBREG_REG into a new register rather than the SUBREG. */ | |
2038 | rtx r = copy_to_mode_reg (word_mode, SUBREG_REG (read_reg)); | |
2039 | read_reg = shallow_copy_rtx (read_reg); | |
2040 | SUBREG_REG (read_reg) = r; | |
2041 | } | |
2042 | else | |
2043 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
18b526e8 RS |
2044 | insns = get_insns (); |
2045 | end_sequence (); | |
8660aaae | 2046 | |
02b47899 KZ |
2047 | if (insns != NULL_RTX) |
2048 | { | |
2049 | /* Now we have to scan the set of new instructions to see if the | |
2050 | sequence contains and sets of hardregs that happened to be | |
2051 | live at this point. For instance, this can happen if one of | |
2052 | the insns sets the CC and the CC happened to be live at that | |
2053 | point. This does occasionally happen, see PR 37922. */ | |
3f9b14ff | 2054 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
02b47899 | 2055 | |
2c0fa3ec JJ |
2056 | for (this_insn = insns; |
2057 | this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
2058 | { | |
2059 | if (insn_invalid_p (this_insn, false)) | |
2060 | { | |
2061 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2062 | { | |
2063 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2064 | print_simple_rtl (dump_file, read_reg); | |
2065 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2066 | } | |
2067 | BITMAP_FREE (regs_set); | |
2068 | return false; | |
2069 | } | |
2070 | note_stores (this_insn, look_for_hardregs, regs_set); | |
2071 | } | |
b8698a0f | 2072 | |
9c747368 JJ |
2073 | if (store_insn->fixed_regs_live) |
2074 | bitmap_and_into (regs_set, store_insn->fixed_regs_live); | |
02b47899 KZ |
2075 | if (!bitmap_empty_p (regs_set)) |
2076 | { | |
456610d3 | 2077 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02b47899 | 2078 | { |
2c0fa3ec JJ |
2079 | fprintf (dump_file, "abandoning replacement because sequence " |
2080 | "clobbers live hardregs:"); | |
02b47899 KZ |
2081 | df_print_regset (dump_file, regs_set); |
2082 | } | |
b8698a0f | 2083 | |
02b47899 KZ |
2084 | BITMAP_FREE (regs_set); |
2085 | return false; | |
2086 | } | |
2087 | BITMAP_FREE (regs_set); | |
2088 | } | |
2089 | ||
2c0fa3ec JJ |
2090 | subrtx_iterator::array_type array; |
2091 | FOR_EACH_SUBRTX (iter, array, *loc, NONCONST) | |
2092 | { | |
2093 | const_rtx x = *iter; | |
2094 | if (GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) | |
2095 | { | |
2096 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2097 | fprintf (dump_file, " -- replacing the MEM failed due to address " | |
2098 | "side-effects\n"); | |
2099 | return false; | |
2100 | } | |
2101 | } | |
2102 | ||
8660aaae | 2103 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
6fb5fa3c | 2104 | { |
fd6daec9 | 2105 | deferred_change *change = deferred_change_pool.allocate (); |
b8698a0f | 2106 | |
8660aaae EC |
2107 | /* Insert this right before the store insn where it will be safe |
2108 | from later insns that might change it before the read. */ | |
2109 | emit_insn_before (insns, store_insn->insn); | |
b8698a0f | 2110 | |
8660aaae EC |
2111 | /* And now for the kludge part: cselib croaks if you just |
2112 | return at this point. There are two reasons for this: | |
b8698a0f | 2113 | |
8660aaae EC |
2114 | 1) Cselib has an idea of how many pseudos there are and |
2115 | that does not include the new ones we just added. | |
b8698a0f | 2116 | |
8660aaae EC |
2117 | 2) Cselib does not know about the move insn we added |
2118 | above the store_info, and there is no way to tell it | |
2119 | about it, because it has "moved on". | |
b8698a0f | 2120 | |
8660aaae EC |
2121 | Problem (1) is fixable with a certain amount of engineering. |
2122 | Problem (2) is requires starting the bb from scratch. This | |
2123 | could be expensive. | |
b8698a0f | 2124 | |
8660aaae EC |
2125 | So we are just going to have to lie. The move/extraction |
2126 | insns are not really an issue, cselib did not see them. But | |
2127 | the use of the new pseudo read_insn is a real problem because | |
2128 | cselib has not scanned this insn. The way that we solve this | |
2129 | problem is that we are just going to put the mem back for now | |
2130 | and when we are finished with the block, we undo this. We | |
2131 | keep a table of mems to get rid of. At the end of the basic | |
2132 | block we can put them back. */ | |
b8698a0f | 2133 | |
8660aaae | 2134 | *loc = read_info->mem; |
da6603c6 ML |
2135 | change->next = deferred_change_list; |
2136 | deferred_change_list = change; | |
2137 | change->loc = loc; | |
2138 | change->reg = read_reg; | |
b8698a0f | 2139 | |
8660aaae EC |
2140 | /* Get rid of the read_info, from the point of view of the |
2141 | rest of dse, play like this read never happened. */ | |
2142 | read_insn->read_rec = read_info->next; | |
fb0b2914 | 2143 | read_info_type_pool.remove (read_info); |
456610d3 | 2144 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2145 | { |
2146 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2147 | print_simple_rtl (dump_file, read_reg); | |
2148 | fprintf (dump_file, "\n"); | |
2149 | } | |
8660aaae | 2150 | return true; |
6fb5fa3c | 2151 | } |
b8698a0f | 2152 | else |
6fb5fa3c | 2153 | { |
456610d3 | 2154 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2155 | { |
2156 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2157 | print_simple_rtl (dump_file, read_reg); | |
2158 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2159 | } | |
6fb5fa3c DB |
2160 | return false; |
2161 | } | |
2162 | } | |
2163 | ||
d7111da8 RS |
2164 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2165 | be active. */ | |
6fb5fa3c | 2166 | |
d7111da8 | 2167 | static void |
47593832 | 2168 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info, bool used_in_call = false) |
6fb5fa3c | 2169 | { |
6216f94e | 2170 | rtx mem = *loc, mem_addr; |
6fb5fa3c | 2171 | insn_info_t insn_info; |
02ce5d90 RS |
2172 | poly_int64 offset = 0; |
2173 | poly_int64 width = 0; | |
b8698a0f | 2174 | cselib_val *base = NULL; |
6fb5fa3c DB |
2175 | int group_id; |
2176 | read_info_t read_info; | |
2177 | ||
6fb5fa3c DB |
2178 | insn_info = bb_info->last_insn; |
2179 | ||
2180 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
a49a975f | 2181 | || MEM_VOLATILE_P (mem)) |
6fb5fa3c | 2182 | { |
a49a975f JJ |
2183 | if (crtl->stack_protect_guard |
2184 | && (MEM_EXPR (mem) == crtl->stack_protect_guard | |
2185 | || (crtl->stack_protect_guard_decl | |
2186 | && MEM_EXPR (mem) == crtl->stack_protect_guard_decl)) | |
2187 | && MEM_VOLATILE_P (mem)) | |
2188 | { | |
2189 | /* This is either the stack protector canary on the stack, | |
2190 | which ought to be written by a MEM_VOLATILE_P store and | |
2191 | thus shouldn't be deleted and is read at the very end of | |
2192 | function, but shouldn't conflict with any other store. | |
2193 | Or it is __stack_chk_guard variable or TLS or whatever else | |
2194 | MEM holding the canary value, which really shouldn't be | |
2195 | ever modified in -fstack-protector* protected functions, | |
2196 | otherwise the prologue store wouldn't match the epilogue | |
2197 | check. */ | |
2198 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2199 | fprintf (dump_file, " stack protector canary read ignored.\n"); | |
2200 | insn_info->cannot_delete = true; | |
2201 | return; | |
2202 | } | |
2203 | ||
456610d3 | 2204 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2205 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
2206 | add_wild_read (bb_info); | |
2207 | insn_info->cannot_delete = true; | |
d7111da8 | 2208 | return; |
6fb5fa3c DB |
2209 | } |
2210 | ||
2211 | /* If it is reading readonly mem, then there can be no conflict with | |
2212 | another write. */ | |
2213 | if (MEM_READONLY_P (mem)) | |
d7111da8 | 2214 | return; |
6fb5fa3c | 2215 | |
ac6929b5 | 2216 | if (!canon_address (mem, &group_id, &offset, &base)) |
6fb5fa3c | 2217 | { |
456610d3 | 2218 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2219 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
2220 | add_wild_read (bb_info); | |
d7111da8 | 2221 | return; |
6fb5fa3c DB |
2222 | } |
2223 | ||
2224 | if (GET_MODE (mem) == BLKmode) | |
2225 | width = -1; | |
2226 | else | |
2227 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2228 | ||
02ce5d90 | 2229 | if (!endpoint_representable_p (offset, known_eq (width, -1) ? 1 : width)) |
938f9248 | 2230 | { |
213ffde2 JJ |
2231 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2232 | fprintf (dump_file, " adding wild read, due to overflow.\n"); | |
2233 | add_wild_read (bb_info); | |
938f9248 ML |
2234 | return; |
2235 | } | |
2236 | ||
fb0b2914 | 2237 | read_info = read_info_type_pool.allocate (); |
6fb5fa3c DB |
2238 | read_info->group_id = group_id; |
2239 | read_info->mem = mem; | |
71d6a386 RS |
2240 | read_info->offset = offset; |
2241 | read_info->width = width; | |
6fb5fa3c DB |
2242 | read_info->next = insn_info->read_rec; |
2243 | insn_info->read_rec = read_info; | |
ac6929b5 RB |
2244 | if (group_id < 0) |
2245 | mem_addr = base->val_rtx; | |
6216f94e JJ |
2246 | else |
2247 | { | |
ac6929b5 RB |
2248 | group_info *group = rtx_group_vec[group_id]; |
2249 | mem_addr = group->canon_base_addr; | |
6216f94e | 2250 | } |
02ce5d90 | 2251 | if (maybe_ne (offset, 0)) |
ac6929b5 | 2252 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
a523add3 RB |
2253 | /* Avoid passing VALUE RTXen as mem_addr to canon_true_dependence |
2254 | which will over and over re-create proper RTL and re-apply the | |
2255 | offset above. See PR80960 where we almost allocate 1.6GB of PLUS | |
2256 | RTXen that way. */ | |
2257 | mem_addr = get_addr (mem_addr); | |
6fb5fa3c | 2258 | |
ac6929b5 | 2259 | if (group_id >= 0) |
6fb5fa3c DB |
2260 | { |
2261 | /* This is the restricted case where the base is a constant or | |
2262 | the frame pointer and offset is a constant. */ | |
2263 | insn_info_t i_ptr = active_local_stores; | |
2264 | insn_info_t last = NULL; | |
b8698a0f | 2265 | |
456610d3 | 2266 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2267 | { |
02ce5d90 | 2268 | if (!known_size_p (width)) |
6fb5fa3c DB |
2269 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", |
2270 | group_id); | |
2271 | else | |
71d6a386 RS |
2272 | { |
2273 | fprintf (dump_file, " processing const load gid=%d", group_id); | |
2274 | print_range (dump_file, offset, width); | |
2275 | fprintf (dump_file, "\n"); | |
2276 | } | |
6fb5fa3c DB |
2277 | } |
2278 | ||
2279 | while (i_ptr) | |
2280 | { | |
2281 | bool remove = false; | |
fd6daec9 | 2282 | store_info *store_info = i_ptr->store_rec; |
b8698a0f | 2283 | |
6fb5fa3c DB |
2284 | /* Skip the clobbers. */ |
2285 | while (!store_info->is_set) | |
2286 | store_info = store_info->next; | |
b8698a0f | 2287 | |
6fb5fa3c DB |
2288 | /* There are three cases here. */ |
2289 | if (store_info->group_id < 0) | |
2290 | /* We have a cselib store followed by a read from a | |
2291 | const base. */ | |
b8698a0f L |
2292 | remove |
2293 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2294 | GET_MODE (store_info->mem), |
2295 | store_info->mem_addr, | |
53d9622b | 2296 | mem, mem_addr); |
b8698a0f | 2297 | |
6fb5fa3c DB |
2298 | else if (group_id == store_info->group_id) |
2299 | { | |
2300 | /* This is a block mode load. We may get lucky and | |
2301 | canon_true_dependence may save the day. */ | |
02ce5d90 | 2302 | if (!known_size_p (width)) |
b8698a0f L |
2303 | remove |
2304 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2305 | GET_MODE (store_info->mem), |
2306 | store_info->mem_addr, | |
53d9622b | 2307 | mem, mem_addr); |
b8698a0f | 2308 | |
6fb5fa3c DB |
2309 | /* If this read is just reading back something that we just |
2310 | stored, rewrite the read. */ | |
b8698a0f | 2311 | else |
6fb5fa3c | 2312 | { |
47593832 JG |
2313 | if (!used_in_call |
2314 | && store_info->rhs | |
71d6a386 RS |
2315 | && known_subrange_p (offset, width, store_info->offset, |
2316 | store_info->width) | |
8dd5516b | 2317 | && all_positions_needed_p (store_info, |
71d6a386 | 2318 | offset - store_info->offset, |
8dd5516b JJ |
2319 | width) |
2320 | && replace_read (store_info, i_ptr, read_info, | |
9c747368 | 2321 | insn_info, loc)) |
d7111da8 | 2322 | return; |
8dd5516b | 2323 | |
6fb5fa3c | 2324 | /* The bases are the same, just see if the offsets |
71d6a386 RS |
2325 | could overlap. */ |
2326 | if (ranges_maybe_overlap_p (offset, width, | |
2327 | store_info->offset, | |
2328 | store_info->width)) | |
6fb5fa3c DB |
2329 | remove = true; |
2330 | } | |
2331 | } | |
b8698a0f L |
2332 | |
2333 | /* else | |
6fb5fa3c DB |
2334 | The else case that is missing here is that the |
2335 | bases are constant but different. There is nothing | |
2336 | to do here because there is no overlap. */ | |
b8698a0f | 2337 | |
6fb5fa3c DB |
2338 | if (remove) |
2339 | { | |
456610d3 | 2340 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2341 | dump_insn_info ("removing from active", i_ptr); |
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 | } | |
b8698a0f | 2354 | else |
6fb5fa3c DB |
2355 | { |
2356 | insn_info_t i_ptr = active_local_stores; | |
2357 | insn_info_t last = NULL; | |
456610d3 | 2358 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2359 | { |
2360 | fprintf (dump_file, " processing cselib load mem:"); | |
2361 | print_inline_rtx (dump_file, mem, 0); | |
2362 | fprintf (dump_file, "\n"); | |
2363 | } | |
2364 | ||
2365 | while (i_ptr) | |
2366 | { | |
2367 | bool remove = false; | |
fd6daec9 | 2368 | store_info *store_info = i_ptr->store_rec; |
b8698a0f | 2369 | |
456610d3 | 2370 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2371 | fprintf (dump_file, " processing cselib load against insn %d\n", |
2372 | INSN_UID (i_ptr->insn)); | |
2373 | ||
2374 | /* Skip the clobbers. */ | |
2375 | while (!store_info->is_set) | |
2376 | store_info = store_info->next; | |
2377 | ||
2378 | /* If this read is just reading back something that we just | |
2379 | stored, rewrite the read. */ | |
47593832 JG |
2380 | if (!used_in_call |
2381 | && store_info->rhs | |
6fb5fa3c DB |
2382 | && store_info->group_id == -1 |
2383 | && store_info->cse_base == base | |
71d6a386 RS |
2384 | && known_subrange_p (offset, width, store_info->offset, |
2385 | store_info->width) | |
8dd5516b | 2386 | && all_positions_needed_p (store_info, |
71d6a386 | 2387 | offset - store_info->offset, width) |
9c747368 | 2388 | && replace_read (store_info, i_ptr, read_info, insn_info, loc)) |
d7111da8 | 2389 | return; |
6fb5fa3c | 2390 | |
ac6929b5 RB |
2391 | remove = canon_true_dependence (store_info->mem, |
2392 | GET_MODE (store_info->mem), | |
2393 | store_info->mem_addr, | |
2394 | mem, mem_addr); | |
b8698a0f | 2395 | |
6fb5fa3c DB |
2396 | if (remove) |
2397 | { | |
456610d3 | 2398 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2399 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2400 | |
dabd47e7 | 2401 | active_local_stores_len--; |
6fb5fa3c DB |
2402 | if (last) |
2403 | last->next_local_store = i_ptr->next_local_store; | |
2404 | else | |
2405 | active_local_stores = i_ptr->next_local_store; | |
2406 | } | |
2407 | else | |
2408 | last = i_ptr; | |
2409 | i_ptr = i_ptr->next_local_store; | |
2410 | } | |
2411 | } | |
6fb5fa3c DB |
2412 | } |
2413 | ||
d7111da8 | 2414 | /* A note_uses callback in which DATA points the INSN_INFO for |
6fb5fa3c DB |
2415 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2416 | true for any part of *LOC. */ | |
2417 | ||
2418 | static void | |
2419 | check_mem_read_use (rtx *loc, void *data) | |
2420 | { | |
d7111da8 RS |
2421 | subrtx_ptr_iterator::array_type array; |
2422 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
2423 | { | |
2424 | rtx *loc = *iter; | |
2425 | if (MEM_P (*loc)) | |
2426 | check_mem_read_rtx (loc, (bb_info_t) data); | |
2427 | } | |
6fb5fa3c DB |
2428 | } |
2429 | ||
8dd5516b JJ |
2430 | |
2431 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2432 | So far it only handles arguments passed in registers. */ | |
2433 | ||
2434 | static bool | |
2435 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2436 | { | |
d5cc9181 JR |
2437 | CUMULATIVE_ARGS args_so_far_v; |
2438 | cumulative_args_t args_so_far; | |
8dd5516b JJ |
2439 | tree arg; |
2440 | int idx; | |
2441 | ||
d5cc9181 JR |
2442 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2443 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
8dd5516b JJ |
2444 | |
2445 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2446 | for (idx = 0; | |
2447 | arg != void_list_node && idx < nargs; | |
2448 | arg = TREE_CHAIN (arg), idx++) | |
2449 | { | |
b4206259 | 2450 | scalar_int_mode mode; |
3c07301f | 2451 | rtx reg, link, tmp; |
b4206259 RS |
2452 | |
2453 | if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg)), &mode)) | |
2454 | return false; | |
2455 | ||
6783fdb7 RS |
2456 | function_arg_info arg (mode, /*named=*/true); |
2457 | reg = targetm.calls.function_arg (args_so_far, arg); | |
b4206259 | 2458 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode) |
8dd5516b JJ |
2459 | return false; |
2460 | ||
2461 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2462 | link; | |
2463 | link = XEXP (link, 1)) | |
2464 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2465 | { | |
b4206259 | 2466 | scalar_int_mode arg_mode; |
8dd5516b JJ |
2467 | args[idx] = XEXP (XEXP (link, 0), 0); |
2468 | if (REG_P (args[idx]) | |
2469 | && REGNO (args[idx]) == REGNO (reg) | |
2470 | && (GET_MODE (args[idx]) == mode | |
b4206259 RS |
2471 | || (is_int_mode (GET_MODE (args[idx]), &arg_mode) |
2472 | && (GET_MODE_SIZE (arg_mode) <= UNITS_PER_WORD) | |
2473 | && (GET_MODE_SIZE (arg_mode) > GET_MODE_SIZE (mode))))) | |
8dd5516b JJ |
2474 | break; |
2475 | } | |
2476 | if (!link) | |
2477 | return false; | |
2478 | ||
2479 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2480 | if (GET_MODE (args[idx]) != mode) | |
2481 | { | |
2482 | if (!tmp || !CONST_INT_P (tmp)) | |
2483 | return false; | |
6d26322f | 2484 | tmp = gen_int_mode (INTVAL (tmp), mode); |
8dd5516b JJ |
2485 | } |
2486 | if (tmp) | |
2487 | args[idx] = tmp; | |
2488 | ||
6930c98c | 2489 | targetm.calls.function_arg_advance (args_so_far, arg); |
8dd5516b JJ |
2490 | } |
2491 | if (arg != void_list_node || idx != nargs) | |
2492 | return false; | |
2493 | return true; | |
2494 | } | |
2495 | ||
9e582b1d JR |
2496 | /* Return a bitmap of the fixed registers contained in IN. */ |
2497 | ||
2498 | static bitmap | |
2499 | copy_fixed_regs (const_bitmap in) | |
2500 | { | |
2501 | bitmap ret; | |
2502 | ||
2503 | ret = ALLOC_REG_SET (NULL); | |
0b0310e9 | 2504 | bitmap_and (ret, in, bitmap_view<HARD_REG_SET> (fixed_reg_set)); |
9e582b1d JR |
2505 | return ret; |
2506 | } | |
8dd5516b | 2507 | |
6fb5fa3c DB |
2508 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2509 | if some part of it is not a candidate store and assigns to a | |
2510 | non-register target. */ | |
2511 | ||
2512 | static void | |
fc2d7303 | 2513 | scan_insn (bb_info_t bb_info, rtx_insn *insn, int max_active_local_stores) |
6fb5fa3c DB |
2514 | { |
2515 | rtx body; | |
fb0b2914 | 2516 | insn_info_type *insn_info = insn_info_type_pool.allocate (); |
6fb5fa3c | 2517 | int mems_found = 0; |
da6603c6 | 2518 | memset (insn_info, 0, sizeof (struct insn_info_type)); |
6fb5fa3c | 2519 | |
456610d3 | 2520 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2521 | fprintf (dump_file, "\n**scanning insn=%d\n", |
2522 | INSN_UID (insn)); | |
2523 | ||
2524 | insn_info->prev_insn = bb_info->last_insn; | |
2525 | insn_info->insn = insn; | |
2526 | bb_info->last_insn = insn_info; | |
b8698a0f | 2527 | |
b5b8b0ac AO |
2528 | if (DEBUG_INSN_P (insn)) |
2529 | { | |
2530 | insn_info->cannot_delete = true; | |
2531 | return; | |
2532 | } | |
6fb5fa3c | 2533 | |
6fb5fa3c DB |
2534 | /* Look at all of the uses in the insn. */ |
2535 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2536 | ||
2537 | if (CALL_P (insn)) | |
2538 | { | |
8dd5516b | 2539 | bool const_call; |
ee516de9 | 2540 | rtx call, sym; |
8dd5516b JJ |
2541 | tree memset_call = NULL_TREE; |
2542 | ||
6fb5fa3c | 2543 | insn_info->cannot_delete = true; |
50f0f366 | 2544 | |
6fb5fa3c | 2545 | /* Const functions cannot do anything bad i.e. read memory, |
50f0f366 | 2546 | however, they can read their parameters which may have |
8dd5516b JJ |
2547 | been pushed onto the stack. |
2548 | memset and bzero don't read memory either. */ | |
2549 | const_call = RTL_CONST_CALL_P (insn); | |
ee516de9 EB |
2550 | if (!const_call |
2551 | && (call = get_call_rtx_from (insn)) | |
2552 | && (sym = XEXP (XEXP (call, 0), 0)) | |
2553 | && GET_CODE (sym) == SYMBOL_REF | |
2554 | && SYMBOL_REF_DECL (sym) | |
2555 | && TREE_CODE (SYMBOL_REF_DECL (sym)) == FUNCTION_DECL | |
3d78e008 | 2556 | && fndecl_built_in_p (SYMBOL_REF_DECL (sym), BUILT_IN_MEMSET)) |
ee516de9 EB |
2557 | memset_call = SYMBOL_REF_DECL (sym); |
2558 | ||
8dd5516b | 2559 | if (const_call || memset_call) |
6fb5fa3c DB |
2560 | { |
2561 | insn_info_t i_ptr = active_local_stores; | |
2562 | insn_info_t last = NULL; | |
2563 | ||
456610d3 | 2564 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2565 | fprintf (dump_file, "%s call %d\n", |
2566 | const_call ? "const" : "memset", INSN_UID (insn)); | |
6fb5fa3c | 2567 | |
64520bdc | 2568 | /* See the head comment of the frame_read field. */ |
57534689 JJ |
2569 | if (reload_completed |
2570 | /* Tail calls are storing their arguments using | |
2571 | arg pointer. If it is a frame pointer on the target, | |
2572 | even before reload we need to kill frame pointer based | |
2573 | stores. */ | |
2574 | || (SIBLING_CALL_P (insn) | |
2575 | && HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
64520bdc EB |
2576 | insn_info->frame_read = true; |
2577 | ||
2578 | /* Loop over the active stores and remove those which are | |
2579 | killed by the const function call. */ | |
6fb5fa3c DB |
2580 | while (i_ptr) |
2581 | { | |
64520bdc EB |
2582 | bool remove_store = false; |
2583 | ||
2584 | /* The stack pointer based stores are always killed. */ | |
50f0f366 | 2585 | if (i_ptr->stack_pointer_based) |
64520bdc EB |
2586 | remove_store = true; |
2587 | ||
2588 | /* If the frame is read, the frame related stores are killed. */ | |
2589 | else if (insn_info->frame_read) | |
2590 | { | |
fd6daec9 | 2591 | store_info *store_info = i_ptr->store_rec; |
64520bdc EB |
2592 | |
2593 | /* Skip the clobbers. */ | |
2594 | while (!store_info->is_set) | |
2595 | store_info = store_info->next; | |
2596 | ||
2597 | if (store_info->group_id >= 0 | |
9771b263 | 2598 | && rtx_group_vec[store_info->group_id]->frame_related) |
64520bdc EB |
2599 | remove_store = true; |
2600 | } | |
2601 | ||
2602 | if (remove_store) | |
6fb5fa3c | 2603 | { |
456610d3 | 2604 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2605 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2606 | |
dabd47e7 | 2607 | active_local_stores_len--; |
6fb5fa3c DB |
2608 | if (last) |
2609 | last->next_local_store = i_ptr->next_local_store; | |
2610 | else | |
2611 | active_local_stores = i_ptr->next_local_store; | |
2612 | } | |
2613 | else | |
2614 | last = i_ptr; | |
64520bdc | 2615 | |
6fb5fa3c DB |
2616 | i_ptr = i_ptr->next_local_store; |
2617 | } | |
8dd5516b JJ |
2618 | |
2619 | if (memset_call) | |
2620 | { | |
2621 | rtx args[3]; | |
2622 | if (get_call_args (insn, memset_call, args, 3) | |
2623 | && CONST_INT_P (args[1]) | |
2624 | && CONST_INT_P (args[2]) | |
2625 | && INTVAL (args[2]) > 0) | |
2626 | { | |
2627 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
f5541398 | 2628 | set_mem_size (mem, INTVAL (args[2])); |
f7df4a84 | 2629 | body = gen_rtx_SET (mem, args[1]); |
8dd5516b | 2630 | mems_found += record_store (body, bb_info); |
456610d3 | 2631 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2632 | fprintf (dump_file, "handling memset as BLKmode store\n"); |
2633 | if (mems_found == 1) | |
2634 | { | |
fc2d7303 | 2635 | if (active_local_stores_len++ >= max_active_local_stores) |
dabd47e7 JJ |
2636 | { |
2637 | active_local_stores_len = 1; | |
2638 | active_local_stores = NULL; | |
2639 | } | |
9e582b1d JR |
2640 | insn_info->fixed_regs_live |
2641 | = copy_fixed_regs (bb_info->regs_live); | |
8dd5516b JJ |
2642 | insn_info->next_local_store = active_local_stores; |
2643 | active_local_stores = insn_info; | |
2644 | } | |
2645 | } | |
a2ad54a0 JJ |
2646 | else |
2647 | clear_rhs_from_active_local_stores (); | |
8dd5516b | 2648 | } |
6fb5fa3c | 2649 | } |
7ee98586 EB |
2650 | else if (SIBLING_CALL_P (insn) |
2651 | && (reload_completed || HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
57534689 JJ |
2652 | /* Arguments for a sibling call that are pushed to memory are passed |
2653 | using the incoming argument pointer of the current function. After | |
7ee98586 EB |
2654 | reload that might be (and likely is) frame pointer based. And, if |
2655 | it is a frame pointer on the target, even before reload we need to | |
2656 | kill frame pointer based stores. */ | |
57534689 | 2657 | add_wild_read (bb_info); |
50f0f366 | 2658 | else |
d26c7090 ER |
2659 | /* Every other call, including pure functions, may read any memory |
2660 | that is not relative to the frame. */ | |
2661 | add_non_frame_wild_read (bb_info); | |
50f0f366 | 2662 | |
47593832 JG |
2663 | for (rtx link = CALL_INSN_FUNCTION_USAGE (insn); |
2664 | link != NULL_RTX; | |
2665 | link = XEXP (link, 1)) | |
2666 | if (GET_CODE (XEXP (link, 0)) == USE && MEM_P (XEXP (XEXP (link, 0),0))) | |
2667 | check_mem_read_rtx (&XEXP (XEXP (link, 0),0), bb_info, true); | |
2668 | ||
6fb5fa3c DB |
2669 | return; |
2670 | } | |
2671 | ||
2672 | /* Assuming that there are sets in these insns, we cannot delete | |
2673 | them. */ | |
2674 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
0a64eeca | 2675 | || volatile_refs_p (PATTERN (insn)) |
2da02156 | 2676 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
6fb5fa3c DB |
2677 | || (RTX_FRAME_RELATED_P (insn)) |
2678 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2679 | insn_info->cannot_delete = true; | |
b8698a0f | 2680 | |
6fb5fa3c DB |
2681 | body = PATTERN (insn); |
2682 | if (GET_CODE (body) == PARALLEL) | |
2683 | { | |
2684 | int i; | |
2685 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2686 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2687 | } | |
2688 | else | |
2689 | mems_found += record_store (body, bb_info); | |
2690 | ||
456610d3 | 2691 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2692 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
6fb5fa3c DB |
2693 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2694 | ||
8dd5516b JJ |
2695 | /* If we found some sets of mems, add it into the active_local_stores so |
2696 | that it can be locally deleted if found dead or used for | |
2697 | replace_read and redundant constant store elimination. Otherwise mark | |
2698 | it as cannot delete. This simplifies the processing later. */ | |
2699 | if (mems_found == 1) | |
6fb5fa3c | 2700 | { |
fc2d7303 | 2701 | if (active_local_stores_len++ >= max_active_local_stores) |
dabd47e7 JJ |
2702 | { |
2703 | active_local_stores_len = 1; | |
2704 | active_local_stores = NULL; | |
2705 | } | |
9e582b1d | 2706 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
6fb5fa3c DB |
2707 | insn_info->next_local_store = active_local_stores; |
2708 | active_local_stores = insn_info; | |
2709 | } | |
2710 | else | |
2711 | insn_info->cannot_delete = true; | |
2712 | } | |
2713 | ||
2714 | ||
2715 | /* Remove BASE from the set of active_local_stores. This is a | |
2716 | callback from cselib that is used to get rid of the stores in | |
2717 | active_local_stores. */ | |
2718 | ||
2719 | static void | |
2720 | remove_useless_values (cselib_val *base) | |
2721 | { | |
2722 | insn_info_t insn_info = active_local_stores; | |
2723 | insn_info_t last = NULL; | |
2724 | ||
2725 | while (insn_info) | |
2726 | { | |
fd6daec9 | 2727 | store_info *store_info = insn_info->store_rec; |
60564289 | 2728 | bool del = false; |
6fb5fa3c DB |
2729 | |
2730 | /* If ANY of the store_infos match the cselib group that is | |
67914693 | 2731 | being deleted, then the insn cannot be deleted. */ |
6fb5fa3c DB |
2732 | while (store_info) |
2733 | { | |
b8698a0f | 2734 | if ((store_info->group_id == -1) |
6fb5fa3c DB |
2735 | && (store_info->cse_base == base)) |
2736 | { | |
60564289 | 2737 | del = true; |
6fb5fa3c DB |
2738 | break; |
2739 | } | |
2740 | store_info = store_info->next; | |
2741 | } | |
2742 | ||
60564289 | 2743 | if (del) |
6fb5fa3c | 2744 | { |
dabd47e7 | 2745 | active_local_stores_len--; |
6fb5fa3c DB |
2746 | if (last) |
2747 | last->next_local_store = insn_info->next_local_store; | |
2748 | else | |
2749 | active_local_stores = insn_info->next_local_store; | |
2750 | free_store_info (insn_info); | |
2751 | } | |
2752 | else | |
2753 | last = insn_info; | |
b8698a0f | 2754 | |
6fb5fa3c DB |
2755 | insn_info = insn_info->next_local_store; |
2756 | } | |
2757 | } | |
2758 | ||
2759 | ||
2760 | /* Do all of step 1. */ | |
2761 | ||
2762 | static void | |
2763 | dse_step1 (void) | |
2764 | { | |
2765 | basic_block bb; | |
3f9b14ff | 2766 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
b8698a0f | 2767 | |
457eeaae | 2768 | cselib_init (0); |
6fb5fa3c DB |
2769 | all_blocks = BITMAP_ALLOC (NULL); |
2770 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2771 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2772 | ||
fc2d7303 RB |
2773 | /* For -O1 reduce the maximum number of active local stores for RTL DSE |
2774 | since this can consume huge amounts of memory (PR89115). */ | |
028d4092 | 2775 | int max_active_local_stores = param_max_dse_active_local_stores; |
fc2d7303 RB |
2776 | if (optimize < 2) |
2777 | max_active_local_stores /= 10; | |
2778 | ||
04a90bec | 2779 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
2780 | { |
2781 | insn_info_t ptr; | |
fb0b2914 | 2782 | bb_info_t bb_info = dse_bb_info_type_pool.allocate (); |
6fb5fa3c | 2783 | |
da6603c6 | 2784 | memset (bb_info, 0, sizeof (dse_bb_info_type)); |
6fb5fa3c | 2785 | bitmap_set_bit (all_blocks, bb->index); |
02b47899 KZ |
2786 | bb_info->regs_live = regs_live; |
2787 | ||
2788 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2789 | df_simulate_initialize_forwards (bb, regs_live); | |
6fb5fa3c DB |
2790 | |
2791 | bb_table[bb->index] = bb_info; | |
2792 | cselib_discard_hook = remove_useless_values; | |
2793 | ||
2794 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2795 | { | |
dd60a84c | 2796 | rtx_insn *insn; |
6fb5fa3c | 2797 | |
6fb5fa3c | 2798 | active_local_stores = NULL; |
dabd47e7 | 2799 | active_local_stores_len = 0; |
6fb5fa3c | 2800 | cselib_clear_table (); |
b8698a0f | 2801 | |
6fb5fa3c DB |
2802 | /* Scan the insns. */ |
2803 | FOR_BB_INSNS (bb, insn) | |
2804 | { | |
2805 | if (INSN_P (insn)) | |
fc2d7303 | 2806 | scan_insn (bb_info, insn, max_active_local_stores); |
6fb5fa3c | 2807 | cselib_process_insn (insn); |
02b47899 KZ |
2808 | if (INSN_P (insn)) |
2809 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
6fb5fa3c | 2810 | } |
b8698a0f | 2811 | |
6fb5fa3c DB |
2812 | /* This is something of a hack, because the global algorithm |
2813 | is supposed to take care of the case where stores go dead | |
2814 | at the end of the function. However, the global | |
2815 | algorithm must take a more conservative view of block | |
2816 | mode reads than the local alg does. So to get the case | |
2817 | where you have a store to the frame followed by a non | |
0d52bcc1 | 2818 | overlapping block more read, we look at the active local |
6fb5fa3c DB |
2819 | stores at the end of the function and delete all of the |
2820 | frame and spill based ones. */ | |
2821 | if (stores_off_frame_dead_at_return | |
2822 | && (EDGE_COUNT (bb->succs) == 0 | |
2823 | || (single_succ_p (bb) | |
fefa31b5 | 2824 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
e3b5732b | 2825 | && ! crtl->calls_eh_return))) |
6fb5fa3c DB |
2826 | { |
2827 | insn_info_t i_ptr = active_local_stores; | |
2828 | while (i_ptr) | |
2829 | { | |
fd6daec9 | 2830 | store_info *store_info = i_ptr->store_rec; |
6fb5fa3c DB |
2831 | |
2832 | /* Skip the clobbers. */ | |
2833 | while (!store_info->is_set) | |
2834 | store_info = store_info->next; | |
ac6929b5 RB |
2835 | if (store_info->group_id >= 0) |
2836 | { | |
2837 | group_info *group = rtx_group_vec[store_info->group_id]; | |
2838 | if (group->frame_related && !i_ptr->cannot_delete) | |
2839 | delete_dead_store_insn (i_ptr); | |
2840 | } | |
6fb5fa3c DB |
2841 | |
2842 | i_ptr = i_ptr->next_local_store; | |
2843 | } | |
2844 | } | |
2845 | ||
2846 | /* Get rid of the loads that were discovered in | |
2847 | replace_read. Cselib is finished with this block. */ | |
2848 | while (deferred_change_list) | |
2849 | { | |
fd6daec9 | 2850 | deferred_change *next = deferred_change_list->next; |
6fb5fa3c DB |
2851 | |
2852 | /* There is no reason to validate this change. That was | |
2853 | done earlier. */ | |
2854 | *deferred_change_list->loc = deferred_change_list->reg; | |
fb0b2914 | 2855 | deferred_change_pool.remove (deferred_change_list); |
6fb5fa3c DB |
2856 | deferred_change_list = next; |
2857 | } | |
2858 | ||
2859 | /* Get rid of all of the cselib based store_infos in this | |
2860 | block and mark the containing insns as not being | |
2861 | deletable. */ | |
2862 | ptr = bb_info->last_insn; | |
2863 | while (ptr) | |
2864 | { | |
2865 | if (ptr->contains_cselib_groups) | |
8dd5516b | 2866 | { |
fd6daec9 | 2867 | store_info *s_info = ptr->store_rec; |
8dd5516b JJ |
2868 | while (s_info && !s_info->is_set) |
2869 | s_info = s_info->next; | |
2870 | if (s_info | |
2871 | && s_info->redundant_reason | |
2872 | && s_info->redundant_reason->insn | |
2873 | && !ptr->cannot_delete) | |
2874 | { | |
456610d3 | 2875 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2876 | fprintf (dump_file, "Locally deleting insn %d " |
2877 | "because insn %d stores the " | |
2878 | "same value and couldn't be " | |
2879 | "eliminated\n", | |
2880 | INSN_UID (ptr->insn), | |
2881 | INSN_UID (s_info->redundant_reason->insn)); | |
2882 | delete_dead_store_insn (ptr); | |
2883 | } | |
8dd5516b JJ |
2884 | free_store_info (ptr); |
2885 | } | |
2886 | else | |
2887 | { | |
fd6daec9 | 2888 | store_info *s_info; |
8dd5516b JJ |
2889 | |
2890 | /* Free at least positions_needed bitmaps. */ | |
2891 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2892 | if (s_info->is_large) | |
2893 | { | |
dc491a25 | 2894 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
8dd5516b JJ |
2895 | s_info->is_large = false; |
2896 | } | |
2897 | } | |
6fb5fa3c DB |
2898 | ptr = ptr->prev_insn; |
2899 | } | |
2900 | ||
da6603c6 | 2901 | cse_store_info_pool.release (); |
6fb5fa3c | 2902 | } |
02b47899 | 2903 | bb_info->regs_live = NULL; |
6fb5fa3c DB |
2904 | } |
2905 | ||
02b47899 | 2906 | BITMAP_FREE (regs_live); |
6fb5fa3c | 2907 | cselib_finish (); |
c203e8a7 | 2908 | rtx_group_table->empty (); |
6fb5fa3c DB |
2909 | } |
2910 | ||
2911 | \f | |
2912 | /*---------------------------------------------------------------------------- | |
2913 | Second step. | |
2914 | ||
2915 | Assign each byte position in the stores that we are going to | |
2916 | analyze globally to a position in the bitmaps. Returns true if | |
6ed3da00 | 2917 | there are any bit positions assigned. |
6fb5fa3c DB |
2918 | ----------------------------------------------------------------------------*/ |
2919 | ||
2920 | static void | |
2921 | dse_step2_init (void) | |
2922 | { | |
2923 | unsigned int i; | |
fd6daec9 | 2924 | group_info *group; |
6fb5fa3c | 2925 | |
9771b263 | 2926 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2927 | { |
2928 | /* For all non stack related bases, we only consider a store to | |
2929 | be deletable if there are two or more stores for that | |
2930 | position. This is because it takes one store to make the | |
2931 | other store redundant. However, for the stores that are | |
2932 | stack related, we consider them if there is only one store | |
2933 | for the position. We do this because the stack related | |
2934 | stores can be deleted if their is no read between them and | |
2935 | the end of the function. | |
b8698a0f | 2936 | |
6fb5fa3c DB |
2937 | To make this work in the current framework, we take the stack |
2938 | related bases add all of the bits from store1 into store2. | |
2939 | This has the effect of making the eligible even if there is | |
2940 | only one store. */ | |
2941 | ||
2942 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2943 | { | |
2944 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2945 | bitmap_ior_into (group->store2_p, group->store1_p); | |
456610d3 | 2946 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2947 | fprintf (dump_file, "group %d is frame related ", i); |
6fb5fa3c DB |
2948 | } |
2949 | ||
2950 | group->offset_map_size_n++; | |
3f9b14ff SB |
2951 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2952 | group->offset_map_size_n); | |
6fb5fa3c | 2953 | group->offset_map_size_p++; |
3f9b14ff SB |
2954 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2955 | group->offset_map_size_p); | |
6fb5fa3c | 2956 | group->process_globally = false; |
456610d3 | 2957 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2958 | { |
b8698a0f | 2959 | fprintf (dump_file, "group %d(%d+%d): ", i, |
6fb5fa3c DB |
2960 | (int)bitmap_count_bits (group->store2_n), |
2961 | (int)bitmap_count_bits (group->store2_p)); | |
2962 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2963 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2964 | } | |
2965 | } | |
2966 | } | |
2967 | ||
2968 | ||
ac6929b5 | 2969 | /* Init the offset tables. */ |
6fb5fa3c DB |
2970 | |
2971 | static bool | |
ac6929b5 | 2972 | dse_step2 (void) |
6fb5fa3c DB |
2973 | { |
2974 | unsigned int i; | |
fd6daec9 | 2975 | group_info *group; |
6fb5fa3c DB |
2976 | /* Position 0 is unused because 0 is used in the maps to mean |
2977 | unused. */ | |
2978 | current_position = 1; | |
9771b263 | 2979 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2980 | { |
2981 | bitmap_iterator bi; | |
2982 | unsigned int j; | |
2983 | ||
c3284718 RS |
2984 | memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
2985 | memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
6fb5fa3c DB |
2986 | bitmap_clear (group->group_kill); |
2987 | ||
2988 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2989 | { | |
2990 | bitmap_set_bit (group->group_kill, current_position); | |
d26c7090 ER |
2991 | if (bitmap_bit_p (group->escaped_n, j)) |
2992 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2993 | group->offset_map_n[j] = current_position++; |
2994 | group->process_globally = true; | |
2995 | } | |
2996 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2997 | { | |
b8698a0f | 2998 | bitmap_set_bit (group->group_kill, current_position); |
d26c7090 ER |
2999 | if (bitmap_bit_p (group->escaped_p, j)) |
3000 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
3001 | group->offset_map_p[j] = current_position++; |
3002 | group->process_globally = true; | |
3003 | } | |
3004 | } | |
3005 | return current_position != 1; | |
3006 | } | |
3007 | ||
3008 | ||
6fb5fa3c DB |
3009 | \f |
3010 | /*---------------------------------------------------------------------------- | |
3011 | Third step. | |
b8698a0f | 3012 | |
6fb5fa3c DB |
3013 | Build the bit vectors for the transfer functions. |
3014 | ----------------------------------------------------------------------------*/ | |
3015 | ||
3016 | ||
6fb5fa3c DB |
3017 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
3018 | there, return 0. */ | |
3019 | ||
3020 | static int | |
fd6daec9 | 3021 | get_bitmap_index (group_info *group_info, HOST_WIDE_INT offset) |
6fb5fa3c DB |
3022 | { |
3023 | if (offset < 0) | |
3024 | { | |
3025 | HOST_WIDE_INT offset_p = -offset; | |
3026 | if (offset_p >= group_info->offset_map_size_n) | |
3027 | return 0; | |
3028 | return group_info->offset_map_n[offset_p]; | |
3029 | } | |
3030 | else | |
3031 | { | |
3032 | if (offset >= group_info->offset_map_size_p) | |
3033 | return 0; | |
3034 | return group_info->offset_map_p[offset]; | |
3035 | } | |
3036 | } | |
3037 | ||
3038 | ||
3039 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
3040 | may be NULL. */ | |
3041 | ||
b8698a0f | 3042 | static void |
ac6929b5 | 3043 | scan_stores (store_info *store_info, bitmap gen, bitmap kill) |
6fb5fa3c DB |
3044 | { |
3045 | while (store_info) | |
3046 | { | |
02ce5d90 | 3047 | HOST_WIDE_INT i, offset, width; |
fd6daec9 | 3048 | group_info *group_info |
9771b263 | 3049 | = rtx_group_vec[store_info->group_id]; |
02ce5d90 RS |
3050 | /* We can (conservatively) ignore stores whose bounds aren't known; |
3051 | they simply don't generate new global dse opportunities. */ | |
3052 | if (group_info->process_globally | |
3053 | && store_info->offset.is_constant (&offset) | |
3054 | && store_info->width.is_constant (&width)) | |
71d6a386 | 3055 | { |
02ce5d90 RS |
3056 | HOST_WIDE_INT end = offset + width; |
3057 | for (i = offset; i < end; i++) | |
71d6a386 RS |
3058 | { |
3059 | int index = get_bitmap_index (group_info, i); | |
3060 | if (index != 0) | |
3061 | { | |
3062 | bitmap_set_bit (gen, index); | |
3063 | if (kill) | |
3064 | bitmap_clear_bit (kill, index); | |
3065 | } | |
3066 | } | |
3067 | } | |
6fb5fa3c DB |
3068 | store_info = store_info->next; |
3069 | } | |
3070 | } | |
3071 | ||
3072 | ||
6fb5fa3c DB |
3073 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL |
3074 | may be NULL. */ | |
3075 | ||
3076 | static void | |
ac6929b5 | 3077 | scan_reads (insn_info_t insn_info, bitmap gen, bitmap kill) |
6fb5fa3c DB |
3078 | { |
3079 | read_info_t read_info = insn_info->read_rec; | |
3080 | int i; | |
fd6daec9 | 3081 | group_info *group; |
6fb5fa3c | 3082 | |
64520bdc EB |
3083 | /* If this insn reads the frame, kill all the frame related stores. */ |
3084 | if (insn_info->frame_read) | |
3085 | { | |
9771b263 | 3086 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
64520bdc EB |
3087 | if (group->process_globally && group->frame_related) |
3088 | { | |
3089 | if (kill) | |
3090 | bitmap_ior_into (kill, group->group_kill); | |
b8698a0f | 3091 | bitmap_and_compl_into (gen, group->group_kill); |
64520bdc EB |
3092 | } |
3093 | } | |
d26c7090 ER |
3094 | if (insn_info->non_frame_wild_read) |
3095 | { | |
3096 | /* Kill all non-frame related stores. Kill all stores of variables that | |
3097 | escape. */ | |
3098 | if (kill) | |
3099 | bitmap_ior_into (kill, kill_on_calls); | |
3100 | bitmap_and_compl_into (gen, kill_on_calls); | |
9771b263 | 3101 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
d26c7090 ER |
3102 | if (group->process_globally && !group->frame_related) |
3103 | { | |
3104 | if (kill) | |
3105 | bitmap_ior_into (kill, group->group_kill); | |
3106 | bitmap_and_compl_into (gen, group->group_kill); | |
3107 | } | |
3108 | } | |
6fb5fa3c DB |
3109 | while (read_info) |
3110 | { | |
9771b263 | 3111 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3112 | { |
3113 | if (group->process_globally) | |
3114 | { | |
3115 | if (i == read_info->group_id) | |
3116 | { | |
02ce5d90 RS |
3117 | HOST_WIDE_INT offset, width; |
3118 | /* Reads with non-constant size kill all DSE opportunities | |
3119 | in the group. */ | |
3120 | if (!read_info->offset.is_constant (&offset) | |
3121 | || !read_info->width.is_constant (&width) | |
3122 | || !known_size_p (width)) | |
6fb5fa3c | 3123 | { |
71d6a386 | 3124 | /* Handle block mode reads. */ |
6fb5fa3c DB |
3125 | if (kill) |
3126 | bitmap_ior_into (kill, group->group_kill); | |
3127 | bitmap_and_compl_into (gen, group->group_kill); | |
3128 | } | |
3129 | else | |
3130 | { | |
3131 | /* The groups are the same, just process the | |
3132 | offsets. */ | |
3133 | HOST_WIDE_INT j; | |
02ce5d90 RS |
3134 | HOST_WIDE_INT end = offset + width; |
3135 | for (j = offset; j < end; j++) | |
6fb5fa3c DB |
3136 | { |
3137 | int index = get_bitmap_index (group, j); | |
3138 | if (index != 0) | |
3139 | { | |
3140 | if (kill) | |
3141 | bitmap_set_bit (kill, index); | |
3142 | bitmap_clear_bit (gen, index); | |
3143 | } | |
3144 | } | |
3145 | } | |
3146 | } | |
3147 | else | |
3148 | { | |
3149 | /* The groups are different, if the alias sets | |
3150 | conflict, clear the entire group. We only need | |
3151 | to apply this test if the read_info is a cselib | |
3152 | read. Anything with a constant base cannot alias | |
3153 | something else with a different constant | |
3154 | base. */ | |
3155 | if ((read_info->group_id < 0) | |
b8698a0f | 3156 | && canon_true_dependence (group->base_mem, |
d32f725a | 3157 | GET_MODE (group->base_mem), |
6216f94e | 3158 | group->canon_base_addr, |
53d9622b | 3159 | read_info->mem, NULL_RTX)) |
6fb5fa3c DB |
3160 | { |
3161 | if (kill) | |
3162 | bitmap_ior_into (kill, group->group_kill); | |
3163 | bitmap_and_compl_into (gen, group->group_kill); | |
3164 | } | |
3165 | } | |
3166 | } | |
3167 | } | |
b8698a0f | 3168 | |
6fb5fa3c DB |
3169 | read_info = read_info->next; |
3170 | } | |
3171 | } | |
3172 | ||
6fb5fa3c DB |
3173 | |
3174 | /* Return the insn in BB_INFO before the first wild read or if there | |
3175 | are no wild reads in the block, return the last insn. */ | |
3176 | ||
3177 | static insn_info_t | |
3178 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3179 | { | |
3180 | insn_info_t insn_info = bb_info->last_insn; | |
3181 | insn_info_t last_wild_read = NULL; | |
3182 | ||
3183 | while (insn_info) | |
3184 | { | |
3185 | if (insn_info->wild_read) | |
3186 | { | |
3187 | last_wild_read = insn_info->prev_insn; | |
3188 | /* Block starts with wild read. */ | |
3189 | if (!last_wild_read) | |
3190 | return NULL; | |
3191 | } | |
3192 | ||
3193 | insn_info = insn_info->prev_insn; | |
3194 | } | |
3195 | ||
3196 | if (last_wild_read) | |
3197 | return last_wild_read; | |
3198 | else | |
3199 | return bb_info->last_insn; | |
3200 | } | |
3201 | ||
3202 | ||
3203 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3204 | the block in order to build the gen and kill sets for the block. | |
3205 | We start at ptr which may be the last insn in the block or may be | |
3206 | the first insn with a wild read. In the latter case we are able to | |
3207 | skip the rest of the block because it just does not matter: | |
3208 | anything that happens is hidden by the wild read. */ | |
3209 | ||
3210 | static void | |
ac6929b5 | 3211 | dse_step3_scan (basic_block bb) |
6fb5fa3c DB |
3212 | { |
3213 | bb_info_t bb_info = bb_table[bb->index]; | |
3214 | insn_info_t insn_info; | |
3215 | ||
ac6929b5 | 3216 | insn_info = find_insn_before_first_wild_read (bb_info); |
b8698a0f | 3217 | |
6fb5fa3c DB |
3218 | /* In the spill case or in the no_spill case if there is no wild |
3219 | read in the block, we will need a kill set. */ | |
3220 | if (insn_info == bb_info->last_insn) | |
3221 | { | |
3222 | if (bb_info->kill) | |
3223 | bitmap_clear (bb_info->kill); | |
3224 | else | |
3f9b14ff | 3225 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c | 3226 | } |
b8698a0f | 3227 | else |
6fb5fa3c DB |
3228 | if (bb_info->kill) |
3229 | BITMAP_FREE (bb_info->kill); | |
3230 | ||
3231 | while (insn_info) | |
3232 | { | |
3233 | /* There may have been code deleted by the dce pass run before | |
3234 | this phase. */ | |
3235 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3236 | { | |
ac6929b5 RB |
3237 | scan_stores (insn_info->store_rec, bb_info->gen, bb_info->kill); |
3238 | scan_reads (insn_info, bb_info->gen, bb_info->kill); | |
b8698a0f | 3239 | } |
6fb5fa3c DB |
3240 | |
3241 | insn_info = insn_info->prev_insn; | |
3242 | } | |
3243 | } | |
3244 | ||
3245 | ||
3246 | /* Set the gen set of the exit block, and also any block with no | |
3247 | successors that does not have a wild read. */ | |
3248 | ||
3249 | static void | |
3250 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3251 | { | |
3252 | /* The gen set is all 0's for the exit block except for the | |
3253 | frame_pointer_group. */ | |
b8698a0f | 3254 | |
6fb5fa3c DB |
3255 | if (stores_off_frame_dead_at_return) |
3256 | { | |
3257 | unsigned int i; | |
fd6daec9 | 3258 | group_info *group; |
b8698a0f | 3259 | |
9771b263 | 3260 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3261 | { |
3262 | if (group->process_globally && group->frame_related) | |
3263 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3264 | } | |
3265 | } | |
3266 | } | |
3267 | ||
3268 | ||
3269 | /* Find all of the blocks that are not backwards reachable from the | |
3270 | exit block or any block with no successors (BB). These are the | |
3271 | infinite loops or infinite self loops. These blocks will still | |
3272 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3273 | ||
3274 | static void | |
3275 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3276 | { | |
3277 | edge e; | |
3278 | edge_iterator ei; | |
3279 | ||
d7c028c0 | 3280 | if (bitmap_bit_p (unreachable_blocks, bb->index)) |
6fb5fa3c | 3281 | { |
d7c028c0 | 3282 | bitmap_clear_bit (unreachable_blocks, bb->index); |
6fb5fa3c | 3283 | FOR_EACH_EDGE (e, ei, bb->preds) |
b8698a0f | 3284 | { |
6fb5fa3c | 3285 | mark_reachable_blocks (unreachable_blocks, e->src); |
b8698a0f | 3286 | } |
6fb5fa3c DB |
3287 | } |
3288 | } | |
3289 | ||
3290 | /* Build the transfer functions for the function. */ | |
3291 | ||
3292 | static void | |
ac6929b5 | 3293 | dse_step3 () |
6fb5fa3c DB |
3294 | { |
3295 | basic_block bb; | |
6fb5fa3c DB |
3296 | sbitmap_iterator sbi; |
3297 | bitmap all_ones = NULL; | |
3298 | unsigned int i; | |
b8698a0f | 3299 | |
7ba9e72d | 3300 | auto_sbitmap unreachable_blocks (last_basic_block_for_fn (cfun)); |
f61e445a | 3301 | bitmap_ones (unreachable_blocks); |
6fb5fa3c | 3302 | |
04a90bec | 3303 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3304 | { |
3305 | bb_info_t bb_info = bb_table[bb->index]; | |
3306 | if (bb_info->gen) | |
3307 | bitmap_clear (bb_info->gen); | |
3308 | else | |
3f9b14ff | 3309 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3310 | |
3311 | if (bb->index == ENTRY_BLOCK) | |
3312 | ; | |
3313 | else if (bb->index == EXIT_BLOCK) | |
3314 | dse_step3_exit_block_scan (bb_info); | |
3315 | else | |
ac6929b5 | 3316 | dse_step3_scan (bb); |
6fb5fa3c DB |
3317 | if (EDGE_COUNT (bb->succs) == 0) |
3318 | mark_reachable_blocks (unreachable_blocks, bb); | |
3319 | ||
3320 | /* If this is the second time dataflow is run, delete the old | |
3321 | sets. */ | |
3322 | if (bb_info->in) | |
3323 | BITMAP_FREE (bb_info->in); | |
3324 | if (bb_info->out) | |
3325 | BITMAP_FREE (bb_info->out); | |
3326 | } | |
3327 | ||
3328 | /* For any block in an infinite loop, we must initialize the out set | |
3329 | to all ones. This could be expensive, but almost never occurs in | |
3330 | practice. However, it is common in regression tests. */ | |
d4ac4ce2 | 3331 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
6fb5fa3c DB |
3332 | { |
3333 | if (bitmap_bit_p (all_blocks, i)) | |
3334 | { | |
3335 | bb_info_t bb_info = bb_table[i]; | |
3336 | if (!all_ones) | |
3337 | { | |
3338 | unsigned int j; | |
fd6daec9 | 3339 | group_info *group; |
6fb5fa3c | 3340 | |
3f9b14ff | 3341 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
9771b263 | 3342 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
6fb5fa3c DB |
3343 | bitmap_ior_into (all_ones, group->group_kill); |
3344 | } | |
3345 | if (!bb_info->out) | |
3346 | { | |
3f9b14ff | 3347 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3348 | bitmap_copy (bb_info->out, all_ones); |
3349 | } | |
3350 | } | |
3351 | } | |
3352 | ||
3353 | if (all_ones) | |
3354 | BITMAP_FREE (all_ones); | |
6fb5fa3c DB |
3355 | } |
3356 | ||
3357 | ||
3358 | \f | |
3359 | /*---------------------------------------------------------------------------- | |
3360 | Fourth step. | |
3361 | ||
3362 | Solve the bitvector equations. | |
3363 | ----------------------------------------------------------------------------*/ | |
3364 | ||
3365 | ||
3366 | /* Confluence function for blocks with no successors. Create an out | |
3367 | set from the gen set of the exit block. This block logically has | |
3368 | the exit block as a successor. */ | |
3369 | ||
3370 | ||
3371 | ||
3372 | static void | |
3373 | dse_confluence_0 (basic_block bb) | |
3374 | { | |
3375 | bb_info_t bb_info = bb_table[bb->index]; | |
3376 | ||
3377 | if (bb->index == EXIT_BLOCK) | |
3378 | return; | |
3379 | ||
3380 | if (!bb_info->out) | |
3381 | { | |
3f9b14ff | 3382 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3383 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3384 | } | |
3385 | } | |
3386 | ||
3387 | /* Propagate the information from the in set of the dest of E to the | |
3388 | out set of the src of E. If the various in or out sets are not | |
3389 | there, that means they are all ones. */ | |
3390 | ||
1a0f3fa1 | 3391 | static bool |
6fb5fa3c DB |
3392 | dse_confluence_n (edge e) |
3393 | { | |
3394 | bb_info_t src_info = bb_table[e->src->index]; | |
3395 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3396 | ||
3397 | if (dest_info->in) | |
3398 | { | |
3399 | if (src_info->out) | |
3400 | bitmap_and_into (src_info->out, dest_info->in); | |
3401 | else | |
3402 | { | |
3f9b14ff | 3403 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3404 | bitmap_copy (src_info->out, dest_info->in); |
3405 | } | |
3406 | } | |
1a0f3fa1 | 3407 | return true; |
6fb5fa3c DB |
3408 | } |
3409 | ||
3410 | ||
3411 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
b8698a0f | 3412 | block. There are three cases: |
6fb5fa3c DB |
3413 | |
3414 | 1) The block has no kill set. In this case the kill set is all | |
3415 | ones. It does not matter what the out set of the block is, none of | |
3416 | the info can reach the top. The only thing that reaches the top is | |
3417 | the gen set and we just copy the set. | |
3418 | ||
3419 | 2) There is a kill set but no out set and bb has successors. In | |
3420 | this case we just return. Eventually an out set will be created and | |
3421 | it is better to wait than to create a set of ones. | |
3422 | ||
3423 | 3) There is both a kill and out set. We apply the obvious transfer | |
3424 | function. | |
3425 | */ | |
3426 | ||
3427 | static bool | |
3428 | dse_transfer_function (int bb_index) | |
3429 | { | |
3430 | bb_info_t bb_info = bb_table[bb_index]; | |
3431 | ||
3432 | if (bb_info->kill) | |
3433 | { | |
3434 | if (bb_info->out) | |
3435 | { | |
3436 | /* Case 3 above. */ | |
3437 | if (bb_info->in) | |
b8698a0f | 3438 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3439 | bb_info->out, bb_info->kill); |
3440 | else | |
3441 | { | |
3f9b14ff | 3442 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
b8698a0f | 3443 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3444 | bb_info->out, bb_info->kill); |
3445 | return true; | |
3446 | } | |
3447 | } | |
3448 | else | |
3449 | /* Case 2 above. */ | |
3450 | return false; | |
3451 | } | |
3452 | else | |
3453 | { | |
3454 | /* Case 1 above. If there is already an in set, nothing | |
3455 | happens. */ | |
3456 | if (bb_info->in) | |
3457 | return false; | |
3458 | else | |
3459 | { | |
3f9b14ff | 3460 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3461 | bitmap_copy (bb_info->in, bb_info->gen); |
3462 | return true; | |
3463 | } | |
3464 | } | |
3465 | } | |
3466 | ||
3467 | /* Solve the dataflow equations. */ | |
3468 | ||
3469 | static void | |
3470 | dse_step4 (void) | |
3471 | { | |
b8698a0f L |
3472 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3473 | dse_confluence_n, dse_transfer_function, | |
3474 | all_blocks, df_get_postorder (DF_BACKWARD), | |
6fb5fa3c | 3475 | df_get_n_blocks (DF_BACKWARD)); |
456610d3 | 3476 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3477 | { |
3478 | basic_block bb; | |
3479 | ||
3480 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
04a90bec | 3481 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3482 | { |
3483 | bb_info_t bb_info = bb_table[bb->index]; | |
3484 | ||
3485 | df_print_bb_index (bb, dump_file); | |
3486 | if (bb_info->in) | |
3487 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3488 | else | |
3489 | fprintf (dump_file, " in: *MISSING*\n"); | |
3490 | if (bb_info->gen) | |
3491 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3492 | else | |
3493 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3494 | if (bb_info->kill) | |
3495 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3496 | else | |
3497 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3498 | if (bb_info->out) | |
3499 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3500 | else | |
3501 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3502 | } | |
3503 | } | |
3504 | } | |
3505 | ||
3506 | ||
3507 | \f | |
3508 | /*---------------------------------------------------------------------------- | |
3509 | Fifth step. | |
3510 | ||
0d52bcc1 | 3511 | Delete the stores that can only be deleted using the global information. |
6fb5fa3c DB |
3512 | ----------------------------------------------------------------------------*/ |
3513 | ||
3514 | ||
3515 | static void | |
ac6929b5 | 3516 | dse_step5 (void) |
6fb5fa3c DB |
3517 | { |
3518 | basic_block bb; | |
11cd3bed | 3519 | FOR_EACH_BB_FN (bb, cfun) |
6fb5fa3c DB |
3520 | { |
3521 | bb_info_t bb_info = bb_table[bb->index]; | |
3522 | insn_info_t insn_info = bb_info->last_insn; | |
3523 | bitmap v = bb_info->out; | |
3524 | ||
3525 | while (insn_info) | |
3526 | { | |
3527 | bool deleted = false; | |
3528 | if (dump_file && insn_info->insn) | |
3529 | { | |
3530 | fprintf (dump_file, "starting to process insn %d\n", | |
3531 | INSN_UID (insn_info->insn)); | |
3532 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3533 | } | |
3534 | ||
3535 | /* There may have been code deleted by the dce pass run before | |
3536 | this phase. */ | |
b8698a0f | 3537 | if (insn_info->insn |
6fb5fa3c DB |
3538 | && INSN_P (insn_info->insn) |
3539 | && (!insn_info->cannot_delete) | |
3540 | && (!bitmap_empty_p (v))) | |
3541 | { | |
fd6daec9 | 3542 | store_info *store_info = insn_info->store_rec; |
6fb5fa3c DB |
3543 | |
3544 | /* Try to delete the current insn. */ | |
3545 | deleted = true; | |
b8698a0f | 3546 | |
6fb5fa3c DB |
3547 | /* Skip the clobbers. */ |
3548 | while (!store_info->is_set) | |
3549 | store_info = store_info->next; | |
3550 | ||
02ce5d90 | 3551 | HOST_WIDE_INT i, offset, width; |
ac6929b5 RB |
3552 | group_info *group_info = rtx_group_vec[store_info->group_id]; |
3553 | ||
02ce5d90 RS |
3554 | if (!store_info->offset.is_constant (&offset) |
3555 | || !store_info->width.is_constant (&width)) | |
3556 | deleted = false; | |
3557 | else | |
6fb5fa3c | 3558 | { |
02ce5d90 RS |
3559 | HOST_WIDE_INT end = offset + width; |
3560 | for (i = offset; i < end; i++) | |
6fb5fa3c | 3561 | { |
02ce5d90 RS |
3562 | int index = get_bitmap_index (group_info, i); |
3563 | ||
456610d3 | 3564 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
3565 | fprintf (dump_file, "i = %d, index = %d\n", |
3566 | (int) i, index); | |
3567 | if (index == 0 || !bitmap_bit_p (v, index)) | |
3568 | { | |
3569 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3570 | fprintf (dump_file, "failing at i = %d\n", | |
3571 | (int) i); | |
3572 | deleted = false; | |
3573 | break; | |
3574 | } | |
6fb5fa3c DB |
3575 | } |
3576 | } | |
3577 | if (deleted) | |
3578 | { | |
9e582b1d JR |
3579 | if (dbg_cnt (dse) |
3580 | && check_for_inc_dec_1 (insn_info)) | |
6fb5fa3c | 3581 | { |
6fb5fa3c DB |
3582 | delete_insn (insn_info->insn); |
3583 | insn_info->insn = NULL; | |
3584 | globally_deleted++; | |
3585 | } | |
3586 | } | |
3587 | } | |
3588 | /* We do want to process the local info if the insn was | |
6ed3da00 | 3589 | deleted. For instance, if the insn did a wild read, we |
6fb5fa3c | 3590 | no longer need to trash the info. */ |
b8698a0f | 3591 | if (insn_info->insn |
6fb5fa3c DB |
3592 | && INSN_P (insn_info->insn) |
3593 | && (!deleted)) | |
3594 | { | |
ac6929b5 | 3595 | scan_stores (insn_info->store_rec, v, NULL); |
6fb5fa3c DB |
3596 | if (insn_info->wild_read) |
3597 | { | |
456610d3 | 3598 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3599 | fprintf (dump_file, "wild read\n"); |
3600 | bitmap_clear (v); | |
3601 | } | |
d26c7090 | 3602 | else if (insn_info->read_rec |
a8404b9a JJ |
3603 | || insn_info->non_frame_wild_read |
3604 | || insn_info->frame_read) | |
6fb5fa3c | 3605 | { |
a8404b9a JJ |
3606 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3607 | { | |
3608 | if (!insn_info->non_frame_wild_read | |
3609 | && !insn_info->frame_read) | |
3610 | fprintf (dump_file, "regular read\n"); | |
3611 | if (insn_info->non_frame_wild_read) | |
3612 | fprintf (dump_file, "non-frame wild read\n"); | |
3613 | if (insn_info->frame_read) | |
3614 | fprintf (dump_file, "frame read\n"); | |
3615 | } | |
ac6929b5 | 3616 | scan_reads (insn_info, v, NULL); |
6fb5fa3c DB |
3617 | } |
3618 | } | |
b8698a0f | 3619 | |
6fb5fa3c DB |
3620 | insn_info = insn_info->prev_insn; |
3621 | } | |
3622 | } | |
3623 | } | |
3624 | ||
3625 | ||
6fb5fa3c DB |
3626 | \f |
3627 | /*---------------------------------------------------------------------------- | |
3628 | Sixth step. | |
3629 | ||
8dd5516b JJ |
3630 | Delete stores made redundant by earlier stores (which store the same |
3631 | value) that couldn't be eliminated. | |
3632 | ----------------------------------------------------------------------------*/ | |
3633 | ||
3634 | static void | |
3635 | dse_step6 (void) | |
3636 | { | |
3637 | basic_block bb; | |
3638 | ||
04a90bec | 3639 | FOR_ALL_BB_FN (bb, cfun) |
8dd5516b JJ |
3640 | { |
3641 | bb_info_t bb_info = bb_table[bb->index]; | |
3642 | insn_info_t insn_info = bb_info->last_insn; | |
3643 | ||
3644 | while (insn_info) | |
3645 | { | |
3646 | /* There may have been code deleted by the dce pass run before | |
3647 | this phase. */ | |
3648 | if (insn_info->insn | |
3649 | && INSN_P (insn_info->insn) | |
3650 | && !insn_info->cannot_delete) | |
3651 | { | |
fd6daec9 | 3652 | store_info *s_info = insn_info->store_rec; |
8dd5516b JJ |
3653 | |
3654 | while (s_info && !s_info->is_set) | |
3655 | s_info = s_info->next; | |
3656 | if (s_info | |
3657 | && s_info->redundant_reason | |
3658 | && s_info->redundant_reason->insn | |
3659 | && INSN_P (s_info->redundant_reason->insn)) | |
3660 | { | |
eb92d49a | 3661 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
456610d3 | 3662 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
3663 | fprintf (dump_file, "Locally deleting insn %d " |
3664 | "because insn %d stores the " | |
3665 | "same value and couldn't be " | |
3666 | "eliminated\n", | |
3667 | INSN_UID (insn_info->insn), | |
3668 | INSN_UID (rinsn)); | |
3669 | delete_dead_store_insn (insn_info); | |
3670 | } | |
3671 | } | |
3672 | insn_info = insn_info->prev_insn; | |
3673 | } | |
3674 | } | |
3675 | } | |
3676 | \f | |
3677 | /*---------------------------------------------------------------------------- | |
3678 | Seventh step. | |
3679 | ||
b8698a0f | 3680 | Destroy everything left standing. |
6fb5fa3c DB |
3681 | ----------------------------------------------------------------------------*/ |
3682 | ||
b8698a0f | 3683 | static void |
3f9b14ff | 3684 | dse_step7 (void) |
6fb5fa3c | 3685 | { |
3f9b14ff SB |
3686 | bitmap_obstack_release (&dse_bitmap_obstack); |
3687 | obstack_free (&dse_obstack, NULL); | |
370f38e8 | 3688 | |
6fb5fa3c DB |
3689 | end_alias_analysis (); |
3690 | free (bb_table); | |
c203e8a7 TS |
3691 | delete rtx_group_table; |
3692 | rtx_group_table = NULL; | |
9771b263 | 3693 | rtx_group_vec.release (); |
6fb5fa3c DB |
3694 | BITMAP_FREE (all_blocks); |
3695 | BITMAP_FREE (scratch); | |
3696 | ||
da6603c6 | 3697 | rtx_store_info_pool.release (); |
fb0b2914 ML |
3698 | read_info_type_pool.release (); |
3699 | insn_info_type_pool.release (); | |
3700 | dse_bb_info_type_pool.release (); | |
3701 | group_info_pool.release (); | |
3702 | deferred_change_pool.release (); | |
6fb5fa3c DB |
3703 | } |
3704 | ||
3705 | ||
6fb5fa3c DB |
3706 | /* ------------------------------------------------------------------------- |
3707 | DSE | |
3708 | ------------------------------------------------------------------------- */ | |
3709 | ||
3710 | /* Callback for running pass_rtl_dse. */ | |
3711 | ||
3712 | static unsigned int | |
3713 | rest_of_handle_dse (void) | |
3714 | { | |
6fb5fa3c DB |
3715 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3716 | ||
02b47899 KZ |
3717 | /* Need the notes since we must track live hardregs in the forwards |
3718 | direction. */ | |
3719 | df_note_add_problem (); | |
3720 | df_analyze (); | |
3721 | ||
6fb5fa3c DB |
3722 | dse_step0 (); |
3723 | dse_step1 (); | |
dfda40f8 RB |
3724 | /* DSE can eliminate potentially-trapping MEMs. |
3725 | Remove any EH edges associated with them, since otherwise | |
3726 | DF_LR_RUN_DCE will complain later. */ | |
3727 | if ((locally_deleted || globally_deleted) | |
3728 | && cfun->can_throw_non_call_exceptions | |
3729 | && purge_all_dead_edges ()) | |
3730 | { | |
3731 | free_dominance_info (CDI_DOMINATORS); | |
3732 | delete_unreachable_blocks (); | |
3733 | } | |
6fb5fa3c | 3734 | dse_step2_init (); |
ac6929b5 | 3735 | if (dse_step2 ()) |
6fb5fa3c DB |
3736 | { |
3737 | df_set_flags (DF_LR_RUN_DCE); | |
3738 | df_analyze (); | |
456610d3 | 3739 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 3740 | fprintf (dump_file, "doing global processing\n"); |
ac6929b5 | 3741 | dse_step3 (); |
6fb5fa3c | 3742 | dse_step4 (); |
ac6929b5 | 3743 | dse_step5 (); |
6fb5fa3c DB |
3744 | } |
3745 | ||
8dd5516b | 3746 | dse_step6 (); |
3f9b14ff | 3747 | dse_step7 (); |
6fb5fa3c DB |
3748 | |
3749 | if (dump_file) | |
ac6929b5 RB |
3750 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d\n", |
3751 | locally_deleted, globally_deleted); | |
6c5ed3f1 JJ |
3752 | |
3753 | /* DSE can eliminate potentially-trapping MEMs. | |
3754 | Remove any EH edges associated with them. */ | |
3755 | if ((locally_deleted || globally_deleted) | |
3756 | && cfun->can_throw_non_call_exceptions | |
3757 | && purge_all_dead_edges ()) | |
7f5ddfcb JJ |
3758 | { |
3759 | free_dominance_info (CDI_DOMINATORS); | |
3760 | cleanup_cfg (0); | |
3761 | } | |
6c5ed3f1 | 3762 | |
6fb5fa3c DB |
3763 | return 0; |
3764 | } | |
3765 | ||
17795822 TS |
3766 | namespace { |
3767 | ||
3768 | const pass_data pass_data_rtl_dse1 = | |
27a4cd48 DM |
3769 | { |
3770 | RTL_PASS, /* type */ | |
3771 | "dse1", /* name */ | |
3772 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3773 | TV_DSE1, /* tv_id */ |
3774 | 0, /* properties_required */ | |
3775 | 0, /* properties_provided */ | |
3776 | 0, /* properties_destroyed */ | |
3777 | 0, /* todo_flags_start */ | |
3bea341f | 3778 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c DB |
3779 | }; |
3780 | ||
17795822 | 3781 | class pass_rtl_dse1 : public rtl_opt_pass |
27a4cd48 DM |
3782 | { |
3783 | public: | |
c3284718 RS |
3784 | pass_rtl_dse1 (gcc::context *ctxt) |
3785 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
27a4cd48 DM |
3786 | {} |
3787 | ||
3788 | /* opt_pass methods: */ | |
725793af | 3789 | bool gate (function *) final override |
1a3d085c TS |
3790 | { |
3791 | return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
3792 | } | |
3793 | ||
725793af DM |
3794 | unsigned int execute (function *) final override |
3795 | { | |
3796 | return rest_of_handle_dse (); | |
3797 | } | |
27a4cd48 DM |
3798 | |
3799 | }; // class pass_rtl_dse1 | |
3800 | ||
17795822 TS |
3801 | } // anon namespace |
3802 | ||
27a4cd48 DM |
3803 | rtl_opt_pass * |
3804 | make_pass_rtl_dse1 (gcc::context *ctxt) | |
3805 | { | |
3806 | return new pass_rtl_dse1 (ctxt); | |
3807 | } | |
3808 | ||
17795822 TS |
3809 | namespace { |
3810 | ||
3811 | const pass_data pass_data_rtl_dse2 = | |
27a4cd48 DM |
3812 | { |
3813 | RTL_PASS, /* type */ | |
3814 | "dse2", /* name */ | |
3815 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3816 | TV_DSE2, /* tv_id */ |
3817 | 0, /* properties_required */ | |
3818 | 0, /* properties_provided */ | |
3819 | 0, /* properties_destroyed */ | |
3820 | 0, /* todo_flags_start */ | |
3bea341f | 3821 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c | 3822 | }; |
27a4cd48 | 3823 | |
17795822 | 3824 | class pass_rtl_dse2 : public rtl_opt_pass |
27a4cd48 DM |
3825 | { |
3826 | public: | |
c3284718 RS |
3827 | pass_rtl_dse2 (gcc::context *ctxt) |
3828 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
27a4cd48 DM |
3829 | {} |
3830 | ||
3831 | /* opt_pass methods: */ | |
725793af | 3832 | bool gate (function *) final override |
1a3d085c TS |
3833 | { |
3834 | return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
3835 | } | |
3836 | ||
725793af DM |
3837 | unsigned int execute (function *) final override |
3838 | { | |
3839 | return rest_of_handle_dse (); | |
3840 | } | |
27a4cd48 DM |
3841 | |
3842 | }; // class pass_rtl_dse2 | |
3843 | ||
17795822 TS |
3844 | } // anon namespace |
3845 | ||
27a4cd48 DM |
3846 | rtl_opt_pass * |
3847 | make_pass_rtl_dse2 (gcc::context *ctxt) | |
3848 | { | |
3849 | return new pass_rtl_dse2 (ctxt); | |
3850 | } |