]>
Commit | Line | Data |
---|---|---|
6fb5fa3c | 1 | /* RTL dead store elimination. |
757e8ba2 | 2 | Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012 |
457eeaae | 3 | Free Software Foundation, Inc. |
6fb5fa3c DB |
4 | |
5 | Contributed by Richard Sandiford <rsandifor@codesourcery.com> | |
6 | and Kenneth Zadeck <zadeck@naturalbridge.com> | |
7 | ||
8 | This file is part of GCC. | |
9 | ||
10 | GCC is free software; you can redistribute it and/or modify it under | |
11 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 12 | Software Foundation; either version 3, or (at your option) any later |
6fb5fa3c DB |
13 | version. |
14 | ||
15 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
16 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
17 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
18 | for more details. | |
19 | ||
20 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
21 | along with GCC; see the file COPYING3. If not see |
22 | <http://www.gnu.org/licenses/>. */ | |
6fb5fa3c DB |
23 | |
24 | #undef BASELINE | |
25 | ||
26 | #include "config.h" | |
27 | #include "system.h" | |
28 | #include "coretypes.h" | |
29 | #include "hashtab.h" | |
30 | #include "tm.h" | |
31 | #include "rtl.h" | |
32 | #include "tree.h" | |
18b526e8 | 33 | #include "tm_p.h" |
6fb5fa3c DB |
34 | #include "regs.h" |
35 | #include "hard-reg-set.h" | |
9e582b1d | 36 | #include "regset.h" |
6fb5fa3c DB |
37 | #include "flags.h" |
38 | #include "df.h" | |
39 | #include "cselib.h" | |
6fb5fa3c DB |
40 | #include "tree-pass.h" |
41 | #include "alloc-pool.h" | |
42 | #include "alias.h" | |
43 | #include "insn-config.h" | |
44 | #include "expr.h" | |
45 | #include "recog.h" | |
8660aaae | 46 | #include "optabs.h" |
6fb5fa3c | 47 | #include "dbgcnt.h" |
67d0afe9 | 48 | #include "target.h" |
dabd47e7 | 49 | #include "params.h" |
6f4185d7 | 50 | #include "tree-flow.h" /* for may_be_aliased */ |
6fb5fa3c DB |
51 | |
52 | /* This file contains three techniques for performing Dead Store | |
b8698a0f | 53 | Elimination (dse). |
6fb5fa3c DB |
54 | |
55 | * The first technique performs dse locally on any base address. It | |
56 | is based on the cselib which is a local value numbering technique. | |
57 | This technique is local to a basic block but deals with a fairly | |
58 | general addresses. | |
b8698a0f | 59 | |
6fb5fa3c DB |
60 | * The second technique performs dse globally but is restricted to |
61 | base addresses that are either constant or are relative to the | |
62 | frame_pointer. | |
63 | ||
64 | * The third technique, (which is only done after register allocation) | |
65 | processes the spill spill slots. This differs from the second | |
66 | technique because it takes advantage of the fact that spilling is | |
67 | completely free from the effects of aliasing. | |
68 | ||
69 | Logically, dse is a backwards dataflow problem. A store can be | |
70 | deleted if it if cannot be reached in the backward direction by any | |
71 | use of the value being stored. However, the local technique uses a | |
72 | forwards scan of the basic block because cselib requires that the | |
73 | block be processed in that order. | |
74 | ||
75 | The pass is logically broken into 7 steps: | |
76 | ||
77 | 0) Initialization. | |
78 | ||
79 | 1) The local algorithm, as well as scanning the insns for the two | |
80 | global algorithms. | |
81 | ||
82 | 2) Analysis to see if the global algs are necessary. In the case | |
83 | of stores base on a constant address, there must be at least two | |
84 | stores to that address, to make it possible to delete some of the | |
85 | stores. In the case of stores off of the frame or spill related | |
86 | stores, only one store to an address is necessary because those | |
87 | stores die at the end of the function. | |
88 | ||
b8698a0f | 89 | 3) Set up the global dataflow equations based on processing the |
6fb5fa3c DB |
90 | info parsed in the first step. |
91 | ||
92 | 4) Solve the dataflow equations. | |
93 | ||
94 | 5) Delete the insns that the global analysis has indicated are | |
95 | unnecessary. | |
96 | ||
073a8998 | 97 | 6) Delete insns that store the same value as preceding store |
8dd5516b JJ |
98 | where the earlier store couldn't be eliminated. |
99 | ||
100 | 7) Cleanup. | |
6fb5fa3c DB |
101 | |
102 | This step uses cselib and canon_rtx to build the largest expression | |
103 | possible for each address. This pass is a forwards pass through | |
104 | each basic block. From the point of view of the global technique, | |
105 | the first pass could examine a block in either direction. The | |
0d52bcc1 | 106 | forwards ordering is to accommodate cselib. |
6fb5fa3c DB |
107 | |
108 | We a simplifying assumption: addresses fall into four broad | |
109 | categories: | |
110 | ||
111 | 1) base has rtx_varies_p == false, offset is constant. | |
112 | 2) base has rtx_varies_p == false, offset variable. | |
113 | 3) base has rtx_varies_p == true, offset constant. | |
114 | 4) base has rtx_varies_p == true, offset variable. | |
115 | ||
116 | The local passes are able to process all 4 kinds of addresses. The | |
117 | global pass only handles (1). | |
118 | ||
119 | The global problem is formulated as follows: | |
120 | ||
121 | A store, S1, to address A, where A is not relative to the stack | |
122 | frame, can be eliminated if all paths from S1 to the end of the | |
123 | of the function contain another store to A before a read to A. | |
124 | ||
125 | If the address A is relative to the stack frame, a store S2 to A | |
126 | can be eliminated if there are no paths from S1 that reach the | |
127 | end of the function that read A before another store to A. In | |
128 | this case S2 can be deleted if there are paths to from S2 to the | |
129 | end of the function that have no reads or writes to A. This | |
130 | second case allows stores to the stack frame to be deleted that | |
131 | would otherwise die when the function returns. This cannot be | |
132 | done if stores_off_frame_dead_at_return is not true. See the doc | |
133 | for that variable for when this variable is false. | |
134 | ||
135 | The global problem is formulated as a backwards set union | |
136 | dataflow problem where the stores are the gens and reads are the | |
137 | kills. Set union problems are rare and require some special | |
138 | handling given our representation of bitmaps. A straightforward | |
139 | implementation of requires a lot of bitmaps filled with 1s. | |
140 | These are expensive and cumbersome in our bitmap formulation so | |
141 | care has been taken to avoid large vectors filled with 1s. See | |
142 | the comments in bb_info and in the dataflow confluence functions | |
b8698a0f | 143 | for details. |
6fb5fa3c DB |
144 | |
145 | There are two places for further enhancements to this algorithm: | |
b8698a0f | 146 | |
6fb5fa3c DB |
147 | 1) The original dse which was embedded in a pass called flow also |
148 | did local address forwarding. For example in | |
149 | ||
150 | A <- r100 | |
151 | ... <- A | |
152 | ||
153 | flow would replace the right hand side of the second insn with a | |
6ed3da00 | 154 | reference to r100. Most of the information is available to add this |
6fb5fa3c DB |
155 | to this pass. It has not done it because it is a lot of work in |
156 | the case that either r100 is assigned to between the first and | |
157 | second insn and/or the second insn is a load of part of the value | |
158 | stored by the first insn. | |
159 | ||
160 | insn 5 in gcc.c-torture/compile/990203-1.c simple case. | |
161 | insn 15 in gcc.c-torture/execute/20001017-2.c simple case. | |
162 | insn 25 in gcc.c-torture/execute/20001026-1.c simple case. | |
163 | insn 44 in gcc.c-torture/execute/20010910-1.c simple case. | |
164 | ||
165 | 2) The cleaning up of spill code is quite profitable. It currently | |
166 | depends on reading tea leaves and chicken entrails left by reload. | |
167 | This pass depends on reload creating a singleton alias set for each | |
168 | spill slot and telling the next dse pass which of these alias sets | |
169 | are the singletons. Rather than analyze the addresses of the | |
170 | spills, dse's spill processing just does analysis of the loads and | |
171 | stores that use those alias sets. There are three cases where this | |
172 | falls short: | |
173 | ||
174 | a) Reload sometimes creates the slot for one mode of access, and | |
175 | then inserts loads and/or stores for a smaller mode. In this | |
176 | case, the current code just punts on the slot. The proper thing | |
177 | to do is to back out and use one bit vector position for each | |
178 | byte of the entity associated with the slot. This depends on | |
179 | KNOWING that reload always generates the accesses for each of the | |
180 | bytes in some canonical (read that easy to understand several | |
181 | passes after reload happens) way. | |
182 | ||
183 | b) Reload sometimes decides that spill slot it allocated was not | |
184 | large enough for the mode and goes back and allocates more slots | |
185 | with the same mode and alias set. The backout in this case is a | |
186 | little more graceful than (a). In this case the slot is unmarked | |
187 | as being a spill slot and if final address comes out to be based | |
b8698a0f | 188 | off the frame pointer, the global algorithm handles this slot. |
6fb5fa3c DB |
189 | |
190 | c) For any pass that may prespill, there is currently no | |
191 | mechanism to tell the dse pass that the slot being used has the | |
192 | special properties that reload uses. It may be that all that is | |
0d52bcc1 | 193 | required is to have those passes make the same calls that reload |
6fb5fa3c DB |
194 | does, assuming that the alias sets can be manipulated in the same |
195 | way. */ | |
196 | ||
197 | /* There are limits to the size of constant offsets we model for the | |
198 | global problem. There are certainly test cases, that exceed this | |
199 | limit, however, it is unlikely that there are important programs | |
200 | that really have constant offsets this size. */ | |
201 | #define MAX_OFFSET (64 * 1024) | |
202 | ||
3f9b14ff SB |
203 | /* Obstack for the DSE dataflow bitmaps. We don't want to put these |
204 | on the default obstack because these bitmaps can grow quite large | |
205 | (~2GB for the small (!) test case of PR54146) and we'll hold on to | |
206 | all that memory until the end of the compiler run. | |
207 | As a bonus, delete_tree_live_info can destroy all the bitmaps by just | |
208 | releasing the whole obstack. */ | |
209 | static bitmap_obstack dse_bitmap_obstack; | |
210 | ||
211 | /* Obstack for other data. As for above: Kinda nice to be able to | |
212 | throw it all away at the end in one big sweep. */ | |
213 | static struct obstack dse_obstack; | |
214 | ||
215 | /* Scratch bitmap for cselib's cselib_expand_value_rtx. */ | |
6fb5fa3c | 216 | static bitmap scratch = NULL; |
3f9b14ff | 217 | |
6fb5fa3c DB |
218 | struct insn_info; |
219 | ||
220 | /* This structure holds information about a candidate store. */ | |
b8698a0f | 221 | struct store_info |
6fb5fa3c DB |
222 | { |
223 | ||
224 | /* False means this is a clobber. */ | |
225 | bool is_set; | |
226 | ||
8dd5516b JJ |
227 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
228 | bool is_large; | |
229 | ||
6fb5fa3c DB |
230 | /* The id of the mem group of the base address. If rtx_varies_p is |
231 | true, this is -1. Otherwise, it is the index into the group | |
232 | table. */ | |
233 | int group_id; | |
b8698a0f | 234 | |
6fb5fa3c DB |
235 | /* This is the cselib value. */ |
236 | cselib_val *cse_base; | |
237 | ||
238 | /* This canonized mem. */ | |
239 | rtx mem; | |
240 | ||
6216f94e | 241 | /* Canonized MEM address for use by canon_true_dependence. */ |
6fb5fa3c DB |
242 | rtx mem_addr; |
243 | ||
244 | /* If this is non-zero, it is the alias set of a spill location. */ | |
4862826d | 245 | alias_set_type alias_set; |
6fb5fa3c DB |
246 | |
247 | /* The offset of the first and byte before the last byte associated | |
248 | with the operation. */ | |
8dd5516b JJ |
249 | HOST_WIDE_INT begin, end; |
250 | ||
251 | union | |
252 | { | |
253 | /* A bitmask as wide as the number of bytes in the word that | |
254 | contains a 1 if the byte may be needed. The store is unused if | |
255 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
256 | unsigned HOST_WIDE_INT small_bitmask; | |
257 | ||
258 | struct | |
259 | { | |
260 | /* A bitmap with one bit per byte. Cleared bit means the position | |
261 | is needed. Used if IS_LARGE is false. */ | |
dc491a25 | 262 | bitmap bmap; |
6fb5fa3c | 263 | |
8dd5516b JJ |
264 | /* Number of set bits (i.e. unneeded bytes) in BITMAP. If it is |
265 | equal to END - BEGIN, the whole store is unused. */ | |
266 | int count; | |
267 | } large; | |
268 | } positions_needed; | |
6fb5fa3c DB |
269 | |
270 | /* The next store info for this insn. */ | |
271 | struct store_info *next; | |
272 | ||
273 | /* The right hand side of the store. This is used if there is a | |
274 | subsequent reload of the mems address somewhere later in the | |
275 | basic block. */ | |
8dd5516b JJ |
276 | rtx rhs; |
277 | ||
278 | /* If rhs is or holds a constant, this contains that constant, | |
279 | otherwise NULL. */ | |
280 | rtx const_rhs; | |
281 | ||
282 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
283 | insn stored. These aren't eliminated early, because doing that | |
284 | might prevent the earlier larger store to be eliminated. */ | |
285 | struct insn_info *redundant_reason; | |
6fb5fa3c DB |
286 | }; |
287 | ||
4fe663b0 L |
288 | /* Return a bitmask with the first N low bits set. */ |
289 | ||
290 | static unsigned HOST_WIDE_INT | |
291 | lowpart_bitmask (int n) | |
292 | { | |
293 | unsigned HOST_WIDE_INT mask = ~(unsigned HOST_WIDE_INT) 0; | |
294 | return mask >> (HOST_BITS_PER_WIDE_INT - n); | |
295 | } | |
296 | ||
6fb5fa3c DB |
297 | typedef struct store_info *store_info_t; |
298 | static alloc_pool cse_store_info_pool; | |
299 | static alloc_pool rtx_store_info_pool; | |
300 | ||
301 | /* This structure holds information about a load. These are only | |
302 | built for rtx bases. */ | |
b8698a0f | 303 | struct read_info |
6fb5fa3c DB |
304 | { |
305 | /* The id of the mem group of the base address. */ | |
306 | int group_id; | |
307 | ||
308 | /* If this is non-zero, it is the alias set of a spill location. */ | |
4862826d | 309 | alias_set_type alias_set; |
6fb5fa3c DB |
310 | |
311 | /* The offset of the first and byte after the last byte associated | |
312 | with the operation. If begin == end == 0, the read did not have | |
313 | a constant offset. */ | |
314 | int begin, end; | |
315 | ||
316 | /* The mem being read. */ | |
317 | rtx mem; | |
318 | ||
319 | /* The next read_info for this insn. */ | |
320 | struct read_info *next; | |
321 | }; | |
322 | typedef struct read_info *read_info_t; | |
323 | static alloc_pool read_info_pool; | |
324 | ||
325 | ||
326 | /* One of these records is created for each insn. */ | |
327 | ||
b8698a0f | 328 | struct insn_info |
6fb5fa3c DB |
329 | { |
330 | /* Set true if the insn contains a store but the insn itself cannot | |
331 | be deleted. This is set if the insn is a parallel and there is | |
332 | more than one non dead output or if the insn is in some way | |
333 | volatile. */ | |
334 | bool cannot_delete; | |
335 | ||
336 | /* This field is only used by the global algorithm. It is set true | |
337 | if the insn contains any read of mem except for a (1). This is | |
338 | also set if the insn is a call or has a clobber mem. If the insn | |
339 | contains a wild read, the use_rec will be null. */ | |
340 | bool wild_read; | |
341 | ||
d26c7090 ER |
342 | /* This is true only for CALL instructions which could potentially read |
343 | any non-frame memory location. This field is used by the global | |
344 | algorithm. */ | |
345 | bool non_frame_wild_read; | |
346 | ||
50f0f366 EB |
347 | /* This field is only used for the processing of const functions. |
348 | These functions cannot read memory, but they can read the stack | |
64520bdc EB |
349 | because that is where they may get their parms. We need to be |
350 | this conservative because, like the store motion pass, we don't | |
351 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
352 | Moreover, we need to distinguish two cases: | |
353 | 1. Before reload (register elimination), the stores related to | |
354 | outgoing arguments are stack pointer based and thus deemed | |
355 | of non-constant base in this pass. This requires special | |
356 | handling but also means that the frame pointer based stores | |
357 | need not be killed upon encountering a const function call. | |
358 | 2. After reload, the stores related to outgoing arguments can be | |
359 | either stack pointer or hard frame pointer based. This means | |
360 | that we have no other choice than also killing all the frame | |
361 | pointer based stores upon encountering a const function call. | |
362 | This field is set after reload for const function calls. Having | |
363 | this set is less severe than a wild read, it just means that all | |
364 | the frame related stores are killed rather than all the stores. */ | |
365 | bool frame_read; | |
366 | ||
367 | /* This field is only used for the processing of const functions. | |
368 | It is set if the insn may contain a stack pointer based store. */ | |
50f0f366 | 369 | bool stack_pointer_based; |
6fb5fa3c DB |
370 | |
371 | /* This is true if any of the sets within the store contains a | |
372 | cselib base. Such stores can only be deleted by the local | |
373 | algorithm. */ | |
374 | bool contains_cselib_groups; | |
375 | ||
376 | /* The insn. */ | |
377 | rtx insn; | |
378 | ||
379 | /* The list of mem sets or mem clobbers that are contained in this | |
380 | insn. If the insn is deletable, it contains only one mem set. | |
381 | But it could also contain clobbers. Insns that contain more than | |
382 | one mem set are not deletable, but each of those mems are here in | |
6ed3da00 | 383 | order to provide info to delete other insns. */ |
6fb5fa3c DB |
384 | store_info_t store_rec; |
385 | ||
386 | /* The linked list of mem uses in this insn. Only the reads from | |
387 | rtx bases are listed here. The reads to cselib bases are | |
388 | completely processed during the first scan and so are never | |
389 | created. */ | |
390 | read_info_t read_rec; | |
391 | ||
9e582b1d JR |
392 | /* The live fixed registers. We assume only fixed registers can |
393 | cause trouble by being clobbered from an expanded pattern; | |
394 | storing only the live fixed registers (rather than all registers) | |
395 | means less memory needs to be allocated / copied for the individual | |
396 | stores. */ | |
397 | regset fixed_regs_live; | |
398 | ||
6fb5fa3c DB |
399 | /* The prev insn in the basic block. */ |
400 | struct insn_info * prev_insn; | |
401 | ||
402 | /* The linked list of insns that are in consideration for removal in | |
073a8998 | 403 | the forwards pass through the basic block. This pointer may be |
6fb5fa3c | 404 | trash as it is not cleared when a wild read occurs. The only |
fa10beec | 405 | time it is guaranteed to be correct is when the traversal starts |
6fb5fa3c DB |
406 | at active_local_stores. */ |
407 | struct insn_info * next_local_store; | |
408 | }; | |
409 | ||
410 | typedef struct insn_info *insn_info_t; | |
411 | static alloc_pool insn_info_pool; | |
412 | ||
413 | /* The linked list of stores that are under consideration in this | |
b8698a0f | 414 | basic block. */ |
6fb5fa3c | 415 | static insn_info_t active_local_stores; |
dabd47e7 | 416 | static int active_local_stores_len; |
6fb5fa3c | 417 | |
b8698a0f | 418 | struct bb_info |
6fb5fa3c DB |
419 | { |
420 | ||
421 | /* Pointer to the insn info for the last insn in the block. These | |
422 | are linked so this is how all of the insns are reached. During | |
423 | scanning this is the current insn being scanned. */ | |
424 | insn_info_t last_insn; | |
425 | ||
426 | /* The info for the global dataflow problem. */ | |
427 | ||
428 | ||
429 | /* This is set if the transfer function should and in the wild_read | |
430 | bitmap before applying the kill and gen sets. That vector knocks | |
431 | out most of the bits in the bitmap and thus speeds up the | |
432 | operations. */ | |
433 | bool apply_wild_read; | |
434 | ||
02b47899 KZ |
435 | /* The following 4 bitvectors hold information about which positions |
436 | of which stores are live or dead. They are indexed by | |
437 | get_bitmap_index. */ | |
438 | ||
6fb5fa3c DB |
439 | /* The set of store positions that exist in this block before a wild read. */ |
440 | bitmap gen; | |
b8698a0f | 441 | |
6fb5fa3c DB |
442 | /* The set of load positions that exist in this block above the |
443 | same position of a store. */ | |
444 | bitmap kill; | |
445 | ||
446 | /* The set of stores that reach the top of the block without being | |
447 | killed by a read. | |
448 | ||
449 | Do not represent the in if it is all ones. Note that this is | |
450 | what the bitvector should logically be initialized to for a set | |
451 | intersection problem. However, like the kill set, this is too | |
452 | expensive. So initially, the in set will only be created for the | |
453 | exit block and any block that contains a wild read. */ | |
454 | bitmap in; | |
455 | ||
456 | /* The set of stores that reach the bottom of the block from it's | |
457 | successors. | |
458 | ||
459 | Do not represent the in if it is all ones. Note that this is | |
460 | what the bitvector should logically be initialized to for a set | |
461 | intersection problem. However, like the kill and in set, this is | |
462 | too expensive. So what is done is that the confluence operator | |
463 | just initializes the vector from one of the out sets of the | |
464 | successors of the block. */ | |
465 | bitmap out; | |
02b47899 KZ |
466 | |
467 | /* The following bitvector is indexed by the reg number. It | |
468 | contains the set of regs that are live at the current instruction | |
469 | being processed. While it contains info for all of the | |
9e582b1d JR |
470 | registers, only the hard registers are actually examined. It is used |
471 | to assure that shift and/or add sequences that are inserted do not | |
073a8998 | 472 | accidentally clobber live hard regs. */ |
02b47899 | 473 | bitmap regs_live; |
6fb5fa3c DB |
474 | }; |
475 | ||
476 | typedef struct bb_info *bb_info_t; | |
477 | static alloc_pool bb_info_pool; | |
478 | ||
479 | /* Table to hold all bb_infos. */ | |
480 | static bb_info_t *bb_table; | |
481 | ||
482 | /* There is a group_info for each rtx base that is used to reference | |
483 | memory. There are also not many of the rtx bases because they are | |
484 | very limited in scope. */ | |
485 | ||
b8698a0f | 486 | struct group_info |
6fb5fa3c DB |
487 | { |
488 | /* The actual base of the address. */ | |
489 | rtx rtx_base; | |
490 | ||
491 | /* The sequential id of the base. This allows us to have a | |
492 | canonical ordering of these that is not based on addresses. */ | |
493 | int id; | |
494 | ||
8f5929e1 JJ |
495 | /* True if there are any positions that are to be processed |
496 | globally. */ | |
497 | bool process_globally; | |
498 | ||
499 | /* True if the base of this group is either the frame_pointer or | |
500 | hard_frame_pointer. */ | |
501 | bool frame_related; | |
502 | ||
d32f725a OH |
503 | /* A mem wrapped around the base pointer for the group in order to do |
504 | read dependency. It must be given BLKmode in order to encompass all | |
505 | the possible offsets from the base. */ | |
6fb5fa3c | 506 | rtx base_mem; |
b8698a0f | 507 | |
6216f94e JJ |
508 | /* Canonized version of base_mem's address. */ |
509 | rtx canon_base_addr; | |
6fb5fa3c DB |
510 | |
511 | /* These two sets of two bitmaps are used to keep track of how many | |
6ed3da00 | 512 | stores are actually referencing that position from this base. We |
6fb5fa3c | 513 | only do this for rtx bases as this will be used to assign |
6ed3da00 | 514 | positions in the bitmaps for the global problem. Bit N is set in |
6fb5fa3c DB |
515 | store1 on the first store for offset N. Bit N is set in store2 |
516 | for the second store to offset N. This is all we need since we | |
517 | only care about offsets that have two or more stores for them. | |
518 | ||
519 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
520 | for 0 and greater offsets. | |
521 | ||
522 | There is one special case here, for stores into the stack frame, | |
523 | we will or store1 into store2 before deciding which stores look | |
524 | at globally. This is because stores to the stack frame that have | |
525 | no other reads before the end of the function can also be | |
526 | deleted. */ | |
527 | bitmap store1_n, store1_p, store2_n, store2_p; | |
528 | ||
d26c7090 ER |
529 | /* These bitmaps keep track of offsets in this group escape this function. |
530 | An offset escapes if it corresponds to a named variable whose | |
531 | addressable flag is set. */ | |
532 | bitmap escaped_n, escaped_p; | |
533 | ||
6ed3da00 | 534 | /* The positions in this bitmap have the same assignments as the in, |
6fb5fa3c | 535 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6ed3da00 | 536 | the positions that are occupied by stores for this group. */ |
6fb5fa3c DB |
537 | bitmap group_kill; |
538 | ||
6fb5fa3c | 539 | /* The offset_map is used to map the offsets from this base into |
6ed3da00 | 540 | positions in the global bitmaps. It is only created after all of |
6fb5fa3c DB |
541 | the all of stores have been scanned and we know which ones we |
542 | care about. */ | |
b8698a0f L |
543 | int *offset_map_n, *offset_map_p; |
544 | int offset_map_size_n, offset_map_size_p; | |
6fb5fa3c DB |
545 | }; |
546 | typedef struct group_info *group_info_t; | |
5f754896 | 547 | typedef const struct group_info *const_group_info_t; |
6fb5fa3c DB |
548 | static alloc_pool rtx_group_info_pool; |
549 | ||
550 | /* Tables of group_info structures, hashed by base value. */ | |
551 | static htab_t rtx_group_table; | |
552 | ||
553 | /* Index into the rtx_group_vec. */ | |
554 | static int rtx_group_next_id; | |
555 | ||
556 | DEF_VEC_P(group_info_t); | |
557 | DEF_VEC_ALLOC_P(group_info_t,heap); | |
558 | ||
559 | static VEC(group_info_t,heap) *rtx_group_vec; | |
560 | ||
561 | ||
562 | /* This structure holds the set of changes that are being deferred | |
563 | when removing read operation. See replace_read. */ | |
b8698a0f | 564 | struct deferred_change |
6fb5fa3c DB |
565 | { |
566 | ||
567 | /* The mem that is being replaced. */ | |
568 | rtx *loc; | |
569 | ||
570 | /* The reg it is being replaced with. */ | |
571 | rtx reg; | |
572 | ||
573 | struct deferred_change *next; | |
574 | }; | |
575 | ||
576 | typedef struct deferred_change *deferred_change_t; | |
577 | static alloc_pool deferred_change_pool; | |
578 | ||
579 | static deferred_change_t deferred_change_list = NULL; | |
580 | ||
581 | /* This are used to hold the alias sets of spill variables. Since | |
582 | these are never aliased and there may be a lot of them, it makes | |
583 | sense to treat them specially. This bitvector is only allocated in | |
584 | calls from dse_record_singleton_alias_set which currently is only | |
585 | made during reload1. So when dse is called before reload this | |
586 | mechanism does nothing. */ | |
587 | ||
588 | static bitmap clear_alias_sets = NULL; | |
589 | ||
590 | /* The set of clear_alias_sets that have been disqualified because | |
591 | there are loads or stores using a different mode than the alias set | |
b8698a0f | 592 | was registered with. */ |
6fb5fa3c DB |
593 | static bitmap disqualified_clear_alias_sets = NULL; |
594 | ||
595 | /* The group that holds all of the clear_alias_sets. */ | |
596 | static group_info_t clear_alias_group; | |
597 | ||
598 | /* The modes of the clear_alias_sets. */ | |
599 | static htab_t clear_alias_mode_table; | |
600 | ||
601 | /* Hash table element to look up the mode for an alias set. */ | |
602 | struct clear_alias_mode_holder | |
603 | { | |
4862826d | 604 | alias_set_type alias_set; |
6fb5fa3c DB |
605 | enum machine_mode mode; |
606 | }; | |
607 | ||
608 | static alloc_pool clear_alias_mode_pool; | |
609 | ||
e3b5732b | 610 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
9dd9bf80 | 611 | this for vararg functions because they play games with the frame. */ |
6fb5fa3c DB |
612 | static bool stores_off_frame_dead_at_return; |
613 | ||
614 | /* Counter for stats. */ | |
b8698a0f L |
615 | static int globally_deleted; |
616 | static int locally_deleted; | |
617 | static int spill_deleted; | |
618 | ||
6fb5fa3c DB |
619 | static bitmap all_blocks; |
620 | ||
d26c7090 ER |
621 | /* Locations that are killed by calls in the global phase. */ |
622 | static bitmap kill_on_calls; | |
623 | ||
6fb5fa3c DB |
624 | /* The number of bits used in the global bitmaps. */ |
625 | static unsigned int current_position; | |
626 | ||
627 | ||
7d817ebc DE |
628 | static bool gate_dse1 (void); |
629 | static bool gate_dse2 (void); | |
6fb5fa3c DB |
630 | |
631 | \f | |
632 | /*---------------------------------------------------------------------------- | |
633 | Zeroth step. | |
634 | ||
b8698a0f | 635 | Initialization. |
6fb5fa3c DB |
636 | ----------------------------------------------------------------------------*/ |
637 | ||
6fb5fa3c DB |
638 | |
639 | /* Find the entry associated with ALIAS_SET. */ | |
640 | ||
641 | static struct clear_alias_mode_holder * | |
4862826d | 642 | clear_alias_set_lookup (alias_set_type alias_set) |
6fb5fa3c DB |
643 | { |
644 | struct clear_alias_mode_holder tmp_holder; | |
645 | void **slot; | |
b8698a0f | 646 | |
6fb5fa3c DB |
647 | tmp_holder.alias_set = alias_set; |
648 | slot = htab_find_slot (clear_alias_mode_table, &tmp_holder, NO_INSERT); | |
649 | gcc_assert (*slot); | |
b8698a0f | 650 | |
f883e0a7 | 651 | return (struct clear_alias_mode_holder *) *slot; |
6fb5fa3c DB |
652 | } |
653 | ||
654 | ||
655 | /* Hashtable callbacks for maintaining the "bases" field of | |
656 | store_group_info, given that the addresses are function invariants. */ | |
657 | ||
658 | static int | |
659 | invariant_group_base_eq (const void *p1, const void *p2) | |
660 | { | |
5f754896 KG |
661 | const_group_info_t gi1 = (const_group_info_t) p1; |
662 | const_group_info_t gi2 = (const_group_info_t) p2; | |
6fb5fa3c DB |
663 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
664 | } | |
665 | ||
666 | ||
667 | static hashval_t | |
668 | invariant_group_base_hash (const void *p) | |
669 | { | |
5f754896 | 670 | const_group_info_t gi = (const_group_info_t) p; |
6fb5fa3c DB |
671 | int do_not_record; |
672 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
673 | } | |
674 | ||
675 | ||
676 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
677 | ||
678 | static group_info_t | |
679 | get_group_info (rtx base) | |
680 | { | |
b8698a0f L |
681 | struct group_info tmp_gi; |
682 | group_info_t gi; | |
6fb5fa3c DB |
683 | void **slot; |
684 | ||
685 | if (base) | |
686 | { | |
687 | /* Find the store_base_info structure for BASE, creating a new one | |
688 | if necessary. */ | |
689 | tmp_gi.rtx_base = base; | |
690 | slot = htab_find_slot (rtx_group_table, &tmp_gi, INSERT); | |
691 | gi = (group_info_t) *slot; | |
692 | } | |
693 | else | |
694 | { | |
695 | if (!clear_alias_group) | |
696 | { | |
f883e0a7 KG |
697 | clear_alias_group = gi = |
698 | (group_info_t) pool_alloc (rtx_group_info_pool); | |
6fb5fa3c DB |
699 | memset (gi, 0, sizeof (struct group_info)); |
700 | gi->id = rtx_group_next_id++; | |
3f9b14ff SB |
701 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
702 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
703 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
704 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
705 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
706 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
707 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c DB |
708 | gi->process_globally = false; |
709 | gi->offset_map_size_n = 0; | |
710 | gi->offset_map_size_p = 0; | |
711 | gi->offset_map_n = NULL; | |
712 | gi->offset_map_p = NULL; | |
713 | VEC_safe_push (group_info_t, heap, rtx_group_vec, gi); | |
714 | } | |
715 | return clear_alias_group; | |
716 | } | |
717 | ||
718 | if (gi == NULL) | |
719 | { | |
f883e0a7 | 720 | *slot = gi = (group_info_t) pool_alloc (rtx_group_info_pool); |
6fb5fa3c DB |
721 | gi->rtx_base = base; |
722 | gi->id = rtx_group_next_id++; | |
d32f725a | 723 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
6216f94e | 724 | gi->canon_base_addr = canon_rtx (base); |
3f9b14ff SB |
725 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
726 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
727 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
728 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
729 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
730 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
731 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 732 | gi->process_globally = false; |
b8698a0f | 733 | gi->frame_related = |
6fb5fa3c DB |
734 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
735 | gi->offset_map_size_n = 0; | |
736 | gi->offset_map_size_p = 0; | |
737 | gi->offset_map_n = NULL; | |
738 | gi->offset_map_p = NULL; | |
739 | VEC_safe_push (group_info_t, heap, rtx_group_vec, gi); | |
740 | } | |
741 | ||
742 | return gi; | |
743 | } | |
744 | ||
745 | ||
746 | /* Initialization of data structures. */ | |
747 | ||
748 | static void | |
749 | dse_step0 (void) | |
750 | { | |
751 | locally_deleted = 0; | |
752 | globally_deleted = 0; | |
753 | spill_deleted = 0; | |
754 | ||
3f9b14ff SB |
755 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
756 | gcc_obstack_init (&dse_obstack); | |
757 | ||
758 | scratch = BITMAP_ALLOC (®_obstack); | |
759 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c DB |
760 | |
761 | rtx_store_info_pool | |
b8698a0f | 762 | = create_alloc_pool ("rtx_store_info_pool", |
6fb5fa3c DB |
763 | sizeof (struct store_info), 100); |
764 | read_info_pool | |
b8698a0f | 765 | = create_alloc_pool ("read_info_pool", |
6fb5fa3c DB |
766 | sizeof (struct read_info), 100); |
767 | insn_info_pool | |
b8698a0f | 768 | = create_alloc_pool ("insn_info_pool", |
6fb5fa3c DB |
769 | sizeof (struct insn_info), 100); |
770 | bb_info_pool | |
b8698a0f | 771 | = create_alloc_pool ("bb_info_pool", |
6fb5fa3c DB |
772 | sizeof (struct bb_info), 100); |
773 | rtx_group_info_pool | |
b8698a0f | 774 | = create_alloc_pool ("rtx_group_info_pool", |
6fb5fa3c DB |
775 | sizeof (struct group_info), 100); |
776 | deferred_change_pool | |
b8698a0f | 777 | = create_alloc_pool ("deferred_change_pool", |
6fb5fa3c DB |
778 | sizeof (struct deferred_change), 10); |
779 | ||
780 | rtx_group_table = htab_create (11, invariant_group_base_hash, | |
781 | invariant_group_base_eq, NULL); | |
782 | ||
3f9b14ff | 783 | bb_table = XNEWVEC (bb_info_t, last_basic_block); |
6fb5fa3c DB |
784 | rtx_group_next_id = 0; |
785 | ||
e3b5732b | 786 | stores_off_frame_dead_at_return = !cfun->stdarg; |
6fb5fa3c DB |
787 | |
788 | init_alias_analysis (); | |
b8698a0f | 789 | |
6fb5fa3c DB |
790 | if (clear_alias_sets) |
791 | clear_alias_group = get_group_info (NULL); | |
792 | else | |
793 | clear_alias_group = NULL; | |
794 | } | |
795 | ||
796 | ||
797 | \f | |
798 | /*---------------------------------------------------------------------------- | |
799 | First step. | |
800 | ||
801 | Scan all of the insns. Any random ordering of the blocks is fine. | |
0d52bcc1 | 802 | Each block is scanned in forward order to accommodate cselib which |
6fb5fa3c DB |
803 | is used to remove stores with non-constant bases. |
804 | ----------------------------------------------------------------------------*/ | |
805 | ||
806 | /* Delete all of the store_info recs from INSN_INFO. */ | |
807 | ||
b8698a0f | 808 | static void |
6fb5fa3c DB |
809 | free_store_info (insn_info_t insn_info) |
810 | { | |
811 | store_info_t store_info = insn_info->store_rec; | |
812 | while (store_info) | |
813 | { | |
814 | store_info_t next = store_info->next; | |
8dd5516b | 815 | if (store_info->is_large) |
dc491a25 | 816 | BITMAP_FREE (store_info->positions_needed.large.bmap); |
6fb5fa3c DB |
817 | if (store_info->cse_base) |
818 | pool_free (cse_store_info_pool, store_info); | |
819 | else | |
820 | pool_free (rtx_store_info_pool, store_info); | |
821 | store_info = next; | |
822 | } | |
823 | ||
824 | insn_info->cannot_delete = true; | |
825 | insn_info->contains_cselib_groups = false; | |
826 | insn_info->store_rec = NULL; | |
827 | } | |
828 | ||
9e582b1d JR |
829 | typedef struct |
830 | { | |
831 | rtx first, current; | |
832 | regset fixed_regs_live; | |
833 | bool failure; | |
834 | } note_add_store_info; | |
835 | ||
836 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
837 | Check if a register is clobbered which is live afterwards. */ | |
838 | ||
839 | static void | |
840 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
841 | { | |
842 | rtx insn; | |
843 | note_add_store_info *info = (note_add_store_info *) data; | |
844 | int r, n; | |
845 | ||
846 | if (!REG_P (loc)) | |
847 | return; | |
848 | ||
849 | /* If this register is referenced by the current or an earlier insn, | |
850 | that's OK. E.g. this applies to the register that is being incremented | |
851 | with this addition. */ | |
852 | for (insn = info->first; | |
853 | insn != NEXT_INSN (info->current); | |
854 | insn = NEXT_INSN (insn)) | |
855 | if (reg_referenced_p (loc, PATTERN (insn))) | |
856 | return; | |
857 | ||
858 | /* If we come here, we have a clobber of a register that's only OK | |
859 | if that register is not live. If we don't have liveness information | |
860 | available, fail now. */ | |
861 | if (!info->fixed_regs_live) | |
862 | { | |
863 | info->failure = true; | |
864 | return; | |
865 | } | |
866 | /* Now check if this is a live fixed register. */ | |
867 | r = REGNO (loc); | |
868 | n = hard_regno_nregs[r][GET_MODE (loc)]; | |
869 | while (--n >= 0) | |
870 | if (REGNO_REG_SET_P (info->fixed_regs_live, r+n)) | |
871 | info->failure = true; | |
872 | } | |
873 | ||
4deef538 AO |
874 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
875 | SRC + SRCOFF before insn ARG. */ | |
6fb5fa3c DB |
876 | |
877 | static int | |
4deef538 AO |
878 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
879 | rtx op ATTRIBUTE_UNUSED, | |
880 | rtx dest, rtx src, rtx srcoff, void *arg) | |
6fb5fa3c | 881 | { |
9e582b1d JR |
882 | insn_info_t insn_info = (insn_info_t) arg; |
883 | rtx insn = insn_info->insn, new_insn, cur; | |
884 | note_add_store_info info; | |
b8698a0f | 885 | |
4deef538 AO |
886 | /* We can reuse all operands without copying, because we are about |
887 | to delete the insn that contained it. */ | |
9e582b1d | 888 | if (srcoff) |
ed079c4b JJ |
889 | { |
890 | start_sequence (); | |
891 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
892 | new_insn = get_insns (); | |
893 | end_sequence (); | |
894 | } | |
9e582b1d JR |
895 | else |
896 | new_insn = gen_move_insn (dest, src); | |
897 | info.first = new_insn; | |
898 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
899 | info.failure = false; | |
900 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
901 | { | |
902 | info.current = cur; | |
903 | note_stores (PATTERN (cur), note_add_store, &info); | |
904 | } | |
6fb5fa3c | 905 | |
9e582b1d JR |
906 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
907 | return it immediately, communicating the failure to its caller. */ | |
908 | if (info.failure) | |
909 | return 1; | |
910 | ||
911 | emit_insn_before (new_insn, insn); | |
6fb5fa3c | 912 | |
4deef538 | 913 | return -1; |
6fb5fa3c DB |
914 | } |
915 | ||
9e582b1d JR |
916 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
917 | is there, is split into a separate insn. | |
918 | Return true on success (or if there was nothing to do), false on failure. */ | |
6fb5fa3c | 919 | |
9e582b1d JR |
920 | static bool |
921 | check_for_inc_dec_1 (insn_info_t insn_info) | |
6fb5fa3c | 922 | { |
9e582b1d | 923 | rtx insn = insn_info->insn; |
6fb5fa3c DB |
924 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
925 | if (note) | |
9e582b1d JR |
926 | return for_each_inc_dec (&insn, emit_inc_dec_insn_before, insn_info) == 0; |
927 | return true; | |
6fb5fa3c DB |
928 | } |
929 | ||
930 | ||
9e582b1d JR |
931 | /* Entry point for postreload. If you work on reload_cse, or you need this |
932 | anywhere else, consider if you can provide register liveness information | |
933 | and add a parameter to this function so that it can be passed down in | |
934 | insn_info.fixed_regs_live. */ | |
935 | bool | |
936 | check_for_inc_dec (rtx insn) | |
937 | { | |
938 | struct insn_info insn_info; | |
939 | rtx note; | |
940 | ||
941 | insn_info.insn = insn; | |
942 | insn_info.fixed_regs_live = NULL; | |
943 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
944 | if (note) | |
945 | return for_each_inc_dec (&insn, emit_inc_dec_insn_before, &insn_info) == 0; | |
946 | return true; | |
947 | } | |
948 | ||
b8698a0f | 949 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
6fb5fa3c DB |
950 | |
951 | static void | |
952 | delete_dead_store_insn (insn_info_t insn_info) | |
953 | { | |
954 | read_info_t read_info; | |
955 | ||
956 | if (!dbg_cnt (dse)) | |
957 | return; | |
958 | ||
9e582b1d JR |
959 | if (!check_for_inc_dec_1 (insn_info)) |
960 | return; | |
6fb5fa3c DB |
961 | if (dump_file) |
962 | { | |
b8698a0f | 963 | fprintf (dump_file, "Locally deleting insn %d ", |
6fb5fa3c DB |
964 | INSN_UID (insn_info->insn)); |
965 | if (insn_info->store_rec->alias_set) | |
b8698a0f | 966 | fprintf (dump_file, "alias set %d\n", |
4862826d | 967 | (int) insn_info->store_rec->alias_set); |
6fb5fa3c DB |
968 | else |
969 | fprintf (dump_file, "\n"); | |
970 | } | |
971 | ||
972 | free_store_info (insn_info); | |
973 | read_info = insn_info->read_rec; | |
b8698a0f | 974 | |
6fb5fa3c DB |
975 | while (read_info) |
976 | { | |
977 | read_info_t next = read_info->next; | |
978 | pool_free (read_info_pool, read_info); | |
979 | read_info = next; | |
980 | } | |
981 | insn_info->read_rec = NULL; | |
982 | ||
983 | delete_insn (insn_info->insn); | |
984 | locally_deleted++; | |
985 | insn_info->insn = NULL; | |
986 | ||
987 | insn_info->wild_read = false; | |
988 | } | |
989 | ||
d26c7090 ER |
990 | /* Check if EXPR can possibly escape the current function scope. */ |
991 | static bool | |
992 | can_escape (tree expr) | |
993 | { | |
994 | tree base; | |
995 | if (!expr) | |
996 | return true; | |
997 | base = get_base_address (expr); | |
998 | if (DECL_P (base) | |
999 | && !may_be_aliased (base)) | |
1000 | return false; | |
1001 | return true; | |
1002 | } | |
6fb5fa3c DB |
1003 | |
1004 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
1005 | OFFSET and WIDTH. */ | |
1006 | ||
1007 | static void | |
d26c7090 ER |
1008 | set_usage_bits (group_info_t group, HOST_WIDE_INT offset, HOST_WIDE_INT width, |
1009 | tree expr) | |
6fb5fa3c DB |
1010 | { |
1011 | HOST_WIDE_INT i; | |
d26c7090 | 1012 | bool expr_escapes = can_escape (expr); |
8dd5516b | 1013 | if (offset > -MAX_OFFSET && offset + width < MAX_OFFSET) |
6fb5fa3c DB |
1014 | for (i=offset; i<offset+width; i++) |
1015 | { | |
1016 | bitmap store1; | |
1017 | bitmap store2; | |
d26c7090 | 1018 | bitmap escaped; |
6fb5fa3c DB |
1019 | int ai; |
1020 | if (i < 0) | |
1021 | { | |
1022 | store1 = group->store1_n; | |
1023 | store2 = group->store2_n; | |
d26c7090 | 1024 | escaped = group->escaped_n; |
6fb5fa3c DB |
1025 | ai = -i; |
1026 | } | |
1027 | else | |
1028 | { | |
1029 | store1 = group->store1_p; | |
1030 | store2 = group->store2_p; | |
d26c7090 | 1031 | escaped = group->escaped_p; |
6fb5fa3c DB |
1032 | ai = i; |
1033 | } | |
b8698a0f | 1034 | |
fcaa4ca4 | 1035 | if (!bitmap_set_bit (store1, ai)) |
6fb5fa3c | 1036 | bitmap_set_bit (store2, ai); |
b8698a0f | 1037 | else |
6fb5fa3c | 1038 | { |
6fb5fa3c DB |
1039 | if (i < 0) |
1040 | { | |
1041 | if (group->offset_map_size_n < ai) | |
1042 | group->offset_map_size_n = ai; | |
1043 | } | |
1044 | else | |
1045 | { | |
1046 | if (group->offset_map_size_p < ai) | |
1047 | group->offset_map_size_p = ai; | |
1048 | } | |
1049 | } | |
d26c7090 ER |
1050 | if (expr_escapes) |
1051 | bitmap_set_bit (escaped, ai); | |
6fb5fa3c DB |
1052 | } |
1053 | } | |
1054 | ||
d26c7090 ER |
1055 | static void |
1056 | reset_active_stores (void) | |
1057 | { | |
1058 | active_local_stores = NULL; | |
1059 | active_local_stores_len = 0; | |
1060 | } | |
6fb5fa3c | 1061 | |
d26c7090 | 1062 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
6fb5fa3c DB |
1063 | |
1064 | static void | |
d26c7090 | 1065 | free_read_records (bb_info_t bb_info) |
6fb5fa3c DB |
1066 | { |
1067 | insn_info_t insn_info = bb_info->last_insn; | |
1068 | read_info_t *ptr = &insn_info->read_rec; | |
6fb5fa3c DB |
1069 | while (*ptr) |
1070 | { | |
1071 | read_info_t next = (*ptr)->next; | |
4862826d | 1072 | if ((*ptr)->alias_set == 0) |
6fb5fa3c DB |
1073 | { |
1074 | pool_free (read_info_pool, *ptr); | |
1075 | *ptr = next; | |
d26c7090 | 1076 | } |
b8698a0f | 1077 | else |
d26c7090 | 1078 | ptr = &(*ptr)->next; |
6fb5fa3c | 1079 | } |
d26c7090 ER |
1080 | } |
1081 | ||
1082 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1083 | ||
1084 | static void | |
1085 | add_wild_read (bb_info_t bb_info) | |
1086 | { | |
1087 | insn_info_t insn_info = bb_info->last_insn; | |
6fb5fa3c | 1088 | insn_info->wild_read = true; |
d26c7090 ER |
1089 | free_read_records (bb_info); |
1090 | reset_active_stores (); | |
6fb5fa3c DB |
1091 | } |
1092 | ||
d26c7090 ER |
1093 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1094 | non-frame locations. */ | |
1095 | ||
1096 | static void | |
1097 | add_non_frame_wild_read (bb_info_t bb_info) | |
1098 | { | |
1099 | insn_info_t insn_info = bb_info->last_insn; | |
1100 | insn_info->non_frame_wild_read = true; | |
1101 | free_read_records (bb_info); | |
1102 | reset_active_stores (); | |
1103 | } | |
6fb5fa3c | 1104 | |
50f0f366 EB |
1105 | /* Return true if X is a constant or one of the registers that behave |
1106 | as a constant over the life of a function. This is equivalent to | |
1107 | !rtx_varies_p for memory addresses. */ | |
6fb5fa3c DB |
1108 | |
1109 | static bool | |
1110 | const_or_frame_p (rtx x) | |
1111 | { | |
1112 | switch (GET_CODE (x)) | |
1113 | { | |
6fb5fa3c DB |
1114 | case CONST: |
1115 | case CONST_INT: | |
1116 | case CONST_DOUBLE: | |
1117 | case CONST_VECTOR: | |
1118 | case SYMBOL_REF: | |
1119 | case LABEL_REF: | |
1120 | return true; | |
1121 | ||
1122 | case REG: | |
1123 | /* Note that we have to test for the actual rtx used for the frame | |
1124 | and arg pointers and not just the register number in case we have | |
1125 | eliminated the frame and/or arg pointer and are using it | |
1126 | for pseudos. */ | |
1127 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1128 | /* The arg pointer varies if it is not a fixed register. */ | |
1129 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1130 | || x == pic_offset_table_rtx) | |
1131 | return true; | |
1132 | return false; | |
1133 | ||
1134 | default: | |
1135 | return false; | |
1136 | } | |
1137 | } | |
1138 | ||
b8698a0f L |
1139 | /* Take all reasonable action to put the address of MEM into the form |
1140 | that we can do analysis on. | |
6fb5fa3c DB |
1141 | |
1142 | The gold standard is to get the address into the form: address + | |
1143 | OFFSET where address is something that rtx_varies_p considers a | |
1144 | constant. When we can get the address in this form, we can do | |
1145 | global analysis on it. Note that for constant bases, address is | |
1146 | not actually returned, only the group_id. The address can be | |
1147 | obtained from that. | |
1148 | ||
1149 | If that fails, we try cselib to get a value we can at least use | |
b8698a0f L |
1150 | locally. If that fails we return false. |
1151 | ||
6fb5fa3c DB |
1152 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1153 | group for non_varying bases. | |
1154 | ||
1155 | FOR_READ is true if this is a mem read and false if not. */ | |
1156 | ||
1157 | static bool | |
1158 | canon_address (rtx mem, | |
4862826d | 1159 | alias_set_type *alias_set_out, |
6fb5fa3c | 1160 | int *group_id, |
b8698a0f | 1161 | HOST_WIDE_INT *offset, |
6fb5fa3c DB |
1162 | cselib_val **base) |
1163 | { | |
372d6395 | 1164 | enum machine_mode address_mode = get_address_mode (mem); |
6fb5fa3c DB |
1165 | rtx mem_address = XEXP (mem, 0); |
1166 | rtx expanded_address, address; | |
403c7520 JJ |
1167 | int expanded; |
1168 | ||
6fb5fa3c DB |
1169 | /* Make sure that cselib is has initialized all of the operands of |
1170 | the address before asking it to do the subst. */ | |
1171 | ||
1172 | if (clear_alias_sets) | |
1173 | { | |
1174 | /* If this is a spill, do not do any further processing. */ | |
4862826d | 1175 | alias_set_type alias_set = MEM_ALIAS_SET (mem); |
6fb5fa3c | 1176 | if (dump_file) |
4862826d | 1177 | fprintf (dump_file, "found alias set %d\n", (int) alias_set); |
6fb5fa3c DB |
1178 | if (bitmap_bit_p (clear_alias_sets, alias_set)) |
1179 | { | |
b8698a0f | 1180 | struct clear_alias_mode_holder *entry |
6fb5fa3c DB |
1181 | = clear_alias_set_lookup (alias_set); |
1182 | ||
1183 | /* If the modes do not match, we cannot process this set. */ | |
1184 | if (entry->mode != GET_MODE (mem)) | |
1185 | { | |
1186 | if (dump_file) | |
b8698a0f L |
1187 | fprintf (dump_file, |
1188 | "disqualifying alias set %d, (%s) != (%s)\n", | |
1189 | (int) alias_set, GET_MODE_NAME (entry->mode), | |
6fb5fa3c | 1190 | GET_MODE_NAME (GET_MODE (mem))); |
b8698a0f | 1191 | |
6fb5fa3c DB |
1192 | bitmap_set_bit (disqualified_clear_alias_sets, alias_set); |
1193 | return false; | |
1194 | } | |
1195 | ||
1196 | *alias_set_out = alias_set; | |
1197 | *group_id = clear_alias_group->id; | |
1198 | return true; | |
1199 | } | |
1200 | } | |
1201 | ||
1202 | *alias_set_out = 0; | |
1203 | ||
4deef538 | 1204 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
6fb5fa3c DB |
1205 | |
1206 | if (dump_file) | |
1207 | { | |
1208 | fprintf (dump_file, " mem: "); | |
1209 | print_inline_rtx (dump_file, mem_address, 0); | |
1210 | fprintf (dump_file, "\n"); | |
1211 | } | |
1212 | ||
403c7520 JJ |
1213 | /* First see if just canon_rtx (mem_address) is const or frame, |
1214 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1215 | address = NULL_RTX; | |
1216 | for (expanded = 0; expanded < 2; expanded++) | |
1217 | { | |
1218 | if (expanded) | |
1219 | { | |
1220 | /* Use cselib to replace all of the reg references with the full | |
b8698a0f | 1221 | expression. This will take care of the case where we have |
6fb5fa3c | 1222 | |
403c7520 JJ |
1223 | r_x = base + offset; |
1224 | val = *r_x; | |
b8698a0f L |
1225 | |
1226 | by making it into | |
6fb5fa3c | 1227 | |
403c7520 | 1228 | val = *(base + offset); */ |
6fb5fa3c | 1229 | |
403c7520 JJ |
1230 | expanded_address = cselib_expand_value_rtx (mem_address, |
1231 | scratch, 5); | |
6fb5fa3c | 1232 | |
403c7520 JJ |
1233 | /* If this fails, just go with the address from first |
1234 | iteration. */ | |
1235 | if (!expanded_address) | |
1236 | break; | |
1237 | } | |
1238 | else | |
1239 | expanded_address = mem_address; | |
6fb5fa3c | 1240 | |
403c7520 JJ |
1241 | /* Split the address into canonical BASE + OFFSET terms. */ |
1242 | address = canon_rtx (expanded_address); | |
6fb5fa3c | 1243 | |
403c7520 | 1244 | *offset = 0; |
6fb5fa3c | 1245 | |
403c7520 JJ |
1246 | if (dump_file) |
1247 | { | |
1248 | if (expanded) | |
1249 | { | |
1250 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1251 | print_inline_rtx (dump_file, expanded_address, 0); | |
1252 | fprintf (dump_file, "\n"); | |
1253 | } | |
6fb5fa3c | 1254 | |
403c7520 JJ |
1255 | fprintf (dump_file, "\n after canon_rtx address: "); |
1256 | print_inline_rtx (dump_file, address, 0); | |
1257 | fprintf (dump_file, "\n"); | |
1258 | } | |
6fb5fa3c | 1259 | |
403c7520 JJ |
1260 | if (GET_CODE (address) == CONST) |
1261 | address = XEXP (address, 0); | |
6fb5fa3c | 1262 | |
403c7520 JJ |
1263 | if (GET_CODE (address) == PLUS |
1264 | && CONST_INT_P (XEXP (address, 1))) | |
1265 | { | |
1266 | *offset = INTVAL (XEXP (address, 1)); | |
1267 | address = XEXP (address, 0); | |
1268 | } | |
6fb5fa3c | 1269 | |
09e881c9 BE |
1270 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1271 | && const_or_frame_p (address)) | |
6fb5fa3c | 1272 | { |
403c7520 JJ |
1273 | group_info_t group = get_group_info (address); |
1274 | ||
6fb5fa3c | 1275 | if (dump_file) |
403c7520 JJ |
1276 | fprintf (dump_file, " gid=%d offset=%d \n", |
1277 | group->id, (int)*offset); | |
1278 | *base = NULL; | |
1279 | *group_id = group->id; | |
1280 | return true; | |
6fb5fa3c | 1281 | } |
403c7520 JJ |
1282 | } |
1283 | ||
4deef538 | 1284 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
403c7520 JJ |
1285 | *group_id = -1; |
1286 | ||
1287 | if (*base == NULL) | |
1288 | { | |
6fb5fa3c | 1289 | if (dump_file) |
403c7520 JJ |
1290 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1291 | return false; | |
6fb5fa3c | 1292 | } |
403c7520 | 1293 | if (dump_file) |
5440c0e7 AO |
1294 | fprintf (dump_file, " varying cselib base=%u:%u offset = %d\n", |
1295 | (*base)->uid, (*base)->hash, (int)*offset); | |
6fb5fa3c DB |
1296 | return true; |
1297 | } | |
1298 | ||
1299 | ||
1300 | /* Clear the rhs field from the active_local_stores array. */ | |
1301 | ||
1302 | static void | |
1303 | clear_rhs_from_active_local_stores (void) | |
1304 | { | |
1305 | insn_info_t ptr = active_local_stores; | |
1306 | ||
1307 | while (ptr) | |
1308 | { | |
1309 | store_info_t store_info = ptr->store_rec; | |
1310 | /* Skip the clobbers. */ | |
1311 | while (!store_info->is_set) | |
1312 | store_info = store_info->next; | |
1313 | ||
1314 | store_info->rhs = NULL; | |
8dd5516b | 1315 | store_info->const_rhs = NULL; |
6fb5fa3c DB |
1316 | |
1317 | ptr = ptr->next_local_store; | |
1318 | } | |
1319 | } | |
1320 | ||
1321 | ||
8dd5516b JJ |
1322 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1323 | ||
1324 | static inline void | |
1325 | set_position_unneeded (store_info_t s_info, int pos) | |
1326 | { | |
1327 | if (__builtin_expect (s_info->is_large, false)) | |
1328 | { | |
fcaa4ca4 NF |
1329 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1330 | s_info->positions_needed.large.count++; | |
8dd5516b JJ |
1331 | } |
1332 | else | |
1333 | s_info->positions_needed.small_bitmask | |
1334 | &= ~(((unsigned HOST_WIDE_INT) 1) << pos); | |
1335 | } | |
1336 | ||
1337 | /* Mark the whole store S_INFO as unneeded. */ | |
1338 | ||
1339 | static inline void | |
1340 | set_all_positions_unneeded (store_info_t s_info) | |
1341 | { | |
1342 | if (__builtin_expect (s_info->is_large, false)) | |
1343 | { | |
1344 | int pos, end = s_info->end - s_info->begin; | |
1345 | for (pos = 0; pos < end; pos++) | |
dc491a25 | 1346 | bitmap_set_bit (s_info->positions_needed.large.bmap, pos); |
8dd5516b JJ |
1347 | s_info->positions_needed.large.count = end; |
1348 | } | |
1349 | else | |
1350 | s_info->positions_needed.small_bitmask = (unsigned HOST_WIDE_INT) 0; | |
1351 | } | |
1352 | ||
1353 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1354 | ||
1355 | static inline bool | |
1356 | any_positions_needed_p (store_info_t s_info) | |
1357 | { | |
1358 | if (__builtin_expect (s_info->is_large, false)) | |
1359 | return (s_info->positions_needed.large.count | |
1360 | < s_info->end - s_info->begin); | |
1361 | else | |
1362 | return (s_info->positions_needed.small_bitmask | |
1363 | != (unsigned HOST_WIDE_INT) 0); | |
1364 | } | |
1365 | ||
1366 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
1367 | store are needed. */ | |
1368 | ||
1369 | static inline bool | |
1370 | all_positions_needed_p (store_info_t s_info, int start, int width) | |
1371 | { | |
1372 | if (__builtin_expect (s_info->is_large, false)) | |
1373 | { | |
1374 | int end = start + width; | |
1375 | while (start < end) | |
dc491a25 | 1376 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, start++)) |
8dd5516b JJ |
1377 | return false; |
1378 | return true; | |
1379 | } | |
1380 | else | |
1381 | { | |
1382 | unsigned HOST_WIDE_INT mask = lowpart_bitmask (width) << start; | |
1383 | return (s_info->positions_needed.small_bitmask & mask) == mask; | |
1384 | } | |
1385 | } | |
1386 | ||
1387 | ||
1388 | static rtx get_stored_val (store_info_t, enum machine_mode, HOST_WIDE_INT, | |
1389 | HOST_WIDE_INT, basic_block, bool); | |
1390 | ||
1391 | ||
6fb5fa3c DB |
1392 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1393 | there is a candidate store, after adding it to the appropriate | |
1394 | local store group if so. */ | |
1395 | ||
1396 | static int | |
1397 | record_store (rtx body, bb_info_t bb_info) | |
1398 | { | |
6216f94e | 1399 | rtx mem, rhs, const_rhs, mem_addr; |
6fb5fa3c DB |
1400 | HOST_WIDE_INT offset = 0; |
1401 | HOST_WIDE_INT width = 0; | |
4862826d | 1402 | alias_set_type spill_alias_set; |
6fb5fa3c DB |
1403 | insn_info_t insn_info = bb_info->last_insn; |
1404 | store_info_t store_info = NULL; | |
1405 | int group_id; | |
1406 | cselib_val *base = NULL; | |
8dd5516b | 1407 | insn_info_t ptr, last, redundant_reason; |
6fb5fa3c DB |
1408 | bool store_is_unused; |
1409 | ||
1410 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1411 | return 0; | |
1412 | ||
8dd5516b JJ |
1413 | mem = SET_DEST (body); |
1414 | ||
6fb5fa3c DB |
1415 | /* If this is not used, then this cannot be used to keep the insn |
1416 | from being deleted. On the other hand, it does provide something | |
1417 | that can be used to prove that another store is dead. */ | |
1418 | store_is_unused | |
8dd5516b | 1419 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
6fb5fa3c DB |
1420 | |
1421 | /* Check whether that value is a suitable memory location. */ | |
6fb5fa3c DB |
1422 | if (!MEM_P (mem)) |
1423 | { | |
1424 | /* If the set or clobber is unused, then it does not effect our | |
1425 | ability to get rid of the entire insn. */ | |
1426 | if (!store_is_unused) | |
1427 | insn_info->cannot_delete = true; | |
1428 | return 0; | |
1429 | } | |
1430 | ||
1431 | /* At this point we know mem is a mem. */ | |
1432 | if (GET_MODE (mem) == BLKmode) | |
1433 | { | |
1434 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
1435 | { | |
b8698a0f | 1436 | if (dump_file) |
6fb5fa3c DB |
1437 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1438 | add_wild_read (bb_info); | |
1439 | insn_info->cannot_delete = true; | |
8dd5516b | 1440 | return 0; |
6fb5fa3c | 1441 | } |
8dd5516b JJ |
1442 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1443 | as memset (addr, 0, 36); */ | |
f5541398 RS |
1444 | else if (!MEM_SIZE_KNOWN_P (mem) |
1445 | || MEM_SIZE (mem) <= 0 | |
1446 | || MEM_SIZE (mem) > MAX_OFFSET | |
8dd5516b | 1447 | || GET_CODE (body) != SET |
8dd5516b | 1448 | || !CONST_INT_P (SET_SRC (body))) |
6fb5fa3c | 1449 | { |
8dd5516b JJ |
1450 | if (!store_is_unused) |
1451 | { | |
1452 | /* If the set or clobber is unused, then it does not effect our | |
1453 | ability to get rid of the entire insn. */ | |
1454 | insn_info->cannot_delete = true; | |
1455 | clear_rhs_from_active_local_stores (); | |
1456 | } | |
1457 | return 0; | |
6fb5fa3c | 1458 | } |
6fb5fa3c DB |
1459 | } |
1460 | ||
1461 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1462 | if (MEM_VOLATILE_P (mem)) | |
8dd5516b | 1463 | insn_info->cannot_delete = true; |
6fb5fa3c DB |
1464 | |
1465 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
1466 | { | |
1467 | clear_rhs_from_active_local_stores (); | |
1468 | return 0; | |
1469 | } | |
1470 | ||
8dd5516b | 1471 | if (GET_MODE (mem) == BLKmode) |
f5541398 | 1472 | width = MEM_SIZE (mem); |
8dd5516b JJ |
1473 | else |
1474 | { | |
1475 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
1476 | gcc_assert ((unsigned) width <= HOST_BITS_PER_WIDE_INT); | |
1477 | } | |
6fb5fa3c DB |
1478 | |
1479 | if (spill_alias_set) | |
1480 | { | |
1481 | bitmap store1 = clear_alias_group->store1_p; | |
1482 | bitmap store2 = clear_alias_group->store2_p; | |
8dd5516b JJ |
1483 | |
1484 | gcc_assert (GET_MODE (mem) != BLKmode); | |
b8698a0f | 1485 | |
fcaa4ca4 | 1486 | if (!bitmap_set_bit (store1, spill_alias_set)) |
6fb5fa3c | 1487 | bitmap_set_bit (store2, spill_alias_set); |
b8698a0f | 1488 | |
6fb5fa3c DB |
1489 | if (clear_alias_group->offset_map_size_p < spill_alias_set) |
1490 | clear_alias_group->offset_map_size_p = spill_alias_set; | |
b8698a0f | 1491 | |
f883e0a7 | 1492 | store_info = (store_info_t) pool_alloc (rtx_store_info_pool); |
6fb5fa3c DB |
1493 | |
1494 | if (dump_file) | |
1495 | fprintf (dump_file, " processing spill store %d(%s)\n", | |
4862826d | 1496 | (int) spill_alias_set, GET_MODE_NAME (GET_MODE (mem))); |
6fb5fa3c DB |
1497 | } |
1498 | else if (group_id >= 0) | |
1499 | { | |
1500 | /* In the restrictive case where the base is a constant or the | |
1501 | frame pointer we can do global analysis. */ | |
b8698a0f L |
1502 | |
1503 | group_info_t group | |
6fb5fa3c | 1504 | = VEC_index (group_info_t, rtx_group_vec, group_id); |
d26c7090 | 1505 | tree expr = MEM_EXPR (mem); |
b8698a0f | 1506 | |
f883e0a7 | 1507 | store_info = (store_info_t) pool_alloc (rtx_store_info_pool); |
d26c7090 | 1508 | set_usage_bits (group, offset, width, expr); |
6fb5fa3c DB |
1509 | |
1510 | if (dump_file) | |
1511 | fprintf (dump_file, " processing const base store gid=%d[%d..%d)\n", | |
1512 | group_id, (int)offset, (int)(offset+width)); | |
1513 | } | |
1514 | else | |
1515 | { | |
9e412ca3 | 1516 | if (may_be_sp_based_p (XEXP (mem, 0))) |
50f0f366 | 1517 | insn_info->stack_pointer_based = true; |
6fb5fa3c | 1518 | insn_info->contains_cselib_groups = true; |
50f0f366 | 1519 | |
f883e0a7 | 1520 | store_info = (store_info_t) pool_alloc (cse_store_info_pool); |
6fb5fa3c DB |
1521 | group_id = -1; |
1522 | ||
1523 | if (dump_file) | |
1524 | fprintf (dump_file, " processing cselib store [%d..%d)\n", | |
1525 | (int)offset, (int)(offset+width)); | |
1526 | } | |
1527 | ||
8dd5516b JJ |
1528 | const_rhs = rhs = NULL_RTX; |
1529 | if (GET_CODE (body) == SET | |
1530 | /* No place to keep the value after ra. */ | |
1531 | && !reload_completed | |
1532 | && (REG_P (SET_SRC (body)) | |
1533 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1534 | || CONSTANT_P (SET_SRC (body))) | |
1535 | && !MEM_VOLATILE_P (mem) | |
1536 | /* Sometimes the store and reload is used for truncation and | |
1537 | rounding. */ | |
1538 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1539 | { | |
1540 | rhs = SET_SRC (body); | |
1541 | if (CONSTANT_P (rhs)) | |
1542 | const_rhs = rhs; | |
1543 | else if (body == PATTERN (insn_info->insn)) | |
1544 | { | |
1545 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1546 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1547 | const_rhs = XEXP (tem, 0); | |
1548 | } | |
1549 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1550 | { | |
1551 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1552 | ||
1553 | if (tem && CONSTANT_P (tem)) | |
1554 | const_rhs = tem; | |
1555 | } | |
1556 | } | |
1557 | ||
6fb5fa3c DB |
1558 | /* Check to see if this stores causes some other stores to be |
1559 | dead. */ | |
1560 | ptr = active_local_stores; | |
1561 | last = NULL; | |
8dd5516b | 1562 | redundant_reason = NULL; |
6216f94e JJ |
1563 | mem = canon_rtx (mem); |
1564 | /* For alias_set != 0 canon_true_dependence should be never called. */ | |
1565 | if (spill_alias_set) | |
1566 | mem_addr = NULL_RTX; | |
1567 | else | |
1568 | { | |
1569 | if (group_id < 0) | |
1570 | mem_addr = base->val_rtx; | |
1571 | else | |
1572 | { | |
1573 | group_info_t group | |
1574 | = VEC_index (group_info_t, rtx_group_vec, group_id); | |
1575 | mem_addr = group->canon_base_addr; | |
1576 | } | |
1577 | if (offset) | |
372d6395 | 1578 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6216f94e | 1579 | } |
6fb5fa3c DB |
1580 | |
1581 | while (ptr) | |
1582 | { | |
1583 | insn_info_t next = ptr->next_local_store; | |
1584 | store_info_t s_info = ptr->store_rec; | |
60564289 | 1585 | bool del = true; |
6fb5fa3c DB |
1586 | |
1587 | /* Skip the clobbers. We delete the active insn if this insn | |
6ed3da00 | 1588 | shadows the set. To have been put on the active list, it |
6fb5fa3c DB |
1589 | has exactly on set. */ |
1590 | while (!s_info->is_set) | |
1591 | s_info = s_info->next; | |
1592 | ||
1593 | if (s_info->alias_set != spill_alias_set) | |
60564289 | 1594 | del = false; |
6fb5fa3c DB |
1595 | else if (s_info->alias_set) |
1596 | { | |
b8698a0f | 1597 | struct clear_alias_mode_holder *entry |
6fb5fa3c DB |
1598 | = clear_alias_set_lookup (s_info->alias_set); |
1599 | /* Generally, spills cannot be processed if and of the | |
1600 | references to the slot have a different mode. But if | |
1601 | we are in the same block and mode is exactly the same | |
1602 | between this store and one before in the same block, | |
1603 | we can still delete it. */ | |
1604 | if ((GET_MODE (mem) == GET_MODE (s_info->mem)) | |
1605 | && (GET_MODE (mem) == entry->mode)) | |
1606 | { | |
60564289 | 1607 | del = true; |
8dd5516b | 1608 | set_all_positions_unneeded (s_info); |
6fb5fa3c DB |
1609 | } |
1610 | if (dump_file) | |
1611 | fprintf (dump_file, " trying spill store in insn=%d alias_set=%d\n", | |
4862826d | 1612 | INSN_UID (ptr->insn), (int) s_info->alias_set); |
6fb5fa3c | 1613 | } |
b8698a0f | 1614 | else if ((s_info->group_id == group_id) |
6fb5fa3c DB |
1615 | && (s_info->cse_base == base)) |
1616 | { | |
1617 | HOST_WIDE_INT i; | |
1618 | if (dump_file) | |
1619 | fprintf (dump_file, " trying store in insn=%d gid=%d[%d..%d)\n", | |
b8698a0f | 1620 | INSN_UID (ptr->insn), s_info->group_id, |
6fb5fa3c | 1621 | (int)s_info->begin, (int)s_info->end); |
8dd5516b JJ |
1622 | |
1623 | /* Even if PTR won't be eliminated as unneeded, if both | |
1624 | PTR and this insn store the same constant value, we might | |
1625 | eliminate this insn instead. */ | |
1626 | if (s_info->const_rhs | |
1627 | && const_rhs | |
1628 | && offset >= s_info->begin | |
1629 | && offset + width <= s_info->end | |
1630 | && all_positions_needed_p (s_info, offset - s_info->begin, | |
1631 | width)) | |
1632 | { | |
1633 | if (GET_MODE (mem) == BLKmode) | |
1634 | { | |
1635 | if (GET_MODE (s_info->mem) == BLKmode | |
1636 | && s_info->const_rhs == const_rhs) | |
1637 | redundant_reason = ptr; | |
1638 | } | |
1639 | else if (s_info->const_rhs == const0_rtx | |
1640 | && const_rhs == const0_rtx) | |
1641 | redundant_reason = ptr; | |
1642 | else | |
1643 | { | |
1644 | rtx val; | |
1645 | start_sequence (); | |
1646 | val = get_stored_val (s_info, GET_MODE (mem), | |
1647 | offset, offset + width, | |
1648 | BLOCK_FOR_INSN (insn_info->insn), | |
1649 | true); | |
1650 | if (get_insns () != NULL) | |
1651 | val = NULL_RTX; | |
1652 | end_sequence (); | |
1653 | if (val && rtx_equal_p (val, const_rhs)) | |
1654 | redundant_reason = ptr; | |
1655 | } | |
1656 | } | |
1657 | ||
1658 | for (i = MAX (offset, s_info->begin); | |
1659 | i < offset + width && i < s_info->end; | |
1660 | i++) | |
1661 | set_position_unneeded (s_info, i - s_info->begin); | |
6fb5fa3c DB |
1662 | } |
1663 | else if (s_info->rhs) | |
1664 | /* Need to see if it is possible for this store to overwrite | |
1665 | the value of store_info. If it is, set the rhs to NULL to | |
1666 | keep it from being used to remove a load. */ | |
1667 | { | |
b8698a0f | 1668 | if (canon_true_dependence (s_info->mem, |
6fb5fa3c DB |
1669 | GET_MODE (s_info->mem), |
1670 | s_info->mem_addr, | |
53d9622b | 1671 | mem, mem_addr)) |
8dd5516b JJ |
1672 | { |
1673 | s_info->rhs = NULL; | |
1674 | s_info->const_rhs = NULL; | |
1675 | } | |
6fb5fa3c | 1676 | } |
6216f94e | 1677 | |
6fb5fa3c DB |
1678 | /* An insn can be deleted if every position of every one of |
1679 | its s_infos is zero. */ | |
1b6fa860 | 1680 | if (any_positions_needed_p (s_info)) |
60564289 | 1681 | del = false; |
8dd5516b | 1682 | |
60564289 | 1683 | if (del) |
6fb5fa3c DB |
1684 | { |
1685 | insn_info_t insn_to_delete = ptr; | |
b8698a0f | 1686 | |
dabd47e7 | 1687 | active_local_stores_len--; |
6fb5fa3c DB |
1688 | if (last) |
1689 | last->next_local_store = ptr->next_local_store; | |
1690 | else | |
1691 | active_local_stores = ptr->next_local_store; | |
b8698a0f | 1692 | |
1b6fa860 JJ |
1693 | if (!insn_to_delete->cannot_delete) |
1694 | delete_dead_store_insn (insn_to_delete); | |
6fb5fa3c DB |
1695 | } |
1696 | else | |
1697 | last = ptr; | |
b8698a0f | 1698 | |
6fb5fa3c DB |
1699 | ptr = next; |
1700 | } | |
b8698a0f | 1701 | |
6fb5fa3c DB |
1702 | /* Finish filling in the store_info. */ |
1703 | store_info->next = insn_info->store_rec; | |
1704 | insn_info->store_rec = store_info; | |
6216f94e | 1705 | store_info->mem = mem; |
6fb5fa3c | 1706 | store_info->alias_set = spill_alias_set; |
6216f94e | 1707 | store_info->mem_addr = mem_addr; |
6fb5fa3c | 1708 | store_info->cse_base = base; |
8dd5516b JJ |
1709 | if (width > HOST_BITS_PER_WIDE_INT) |
1710 | { | |
1711 | store_info->is_large = true; | |
1712 | store_info->positions_needed.large.count = 0; | |
3f9b14ff | 1713 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
8dd5516b JJ |
1714 | } |
1715 | else | |
1716 | { | |
1717 | store_info->is_large = false; | |
1718 | store_info->positions_needed.small_bitmask = lowpart_bitmask (width); | |
1719 | } | |
6fb5fa3c DB |
1720 | store_info->group_id = group_id; |
1721 | store_info->begin = offset; | |
1722 | store_info->end = offset + width; | |
1723 | store_info->is_set = GET_CODE (body) == SET; | |
8dd5516b JJ |
1724 | store_info->rhs = rhs; |
1725 | store_info->const_rhs = const_rhs; | |
1726 | store_info->redundant_reason = redundant_reason; | |
6fb5fa3c | 1727 | |
6fb5fa3c DB |
1728 | /* If this is a clobber, we return 0. We will only be able to |
1729 | delete this insn if there is only one store USED store, but we | |
1730 | can use the clobber to delete other stores earlier. */ | |
1731 | return store_info->is_set ? 1 : 0; | |
1732 | } | |
1733 | ||
1734 | ||
1735 | static void | |
1736 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1737 | { | |
b8698a0f | 1738 | fprintf (dump_file, "%s insn=%d %s\n", start, |
6fb5fa3c DB |
1739 | INSN_UID (insn_info->insn), |
1740 | insn_info->store_rec ? "has store" : "naked"); | |
1741 | } | |
1742 | ||
1743 | ||
8660aaae EC |
1744 | /* If the modes are different and the value's source and target do not |
1745 | line up, we need to extract the value from lower part of the rhs of | |
1746 | the store, shift it, and then put it into a form that can be shoved | |
1747 | into the read_insn. This function generates a right SHIFT of a | |
1748 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1749 | shift sequence is returned or NULL if we failed to find a | |
1750 | shift. */ | |
1751 | ||
1752 | static rtx | |
18b526e8 | 1753 | find_shift_sequence (int access_size, |
8660aaae | 1754 | store_info_t store_info, |
8dd5516b JJ |
1755 | enum machine_mode read_mode, |
1756 | int shift, bool speed, bool require_cst) | |
8660aaae EC |
1757 | { |
1758 | enum machine_mode store_mode = GET_MODE (store_info->mem); | |
18b526e8 RS |
1759 | enum machine_mode new_mode; |
1760 | rtx read_reg = NULL; | |
8660aaae EC |
1761 | |
1762 | /* Some machines like the x86 have shift insns for each size of | |
1763 | operand. Other machines like the ppc or the ia-64 may only have | |
1764 | shift insns that shift values within 32 or 64 bit registers. | |
1765 | This loop tries to find the smallest shift insn that will right | |
1766 | justify the value we want to read but is available in one insn on | |
1767 | the machine. */ | |
1768 | ||
18b526e8 RS |
1769 | for (new_mode = smallest_mode_for_size (access_size * BITS_PER_UNIT, |
1770 | MODE_INT); | |
1771 | GET_MODE_BITSIZE (new_mode) <= BITS_PER_WORD; | |
1772 | new_mode = GET_MODE_WIDER_MODE (new_mode)) | |
8660aaae | 1773 | { |
18b526e8 | 1774 | rtx target, new_reg, shift_seq, insn, new_lhs; |
d898d29b | 1775 | int cost; |
348eea5f | 1776 | |
72a2609f JJ |
1777 | /* If a constant was stored into memory, try to simplify it here, |
1778 | otherwise the cost of the shift might preclude this optimization | |
1779 | e.g. at -Os, even when no actual shift will be needed. */ | |
8dd5516b | 1780 | if (store_info->const_rhs) |
72a2609f JJ |
1781 | { |
1782 | unsigned int byte = subreg_lowpart_offset (new_mode, store_mode); | |
8dd5516b JJ |
1783 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1784 | store_mode, byte); | |
72a2609f JJ |
1785 | if (ret && CONSTANT_P (ret)) |
1786 | { | |
1787 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, | |
1788 | ret, GEN_INT (shift)); | |
1789 | if (ret && CONSTANT_P (ret)) | |
1790 | { | |
1791 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1792 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1793 | if (ret && CONSTANT_P (ret) | |
5e8f01f4 | 1794 | && set_src_cost (ret, speed) <= COSTS_N_INSNS (1)) |
72a2609f JJ |
1795 | return ret; |
1796 | } | |
1797 | } | |
1798 | } | |
1799 | ||
8dd5516b JJ |
1800 | if (require_cst) |
1801 | return NULL_RTX; | |
1802 | ||
18b526e8 RS |
1803 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1804 | requires a real instruction. */ | |
1805 | if (GET_MODE_BITSIZE (new_mode) < GET_MODE_BITSIZE (store_mode) | |
d0edd768 | 1806 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
348eea5f RS |
1807 | continue; |
1808 | ||
18b526e8 RS |
1809 | /* Also try a wider mode if the necessary punning is either not |
1810 | desirable or not possible. */ | |
1811 | if (!CONSTANT_P (store_info->rhs) | |
1812 | && !MODES_TIEABLE_P (new_mode, store_mode)) | |
1813 | continue; | |
18b526e8 | 1814 | |
348eea5f | 1815 | new_reg = gen_reg_rtx (new_mode); |
8660aaae EC |
1816 | |
1817 | start_sequence (); | |
1818 | ||
1819 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1820 | of one dsp where the cost of these two was not the same. But | |
1821 | this really is a rare case anyway. */ | |
1822 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
1823 | GEN_INT (shift), new_reg, 1, OPTAB_DIRECT); | |
1824 | ||
c6f3019a RS |
1825 | shift_seq = get_insns (); |
1826 | end_sequence (); | |
8660aaae | 1827 | |
c6f3019a RS |
1828 | if (target != new_reg || shift_seq == NULL) |
1829 | continue; | |
1830 | ||
1831 | cost = 0; | |
1832 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1833 | if (INSN_P (insn)) | |
f40751dd | 1834 | cost += insn_rtx_cost (PATTERN (insn), speed); |
c6f3019a RS |
1835 | |
1836 | /* The computation up to here is essentially independent | |
1837 | of the arguments and could be precomputed. It may | |
1838 | not be worth doing so. We could precompute if | |
1839 | worthwhile or at least cache the results. The result | |
06acf7d0 RS |
1840 | technically depends on both SHIFT and ACCESS_SIZE, |
1841 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
c6f3019a RS |
1842 | |
1843 | if (cost > COSTS_N_INSNS (1)) | |
1844 | continue; | |
1845 | ||
d898d29b JJ |
1846 | new_lhs = extract_low_bits (new_mode, store_mode, |
1847 | copy_rtx (store_info->rhs)); | |
1848 | if (new_lhs == NULL_RTX) | |
1849 | continue; | |
1850 | ||
c6f3019a RS |
1851 | /* We found an acceptable shift. Generate a move to |
1852 | take the value from the store and put it into the | |
1853 | shift pseudo, then shift it, then generate another | |
1854 | move to put in into the target of the read. */ | |
18b526e8 | 1855 | emit_move_insn (new_reg, new_lhs); |
c6f3019a | 1856 | emit_insn (shift_seq); |
18b526e8 | 1857 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
c6f3019a | 1858 | break; |
8660aaae EC |
1859 | } |
1860 | ||
18b526e8 | 1861 | return read_reg; |
8660aaae EC |
1862 | } |
1863 | ||
1864 | ||
02b47899 KZ |
1865 | /* Call back for note_stores to find the hard regs set or clobbered by |
1866 | insn. Data is a bitmap of the hardregs set so far. */ | |
1867 | ||
1868 | static void | |
1869 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1870 | { | |
1871 | bitmap regs_set = (bitmap) data; | |
1872 | ||
1873 | if (REG_P (x) | |
f773c2bd | 1874 | && HARD_REGISTER_P (x)) |
02b47899 | 1875 | { |
f773c2bd AS |
1876 | unsigned int regno = REGNO (x); |
1877 | bitmap_set_range (regs_set, regno, | |
1878 | hard_regno_nregs[regno][GET_MODE (x)]); | |
02b47899 KZ |
1879 | } |
1880 | } | |
1881 | ||
8dd5516b JJ |
1882 | /* Helper function for replace_read and record_store. |
1883 | Attempt to return a value stored in STORE_INFO, from READ_BEGIN | |
1884 | to one before READ_END bytes read in READ_MODE. Return NULL | |
1885 | if not successful. If REQUIRE_CST is true, return always constant. */ | |
1886 | ||
1887 | static rtx | |
1888 | get_stored_val (store_info_t store_info, enum machine_mode read_mode, | |
1889 | HOST_WIDE_INT read_begin, HOST_WIDE_INT read_end, | |
1890 | basic_block bb, bool require_cst) | |
1891 | { | |
1892 | enum machine_mode store_mode = GET_MODE (store_info->mem); | |
1893 | int shift; | |
1894 | int access_size; /* In bytes. */ | |
1895 | rtx read_reg; | |
1896 | ||
1897 | /* To get here the read is within the boundaries of the write so | |
1898 | shift will never be negative. Start out with the shift being in | |
1899 | bytes. */ | |
1900 | if (store_mode == BLKmode) | |
1901 | shift = 0; | |
1902 | else if (BYTES_BIG_ENDIAN) | |
1903 | shift = store_info->end - read_end; | |
1904 | else | |
1905 | shift = read_begin - store_info->begin; | |
1906 | ||
1907 | access_size = shift + GET_MODE_SIZE (read_mode); | |
1908 | ||
1909 | /* From now on it is bits. */ | |
1910 | shift *= BITS_PER_UNIT; | |
1911 | ||
1912 | if (shift) | |
1913 | read_reg = find_shift_sequence (access_size, store_info, read_mode, shift, | |
1914 | optimize_bb_for_speed_p (bb), | |
1915 | require_cst); | |
1916 | else if (store_mode == BLKmode) | |
1917 | { | |
1918 | /* The store is a memset (addr, const_val, const_size). */ | |
1919 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
1920 | store_mode = int_mode_for_mode (read_mode); | |
1921 | if (store_mode == BLKmode) | |
1922 | read_reg = NULL_RTX; | |
1923 | else if (store_info->rhs == const0_rtx) | |
1924 | read_reg = extract_low_bits (read_mode, store_mode, const0_rtx); | |
1925 | else if (GET_MODE_BITSIZE (store_mode) > HOST_BITS_PER_WIDE_INT | |
1926 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) | |
1927 | read_reg = NULL_RTX; | |
1928 | else | |
1929 | { | |
1930 | unsigned HOST_WIDE_INT c | |
1931 | = INTVAL (store_info->rhs) | |
1932 | & (((HOST_WIDE_INT) 1 << BITS_PER_UNIT) - 1); | |
1933 | int shift = BITS_PER_UNIT; | |
1934 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1935 | { | |
1936 | c |= (c << shift); | |
1937 | shift <<= 1; | |
1938 | } | |
6d26322f | 1939 | read_reg = gen_int_mode (c, store_mode); |
8dd5516b JJ |
1940 | read_reg = extract_low_bits (read_mode, store_mode, read_reg); |
1941 | } | |
1942 | } | |
1943 | else if (store_info->const_rhs | |
1944 | && (require_cst | |
1945 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1946 | read_reg = extract_low_bits (read_mode, store_mode, | |
1947 | copy_rtx (store_info->const_rhs)); | |
1948 | else | |
1949 | read_reg = extract_low_bits (read_mode, store_mode, | |
1950 | copy_rtx (store_info->rhs)); | |
1951 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1952 | read_reg = NULL_RTX; | |
1953 | return read_reg; | |
1954 | } | |
02b47899 | 1955 | |
6fb5fa3c DB |
1956 | /* Take a sequence of: |
1957 | A <- r1 | |
1958 | ... | |
1959 | ... <- A | |
1960 | ||
b8698a0f | 1961 | and change it into |
6fb5fa3c DB |
1962 | r2 <- r1 |
1963 | A <- r1 | |
1964 | ... | |
1965 | ... <- r2 | |
1966 | ||
8660aaae EC |
1967 | or |
1968 | ||
1969 | r3 <- extract (r1) | |
1970 | r3 <- r3 >> shift | |
1971 | r2 <- extract (r3) | |
1972 | ... <- r2 | |
1973 | ||
1974 | or | |
1975 | ||
1976 | r2 <- extract (r1) | |
1977 | ... <- r2 | |
1978 | ||
1979 | Depending on the alignment and the mode of the store and | |
1980 | subsequent load. | |
1981 | ||
1982 | ||
1983 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
6fb5fa3c DB |
1984 | and READ_INSN are for the read. Return true if the replacement |
1985 | went ok. */ | |
1986 | ||
1987 | static bool | |
b8698a0f | 1988 | replace_read (store_info_t store_info, insn_info_t store_insn, |
8dd5516b JJ |
1989 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
1990 | bitmap regs_live) | |
6fb5fa3c | 1991 | { |
8660aaae EC |
1992 | enum machine_mode store_mode = GET_MODE (store_info->mem); |
1993 | enum machine_mode read_mode = GET_MODE (read_info->mem); | |
02b47899 | 1994 | rtx insns, this_insn, read_reg; |
8dd5516b | 1995 | basic_block bb; |
8660aaae | 1996 | |
6fb5fa3c DB |
1997 | if (!dbg_cnt (dse)) |
1998 | return false; | |
1999 | ||
18b526e8 RS |
2000 | /* Create a sequence of instructions to set up the read register. |
2001 | This sequence goes immediately before the store and its result | |
2002 | is read by the load. | |
2003 | ||
2004 | We need to keep this in perspective. We are replacing a read | |
8660aaae EC |
2005 | with a sequence of insns, but the read will almost certainly be |
2006 | in cache, so it is not going to be an expensive one. Thus, we | |
2007 | are not willing to do a multi insn shift or worse a subroutine | |
2008 | call to get rid of the read. */ | |
18b526e8 RS |
2009 | if (dump_file) |
2010 | fprintf (dump_file, "trying to replace %smode load in insn %d" | |
2011 | " from %smode store in insn %d\n", | |
2012 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
2013 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
2014 | start_sequence (); | |
8dd5516b JJ |
2015 | bb = BLOCK_FOR_INSN (read_insn->insn); |
2016 | read_reg = get_stored_val (store_info, | |
2017 | read_mode, read_info->begin, read_info->end, | |
2018 | bb, false); | |
18b526e8 | 2019 | if (read_reg == NULL_RTX) |
8660aaae | 2020 | { |
18b526e8 RS |
2021 | end_sequence (); |
2022 | if (dump_file) | |
2023 | fprintf (dump_file, " -- could not extract bits of stored value\n"); | |
2024 | return false; | |
8660aaae | 2025 | } |
18b526e8 RS |
2026 | /* Force the value into a new register so that it won't be clobbered |
2027 | between the store and the load. */ | |
2028 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
2029 | insns = get_insns (); | |
2030 | end_sequence (); | |
8660aaae | 2031 | |
02b47899 KZ |
2032 | if (insns != NULL_RTX) |
2033 | { | |
2034 | /* Now we have to scan the set of new instructions to see if the | |
2035 | sequence contains and sets of hardregs that happened to be | |
2036 | live at this point. For instance, this can happen if one of | |
2037 | the insns sets the CC and the CC happened to be live at that | |
2038 | point. This does occasionally happen, see PR 37922. */ | |
3f9b14ff | 2039 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
02b47899 KZ |
2040 | |
2041 | for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
2042 | note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
b8698a0f | 2043 | |
02b47899 KZ |
2044 | bitmap_and_into (regs_set, regs_live); |
2045 | if (!bitmap_empty_p (regs_set)) | |
2046 | { | |
2047 | if (dump_file) | |
2048 | { | |
b8698a0f | 2049 | fprintf (dump_file, |
02b47899 KZ |
2050 | "abandoning replacement because sequence clobbers live hardregs:"); |
2051 | df_print_regset (dump_file, regs_set); | |
2052 | } | |
b8698a0f | 2053 | |
02b47899 KZ |
2054 | BITMAP_FREE (regs_set); |
2055 | return false; | |
2056 | } | |
2057 | BITMAP_FREE (regs_set); | |
2058 | } | |
2059 | ||
8660aaae | 2060 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
6fb5fa3c | 2061 | { |
f883e0a7 KG |
2062 | deferred_change_t deferred_change = |
2063 | (deferred_change_t) pool_alloc (deferred_change_pool); | |
b8698a0f | 2064 | |
8660aaae EC |
2065 | /* Insert this right before the store insn where it will be safe |
2066 | from later insns that might change it before the read. */ | |
2067 | emit_insn_before (insns, store_insn->insn); | |
b8698a0f | 2068 | |
8660aaae EC |
2069 | /* And now for the kludge part: cselib croaks if you just |
2070 | return at this point. There are two reasons for this: | |
b8698a0f | 2071 | |
8660aaae EC |
2072 | 1) Cselib has an idea of how many pseudos there are and |
2073 | that does not include the new ones we just added. | |
b8698a0f | 2074 | |
8660aaae EC |
2075 | 2) Cselib does not know about the move insn we added |
2076 | above the store_info, and there is no way to tell it | |
2077 | about it, because it has "moved on". | |
b8698a0f | 2078 | |
8660aaae EC |
2079 | Problem (1) is fixable with a certain amount of engineering. |
2080 | Problem (2) is requires starting the bb from scratch. This | |
2081 | could be expensive. | |
b8698a0f | 2082 | |
8660aaae EC |
2083 | So we are just going to have to lie. The move/extraction |
2084 | insns are not really an issue, cselib did not see them. But | |
2085 | the use of the new pseudo read_insn is a real problem because | |
2086 | cselib has not scanned this insn. The way that we solve this | |
2087 | problem is that we are just going to put the mem back for now | |
2088 | and when we are finished with the block, we undo this. We | |
2089 | keep a table of mems to get rid of. At the end of the basic | |
2090 | block we can put them back. */ | |
b8698a0f | 2091 | |
8660aaae EC |
2092 | *loc = read_info->mem; |
2093 | deferred_change->next = deferred_change_list; | |
2094 | deferred_change_list = deferred_change; | |
2095 | deferred_change->loc = loc; | |
2096 | deferred_change->reg = read_reg; | |
b8698a0f | 2097 | |
8660aaae EC |
2098 | /* Get rid of the read_info, from the point of view of the |
2099 | rest of dse, play like this read never happened. */ | |
2100 | read_insn->read_rec = read_info->next; | |
2101 | pool_free (read_info_pool, read_info); | |
18b526e8 RS |
2102 | if (dump_file) |
2103 | { | |
2104 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2105 | print_simple_rtl (dump_file, read_reg); | |
2106 | fprintf (dump_file, "\n"); | |
2107 | } | |
8660aaae | 2108 | return true; |
6fb5fa3c | 2109 | } |
b8698a0f | 2110 | else |
6fb5fa3c | 2111 | { |
6fb5fa3c | 2112 | if (dump_file) |
18b526e8 RS |
2113 | { |
2114 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2115 | print_simple_rtl (dump_file, read_reg); | |
2116 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2117 | } | |
6fb5fa3c DB |
2118 | return false; |
2119 | } | |
2120 | } | |
2121 | ||
6fb5fa3c DB |
2122 | /* A for_each_rtx callback in which DATA is the bb_info. Check to see |
2123 | if LOC is a mem and if it is look at the address and kill any | |
2124 | appropriate stores that may be active. */ | |
2125 | ||
2126 | static int | |
2127 | check_mem_read_rtx (rtx *loc, void *data) | |
2128 | { | |
6216f94e | 2129 | rtx mem = *loc, mem_addr; |
6fb5fa3c DB |
2130 | bb_info_t bb_info; |
2131 | insn_info_t insn_info; | |
2132 | HOST_WIDE_INT offset = 0; | |
2133 | HOST_WIDE_INT width = 0; | |
4862826d | 2134 | alias_set_type spill_alias_set = 0; |
b8698a0f | 2135 | cselib_val *base = NULL; |
6fb5fa3c DB |
2136 | int group_id; |
2137 | read_info_t read_info; | |
2138 | ||
2139 | if (!mem || !MEM_P (mem)) | |
2140 | return 0; | |
2141 | ||
2142 | bb_info = (bb_info_t) data; | |
2143 | insn_info = bb_info->last_insn; | |
2144 | ||
2145 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2146 | || (MEM_VOLATILE_P (mem))) | |
2147 | { | |
2148 | if (dump_file) | |
2149 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); | |
2150 | add_wild_read (bb_info); | |
2151 | insn_info->cannot_delete = true; | |
2152 | return 0; | |
2153 | } | |
2154 | ||
2155 | /* If it is reading readonly mem, then there can be no conflict with | |
2156 | another write. */ | |
2157 | if (MEM_READONLY_P (mem)) | |
2158 | return 0; | |
2159 | ||
2160 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
2161 | { | |
2162 | if (dump_file) | |
2163 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); | |
2164 | add_wild_read (bb_info); | |
2165 | return 0; | |
2166 | } | |
2167 | ||
2168 | if (GET_MODE (mem) == BLKmode) | |
2169 | width = -1; | |
2170 | else | |
2171 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2172 | ||
f883e0a7 | 2173 | read_info = (read_info_t) pool_alloc (read_info_pool); |
6fb5fa3c DB |
2174 | read_info->group_id = group_id; |
2175 | read_info->mem = mem; | |
2176 | read_info->alias_set = spill_alias_set; | |
2177 | read_info->begin = offset; | |
2178 | read_info->end = offset + width; | |
2179 | read_info->next = insn_info->read_rec; | |
2180 | insn_info->read_rec = read_info; | |
6216f94e JJ |
2181 | /* For alias_set != 0 canon_true_dependence should be never called. */ |
2182 | if (spill_alias_set) | |
2183 | mem_addr = NULL_RTX; | |
2184 | else | |
2185 | { | |
2186 | if (group_id < 0) | |
2187 | mem_addr = base->val_rtx; | |
2188 | else | |
2189 | { | |
2190 | group_info_t group | |
2191 | = VEC_index (group_info_t, rtx_group_vec, group_id); | |
2192 | mem_addr = group->canon_base_addr; | |
2193 | } | |
2194 | if (offset) | |
372d6395 | 2195 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6216f94e | 2196 | } |
6fb5fa3c | 2197 | |
0d52bcc1 | 2198 | /* We ignore the clobbers in store_info. The is mildly aggressive, |
6fb5fa3c DB |
2199 | but there really should not be a clobber followed by a read. */ |
2200 | ||
2201 | if (spill_alias_set) | |
2202 | { | |
2203 | insn_info_t i_ptr = active_local_stores; | |
2204 | insn_info_t last = NULL; | |
2205 | ||
2206 | if (dump_file) | |
2207 | fprintf (dump_file, " processing spill load %d\n", | |
4862826d | 2208 | (int) spill_alias_set); |
6fb5fa3c DB |
2209 | |
2210 | while (i_ptr) | |
2211 | { | |
2212 | store_info_t store_info = i_ptr->store_rec; | |
2213 | ||
2214 | /* Skip the clobbers. */ | |
2215 | while (!store_info->is_set) | |
2216 | store_info = store_info->next; | |
b8698a0f | 2217 | |
6fb5fa3c DB |
2218 | if (store_info->alias_set == spill_alias_set) |
2219 | { | |
2220 | if (dump_file) | |
2221 | dump_insn_info ("removing from active", i_ptr); | |
2222 | ||
dabd47e7 | 2223 | active_local_stores_len--; |
6fb5fa3c DB |
2224 | if (last) |
2225 | last->next_local_store = i_ptr->next_local_store; | |
2226 | else | |
2227 | active_local_stores = i_ptr->next_local_store; | |
2228 | } | |
2229 | else | |
2230 | last = i_ptr; | |
2231 | i_ptr = i_ptr->next_local_store; | |
2232 | } | |
2233 | } | |
2234 | else if (group_id >= 0) | |
2235 | { | |
2236 | /* This is the restricted case where the base is a constant or | |
2237 | the frame pointer and offset is a constant. */ | |
2238 | insn_info_t i_ptr = active_local_stores; | |
2239 | insn_info_t last = NULL; | |
b8698a0f | 2240 | |
6fb5fa3c DB |
2241 | if (dump_file) |
2242 | { | |
2243 | if (width == -1) | |
2244 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", | |
2245 | group_id); | |
2246 | else | |
2247 | fprintf (dump_file, " processing const load gid=%d[%d..%d)\n", | |
2248 | group_id, (int)offset, (int)(offset+width)); | |
2249 | } | |
2250 | ||
2251 | while (i_ptr) | |
2252 | { | |
2253 | bool remove = false; | |
2254 | store_info_t store_info = i_ptr->store_rec; | |
b8698a0f | 2255 | |
6fb5fa3c DB |
2256 | /* Skip the clobbers. */ |
2257 | while (!store_info->is_set) | |
2258 | store_info = store_info->next; | |
b8698a0f | 2259 | |
6fb5fa3c DB |
2260 | /* There are three cases here. */ |
2261 | if (store_info->group_id < 0) | |
2262 | /* We have a cselib store followed by a read from a | |
2263 | const base. */ | |
b8698a0f L |
2264 | remove |
2265 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2266 | GET_MODE (store_info->mem), |
2267 | store_info->mem_addr, | |
53d9622b | 2268 | mem, mem_addr); |
b8698a0f | 2269 | |
6fb5fa3c DB |
2270 | else if (group_id == store_info->group_id) |
2271 | { | |
2272 | /* This is a block mode load. We may get lucky and | |
2273 | canon_true_dependence may save the day. */ | |
2274 | if (width == -1) | |
b8698a0f L |
2275 | remove |
2276 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2277 | GET_MODE (store_info->mem), |
2278 | store_info->mem_addr, | |
53d9622b | 2279 | mem, mem_addr); |
b8698a0f | 2280 | |
6fb5fa3c DB |
2281 | /* If this read is just reading back something that we just |
2282 | stored, rewrite the read. */ | |
b8698a0f | 2283 | else |
6fb5fa3c DB |
2284 | { |
2285 | if (store_info->rhs | |
8dd5516b JJ |
2286 | && offset >= store_info->begin |
2287 | && offset + width <= store_info->end | |
2288 | && all_positions_needed_p (store_info, | |
2289 | offset - store_info->begin, | |
2290 | width) | |
2291 | && replace_read (store_info, i_ptr, read_info, | |
2292 | insn_info, loc, bb_info->regs_live)) | |
2293 | return 0; | |
2294 | ||
6fb5fa3c DB |
2295 | /* The bases are the same, just see if the offsets |
2296 | overlap. */ | |
b8698a0f | 2297 | if ((offset < store_info->end) |
6fb5fa3c DB |
2298 | && (offset + width > store_info->begin)) |
2299 | remove = true; | |
2300 | } | |
2301 | } | |
b8698a0f L |
2302 | |
2303 | /* else | |
6fb5fa3c DB |
2304 | The else case that is missing here is that the |
2305 | bases are constant but different. There is nothing | |
2306 | to do here because there is no overlap. */ | |
b8698a0f | 2307 | |
6fb5fa3c DB |
2308 | if (remove) |
2309 | { | |
2310 | if (dump_file) | |
2311 | dump_insn_info ("removing from active", i_ptr); | |
2312 | ||
dabd47e7 | 2313 | active_local_stores_len--; |
6fb5fa3c DB |
2314 | if (last) |
2315 | last->next_local_store = i_ptr->next_local_store; | |
2316 | else | |
2317 | active_local_stores = i_ptr->next_local_store; | |
2318 | } | |
2319 | else | |
2320 | last = i_ptr; | |
2321 | i_ptr = i_ptr->next_local_store; | |
2322 | } | |
2323 | } | |
b8698a0f | 2324 | else |
6fb5fa3c DB |
2325 | { |
2326 | insn_info_t i_ptr = active_local_stores; | |
2327 | insn_info_t last = NULL; | |
2328 | if (dump_file) | |
2329 | { | |
2330 | fprintf (dump_file, " processing cselib load mem:"); | |
2331 | print_inline_rtx (dump_file, mem, 0); | |
2332 | fprintf (dump_file, "\n"); | |
2333 | } | |
2334 | ||
2335 | while (i_ptr) | |
2336 | { | |
2337 | bool remove = false; | |
2338 | store_info_t store_info = i_ptr->store_rec; | |
b8698a0f | 2339 | |
6fb5fa3c DB |
2340 | if (dump_file) |
2341 | fprintf (dump_file, " processing cselib load against insn %d\n", | |
2342 | INSN_UID (i_ptr->insn)); | |
2343 | ||
2344 | /* Skip the clobbers. */ | |
2345 | while (!store_info->is_set) | |
2346 | store_info = store_info->next; | |
2347 | ||
2348 | /* If this read is just reading back something that we just | |
2349 | stored, rewrite the read. */ | |
2350 | if (store_info->rhs | |
2351 | && store_info->group_id == -1 | |
2352 | && store_info->cse_base == base | |
efc3527a | 2353 | && width != -1 |
8dd5516b JJ |
2354 | && offset >= store_info->begin |
2355 | && offset + width <= store_info->end | |
2356 | && all_positions_needed_p (store_info, | |
2357 | offset - store_info->begin, width) | |
2358 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, | |
2359 | bb_info->regs_live)) | |
2360 | return 0; | |
6fb5fa3c DB |
2361 | |
2362 | if (!store_info->alias_set) | |
b8698a0f | 2363 | remove = canon_true_dependence (store_info->mem, |
6fb5fa3c DB |
2364 | GET_MODE (store_info->mem), |
2365 | store_info->mem_addr, | |
53d9622b | 2366 | mem, mem_addr); |
b8698a0f | 2367 | |
6fb5fa3c DB |
2368 | if (remove) |
2369 | { | |
2370 | if (dump_file) | |
2371 | dump_insn_info ("removing from active", i_ptr); | |
b8698a0f | 2372 | |
dabd47e7 | 2373 | active_local_stores_len--; |
6fb5fa3c DB |
2374 | if (last) |
2375 | last->next_local_store = i_ptr->next_local_store; | |
2376 | else | |
2377 | active_local_stores = i_ptr->next_local_store; | |
2378 | } | |
2379 | else | |
2380 | last = i_ptr; | |
2381 | i_ptr = i_ptr->next_local_store; | |
2382 | } | |
2383 | } | |
2384 | return 0; | |
2385 | } | |
2386 | ||
b8698a0f | 2387 | /* A for_each_rtx callback in which DATA points the INSN_INFO for |
6fb5fa3c DB |
2388 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2389 | true for any part of *LOC. */ | |
2390 | ||
2391 | static void | |
2392 | check_mem_read_use (rtx *loc, void *data) | |
2393 | { | |
2394 | for_each_rtx (loc, check_mem_read_rtx, data); | |
2395 | } | |
2396 | ||
8dd5516b JJ |
2397 | |
2398 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2399 | So far it only handles arguments passed in registers. */ | |
2400 | ||
2401 | static bool | |
2402 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2403 | { | |
d5cc9181 JR |
2404 | CUMULATIVE_ARGS args_so_far_v; |
2405 | cumulative_args_t args_so_far; | |
8dd5516b JJ |
2406 | tree arg; |
2407 | int idx; | |
2408 | ||
d5cc9181 JR |
2409 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2410 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
8dd5516b JJ |
2411 | |
2412 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2413 | for (idx = 0; | |
2414 | arg != void_list_node && idx < nargs; | |
2415 | arg = TREE_CHAIN (arg), idx++) | |
2416 | { | |
2417 | enum machine_mode mode = TYPE_MODE (TREE_VALUE (arg)); | |
3c07301f | 2418 | rtx reg, link, tmp; |
d5cc9181 | 2419 | reg = targetm.calls.function_arg (args_so_far, mode, NULL_TREE, true); |
8dd5516b JJ |
2420 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode |
2421 | || GET_MODE_CLASS (mode) != MODE_INT) | |
2422 | return false; | |
2423 | ||
2424 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2425 | link; | |
2426 | link = XEXP (link, 1)) | |
2427 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2428 | { | |
2429 | args[idx] = XEXP (XEXP (link, 0), 0); | |
2430 | if (REG_P (args[idx]) | |
2431 | && REGNO (args[idx]) == REGNO (reg) | |
2432 | && (GET_MODE (args[idx]) == mode | |
2433 | || (GET_MODE_CLASS (GET_MODE (args[idx])) == MODE_INT | |
2434 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2435 | <= UNITS_PER_WORD) | |
2436 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2437 | > GET_MODE_SIZE (mode))))) | |
2438 | break; | |
2439 | } | |
2440 | if (!link) | |
2441 | return false; | |
2442 | ||
2443 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2444 | if (GET_MODE (args[idx]) != mode) | |
2445 | { | |
2446 | if (!tmp || !CONST_INT_P (tmp)) | |
2447 | return false; | |
6d26322f | 2448 | tmp = gen_int_mode (INTVAL (tmp), mode); |
8dd5516b JJ |
2449 | } |
2450 | if (tmp) | |
2451 | args[idx] = tmp; | |
2452 | ||
d5cc9181 | 2453 | targetm.calls.function_arg_advance (args_so_far, mode, NULL_TREE, true); |
8dd5516b JJ |
2454 | } |
2455 | if (arg != void_list_node || idx != nargs) | |
2456 | return false; | |
2457 | return true; | |
2458 | } | |
2459 | ||
9e582b1d JR |
2460 | /* Return a bitmap of the fixed registers contained in IN. */ |
2461 | ||
2462 | static bitmap | |
2463 | copy_fixed_regs (const_bitmap in) | |
2464 | { | |
2465 | bitmap ret; | |
2466 | ||
2467 | ret = ALLOC_REG_SET (NULL); | |
2468 | bitmap_and (ret, in, fixed_reg_set_regset); | |
2469 | return ret; | |
2470 | } | |
8dd5516b | 2471 | |
6fb5fa3c DB |
2472 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2473 | if some part of it is not a candidate store and assigns to a | |
2474 | non-register target. */ | |
2475 | ||
2476 | static void | |
2477 | scan_insn (bb_info_t bb_info, rtx insn) | |
2478 | { | |
2479 | rtx body; | |
f883e0a7 | 2480 | insn_info_t insn_info = (insn_info_t) pool_alloc (insn_info_pool); |
6fb5fa3c DB |
2481 | int mems_found = 0; |
2482 | memset (insn_info, 0, sizeof (struct insn_info)); | |
2483 | ||
2484 | if (dump_file) | |
2485 | fprintf (dump_file, "\n**scanning insn=%d\n", | |
2486 | INSN_UID (insn)); | |
2487 | ||
2488 | insn_info->prev_insn = bb_info->last_insn; | |
2489 | insn_info->insn = insn; | |
2490 | bb_info->last_insn = insn_info; | |
b8698a0f | 2491 | |
b5b8b0ac AO |
2492 | if (DEBUG_INSN_P (insn)) |
2493 | { | |
2494 | insn_info->cannot_delete = true; | |
2495 | return; | |
2496 | } | |
6fb5fa3c | 2497 | |
0d52bcc1 | 2498 | /* Cselib clears the table for this case, so we have to essentially |
6fb5fa3c DB |
2499 | do the same. */ |
2500 | if (NONJUMP_INSN_P (insn) | |
2501 | && GET_CODE (PATTERN (insn)) == ASM_OPERANDS | |
2502 | && MEM_VOLATILE_P (PATTERN (insn))) | |
2503 | { | |
2504 | add_wild_read (bb_info); | |
2505 | insn_info->cannot_delete = true; | |
2506 | return; | |
2507 | } | |
2508 | ||
2509 | /* Look at all of the uses in the insn. */ | |
2510 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2511 | ||
2512 | if (CALL_P (insn)) | |
2513 | { | |
8dd5516b JJ |
2514 | bool const_call; |
2515 | tree memset_call = NULL_TREE; | |
2516 | ||
6fb5fa3c | 2517 | insn_info->cannot_delete = true; |
50f0f366 | 2518 | |
6fb5fa3c | 2519 | /* Const functions cannot do anything bad i.e. read memory, |
50f0f366 | 2520 | however, they can read their parameters which may have |
8dd5516b JJ |
2521 | been pushed onto the stack. |
2522 | memset and bzero don't read memory either. */ | |
2523 | const_call = RTL_CONST_CALL_P (insn); | |
2524 | if (!const_call) | |
2525 | { | |
2526 | rtx call = PATTERN (insn); | |
2527 | if (GET_CODE (call) == PARALLEL) | |
2528 | call = XVECEXP (call, 0, 0); | |
2529 | if (GET_CODE (call) == SET) | |
2530 | call = SET_SRC (call); | |
2531 | if (GET_CODE (call) == CALL | |
2532 | && MEM_P (XEXP (call, 0)) | |
2533 | && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF) | |
2534 | { | |
2535 | rtx symbol = XEXP (XEXP (call, 0), 0); | |
2536 | if (SYMBOL_REF_DECL (symbol) | |
2537 | && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL) | |
2538 | { | |
2539 | if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol)) | |
2540 | == BUILT_IN_NORMAL | |
2541 | && (DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol)) | |
2542 | == BUILT_IN_MEMSET)) | |
2543 | || SYMBOL_REF_DECL (symbol) == block_clear_fn) | |
2544 | memset_call = SYMBOL_REF_DECL (symbol); | |
2545 | } | |
2546 | } | |
2547 | } | |
2548 | if (const_call || memset_call) | |
6fb5fa3c DB |
2549 | { |
2550 | insn_info_t i_ptr = active_local_stores; | |
2551 | insn_info_t last = NULL; | |
2552 | ||
2553 | if (dump_file) | |
8dd5516b JJ |
2554 | fprintf (dump_file, "%s call %d\n", |
2555 | const_call ? "const" : "memset", INSN_UID (insn)); | |
6fb5fa3c | 2556 | |
64520bdc EB |
2557 | /* See the head comment of the frame_read field. */ |
2558 | if (reload_completed) | |
2559 | insn_info->frame_read = true; | |
2560 | ||
2561 | /* Loop over the active stores and remove those which are | |
2562 | killed by the const function call. */ | |
6fb5fa3c DB |
2563 | while (i_ptr) |
2564 | { | |
64520bdc EB |
2565 | bool remove_store = false; |
2566 | ||
2567 | /* The stack pointer based stores are always killed. */ | |
50f0f366 | 2568 | if (i_ptr->stack_pointer_based) |
64520bdc EB |
2569 | remove_store = true; |
2570 | ||
2571 | /* If the frame is read, the frame related stores are killed. */ | |
2572 | else if (insn_info->frame_read) | |
2573 | { | |
2574 | store_info_t store_info = i_ptr->store_rec; | |
2575 | ||
2576 | /* Skip the clobbers. */ | |
2577 | while (!store_info->is_set) | |
2578 | store_info = store_info->next; | |
2579 | ||
2580 | if (store_info->group_id >= 0 | |
2581 | && VEC_index (group_info_t, rtx_group_vec, | |
2582 | store_info->group_id)->frame_related) | |
2583 | remove_store = true; | |
2584 | } | |
2585 | ||
2586 | if (remove_store) | |
6fb5fa3c DB |
2587 | { |
2588 | if (dump_file) | |
2589 | dump_insn_info ("removing from active", i_ptr); | |
b8698a0f | 2590 | |
dabd47e7 | 2591 | active_local_stores_len--; |
6fb5fa3c DB |
2592 | if (last) |
2593 | last->next_local_store = i_ptr->next_local_store; | |
2594 | else | |
2595 | active_local_stores = i_ptr->next_local_store; | |
2596 | } | |
2597 | else | |
2598 | last = i_ptr; | |
64520bdc | 2599 | |
6fb5fa3c DB |
2600 | i_ptr = i_ptr->next_local_store; |
2601 | } | |
8dd5516b JJ |
2602 | |
2603 | if (memset_call) | |
2604 | { | |
2605 | rtx args[3]; | |
2606 | if (get_call_args (insn, memset_call, args, 3) | |
2607 | && CONST_INT_P (args[1]) | |
2608 | && CONST_INT_P (args[2]) | |
2609 | && INTVAL (args[2]) > 0) | |
2610 | { | |
2611 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
f5541398 | 2612 | set_mem_size (mem, INTVAL (args[2])); |
8dd5516b JJ |
2613 | body = gen_rtx_SET (VOIDmode, mem, args[1]); |
2614 | mems_found += record_store (body, bb_info); | |
2615 | if (dump_file) | |
2616 | fprintf (dump_file, "handling memset as BLKmode store\n"); | |
2617 | if (mems_found == 1) | |
2618 | { | |
dabd47e7 JJ |
2619 | if (active_local_stores_len++ |
2620 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2621 | { | |
2622 | active_local_stores_len = 1; | |
2623 | active_local_stores = NULL; | |
2624 | } | |
9e582b1d JR |
2625 | insn_info->fixed_regs_live |
2626 | = copy_fixed_regs (bb_info->regs_live); | |
8dd5516b JJ |
2627 | insn_info->next_local_store = active_local_stores; |
2628 | active_local_stores = insn_info; | |
2629 | } | |
2630 | } | |
2631 | } | |
6fb5fa3c DB |
2632 | } |
2633 | ||
50f0f366 | 2634 | else |
d26c7090 ER |
2635 | /* Every other call, including pure functions, may read any memory |
2636 | that is not relative to the frame. */ | |
2637 | add_non_frame_wild_read (bb_info); | |
50f0f366 | 2638 | |
6fb5fa3c DB |
2639 | return; |
2640 | } | |
2641 | ||
2642 | /* Assuming that there are sets in these insns, we cannot delete | |
2643 | them. */ | |
2644 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
0a64eeca | 2645 | || volatile_refs_p (PATTERN (insn)) |
2da02156 | 2646 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
6fb5fa3c DB |
2647 | || (RTX_FRAME_RELATED_P (insn)) |
2648 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2649 | insn_info->cannot_delete = true; | |
b8698a0f | 2650 | |
6fb5fa3c DB |
2651 | body = PATTERN (insn); |
2652 | if (GET_CODE (body) == PARALLEL) | |
2653 | { | |
2654 | int i; | |
2655 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2656 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2657 | } | |
2658 | else | |
2659 | mems_found += record_store (body, bb_info); | |
2660 | ||
2661 | if (dump_file) | |
b8698a0f | 2662 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
6fb5fa3c DB |
2663 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2664 | ||
8dd5516b JJ |
2665 | /* If we found some sets of mems, add it into the active_local_stores so |
2666 | that it can be locally deleted if found dead or used for | |
2667 | replace_read and redundant constant store elimination. Otherwise mark | |
2668 | it as cannot delete. This simplifies the processing later. */ | |
2669 | if (mems_found == 1) | |
6fb5fa3c | 2670 | { |
dabd47e7 JJ |
2671 | if (active_local_stores_len++ |
2672 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2673 | { | |
2674 | active_local_stores_len = 1; | |
2675 | active_local_stores = NULL; | |
2676 | } | |
9e582b1d | 2677 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
6fb5fa3c DB |
2678 | insn_info->next_local_store = active_local_stores; |
2679 | active_local_stores = insn_info; | |
2680 | } | |
2681 | else | |
2682 | insn_info->cannot_delete = true; | |
2683 | } | |
2684 | ||
2685 | ||
2686 | /* Remove BASE from the set of active_local_stores. This is a | |
2687 | callback from cselib that is used to get rid of the stores in | |
2688 | active_local_stores. */ | |
2689 | ||
2690 | static void | |
2691 | remove_useless_values (cselib_val *base) | |
2692 | { | |
2693 | insn_info_t insn_info = active_local_stores; | |
2694 | insn_info_t last = NULL; | |
2695 | ||
2696 | while (insn_info) | |
2697 | { | |
2698 | store_info_t store_info = insn_info->store_rec; | |
60564289 | 2699 | bool del = false; |
6fb5fa3c DB |
2700 | |
2701 | /* If ANY of the store_infos match the cselib group that is | |
2702 | being deleted, then the insn can not be deleted. */ | |
2703 | while (store_info) | |
2704 | { | |
b8698a0f | 2705 | if ((store_info->group_id == -1) |
6fb5fa3c DB |
2706 | && (store_info->cse_base == base)) |
2707 | { | |
60564289 | 2708 | del = true; |
6fb5fa3c DB |
2709 | break; |
2710 | } | |
2711 | store_info = store_info->next; | |
2712 | } | |
2713 | ||
60564289 | 2714 | if (del) |
6fb5fa3c | 2715 | { |
dabd47e7 | 2716 | active_local_stores_len--; |
6fb5fa3c DB |
2717 | if (last) |
2718 | last->next_local_store = insn_info->next_local_store; | |
2719 | else | |
2720 | active_local_stores = insn_info->next_local_store; | |
2721 | free_store_info (insn_info); | |
2722 | } | |
2723 | else | |
2724 | last = insn_info; | |
b8698a0f | 2725 | |
6fb5fa3c DB |
2726 | insn_info = insn_info->next_local_store; |
2727 | } | |
2728 | } | |
2729 | ||
2730 | ||
2731 | /* Do all of step 1. */ | |
2732 | ||
2733 | static void | |
2734 | dse_step1 (void) | |
2735 | { | |
2736 | basic_block bb; | |
3f9b14ff | 2737 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
b8698a0f | 2738 | |
457eeaae | 2739 | cselib_init (0); |
6fb5fa3c DB |
2740 | all_blocks = BITMAP_ALLOC (NULL); |
2741 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2742 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2743 | ||
2744 | FOR_ALL_BB (bb) | |
2745 | { | |
2746 | insn_info_t ptr; | |
f883e0a7 | 2747 | bb_info_t bb_info = (bb_info_t) pool_alloc (bb_info_pool); |
6fb5fa3c DB |
2748 | |
2749 | memset (bb_info, 0, sizeof (struct bb_info)); | |
2750 | bitmap_set_bit (all_blocks, bb->index); | |
02b47899 KZ |
2751 | bb_info->regs_live = regs_live; |
2752 | ||
2753 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2754 | df_simulate_initialize_forwards (bb, regs_live); | |
6fb5fa3c DB |
2755 | |
2756 | bb_table[bb->index] = bb_info; | |
2757 | cselib_discard_hook = remove_useless_values; | |
2758 | ||
2759 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2760 | { | |
2761 | rtx insn; | |
2762 | ||
2763 | cse_store_info_pool | |
b8698a0f | 2764 | = create_alloc_pool ("cse_store_info_pool", |
6fb5fa3c DB |
2765 | sizeof (struct store_info), 100); |
2766 | active_local_stores = NULL; | |
dabd47e7 | 2767 | active_local_stores_len = 0; |
6fb5fa3c | 2768 | cselib_clear_table (); |
b8698a0f | 2769 | |
6fb5fa3c DB |
2770 | /* Scan the insns. */ |
2771 | FOR_BB_INSNS (bb, insn) | |
2772 | { | |
2773 | if (INSN_P (insn)) | |
2774 | scan_insn (bb_info, insn); | |
2775 | cselib_process_insn (insn); | |
02b47899 KZ |
2776 | if (INSN_P (insn)) |
2777 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
6fb5fa3c | 2778 | } |
b8698a0f | 2779 | |
6fb5fa3c DB |
2780 | /* This is something of a hack, because the global algorithm |
2781 | is supposed to take care of the case where stores go dead | |
2782 | at the end of the function. However, the global | |
2783 | algorithm must take a more conservative view of block | |
2784 | mode reads than the local alg does. So to get the case | |
2785 | where you have a store to the frame followed by a non | |
0d52bcc1 | 2786 | overlapping block more read, we look at the active local |
6fb5fa3c DB |
2787 | stores at the end of the function and delete all of the |
2788 | frame and spill based ones. */ | |
2789 | if (stores_off_frame_dead_at_return | |
2790 | && (EDGE_COUNT (bb->succs) == 0 | |
2791 | || (single_succ_p (bb) | |
2792 | && single_succ (bb) == EXIT_BLOCK_PTR | |
e3b5732b | 2793 | && ! crtl->calls_eh_return))) |
6fb5fa3c DB |
2794 | { |
2795 | insn_info_t i_ptr = active_local_stores; | |
2796 | while (i_ptr) | |
2797 | { | |
2798 | store_info_t store_info = i_ptr->store_rec; | |
2799 | ||
2800 | /* Skip the clobbers. */ | |
2801 | while (!store_info->is_set) | |
2802 | store_info = store_info->next; | |
8dd5516b | 2803 | if (store_info->alias_set && !i_ptr->cannot_delete) |
6fb5fa3c | 2804 | delete_dead_store_insn (i_ptr); |
b8698a0f | 2805 | else |
6fb5fa3c DB |
2806 | if (store_info->group_id >= 0) |
2807 | { | |
b8698a0f | 2808 | group_info_t group |
6fb5fa3c | 2809 | = VEC_index (group_info_t, rtx_group_vec, store_info->group_id); |
8dd5516b | 2810 | if (group->frame_related && !i_ptr->cannot_delete) |
6fb5fa3c DB |
2811 | delete_dead_store_insn (i_ptr); |
2812 | } | |
2813 | ||
2814 | i_ptr = i_ptr->next_local_store; | |
2815 | } | |
2816 | } | |
2817 | ||
2818 | /* Get rid of the loads that were discovered in | |
2819 | replace_read. Cselib is finished with this block. */ | |
2820 | while (deferred_change_list) | |
2821 | { | |
2822 | deferred_change_t next = deferred_change_list->next; | |
2823 | ||
2824 | /* There is no reason to validate this change. That was | |
2825 | done earlier. */ | |
2826 | *deferred_change_list->loc = deferred_change_list->reg; | |
2827 | pool_free (deferred_change_pool, deferred_change_list); | |
2828 | deferred_change_list = next; | |
2829 | } | |
2830 | ||
2831 | /* Get rid of all of the cselib based store_infos in this | |
2832 | block and mark the containing insns as not being | |
2833 | deletable. */ | |
2834 | ptr = bb_info->last_insn; | |
2835 | while (ptr) | |
2836 | { | |
2837 | if (ptr->contains_cselib_groups) | |
8dd5516b JJ |
2838 | { |
2839 | store_info_t s_info = ptr->store_rec; | |
2840 | while (s_info && !s_info->is_set) | |
2841 | s_info = s_info->next; | |
2842 | if (s_info | |
2843 | && s_info->redundant_reason | |
2844 | && s_info->redundant_reason->insn | |
2845 | && !ptr->cannot_delete) | |
2846 | { | |
2847 | if (dump_file) | |
2848 | fprintf (dump_file, "Locally deleting insn %d " | |
2849 | "because insn %d stores the " | |
2850 | "same value and couldn't be " | |
2851 | "eliminated\n", | |
2852 | INSN_UID (ptr->insn), | |
2853 | INSN_UID (s_info->redundant_reason->insn)); | |
2854 | delete_dead_store_insn (ptr); | |
2855 | } | |
2856 | if (s_info) | |
2857 | s_info->redundant_reason = NULL; | |
2858 | free_store_info (ptr); | |
2859 | } | |
2860 | else | |
2861 | { | |
2862 | store_info_t s_info; | |
2863 | ||
2864 | /* Free at least positions_needed bitmaps. */ | |
2865 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2866 | if (s_info->is_large) | |
2867 | { | |
dc491a25 | 2868 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
8dd5516b JJ |
2869 | s_info->is_large = false; |
2870 | } | |
2871 | } | |
6fb5fa3c DB |
2872 | ptr = ptr->prev_insn; |
2873 | } | |
2874 | ||
2875 | free_alloc_pool (cse_store_info_pool); | |
2876 | } | |
02b47899 | 2877 | bb_info->regs_live = NULL; |
6fb5fa3c DB |
2878 | } |
2879 | ||
02b47899 | 2880 | BITMAP_FREE (regs_live); |
6fb5fa3c DB |
2881 | cselib_finish (); |
2882 | htab_empty (rtx_group_table); | |
2883 | } | |
2884 | ||
2885 | \f | |
2886 | /*---------------------------------------------------------------------------- | |
2887 | Second step. | |
2888 | ||
2889 | Assign each byte position in the stores that we are going to | |
2890 | analyze globally to a position in the bitmaps. Returns true if | |
6ed3da00 | 2891 | there are any bit positions assigned. |
6fb5fa3c DB |
2892 | ----------------------------------------------------------------------------*/ |
2893 | ||
2894 | static void | |
2895 | dse_step2_init (void) | |
2896 | { | |
2897 | unsigned int i; | |
2898 | group_info_t group; | |
2899 | ||
ac47786e | 2900 | FOR_EACH_VEC_ELT (group_info_t, rtx_group_vec, i, group) |
6fb5fa3c DB |
2901 | { |
2902 | /* For all non stack related bases, we only consider a store to | |
2903 | be deletable if there are two or more stores for that | |
2904 | position. This is because it takes one store to make the | |
2905 | other store redundant. However, for the stores that are | |
2906 | stack related, we consider them if there is only one store | |
2907 | for the position. We do this because the stack related | |
2908 | stores can be deleted if their is no read between them and | |
2909 | the end of the function. | |
b8698a0f | 2910 | |
6fb5fa3c DB |
2911 | To make this work in the current framework, we take the stack |
2912 | related bases add all of the bits from store1 into store2. | |
2913 | This has the effect of making the eligible even if there is | |
2914 | only one store. */ | |
2915 | ||
2916 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2917 | { | |
2918 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2919 | bitmap_ior_into (group->store2_p, group->store1_p); | |
2920 | if (dump_file) | |
b8698a0f | 2921 | fprintf (dump_file, "group %d is frame related ", i); |
6fb5fa3c DB |
2922 | } |
2923 | ||
2924 | group->offset_map_size_n++; | |
3f9b14ff SB |
2925 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2926 | group->offset_map_size_n); | |
6fb5fa3c | 2927 | group->offset_map_size_p++; |
3f9b14ff SB |
2928 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2929 | group->offset_map_size_p); | |
6fb5fa3c DB |
2930 | group->process_globally = false; |
2931 | if (dump_file) | |
2932 | { | |
b8698a0f | 2933 | fprintf (dump_file, "group %d(%d+%d): ", i, |
6fb5fa3c DB |
2934 | (int)bitmap_count_bits (group->store2_n), |
2935 | (int)bitmap_count_bits (group->store2_p)); | |
2936 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2937 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2938 | } | |
2939 | } | |
2940 | } | |
2941 | ||
2942 | ||
2943 | /* Init the offset tables for the normal case. */ | |
2944 | ||
2945 | static bool | |
2946 | dse_step2_nospill (void) | |
2947 | { | |
2948 | unsigned int i; | |
2949 | group_info_t group; | |
2950 | /* Position 0 is unused because 0 is used in the maps to mean | |
2951 | unused. */ | |
2952 | current_position = 1; | |
ac47786e | 2953 | FOR_EACH_VEC_ELT (group_info_t, rtx_group_vec, i, group) |
6fb5fa3c DB |
2954 | { |
2955 | bitmap_iterator bi; | |
2956 | unsigned int j; | |
2957 | ||
2958 | if (group == clear_alias_group) | |
2959 | continue; | |
2960 | ||
2961 | memset (group->offset_map_n, 0, sizeof(int) * group->offset_map_size_n); | |
2962 | memset (group->offset_map_p, 0, sizeof(int) * group->offset_map_size_p); | |
2963 | bitmap_clear (group->group_kill); | |
2964 | ||
2965 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2966 | { | |
2967 | bitmap_set_bit (group->group_kill, current_position); | |
d26c7090 ER |
2968 | if (bitmap_bit_p (group->escaped_n, j)) |
2969 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2970 | group->offset_map_n[j] = current_position++; |
2971 | group->process_globally = true; | |
2972 | } | |
2973 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2974 | { | |
b8698a0f | 2975 | bitmap_set_bit (group->group_kill, current_position); |
d26c7090 ER |
2976 | if (bitmap_bit_p (group->escaped_p, j)) |
2977 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2978 | group->offset_map_p[j] = current_position++; |
2979 | group->process_globally = true; | |
2980 | } | |
2981 | } | |
2982 | return current_position != 1; | |
2983 | } | |
2984 | ||
2985 | ||
2986 | /* Init the offset tables for the spill case. */ | |
2987 | ||
2988 | static bool | |
2989 | dse_step2_spill (void) | |
2990 | { | |
2991 | unsigned int j; | |
2992 | group_info_t group = clear_alias_group; | |
2993 | bitmap_iterator bi; | |
2994 | ||
2995 | /* Position 0 is unused because 0 is used in the maps to mean | |
2996 | unused. */ | |
2997 | current_position = 1; | |
2998 | ||
2999 | if (dump_file) | |
3000 | { | |
b8698a0f | 3001 | bitmap_print (dump_file, clear_alias_sets, |
6fb5fa3c | 3002 | "clear alias sets ", "\n"); |
b8698a0f | 3003 | bitmap_print (dump_file, disqualified_clear_alias_sets, |
6fb5fa3c DB |
3004 | "disqualified clear alias sets ", "\n"); |
3005 | } | |
3006 | ||
3007 | memset (group->offset_map_n, 0, sizeof(int) * group->offset_map_size_n); | |
3008 | memset (group->offset_map_p, 0, sizeof(int) * group->offset_map_size_p); | |
3009 | bitmap_clear (group->group_kill); | |
b8698a0f | 3010 | |
6fb5fa3c DB |
3011 | /* Remove the disqualified positions from the store2_p set. */ |
3012 | bitmap_and_compl_into (group->store2_p, disqualified_clear_alias_sets); | |
b8698a0f | 3013 | |
6fb5fa3c DB |
3014 | /* We do not need to process the store2_n set because |
3015 | alias_sets are always positive. */ | |
3016 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
3017 | { | |
b8698a0f | 3018 | bitmap_set_bit (group->group_kill, current_position); |
6fb5fa3c DB |
3019 | group->offset_map_p[j] = current_position++; |
3020 | group->process_globally = true; | |
3021 | } | |
3022 | ||
3023 | return current_position != 1; | |
3024 | } | |
3025 | ||
3026 | ||
3027 | \f | |
3028 | /*---------------------------------------------------------------------------- | |
3029 | Third step. | |
b8698a0f | 3030 | |
6fb5fa3c DB |
3031 | Build the bit vectors for the transfer functions. |
3032 | ----------------------------------------------------------------------------*/ | |
3033 | ||
3034 | ||
6fb5fa3c DB |
3035 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
3036 | there, return 0. */ | |
3037 | ||
3038 | static int | |
3039 | get_bitmap_index (group_info_t group_info, HOST_WIDE_INT offset) | |
3040 | { | |
3041 | if (offset < 0) | |
3042 | { | |
3043 | HOST_WIDE_INT offset_p = -offset; | |
3044 | if (offset_p >= group_info->offset_map_size_n) | |
3045 | return 0; | |
3046 | return group_info->offset_map_n[offset_p]; | |
3047 | } | |
3048 | else | |
3049 | { | |
3050 | if (offset >= group_info->offset_map_size_p) | |
3051 | return 0; | |
3052 | return group_info->offset_map_p[offset]; | |
3053 | } | |
3054 | } | |
3055 | ||
3056 | ||
3057 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
3058 | may be NULL. */ | |
3059 | ||
b8698a0f | 3060 | static void |
6fb5fa3c DB |
3061 | scan_stores_nospill (store_info_t store_info, bitmap gen, bitmap kill) |
3062 | { | |
3063 | while (store_info) | |
3064 | { | |
3065 | HOST_WIDE_INT i; | |
b8698a0f | 3066 | group_info_t group_info |
6fb5fa3c DB |
3067 | = VEC_index (group_info_t, rtx_group_vec, store_info->group_id); |
3068 | if (group_info->process_globally) | |
3069 | for (i = store_info->begin; i < store_info->end; i++) | |
3070 | { | |
3071 | int index = get_bitmap_index (group_info, i); | |
3072 | if (index != 0) | |
3073 | { | |
3074 | bitmap_set_bit (gen, index); | |
3075 | if (kill) | |
3076 | bitmap_clear_bit (kill, index); | |
3077 | } | |
3078 | } | |
3079 | store_info = store_info->next; | |
3080 | } | |
3081 | } | |
3082 | ||
3083 | ||
3084 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
3085 | may be NULL. */ | |
3086 | ||
b8698a0f | 3087 | static void |
6fb5fa3c DB |
3088 | scan_stores_spill (store_info_t store_info, bitmap gen, bitmap kill) |
3089 | { | |
3090 | while (store_info) | |
3091 | { | |
3092 | if (store_info->alias_set) | |
3093 | { | |
b8698a0f | 3094 | int index = get_bitmap_index (clear_alias_group, |
6fb5fa3c DB |
3095 | store_info->alias_set); |
3096 | if (index != 0) | |
3097 | { | |
3098 | bitmap_set_bit (gen, index); | |
3099 | if (kill) | |
3100 | bitmap_clear_bit (kill, index); | |
3101 | } | |
3102 | } | |
3103 | store_info = store_info->next; | |
3104 | } | |
3105 | } | |
3106 | ||
3107 | ||
3108 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3109 | may be NULL. */ | |
3110 | ||
3111 | static void | |
3112 | scan_reads_nospill (insn_info_t insn_info, bitmap gen, bitmap kill) | |
3113 | { | |
3114 | read_info_t read_info = insn_info->read_rec; | |
3115 | int i; | |
3116 | group_info_t group; | |
3117 | ||
64520bdc EB |
3118 | /* If this insn reads the frame, kill all the frame related stores. */ |
3119 | if (insn_info->frame_read) | |
3120 | { | |
ac47786e | 3121 | FOR_EACH_VEC_ELT (group_info_t, rtx_group_vec, i, group) |
64520bdc EB |
3122 | if (group->process_globally && group->frame_related) |
3123 | { | |
3124 | if (kill) | |
3125 | bitmap_ior_into (kill, group->group_kill); | |
b8698a0f | 3126 | bitmap_and_compl_into (gen, group->group_kill); |
64520bdc EB |
3127 | } |
3128 | } | |
d26c7090 ER |
3129 | if (insn_info->non_frame_wild_read) |
3130 | { | |
3131 | /* Kill all non-frame related stores. Kill all stores of variables that | |
3132 | escape. */ | |
3133 | if (kill) | |
3134 | bitmap_ior_into (kill, kill_on_calls); | |
3135 | bitmap_and_compl_into (gen, kill_on_calls); | |
3136 | FOR_EACH_VEC_ELT (group_info_t, rtx_group_vec, i, group) | |
3137 | if (group->process_globally && !group->frame_related) | |
3138 | { | |
3139 | if (kill) | |
3140 | bitmap_ior_into (kill, group->group_kill); | |
3141 | bitmap_and_compl_into (gen, group->group_kill); | |
3142 | } | |
3143 | } | |
6fb5fa3c DB |
3144 | while (read_info) |
3145 | { | |
ac47786e | 3146 | FOR_EACH_VEC_ELT (group_info_t, rtx_group_vec, i, group) |
6fb5fa3c DB |
3147 | { |
3148 | if (group->process_globally) | |
3149 | { | |
3150 | if (i == read_info->group_id) | |
3151 | { | |
3152 | if (read_info->begin > read_info->end) | |
3153 | { | |
3154 | /* Begin > end for block mode reads. */ | |
3155 | if (kill) | |
3156 | bitmap_ior_into (kill, group->group_kill); | |
3157 | bitmap_and_compl_into (gen, group->group_kill); | |
3158 | } | |
3159 | else | |
3160 | { | |
3161 | /* The groups are the same, just process the | |
3162 | offsets. */ | |
3163 | HOST_WIDE_INT j; | |
3164 | for (j = read_info->begin; j < read_info->end; j++) | |
3165 | { | |
3166 | int index = get_bitmap_index (group, j); | |
3167 | if (index != 0) | |
3168 | { | |
3169 | if (kill) | |
3170 | bitmap_set_bit (kill, index); | |
3171 | bitmap_clear_bit (gen, index); | |
3172 | } | |
3173 | } | |
3174 | } | |
3175 | } | |
3176 | else | |
3177 | { | |
3178 | /* The groups are different, if the alias sets | |
3179 | conflict, clear the entire group. We only need | |
3180 | to apply this test if the read_info is a cselib | |
3181 | read. Anything with a constant base cannot alias | |
3182 | something else with a different constant | |
3183 | base. */ | |
3184 | if ((read_info->group_id < 0) | |
b8698a0f | 3185 | && canon_true_dependence (group->base_mem, |
d32f725a | 3186 | GET_MODE (group->base_mem), |
6216f94e | 3187 | group->canon_base_addr, |
53d9622b | 3188 | read_info->mem, NULL_RTX)) |
6fb5fa3c DB |
3189 | { |
3190 | if (kill) | |
3191 | bitmap_ior_into (kill, group->group_kill); | |
3192 | bitmap_and_compl_into (gen, group->group_kill); | |
3193 | } | |
3194 | } | |
3195 | } | |
3196 | } | |
b8698a0f | 3197 | |
6fb5fa3c DB |
3198 | read_info = read_info->next; |
3199 | } | |
3200 | } | |
3201 | ||
3202 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3203 | may be NULL. */ | |
3204 | ||
3205 | static void | |
3206 | scan_reads_spill (read_info_t read_info, bitmap gen, bitmap kill) | |
3207 | { | |
3208 | while (read_info) | |
3209 | { | |
3210 | if (read_info->alias_set) | |
3211 | { | |
b8698a0f | 3212 | int index = get_bitmap_index (clear_alias_group, |
6fb5fa3c DB |
3213 | read_info->alias_set); |
3214 | if (index != 0) | |
3215 | { | |
3216 | if (kill) | |
3217 | bitmap_set_bit (kill, index); | |
3218 | bitmap_clear_bit (gen, index); | |
3219 | } | |
3220 | } | |
b8698a0f | 3221 | |
6fb5fa3c DB |
3222 | read_info = read_info->next; |
3223 | } | |
3224 | } | |
3225 | ||
3226 | ||
3227 | /* Return the insn in BB_INFO before the first wild read or if there | |
3228 | are no wild reads in the block, return the last insn. */ | |
3229 | ||
3230 | static insn_info_t | |
3231 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3232 | { | |
3233 | insn_info_t insn_info = bb_info->last_insn; | |
3234 | insn_info_t last_wild_read = NULL; | |
3235 | ||
3236 | while (insn_info) | |
3237 | { | |
3238 | if (insn_info->wild_read) | |
3239 | { | |
3240 | last_wild_read = insn_info->prev_insn; | |
3241 | /* Block starts with wild read. */ | |
3242 | if (!last_wild_read) | |
3243 | return NULL; | |
3244 | } | |
3245 | ||
3246 | insn_info = insn_info->prev_insn; | |
3247 | } | |
3248 | ||
3249 | if (last_wild_read) | |
3250 | return last_wild_read; | |
3251 | else | |
3252 | return bb_info->last_insn; | |
3253 | } | |
3254 | ||
3255 | ||
3256 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3257 | the block in order to build the gen and kill sets for the block. | |
3258 | We start at ptr which may be the last insn in the block or may be | |
3259 | the first insn with a wild read. In the latter case we are able to | |
3260 | skip the rest of the block because it just does not matter: | |
3261 | anything that happens is hidden by the wild read. */ | |
3262 | ||
3263 | static void | |
3264 | dse_step3_scan (bool for_spills, basic_block bb) | |
3265 | { | |
3266 | bb_info_t bb_info = bb_table[bb->index]; | |
3267 | insn_info_t insn_info; | |
3268 | ||
3269 | if (for_spills) | |
3270 | /* There are no wild reads in the spill case. */ | |
3271 | insn_info = bb_info->last_insn; | |
3272 | else | |
3273 | insn_info = find_insn_before_first_wild_read (bb_info); | |
b8698a0f | 3274 | |
6fb5fa3c DB |
3275 | /* In the spill case or in the no_spill case if there is no wild |
3276 | read in the block, we will need a kill set. */ | |
3277 | if (insn_info == bb_info->last_insn) | |
3278 | { | |
3279 | if (bb_info->kill) | |
3280 | bitmap_clear (bb_info->kill); | |
3281 | else | |
3f9b14ff | 3282 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c | 3283 | } |
b8698a0f | 3284 | else |
6fb5fa3c DB |
3285 | if (bb_info->kill) |
3286 | BITMAP_FREE (bb_info->kill); | |
3287 | ||
3288 | while (insn_info) | |
3289 | { | |
3290 | /* There may have been code deleted by the dce pass run before | |
3291 | this phase. */ | |
3292 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3293 | { | |
b8698a0f | 3294 | /* Process the read(s) last. */ |
6fb5fa3c DB |
3295 | if (for_spills) |
3296 | { | |
3297 | scan_stores_spill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3298 | scan_reads_spill (insn_info->read_rec, bb_info->gen, bb_info->kill); | |
3299 | } | |
3300 | else | |
3301 | { | |
3302 | scan_stores_nospill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3303 | scan_reads_nospill (insn_info, bb_info->gen, bb_info->kill); | |
3304 | } | |
b8698a0f | 3305 | } |
6fb5fa3c DB |
3306 | |
3307 | insn_info = insn_info->prev_insn; | |
3308 | } | |
3309 | } | |
3310 | ||
3311 | ||
3312 | /* Set the gen set of the exit block, and also any block with no | |
3313 | successors that does not have a wild read. */ | |
3314 | ||
3315 | static void | |
3316 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3317 | { | |
3318 | /* The gen set is all 0's for the exit block except for the | |
3319 | frame_pointer_group. */ | |
b8698a0f | 3320 | |
6fb5fa3c DB |
3321 | if (stores_off_frame_dead_at_return) |
3322 | { | |
3323 | unsigned int i; | |
3324 | group_info_t group; | |
b8698a0f | 3325 | |
ac47786e | 3326 | FOR_EACH_VEC_ELT (group_info_t, rtx_group_vec, i, group) |
6fb5fa3c DB |
3327 | { |
3328 | if (group->process_globally && group->frame_related) | |
3329 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3330 | } | |
3331 | } | |
3332 | } | |
3333 | ||
3334 | ||
3335 | /* Find all of the blocks that are not backwards reachable from the | |
3336 | exit block or any block with no successors (BB). These are the | |
3337 | infinite loops or infinite self loops. These blocks will still | |
3338 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3339 | ||
3340 | static void | |
3341 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3342 | { | |
3343 | edge e; | |
3344 | edge_iterator ei; | |
3345 | ||
3346 | if (TEST_BIT (unreachable_blocks, bb->index)) | |
3347 | { | |
3348 | RESET_BIT (unreachable_blocks, bb->index); | |
3349 | FOR_EACH_EDGE (e, ei, bb->preds) | |
b8698a0f | 3350 | { |
6fb5fa3c | 3351 | mark_reachable_blocks (unreachable_blocks, e->src); |
b8698a0f | 3352 | } |
6fb5fa3c DB |
3353 | } |
3354 | } | |
3355 | ||
3356 | /* Build the transfer functions for the function. */ | |
3357 | ||
3358 | static void | |
3359 | dse_step3 (bool for_spills) | |
3360 | { | |
3361 | basic_block bb; | |
3362 | sbitmap unreachable_blocks = sbitmap_alloc (last_basic_block); | |
3363 | sbitmap_iterator sbi; | |
3364 | bitmap all_ones = NULL; | |
3365 | unsigned int i; | |
b8698a0f | 3366 | |
6fb5fa3c DB |
3367 | sbitmap_ones (unreachable_blocks); |
3368 | ||
3369 | FOR_ALL_BB (bb) | |
3370 | { | |
3371 | bb_info_t bb_info = bb_table[bb->index]; | |
3372 | if (bb_info->gen) | |
3373 | bitmap_clear (bb_info->gen); | |
3374 | else | |
3f9b14ff | 3375 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3376 | |
3377 | if (bb->index == ENTRY_BLOCK) | |
3378 | ; | |
3379 | else if (bb->index == EXIT_BLOCK) | |
3380 | dse_step3_exit_block_scan (bb_info); | |
3381 | else | |
3382 | dse_step3_scan (for_spills, bb); | |
3383 | if (EDGE_COUNT (bb->succs) == 0) | |
3384 | mark_reachable_blocks (unreachable_blocks, bb); | |
3385 | ||
3386 | /* If this is the second time dataflow is run, delete the old | |
3387 | sets. */ | |
3388 | if (bb_info->in) | |
3389 | BITMAP_FREE (bb_info->in); | |
3390 | if (bb_info->out) | |
3391 | BITMAP_FREE (bb_info->out); | |
3392 | } | |
3393 | ||
3394 | /* For any block in an infinite loop, we must initialize the out set | |
3395 | to all ones. This could be expensive, but almost never occurs in | |
3396 | practice. However, it is common in regression tests. */ | |
3397 | EXECUTE_IF_SET_IN_SBITMAP (unreachable_blocks, 0, i, sbi) | |
3398 | { | |
3399 | if (bitmap_bit_p (all_blocks, i)) | |
3400 | { | |
3401 | bb_info_t bb_info = bb_table[i]; | |
3402 | if (!all_ones) | |
3403 | { | |
3404 | unsigned int j; | |
3405 | group_info_t group; | |
3406 | ||
3f9b14ff | 3407 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
ac47786e | 3408 | FOR_EACH_VEC_ELT (group_info_t, rtx_group_vec, j, group) |
6fb5fa3c DB |
3409 | bitmap_ior_into (all_ones, group->group_kill); |
3410 | } | |
3411 | if (!bb_info->out) | |
3412 | { | |
3f9b14ff | 3413 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3414 | bitmap_copy (bb_info->out, all_ones); |
3415 | } | |
3416 | } | |
3417 | } | |
3418 | ||
3419 | if (all_ones) | |
3420 | BITMAP_FREE (all_ones); | |
3421 | sbitmap_free (unreachable_blocks); | |
3422 | } | |
3423 | ||
3424 | ||
3425 | \f | |
3426 | /*---------------------------------------------------------------------------- | |
3427 | Fourth step. | |
3428 | ||
3429 | Solve the bitvector equations. | |
3430 | ----------------------------------------------------------------------------*/ | |
3431 | ||
3432 | ||
3433 | /* Confluence function for blocks with no successors. Create an out | |
3434 | set from the gen set of the exit block. This block logically has | |
3435 | the exit block as a successor. */ | |
3436 | ||
3437 | ||
3438 | ||
3439 | static void | |
3440 | dse_confluence_0 (basic_block bb) | |
3441 | { | |
3442 | bb_info_t bb_info = bb_table[bb->index]; | |
3443 | ||
3444 | if (bb->index == EXIT_BLOCK) | |
3445 | return; | |
3446 | ||
3447 | if (!bb_info->out) | |
3448 | { | |
3f9b14ff | 3449 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3450 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3451 | } | |
3452 | } | |
3453 | ||
3454 | /* Propagate the information from the in set of the dest of E to the | |
3455 | out set of the src of E. If the various in or out sets are not | |
3456 | there, that means they are all ones. */ | |
3457 | ||
1a0f3fa1 | 3458 | static bool |
6fb5fa3c DB |
3459 | dse_confluence_n (edge e) |
3460 | { | |
3461 | bb_info_t src_info = bb_table[e->src->index]; | |
3462 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3463 | ||
3464 | if (dest_info->in) | |
3465 | { | |
3466 | if (src_info->out) | |
3467 | bitmap_and_into (src_info->out, dest_info->in); | |
3468 | else | |
3469 | { | |
3f9b14ff | 3470 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3471 | bitmap_copy (src_info->out, dest_info->in); |
3472 | } | |
3473 | } | |
1a0f3fa1 | 3474 | return true; |
6fb5fa3c DB |
3475 | } |
3476 | ||
3477 | ||
3478 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
b8698a0f | 3479 | block. There are three cases: |
6fb5fa3c DB |
3480 | |
3481 | 1) The block has no kill set. In this case the kill set is all | |
3482 | ones. It does not matter what the out set of the block is, none of | |
3483 | the info can reach the top. The only thing that reaches the top is | |
3484 | the gen set and we just copy the set. | |
3485 | ||
3486 | 2) There is a kill set but no out set and bb has successors. In | |
3487 | this case we just return. Eventually an out set will be created and | |
3488 | it is better to wait than to create a set of ones. | |
3489 | ||
3490 | 3) There is both a kill and out set. We apply the obvious transfer | |
3491 | function. | |
3492 | */ | |
3493 | ||
3494 | static bool | |
3495 | dse_transfer_function (int bb_index) | |
3496 | { | |
3497 | bb_info_t bb_info = bb_table[bb_index]; | |
3498 | ||
3499 | if (bb_info->kill) | |
3500 | { | |
3501 | if (bb_info->out) | |
3502 | { | |
3503 | /* Case 3 above. */ | |
3504 | if (bb_info->in) | |
b8698a0f | 3505 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3506 | bb_info->out, bb_info->kill); |
3507 | else | |
3508 | { | |
3f9b14ff | 3509 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
b8698a0f | 3510 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3511 | bb_info->out, bb_info->kill); |
3512 | return true; | |
3513 | } | |
3514 | } | |
3515 | else | |
3516 | /* Case 2 above. */ | |
3517 | return false; | |
3518 | } | |
3519 | else | |
3520 | { | |
3521 | /* Case 1 above. If there is already an in set, nothing | |
3522 | happens. */ | |
3523 | if (bb_info->in) | |
3524 | return false; | |
3525 | else | |
3526 | { | |
3f9b14ff | 3527 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3528 | bitmap_copy (bb_info->in, bb_info->gen); |
3529 | return true; | |
3530 | } | |
3531 | } | |
3532 | } | |
3533 | ||
3534 | /* Solve the dataflow equations. */ | |
3535 | ||
3536 | static void | |
3537 | dse_step4 (void) | |
3538 | { | |
b8698a0f L |
3539 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3540 | dse_confluence_n, dse_transfer_function, | |
3541 | all_blocks, df_get_postorder (DF_BACKWARD), | |
6fb5fa3c DB |
3542 | df_get_n_blocks (DF_BACKWARD)); |
3543 | if (dump_file) | |
3544 | { | |
3545 | basic_block bb; | |
3546 | ||
3547 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
3548 | FOR_ALL_BB (bb) | |
3549 | { | |
3550 | bb_info_t bb_info = bb_table[bb->index]; | |
3551 | ||
3552 | df_print_bb_index (bb, dump_file); | |
3553 | if (bb_info->in) | |
3554 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3555 | else | |
3556 | fprintf (dump_file, " in: *MISSING*\n"); | |
3557 | if (bb_info->gen) | |
3558 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3559 | else | |
3560 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3561 | if (bb_info->kill) | |
3562 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3563 | else | |
3564 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3565 | if (bb_info->out) | |
3566 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3567 | else | |
3568 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3569 | } | |
3570 | } | |
3571 | } | |
3572 | ||
3573 | ||
3574 | \f | |
3575 | /*---------------------------------------------------------------------------- | |
3576 | Fifth step. | |
3577 | ||
0d52bcc1 | 3578 | Delete the stores that can only be deleted using the global information. |
6fb5fa3c DB |
3579 | ----------------------------------------------------------------------------*/ |
3580 | ||
3581 | ||
3582 | static void | |
3583 | dse_step5_nospill (void) | |
3584 | { | |
3585 | basic_block bb; | |
3586 | FOR_EACH_BB (bb) | |
3587 | { | |
3588 | bb_info_t bb_info = bb_table[bb->index]; | |
3589 | insn_info_t insn_info = bb_info->last_insn; | |
3590 | bitmap v = bb_info->out; | |
3591 | ||
3592 | while (insn_info) | |
3593 | { | |
3594 | bool deleted = false; | |
3595 | if (dump_file && insn_info->insn) | |
3596 | { | |
3597 | fprintf (dump_file, "starting to process insn %d\n", | |
3598 | INSN_UID (insn_info->insn)); | |
3599 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3600 | } | |
3601 | ||
3602 | /* There may have been code deleted by the dce pass run before | |
3603 | this phase. */ | |
b8698a0f | 3604 | if (insn_info->insn |
6fb5fa3c DB |
3605 | && INSN_P (insn_info->insn) |
3606 | && (!insn_info->cannot_delete) | |
3607 | && (!bitmap_empty_p (v))) | |
3608 | { | |
3609 | store_info_t store_info = insn_info->store_rec; | |
3610 | ||
3611 | /* Try to delete the current insn. */ | |
3612 | deleted = true; | |
b8698a0f | 3613 | |
6fb5fa3c DB |
3614 | /* Skip the clobbers. */ |
3615 | while (!store_info->is_set) | |
3616 | store_info = store_info->next; | |
3617 | ||
3618 | if (store_info->alias_set) | |
3619 | deleted = false; | |
3620 | else | |
3621 | { | |
3622 | HOST_WIDE_INT i; | |
b8698a0f | 3623 | group_info_t group_info |
6fb5fa3c | 3624 | = VEC_index (group_info_t, rtx_group_vec, store_info->group_id); |
b8698a0f | 3625 | |
6fb5fa3c DB |
3626 | for (i = store_info->begin; i < store_info->end; i++) |
3627 | { | |
3628 | int index = get_bitmap_index (group_info, i); | |
b8698a0f | 3629 | |
6fb5fa3c | 3630 | if (dump_file) |
b8698a0f | 3631 | fprintf (dump_file, "i = %d, index = %d\n", (int)i, index); |
6fb5fa3c DB |
3632 | if (index == 0 || !bitmap_bit_p (v, index)) |
3633 | { | |
3634 | if (dump_file) | |
b8698a0f | 3635 | fprintf (dump_file, "failing at i = %d\n", (int)i); |
6fb5fa3c DB |
3636 | deleted = false; |
3637 | break; | |
3638 | } | |
3639 | } | |
3640 | } | |
3641 | if (deleted) | |
3642 | { | |
9e582b1d JR |
3643 | if (dbg_cnt (dse) |
3644 | && check_for_inc_dec_1 (insn_info)) | |
6fb5fa3c | 3645 | { |
6fb5fa3c DB |
3646 | delete_insn (insn_info->insn); |
3647 | insn_info->insn = NULL; | |
3648 | globally_deleted++; | |
3649 | } | |
3650 | } | |
3651 | } | |
3652 | /* We do want to process the local info if the insn was | |
6ed3da00 | 3653 | deleted. For instance, if the insn did a wild read, we |
6fb5fa3c | 3654 | no longer need to trash the info. */ |
b8698a0f | 3655 | if (insn_info->insn |
6fb5fa3c DB |
3656 | && INSN_P (insn_info->insn) |
3657 | && (!deleted)) | |
3658 | { | |
3659 | scan_stores_nospill (insn_info->store_rec, v, NULL); | |
3660 | if (insn_info->wild_read) | |
3661 | { | |
3662 | if (dump_file) | |
3663 | fprintf (dump_file, "wild read\n"); | |
3664 | bitmap_clear (v); | |
3665 | } | |
d26c7090 ER |
3666 | else if (insn_info->read_rec |
3667 | || insn_info->non_frame_wild_read) | |
6fb5fa3c | 3668 | { |
d26c7090 | 3669 | if (dump_file && !insn_info->non_frame_wild_read) |
6fb5fa3c | 3670 | fprintf (dump_file, "regular read\n"); |
d26c7090 ER |
3671 | else if (dump_file) |
3672 | fprintf (dump_file, "non-frame wild read\n"); | |
6fb5fa3c DB |
3673 | scan_reads_nospill (insn_info, v, NULL); |
3674 | } | |
3675 | } | |
b8698a0f | 3676 | |
6fb5fa3c DB |
3677 | insn_info = insn_info->prev_insn; |
3678 | } | |
3679 | } | |
3680 | } | |
3681 | ||
3682 | ||
3683 | static void | |
3684 | dse_step5_spill (void) | |
3685 | { | |
3686 | basic_block bb; | |
3687 | FOR_EACH_BB (bb) | |
3688 | { | |
3689 | bb_info_t bb_info = bb_table[bb->index]; | |
3690 | insn_info_t insn_info = bb_info->last_insn; | |
3691 | bitmap v = bb_info->out; | |
3692 | ||
3693 | while (insn_info) | |
3694 | { | |
3695 | bool deleted = false; | |
3696 | /* There may have been code deleted by the dce pass run before | |
3697 | this phase. */ | |
b8698a0f | 3698 | if (insn_info->insn |
6fb5fa3c DB |
3699 | && INSN_P (insn_info->insn) |
3700 | && (!insn_info->cannot_delete) | |
3701 | && (!bitmap_empty_p (v))) | |
3702 | { | |
3703 | /* Try to delete the current insn. */ | |
3704 | store_info_t store_info = insn_info->store_rec; | |
3705 | deleted = true; | |
b8698a0f | 3706 | |
6fb5fa3c DB |
3707 | while (store_info) |
3708 | { | |
3709 | if (store_info->alias_set) | |
3710 | { | |
b8698a0f | 3711 | int index = get_bitmap_index (clear_alias_group, |
6fb5fa3c DB |
3712 | store_info->alias_set); |
3713 | if (index == 0 || !bitmap_bit_p (v, index)) | |
3714 | { | |
3715 | deleted = false; | |
3716 | break; | |
3717 | } | |
3718 | } | |
b8698a0f | 3719 | else |
6fb5fa3c DB |
3720 | deleted = false; |
3721 | store_info = store_info->next; | |
3722 | } | |
9e582b1d JR |
3723 | if (deleted && dbg_cnt (dse) |
3724 | && check_for_inc_dec_1 (insn_info)) | |
6fb5fa3c DB |
3725 | { |
3726 | if (dump_file) | |
b8698a0f | 3727 | fprintf (dump_file, "Spill deleting insn %d\n", |
6fb5fa3c | 3728 | INSN_UID (insn_info->insn)); |
6fb5fa3c DB |
3729 | delete_insn (insn_info->insn); |
3730 | spill_deleted++; | |
3731 | insn_info->insn = NULL; | |
3732 | } | |
3733 | } | |
b8698a0f L |
3734 | |
3735 | if (insn_info->insn | |
6fb5fa3c DB |
3736 | && INSN_P (insn_info->insn) |
3737 | && (!deleted)) | |
3738 | { | |
3739 | scan_stores_spill (insn_info->store_rec, v, NULL); | |
3740 | scan_reads_spill (insn_info->read_rec, v, NULL); | |
3741 | } | |
b8698a0f | 3742 | |
6fb5fa3c DB |
3743 | insn_info = insn_info->prev_insn; |
3744 | } | |
3745 | } | |
3746 | } | |
3747 | ||
3748 | ||
3749 | \f | |
3750 | /*---------------------------------------------------------------------------- | |
3751 | Sixth step. | |
3752 | ||
8dd5516b JJ |
3753 | Delete stores made redundant by earlier stores (which store the same |
3754 | value) that couldn't be eliminated. | |
3755 | ----------------------------------------------------------------------------*/ | |
3756 | ||
3757 | static void | |
3758 | dse_step6 (void) | |
3759 | { | |
3760 | basic_block bb; | |
3761 | ||
3762 | FOR_ALL_BB (bb) | |
3763 | { | |
3764 | bb_info_t bb_info = bb_table[bb->index]; | |
3765 | insn_info_t insn_info = bb_info->last_insn; | |
3766 | ||
3767 | while (insn_info) | |
3768 | { | |
3769 | /* There may have been code deleted by the dce pass run before | |
3770 | this phase. */ | |
3771 | if (insn_info->insn | |
3772 | && INSN_P (insn_info->insn) | |
3773 | && !insn_info->cannot_delete) | |
3774 | { | |
3775 | store_info_t s_info = insn_info->store_rec; | |
3776 | ||
3777 | while (s_info && !s_info->is_set) | |
3778 | s_info = s_info->next; | |
3779 | if (s_info | |
3780 | && s_info->redundant_reason | |
3781 | && s_info->redundant_reason->insn | |
3782 | && INSN_P (s_info->redundant_reason->insn)) | |
3783 | { | |
3784 | rtx rinsn = s_info->redundant_reason->insn; | |
3785 | if (dump_file) | |
3786 | fprintf (dump_file, "Locally deleting insn %d " | |
3787 | "because insn %d stores the " | |
3788 | "same value and couldn't be " | |
3789 | "eliminated\n", | |
3790 | INSN_UID (insn_info->insn), | |
3791 | INSN_UID (rinsn)); | |
3792 | delete_dead_store_insn (insn_info); | |
3793 | } | |
3794 | } | |
3795 | insn_info = insn_info->prev_insn; | |
3796 | } | |
3797 | } | |
3798 | } | |
3799 | \f | |
3800 | /*---------------------------------------------------------------------------- | |
3801 | Seventh step. | |
3802 | ||
b8698a0f | 3803 | Destroy everything left standing. |
6fb5fa3c DB |
3804 | ----------------------------------------------------------------------------*/ |
3805 | ||
b8698a0f | 3806 | static void |
3f9b14ff | 3807 | dse_step7 (void) |
6fb5fa3c | 3808 | { |
3f9b14ff SB |
3809 | bitmap_obstack_release (&dse_bitmap_obstack); |
3810 | obstack_free (&dse_obstack, NULL); | |
370f38e8 | 3811 | |
6fb5fa3c DB |
3812 | if (clear_alias_sets) |
3813 | { | |
3814 | BITMAP_FREE (clear_alias_sets); | |
3815 | BITMAP_FREE (disqualified_clear_alias_sets); | |
3816 | free_alloc_pool (clear_alias_mode_pool); | |
3817 | htab_delete (clear_alias_mode_table); | |
3818 | } | |
3819 | ||
3820 | end_alias_analysis (); | |
3821 | free (bb_table); | |
3822 | htab_delete (rtx_group_table); | |
3823 | VEC_free (group_info_t, heap, rtx_group_vec); | |
3824 | BITMAP_FREE (all_blocks); | |
3825 | BITMAP_FREE (scratch); | |
3826 | ||
3827 | free_alloc_pool (rtx_store_info_pool); | |
3828 | free_alloc_pool (read_info_pool); | |
3829 | free_alloc_pool (insn_info_pool); | |
3830 | free_alloc_pool (bb_info_pool); | |
3831 | free_alloc_pool (rtx_group_info_pool); | |
3832 | free_alloc_pool (deferred_change_pool); | |
3833 | } | |
3834 | ||
3835 | ||
6fb5fa3c DB |
3836 | /* ------------------------------------------------------------------------- |
3837 | DSE | |
3838 | ------------------------------------------------------------------------- */ | |
3839 | ||
3840 | /* Callback for running pass_rtl_dse. */ | |
3841 | ||
3842 | static unsigned int | |
3843 | rest_of_handle_dse (void) | |
3844 | { | |
3845 | bool did_global = false; | |
3846 | ||
3847 | df_set_flags (DF_DEFER_INSN_RESCAN); | |
3848 | ||
02b47899 KZ |
3849 | /* Need the notes since we must track live hardregs in the forwards |
3850 | direction. */ | |
3851 | df_note_add_problem (); | |
3852 | df_analyze (); | |
3853 | ||
6fb5fa3c DB |
3854 | dse_step0 (); |
3855 | dse_step1 (); | |
3856 | dse_step2_init (); | |
3857 | if (dse_step2_nospill ()) | |
3858 | { | |
3859 | df_set_flags (DF_LR_RUN_DCE); | |
3860 | df_analyze (); | |
3861 | did_global = true; | |
3862 | if (dump_file) | |
3863 | fprintf (dump_file, "doing global processing\n"); | |
3864 | dse_step3 (false); | |
3865 | dse_step4 (); | |
3866 | dse_step5_nospill (); | |
3867 | } | |
3868 | ||
3869 | /* For the instance of dse that runs after reload, we make a special | |
3870 | pass to process the spills. These are special in that they are | |
3871 | totally transparent, i.e, there is no aliasing issues that need | |
3872 | to be considered. This means that the wild reads that kill | |
b8698a0f | 3873 | everything else do not apply here. */ |
6fb5fa3c DB |
3874 | if (clear_alias_sets && dse_step2_spill ()) |
3875 | { | |
3876 | if (!did_global) | |
3877 | { | |
3878 | df_set_flags (DF_LR_RUN_DCE); | |
3879 | df_analyze (); | |
3880 | } | |
3881 | did_global = true; | |
3882 | if (dump_file) | |
3883 | fprintf (dump_file, "doing global spill processing\n"); | |
3884 | dse_step3 (true); | |
3885 | dse_step4 (); | |
3886 | dse_step5_spill (); | |
3887 | } | |
8dd5516b JJ |
3888 | |
3889 | dse_step6 (); | |
3f9b14ff | 3890 | dse_step7 (); |
6fb5fa3c DB |
3891 | |
3892 | if (dump_file) | |
3893 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d, spill deletions = %d\n", | |
3894 | locally_deleted, globally_deleted, spill_deleted); | |
3895 | return 0; | |
3896 | } | |
3897 | ||
7d817ebc DE |
3898 | static bool |
3899 | gate_dse1 (void) | |
3900 | { | |
3901 | return optimize > 0 && flag_dse | |
3902 | && dbg_cnt (dse1); | |
3903 | } | |
3904 | ||
3905 | static bool | |
3906 | gate_dse2 (void) | |
3907 | { | |
3908 | return optimize > 0 && flag_dse | |
3909 | && dbg_cnt (dse2); | |
6fb5fa3c DB |
3910 | } |
3911 | ||
8ddbbcae | 3912 | struct rtl_opt_pass pass_rtl_dse1 = |
6fb5fa3c | 3913 | { |
8ddbbcae JH |
3914 | { |
3915 | RTL_PASS, | |
6fb5fa3c | 3916 | "dse1", /* name */ |
7d817ebc | 3917 | gate_dse1, /* gate */ |
6fb5fa3c DB |
3918 | rest_of_handle_dse, /* execute */ |
3919 | NULL, /* sub */ | |
3920 | NULL, /* next */ | |
3921 | 0, /* static_pass_number */ | |
3922 | TV_DSE1, /* tv_id */ | |
3923 | 0, /* properties_required */ | |
3924 | 0, /* properties_provided */ | |
3925 | 0, /* properties_destroyed */ | |
3926 | 0, /* todo_flags_start */ | |
a36b8a1e | 3927 | TODO_df_finish | TODO_verify_rtl_sharing | |
8ddbbcae JH |
3928 | TODO_ggc_collect /* todo_flags_finish */ |
3929 | } | |
6fb5fa3c DB |
3930 | }; |
3931 | ||
8ddbbcae | 3932 | struct rtl_opt_pass pass_rtl_dse2 = |
6fb5fa3c | 3933 | { |
8ddbbcae JH |
3934 | { |
3935 | RTL_PASS, | |
6fb5fa3c | 3936 | "dse2", /* name */ |
7d817ebc | 3937 | gate_dse2, /* gate */ |
6fb5fa3c DB |
3938 | rest_of_handle_dse, /* execute */ |
3939 | NULL, /* sub */ | |
3940 | NULL, /* next */ | |
3941 | 0, /* static_pass_number */ | |
3942 | TV_DSE2, /* tv_id */ | |
3943 | 0, /* properties_required */ | |
3944 | 0, /* properties_provided */ | |
3945 | 0, /* properties_destroyed */ | |
3946 | 0, /* todo_flags_start */ | |
a36b8a1e | 3947 | TODO_df_finish | TODO_verify_rtl_sharing | |
8ddbbcae JH |
3948 | TODO_ggc_collect /* todo_flags_finish */ |
3949 | } | |
6fb5fa3c | 3950 | }; |