]>
Commit | Line | Data |
---|---|---|
6fb5fa3c | 1 | /* RTL dead store elimination. |
a5544970 | 2 | Copyright (C) 2005-2019 Free Software Foundation, Inc. |
6fb5fa3c DB |
3 | |
4 | Contributed by Richard Sandiford <rsandifor@codesourcery.com> | |
5 | and Kenneth Zadeck <zadeck@naturalbridge.com> | |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 11 | Software Foundation; either version 3, or (at your option) any later |
6fb5fa3c DB |
12 | version. |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ | |
6fb5fa3c DB |
22 | |
23 | #undef BASELINE | |
24 | ||
25 | #include "config.h" | |
26 | #include "system.h" | |
27 | #include "coretypes.h" | |
c7131fb2 | 28 | #include "backend.h" |
957060b5 AM |
29 | #include "target.h" |
30 | #include "rtl.h" | |
c7131fb2 AM |
31 | #include "tree.h" |
32 | #include "gimple.h" | |
957060b5 | 33 | #include "predict.h" |
c7131fb2 | 34 | #include "df.h" |
4d0cdd0c | 35 | #include "memmodel.h" |
957060b5 AM |
36 | #include "tm_p.h" |
37 | #include "gimple-ssa.h" | |
38 | #include "expmed.h" | |
39 | #include "optabs.h" | |
957060b5 AM |
40 | #include "emit-rtl.h" |
41 | #include "recog.h" | |
40e23961 | 42 | #include "alias.h" |
d8a2d370 | 43 | #include "stor-layout.h" |
60393bbc | 44 | #include "cfgrtl.h" |
6fb5fa3c | 45 | #include "cselib.h" |
6fb5fa3c | 46 | #include "tree-pass.h" |
36566b39 | 47 | #include "explow.h" |
6fb5fa3c | 48 | #include "expr.h" |
6fb5fa3c | 49 | #include "dbgcnt.h" |
dabd47e7 | 50 | #include "params.h" |
d7111da8 | 51 | #include "rtl-iter.h" |
6c5ed3f1 | 52 | #include "cfgcleanup.h" |
6783fdb7 | 53 | #include "calls.h" |
6fb5fa3c DB |
54 | |
55 | /* This file contains three techniques for performing Dead Store | |
b8698a0f | 56 | Elimination (dse). |
6fb5fa3c DB |
57 | |
58 | * The first technique performs dse locally on any base address. It | |
59 | is based on the cselib which is a local value numbering technique. | |
60 | This technique is local to a basic block but deals with a fairly | |
61 | general addresses. | |
b8698a0f | 62 | |
6fb5fa3c DB |
63 | * The second technique performs dse globally but is restricted to |
64 | base addresses that are either constant or are relative to the | |
65 | frame_pointer. | |
66 | ||
67 | * The third technique, (which is only done after register allocation) | |
026c3cfd | 68 | processes the spill slots. This differs from the second |
6fb5fa3c DB |
69 | technique because it takes advantage of the fact that spilling is |
70 | completely free from the effects of aliasing. | |
71 | ||
72 | Logically, dse is a backwards dataflow problem. A store can be | |
73 | deleted if it if cannot be reached in the backward direction by any | |
74 | use of the value being stored. However, the local technique uses a | |
75 | forwards scan of the basic block because cselib requires that the | |
76 | block be processed in that order. | |
77 | ||
78 | The pass is logically broken into 7 steps: | |
79 | ||
80 | 0) Initialization. | |
81 | ||
82 | 1) The local algorithm, as well as scanning the insns for the two | |
83 | global algorithms. | |
84 | ||
85 | 2) Analysis to see if the global algs are necessary. In the case | |
86 | of stores base on a constant address, there must be at least two | |
87 | stores to that address, to make it possible to delete some of the | |
88 | stores. In the case of stores off of the frame or spill related | |
89 | stores, only one store to an address is necessary because those | |
90 | stores die at the end of the function. | |
91 | ||
b8698a0f | 92 | 3) Set up the global dataflow equations based on processing the |
6fb5fa3c DB |
93 | info parsed in the first step. |
94 | ||
95 | 4) Solve the dataflow equations. | |
96 | ||
97 | 5) Delete the insns that the global analysis has indicated are | |
98 | unnecessary. | |
99 | ||
073a8998 | 100 | 6) Delete insns that store the same value as preceding store |
8dd5516b JJ |
101 | where the earlier store couldn't be eliminated. |
102 | ||
103 | 7) Cleanup. | |
6fb5fa3c DB |
104 | |
105 | This step uses cselib and canon_rtx to build the largest expression | |
106 | possible for each address. This pass is a forwards pass through | |
107 | each basic block. From the point of view of the global technique, | |
108 | the first pass could examine a block in either direction. The | |
0d52bcc1 | 109 | forwards ordering is to accommodate cselib. |
6fb5fa3c | 110 | |
5ef0b50d | 111 | We make a simplifying assumption: addresses fall into four broad |
6fb5fa3c DB |
112 | categories: |
113 | ||
114 | 1) base has rtx_varies_p == false, offset is constant. | |
115 | 2) base has rtx_varies_p == false, offset variable. | |
116 | 3) base has rtx_varies_p == true, offset constant. | |
117 | 4) base has rtx_varies_p == true, offset variable. | |
118 | ||
119 | The local passes are able to process all 4 kinds of addresses. The | |
5ef0b50d | 120 | global pass only handles 1). |
6fb5fa3c DB |
121 | |
122 | The global problem is formulated as follows: | |
123 | ||
124 | A store, S1, to address A, where A is not relative to the stack | |
125 | frame, can be eliminated if all paths from S1 to the end of the | |
5ef0b50d | 126 | function contain another store to A before a read to A. |
6fb5fa3c DB |
127 | |
128 | If the address A is relative to the stack frame, a store S2 to A | |
5ef0b50d | 129 | can be eliminated if there are no paths from S2 that reach the |
6fb5fa3c | 130 | end of the function that read A before another store to A. In |
5ef0b50d | 131 | this case S2 can be deleted if there are paths from S2 to the |
6fb5fa3c DB |
132 | end of the function that have no reads or writes to A. This |
133 | second case allows stores to the stack frame to be deleted that | |
134 | would otherwise die when the function returns. This cannot be | |
135 | done if stores_off_frame_dead_at_return is not true. See the doc | |
136 | for that variable for when this variable is false. | |
137 | ||
138 | The global problem is formulated as a backwards set union | |
139 | dataflow problem where the stores are the gens and reads are the | |
140 | kills. Set union problems are rare and require some special | |
141 | handling given our representation of bitmaps. A straightforward | |
5ef0b50d | 142 | implementation requires a lot of bitmaps filled with 1s. |
6fb5fa3c DB |
143 | These are expensive and cumbersome in our bitmap formulation so |
144 | care has been taken to avoid large vectors filled with 1s. See | |
145 | the comments in bb_info and in the dataflow confluence functions | |
b8698a0f | 146 | for details. |
6fb5fa3c DB |
147 | |
148 | There are two places for further enhancements to this algorithm: | |
b8698a0f | 149 | |
6fb5fa3c DB |
150 | 1) The original dse which was embedded in a pass called flow also |
151 | did local address forwarding. For example in | |
152 | ||
153 | A <- r100 | |
154 | ... <- A | |
155 | ||
156 | flow would replace the right hand side of the second insn with a | |
6ed3da00 | 157 | reference to r100. Most of the information is available to add this |
6fb5fa3c DB |
158 | to this pass. It has not done it because it is a lot of work in |
159 | the case that either r100 is assigned to between the first and | |
160 | second insn and/or the second insn is a load of part of the value | |
161 | stored by the first insn. | |
162 | ||
163 | insn 5 in gcc.c-torture/compile/990203-1.c simple case. | |
164 | insn 15 in gcc.c-torture/execute/20001017-2.c simple case. | |
165 | insn 25 in gcc.c-torture/execute/20001026-1.c simple case. | |
166 | insn 44 in gcc.c-torture/execute/20010910-1.c simple case. | |
167 | ||
168 | 2) The cleaning up of spill code is quite profitable. It currently | |
169 | depends on reading tea leaves and chicken entrails left by reload. | |
170 | This pass depends on reload creating a singleton alias set for each | |
171 | spill slot and telling the next dse pass which of these alias sets | |
172 | are the singletons. Rather than analyze the addresses of the | |
173 | spills, dse's spill processing just does analysis of the loads and | |
174 | stores that use those alias sets. There are three cases where this | |
175 | falls short: | |
176 | ||
177 | a) Reload sometimes creates the slot for one mode of access, and | |
178 | then inserts loads and/or stores for a smaller mode. In this | |
179 | case, the current code just punts on the slot. The proper thing | |
180 | to do is to back out and use one bit vector position for each | |
181 | byte of the entity associated with the slot. This depends on | |
182 | KNOWING that reload always generates the accesses for each of the | |
183 | bytes in some canonical (read that easy to understand several | |
184 | passes after reload happens) way. | |
185 | ||
186 | b) Reload sometimes decides that spill slot it allocated was not | |
187 | large enough for the mode and goes back and allocates more slots | |
188 | with the same mode and alias set. The backout in this case is a | |
189 | little more graceful than (a). In this case the slot is unmarked | |
190 | as being a spill slot and if final address comes out to be based | |
b8698a0f | 191 | off the frame pointer, the global algorithm handles this slot. |
6fb5fa3c DB |
192 | |
193 | c) For any pass that may prespill, there is currently no | |
194 | mechanism to tell the dse pass that the slot being used has the | |
195 | special properties that reload uses. It may be that all that is | |
0d52bcc1 | 196 | required is to have those passes make the same calls that reload |
6fb5fa3c DB |
197 | does, assuming that the alias sets can be manipulated in the same |
198 | way. */ | |
199 | ||
200 | /* There are limits to the size of constant offsets we model for the | |
201 | global problem. There are certainly test cases, that exceed this | |
202 | limit, however, it is unlikely that there are important programs | |
203 | that really have constant offsets this size. */ | |
204 | #define MAX_OFFSET (64 * 1024) | |
205 | ||
3f9b14ff SB |
206 | /* Obstack for the DSE dataflow bitmaps. We don't want to put these |
207 | on the default obstack because these bitmaps can grow quite large | |
208 | (~2GB for the small (!) test case of PR54146) and we'll hold on to | |
209 | all that memory until the end of the compiler run. | |
210 | As a bonus, delete_tree_live_info can destroy all the bitmaps by just | |
211 | releasing the whole obstack. */ | |
212 | static bitmap_obstack dse_bitmap_obstack; | |
213 | ||
214 | /* Obstack for other data. As for above: Kinda nice to be able to | |
215 | throw it all away at the end in one big sweep. */ | |
216 | static struct obstack dse_obstack; | |
217 | ||
218 | /* Scratch bitmap for cselib's cselib_expand_value_rtx. */ | |
6fb5fa3c | 219 | static bitmap scratch = NULL; |
3f9b14ff | 220 | |
da6603c6 | 221 | struct insn_info_type; |
6fb5fa3c DB |
222 | |
223 | /* This structure holds information about a candidate store. */ | |
6c1dae73 | 224 | class store_info |
6fb5fa3c | 225 | { |
6c1dae73 | 226 | public: |
6fb5fa3c DB |
227 | |
228 | /* False means this is a clobber. */ | |
229 | bool is_set; | |
230 | ||
8dd5516b JJ |
231 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
232 | bool is_large; | |
233 | ||
6fb5fa3c DB |
234 | /* The id of the mem group of the base address. If rtx_varies_p is |
235 | true, this is -1. Otherwise, it is the index into the group | |
236 | table. */ | |
237 | int group_id; | |
b8698a0f | 238 | |
6fb5fa3c DB |
239 | /* This is the cselib value. */ |
240 | cselib_val *cse_base; | |
241 | ||
242 | /* This canonized mem. */ | |
243 | rtx mem; | |
244 | ||
6216f94e | 245 | /* Canonized MEM address for use by canon_true_dependence. */ |
6fb5fa3c DB |
246 | rtx mem_addr; |
247 | ||
71d6a386 | 248 | /* The offset of the first byte associated with the operation. */ |
02ce5d90 | 249 | poly_int64 offset; |
71d6a386 RS |
250 | |
251 | /* The number of bytes covered by the operation. This is always exact | |
252 | and known (rather than -1). */ | |
02ce5d90 | 253 | poly_int64 width; |
8dd5516b JJ |
254 | |
255 | union | |
256 | { | |
257 | /* A bitmask as wide as the number of bytes in the word that | |
258 | contains a 1 if the byte may be needed. The store is unused if | |
259 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
260 | unsigned HOST_WIDE_INT small_bitmask; | |
261 | ||
262 | struct | |
263 | { | |
02ce5d90 RS |
264 | /* A bitmap with one bit per byte, or null if the number of |
265 | bytes isn't known at compile time. A cleared bit means | |
266 | the position is needed. Used if IS_LARGE is true. */ | |
dc491a25 | 267 | bitmap bmap; |
6fb5fa3c | 268 | |
02ce5d90 RS |
269 | /* When BITMAP is nonnull, this counts the number of set bits |
270 | (i.e. unneeded bytes) in the bitmap. If it is equal to | |
271 | WIDTH, the whole store is unused. | |
272 | ||
273 | When BITMAP is null: | |
274 | - the store is definitely not needed when COUNT == 1 | |
275 | - all the store is needed when COUNT == 0 and RHS is nonnull | |
276 | - otherwise we don't know which parts of the store are needed. */ | |
8dd5516b JJ |
277 | int count; |
278 | } large; | |
279 | } positions_needed; | |
6fb5fa3c DB |
280 | |
281 | /* The next store info for this insn. */ | |
99b1c316 | 282 | class store_info *next; |
6fb5fa3c DB |
283 | |
284 | /* The right hand side of the store. This is used if there is a | |
285 | subsequent reload of the mems address somewhere later in the | |
286 | basic block. */ | |
8dd5516b JJ |
287 | rtx rhs; |
288 | ||
289 | /* If rhs is or holds a constant, this contains that constant, | |
290 | otherwise NULL. */ | |
291 | rtx const_rhs; | |
292 | ||
293 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
294 | insn stored. These aren't eliminated early, because doing that | |
295 | might prevent the earlier larger store to be eliminated. */ | |
da6603c6 | 296 | struct insn_info_type *redundant_reason; |
6fb5fa3c DB |
297 | }; |
298 | ||
4fe663b0 L |
299 | /* Return a bitmask with the first N low bits set. */ |
300 | ||
301 | static unsigned HOST_WIDE_INT | |
302 | lowpart_bitmask (int n) | |
303 | { | |
dd4786fe | 304 | unsigned HOST_WIDE_INT mask = HOST_WIDE_INT_M1U; |
4fe663b0 L |
305 | return mask >> (HOST_BITS_PER_WIDE_INT - n); |
306 | } | |
307 | ||
fcb87c50 | 308 | static object_allocator<store_info> cse_store_info_pool ("cse_store_info_pool"); |
da6603c6 | 309 | |
fcb87c50 | 310 | static object_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool"); |
6fb5fa3c DB |
311 | |
312 | /* This structure holds information about a load. These are only | |
313 | built for rtx bases. */ | |
6c1dae73 | 314 | class read_info_type |
6fb5fa3c | 315 | { |
6c1dae73 | 316 | public: |
6fb5fa3c DB |
317 | /* The id of the mem group of the base address. */ |
318 | int group_id; | |
319 | ||
71d6a386 | 320 | /* The offset of the first byte associated with the operation. */ |
02ce5d90 | 321 | poly_int64 offset; |
71d6a386 RS |
322 | |
323 | /* The number of bytes covered by the operation, or -1 if not known. */ | |
02ce5d90 | 324 | poly_int64 width; |
6fb5fa3c DB |
325 | |
326 | /* The mem being read. */ | |
327 | rtx mem; | |
328 | ||
329 | /* The next read_info for this insn. */ | |
99b1c316 | 330 | class read_info_type *next; |
6fb5fa3c | 331 | }; |
99b1c316 | 332 | typedef class read_info_type *read_info_t; |
6fb5fa3c | 333 | |
fcb87c50 | 334 | static object_allocator<read_info_type> read_info_type_pool ("read_info_pool"); |
6fb5fa3c DB |
335 | |
336 | /* One of these records is created for each insn. */ | |
337 | ||
da6603c6 | 338 | struct insn_info_type |
6fb5fa3c DB |
339 | { |
340 | /* Set true if the insn contains a store but the insn itself cannot | |
341 | be deleted. This is set if the insn is a parallel and there is | |
342 | more than one non dead output or if the insn is in some way | |
343 | volatile. */ | |
344 | bool cannot_delete; | |
345 | ||
346 | /* This field is only used by the global algorithm. It is set true | |
347 | if the insn contains any read of mem except for a (1). This is | |
348 | also set if the insn is a call or has a clobber mem. If the insn | |
349 | contains a wild read, the use_rec will be null. */ | |
350 | bool wild_read; | |
351 | ||
d26c7090 ER |
352 | /* This is true only for CALL instructions which could potentially read |
353 | any non-frame memory location. This field is used by the global | |
354 | algorithm. */ | |
355 | bool non_frame_wild_read; | |
356 | ||
50f0f366 EB |
357 | /* This field is only used for the processing of const functions. |
358 | These functions cannot read memory, but they can read the stack | |
64520bdc EB |
359 | because that is where they may get their parms. We need to be |
360 | this conservative because, like the store motion pass, we don't | |
361 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
362 | Moreover, we need to distinguish two cases: | |
363 | 1. Before reload (register elimination), the stores related to | |
364 | outgoing arguments are stack pointer based and thus deemed | |
365 | of non-constant base in this pass. This requires special | |
366 | handling but also means that the frame pointer based stores | |
367 | need not be killed upon encountering a const function call. | |
368 | 2. After reload, the stores related to outgoing arguments can be | |
369 | either stack pointer or hard frame pointer based. This means | |
370 | that we have no other choice than also killing all the frame | |
371 | pointer based stores upon encountering a const function call. | |
57534689 JJ |
372 | This field is set after reload for const function calls and before |
373 | reload for const tail function calls on targets where arg pointer | |
374 | is the frame pointer. Having this set is less severe than a wild | |
375 | read, it just means that all the frame related stores are killed | |
376 | rather than all the stores. */ | |
64520bdc EB |
377 | bool frame_read; |
378 | ||
379 | /* This field is only used for the processing of const functions. | |
380 | It is set if the insn may contain a stack pointer based store. */ | |
50f0f366 | 381 | bool stack_pointer_based; |
6fb5fa3c DB |
382 | |
383 | /* This is true if any of the sets within the store contains a | |
384 | cselib base. Such stores can only be deleted by the local | |
385 | algorithm. */ | |
386 | bool contains_cselib_groups; | |
387 | ||
388 | /* The insn. */ | |
dd60a84c | 389 | rtx_insn *insn; |
6fb5fa3c DB |
390 | |
391 | /* The list of mem sets or mem clobbers that are contained in this | |
392 | insn. If the insn is deletable, it contains only one mem set. | |
393 | But it could also contain clobbers. Insns that contain more than | |
394 | one mem set are not deletable, but each of those mems are here in | |
6ed3da00 | 395 | order to provide info to delete other insns. */ |
fd6daec9 | 396 | store_info *store_rec; |
6fb5fa3c DB |
397 | |
398 | /* The linked list of mem uses in this insn. Only the reads from | |
399 | rtx bases are listed here. The reads to cselib bases are | |
400 | completely processed during the first scan and so are never | |
401 | created. */ | |
402 | read_info_t read_rec; | |
403 | ||
9e582b1d JR |
404 | /* The live fixed registers. We assume only fixed registers can |
405 | cause trouble by being clobbered from an expanded pattern; | |
406 | storing only the live fixed registers (rather than all registers) | |
407 | means less memory needs to be allocated / copied for the individual | |
408 | stores. */ | |
409 | regset fixed_regs_live; | |
410 | ||
6fb5fa3c | 411 | /* The prev insn in the basic block. */ |
da6603c6 | 412 | struct insn_info_type * prev_insn; |
6fb5fa3c DB |
413 | |
414 | /* The linked list of insns that are in consideration for removal in | |
073a8998 | 415 | the forwards pass through the basic block. This pointer may be |
6fb5fa3c | 416 | trash as it is not cleared when a wild read occurs. The only |
fa10beec | 417 | time it is guaranteed to be correct is when the traversal starts |
6fb5fa3c | 418 | at active_local_stores. */ |
da6603c6 | 419 | struct insn_info_type * next_local_store; |
6fb5fa3c | 420 | }; |
da6603c6 | 421 | typedef struct insn_info_type *insn_info_t; |
6fb5fa3c | 422 | |
fcb87c50 | 423 | static object_allocator<insn_info_type> insn_info_type_pool ("insn_info_pool"); |
6fb5fa3c DB |
424 | |
425 | /* The linked list of stores that are under consideration in this | |
b8698a0f | 426 | basic block. */ |
6fb5fa3c | 427 | static insn_info_t active_local_stores; |
dabd47e7 | 428 | static int active_local_stores_len; |
6fb5fa3c | 429 | |
da6603c6 | 430 | struct dse_bb_info_type |
6fb5fa3c | 431 | { |
6fb5fa3c DB |
432 | /* Pointer to the insn info for the last insn in the block. These |
433 | are linked so this is how all of the insns are reached. During | |
434 | scanning this is the current insn being scanned. */ | |
435 | insn_info_t last_insn; | |
436 | ||
437 | /* The info for the global dataflow problem. */ | |
438 | ||
439 | ||
440 | /* This is set if the transfer function should and in the wild_read | |
441 | bitmap before applying the kill and gen sets. That vector knocks | |
442 | out most of the bits in the bitmap and thus speeds up the | |
443 | operations. */ | |
444 | bool apply_wild_read; | |
445 | ||
02b47899 KZ |
446 | /* The following 4 bitvectors hold information about which positions |
447 | of which stores are live or dead. They are indexed by | |
448 | get_bitmap_index. */ | |
449 | ||
6fb5fa3c DB |
450 | /* The set of store positions that exist in this block before a wild read. */ |
451 | bitmap gen; | |
b8698a0f | 452 | |
6fb5fa3c DB |
453 | /* The set of load positions that exist in this block above the |
454 | same position of a store. */ | |
455 | bitmap kill; | |
456 | ||
457 | /* The set of stores that reach the top of the block without being | |
458 | killed by a read. | |
459 | ||
460 | Do not represent the in if it is all ones. Note that this is | |
461 | what the bitvector should logically be initialized to for a set | |
462 | intersection problem. However, like the kill set, this is too | |
463 | expensive. So initially, the in set will only be created for the | |
464 | exit block and any block that contains a wild read. */ | |
465 | bitmap in; | |
466 | ||
467 | /* The set of stores that reach the bottom of the block from it's | |
468 | successors. | |
469 | ||
470 | Do not represent the in if it is all ones. Note that this is | |
471 | what the bitvector should logically be initialized to for a set | |
472 | intersection problem. However, like the kill and in set, this is | |
473 | too expensive. So what is done is that the confluence operator | |
474 | just initializes the vector from one of the out sets of the | |
475 | successors of the block. */ | |
476 | bitmap out; | |
02b47899 KZ |
477 | |
478 | /* The following bitvector is indexed by the reg number. It | |
479 | contains the set of regs that are live at the current instruction | |
480 | being processed. While it contains info for all of the | |
9e582b1d JR |
481 | registers, only the hard registers are actually examined. It is used |
482 | to assure that shift and/or add sequences that are inserted do not | |
073a8998 | 483 | accidentally clobber live hard regs. */ |
02b47899 | 484 | bitmap regs_live; |
6fb5fa3c DB |
485 | }; |
486 | ||
da6603c6 | 487 | typedef struct dse_bb_info_type *bb_info_t; |
fb0b2914 ML |
488 | |
489 | static object_allocator<dse_bb_info_type> dse_bb_info_type_pool | |
fcb87c50 | 490 | ("bb_info_pool"); |
6fb5fa3c DB |
491 | |
492 | /* Table to hold all bb_infos. */ | |
493 | static bb_info_t *bb_table; | |
494 | ||
495 | /* There is a group_info for each rtx base that is used to reference | |
496 | memory. There are also not many of the rtx bases because they are | |
497 | very limited in scope. */ | |
498 | ||
b8698a0f | 499 | struct group_info |
6fb5fa3c DB |
500 | { |
501 | /* The actual base of the address. */ | |
502 | rtx rtx_base; | |
503 | ||
504 | /* The sequential id of the base. This allows us to have a | |
505 | canonical ordering of these that is not based on addresses. */ | |
506 | int id; | |
507 | ||
8f5929e1 JJ |
508 | /* True if there are any positions that are to be processed |
509 | globally. */ | |
510 | bool process_globally; | |
511 | ||
512 | /* True if the base of this group is either the frame_pointer or | |
513 | hard_frame_pointer. */ | |
514 | bool frame_related; | |
515 | ||
d32f725a OH |
516 | /* A mem wrapped around the base pointer for the group in order to do |
517 | read dependency. It must be given BLKmode in order to encompass all | |
518 | the possible offsets from the base. */ | |
6fb5fa3c | 519 | rtx base_mem; |
b8698a0f | 520 | |
6216f94e JJ |
521 | /* Canonized version of base_mem's address. */ |
522 | rtx canon_base_addr; | |
6fb5fa3c DB |
523 | |
524 | /* These two sets of two bitmaps are used to keep track of how many | |
6ed3da00 | 525 | stores are actually referencing that position from this base. We |
6fb5fa3c | 526 | only do this for rtx bases as this will be used to assign |
6ed3da00 | 527 | positions in the bitmaps for the global problem. Bit N is set in |
6fb5fa3c DB |
528 | store1 on the first store for offset N. Bit N is set in store2 |
529 | for the second store to offset N. This is all we need since we | |
530 | only care about offsets that have two or more stores for them. | |
531 | ||
532 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
533 | for 0 and greater offsets. | |
534 | ||
535 | There is one special case here, for stores into the stack frame, | |
536 | we will or store1 into store2 before deciding which stores look | |
537 | at globally. This is because stores to the stack frame that have | |
538 | no other reads before the end of the function can also be | |
539 | deleted. */ | |
540 | bitmap store1_n, store1_p, store2_n, store2_p; | |
541 | ||
d26c7090 ER |
542 | /* These bitmaps keep track of offsets in this group escape this function. |
543 | An offset escapes if it corresponds to a named variable whose | |
544 | addressable flag is set. */ | |
545 | bitmap escaped_n, escaped_p; | |
546 | ||
6ed3da00 | 547 | /* The positions in this bitmap have the same assignments as the in, |
6fb5fa3c | 548 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6ed3da00 | 549 | the positions that are occupied by stores for this group. */ |
6fb5fa3c DB |
550 | bitmap group_kill; |
551 | ||
6fb5fa3c | 552 | /* The offset_map is used to map the offsets from this base into |
6ed3da00 | 553 | positions in the global bitmaps. It is only created after all of |
6fb5fa3c DB |
554 | the all of stores have been scanned and we know which ones we |
555 | care about. */ | |
b8698a0f L |
556 | int *offset_map_n, *offset_map_p; |
557 | int offset_map_size_n, offset_map_size_p; | |
6fb5fa3c | 558 | }; |
da6603c6 | 559 | |
fcb87c50 | 560 | static object_allocator<group_info> group_info_pool ("rtx_group_info_pool"); |
6fb5fa3c | 561 | |
6fb5fa3c DB |
562 | /* Index into the rtx_group_vec. */ |
563 | static int rtx_group_next_id; | |
564 | ||
6fb5fa3c | 565 | |
fd6daec9 | 566 | static vec<group_info *> rtx_group_vec; |
6fb5fa3c DB |
567 | |
568 | ||
569 | /* This structure holds the set of changes that are being deferred | |
570 | when removing read operation. See replace_read. */ | |
b8698a0f | 571 | struct deferred_change |
6fb5fa3c DB |
572 | { |
573 | ||
574 | /* The mem that is being replaced. */ | |
575 | rtx *loc; | |
576 | ||
577 | /* The reg it is being replaced with. */ | |
578 | rtx reg; | |
579 | ||
580 | struct deferred_change *next; | |
581 | }; | |
582 | ||
fb0b2914 | 583 | static object_allocator<deferred_change> deferred_change_pool |
fcb87c50 | 584 | ("deferred_change_pool"); |
6fb5fa3c | 585 | |
fd6daec9 | 586 | static deferred_change *deferred_change_list = NULL; |
6fb5fa3c | 587 | |
e3b5732b | 588 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
9dd9bf80 | 589 | this for vararg functions because they play games with the frame. */ |
6fb5fa3c DB |
590 | static bool stores_off_frame_dead_at_return; |
591 | ||
592 | /* Counter for stats. */ | |
b8698a0f L |
593 | static int globally_deleted; |
594 | static int locally_deleted; | |
b8698a0f | 595 | |
6fb5fa3c DB |
596 | static bitmap all_blocks; |
597 | ||
d26c7090 ER |
598 | /* Locations that are killed by calls in the global phase. */ |
599 | static bitmap kill_on_calls; | |
600 | ||
6fb5fa3c DB |
601 | /* The number of bits used in the global bitmaps. */ |
602 | static unsigned int current_position; | |
71d6a386 RS |
603 | |
604 | /* Print offset range [OFFSET, OFFSET + WIDTH) to FILE. */ | |
605 | ||
606 | static void | |
607 | print_range (FILE *file, poly_int64 offset, poly_int64 width) | |
608 | { | |
609 | fprintf (file, "["); | |
610 | print_dec (offset, file, SIGNED); | |
611 | fprintf (file, ".."); | |
612 | print_dec (offset + width, file, SIGNED); | |
613 | fprintf (file, ")"); | |
614 | } | |
6fb5fa3c DB |
615 | \f |
616 | /*---------------------------------------------------------------------------- | |
617 | Zeroth step. | |
618 | ||
b8698a0f | 619 | Initialization. |
6fb5fa3c DB |
620 | ----------------------------------------------------------------------------*/ |
621 | ||
6fb5fa3c | 622 | |
6fb5fa3c DB |
623 | /* Hashtable callbacks for maintaining the "bases" field of |
624 | store_group_info, given that the addresses are function invariants. */ | |
625 | ||
8d67ee55 | 626 | struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
703c8606 | 627 | { |
67f58944 TS |
628 | static inline hashval_t hash (const group_info *); |
629 | static inline bool equal (const group_info *, const group_info *); | |
703c8606 LC |
630 | }; |
631 | ||
632 | inline bool | |
67f58944 TS |
633 | invariant_group_base_hasher::equal (const group_info *gi1, |
634 | const group_info *gi2) | |
6fb5fa3c | 635 | { |
6fb5fa3c DB |
636 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
637 | } | |
638 | ||
703c8606 | 639 | inline hashval_t |
67f58944 | 640 | invariant_group_base_hasher::hash (const group_info *gi) |
6fb5fa3c | 641 | { |
6fb5fa3c DB |
642 | int do_not_record; |
643 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
644 | } | |
645 | ||
703c8606 | 646 | /* Tables of group_info structures, hashed by base value. */ |
c203e8a7 | 647 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
703c8606 | 648 | |
6fb5fa3c DB |
649 | |
650 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
651 | ||
fd6daec9 | 652 | static group_info * |
6fb5fa3c DB |
653 | get_group_info (rtx base) |
654 | { | |
b8698a0f | 655 | struct group_info tmp_gi; |
fd6daec9 | 656 | group_info *gi; |
703c8606 | 657 | group_info **slot; |
6fb5fa3c | 658 | |
ac6929b5 RB |
659 | gcc_assert (base != NULL_RTX); |
660 | ||
661 | /* Find the store_base_info structure for BASE, creating a new one | |
662 | if necessary. */ | |
663 | tmp_gi.rtx_base = base; | |
664 | slot = rtx_group_table->find_slot (&tmp_gi, INSERT); | |
665 | gi = *slot; | |
6fb5fa3c DB |
666 | |
667 | if (gi == NULL) | |
668 | { | |
fb0b2914 | 669 | *slot = gi = group_info_pool.allocate (); |
6fb5fa3c DB |
670 | gi->rtx_base = base; |
671 | gi->id = rtx_group_next_id++; | |
d32f725a | 672 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
6216f94e | 673 | gi->canon_base_addr = canon_rtx (base); |
3f9b14ff SB |
674 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
675 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
676 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
677 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
678 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
679 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
680 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 681 | gi->process_globally = false; |
b8698a0f | 682 | gi->frame_related = |
6fb5fa3c DB |
683 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
684 | gi->offset_map_size_n = 0; | |
685 | gi->offset_map_size_p = 0; | |
686 | gi->offset_map_n = NULL; | |
687 | gi->offset_map_p = NULL; | |
9771b263 | 688 | rtx_group_vec.safe_push (gi); |
6fb5fa3c DB |
689 | } |
690 | ||
691 | return gi; | |
692 | } | |
693 | ||
694 | ||
695 | /* Initialization of data structures. */ | |
696 | ||
697 | static void | |
698 | dse_step0 (void) | |
699 | { | |
700 | locally_deleted = 0; | |
701 | globally_deleted = 0; | |
6fb5fa3c | 702 | |
3f9b14ff SB |
703 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
704 | gcc_obstack_init (&dse_obstack); | |
705 | ||
706 | scratch = BITMAP_ALLOC (®_obstack); | |
707 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 708 | |
6fb5fa3c | 709 | |
c203e8a7 | 710 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
6fb5fa3c | 711 | |
8b1c6fd7 | 712 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
6fb5fa3c DB |
713 | rtx_group_next_id = 0; |
714 | ||
e3b5732b | 715 | stores_off_frame_dead_at_return = !cfun->stdarg; |
6fb5fa3c DB |
716 | |
717 | init_alias_analysis (); | |
6fb5fa3c DB |
718 | } |
719 | ||
720 | ||
721 | \f | |
722 | /*---------------------------------------------------------------------------- | |
723 | First step. | |
724 | ||
725 | Scan all of the insns. Any random ordering of the blocks is fine. | |
0d52bcc1 | 726 | Each block is scanned in forward order to accommodate cselib which |
6fb5fa3c DB |
727 | is used to remove stores with non-constant bases. |
728 | ----------------------------------------------------------------------------*/ | |
729 | ||
730 | /* Delete all of the store_info recs from INSN_INFO. */ | |
731 | ||
b8698a0f | 732 | static void |
6fb5fa3c DB |
733 | free_store_info (insn_info_t insn_info) |
734 | { | |
fd6daec9 TS |
735 | store_info *cur = insn_info->store_rec; |
736 | while (cur) | |
6fb5fa3c | 737 | { |
fd6daec9 TS |
738 | store_info *next = cur->next; |
739 | if (cur->is_large) | |
740 | BITMAP_FREE (cur->positions_needed.large.bmap); | |
741 | if (cur->cse_base) | |
742 | cse_store_info_pool.remove (cur); | |
6fb5fa3c | 743 | else |
fd6daec9 TS |
744 | rtx_store_info_pool.remove (cur); |
745 | cur = next; | |
6fb5fa3c DB |
746 | } |
747 | ||
748 | insn_info->cannot_delete = true; | |
749 | insn_info->contains_cselib_groups = false; | |
750 | insn_info->store_rec = NULL; | |
751 | } | |
752 | ||
a79683d5 | 753 | struct note_add_store_info |
9e582b1d | 754 | { |
dc01c3d1 | 755 | rtx_insn *first, *current; |
9e582b1d JR |
756 | regset fixed_regs_live; |
757 | bool failure; | |
a79683d5 | 758 | }; |
9e582b1d JR |
759 | |
760 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
761 | Check if a register is clobbered which is live afterwards. */ | |
762 | ||
763 | static void | |
764 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
765 | { | |
dc01c3d1 | 766 | rtx_insn *insn; |
9e582b1d | 767 | note_add_store_info *info = (note_add_store_info *) data; |
9e582b1d JR |
768 | |
769 | if (!REG_P (loc)) | |
770 | return; | |
771 | ||
772 | /* If this register is referenced by the current or an earlier insn, | |
773 | that's OK. E.g. this applies to the register that is being incremented | |
774 | with this addition. */ | |
775 | for (insn = info->first; | |
776 | insn != NEXT_INSN (info->current); | |
777 | insn = NEXT_INSN (insn)) | |
778 | if (reg_referenced_p (loc, PATTERN (insn))) | |
779 | return; | |
780 | ||
781 | /* If we come here, we have a clobber of a register that's only OK | |
782 | if that register is not live. If we don't have liveness information | |
783 | available, fail now. */ | |
784 | if (!info->fixed_regs_live) | |
785 | { | |
53d1bae9 | 786 | info->failure = true; |
9e582b1d JR |
787 | return; |
788 | } | |
789 | /* Now check if this is a live fixed register. */ | |
53d1bae9 RS |
790 | unsigned int end_regno = END_REGNO (loc); |
791 | for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) | |
792 | if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) | |
793 | info->failure = true; | |
9e582b1d JR |
794 | } |
795 | ||
4deef538 AO |
796 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
797 | SRC + SRCOFF before insn ARG. */ | |
6fb5fa3c DB |
798 | |
799 | static int | |
4deef538 AO |
800 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
801 | rtx op ATTRIBUTE_UNUSED, | |
802 | rtx dest, rtx src, rtx srcoff, void *arg) | |
6fb5fa3c | 803 | { |
9e582b1d | 804 | insn_info_t insn_info = (insn_info_t) arg; |
dc01c3d1 | 805 | rtx_insn *insn = insn_info->insn, *new_insn, *cur; |
9e582b1d | 806 | note_add_store_info info; |
b8698a0f | 807 | |
4deef538 AO |
808 | /* We can reuse all operands without copying, because we are about |
809 | to delete the insn that contained it. */ | |
9e582b1d | 810 | if (srcoff) |
ed079c4b JJ |
811 | { |
812 | start_sequence (); | |
813 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
814 | new_insn = get_insns (); | |
815 | end_sequence (); | |
816 | } | |
9e582b1d | 817 | else |
1476d1bd | 818 | new_insn = gen_move_insn (dest, src); |
9e582b1d JR |
819 | info.first = new_insn; |
820 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
821 | info.failure = false; | |
822 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
823 | { | |
824 | info.current = cur; | |
825 | note_stores (PATTERN (cur), note_add_store, &info); | |
826 | } | |
6fb5fa3c | 827 | |
9e582b1d JR |
828 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
829 | return it immediately, communicating the failure to its caller. */ | |
830 | if (info.failure) | |
831 | return 1; | |
832 | ||
833 | emit_insn_before (new_insn, insn); | |
6fb5fa3c | 834 | |
8d8e205b | 835 | return 0; |
6fb5fa3c DB |
836 | } |
837 | ||
9e582b1d JR |
838 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
839 | is there, is split into a separate insn. | |
840 | Return true on success (or if there was nothing to do), false on failure. */ | |
6fb5fa3c | 841 | |
9e582b1d JR |
842 | static bool |
843 | check_for_inc_dec_1 (insn_info_t insn_info) | |
6fb5fa3c | 844 | { |
dd60a84c | 845 | rtx_insn *insn = insn_info->insn; |
6fb5fa3c DB |
846 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
847 | if (note) | |
8d8e205b RS |
848 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
849 | insn_info) == 0; | |
9e582b1d | 850 | return true; |
6fb5fa3c DB |
851 | } |
852 | ||
853 | ||
9e582b1d JR |
854 | /* Entry point for postreload. If you work on reload_cse, or you need this |
855 | anywhere else, consider if you can provide register liveness information | |
856 | and add a parameter to this function so that it can be passed down in | |
857 | insn_info.fixed_regs_live. */ | |
858 | bool | |
dd60a84c | 859 | check_for_inc_dec (rtx_insn *insn) |
9e582b1d | 860 | { |
da6603c6 | 861 | insn_info_type insn_info; |
9e582b1d JR |
862 | rtx note; |
863 | ||
864 | insn_info.insn = insn; | |
865 | insn_info.fixed_regs_live = NULL; | |
866 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
867 | if (note) | |
8d8e205b RS |
868 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
869 | &insn_info) == 0; | |
9e582b1d JR |
870 | return true; |
871 | } | |
872 | ||
b8698a0f | 873 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
6fb5fa3c DB |
874 | |
875 | static void | |
876 | delete_dead_store_insn (insn_info_t insn_info) | |
877 | { | |
878 | read_info_t read_info; | |
879 | ||
880 | if (!dbg_cnt (dse)) | |
881 | return; | |
882 | ||
9e582b1d JR |
883 | if (!check_for_inc_dec_1 (insn_info)) |
884 | return; | |
456610d3 | 885 | if (dump_file && (dump_flags & TDF_DETAILS)) |
ac6929b5 RB |
886 | fprintf (dump_file, "Locally deleting insn %d\n", |
887 | INSN_UID (insn_info->insn)); | |
6fb5fa3c DB |
888 | |
889 | free_store_info (insn_info); | |
890 | read_info = insn_info->read_rec; | |
b8698a0f | 891 | |
6fb5fa3c DB |
892 | while (read_info) |
893 | { | |
894 | read_info_t next = read_info->next; | |
fb0b2914 | 895 | read_info_type_pool.remove (read_info); |
6fb5fa3c DB |
896 | read_info = next; |
897 | } | |
898 | insn_info->read_rec = NULL; | |
899 | ||
900 | delete_insn (insn_info->insn); | |
901 | locally_deleted++; | |
902 | insn_info->insn = NULL; | |
903 | ||
904 | insn_info->wild_read = false; | |
905 | } | |
906 | ||
88d8330d EB |
907 | /* Return whether DECL, a local variable, can possibly escape the current |
908 | function scope. */ | |
909 | ||
910 | static bool | |
911 | local_variable_can_escape (tree decl) | |
912 | { | |
913 | if (TREE_ADDRESSABLE (decl)) | |
914 | return true; | |
915 | ||
916 | /* If this is a partitioned variable, we need to consider all the variables | |
917 | in the partition. This is necessary because a store into one of them can | |
918 | be replaced with a store into another and this may not change the outcome | |
919 | of the escape analysis. */ | |
920 | if (cfun->gimple_df->decls_to_pointers != NULL) | |
921 | { | |
39c8aaa4 | 922 | tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); |
88d8330d | 923 | if (namep) |
39c8aaa4 | 924 | return TREE_ADDRESSABLE (*namep); |
88d8330d EB |
925 | } |
926 | ||
927 | return false; | |
928 | } | |
929 | ||
930 | /* Return whether EXPR can possibly escape the current function scope. */ | |
931 | ||
d26c7090 ER |
932 | static bool |
933 | can_escape (tree expr) | |
934 | { | |
935 | tree base; | |
936 | if (!expr) | |
937 | return true; | |
938 | base = get_base_address (expr); | |
939 | if (DECL_P (base) | |
88d8330d | 940 | && !may_be_aliased (base) |
8813a647 | 941 | && !(VAR_P (base) |
88d8330d EB |
942 | && !DECL_EXTERNAL (base) |
943 | && !TREE_STATIC (base) | |
944 | && local_variable_can_escape (base))) | |
d26c7090 ER |
945 | return false; |
946 | return true; | |
947 | } | |
6fb5fa3c DB |
948 | |
949 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
950 | OFFSET and WIDTH. */ | |
951 | ||
952 | static void | |
02ce5d90 | 953 | set_usage_bits (group_info *group, poly_int64 offset, poly_int64 width, |
d26c7090 | 954 | tree expr) |
6fb5fa3c | 955 | { |
02ce5d90 RS |
956 | /* Non-constant offsets and widths act as global kills, so there's no point |
957 | trying to use them to derive global DSE candidates. */ | |
958 | HOST_WIDE_INT i, const_offset, const_width; | |
d26c7090 | 959 | bool expr_escapes = can_escape (expr); |
02ce5d90 RS |
960 | if (offset.is_constant (&const_offset) |
961 | && width.is_constant (&const_width) | |
962 | && const_offset > -MAX_OFFSET | |
963 | && const_offset + const_width < MAX_OFFSET) | |
964 | for (i = const_offset; i < const_offset + const_width; ++i) | |
6fb5fa3c DB |
965 | { |
966 | bitmap store1; | |
967 | bitmap store2; | |
d26c7090 | 968 | bitmap escaped; |
6fb5fa3c DB |
969 | int ai; |
970 | if (i < 0) | |
971 | { | |
972 | store1 = group->store1_n; | |
973 | store2 = group->store2_n; | |
d26c7090 | 974 | escaped = group->escaped_n; |
6fb5fa3c DB |
975 | ai = -i; |
976 | } | |
977 | else | |
978 | { | |
979 | store1 = group->store1_p; | |
980 | store2 = group->store2_p; | |
d26c7090 | 981 | escaped = group->escaped_p; |
6fb5fa3c DB |
982 | ai = i; |
983 | } | |
b8698a0f | 984 | |
fcaa4ca4 | 985 | if (!bitmap_set_bit (store1, ai)) |
6fb5fa3c | 986 | bitmap_set_bit (store2, ai); |
b8698a0f | 987 | else |
6fb5fa3c | 988 | { |
6fb5fa3c DB |
989 | if (i < 0) |
990 | { | |
991 | if (group->offset_map_size_n < ai) | |
992 | group->offset_map_size_n = ai; | |
993 | } | |
994 | else | |
995 | { | |
996 | if (group->offset_map_size_p < ai) | |
997 | group->offset_map_size_p = ai; | |
998 | } | |
999 | } | |
d26c7090 ER |
1000 | if (expr_escapes) |
1001 | bitmap_set_bit (escaped, ai); | |
6fb5fa3c DB |
1002 | } |
1003 | } | |
1004 | ||
d26c7090 ER |
1005 | static void |
1006 | reset_active_stores (void) | |
1007 | { | |
1008 | active_local_stores = NULL; | |
1009 | active_local_stores_len = 0; | |
1010 | } | |
6fb5fa3c | 1011 | |
d26c7090 | 1012 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
6fb5fa3c DB |
1013 | |
1014 | static void | |
d26c7090 | 1015 | free_read_records (bb_info_t bb_info) |
6fb5fa3c DB |
1016 | { |
1017 | insn_info_t insn_info = bb_info->last_insn; | |
1018 | read_info_t *ptr = &insn_info->read_rec; | |
6fb5fa3c DB |
1019 | while (*ptr) |
1020 | { | |
1021 | read_info_t next = (*ptr)->next; | |
ac6929b5 RB |
1022 | read_info_type_pool.remove (*ptr); |
1023 | *ptr = next; | |
6fb5fa3c | 1024 | } |
d26c7090 ER |
1025 | } |
1026 | ||
1027 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1028 | ||
1029 | static void | |
1030 | add_wild_read (bb_info_t bb_info) | |
1031 | { | |
1032 | insn_info_t insn_info = bb_info->last_insn; | |
6fb5fa3c | 1033 | insn_info->wild_read = true; |
d26c7090 ER |
1034 | free_read_records (bb_info); |
1035 | reset_active_stores (); | |
6fb5fa3c DB |
1036 | } |
1037 | ||
d26c7090 ER |
1038 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1039 | non-frame locations. */ | |
1040 | ||
1041 | static void | |
1042 | add_non_frame_wild_read (bb_info_t bb_info) | |
1043 | { | |
1044 | insn_info_t insn_info = bb_info->last_insn; | |
1045 | insn_info->non_frame_wild_read = true; | |
1046 | free_read_records (bb_info); | |
1047 | reset_active_stores (); | |
1048 | } | |
6fb5fa3c | 1049 | |
50f0f366 EB |
1050 | /* Return true if X is a constant or one of the registers that behave |
1051 | as a constant over the life of a function. This is equivalent to | |
1052 | !rtx_varies_p for memory addresses. */ | |
6fb5fa3c DB |
1053 | |
1054 | static bool | |
1055 | const_or_frame_p (rtx x) | |
1056 | { | |
d8116890 KZ |
1057 | if (CONSTANT_P (x)) |
1058 | return true; | |
1059 | ||
1060 | if (GET_CODE (x) == REG) | |
6fb5fa3c | 1061 | { |
6fb5fa3c DB |
1062 | /* Note that we have to test for the actual rtx used for the frame |
1063 | and arg pointers and not just the register number in case we have | |
1064 | eliminated the frame and/or arg pointer and are using it | |
1065 | for pseudos. */ | |
1066 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1067 | /* The arg pointer varies if it is not a fixed register. */ | |
1068 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1069 | || x == pic_offset_table_rtx) | |
1070 | return true; | |
1071 | return false; | |
6fb5fa3c | 1072 | } |
da6603c6 | 1073 | |
d8116890 | 1074 | return false; |
6fb5fa3c DB |
1075 | } |
1076 | ||
b8698a0f L |
1077 | /* Take all reasonable action to put the address of MEM into the form |
1078 | that we can do analysis on. | |
6fb5fa3c DB |
1079 | |
1080 | The gold standard is to get the address into the form: address + | |
1081 | OFFSET where address is something that rtx_varies_p considers a | |
1082 | constant. When we can get the address in this form, we can do | |
1083 | global analysis on it. Note that for constant bases, address is | |
1084 | not actually returned, only the group_id. The address can be | |
1085 | obtained from that. | |
1086 | ||
1087 | If that fails, we try cselib to get a value we can at least use | |
b8698a0f L |
1088 | locally. If that fails we return false. |
1089 | ||
6fb5fa3c DB |
1090 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1091 | group for non_varying bases. | |
1092 | ||
1093 | FOR_READ is true if this is a mem read and false if not. */ | |
1094 | ||
1095 | static bool | |
1096 | canon_address (rtx mem, | |
6fb5fa3c | 1097 | int *group_id, |
02ce5d90 | 1098 | poly_int64 *offset, |
6fb5fa3c DB |
1099 | cselib_val **base) |
1100 | { | |
ef4bddc2 | 1101 | machine_mode address_mode = get_address_mode (mem); |
6fb5fa3c DB |
1102 | rtx mem_address = XEXP (mem, 0); |
1103 | rtx expanded_address, address; | |
403c7520 JJ |
1104 | int expanded; |
1105 | ||
4deef538 | 1106 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
6fb5fa3c | 1107 | |
456610d3 | 1108 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1109 | { |
1110 | fprintf (dump_file, " mem: "); | |
1111 | print_inline_rtx (dump_file, mem_address, 0); | |
1112 | fprintf (dump_file, "\n"); | |
1113 | } | |
1114 | ||
403c7520 JJ |
1115 | /* First see if just canon_rtx (mem_address) is const or frame, |
1116 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1117 | address = NULL_RTX; | |
1118 | for (expanded = 0; expanded < 2; expanded++) | |
1119 | { | |
1120 | if (expanded) | |
1121 | { | |
1122 | /* Use cselib to replace all of the reg references with the full | |
b8698a0f | 1123 | expression. This will take care of the case where we have |
6fb5fa3c | 1124 | |
403c7520 JJ |
1125 | r_x = base + offset; |
1126 | val = *r_x; | |
b8698a0f L |
1127 | |
1128 | by making it into | |
6fb5fa3c | 1129 | |
403c7520 | 1130 | val = *(base + offset); */ |
6fb5fa3c | 1131 | |
403c7520 JJ |
1132 | expanded_address = cselib_expand_value_rtx (mem_address, |
1133 | scratch, 5); | |
6fb5fa3c | 1134 | |
403c7520 JJ |
1135 | /* If this fails, just go with the address from first |
1136 | iteration. */ | |
1137 | if (!expanded_address) | |
1138 | break; | |
1139 | } | |
1140 | else | |
1141 | expanded_address = mem_address; | |
6fb5fa3c | 1142 | |
403c7520 JJ |
1143 | /* Split the address into canonical BASE + OFFSET terms. */ |
1144 | address = canon_rtx (expanded_address); | |
6fb5fa3c | 1145 | |
403c7520 | 1146 | *offset = 0; |
6fb5fa3c | 1147 | |
456610d3 | 1148 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1149 | { |
1150 | if (expanded) | |
1151 | { | |
1152 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1153 | print_inline_rtx (dump_file, expanded_address, 0); | |
1154 | fprintf (dump_file, "\n"); | |
1155 | } | |
6fb5fa3c | 1156 | |
403c7520 JJ |
1157 | fprintf (dump_file, "\n after canon_rtx address: "); |
1158 | print_inline_rtx (dump_file, address, 0); | |
1159 | fprintf (dump_file, "\n"); | |
1160 | } | |
6fb5fa3c | 1161 | |
403c7520 JJ |
1162 | if (GET_CODE (address) == CONST) |
1163 | address = XEXP (address, 0); | |
6fb5fa3c | 1164 | |
02ce5d90 | 1165 | address = strip_offset_and_add (address, offset); |
6fb5fa3c | 1166 | |
09e881c9 BE |
1167 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1168 | && const_or_frame_p (address)) | |
6fb5fa3c | 1169 | { |
fd6daec9 | 1170 | group_info *group = get_group_info (address); |
403c7520 | 1171 | |
456610d3 | 1172 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
1173 | { |
1174 | fprintf (dump_file, " gid=%d offset=", group->id); | |
1175 | print_dec (*offset, dump_file); | |
1176 | fprintf (dump_file, "\n"); | |
1177 | } | |
403c7520 JJ |
1178 | *base = NULL; |
1179 | *group_id = group->id; | |
1180 | return true; | |
6fb5fa3c | 1181 | } |
403c7520 JJ |
1182 | } |
1183 | ||
4deef538 | 1184 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
403c7520 JJ |
1185 | *group_id = -1; |
1186 | ||
1187 | if (*base == NULL) | |
1188 | { | |
456610d3 | 1189 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1190 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1191 | return false; | |
6fb5fa3c | 1192 | } |
456610d3 | 1193 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
1194 | { |
1195 | fprintf (dump_file, " varying cselib base=%u:%u offset = ", | |
1196 | (*base)->uid, (*base)->hash); | |
1197 | print_dec (*offset, dump_file); | |
1198 | fprintf (dump_file, "\n"); | |
1199 | } | |
6fb5fa3c DB |
1200 | return true; |
1201 | } | |
1202 | ||
1203 | ||
1204 | /* Clear the rhs field from the active_local_stores array. */ | |
1205 | ||
1206 | static void | |
1207 | clear_rhs_from_active_local_stores (void) | |
1208 | { | |
1209 | insn_info_t ptr = active_local_stores; | |
1210 | ||
1211 | while (ptr) | |
1212 | { | |
fd6daec9 | 1213 | store_info *store_info = ptr->store_rec; |
6fb5fa3c DB |
1214 | /* Skip the clobbers. */ |
1215 | while (!store_info->is_set) | |
1216 | store_info = store_info->next; | |
1217 | ||
1218 | store_info->rhs = NULL; | |
8dd5516b | 1219 | store_info->const_rhs = NULL; |
6fb5fa3c DB |
1220 | |
1221 | ptr = ptr->next_local_store; | |
1222 | } | |
1223 | } | |
1224 | ||
1225 | ||
8dd5516b JJ |
1226 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1227 | ||
1228 | static inline void | |
fd6daec9 | 1229 | set_position_unneeded (store_info *s_info, int pos) |
8dd5516b JJ |
1230 | { |
1231 | if (__builtin_expect (s_info->is_large, false)) | |
1232 | { | |
fcaa4ca4 NF |
1233 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1234 | s_info->positions_needed.large.count++; | |
8dd5516b JJ |
1235 | } |
1236 | else | |
1237 | s_info->positions_needed.small_bitmask | |
fecfbfa4 | 1238 | &= ~(HOST_WIDE_INT_1U << pos); |
8dd5516b JJ |
1239 | } |
1240 | ||
1241 | /* Mark the whole store S_INFO as unneeded. */ | |
1242 | ||
1243 | static inline void | |
fd6daec9 | 1244 | set_all_positions_unneeded (store_info *s_info) |
8dd5516b JJ |
1245 | { |
1246 | if (__builtin_expect (s_info->is_large, false)) | |
1247 | { | |
02ce5d90 RS |
1248 | HOST_WIDE_INT width; |
1249 | if (s_info->width.is_constant (&width)) | |
1250 | { | |
1251 | bitmap_set_range (s_info->positions_needed.large.bmap, 0, width); | |
1252 | s_info->positions_needed.large.count = width; | |
1253 | } | |
1254 | else | |
1255 | { | |
1256 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1257 | s_info->positions_needed.large.count = 1; | |
1258 | } | |
8dd5516b JJ |
1259 | } |
1260 | else | |
07e96250 | 1261 | s_info->positions_needed.small_bitmask = HOST_WIDE_INT_0U; |
8dd5516b JJ |
1262 | } |
1263 | ||
1264 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1265 | ||
1266 | static inline bool | |
fd6daec9 | 1267 | any_positions_needed_p (store_info *s_info) |
8dd5516b JJ |
1268 | { |
1269 | if (__builtin_expect (s_info->is_large, false)) | |
02ce5d90 RS |
1270 | { |
1271 | HOST_WIDE_INT width; | |
1272 | if (s_info->width.is_constant (&width)) | |
1273 | { | |
1274 | gcc_checking_assert (s_info->positions_needed.large.bmap); | |
1275 | return s_info->positions_needed.large.count < width; | |
1276 | } | |
1277 | else | |
1278 | { | |
1279 | gcc_checking_assert (!s_info->positions_needed.large.bmap); | |
1280 | return s_info->positions_needed.large.count == 0; | |
1281 | } | |
1282 | } | |
8dd5516b | 1283 | else |
07e96250 | 1284 | return (s_info->positions_needed.small_bitmask != HOST_WIDE_INT_0U); |
8dd5516b JJ |
1285 | } |
1286 | ||
1287 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
02ce5d90 | 1288 | store are known to be needed. */ |
8dd5516b JJ |
1289 | |
1290 | static inline bool | |
02ce5d90 RS |
1291 | all_positions_needed_p (store_info *s_info, poly_int64 start, |
1292 | poly_int64 width) | |
8dd5516b | 1293 | { |
02ce5d90 RS |
1294 | gcc_assert (s_info->rhs); |
1295 | if (!s_info->width.is_constant ()) | |
1296 | { | |
1297 | gcc_assert (s_info->is_large | |
1298 | && !s_info->positions_needed.large.bmap); | |
1299 | return s_info->positions_needed.large.count == 0; | |
1300 | } | |
1301 | ||
1302 | /* Otherwise, if START and WIDTH are non-constant, we're asking about | |
1303 | a non-constant region of a constant-sized store. We can't say for | |
1304 | sure that all positions are needed. */ | |
1305 | HOST_WIDE_INT const_start, const_width; | |
1306 | if (!start.is_constant (&const_start) | |
1307 | || !width.is_constant (&const_width)) | |
1308 | return false; | |
1309 | ||
8dd5516b JJ |
1310 | if (__builtin_expect (s_info->is_large, false)) |
1311 | { | |
02ce5d90 RS |
1312 | for (HOST_WIDE_INT i = const_start; i < const_start + const_width; ++i) |
1313 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, i)) | |
8dd5516b JJ |
1314 | return false; |
1315 | return true; | |
1316 | } | |
1317 | else | |
1318 | { | |
02ce5d90 RS |
1319 | unsigned HOST_WIDE_INT mask |
1320 | = lowpart_bitmask (const_width) << const_start; | |
8dd5516b JJ |
1321 | return (s_info->positions_needed.small_bitmask & mask) == mask; |
1322 | } | |
1323 | } | |
1324 | ||
1325 | ||
02ce5d90 RS |
1326 | static rtx get_stored_val (store_info *, machine_mode, poly_int64, |
1327 | poly_int64, basic_block, bool); | |
8dd5516b JJ |
1328 | |
1329 | ||
6fb5fa3c DB |
1330 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1331 | there is a candidate store, after adding it to the appropriate | |
1332 | local store group if so. */ | |
1333 | ||
1334 | static int | |
1335 | record_store (rtx body, bb_info_t bb_info) | |
1336 | { | |
6216f94e | 1337 | rtx mem, rhs, const_rhs, mem_addr; |
02ce5d90 RS |
1338 | poly_int64 offset = 0; |
1339 | poly_int64 width = 0; | |
6fb5fa3c | 1340 | insn_info_t insn_info = bb_info->last_insn; |
fd6daec9 | 1341 | store_info *store_info = NULL; |
6fb5fa3c DB |
1342 | int group_id; |
1343 | cselib_val *base = NULL; | |
8dd5516b | 1344 | insn_info_t ptr, last, redundant_reason; |
6fb5fa3c DB |
1345 | bool store_is_unused; |
1346 | ||
1347 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1348 | return 0; | |
1349 | ||
8dd5516b JJ |
1350 | mem = SET_DEST (body); |
1351 | ||
6fb5fa3c DB |
1352 | /* If this is not used, then this cannot be used to keep the insn |
1353 | from being deleted. On the other hand, it does provide something | |
1354 | that can be used to prove that another store is dead. */ | |
1355 | store_is_unused | |
8dd5516b | 1356 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
6fb5fa3c DB |
1357 | |
1358 | /* Check whether that value is a suitable memory location. */ | |
6fb5fa3c DB |
1359 | if (!MEM_P (mem)) |
1360 | { | |
1361 | /* If the set or clobber is unused, then it does not effect our | |
1362 | ability to get rid of the entire insn. */ | |
1363 | if (!store_is_unused) | |
1364 | insn_info->cannot_delete = true; | |
1365 | return 0; | |
1366 | } | |
1367 | ||
1368 | /* At this point we know mem is a mem. */ | |
1369 | if (GET_MODE (mem) == BLKmode) | |
1370 | { | |
d05d7551 | 1371 | HOST_WIDE_INT const_size; |
6fb5fa3c DB |
1372 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) |
1373 | { | |
456610d3 | 1374 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1375 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1376 | add_wild_read (bb_info); | |
1377 | insn_info->cannot_delete = true; | |
8dd5516b | 1378 | return 0; |
6fb5fa3c | 1379 | } |
8dd5516b JJ |
1380 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1381 | as memset (addr, 0, 36); */ | |
f5541398 | 1382 | else if (!MEM_SIZE_KNOWN_P (mem) |
d05d7551 RS |
1383 | || maybe_le (MEM_SIZE (mem), 0) |
1384 | /* This is a limit on the bitmap size, which is only relevant | |
1385 | for constant-sized MEMs. */ | |
1386 | || (MEM_SIZE (mem).is_constant (&const_size) | |
1387 | && const_size > MAX_OFFSET) | |
8dd5516b | 1388 | || GET_CODE (body) != SET |
8dd5516b | 1389 | || !CONST_INT_P (SET_SRC (body))) |
6fb5fa3c | 1390 | { |
8dd5516b JJ |
1391 | if (!store_is_unused) |
1392 | { | |
1393 | /* If the set or clobber is unused, then it does not effect our | |
1394 | ability to get rid of the entire insn. */ | |
1395 | insn_info->cannot_delete = true; | |
1396 | clear_rhs_from_active_local_stores (); | |
1397 | } | |
1398 | return 0; | |
6fb5fa3c | 1399 | } |
6fb5fa3c DB |
1400 | } |
1401 | ||
1402 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1403 | if (MEM_VOLATILE_P (mem)) | |
8dd5516b | 1404 | insn_info->cannot_delete = true; |
6fb5fa3c | 1405 | |
ac6929b5 | 1406 | if (!canon_address (mem, &group_id, &offset, &base)) |
6fb5fa3c DB |
1407 | { |
1408 | clear_rhs_from_active_local_stores (); | |
1409 | return 0; | |
1410 | } | |
1411 | ||
8dd5516b | 1412 | if (GET_MODE (mem) == BLKmode) |
f5541398 | 1413 | width = MEM_SIZE (mem); |
8dd5516b | 1414 | else |
6f557e0e | 1415 | width = GET_MODE_SIZE (GET_MODE (mem)); |
6fb5fa3c | 1416 | |
02ce5d90 | 1417 | if (!endpoint_representable_p (offset, width)) |
213ffde2 JJ |
1418 | { |
1419 | clear_rhs_from_active_local_stores (); | |
1420 | return 0; | |
1421 | } | |
1422 | ||
cbe679a4 JJ |
1423 | if (known_eq (width, 0)) |
1424 | return 0; | |
1425 | ||
ac6929b5 | 1426 | if (group_id >= 0) |
6fb5fa3c DB |
1427 | { |
1428 | /* In the restrictive case where the base is a constant or the | |
1429 | frame pointer we can do global analysis. */ | |
b8698a0f | 1430 | |
fd6daec9 | 1431 | group_info *group |
9771b263 | 1432 | = rtx_group_vec[group_id]; |
d26c7090 | 1433 | tree expr = MEM_EXPR (mem); |
b8698a0f | 1434 | |
da6603c6 | 1435 | store_info = rtx_store_info_pool.allocate (); |
d26c7090 | 1436 | set_usage_bits (group, offset, width, expr); |
6fb5fa3c | 1437 | |
456610d3 | 1438 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1439 | { |
1440 | fprintf (dump_file, " processing const base store gid=%d", | |
1441 | group_id); | |
1442 | print_range (dump_file, offset, width); | |
1443 | fprintf (dump_file, "\n"); | |
1444 | } | |
6fb5fa3c DB |
1445 | } |
1446 | else | |
1447 | { | |
9e412ca3 | 1448 | if (may_be_sp_based_p (XEXP (mem, 0))) |
50f0f366 | 1449 | insn_info->stack_pointer_based = true; |
6fb5fa3c | 1450 | insn_info->contains_cselib_groups = true; |
50f0f366 | 1451 | |
da6603c6 | 1452 | store_info = cse_store_info_pool.allocate (); |
6fb5fa3c DB |
1453 | group_id = -1; |
1454 | ||
456610d3 | 1455 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1456 | { |
1457 | fprintf (dump_file, " processing cselib store "); | |
1458 | print_range (dump_file, offset, width); | |
1459 | fprintf (dump_file, "\n"); | |
1460 | } | |
6fb5fa3c DB |
1461 | } |
1462 | ||
8dd5516b JJ |
1463 | const_rhs = rhs = NULL_RTX; |
1464 | if (GET_CODE (body) == SET | |
1465 | /* No place to keep the value after ra. */ | |
1466 | && !reload_completed | |
1467 | && (REG_P (SET_SRC (body)) | |
1468 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1469 | || CONSTANT_P (SET_SRC (body))) | |
1470 | && !MEM_VOLATILE_P (mem) | |
1471 | /* Sometimes the store and reload is used for truncation and | |
1472 | rounding. */ | |
1473 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1474 | { | |
1475 | rhs = SET_SRC (body); | |
1476 | if (CONSTANT_P (rhs)) | |
1477 | const_rhs = rhs; | |
1478 | else if (body == PATTERN (insn_info->insn)) | |
1479 | { | |
1480 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1481 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1482 | const_rhs = XEXP (tem, 0); | |
1483 | } | |
1484 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1485 | { | |
1486 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1487 | ||
1488 | if (tem && CONSTANT_P (tem)) | |
1489 | const_rhs = tem; | |
1490 | } | |
1491 | } | |
1492 | ||
6fb5fa3c DB |
1493 | /* Check to see if this stores causes some other stores to be |
1494 | dead. */ | |
1495 | ptr = active_local_stores; | |
1496 | last = NULL; | |
8dd5516b | 1497 | redundant_reason = NULL; |
6216f94e | 1498 | mem = canon_rtx (mem); |
ac6929b5 RB |
1499 | |
1500 | if (group_id < 0) | |
1501 | mem_addr = base->val_rtx; | |
6216f94e JJ |
1502 | else |
1503 | { | |
ac6929b5 RB |
1504 | group_info *group = rtx_group_vec[group_id]; |
1505 | mem_addr = group->canon_base_addr; | |
6216f94e | 1506 | } |
02ce5d90 | 1507 | if (maybe_ne (offset, 0)) |
ac6929b5 | 1508 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6fb5fa3c DB |
1509 | |
1510 | while (ptr) | |
1511 | { | |
1512 | insn_info_t next = ptr->next_local_store; | |
99b1c316 | 1513 | class store_info *s_info = ptr->store_rec; |
60564289 | 1514 | bool del = true; |
6fb5fa3c DB |
1515 | |
1516 | /* Skip the clobbers. We delete the active insn if this insn | |
6ed3da00 | 1517 | shadows the set. To have been put on the active list, it |
6fb5fa3c DB |
1518 | has exactly on set. */ |
1519 | while (!s_info->is_set) | |
1520 | s_info = s_info->next; | |
1521 | ||
ac6929b5 | 1522 | if (s_info->group_id == group_id && s_info->cse_base == base) |
6fb5fa3c DB |
1523 | { |
1524 | HOST_WIDE_INT i; | |
456610d3 | 1525 | if (dump_file && (dump_flags & TDF_DETAILS)) |
71d6a386 RS |
1526 | { |
1527 | fprintf (dump_file, " trying store in insn=%d gid=%d", | |
1528 | INSN_UID (ptr->insn), s_info->group_id); | |
1529 | print_range (dump_file, s_info->offset, s_info->width); | |
1530 | fprintf (dump_file, "\n"); | |
1531 | } | |
8dd5516b JJ |
1532 | |
1533 | /* Even if PTR won't be eliminated as unneeded, if both | |
1534 | PTR and this insn store the same constant value, we might | |
1535 | eliminate this insn instead. */ | |
1536 | if (s_info->const_rhs | |
1537 | && const_rhs | |
71d6a386 RS |
1538 | && known_subrange_p (offset, width, |
1539 | s_info->offset, s_info->width) | |
1540 | && all_positions_needed_p (s_info, offset - s_info->offset, | |
d7e1f499 RB |
1541 | width) |
1542 | /* We can only remove the later store if the earlier aliases | |
1543 | at least all accesses the later one. */ | |
1544 | && (MEM_ALIAS_SET (mem) == MEM_ALIAS_SET (s_info->mem) | |
1545 | || alias_set_subset_of (MEM_ALIAS_SET (mem), | |
1546 | MEM_ALIAS_SET (s_info->mem)))) | |
8dd5516b JJ |
1547 | { |
1548 | if (GET_MODE (mem) == BLKmode) | |
1549 | { | |
1550 | if (GET_MODE (s_info->mem) == BLKmode | |
1551 | && s_info->const_rhs == const_rhs) | |
1552 | redundant_reason = ptr; | |
1553 | } | |
1554 | else if (s_info->const_rhs == const0_rtx | |
1555 | && const_rhs == const0_rtx) | |
1556 | redundant_reason = ptr; | |
1557 | else | |
1558 | { | |
1559 | rtx val; | |
1560 | start_sequence (); | |
71d6a386 | 1561 | val = get_stored_val (s_info, GET_MODE (mem), offset, width, |
8dd5516b JJ |
1562 | BLOCK_FOR_INSN (insn_info->insn), |
1563 | true); | |
1564 | if (get_insns () != NULL) | |
1565 | val = NULL_RTX; | |
1566 | end_sequence (); | |
1567 | if (val && rtx_equal_p (val, const_rhs)) | |
1568 | redundant_reason = ptr; | |
1569 | } | |
1570 | } | |
1571 | ||
02ce5d90 | 1572 | HOST_WIDE_INT begin_unneeded, const_s_width, const_width; |
71d6a386 RS |
1573 | if (known_subrange_p (s_info->offset, s_info->width, offset, width)) |
1574 | /* The new store touches every byte that S_INFO does. */ | |
1575 | set_all_positions_unneeded (s_info); | |
02ce5d90 RS |
1576 | else if ((offset - s_info->offset).is_constant (&begin_unneeded) |
1577 | && s_info->width.is_constant (&const_s_width) | |
1578 | && width.is_constant (&const_width)) | |
71d6a386 | 1579 | { |
02ce5d90 | 1580 | HOST_WIDE_INT end_unneeded = begin_unneeded + const_width; |
71d6a386 | 1581 | begin_unneeded = MAX (begin_unneeded, 0); |
02ce5d90 | 1582 | end_unneeded = MIN (end_unneeded, const_s_width); |
71d6a386 RS |
1583 | for (i = begin_unneeded; i < end_unneeded; ++i) |
1584 | set_position_unneeded (s_info, i); | |
1585 | } | |
02ce5d90 RS |
1586 | else |
1587 | { | |
1588 | /* We don't know which parts of S_INFO are needed and | |
1589 | which aren't, so invalidate the RHS. */ | |
1590 | s_info->rhs = NULL; | |
1591 | s_info->const_rhs = NULL; | |
1592 | } | |
6fb5fa3c DB |
1593 | } |
1594 | else if (s_info->rhs) | |
1595 | /* Need to see if it is possible for this store to overwrite | |
1596 | the value of store_info. If it is, set the rhs to NULL to | |
1597 | keep it from being used to remove a load. */ | |
1598 | { | |
43b9f499 RB |
1599 | if (canon_output_dependence (s_info->mem, true, |
1600 | mem, GET_MODE (mem), | |
1601 | mem_addr)) | |
8dd5516b JJ |
1602 | { |
1603 | s_info->rhs = NULL; | |
1604 | s_info->const_rhs = NULL; | |
1605 | } | |
6fb5fa3c | 1606 | } |
6216f94e | 1607 | |
6fb5fa3c DB |
1608 | /* An insn can be deleted if every position of every one of |
1609 | its s_infos is zero. */ | |
1b6fa860 | 1610 | if (any_positions_needed_p (s_info)) |
60564289 | 1611 | del = false; |
8dd5516b | 1612 | |
60564289 | 1613 | if (del) |
6fb5fa3c DB |
1614 | { |
1615 | insn_info_t insn_to_delete = ptr; | |
b8698a0f | 1616 | |
dabd47e7 | 1617 | active_local_stores_len--; |
6fb5fa3c DB |
1618 | if (last) |
1619 | last->next_local_store = ptr->next_local_store; | |
1620 | else | |
1621 | active_local_stores = ptr->next_local_store; | |
b8698a0f | 1622 | |
1b6fa860 JJ |
1623 | if (!insn_to_delete->cannot_delete) |
1624 | delete_dead_store_insn (insn_to_delete); | |
6fb5fa3c DB |
1625 | } |
1626 | else | |
1627 | last = ptr; | |
b8698a0f | 1628 | |
6fb5fa3c DB |
1629 | ptr = next; |
1630 | } | |
b8698a0f | 1631 | |
6fb5fa3c DB |
1632 | /* Finish filling in the store_info. */ |
1633 | store_info->next = insn_info->store_rec; | |
1634 | insn_info->store_rec = store_info; | |
6216f94e | 1635 | store_info->mem = mem; |
6216f94e | 1636 | store_info->mem_addr = mem_addr; |
6fb5fa3c | 1637 | store_info->cse_base = base; |
02ce5d90 RS |
1638 | HOST_WIDE_INT const_width; |
1639 | if (!width.is_constant (&const_width)) | |
1640 | { | |
1641 | store_info->is_large = true; | |
1642 | store_info->positions_needed.large.count = 0; | |
1643 | store_info->positions_needed.large.bmap = NULL; | |
1644 | } | |
1645 | else if (const_width > HOST_BITS_PER_WIDE_INT) | |
8dd5516b JJ |
1646 | { |
1647 | store_info->is_large = true; | |
1648 | store_info->positions_needed.large.count = 0; | |
3f9b14ff | 1649 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
8dd5516b JJ |
1650 | } |
1651 | else | |
1652 | { | |
1653 | store_info->is_large = false; | |
02ce5d90 RS |
1654 | store_info->positions_needed.small_bitmask |
1655 | = lowpart_bitmask (const_width); | |
8dd5516b | 1656 | } |
6fb5fa3c | 1657 | store_info->group_id = group_id; |
71d6a386 RS |
1658 | store_info->offset = offset; |
1659 | store_info->width = width; | |
6fb5fa3c | 1660 | store_info->is_set = GET_CODE (body) == SET; |
8dd5516b JJ |
1661 | store_info->rhs = rhs; |
1662 | store_info->const_rhs = const_rhs; | |
1663 | store_info->redundant_reason = redundant_reason; | |
6fb5fa3c | 1664 | |
6fb5fa3c DB |
1665 | /* If this is a clobber, we return 0. We will only be able to |
1666 | delete this insn if there is only one store USED store, but we | |
1667 | can use the clobber to delete other stores earlier. */ | |
1668 | return store_info->is_set ? 1 : 0; | |
1669 | } | |
1670 | ||
1671 | ||
1672 | static void | |
1673 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1674 | { | |
b8698a0f | 1675 | fprintf (dump_file, "%s insn=%d %s\n", start, |
6fb5fa3c DB |
1676 | INSN_UID (insn_info->insn), |
1677 | insn_info->store_rec ? "has store" : "naked"); | |
1678 | } | |
1679 | ||
1680 | ||
8660aaae EC |
1681 | /* If the modes are different and the value's source and target do not |
1682 | line up, we need to extract the value from lower part of the rhs of | |
1683 | the store, shift it, and then put it into a form that can be shoved | |
1684 | into the read_insn. This function generates a right SHIFT of a | |
1685 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1686 | shift sequence is returned or NULL if we failed to find a | |
1687 | shift. */ | |
1688 | ||
1689 | static rtx | |
02ce5d90 | 1690 | find_shift_sequence (poly_int64 access_size, |
fd6daec9 | 1691 | store_info *store_info, |
ef4bddc2 | 1692 | machine_mode read_mode, |
02ce5d90 | 1693 | poly_int64 shift, bool speed, bool require_cst) |
8660aaae | 1694 | { |
ef4bddc2 | 1695 | machine_mode store_mode = GET_MODE (store_info->mem); |
f67f4dff | 1696 | scalar_int_mode new_mode; |
18b526e8 | 1697 | rtx read_reg = NULL; |
8660aaae EC |
1698 | |
1699 | /* Some machines like the x86 have shift insns for each size of | |
1700 | operand. Other machines like the ppc or the ia-64 may only have | |
1701 | shift insns that shift values within 32 or 64 bit registers. | |
1702 | This loop tries to find the smallest shift insn that will right | |
1703 | justify the value we want to read but is available in one insn on | |
1704 | the machine. */ | |
1705 | ||
f67f4dff RS |
1706 | opt_scalar_int_mode new_mode_iter; |
1707 | FOR_EACH_MODE_FROM (new_mode_iter, | |
1708 | smallest_int_mode_for_size (access_size * BITS_PER_UNIT)) | |
8660aaae | 1709 | { |
dc01c3d1 DM |
1710 | rtx target, new_reg, new_lhs; |
1711 | rtx_insn *shift_seq, *insn; | |
d898d29b | 1712 | int cost; |
348eea5f | 1713 | |
f67f4dff | 1714 | new_mode = new_mode_iter.require (); |
c94843d2 RS |
1715 | if (GET_MODE_BITSIZE (new_mode) > BITS_PER_WORD) |
1716 | break; | |
1717 | ||
72a2609f JJ |
1718 | /* If a constant was stored into memory, try to simplify it here, |
1719 | otherwise the cost of the shift might preclude this optimization | |
1720 | e.g. at -Os, even when no actual shift will be needed. */ | |
8dd5516b | 1721 | if (store_info->const_rhs) |
72a2609f | 1722 | { |
91914e56 | 1723 | poly_uint64 byte = subreg_lowpart_offset (new_mode, store_mode); |
8dd5516b JJ |
1724 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1725 | store_mode, byte); | |
72a2609f JJ |
1726 | if (ret && CONSTANT_P (ret)) |
1727 | { | |
abd3c800 | 1728 | rtx shift_rtx = gen_int_shift_amount (new_mode, shift); |
72a2609f | 1729 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, |
abd3c800 | 1730 | ret, shift_rtx); |
72a2609f JJ |
1731 | if (ret && CONSTANT_P (ret)) |
1732 | { | |
1733 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1734 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1735 | if (ret && CONSTANT_P (ret) | |
e548c9df AM |
1736 | && (set_src_cost (ret, read_mode, speed) |
1737 | <= COSTS_N_INSNS (1))) | |
72a2609f JJ |
1738 | return ret; |
1739 | } | |
1740 | } | |
1741 | } | |
1742 | ||
8dd5516b JJ |
1743 | if (require_cst) |
1744 | return NULL_RTX; | |
1745 | ||
18b526e8 RS |
1746 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1747 | requires a real instruction. */ | |
73a699ae | 1748 | if (maybe_lt (GET_MODE_SIZE (new_mode), GET_MODE_SIZE (store_mode)) |
d0edd768 | 1749 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
348eea5f RS |
1750 | continue; |
1751 | ||
18b526e8 RS |
1752 | /* Also try a wider mode if the necessary punning is either not |
1753 | desirable or not possible. */ | |
1754 | if (!CONSTANT_P (store_info->rhs) | |
99e1629f | 1755 | && !targetm.modes_tieable_p (new_mode, store_mode)) |
18b526e8 | 1756 | continue; |
18b526e8 | 1757 | |
348eea5f | 1758 | new_reg = gen_reg_rtx (new_mode); |
8660aaae EC |
1759 | |
1760 | start_sequence (); | |
1761 | ||
1762 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1763 | of one dsp where the cost of these two was not the same. But | |
1764 | this really is a rare case anyway. */ | |
1765 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
abd3c800 RS |
1766 | gen_int_shift_amount (new_mode, shift), |
1767 | new_reg, 1, OPTAB_DIRECT); | |
8660aaae | 1768 | |
c6f3019a RS |
1769 | shift_seq = get_insns (); |
1770 | end_sequence (); | |
8660aaae | 1771 | |
c6f3019a RS |
1772 | if (target != new_reg || shift_seq == NULL) |
1773 | continue; | |
1774 | ||
1775 | cost = 0; | |
1776 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1777 | if (INSN_P (insn)) | |
d09e78e8 | 1778 | cost += insn_cost (insn, speed); |
c6f3019a RS |
1779 | |
1780 | /* The computation up to here is essentially independent | |
1781 | of the arguments and could be precomputed. It may | |
1782 | not be worth doing so. We could precompute if | |
1783 | worthwhile or at least cache the results. The result | |
06acf7d0 RS |
1784 | technically depends on both SHIFT and ACCESS_SIZE, |
1785 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
c6f3019a RS |
1786 | |
1787 | if (cost > COSTS_N_INSNS (1)) | |
1788 | continue; | |
1789 | ||
d898d29b JJ |
1790 | new_lhs = extract_low_bits (new_mode, store_mode, |
1791 | copy_rtx (store_info->rhs)); | |
1792 | if (new_lhs == NULL_RTX) | |
1793 | continue; | |
1794 | ||
c6f3019a RS |
1795 | /* We found an acceptable shift. Generate a move to |
1796 | take the value from the store and put it into the | |
1797 | shift pseudo, then shift it, then generate another | |
1798 | move to put in into the target of the read. */ | |
18b526e8 | 1799 | emit_move_insn (new_reg, new_lhs); |
c6f3019a | 1800 | emit_insn (shift_seq); |
18b526e8 | 1801 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
c6f3019a | 1802 | break; |
8660aaae EC |
1803 | } |
1804 | ||
18b526e8 | 1805 | return read_reg; |
8660aaae EC |
1806 | } |
1807 | ||
1808 | ||
02b47899 KZ |
1809 | /* Call back for note_stores to find the hard regs set or clobbered by |
1810 | insn. Data is a bitmap of the hardregs set so far. */ | |
1811 | ||
1812 | static void | |
1813 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1814 | { | |
1815 | bitmap regs_set = (bitmap) data; | |
1816 | ||
1817 | if (REG_P (x) | |
f773c2bd | 1818 | && HARD_REGISTER_P (x)) |
dc8afb70 | 1819 | bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); |
02b47899 KZ |
1820 | } |
1821 | ||
8dd5516b | 1822 | /* Helper function for replace_read and record_store. |
71d6a386 RS |
1823 | Attempt to return a value of mode READ_MODE stored in STORE_INFO, |
1824 | consisting of READ_WIDTH bytes starting from READ_OFFSET. Return NULL | |
8dd5516b JJ |
1825 | if not successful. If REQUIRE_CST is true, return always constant. */ |
1826 | ||
1827 | static rtx | |
fd6daec9 | 1828 | get_stored_val (store_info *store_info, machine_mode read_mode, |
02ce5d90 | 1829 | poly_int64 read_offset, poly_int64 read_width, |
8dd5516b JJ |
1830 | basic_block bb, bool require_cst) |
1831 | { | |
ef4bddc2 | 1832 | machine_mode store_mode = GET_MODE (store_info->mem); |
02ce5d90 | 1833 | poly_int64 gap; |
8dd5516b JJ |
1834 | rtx read_reg; |
1835 | ||
1836 | /* To get here the read is within the boundaries of the write so | |
1837 | shift will never be negative. Start out with the shift being in | |
1838 | bytes. */ | |
1839 | if (store_mode == BLKmode) | |
71d6a386 | 1840 | gap = 0; |
8dd5516b | 1841 | else if (BYTES_BIG_ENDIAN) |
71d6a386 RS |
1842 | gap = ((store_info->offset + store_info->width) |
1843 | - (read_offset + read_width)); | |
8dd5516b | 1844 | else |
71d6a386 | 1845 | gap = read_offset - store_info->offset; |
8dd5516b | 1846 | |
e6575643 | 1847 | if (gap.is_constant () && maybe_ne (gap, 0)) |
71d6a386 | 1848 | { |
02ce5d90 RS |
1849 | poly_int64 shift = gap * BITS_PER_UNIT; |
1850 | poly_int64 access_size = GET_MODE_SIZE (read_mode) + gap; | |
71d6a386 RS |
1851 | read_reg = find_shift_sequence (access_size, store_info, read_mode, |
1852 | shift, optimize_bb_for_speed_p (bb), | |
1853 | require_cst); | |
1854 | } | |
8dd5516b JJ |
1855 | else if (store_mode == BLKmode) |
1856 | { | |
1857 | /* The store is a memset (addr, const_val, const_size). */ | |
1858 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
304b9962 RS |
1859 | scalar_int_mode int_store_mode; |
1860 | if (!int_mode_for_mode (read_mode).exists (&int_store_mode)) | |
8dd5516b JJ |
1861 | read_reg = NULL_RTX; |
1862 | else if (store_info->rhs == const0_rtx) | |
304b9962 RS |
1863 | read_reg = extract_low_bits (read_mode, int_store_mode, const0_rtx); |
1864 | else if (GET_MODE_BITSIZE (int_store_mode) > HOST_BITS_PER_WIDE_INT | |
8dd5516b JJ |
1865 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) |
1866 | read_reg = NULL_RTX; | |
1867 | else | |
1868 | { | |
1869 | unsigned HOST_WIDE_INT c | |
1870 | = INTVAL (store_info->rhs) | |
fecfbfa4 | 1871 | & ((HOST_WIDE_INT_1 << BITS_PER_UNIT) - 1); |
8dd5516b JJ |
1872 | int shift = BITS_PER_UNIT; |
1873 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1874 | { | |
1875 | c |= (c << shift); | |
1876 | shift <<= 1; | |
1877 | } | |
304b9962 RS |
1878 | read_reg = gen_int_mode (c, int_store_mode); |
1879 | read_reg = extract_low_bits (read_mode, int_store_mode, read_reg); | |
8dd5516b JJ |
1880 | } |
1881 | } | |
1882 | else if (store_info->const_rhs | |
1883 | && (require_cst | |
1884 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1885 | read_reg = extract_low_bits (read_mode, store_mode, | |
1886 | copy_rtx (store_info->const_rhs)); | |
1887 | else | |
1888 | read_reg = extract_low_bits (read_mode, store_mode, | |
1889 | copy_rtx (store_info->rhs)); | |
1890 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1891 | read_reg = NULL_RTX; | |
1892 | return read_reg; | |
1893 | } | |
02b47899 | 1894 | |
6fb5fa3c DB |
1895 | /* Take a sequence of: |
1896 | A <- r1 | |
1897 | ... | |
1898 | ... <- A | |
1899 | ||
b8698a0f | 1900 | and change it into |
6fb5fa3c DB |
1901 | r2 <- r1 |
1902 | A <- r1 | |
1903 | ... | |
1904 | ... <- r2 | |
1905 | ||
8660aaae EC |
1906 | or |
1907 | ||
1908 | r3 <- extract (r1) | |
1909 | r3 <- r3 >> shift | |
1910 | r2 <- extract (r3) | |
1911 | ... <- r2 | |
1912 | ||
1913 | or | |
1914 | ||
1915 | r2 <- extract (r1) | |
1916 | ... <- r2 | |
1917 | ||
1918 | Depending on the alignment and the mode of the store and | |
1919 | subsequent load. | |
1920 | ||
1921 | ||
1922 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
6fb5fa3c DB |
1923 | and READ_INSN are for the read. Return true if the replacement |
1924 | went ok. */ | |
1925 | ||
1926 | static bool | |
fd6daec9 | 1927 | replace_read (store_info *store_info, insn_info_t store_insn, |
8dd5516b JJ |
1928 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
1929 | bitmap regs_live) | |
6fb5fa3c | 1930 | { |
ef4bddc2 RS |
1931 | machine_mode store_mode = GET_MODE (store_info->mem); |
1932 | machine_mode read_mode = GET_MODE (read_info->mem); | |
dc01c3d1 DM |
1933 | rtx_insn *insns, *this_insn; |
1934 | rtx read_reg; | |
8dd5516b | 1935 | basic_block bb; |
8660aaae | 1936 | |
6fb5fa3c DB |
1937 | if (!dbg_cnt (dse)) |
1938 | return false; | |
1939 | ||
18b526e8 RS |
1940 | /* Create a sequence of instructions to set up the read register. |
1941 | This sequence goes immediately before the store and its result | |
1942 | is read by the load. | |
1943 | ||
1944 | We need to keep this in perspective. We are replacing a read | |
8660aaae EC |
1945 | with a sequence of insns, but the read will almost certainly be |
1946 | in cache, so it is not going to be an expensive one. Thus, we | |
1947 | are not willing to do a multi insn shift or worse a subroutine | |
1948 | call to get rid of the read. */ | |
456610d3 | 1949 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
1950 | fprintf (dump_file, "trying to replace %smode load in insn %d" |
1951 | " from %smode store in insn %d\n", | |
1952 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
1953 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
1954 | start_sequence (); | |
8dd5516b JJ |
1955 | bb = BLOCK_FOR_INSN (read_insn->insn); |
1956 | read_reg = get_stored_val (store_info, | |
71d6a386 | 1957 | read_mode, read_info->offset, read_info->width, |
8dd5516b | 1958 | bb, false); |
18b526e8 | 1959 | if (read_reg == NULL_RTX) |
8660aaae | 1960 | { |
18b526e8 | 1961 | end_sequence (); |
456610d3 | 1962 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
1963 | fprintf (dump_file, " -- could not extract bits of stored value\n"); |
1964 | return false; | |
8660aaae | 1965 | } |
18b526e8 RS |
1966 | /* Force the value into a new register so that it won't be clobbered |
1967 | between the store and the load. */ | |
1968 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
1969 | insns = get_insns (); | |
1970 | end_sequence (); | |
8660aaae | 1971 | |
02b47899 KZ |
1972 | if (insns != NULL_RTX) |
1973 | { | |
1974 | /* Now we have to scan the set of new instructions to see if the | |
1975 | sequence contains and sets of hardregs that happened to be | |
1976 | live at this point. For instance, this can happen if one of | |
1977 | the insns sets the CC and the CC happened to be live at that | |
1978 | point. This does occasionally happen, see PR 37922. */ | |
3f9b14ff | 1979 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
02b47899 KZ |
1980 | |
1981 | for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
1982 | note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
b8698a0f | 1983 | |
02b47899 KZ |
1984 | bitmap_and_into (regs_set, regs_live); |
1985 | if (!bitmap_empty_p (regs_set)) | |
1986 | { | |
456610d3 | 1987 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02b47899 | 1988 | { |
b8698a0f | 1989 | fprintf (dump_file, |
02b47899 KZ |
1990 | "abandoning replacement because sequence clobbers live hardregs:"); |
1991 | df_print_regset (dump_file, regs_set); | |
1992 | } | |
b8698a0f | 1993 | |
02b47899 KZ |
1994 | BITMAP_FREE (regs_set); |
1995 | return false; | |
1996 | } | |
1997 | BITMAP_FREE (regs_set); | |
1998 | } | |
1999 | ||
8660aaae | 2000 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
6fb5fa3c | 2001 | { |
fd6daec9 | 2002 | deferred_change *change = deferred_change_pool.allocate (); |
b8698a0f | 2003 | |
8660aaae EC |
2004 | /* Insert this right before the store insn where it will be safe |
2005 | from later insns that might change it before the read. */ | |
2006 | emit_insn_before (insns, store_insn->insn); | |
b8698a0f | 2007 | |
8660aaae EC |
2008 | /* And now for the kludge part: cselib croaks if you just |
2009 | return at this point. There are two reasons for this: | |
b8698a0f | 2010 | |
8660aaae EC |
2011 | 1) Cselib has an idea of how many pseudos there are and |
2012 | that does not include the new ones we just added. | |
b8698a0f | 2013 | |
8660aaae EC |
2014 | 2) Cselib does not know about the move insn we added |
2015 | above the store_info, and there is no way to tell it | |
2016 | about it, because it has "moved on". | |
b8698a0f | 2017 | |
8660aaae EC |
2018 | Problem (1) is fixable with a certain amount of engineering. |
2019 | Problem (2) is requires starting the bb from scratch. This | |
2020 | could be expensive. | |
b8698a0f | 2021 | |
8660aaae EC |
2022 | So we are just going to have to lie. The move/extraction |
2023 | insns are not really an issue, cselib did not see them. But | |
2024 | the use of the new pseudo read_insn is a real problem because | |
2025 | cselib has not scanned this insn. The way that we solve this | |
2026 | problem is that we are just going to put the mem back for now | |
2027 | and when we are finished with the block, we undo this. We | |
2028 | keep a table of mems to get rid of. At the end of the basic | |
2029 | block we can put them back. */ | |
b8698a0f | 2030 | |
8660aaae | 2031 | *loc = read_info->mem; |
da6603c6 ML |
2032 | change->next = deferred_change_list; |
2033 | deferred_change_list = change; | |
2034 | change->loc = loc; | |
2035 | change->reg = read_reg; | |
b8698a0f | 2036 | |
8660aaae EC |
2037 | /* Get rid of the read_info, from the point of view of the |
2038 | rest of dse, play like this read never happened. */ | |
2039 | read_insn->read_rec = read_info->next; | |
fb0b2914 | 2040 | read_info_type_pool.remove (read_info); |
456610d3 | 2041 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2042 | { |
2043 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2044 | print_simple_rtl (dump_file, read_reg); | |
2045 | fprintf (dump_file, "\n"); | |
2046 | } | |
8660aaae | 2047 | return true; |
6fb5fa3c | 2048 | } |
b8698a0f | 2049 | else |
6fb5fa3c | 2050 | { |
456610d3 | 2051 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2052 | { |
2053 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2054 | print_simple_rtl (dump_file, read_reg); | |
2055 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2056 | } | |
6fb5fa3c DB |
2057 | return false; |
2058 | } | |
2059 | } | |
2060 | ||
d7111da8 RS |
2061 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2062 | be active. */ | |
6fb5fa3c | 2063 | |
d7111da8 RS |
2064 | static void |
2065 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info) | |
6fb5fa3c | 2066 | { |
6216f94e | 2067 | rtx mem = *loc, mem_addr; |
6fb5fa3c | 2068 | insn_info_t insn_info; |
02ce5d90 RS |
2069 | poly_int64 offset = 0; |
2070 | poly_int64 width = 0; | |
b8698a0f | 2071 | cselib_val *base = NULL; |
6fb5fa3c DB |
2072 | int group_id; |
2073 | read_info_t read_info; | |
2074 | ||
6fb5fa3c DB |
2075 | insn_info = bb_info->last_insn; |
2076 | ||
2077 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
a49a975f | 2078 | || MEM_VOLATILE_P (mem)) |
6fb5fa3c | 2079 | { |
a49a975f JJ |
2080 | if (crtl->stack_protect_guard |
2081 | && (MEM_EXPR (mem) == crtl->stack_protect_guard | |
2082 | || (crtl->stack_protect_guard_decl | |
2083 | && MEM_EXPR (mem) == crtl->stack_protect_guard_decl)) | |
2084 | && MEM_VOLATILE_P (mem)) | |
2085 | { | |
2086 | /* This is either the stack protector canary on the stack, | |
2087 | which ought to be written by a MEM_VOLATILE_P store and | |
2088 | thus shouldn't be deleted and is read at the very end of | |
2089 | function, but shouldn't conflict with any other store. | |
2090 | Or it is __stack_chk_guard variable or TLS or whatever else | |
2091 | MEM holding the canary value, which really shouldn't be | |
2092 | ever modified in -fstack-protector* protected functions, | |
2093 | otherwise the prologue store wouldn't match the epilogue | |
2094 | check. */ | |
2095 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2096 | fprintf (dump_file, " stack protector canary read ignored.\n"); | |
2097 | insn_info->cannot_delete = true; | |
2098 | return; | |
2099 | } | |
2100 | ||
456610d3 | 2101 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2102 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
2103 | add_wild_read (bb_info); | |
2104 | insn_info->cannot_delete = true; | |
d7111da8 | 2105 | return; |
6fb5fa3c DB |
2106 | } |
2107 | ||
2108 | /* If it is reading readonly mem, then there can be no conflict with | |
2109 | another write. */ | |
2110 | if (MEM_READONLY_P (mem)) | |
d7111da8 | 2111 | return; |
6fb5fa3c | 2112 | |
ac6929b5 | 2113 | if (!canon_address (mem, &group_id, &offset, &base)) |
6fb5fa3c | 2114 | { |
456610d3 | 2115 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2116 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
2117 | add_wild_read (bb_info); | |
d7111da8 | 2118 | return; |
6fb5fa3c DB |
2119 | } |
2120 | ||
2121 | if (GET_MODE (mem) == BLKmode) | |
2122 | width = -1; | |
2123 | else | |
2124 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2125 | ||
02ce5d90 | 2126 | if (!endpoint_representable_p (offset, known_eq (width, -1) ? 1 : width)) |
938f9248 | 2127 | { |
213ffde2 JJ |
2128 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2129 | fprintf (dump_file, " adding wild read, due to overflow.\n"); | |
2130 | add_wild_read (bb_info); | |
938f9248 ML |
2131 | return; |
2132 | } | |
2133 | ||
fb0b2914 | 2134 | read_info = read_info_type_pool.allocate (); |
6fb5fa3c DB |
2135 | read_info->group_id = group_id; |
2136 | read_info->mem = mem; | |
71d6a386 RS |
2137 | read_info->offset = offset; |
2138 | read_info->width = width; | |
6fb5fa3c DB |
2139 | read_info->next = insn_info->read_rec; |
2140 | insn_info->read_rec = read_info; | |
ac6929b5 RB |
2141 | if (group_id < 0) |
2142 | mem_addr = base->val_rtx; | |
6216f94e JJ |
2143 | else |
2144 | { | |
ac6929b5 RB |
2145 | group_info *group = rtx_group_vec[group_id]; |
2146 | mem_addr = group->canon_base_addr; | |
6216f94e | 2147 | } |
02ce5d90 | 2148 | if (maybe_ne (offset, 0)) |
ac6929b5 | 2149 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6fb5fa3c | 2150 | |
ac6929b5 | 2151 | if (group_id >= 0) |
6fb5fa3c DB |
2152 | { |
2153 | /* This is the restricted case where the base is a constant or | |
2154 | the frame pointer and offset is a constant. */ | |
2155 | insn_info_t i_ptr = active_local_stores; | |
2156 | insn_info_t last = NULL; | |
b8698a0f | 2157 | |
456610d3 | 2158 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2159 | { |
02ce5d90 | 2160 | if (!known_size_p (width)) |
6fb5fa3c DB |
2161 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", |
2162 | group_id); | |
2163 | else | |
71d6a386 RS |
2164 | { |
2165 | fprintf (dump_file, " processing const load gid=%d", group_id); | |
2166 | print_range (dump_file, offset, width); | |
2167 | fprintf (dump_file, "\n"); | |
2168 | } | |
6fb5fa3c DB |
2169 | } |
2170 | ||
2171 | while (i_ptr) | |
2172 | { | |
2173 | bool remove = false; | |
fd6daec9 | 2174 | store_info *store_info = i_ptr->store_rec; |
b8698a0f | 2175 | |
6fb5fa3c DB |
2176 | /* Skip the clobbers. */ |
2177 | while (!store_info->is_set) | |
2178 | store_info = store_info->next; | |
b8698a0f | 2179 | |
6fb5fa3c DB |
2180 | /* There are three cases here. */ |
2181 | if (store_info->group_id < 0) | |
2182 | /* We have a cselib store followed by a read from a | |
2183 | const base. */ | |
b8698a0f L |
2184 | remove |
2185 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2186 | GET_MODE (store_info->mem), |
2187 | store_info->mem_addr, | |
53d9622b | 2188 | mem, mem_addr); |
b8698a0f | 2189 | |
6fb5fa3c DB |
2190 | else if (group_id == store_info->group_id) |
2191 | { | |
2192 | /* This is a block mode load. We may get lucky and | |
2193 | canon_true_dependence may save the day. */ | |
02ce5d90 | 2194 | if (!known_size_p (width)) |
b8698a0f L |
2195 | remove |
2196 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2197 | GET_MODE (store_info->mem), |
2198 | store_info->mem_addr, | |
53d9622b | 2199 | mem, mem_addr); |
b8698a0f | 2200 | |
6fb5fa3c DB |
2201 | /* If this read is just reading back something that we just |
2202 | stored, rewrite the read. */ | |
b8698a0f | 2203 | else |
6fb5fa3c DB |
2204 | { |
2205 | if (store_info->rhs | |
71d6a386 RS |
2206 | && known_subrange_p (offset, width, store_info->offset, |
2207 | store_info->width) | |
8dd5516b | 2208 | && all_positions_needed_p (store_info, |
71d6a386 | 2209 | offset - store_info->offset, |
8dd5516b JJ |
2210 | width) |
2211 | && replace_read (store_info, i_ptr, read_info, | |
2212 | insn_info, loc, bb_info->regs_live)) | |
d7111da8 | 2213 | return; |
8dd5516b | 2214 | |
6fb5fa3c | 2215 | /* The bases are the same, just see if the offsets |
71d6a386 RS |
2216 | could overlap. */ |
2217 | if (ranges_maybe_overlap_p (offset, width, | |
2218 | store_info->offset, | |
2219 | store_info->width)) | |
6fb5fa3c DB |
2220 | remove = true; |
2221 | } | |
2222 | } | |
b8698a0f L |
2223 | |
2224 | /* else | |
6fb5fa3c DB |
2225 | The else case that is missing here is that the |
2226 | bases are constant but different. There is nothing | |
2227 | to do here because there is no overlap. */ | |
b8698a0f | 2228 | |
6fb5fa3c DB |
2229 | if (remove) |
2230 | { | |
456610d3 | 2231 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2232 | dump_insn_info ("removing from active", i_ptr); |
2233 | ||
dabd47e7 | 2234 | active_local_stores_len--; |
6fb5fa3c DB |
2235 | if (last) |
2236 | last->next_local_store = i_ptr->next_local_store; | |
2237 | else | |
2238 | active_local_stores = i_ptr->next_local_store; | |
2239 | } | |
2240 | else | |
2241 | last = i_ptr; | |
2242 | i_ptr = i_ptr->next_local_store; | |
2243 | } | |
2244 | } | |
b8698a0f | 2245 | else |
6fb5fa3c DB |
2246 | { |
2247 | insn_info_t i_ptr = active_local_stores; | |
2248 | insn_info_t last = NULL; | |
456610d3 | 2249 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2250 | { |
2251 | fprintf (dump_file, " processing cselib load mem:"); | |
2252 | print_inline_rtx (dump_file, mem, 0); | |
2253 | fprintf (dump_file, "\n"); | |
2254 | } | |
2255 | ||
2256 | while (i_ptr) | |
2257 | { | |
2258 | bool remove = false; | |
fd6daec9 | 2259 | store_info *store_info = i_ptr->store_rec; |
b8698a0f | 2260 | |
456610d3 | 2261 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2262 | fprintf (dump_file, " processing cselib load against insn %d\n", |
2263 | INSN_UID (i_ptr->insn)); | |
2264 | ||
2265 | /* Skip the clobbers. */ | |
2266 | while (!store_info->is_set) | |
2267 | store_info = store_info->next; | |
2268 | ||
2269 | /* If this read is just reading back something that we just | |
2270 | stored, rewrite the read. */ | |
2271 | if (store_info->rhs | |
2272 | && store_info->group_id == -1 | |
2273 | && store_info->cse_base == base | |
71d6a386 RS |
2274 | && known_subrange_p (offset, width, store_info->offset, |
2275 | store_info->width) | |
8dd5516b | 2276 | && all_positions_needed_p (store_info, |
71d6a386 | 2277 | offset - store_info->offset, width) |
8dd5516b JJ |
2278 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, |
2279 | bb_info->regs_live)) | |
d7111da8 | 2280 | return; |
6fb5fa3c | 2281 | |
ac6929b5 RB |
2282 | remove = canon_true_dependence (store_info->mem, |
2283 | GET_MODE (store_info->mem), | |
2284 | store_info->mem_addr, | |
2285 | mem, mem_addr); | |
b8698a0f | 2286 | |
6fb5fa3c DB |
2287 | if (remove) |
2288 | { | |
456610d3 | 2289 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2290 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2291 | |
dabd47e7 | 2292 | active_local_stores_len--; |
6fb5fa3c DB |
2293 | if (last) |
2294 | last->next_local_store = i_ptr->next_local_store; | |
2295 | else | |
2296 | active_local_stores = i_ptr->next_local_store; | |
2297 | } | |
2298 | else | |
2299 | last = i_ptr; | |
2300 | i_ptr = i_ptr->next_local_store; | |
2301 | } | |
2302 | } | |
6fb5fa3c DB |
2303 | } |
2304 | ||
d7111da8 | 2305 | /* A note_uses callback in which DATA points the INSN_INFO for |
6fb5fa3c DB |
2306 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2307 | true for any part of *LOC. */ | |
2308 | ||
2309 | static void | |
2310 | check_mem_read_use (rtx *loc, void *data) | |
2311 | { | |
d7111da8 RS |
2312 | subrtx_ptr_iterator::array_type array; |
2313 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
2314 | { | |
2315 | rtx *loc = *iter; | |
2316 | if (MEM_P (*loc)) | |
2317 | check_mem_read_rtx (loc, (bb_info_t) data); | |
2318 | } | |
6fb5fa3c DB |
2319 | } |
2320 | ||
8dd5516b JJ |
2321 | |
2322 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2323 | So far it only handles arguments passed in registers. */ | |
2324 | ||
2325 | static bool | |
2326 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2327 | { | |
d5cc9181 JR |
2328 | CUMULATIVE_ARGS args_so_far_v; |
2329 | cumulative_args_t args_so_far; | |
8dd5516b JJ |
2330 | tree arg; |
2331 | int idx; | |
2332 | ||
d5cc9181 JR |
2333 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2334 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
8dd5516b JJ |
2335 | |
2336 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2337 | for (idx = 0; | |
2338 | arg != void_list_node && idx < nargs; | |
2339 | arg = TREE_CHAIN (arg), idx++) | |
2340 | { | |
b4206259 | 2341 | scalar_int_mode mode; |
3c07301f | 2342 | rtx reg, link, tmp; |
b4206259 RS |
2343 | |
2344 | if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg)), &mode)) | |
2345 | return false; | |
2346 | ||
6783fdb7 RS |
2347 | function_arg_info arg (mode, /*named=*/true); |
2348 | reg = targetm.calls.function_arg (args_so_far, arg); | |
b4206259 | 2349 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode) |
8dd5516b JJ |
2350 | return false; |
2351 | ||
2352 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2353 | link; | |
2354 | link = XEXP (link, 1)) | |
2355 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2356 | { | |
b4206259 | 2357 | scalar_int_mode arg_mode; |
8dd5516b JJ |
2358 | args[idx] = XEXP (XEXP (link, 0), 0); |
2359 | if (REG_P (args[idx]) | |
2360 | && REGNO (args[idx]) == REGNO (reg) | |
2361 | && (GET_MODE (args[idx]) == mode | |
b4206259 RS |
2362 | || (is_int_mode (GET_MODE (args[idx]), &arg_mode) |
2363 | && (GET_MODE_SIZE (arg_mode) <= UNITS_PER_WORD) | |
2364 | && (GET_MODE_SIZE (arg_mode) > GET_MODE_SIZE (mode))))) | |
8dd5516b JJ |
2365 | break; |
2366 | } | |
2367 | if (!link) | |
2368 | return false; | |
2369 | ||
2370 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2371 | if (GET_MODE (args[idx]) != mode) | |
2372 | { | |
2373 | if (!tmp || !CONST_INT_P (tmp)) | |
2374 | return false; | |
6d26322f | 2375 | tmp = gen_int_mode (INTVAL (tmp), mode); |
8dd5516b JJ |
2376 | } |
2377 | if (tmp) | |
2378 | args[idx] = tmp; | |
2379 | ||
6930c98c | 2380 | targetm.calls.function_arg_advance (args_so_far, arg); |
8dd5516b JJ |
2381 | } |
2382 | if (arg != void_list_node || idx != nargs) | |
2383 | return false; | |
2384 | return true; | |
2385 | } | |
2386 | ||
9e582b1d JR |
2387 | /* Return a bitmap of the fixed registers contained in IN. */ |
2388 | ||
2389 | static bitmap | |
2390 | copy_fixed_regs (const_bitmap in) | |
2391 | { | |
2392 | bitmap ret; | |
2393 | ||
2394 | ret = ALLOC_REG_SET (NULL); | |
2395 | bitmap_and (ret, in, fixed_reg_set_regset); | |
2396 | return ret; | |
2397 | } | |
8dd5516b | 2398 | |
6fb5fa3c DB |
2399 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2400 | if some part of it is not a candidate store and assigns to a | |
2401 | non-register target. */ | |
2402 | ||
2403 | static void | |
dd60a84c | 2404 | scan_insn (bb_info_t bb_info, rtx_insn *insn) |
6fb5fa3c DB |
2405 | { |
2406 | rtx body; | |
fb0b2914 | 2407 | insn_info_type *insn_info = insn_info_type_pool.allocate (); |
6fb5fa3c | 2408 | int mems_found = 0; |
da6603c6 | 2409 | memset (insn_info, 0, sizeof (struct insn_info_type)); |
6fb5fa3c | 2410 | |
456610d3 | 2411 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2412 | fprintf (dump_file, "\n**scanning insn=%d\n", |
2413 | INSN_UID (insn)); | |
2414 | ||
2415 | insn_info->prev_insn = bb_info->last_insn; | |
2416 | insn_info->insn = insn; | |
2417 | bb_info->last_insn = insn_info; | |
b8698a0f | 2418 | |
b5b8b0ac AO |
2419 | if (DEBUG_INSN_P (insn)) |
2420 | { | |
2421 | insn_info->cannot_delete = true; | |
2422 | return; | |
2423 | } | |
6fb5fa3c | 2424 | |
6fb5fa3c DB |
2425 | /* Look at all of the uses in the insn. */ |
2426 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2427 | ||
2428 | if (CALL_P (insn)) | |
2429 | { | |
8dd5516b | 2430 | bool const_call; |
ee516de9 | 2431 | rtx call, sym; |
8dd5516b JJ |
2432 | tree memset_call = NULL_TREE; |
2433 | ||
6fb5fa3c | 2434 | insn_info->cannot_delete = true; |
50f0f366 | 2435 | |
6fb5fa3c | 2436 | /* Const functions cannot do anything bad i.e. read memory, |
50f0f366 | 2437 | however, they can read their parameters which may have |
8dd5516b JJ |
2438 | been pushed onto the stack. |
2439 | memset and bzero don't read memory either. */ | |
2440 | const_call = RTL_CONST_CALL_P (insn); | |
ee516de9 EB |
2441 | if (!const_call |
2442 | && (call = get_call_rtx_from (insn)) | |
2443 | && (sym = XEXP (XEXP (call, 0), 0)) | |
2444 | && GET_CODE (sym) == SYMBOL_REF | |
2445 | && SYMBOL_REF_DECL (sym) | |
2446 | && TREE_CODE (SYMBOL_REF_DECL (sym)) == FUNCTION_DECL | |
3d78e008 | 2447 | && fndecl_built_in_p (SYMBOL_REF_DECL (sym), BUILT_IN_MEMSET)) |
ee516de9 EB |
2448 | memset_call = SYMBOL_REF_DECL (sym); |
2449 | ||
8dd5516b | 2450 | if (const_call || memset_call) |
6fb5fa3c DB |
2451 | { |
2452 | insn_info_t i_ptr = active_local_stores; | |
2453 | insn_info_t last = NULL; | |
2454 | ||
456610d3 | 2455 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2456 | fprintf (dump_file, "%s call %d\n", |
2457 | const_call ? "const" : "memset", INSN_UID (insn)); | |
6fb5fa3c | 2458 | |
64520bdc | 2459 | /* See the head comment of the frame_read field. */ |
57534689 JJ |
2460 | if (reload_completed |
2461 | /* Tail calls are storing their arguments using | |
2462 | arg pointer. If it is a frame pointer on the target, | |
2463 | even before reload we need to kill frame pointer based | |
2464 | stores. */ | |
2465 | || (SIBLING_CALL_P (insn) | |
2466 | && HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
64520bdc EB |
2467 | insn_info->frame_read = true; |
2468 | ||
2469 | /* Loop over the active stores and remove those which are | |
2470 | killed by the const function call. */ | |
6fb5fa3c DB |
2471 | while (i_ptr) |
2472 | { | |
64520bdc EB |
2473 | bool remove_store = false; |
2474 | ||
2475 | /* The stack pointer based stores are always killed. */ | |
50f0f366 | 2476 | if (i_ptr->stack_pointer_based) |
64520bdc EB |
2477 | remove_store = true; |
2478 | ||
2479 | /* If the frame is read, the frame related stores are killed. */ | |
2480 | else if (insn_info->frame_read) | |
2481 | { | |
fd6daec9 | 2482 | store_info *store_info = i_ptr->store_rec; |
64520bdc EB |
2483 | |
2484 | /* Skip the clobbers. */ | |
2485 | while (!store_info->is_set) | |
2486 | store_info = store_info->next; | |
2487 | ||
2488 | if (store_info->group_id >= 0 | |
9771b263 | 2489 | && rtx_group_vec[store_info->group_id]->frame_related) |
64520bdc EB |
2490 | remove_store = true; |
2491 | } | |
2492 | ||
2493 | if (remove_store) | |
6fb5fa3c | 2494 | { |
456610d3 | 2495 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2496 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2497 | |
dabd47e7 | 2498 | active_local_stores_len--; |
6fb5fa3c DB |
2499 | if (last) |
2500 | last->next_local_store = i_ptr->next_local_store; | |
2501 | else | |
2502 | active_local_stores = i_ptr->next_local_store; | |
2503 | } | |
2504 | else | |
2505 | last = i_ptr; | |
64520bdc | 2506 | |
6fb5fa3c DB |
2507 | i_ptr = i_ptr->next_local_store; |
2508 | } | |
8dd5516b JJ |
2509 | |
2510 | if (memset_call) | |
2511 | { | |
2512 | rtx args[3]; | |
2513 | if (get_call_args (insn, memset_call, args, 3) | |
2514 | && CONST_INT_P (args[1]) | |
2515 | && CONST_INT_P (args[2]) | |
2516 | && INTVAL (args[2]) > 0) | |
2517 | { | |
2518 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
f5541398 | 2519 | set_mem_size (mem, INTVAL (args[2])); |
f7df4a84 | 2520 | body = gen_rtx_SET (mem, args[1]); |
8dd5516b | 2521 | mems_found += record_store (body, bb_info); |
456610d3 | 2522 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2523 | fprintf (dump_file, "handling memset as BLKmode store\n"); |
2524 | if (mems_found == 1) | |
2525 | { | |
dabd47e7 JJ |
2526 | if (active_local_stores_len++ |
2527 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2528 | { | |
2529 | active_local_stores_len = 1; | |
2530 | active_local_stores = NULL; | |
2531 | } | |
9e582b1d JR |
2532 | insn_info->fixed_regs_live |
2533 | = copy_fixed_regs (bb_info->regs_live); | |
8dd5516b JJ |
2534 | insn_info->next_local_store = active_local_stores; |
2535 | active_local_stores = insn_info; | |
2536 | } | |
2537 | } | |
a2ad54a0 JJ |
2538 | else |
2539 | clear_rhs_from_active_local_stores (); | |
8dd5516b | 2540 | } |
6fb5fa3c | 2541 | } |
7ee98586 EB |
2542 | else if (SIBLING_CALL_P (insn) |
2543 | && (reload_completed || HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
57534689 JJ |
2544 | /* Arguments for a sibling call that are pushed to memory are passed |
2545 | using the incoming argument pointer of the current function. After | |
7ee98586 EB |
2546 | reload that might be (and likely is) frame pointer based. And, if |
2547 | it is a frame pointer on the target, even before reload we need to | |
2548 | kill frame pointer based stores. */ | |
57534689 | 2549 | add_wild_read (bb_info); |
50f0f366 | 2550 | else |
d26c7090 ER |
2551 | /* Every other call, including pure functions, may read any memory |
2552 | that is not relative to the frame. */ | |
2553 | add_non_frame_wild_read (bb_info); | |
50f0f366 | 2554 | |
6fb5fa3c DB |
2555 | return; |
2556 | } | |
2557 | ||
2558 | /* Assuming that there are sets in these insns, we cannot delete | |
2559 | them. */ | |
2560 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
0a64eeca | 2561 | || volatile_refs_p (PATTERN (insn)) |
2da02156 | 2562 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
6fb5fa3c DB |
2563 | || (RTX_FRAME_RELATED_P (insn)) |
2564 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2565 | insn_info->cannot_delete = true; | |
b8698a0f | 2566 | |
6fb5fa3c DB |
2567 | body = PATTERN (insn); |
2568 | if (GET_CODE (body) == PARALLEL) | |
2569 | { | |
2570 | int i; | |
2571 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2572 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2573 | } | |
2574 | else | |
2575 | mems_found += record_store (body, bb_info); | |
2576 | ||
456610d3 | 2577 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2578 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
6fb5fa3c DB |
2579 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2580 | ||
8dd5516b JJ |
2581 | /* If we found some sets of mems, add it into the active_local_stores so |
2582 | that it can be locally deleted if found dead or used for | |
2583 | replace_read and redundant constant store elimination. Otherwise mark | |
2584 | it as cannot delete. This simplifies the processing later. */ | |
2585 | if (mems_found == 1) | |
6fb5fa3c | 2586 | { |
dabd47e7 JJ |
2587 | if (active_local_stores_len++ |
2588 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2589 | { | |
2590 | active_local_stores_len = 1; | |
2591 | active_local_stores = NULL; | |
2592 | } | |
9e582b1d | 2593 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
6fb5fa3c DB |
2594 | insn_info->next_local_store = active_local_stores; |
2595 | active_local_stores = insn_info; | |
2596 | } | |
2597 | else | |
2598 | insn_info->cannot_delete = true; | |
2599 | } | |
2600 | ||
2601 | ||
2602 | /* Remove BASE from the set of active_local_stores. This is a | |
2603 | callback from cselib that is used to get rid of the stores in | |
2604 | active_local_stores. */ | |
2605 | ||
2606 | static void | |
2607 | remove_useless_values (cselib_val *base) | |
2608 | { | |
2609 | insn_info_t insn_info = active_local_stores; | |
2610 | insn_info_t last = NULL; | |
2611 | ||
2612 | while (insn_info) | |
2613 | { | |
fd6daec9 | 2614 | store_info *store_info = insn_info->store_rec; |
60564289 | 2615 | bool del = false; |
6fb5fa3c DB |
2616 | |
2617 | /* If ANY of the store_infos match the cselib group that is | |
67914693 | 2618 | being deleted, then the insn cannot be deleted. */ |
6fb5fa3c DB |
2619 | while (store_info) |
2620 | { | |
b8698a0f | 2621 | if ((store_info->group_id == -1) |
6fb5fa3c DB |
2622 | && (store_info->cse_base == base)) |
2623 | { | |
60564289 | 2624 | del = true; |
6fb5fa3c DB |
2625 | break; |
2626 | } | |
2627 | store_info = store_info->next; | |
2628 | } | |
2629 | ||
60564289 | 2630 | if (del) |
6fb5fa3c | 2631 | { |
dabd47e7 | 2632 | active_local_stores_len--; |
6fb5fa3c DB |
2633 | if (last) |
2634 | last->next_local_store = insn_info->next_local_store; | |
2635 | else | |
2636 | active_local_stores = insn_info->next_local_store; | |
2637 | free_store_info (insn_info); | |
2638 | } | |
2639 | else | |
2640 | last = insn_info; | |
b8698a0f | 2641 | |
6fb5fa3c DB |
2642 | insn_info = insn_info->next_local_store; |
2643 | } | |
2644 | } | |
2645 | ||
2646 | ||
2647 | /* Do all of step 1. */ | |
2648 | ||
2649 | static void | |
2650 | dse_step1 (void) | |
2651 | { | |
2652 | basic_block bb; | |
3f9b14ff | 2653 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
b8698a0f | 2654 | |
457eeaae | 2655 | cselib_init (0); |
6fb5fa3c DB |
2656 | all_blocks = BITMAP_ALLOC (NULL); |
2657 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2658 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2659 | ||
04a90bec | 2660 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
2661 | { |
2662 | insn_info_t ptr; | |
fb0b2914 | 2663 | bb_info_t bb_info = dse_bb_info_type_pool.allocate (); |
6fb5fa3c | 2664 | |
da6603c6 | 2665 | memset (bb_info, 0, sizeof (dse_bb_info_type)); |
6fb5fa3c | 2666 | bitmap_set_bit (all_blocks, bb->index); |
02b47899 KZ |
2667 | bb_info->regs_live = regs_live; |
2668 | ||
2669 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2670 | df_simulate_initialize_forwards (bb, regs_live); | |
6fb5fa3c DB |
2671 | |
2672 | bb_table[bb->index] = bb_info; | |
2673 | cselib_discard_hook = remove_useless_values; | |
2674 | ||
2675 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2676 | { | |
dd60a84c | 2677 | rtx_insn *insn; |
6fb5fa3c | 2678 | |
6fb5fa3c | 2679 | active_local_stores = NULL; |
dabd47e7 | 2680 | active_local_stores_len = 0; |
6fb5fa3c | 2681 | cselib_clear_table (); |
b8698a0f | 2682 | |
6fb5fa3c DB |
2683 | /* Scan the insns. */ |
2684 | FOR_BB_INSNS (bb, insn) | |
2685 | { | |
2686 | if (INSN_P (insn)) | |
2687 | scan_insn (bb_info, insn); | |
2688 | cselib_process_insn (insn); | |
02b47899 KZ |
2689 | if (INSN_P (insn)) |
2690 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
6fb5fa3c | 2691 | } |
b8698a0f | 2692 | |
6fb5fa3c DB |
2693 | /* This is something of a hack, because the global algorithm |
2694 | is supposed to take care of the case where stores go dead | |
2695 | at the end of the function. However, the global | |
2696 | algorithm must take a more conservative view of block | |
2697 | mode reads than the local alg does. So to get the case | |
2698 | where you have a store to the frame followed by a non | |
0d52bcc1 | 2699 | overlapping block more read, we look at the active local |
6fb5fa3c DB |
2700 | stores at the end of the function and delete all of the |
2701 | frame and spill based ones. */ | |
2702 | if (stores_off_frame_dead_at_return | |
2703 | && (EDGE_COUNT (bb->succs) == 0 | |
2704 | || (single_succ_p (bb) | |
fefa31b5 | 2705 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
e3b5732b | 2706 | && ! crtl->calls_eh_return))) |
6fb5fa3c DB |
2707 | { |
2708 | insn_info_t i_ptr = active_local_stores; | |
2709 | while (i_ptr) | |
2710 | { | |
fd6daec9 | 2711 | store_info *store_info = i_ptr->store_rec; |
6fb5fa3c DB |
2712 | |
2713 | /* Skip the clobbers. */ | |
2714 | while (!store_info->is_set) | |
2715 | store_info = store_info->next; | |
ac6929b5 RB |
2716 | if (store_info->group_id >= 0) |
2717 | { | |
2718 | group_info *group = rtx_group_vec[store_info->group_id]; | |
2719 | if (group->frame_related && !i_ptr->cannot_delete) | |
2720 | delete_dead_store_insn (i_ptr); | |
2721 | } | |
6fb5fa3c DB |
2722 | |
2723 | i_ptr = i_ptr->next_local_store; | |
2724 | } | |
2725 | } | |
2726 | ||
2727 | /* Get rid of the loads that were discovered in | |
2728 | replace_read. Cselib is finished with this block. */ | |
2729 | while (deferred_change_list) | |
2730 | { | |
fd6daec9 | 2731 | deferred_change *next = deferred_change_list->next; |
6fb5fa3c DB |
2732 | |
2733 | /* There is no reason to validate this change. That was | |
2734 | done earlier. */ | |
2735 | *deferred_change_list->loc = deferred_change_list->reg; | |
fb0b2914 | 2736 | deferred_change_pool.remove (deferred_change_list); |
6fb5fa3c DB |
2737 | deferred_change_list = next; |
2738 | } | |
2739 | ||
2740 | /* Get rid of all of the cselib based store_infos in this | |
2741 | block and mark the containing insns as not being | |
2742 | deletable. */ | |
2743 | ptr = bb_info->last_insn; | |
2744 | while (ptr) | |
2745 | { | |
2746 | if (ptr->contains_cselib_groups) | |
8dd5516b | 2747 | { |
fd6daec9 | 2748 | store_info *s_info = ptr->store_rec; |
8dd5516b JJ |
2749 | while (s_info && !s_info->is_set) |
2750 | s_info = s_info->next; | |
2751 | if (s_info | |
2752 | && s_info->redundant_reason | |
2753 | && s_info->redundant_reason->insn | |
2754 | && !ptr->cannot_delete) | |
2755 | { | |
456610d3 | 2756 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2757 | fprintf (dump_file, "Locally deleting insn %d " |
2758 | "because insn %d stores the " | |
2759 | "same value and couldn't be " | |
2760 | "eliminated\n", | |
2761 | INSN_UID (ptr->insn), | |
2762 | INSN_UID (s_info->redundant_reason->insn)); | |
2763 | delete_dead_store_insn (ptr); | |
2764 | } | |
8dd5516b JJ |
2765 | free_store_info (ptr); |
2766 | } | |
2767 | else | |
2768 | { | |
fd6daec9 | 2769 | store_info *s_info; |
8dd5516b JJ |
2770 | |
2771 | /* Free at least positions_needed bitmaps. */ | |
2772 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2773 | if (s_info->is_large) | |
2774 | { | |
dc491a25 | 2775 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
8dd5516b JJ |
2776 | s_info->is_large = false; |
2777 | } | |
2778 | } | |
6fb5fa3c DB |
2779 | ptr = ptr->prev_insn; |
2780 | } | |
2781 | ||
da6603c6 | 2782 | cse_store_info_pool.release (); |
6fb5fa3c | 2783 | } |
02b47899 | 2784 | bb_info->regs_live = NULL; |
6fb5fa3c DB |
2785 | } |
2786 | ||
02b47899 | 2787 | BITMAP_FREE (regs_live); |
6fb5fa3c | 2788 | cselib_finish (); |
c203e8a7 | 2789 | rtx_group_table->empty (); |
6fb5fa3c DB |
2790 | } |
2791 | ||
2792 | \f | |
2793 | /*---------------------------------------------------------------------------- | |
2794 | Second step. | |
2795 | ||
2796 | Assign each byte position in the stores that we are going to | |
2797 | analyze globally to a position in the bitmaps. Returns true if | |
6ed3da00 | 2798 | there are any bit positions assigned. |
6fb5fa3c DB |
2799 | ----------------------------------------------------------------------------*/ |
2800 | ||
2801 | static void | |
2802 | dse_step2_init (void) | |
2803 | { | |
2804 | unsigned int i; | |
fd6daec9 | 2805 | group_info *group; |
6fb5fa3c | 2806 | |
9771b263 | 2807 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2808 | { |
2809 | /* For all non stack related bases, we only consider a store to | |
2810 | be deletable if there are two or more stores for that | |
2811 | position. This is because it takes one store to make the | |
2812 | other store redundant. However, for the stores that are | |
2813 | stack related, we consider them if there is only one store | |
2814 | for the position. We do this because the stack related | |
2815 | stores can be deleted if their is no read between them and | |
2816 | the end of the function. | |
b8698a0f | 2817 | |
6fb5fa3c DB |
2818 | To make this work in the current framework, we take the stack |
2819 | related bases add all of the bits from store1 into store2. | |
2820 | This has the effect of making the eligible even if there is | |
2821 | only one store. */ | |
2822 | ||
2823 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2824 | { | |
2825 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2826 | bitmap_ior_into (group->store2_p, group->store1_p); | |
456610d3 | 2827 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2828 | fprintf (dump_file, "group %d is frame related ", i); |
6fb5fa3c DB |
2829 | } |
2830 | ||
2831 | group->offset_map_size_n++; | |
3f9b14ff SB |
2832 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2833 | group->offset_map_size_n); | |
6fb5fa3c | 2834 | group->offset_map_size_p++; |
3f9b14ff SB |
2835 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2836 | group->offset_map_size_p); | |
6fb5fa3c | 2837 | group->process_globally = false; |
456610d3 | 2838 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2839 | { |
b8698a0f | 2840 | fprintf (dump_file, "group %d(%d+%d): ", i, |
6fb5fa3c DB |
2841 | (int)bitmap_count_bits (group->store2_n), |
2842 | (int)bitmap_count_bits (group->store2_p)); | |
2843 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2844 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2845 | } | |
2846 | } | |
2847 | } | |
2848 | ||
2849 | ||
ac6929b5 | 2850 | /* Init the offset tables. */ |
6fb5fa3c DB |
2851 | |
2852 | static bool | |
ac6929b5 | 2853 | dse_step2 (void) |
6fb5fa3c DB |
2854 | { |
2855 | unsigned int i; | |
fd6daec9 | 2856 | group_info *group; |
6fb5fa3c DB |
2857 | /* Position 0 is unused because 0 is used in the maps to mean |
2858 | unused. */ | |
2859 | current_position = 1; | |
9771b263 | 2860 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2861 | { |
2862 | bitmap_iterator bi; | |
2863 | unsigned int j; | |
2864 | ||
c3284718 RS |
2865 | memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
2866 | memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
6fb5fa3c DB |
2867 | bitmap_clear (group->group_kill); |
2868 | ||
2869 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2870 | { | |
2871 | bitmap_set_bit (group->group_kill, current_position); | |
d26c7090 ER |
2872 | if (bitmap_bit_p (group->escaped_n, j)) |
2873 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2874 | group->offset_map_n[j] = current_position++; |
2875 | group->process_globally = true; | |
2876 | } | |
2877 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2878 | { | |
b8698a0f | 2879 | bitmap_set_bit (group->group_kill, current_position); |
d26c7090 ER |
2880 | if (bitmap_bit_p (group->escaped_p, j)) |
2881 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2882 | group->offset_map_p[j] = current_position++; |
2883 | group->process_globally = true; | |
2884 | } | |
2885 | } | |
2886 | return current_position != 1; | |
2887 | } | |
2888 | ||
2889 | ||
6fb5fa3c DB |
2890 | \f |
2891 | /*---------------------------------------------------------------------------- | |
2892 | Third step. | |
b8698a0f | 2893 | |
6fb5fa3c DB |
2894 | Build the bit vectors for the transfer functions. |
2895 | ----------------------------------------------------------------------------*/ | |
2896 | ||
2897 | ||
6fb5fa3c DB |
2898 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
2899 | there, return 0. */ | |
2900 | ||
2901 | static int | |
fd6daec9 | 2902 | get_bitmap_index (group_info *group_info, HOST_WIDE_INT offset) |
6fb5fa3c DB |
2903 | { |
2904 | if (offset < 0) | |
2905 | { | |
2906 | HOST_WIDE_INT offset_p = -offset; | |
2907 | if (offset_p >= group_info->offset_map_size_n) | |
2908 | return 0; | |
2909 | return group_info->offset_map_n[offset_p]; | |
2910 | } | |
2911 | else | |
2912 | { | |
2913 | if (offset >= group_info->offset_map_size_p) | |
2914 | return 0; | |
2915 | return group_info->offset_map_p[offset]; | |
2916 | } | |
2917 | } | |
2918 | ||
2919 | ||
2920 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
2921 | may be NULL. */ | |
2922 | ||
b8698a0f | 2923 | static void |
ac6929b5 | 2924 | scan_stores (store_info *store_info, bitmap gen, bitmap kill) |
6fb5fa3c DB |
2925 | { |
2926 | while (store_info) | |
2927 | { | |
02ce5d90 | 2928 | HOST_WIDE_INT i, offset, width; |
fd6daec9 | 2929 | group_info *group_info |
9771b263 | 2930 | = rtx_group_vec[store_info->group_id]; |
02ce5d90 RS |
2931 | /* We can (conservatively) ignore stores whose bounds aren't known; |
2932 | they simply don't generate new global dse opportunities. */ | |
2933 | if (group_info->process_globally | |
2934 | && store_info->offset.is_constant (&offset) | |
2935 | && store_info->width.is_constant (&width)) | |
71d6a386 | 2936 | { |
02ce5d90 RS |
2937 | HOST_WIDE_INT end = offset + width; |
2938 | for (i = offset; i < end; i++) | |
71d6a386 RS |
2939 | { |
2940 | int index = get_bitmap_index (group_info, i); | |
2941 | if (index != 0) | |
2942 | { | |
2943 | bitmap_set_bit (gen, index); | |
2944 | if (kill) | |
2945 | bitmap_clear_bit (kill, index); | |
2946 | } | |
2947 | } | |
2948 | } | |
6fb5fa3c DB |
2949 | store_info = store_info->next; |
2950 | } | |
2951 | } | |
2952 | ||
2953 | ||
6fb5fa3c DB |
2954 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL |
2955 | may be NULL. */ | |
2956 | ||
2957 | static void | |
ac6929b5 | 2958 | scan_reads (insn_info_t insn_info, bitmap gen, bitmap kill) |
6fb5fa3c DB |
2959 | { |
2960 | read_info_t read_info = insn_info->read_rec; | |
2961 | int i; | |
fd6daec9 | 2962 | group_info *group; |
6fb5fa3c | 2963 | |
64520bdc EB |
2964 | /* If this insn reads the frame, kill all the frame related stores. */ |
2965 | if (insn_info->frame_read) | |
2966 | { | |
9771b263 | 2967 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
64520bdc EB |
2968 | if (group->process_globally && group->frame_related) |
2969 | { | |
2970 | if (kill) | |
2971 | bitmap_ior_into (kill, group->group_kill); | |
b8698a0f | 2972 | bitmap_and_compl_into (gen, group->group_kill); |
64520bdc EB |
2973 | } |
2974 | } | |
d26c7090 ER |
2975 | if (insn_info->non_frame_wild_read) |
2976 | { | |
2977 | /* Kill all non-frame related stores. Kill all stores of variables that | |
2978 | escape. */ | |
2979 | if (kill) | |
2980 | bitmap_ior_into (kill, kill_on_calls); | |
2981 | bitmap_and_compl_into (gen, kill_on_calls); | |
9771b263 | 2982 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
d26c7090 ER |
2983 | if (group->process_globally && !group->frame_related) |
2984 | { | |
2985 | if (kill) | |
2986 | bitmap_ior_into (kill, group->group_kill); | |
2987 | bitmap_and_compl_into (gen, group->group_kill); | |
2988 | } | |
2989 | } | |
6fb5fa3c DB |
2990 | while (read_info) |
2991 | { | |
9771b263 | 2992 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2993 | { |
2994 | if (group->process_globally) | |
2995 | { | |
2996 | if (i == read_info->group_id) | |
2997 | { | |
02ce5d90 RS |
2998 | HOST_WIDE_INT offset, width; |
2999 | /* Reads with non-constant size kill all DSE opportunities | |
3000 | in the group. */ | |
3001 | if (!read_info->offset.is_constant (&offset) | |
3002 | || !read_info->width.is_constant (&width) | |
3003 | || !known_size_p (width)) | |
6fb5fa3c | 3004 | { |
71d6a386 | 3005 | /* Handle block mode reads. */ |
6fb5fa3c DB |
3006 | if (kill) |
3007 | bitmap_ior_into (kill, group->group_kill); | |
3008 | bitmap_and_compl_into (gen, group->group_kill); | |
3009 | } | |
3010 | else | |
3011 | { | |
3012 | /* The groups are the same, just process the | |
3013 | offsets. */ | |
3014 | HOST_WIDE_INT j; | |
02ce5d90 RS |
3015 | HOST_WIDE_INT end = offset + width; |
3016 | for (j = offset; j < end; j++) | |
6fb5fa3c DB |
3017 | { |
3018 | int index = get_bitmap_index (group, j); | |
3019 | if (index != 0) | |
3020 | { | |
3021 | if (kill) | |
3022 | bitmap_set_bit (kill, index); | |
3023 | bitmap_clear_bit (gen, index); | |
3024 | } | |
3025 | } | |
3026 | } | |
3027 | } | |
3028 | else | |
3029 | { | |
3030 | /* The groups are different, if the alias sets | |
3031 | conflict, clear the entire group. We only need | |
3032 | to apply this test if the read_info is a cselib | |
3033 | read. Anything with a constant base cannot alias | |
3034 | something else with a different constant | |
3035 | base. */ | |
3036 | if ((read_info->group_id < 0) | |
b8698a0f | 3037 | && canon_true_dependence (group->base_mem, |
d32f725a | 3038 | GET_MODE (group->base_mem), |
6216f94e | 3039 | group->canon_base_addr, |
53d9622b | 3040 | read_info->mem, NULL_RTX)) |
6fb5fa3c DB |
3041 | { |
3042 | if (kill) | |
3043 | bitmap_ior_into (kill, group->group_kill); | |
3044 | bitmap_and_compl_into (gen, group->group_kill); | |
3045 | } | |
3046 | } | |
3047 | } | |
3048 | } | |
b8698a0f | 3049 | |
6fb5fa3c DB |
3050 | read_info = read_info->next; |
3051 | } | |
3052 | } | |
3053 | ||
6fb5fa3c DB |
3054 | |
3055 | /* Return the insn in BB_INFO before the first wild read or if there | |
3056 | are no wild reads in the block, return the last insn. */ | |
3057 | ||
3058 | static insn_info_t | |
3059 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3060 | { | |
3061 | insn_info_t insn_info = bb_info->last_insn; | |
3062 | insn_info_t last_wild_read = NULL; | |
3063 | ||
3064 | while (insn_info) | |
3065 | { | |
3066 | if (insn_info->wild_read) | |
3067 | { | |
3068 | last_wild_read = insn_info->prev_insn; | |
3069 | /* Block starts with wild read. */ | |
3070 | if (!last_wild_read) | |
3071 | return NULL; | |
3072 | } | |
3073 | ||
3074 | insn_info = insn_info->prev_insn; | |
3075 | } | |
3076 | ||
3077 | if (last_wild_read) | |
3078 | return last_wild_read; | |
3079 | else | |
3080 | return bb_info->last_insn; | |
3081 | } | |
3082 | ||
3083 | ||
3084 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3085 | the block in order to build the gen and kill sets for the block. | |
3086 | We start at ptr which may be the last insn in the block or may be | |
3087 | the first insn with a wild read. In the latter case we are able to | |
3088 | skip the rest of the block because it just does not matter: | |
3089 | anything that happens is hidden by the wild read. */ | |
3090 | ||
3091 | static void | |
ac6929b5 | 3092 | dse_step3_scan (basic_block bb) |
6fb5fa3c DB |
3093 | { |
3094 | bb_info_t bb_info = bb_table[bb->index]; | |
3095 | insn_info_t insn_info; | |
3096 | ||
ac6929b5 | 3097 | insn_info = find_insn_before_first_wild_read (bb_info); |
b8698a0f | 3098 | |
6fb5fa3c DB |
3099 | /* In the spill case or in the no_spill case if there is no wild |
3100 | read in the block, we will need a kill set. */ | |
3101 | if (insn_info == bb_info->last_insn) | |
3102 | { | |
3103 | if (bb_info->kill) | |
3104 | bitmap_clear (bb_info->kill); | |
3105 | else | |
3f9b14ff | 3106 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c | 3107 | } |
b8698a0f | 3108 | else |
6fb5fa3c DB |
3109 | if (bb_info->kill) |
3110 | BITMAP_FREE (bb_info->kill); | |
3111 | ||
3112 | while (insn_info) | |
3113 | { | |
3114 | /* There may have been code deleted by the dce pass run before | |
3115 | this phase. */ | |
3116 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3117 | { | |
ac6929b5 RB |
3118 | scan_stores (insn_info->store_rec, bb_info->gen, bb_info->kill); |
3119 | scan_reads (insn_info, bb_info->gen, bb_info->kill); | |
b8698a0f | 3120 | } |
6fb5fa3c DB |
3121 | |
3122 | insn_info = insn_info->prev_insn; | |
3123 | } | |
3124 | } | |
3125 | ||
3126 | ||
3127 | /* Set the gen set of the exit block, and also any block with no | |
3128 | successors that does not have a wild read. */ | |
3129 | ||
3130 | static void | |
3131 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3132 | { | |
3133 | /* The gen set is all 0's for the exit block except for the | |
3134 | frame_pointer_group. */ | |
b8698a0f | 3135 | |
6fb5fa3c DB |
3136 | if (stores_off_frame_dead_at_return) |
3137 | { | |
3138 | unsigned int i; | |
fd6daec9 | 3139 | group_info *group; |
b8698a0f | 3140 | |
9771b263 | 3141 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3142 | { |
3143 | if (group->process_globally && group->frame_related) | |
3144 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3145 | } | |
3146 | } | |
3147 | } | |
3148 | ||
3149 | ||
3150 | /* Find all of the blocks that are not backwards reachable from the | |
3151 | exit block or any block with no successors (BB). These are the | |
3152 | infinite loops or infinite self loops. These blocks will still | |
3153 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3154 | ||
3155 | static void | |
3156 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3157 | { | |
3158 | edge e; | |
3159 | edge_iterator ei; | |
3160 | ||
d7c028c0 | 3161 | if (bitmap_bit_p (unreachable_blocks, bb->index)) |
6fb5fa3c | 3162 | { |
d7c028c0 | 3163 | bitmap_clear_bit (unreachable_blocks, bb->index); |
6fb5fa3c | 3164 | FOR_EACH_EDGE (e, ei, bb->preds) |
b8698a0f | 3165 | { |
6fb5fa3c | 3166 | mark_reachable_blocks (unreachable_blocks, e->src); |
b8698a0f | 3167 | } |
6fb5fa3c DB |
3168 | } |
3169 | } | |
3170 | ||
3171 | /* Build the transfer functions for the function. */ | |
3172 | ||
3173 | static void | |
ac6929b5 | 3174 | dse_step3 () |
6fb5fa3c DB |
3175 | { |
3176 | basic_block bb; | |
6fb5fa3c DB |
3177 | sbitmap_iterator sbi; |
3178 | bitmap all_ones = NULL; | |
3179 | unsigned int i; | |
b8698a0f | 3180 | |
7ba9e72d | 3181 | auto_sbitmap unreachable_blocks (last_basic_block_for_fn (cfun)); |
f61e445a | 3182 | bitmap_ones (unreachable_blocks); |
6fb5fa3c | 3183 | |
04a90bec | 3184 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3185 | { |
3186 | bb_info_t bb_info = bb_table[bb->index]; | |
3187 | if (bb_info->gen) | |
3188 | bitmap_clear (bb_info->gen); | |
3189 | else | |
3f9b14ff | 3190 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3191 | |
3192 | if (bb->index == ENTRY_BLOCK) | |
3193 | ; | |
3194 | else if (bb->index == EXIT_BLOCK) | |
3195 | dse_step3_exit_block_scan (bb_info); | |
3196 | else | |
ac6929b5 | 3197 | dse_step3_scan (bb); |
6fb5fa3c DB |
3198 | if (EDGE_COUNT (bb->succs) == 0) |
3199 | mark_reachable_blocks (unreachable_blocks, bb); | |
3200 | ||
3201 | /* If this is the second time dataflow is run, delete the old | |
3202 | sets. */ | |
3203 | if (bb_info->in) | |
3204 | BITMAP_FREE (bb_info->in); | |
3205 | if (bb_info->out) | |
3206 | BITMAP_FREE (bb_info->out); | |
3207 | } | |
3208 | ||
3209 | /* For any block in an infinite loop, we must initialize the out set | |
3210 | to all ones. This could be expensive, but almost never occurs in | |
3211 | practice. However, it is common in regression tests. */ | |
d4ac4ce2 | 3212 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
6fb5fa3c DB |
3213 | { |
3214 | if (bitmap_bit_p (all_blocks, i)) | |
3215 | { | |
3216 | bb_info_t bb_info = bb_table[i]; | |
3217 | if (!all_ones) | |
3218 | { | |
3219 | unsigned int j; | |
fd6daec9 | 3220 | group_info *group; |
6fb5fa3c | 3221 | |
3f9b14ff | 3222 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
9771b263 | 3223 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
6fb5fa3c DB |
3224 | bitmap_ior_into (all_ones, group->group_kill); |
3225 | } | |
3226 | if (!bb_info->out) | |
3227 | { | |
3f9b14ff | 3228 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3229 | bitmap_copy (bb_info->out, all_ones); |
3230 | } | |
3231 | } | |
3232 | } | |
3233 | ||
3234 | if (all_ones) | |
3235 | BITMAP_FREE (all_ones); | |
6fb5fa3c DB |
3236 | } |
3237 | ||
3238 | ||
3239 | \f | |
3240 | /*---------------------------------------------------------------------------- | |
3241 | Fourth step. | |
3242 | ||
3243 | Solve the bitvector equations. | |
3244 | ----------------------------------------------------------------------------*/ | |
3245 | ||
3246 | ||
3247 | /* Confluence function for blocks with no successors. Create an out | |
3248 | set from the gen set of the exit block. This block logically has | |
3249 | the exit block as a successor. */ | |
3250 | ||
3251 | ||
3252 | ||
3253 | static void | |
3254 | dse_confluence_0 (basic_block bb) | |
3255 | { | |
3256 | bb_info_t bb_info = bb_table[bb->index]; | |
3257 | ||
3258 | if (bb->index == EXIT_BLOCK) | |
3259 | return; | |
3260 | ||
3261 | if (!bb_info->out) | |
3262 | { | |
3f9b14ff | 3263 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3264 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3265 | } | |
3266 | } | |
3267 | ||
3268 | /* Propagate the information from the in set of the dest of E to the | |
3269 | out set of the src of E. If the various in or out sets are not | |
3270 | there, that means they are all ones. */ | |
3271 | ||
1a0f3fa1 | 3272 | static bool |
6fb5fa3c DB |
3273 | dse_confluence_n (edge e) |
3274 | { | |
3275 | bb_info_t src_info = bb_table[e->src->index]; | |
3276 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3277 | ||
3278 | if (dest_info->in) | |
3279 | { | |
3280 | if (src_info->out) | |
3281 | bitmap_and_into (src_info->out, dest_info->in); | |
3282 | else | |
3283 | { | |
3f9b14ff | 3284 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3285 | bitmap_copy (src_info->out, dest_info->in); |
3286 | } | |
3287 | } | |
1a0f3fa1 | 3288 | return true; |
6fb5fa3c DB |
3289 | } |
3290 | ||
3291 | ||
3292 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
b8698a0f | 3293 | block. There are three cases: |
6fb5fa3c DB |
3294 | |
3295 | 1) The block has no kill set. In this case the kill set is all | |
3296 | ones. It does not matter what the out set of the block is, none of | |
3297 | the info can reach the top. The only thing that reaches the top is | |
3298 | the gen set and we just copy the set. | |
3299 | ||
3300 | 2) There is a kill set but no out set and bb has successors. In | |
3301 | this case we just return. Eventually an out set will be created and | |
3302 | it is better to wait than to create a set of ones. | |
3303 | ||
3304 | 3) There is both a kill and out set. We apply the obvious transfer | |
3305 | function. | |
3306 | */ | |
3307 | ||
3308 | static bool | |
3309 | dse_transfer_function (int bb_index) | |
3310 | { | |
3311 | bb_info_t bb_info = bb_table[bb_index]; | |
3312 | ||
3313 | if (bb_info->kill) | |
3314 | { | |
3315 | if (bb_info->out) | |
3316 | { | |
3317 | /* Case 3 above. */ | |
3318 | if (bb_info->in) | |
b8698a0f | 3319 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3320 | bb_info->out, bb_info->kill); |
3321 | else | |
3322 | { | |
3f9b14ff | 3323 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
b8698a0f | 3324 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3325 | bb_info->out, bb_info->kill); |
3326 | return true; | |
3327 | } | |
3328 | } | |
3329 | else | |
3330 | /* Case 2 above. */ | |
3331 | return false; | |
3332 | } | |
3333 | else | |
3334 | { | |
3335 | /* Case 1 above. If there is already an in set, nothing | |
3336 | happens. */ | |
3337 | if (bb_info->in) | |
3338 | return false; | |
3339 | else | |
3340 | { | |
3f9b14ff | 3341 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3342 | bitmap_copy (bb_info->in, bb_info->gen); |
3343 | return true; | |
3344 | } | |
3345 | } | |
3346 | } | |
3347 | ||
3348 | /* Solve the dataflow equations. */ | |
3349 | ||
3350 | static void | |
3351 | dse_step4 (void) | |
3352 | { | |
b8698a0f L |
3353 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3354 | dse_confluence_n, dse_transfer_function, | |
3355 | all_blocks, df_get_postorder (DF_BACKWARD), | |
6fb5fa3c | 3356 | df_get_n_blocks (DF_BACKWARD)); |
456610d3 | 3357 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3358 | { |
3359 | basic_block bb; | |
3360 | ||
3361 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
04a90bec | 3362 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3363 | { |
3364 | bb_info_t bb_info = bb_table[bb->index]; | |
3365 | ||
3366 | df_print_bb_index (bb, dump_file); | |
3367 | if (bb_info->in) | |
3368 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3369 | else | |
3370 | fprintf (dump_file, " in: *MISSING*\n"); | |
3371 | if (bb_info->gen) | |
3372 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3373 | else | |
3374 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3375 | if (bb_info->kill) | |
3376 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3377 | else | |
3378 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3379 | if (bb_info->out) | |
3380 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3381 | else | |
3382 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3383 | } | |
3384 | } | |
3385 | } | |
3386 | ||
3387 | ||
3388 | \f | |
3389 | /*---------------------------------------------------------------------------- | |
3390 | Fifth step. | |
3391 | ||
0d52bcc1 | 3392 | Delete the stores that can only be deleted using the global information. |
6fb5fa3c DB |
3393 | ----------------------------------------------------------------------------*/ |
3394 | ||
3395 | ||
3396 | static void | |
ac6929b5 | 3397 | dse_step5 (void) |
6fb5fa3c DB |
3398 | { |
3399 | basic_block bb; | |
11cd3bed | 3400 | FOR_EACH_BB_FN (bb, cfun) |
6fb5fa3c DB |
3401 | { |
3402 | bb_info_t bb_info = bb_table[bb->index]; | |
3403 | insn_info_t insn_info = bb_info->last_insn; | |
3404 | bitmap v = bb_info->out; | |
3405 | ||
3406 | while (insn_info) | |
3407 | { | |
3408 | bool deleted = false; | |
3409 | if (dump_file && insn_info->insn) | |
3410 | { | |
3411 | fprintf (dump_file, "starting to process insn %d\n", | |
3412 | INSN_UID (insn_info->insn)); | |
3413 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3414 | } | |
3415 | ||
3416 | /* There may have been code deleted by the dce pass run before | |
3417 | this phase. */ | |
b8698a0f | 3418 | if (insn_info->insn |
6fb5fa3c DB |
3419 | && INSN_P (insn_info->insn) |
3420 | && (!insn_info->cannot_delete) | |
3421 | && (!bitmap_empty_p (v))) | |
3422 | { | |
fd6daec9 | 3423 | store_info *store_info = insn_info->store_rec; |
6fb5fa3c DB |
3424 | |
3425 | /* Try to delete the current insn. */ | |
3426 | deleted = true; | |
b8698a0f | 3427 | |
6fb5fa3c DB |
3428 | /* Skip the clobbers. */ |
3429 | while (!store_info->is_set) | |
3430 | store_info = store_info->next; | |
3431 | ||
02ce5d90 | 3432 | HOST_WIDE_INT i, offset, width; |
ac6929b5 RB |
3433 | group_info *group_info = rtx_group_vec[store_info->group_id]; |
3434 | ||
02ce5d90 RS |
3435 | if (!store_info->offset.is_constant (&offset) |
3436 | || !store_info->width.is_constant (&width)) | |
3437 | deleted = false; | |
3438 | else | |
6fb5fa3c | 3439 | { |
02ce5d90 RS |
3440 | HOST_WIDE_INT end = offset + width; |
3441 | for (i = offset; i < end; i++) | |
6fb5fa3c | 3442 | { |
02ce5d90 RS |
3443 | int index = get_bitmap_index (group_info, i); |
3444 | ||
456610d3 | 3445 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02ce5d90 RS |
3446 | fprintf (dump_file, "i = %d, index = %d\n", |
3447 | (int) i, index); | |
3448 | if (index == 0 || !bitmap_bit_p (v, index)) | |
3449 | { | |
3450 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3451 | fprintf (dump_file, "failing at i = %d\n", | |
3452 | (int) i); | |
3453 | deleted = false; | |
3454 | break; | |
3455 | } | |
6fb5fa3c DB |
3456 | } |
3457 | } | |
3458 | if (deleted) | |
3459 | { | |
9e582b1d JR |
3460 | if (dbg_cnt (dse) |
3461 | && check_for_inc_dec_1 (insn_info)) | |
6fb5fa3c | 3462 | { |
6fb5fa3c DB |
3463 | delete_insn (insn_info->insn); |
3464 | insn_info->insn = NULL; | |
3465 | globally_deleted++; | |
3466 | } | |
3467 | } | |
3468 | } | |
3469 | /* We do want to process the local info if the insn was | |
6ed3da00 | 3470 | deleted. For instance, if the insn did a wild read, we |
6fb5fa3c | 3471 | no longer need to trash the info. */ |
b8698a0f | 3472 | if (insn_info->insn |
6fb5fa3c DB |
3473 | && INSN_P (insn_info->insn) |
3474 | && (!deleted)) | |
3475 | { | |
ac6929b5 | 3476 | scan_stores (insn_info->store_rec, v, NULL); |
6fb5fa3c DB |
3477 | if (insn_info->wild_read) |
3478 | { | |
456610d3 | 3479 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3480 | fprintf (dump_file, "wild read\n"); |
3481 | bitmap_clear (v); | |
3482 | } | |
d26c7090 | 3483 | else if (insn_info->read_rec |
a8404b9a JJ |
3484 | || insn_info->non_frame_wild_read |
3485 | || insn_info->frame_read) | |
6fb5fa3c | 3486 | { |
a8404b9a JJ |
3487 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3488 | { | |
3489 | if (!insn_info->non_frame_wild_read | |
3490 | && !insn_info->frame_read) | |
3491 | fprintf (dump_file, "regular read\n"); | |
3492 | if (insn_info->non_frame_wild_read) | |
3493 | fprintf (dump_file, "non-frame wild read\n"); | |
3494 | if (insn_info->frame_read) | |
3495 | fprintf (dump_file, "frame read\n"); | |
3496 | } | |
ac6929b5 | 3497 | scan_reads (insn_info, v, NULL); |
6fb5fa3c DB |
3498 | } |
3499 | } | |
b8698a0f | 3500 | |
6fb5fa3c DB |
3501 | insn_info = insn_info->prev_insn; |
3502 | } | |
3503 | } | |
3504 | } | |
3505 | ||
3506 | ||
6fb5fa3c DB |
3507 | \f |
3508 | /*---------------------------------------------------------------------------- | |
3509 | Sixth step. | |
3510 | ||
8dd5516b JJ |
3511 | Delete stores made redundant by earlier stores (which store the same |
3512 | value) that couldn't be eliminated. | |
3513 | ----------------------------------------------------------------------------*/ | |
3514 | ||
3515 | static void | |
3516 | dse_step6 (void) | |
3517 | { | |
3518 | basic_block bb; | |
3519 | ||
04a90bec | 3520 | FOR_ALL_BB_FN (bb, cfun) |
8dd5516b JJ |
3521 | { |
3522 | bb_info_t bb_info = bb_table[bb->index]; | |
3523 | insn_info_t insn_info = bb_info->last_insn; | |
3524 | ||
3525 | while (insn_info) | |
3526 | { | |
3527 | /* There may have been code deleted by the dce pass run before | |
3528 | this phase. */ | |
3529 | if (insn_info->insn | |
3530 | && INSN_P (insn_info->insn) | |
3531 | && !insn_info->cannot_delete) | |
3532 | { | |
fd6daec9 | 3533 | store_info *s_info = insn_info->store_rec; |
8dd5516b JJ |
3534 | |
3535 | while (s_info && !s_info->is_set) | |
3536 | s_info = s_info->next; | |
3537 | if (s_info | |
3538 | && s_info->redundant_reason | |
3539 | && s_info->redundant_reason->insn | |
3540 | && INSN_P (s_info->redundant_reason->insn)) | |
3541 | { | |
eb92d49a | 3542 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
456610d3 | 3543 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
3544 | fprintf (dump_file, "Locally deleting insn %d " |
3545 | "because insn %d stores the " | |
3546 | "same value and couldn't be " | |
3547 | "eliminated\n", | |
3548 | INSN_UID (insn_info->insn), | |
3549 | INSN_UID (rinsn)); | |
3550 | delete_dead_store_insn (insn_info); | |
3551 | } | |
3552 | } | |
3553 | insn_info = insn_info->prev_insn; | |
3554 | } | |
3555 | } | |
3556 | } | |
3557 | \f | |
3558 | /*---------------------------------------------------------------------------- | |
3559 | Seventh step. | |
3560 | ||
b8698a0f | 3561 | Destroy everything left standing. |
6fb5fa3c DB |
3562 | ----------------------------------------------------------------------------*/ |
3563 | ||
b8698a0f | 3564 | static void |
3f9b14ff | 3565 | dse_step7 (void) |
6fb5fa3c | 3566 | { |
3f9b14ff SB |
3567 | bitmap_obstack_release (&dse_bitmap_obstack); |
3568 | obstack_free (&dse_obstack, NULL); | |
370f38e8 | 3569 | |
6fb5fa3c DB |
3570 | end_alias_analysis (); |
3571 | free (bb_table); | |
c203e8a7 TS |
3572 | delete rtx_group_table; |
3573 | rtx_group_table = NULL; | |
9771b263 | 3574 | rtx_group_vec.release (); |
6fb5fa3c DB |
3575 | BITMAP_FREE (all_blocks); |
3576 | BITMAP_FREE (scratch); | |
3577 | ||
da6603c6 | 3578 | rtx_store_info_pool.release (); |
fb0b2914 ML |
3579 | read_info_type_pool.release (); |
3580 | insn_info_type_pool.release (); | |
3581 | dse_bb_info_type_pool.release (); | |
3582 | group_info_pool.release (); | |
3583 | deferred_change_pool.release (); | |
6fb5fa3c DB |
3584 | } |
3585 | ||
3586 | ||
6fb5fa3c DB |
3587 | /* ------------------------------------------------------------------------- |
3588 | DSE | |
3589 | ------------------------------------------------------------------------- */ | |
3590 | ||
3591 | /* Callback for running pass_rtl_dse. */ | |
3592 | ||
3593 | static unsigned int | |
3594 | rest_of_handle_dse (void) | |
3595 | { | |
6fb5fa3c DB |
3596 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3597 | ||
02b47899 KZ |
3598 | /* Need the notes since we must track live hardregs in the forwards |
3599 | direction. */ | |
3600 | df_note_add_problem (); | |
3601 | df_analyze (); | |
3602 | ||
6fb5fa3c DB |
3603 | dse_step0 (); |
3604 | dse_step1 (); | |
3605 | dse_step2_init (); | |
ac6929b5 | 3606 | if (dse_step2 ()) |
6fb5fa3c DB |
3607 | { |
3608 | df_set_flags (DF_LR_RUN_DCE); | |
3609 | df_analyze (); | |
456610d3 | 3610 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 3611 | fprintf (dump_file, "doing global processing\n"); |
ac6929b5 | 3612 | dse_step3 (); |
6fb5fa3c | 3613 | dse_step4 (); |
ac6929b5 | 3614 | dse_step5 (); |
6fb5fa3c DB |
3615 | } |
3616 | ||
8dd5516b | 3617 | dse_step6 (); |
3f9b14ff | 3618 | dse_step7 (); |
6fb5fa3c DB |
3619 | |
3620 | if (dump_file) | |
ac6929b5 RB |
3621 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d\n", |
3622 | locally_deleted, globally_deleted); | |
6c5ed3f1 JJ |
3623 | |
3624 | /* DSE can eliminate potentially-trapping MEMs. | |
3625 | Remove any EH edges associated with them. */ | |
3626 | if ((locally_deleted || globally_deleted) | |
3627 | && cfun->can_throw_non_call_exceptions | |
3628 | && purge_all_dead_edges ()) | |
7f5ddfcb JJ |
3629 | { |
3630 | free_dominance_info (CDI_DOMINATORS); | |
3631 | cleanup_cfg (0); | |
3632 | } | |
6c5ed3f1 | 3633 | |
6fb5fa3c DB |
3634 | return 0; |
3635 | } | |
3636 | ||
17795822 TS |
3637 | namespace { |
3638 | ||
3639 | const pass_data pass_data_rtl_dse1 = | |
27a4cd48 DM |
3640 | { |
3641 | RTL_PASS, /* type */ | |
3642 | "dse1", /* name */ | |
3643 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3644 | TV_DSE1, /* tv_id */ |
3645 | 0, /* properties_required */ | |
3646 | 0, /* properties_provided */ | |
3647 | 0, /* properties_destroyed */ | |
3648 | 0, /* todo_flags_start */ | |
3bea341f | 3649 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c DB |
3650 | }; |
3651 | ||
17795822 | 3652 | class pass_rtl_dse1 : public rtl_opt_pass |
27a4cd48 DM |
3653 | { |
3654 | public: | |
c3284718 RS |
3655 | pass_rtl_dse1 (gcc::context *ctxt) |
3656 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
27a4cd48 DM |
3657 | {} |
3658 | ||
3659 | /* opt_pass methods: */ | |
1a3d085c TS |
3660 | virtual bool gate (function *) |
3661 | { | |
3662 | return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
3663 | } | |
3664 | ||
be55bfe6 | 3665 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3666 | |
3667 | }; // class pass_rtl_dse1 | |
3668 | ||
17795822 TS |
3669 | } // anon namespace |
3670 | ||
27a4cd48 DM |
3671 | rtl_opt_pass * |
3672 | make_pass_rtl_dse1 (gcc::context *ctxt) | |
3673 | { | |
3674 | return new pass_rtl_dse1 (ctxt); | |
3675 | } | |
3676 | ||
17795822 TS |
3677 | namespace { |
3678 | ||
3679 | const pass_data pass_data_rtl_dse2 = | |
27a4cd48 DM |
3680 | { |
3681 | RTL_PASS, /* type */ | |
3682 | "dse2", /* name */ | |
3683 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3684 | TV_DSE2, /* tv_id */ |
3685 | 0, /* properties_required */ | |
3686 | 0, /* properties_provided */ | |
3687 | 0, /* properties_destroyed */ | |
3688 | 0, /* todo_flags_start */ | |
3bea341f | 3689 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c | 3690 | }; |
27a4cd48 | 3691 | |
17795822 | 3692 | class pass_rtl_dse2 : public rtl_opt_pass |
27a4cd48 DM |
3693 | { |
3694 | public: | |
c3284718 RS |
3695 | pass_rtl_dse2 (gcc::context *ctxt) |
3696 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
27a4cd48 DM |
3697 | {} |
3698 | ||
3699 | /* opt_pass methods: */ | |
1a3d085c TS |
3700 | virtual bool gate (function *) |
3701 | { | |
3702 | return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
3703 | } | |
3704 | ||
be55bfe6 | 3705 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3706 | |
3707 | }; // class pass_rtl_dse2 | |
3708 | ||
17795822 TS |
3709 | } // anon namespace |
3710 | ||
27a4cd48 DM |
3711 | rtl_opt_pass * |
3712 | make_pass_rtl_dse2 (gcc::context *ctxt) | |
3713 | { | |
3714 | return new pass_rtl_dse2 (ctxt); | |
3715 | } |