]>
Commit | Line | Data |
---|---|---|
6fb5fa3c | 1 | /* RTL dead store elimination. |
23a5b65a | 2 | Copyright (C) 2005-2014 Free Software Foundation, Inc. |
6fb5fa3c DB |
3 | |
4 | Contributed by Richard Sandiford <rsandifor@codesourcery.com> | |
5 | and Kenneth Zadeck <zadeck@naturalbridge.com> | |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 11 | Software Foundation; either version 3, or (at your option) any later |
6fb5fa3c DB |
12 | version. |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ | |
6fb5fa3c DB |
22 | |
23 | #undef BASELINE | |
24 | ||
25 | #include "config.h" | |
26 | #include "system.h" | |
27 | #include "coretypes.h" | |
703c8606 | 28 | #include "hash-table.h" |
6fb5fa3c DB |
29 | #include "tm.h" |
30 | #include "rtl.h" | |
31 | #include "tree.h" | |
d8a2d370 | 32 | #include "stor-layout.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" |
2fb9a547 AM |
50 | #include "tree-ssa-alias.h" |
51 | #include "internal-fn.h" | |
52 | #include "gimple-expr.h" | |
53 | #include "is-a.h" | |
442b4905 AM |
54 | #include "gimple.h" |
55 | #include "gimple-ssa.h" | |
d7111da8 | 56 | #include "rtl-iter.h" |
6fb5fa3c DB |
57 | |
58 | /* This file contains three techniques for performing Dead Store | |
b8698a0f | 59 | Elimination (dse). |
6fb5fa3c DB |
60 | |
61 | * The first technique performs dse locally on any base address. It | |
62 | is based on the cselib which is a local value numbering technique. | |
63 | This technique is local to a basic block but deals with a fairly | |
64 | general addresses. | |
b8698a0f | 65 | |
6fb5fa3c DB |
66 | * The second technique performs dse globally but is restricted to |
67 | base addresses that are either constant or are relative to the | |
68 | frame_pointer. | |
69 | ||
70 | * The third technique, (which is only done after register allocation) | |
71 | processes the spill spill slots. This differs from the second | |
72 | technique because it takes advantage of the fact that spilling is | |
73 | completely free from the effects of aliasing. | |
74 | ||
75 | Logically, dse is a backwards dataflow problem. A store can be | |
76 | deleted if it if cannot be reached in the backward direction by any | |
77 | use of the value being stored. However, the local technique uses a | |
78 | forwards scan of the basic block because cselib requires that the | |
79 | block be processed in that order. | |
80 | ||
81 | The pass is logically broken into 7 steps: | |
82 | ||
83 | 0) Initialization. | |
84 | ||
85 | 1) The local algorithm, as well as scanning the insns for the two | |
86 | global algorithms. | |
87 | ||
88 | 2) Analysis to see if the global algs are necessary. In the case | |
89 | of stores base on a constant address, there must be at least two | |
90 | stores to that address, to make it possible to delete some of the | |
91 | stores. In the case of stores off of the frame or spill related | |
92 | stores, only one store to an address is necessary because those | |
93 | stores die at the end of the function. | |
94 | ||
b8698a0f | 95 | 3) Set up the global dataflow equations based on processing the |
6fb5fa3c DB |
96 | info parsed in the first step. |
97 | ||
98 | 4) Solve the dataflow equations. | |
99 | ||
100 | 5) Delete the insns that the global analysis has indicated are | |
101 | unnecessary. | |
102 | ||
073a8998 | 103 | 6) Delete insns that store the same value as preceding store |
8dd5516b JJ |
104 | where the earlier store couldn't be eliminated. |
105 | ||
106 | 7) Cleanup. | |
6fb5fa3c DB |
107 | |
108 | This step uses cselib and canon_rtx to build the largest expression | |
109 | possible for each address. This pass is a forwards pass through | |
110 | each basic block. From the point of view of the global technique, | |
111 | the first pass could examine a block in either direction. The | |
0d52bcc1 | 112 | forwards ordering is to accommodate cselib. |
6fb5fa3c | 113 | |
5ef0b50d | 114 | We make a simplifying assumption: addresses fall into four broad |
6fb5fa3c DB |
115 | categories: |
116 | ||
117 | 1) base has rtx_varies_p == false, offset is constant. | |
118 | 2) base has rtx_varies_p == false, offset variable. | |
119 | 3) base has rtx_varies_p == true, offset constant. | |
120 | 4) base has rtx_varies_p == true, offset variable. | |
121 | ||
122 | The local passes are able to process all 4 kinds of addresses. The | |
5ef0b50d | 123 | global pass only handles 1). |
6fb5fa3c DB |
124 | |
125 | The global problem is formulated as follows: | |
126 | ||
127 | A store, S1, to address A, where A is not relative to the stack | |
128 | frame, can be eliminated if all paths from S1 to the end of the | |
5ef0b50d | 129 | function contain another store to A before a read to A. |
6fb5fa3c DB |
130 | |
131 | If the address A is relative to the stack frame, a store S2 to A | |
5ef0b50d | 132 | can be eliminated if there are no paths from S2 that reach the |
6fb5fa3c | 133 | end of the function that read A before another store to A. In |
5ef0b50d | 134 | this case S2 can be deleted if there are paths from S2 to the |
6fb5fa3c DB |
135 | end of the function that have no reads or writes to A. This |
136 | second case allows stores to the stack frame to be deleted that | |
137 | would otherwise die when the function returns. This cannot be | |
138 | done if stores_off_frame_dead_at_return is not true. See the doc | |
139 | for that variable for when this variable is false. | |
140 | ||
141 | The global problem is formulated as a backwards set union | |
142 | dataflow problem where the stores are the gens and reads are the | |
143 | kills. Set union problems are rare and require some special | |
144 | handling given our representation of bitmaps. A straightforward | |
5ef0b50d | 145 | implementation requires a lot of bitmaps filled with 1s. |
6fb5fa3c DB |
146 | These are expensive and cumbersome in our bitmap formulation so |
147 | care has been taken to avoid large vectors filled with 1s. See | |
148 | the comments in bb_info and in the dataflow confluence functions | |
b8698a0f | 149 | for details. |
6fb5fa3c DB |
150 | |
151 | There are two places for further enhancements to this algorithm: | |
b8698a0f | 152 | |
6fb5fa3c DB |
153 | 1) The original dse which was embedded in a pass called flow also |
154 | did local address forwarding. For example in | |
155 | ||
156 | A <- r100 | |
157 | ... <- A | |
158 | ||
159 | flow would replace the right hand side of the second insn with a | |
6ed3da00 | 160 | reference to r100. Most of the information is available to add this |
6fb5fa3c DB |
161 | to this pass. It has not done it because it is a lot of work in |
162 | the case that either r100 is assigned to between the first and | |
163 | second insn and/or the second insn is a load of part of the value | |
164 | stored by the first insn. | |
165 | ||
166 | insn 5 in gcc.c-torture/compile/990203-1.c simple case. | |
167 | insn 15 in gcc.c-torture/execute/20001017-2.c simple case. | |
168 | insn 25 in gcc.c-torture/execute/20001026-1.c simple case. | |
169 | insn 44 in gcc.c-torture/execute/20010910-1.c simple case. | |
170 | ||
171 | 2) The cleaning up of spill code is quite profitable. It currently | |
172 | depends on reading tea leaves and chicken entrails left by reload. | |
173 | This pass depends on reload creating a singleton alias set for each | |
174 | spill slot and telling the next dse pass which of these alias sets | |
175 | are the singletons. Rather than analyze the addresses of the | |
176 | spills, dse's spill processing just does analysis of the loads and | |
177 | stores that use those alias sets. There are three cases where this | |
178 | falls short: | |
179 | ||
180 | a) Reload sometimes creates the slot for one mode of access, and | |
181 | then inserts loads and/or stores for a smaller mode. In this | |
182 | case, the current code just punts on the slot. The proper thing | |
183 | to do is to back out and use one bit vector position for each | |
184 | byte of the entity associated with the slot. This depends on | |
185 | KNOWING that reload always generates the accesses for each of the | |
186 | bytes in some canonical (read that easy to understand several | |
187 | passes after reload happens) way. | |
188 | ||
189 | b) Reload sometimes decides that spill slot it allocated was not | |
190 | large enough for the mode and goes back and allocates more slots | |
191 | with the same mode and alias set. The backout in this case is a | |
192 | little more graceful than (a). In this case the slot is unmarked | |
193 | as being a spill slot and if final address comes out to be based | |
b8698a0f | 194 | off the frame pointer, the global algorithm handles this slot. |
6fb5fa3c DB |
195 | |
196 | c) For any pass that may prespill, there is currently no | |
197 | mechanism to tell the dse pass that the slot being used has the | |
198 | special properties that reload uses. It may be that all that is | |
0d52bcc1 | 199 | required is to have those passes make the same calls that reload |
6fb5fa3c DB |
200 | does, assuming that the alias sets can be manipulated in the same |
201 | way. */ | |
202 | ||
203 | /* There are limits to the size of constant offsets we model for the | |
204 | global problem. There are certainly test cases, that exceed this | |
205 | limit, however, it is unlikely that there are important programs | |
206 | that really have constant offsets this size. */ | |
207 | #define MAX_OFFSET (64 * 1024) | |
208 | ||
3f9b14ff SB |
209 | /* Obstack for the DSE dataflow bitmaps. We don't want to put these |
210 | on the default obstack because these bitmaps can grow quite large | |
211 | (~2GB for the small (!) test case of PR54146) and we'll hold on to | |
212 | all that memory until the end of the compiler run. | |
213 | As a bonus, delete_tree_live_info can destroy all the bitmaps by just | |
214 | releasing the whole obstack. */ | |
215 | static bitmap_obstack dse_bitmap_obstack; | |
216 | ||
217 | /* Obstack for other data. As for above: Kinda nice to be able to | |
218 | throw it all away at the end in one big sweep. */ | |
219 | static struct obstack dse_obstack; | |
220 | ||
221 | /* Scratch bitmap for cselib's cselib_expand_value_rtx. */ | |
6fb5fa3c | 222 | static bitmap scratch = NULL; |
3f9b14ff | 223 | |
6fb5fa3c DB |
224 | struct insn_info; |
225 | ||
226 | /* This structure holds information about a candidate store. */ | |
b8698a0f | 227 | struct store_info |
6fb5fa3c DB |
228 | { |
229 | ||
230 | /* False means this is a clobber. */ | |
231 | bool is_set; | |
232 | ||
8dd5516b JJ |
233 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
234 | bool is_large; | |
235 | ||
6fb5fa3c DB |
236 | /* The id of the mem group of the base address. If rtx_varies_p is |
237 | true, this is -1. Otherwise, it is the index into the group | |
238 | table. */ | |
239 | int group_id; | |
b8698a0f | 240 | |
6fb5fa3c DB |
241 | /* This is the cselib value. */ |
242 | cselib_val *cse_base; | |
243 | ||
244 | /* This canonized mem. */ | |
245 | rtx mem; | |
246 | ||
6216f94e | 247 | /* Canonized MEM address for use by canon_true_dependence. */ |
6fb5fa3c DB |
248 | rtx mem_addr; |
249 | ||
250 | /* If this is non-zero, it is the alias set of a spill location. */ | |
4862826d | 251 | alias_set_type alias_set; |
6fb5fa3c DB |
252 | |
253 | /* The offset of the first and byte before the last byte associated | |
254 | with the operation. */ | |
8dd5516b JJ |
255 | HOST_WIDE_INT begin, end; |
256 | ||
257 | union | |
258 | { | |
259 | /* A bitmask as wide as the number of bytes in the word that | |
260 | contains a 1 if the byte may be needed. The store is unused if | |
261 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
262 | unsigned HOST_WIDE_INT small_bitmask; | |
263 | ||
264 | struct | |
265 | { | |
266 | /* A bitmap with one bit per byte. Cleared bit means the position | |
267 | is needed. Used if IS_LARGE is false. */ | |
dc491a25 | 268 | bitmap bmap; |
6fb5fa3c | 269 | |
8dd5516b JJ |
270 | /* Number of set bits (i.e. unneeded bytes) in BITMAP. If it is |
271 | equal to END - BEGIN, the whole store is unused. */ | |
272 | int count; | |
273 | } large; | |
274 | } positions_needed; | |
6fb5fa3c DB |
275 | |
276 | /* The next store info for this insn. */ | |
277 | struct store_info *next; | |
278 | ||
279 | /* The right hand side of the store. This is used if there is a | |
280 | subsequent reload of the mems address somewhere later in the | |
281 | basic block. */ | |
8dd5516b JJ |
282 | rtx rhs; |
283 | ||
284 | /* If rhs is or holds a constant, this contains that constant, | |
285 | otherwise NULL. */ | |
286 | rtx const_rhs; | |
287 | ||
288 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
289 | insn stored. These aren't eliminated early, because doing that | |
290 | might prevent the earlier larger store to be eliminated. */ | |
291 | struct insn_info *redundant_reason; | |
6fb5fa3c DB |
292 | }; |
293 | ||
4fe663b0 L |
294 | /* Return a bitmask with the first N low bits set. */ |
295 | ||
296 | static unsigned HOST_WIDE_INT | |
297 | lowpart_bitmask (int n) | |
298 | { | |
299 | unsigned HOST_WIDE_INT mask = ~(unsigned HOST_WIDE_INT) 0; | |
300 | return mask >> (HOST_BITS_PER_WIDE_INT - n); | |
301 | } | |
302 | ||
6fb5fa3c DB |
303 | typedef struct store_info *store_info_t; |
304 | static alloc_pool cse_store_info_pool; | |
305 | static alloc_pool rtx_store_info_pool; | |
306 | ||
307 | /* This structure holds information about a load. These are only | |
308 | built for rtx bases. */ | |
b8698a0f | 309 | struct read_info |
6fb5fa3c DB |
310 | { |
311 | /* The id of the mem group of the base address. */ | |
312 | int group_id; | |
313 | ||
314 | /* If this is non-zero, it is the alias set of a spill location. */ | |
4862826d | 315 | alias_set_type alias_set; |
6fb5fa3c DB |
316 | |
317 | /* The offset of the first and byte after the last byte associated | |
318 | with the operation. If begin == end == 0, the read did not have | |
319 | a constant offset. */ | |
320 | int begin, end; | |
321 | ||
322 | /* The mem being read. */ | |
323 | rtx mem; | |
324 | ||
325 | /* The next read_info for this insn. */ | |
326 | struct read_info *next; | |
327 | }; | |
328 | typedef struct read_info *read_info_t; | |
329 | static alloc_pool read_info_pool; | |
330 | ||
331 | ||
332 | /* One of these records is created for each insn. */ | |
333 | ||
b8698a0f | 334 | struct insn_info |
6fb5fa3c DB |
335 | { |
336 | /* Set true if the insn contains a store but the insn itself cannot | |
337 | be deleted. This is set if the insn is a parallel and there is | |
338 | more than one non dead output or if the insn is in some way | |
339 | volatile. */ | |
340 | bool cannot_delete; | |
341 | ||
342 | /* This field is only used by the global algorithm. It is set true | |
343 | if the insn contains any read of mem except for a (1). This is | |
344 | also set if the insn is a call or has a clobber mem. If the insn | |
345 | contains a wild read, the use_rec will be null. */ | |
346 | bool wild_read; | |
347 | ||
d26c7090 ER |
348 | /* This is true only for CALL instructions which could potentially read |
349 | any non-frame memory location. This field is used by the global | |
350 | algorithm. */ | |
351 | bool non_frame_wild_read; | |
352 | ||
50f0f366 EB |
353 | /* This field is only used for the processing of const functions. |
354 | These functions cannot read memory, but they can read the stack | |
64520bdc EB |
355 | because that is where they may get their parms. We need to be |
356 | this conservative because, like the store motion pass, we don't | |
357 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
358 | Moreover, we need to distinguish two cases: | |
359 | 1. Before reload (register elimination), the stores related to | |
360 | outgoing arguments are stack pointer based and thus deemed | |
361 | of non-constant base in this pass. This requires special | |
362 | handling but also means that the frame pointer based stores | |
363 | need not be killed upon encountering a const function call. | |
364 | 2. After reload, the stores related to outgoing arguments can be | |
365 | either stack pointer or hard frame pointer based. This means | |
366 | that we have no other choice than also killing all the frame | |
367 | pointer based stores upon encountering a const function call. | |
368 | This field is set after reload for const function calls. Having | |
369 | this set is less severe than a wild read, it just means that all | |
370 | the frame related stores are killed rather than all the stores. */ | |
371 | bool frame_read; | |
372 | ||
373 | /* This field is only used for the processing of const functions. | |
374 | It is set if the insn may contain a stack pointer based store. */ | |
50f0f366 | 375 | bool stack_pointer_based; |
6fb5fa3c DB |
376 | |
377 | /* This is true if any of the sets within the store contains a | |
378 | cselib base. Such stores can only be deleted by the local | |
379 | algorithm. */ | |
380 | bool contains_cselib_groups; | |
381 | ||
382 | /* The insn. */ | |
dd60a84c | 383 | rtx_insn *insn; |
6fb5fa3c DB |
384 | |
385 | /* The list of mem sets or mem clobbers that are contained in this | |
386 | insn. If the insn is deletable, it contains only one mem set. | |
387 | But it could also contain clobbers. Insns that contain more than | |
388 | one mem set are not deletable, but each of those mems are here in | |
6ed3da00 | 389 | order to provide info to delete other insns. */ |
6fb5fa3c DB |
390 | store_info_t store_rec; |
391 | ||
392 | /* The linked list of mem uses in this insn. Only the reads from | |
393 | rtx bases are listed here. The reads to cselib bases are | |
394 | completely processed during the first scan and so are never | |
395 | created. */ | |
396 | read_info_t read_rec; | |
397 | ||
9e582b1d JR |
398 | /* The live fixed registers. We assume only fixed registers can |
399 | cause trouble by being clobbered from an expanded pattern; | |
400 | storing only the live fixed registers (rather than all registers) | |
401 | means less memory needs to be allocated / copied for the individual | |
402 | stores. */ | |
403 | regset fixed_regs_live; | |
404 | ||
6fb5fa3c DB |
405 | /* The prev insn in the basic block. */ |
406 | struct insn_info * prev_insn; | |
407 | ||
408 | /* The linked list of insns that are in consideration for removal in | |
073a8998 | 409 | the forwards pass through the basic block. This pointer may be |
6fb5fa3c | 410 | trash as it is not cleared when a wild read occurs. The only |
fa10beec | 411 | time it is guaranteed to be correct is when the traversal starts |
6fb5fa3c DB |
412 | at active_local_stores. */ |
413 | struct insn_info * next_local_store; | |
414 | }; | |
415 | ||
416 | typedef struct insn_info *insn_info_t; | |
417 | static alloc_pool insn_info_pool; | |
418 | ||
419 | /* The linked list of stores that are under consideration in this | |
b8698a0f | 420 | basic block. */ |
6fb5fa3c | 421 | static insn_info_t active_local_stores; |
dabd47e7 | 422 | static int active_local_stores_len; |
6fb5fa3c | 423 | |
b8698a0f | 424 | struct bb_info |
6fb5fa3c DB |
425 | { |
426 | ||
427 | /* Pointer to the insn info for the last insn in the block. These | |
428 | are linked so this is how all of the insns are reached. During | |
429 | scanning this is the current insn being scanned. */ | |
430 | insn_info_t last_insn; | |
431 | ||
432 | /* The info for the global dataflow problem. */ | |
433 | ||
434 | ||
435 | /* This is set if the transfer function should and in the wild_read | |
436 | bitmap before applying the kill and gen sets. That vector knocks | |
437 | out most of the bits in the bitmap and thus speeds up the | |
438 | operations. */ | |
439 | bool apply_wild_read; | |
440 | ||
02b47899 KZ |
441 | /* The following 4 bitvectors hold information about which positions |
442 | of which stores are live or dead. They are indexed by | |
443 | get_bitmap_index. */ | |
444 | ||
6fb5fa3c DB |
445 | /* The set of store positions that exist in this block before a wild read. */ |
446 | bitmap gen; | |
b8698a0f | 447 | |
6fb5fa3c DB |
448 | /* The set of load positions that exist in this block above the |
449 | same position of a store. */ | |
450 | bitmap kill; | |
451 | ||
452 | /* The set of stores that reach the top of the block without being | |
453 | killed by a read. | |
454 | ||
455 | Do not represent the in if it is all ones. Note that this is | |
456 | what the bitvector should logically be initialized to for a set | |
457 | intersection problem. However, like the kill set, this is too | |
458 | expensive. So initially, the in set will only be created for the | |
459 | exit block and any block that contains a wild read. */ | |
460 | bitmap in; | |
461 | ||
462 | /* The set of stores that reach the bottom of the block from it's | |
463 | successors. | |
464 | ||
465 | Do not represent the in if it is all ones. Note that this is | |
466 | what the bitvector should logically be initialized to for a set | |
467 | intersection problem. However, like the kill and in set, this is | |
468 | too expensive. So what is done is that the confluence operator | |
469 | just initializes the vector from one of the out sets of the | |
470 | successors of the block. */ | |
471 | bitmap out; | |
02b47899 KZ |
472 | |
473 | /* The following bitvector is indexed by the reg number. It | |
474 | contains the set of regs that are live at the current instruction | |
475 | being processed. While it contains info for all of the | |
9e582b1d JR |
476 | registers, only the hard registers are actually examined. It is used |
477 | to assure that shift and/or add sequences that are inserted do not | |
073a8998 | 478 | accidentally clobber live hard regs. */ |
02b47899 | 479 | bitmap regs_live; |
6fb5fa3c DB |
480 | }; |
481 | ||
482 | typedef struct bb_info *bb_info_t; | |
483 | static alloc_pool bb_info_pool; | |
484 | ||
485 | /* Table to hold all bb_infos. */ | |
486 | static bb_info_t *bb_table; | |
487 | ||
488 | /* There is a group_info for each rtx base that is used to reference | |
489 | memory. There are also not many of the rtx bases because they are | |
490 | very limited in scope. */ | |
491 | ||
b8698a0f | 492 | struct group_info |
6fb5fa3c DB |
493 | { |
494 | /* The actual base of the address. */ | |
495 | rtx rtx_base; | |
496 | ||
497 | /* The sequential id of the base. This allows us to have a | |
498 | canonical ordering of these that is not based on addresses. */ | |
499 | int id; | |
500 | ||
8f5929e1 JJ |
501 | /* True if there are any positions that are to be processed |
502 | globally. */ | |
503 | bool process_globally; | |
504 | ||
505 | /* True if the base of this group is either the frame_pointer or | |
506 | hard_frame_pointer. */ | |
507 | bool frame_related; | |
508 | ||
d32f725a OH |
509 | /* A mem wrapped around the base pointer for the group in order to do |
510 | read dependency. It must be given BLKmode in order to encompass all | |
511 | the possible offsets from the base. */ | |
6fb5fa3c | 512 | rtx base_mem; |
b8698a0f | 513 | |
6216f94e JJ |
514 | /* Canonized version of base_mem's address. */ |
515 | rtx canon_base_addr; | |
6fb5fa3c DB |
516 | |
517 | /* These two sets of two bitmaps are used to keep track of how many | |
6ed3da00 | 518 | stores are actually referencing that position from this base. We |
6fb5fa3c | 519 | only do this for rtx bases as this will be used to assign |
6ed3da00 | 520 | positions in the bitmaps for the global problem. Bit N is set in |
6fb5fa3c DB |
521 | store1 on the first store for offset N. Bit N is set in store2 |
522 | for the second store to offset N. This is all we need since we | |
523 | only care about offsets that have two or more stores for them. | |
524 | ||
525 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
526 | for 0 and greater offsets. | |
527 | ||
528 | There is one special case here, for stores into the stack frame, | |
529 | we will or store1 into store2 before deciding which stores look | |
530 | at globally. This is because stores to the stack frame that have | |
531 | no other reads before the end of the function can also be | |
532 | deleted. */ | |
533 | bitmap store1_n, store1_p, store2_n, store2_p; | |
534 | ||
d26c7090 ER |
535 | /* These bitmaps keep track of offsets in this group escape this function. |
536 | An offset escapes if it corresponds to a named variable whose | |
537 | addressable flag is set. */ | |
538 | bitmap escaped_n, escaped_p; | |
539 | ||
6ed3da00 | 540 | /* The positions in this bitmap have the same assignments as the in, |
6fb5fa3c | 541 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6ed3da00 | 542 | the positions that are occupied by stores for this group. */ |
6fb5fa3c DB |
543 | bitmap group_kill; |
544 | ||
6fb5fa3c | 545 | /* The offset_map is used to map the offsets from this base into |
6ed3da00 | 546 | positions in the global bitmaps. It is only created after all of |
6fb5fa3c DB |
547 | the all of stores have been scanned and we know which ones we |
548 | care about. */ | |
b8698a0f L |
549 | int *offset_map_n, *offset_map_p; |
550 | int offset_map_size_n, offset_map_size_p; | |
6fb5fa3c DB |
551 | }; |
552 | typedef struct group_info *group_info_t; | |
5f754896 | 553 | typedef const struct group_info *const_group_info_t; |
6fb5fa3c DB |
554 | static alloc_pool rtx_group_info_pool; |
555 | ||
6fb5fa3c DB |
556 | /* Index into the rtx_group_vec. */ |
557 | static int rtx_group_next_id; | |
558 | ||
6fb5fa3c | 559 | |
9771b263 | 560 | static vec<group_info_t> rtx_group_vec; |
6fb5fa3c DB |
561 | |
562 | ||
563 | /* This structure holds the set of changes that are being deferred | |
564 | when removing read operation. See replace_read. */ | |
b8698a0f | 565 | struct deferred_change |
6fb5fa3c DB |
566 | { |
567 | ||
568 | /* The mem that is being replaced. */ | |
569 | rtx *loc; | |
570 | ||
571 | /* The reg it is being replaced with. */ | |
572 | rtx reg; | |
573 | ||
574 | struct deferred_change *next; | |
575 | }; | |
576 | ||
577 | typedef struct deferred_change *deferred_change_t; | |
578 | static alloc_pool deferred_change_pool; | |
579 | ||
580 | static deferred_change_t deferred_change_list = NULL; | |
581 | ||
6fb5fa3c DB |
582 | /* The group that holds all of the clear_alias_sets. */ |
583 | static group_info_t clear_alias_group; | |
584 | ||
585 | /* The modes of the clear_alias_sets. */ | |
586 | static htab_t clear_alias_mode_table; | |
587 | ||
588 | /* Hash table element to look up the mode for an alias set. */ | |
589 | struct clear_alias_mode_holder | |
590 | { | |
4862826d | 591 | alias_set_type alias_set; |
6fb5fa3c DB |
592 | enum machine_mode mode; |
593 | }; | |
594 | ||
e3b5732b | 595 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
9dd9bf80 | 596 | this for vararg functions because they play games with the frame. */ |
6fb5fa3c DB |
597 | static bool stores_off_frame_dead_at_return; |
598 | ||
599 | /* Counter for stats. */ | |
b8698a0f L |
600 | static int globally_deleted; |
601 | static int locally_deleted; | |
602 | static int spill_deleted; | |
603 | ||
6fb5fa3c DB |
604 | static bitmap all_blocks; |
605 | ||
d26c7090 ER |
606 | /* Locations that are killed by calls in the global phase. */ |
607 | static bitmap kill_on_calls; | |
608 | ||
6fb5fa3c DB |
609 | /* The number of bits used in the global bitmaps. */ |
610 | static unsigned int current_position; | |
6fb5fa3c DB |
611 | \f |
612 | /*---------------------------------------------------------------------------- | |
613 | Zeroth step. | |
614 | ||
b8698a0f | 615 | Initialization. |
6fb5fa3c DB |
616 | ----------------------------------------------------------------------------*/ |
617 | ||
6fb5fa3c DB |
618 | |
619 | /* Find the entry associated with ALIAS_SET. */ | |
620 | ||
621 | static struct clear_alias_mode_holder * | |
4862826d | 622 | clear_alias_set_lookup (alias_set_type alias_set) |
6fb5fa3c DB |
623 | { |
624 | struct clear_alias_mode_holder tmp_holder; | |
625 | void **slot; | |
b8698a0f | 626 | |
6fb5fa3c DB |
627 | tmp_holder.alias_set = alias_set; |
628 | slot = htab_find_slot (clear_alias_mode_table, &tmp_holder, NO_INSERT); | |
629 | gcc_assert (*slot); | |
b8698a0f | 630 | |
f883e0a7 | 631 | return (struct clear_alias_mode_holder *) *slot; |
6fb5fa3c DB |
632 | } |
633 | ||
634 | ||
635 | /* Hashtable callbacks for maintaining the "bases" field of | |
636 | store_group_info, given that the addresses are function invariants. */ | |
637 | ||
703c8606 LC |
638 | struct invariant_group_base_hasher : typed_noop_remove <group_info> |
639 | { | |
5831a5f0 LC |
640 | typedef group_info value_type; |
641 | typedef group_info compare_type; | |
642 | static inline hashval_t hash (const value_type *); | |
643 | static inline bool equal (const value_type *, const compare_type *); | |
703c8606 LC |
644 | }; |
645 | ||
646 | inline bool | |
5831a5f0 LC |
647 | invariant_group_base_hasher::equal (const value_type *gi1, |
648 | const compare_type *gi2) | |
6fb5fa3c | 649 | { |
6fb5fa3c DB |
650 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
651 | } | |
652 | ||
703c8606 | 653 | inline hashval_t |
5831a5f0 | 654 | invariant_group_base_hasher::hash (const value_type *gi) |
6fb5fa3c | 655 | { |
6fb5fa3c DB |
656 | int do_not_record; |
657 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
658 | } | |
659 | ||
703c8606 | 660 | /* Tables of group_info structures, hashed by base value. */ |
c203e8a7 | 661 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
703c8606 | 662 | |
6fb5fa3c DB |
663 | |
664 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
665 | ||
666 | static group_info_t | |
667 | get_group_info (rtx base) | |
668 | { | |
b8698a0f L |
669 | struct group_info tmp_gi; |
670 | group_info_t gi; | |
703c8606 | 671 | group_info **slot; |
6fb5fa3c DB |
672 | |
673 | if (base) | |
674 | { | |
675 | /* Find the store_base_info structure for BASE, creating a new one | |
676 | if necessary. */ | |
677 | tmp_gi.rtx_base = base; | |
c203e8a7 | 678 | slot = rtx_group_table->find_slot (&tmp_gi, INSERT); |
6fb5fa3c DB |
679 | gi = (group_info_t) *slot; |
680 | } | |
681 | else | |
682 | { | |
683 | if (!clear_alias_group) | |
684 | { | |
f883e0a7 KG |
685 | clear_alias_group = gi = |
686 | (group_info_t) pool_alloc (rtx_group_info_pool); | |
6fb5fa3c DB |
687 | memset (gi, 0, sizeof (struct group_info)); |
688 | gi->id = rtx_group_next_id++; | |
3f9b14ff SB |
689 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
690 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
691 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
692 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
693 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
694 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
695 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c DB |
696 | gi->process_globally = false; |
697 | gi->offset_map_size_n = 0; | |
698 | gi->offset_map_size_p = 0; | |
699 | gi->offset_map_n = NULL; | |
700 | gi->offset_map_p = NULL; | |
9771b263 | 701 | rtx_group_vec.safe_push (gi); |
6fb5fa3c DB |
702 | } |
703 | return clear_alias_group; | |
704 | } | |
705 | ||
706 | if (gi == NULL) | |
707 | { | |
f883e0a7 | 708 | *slot = gi = (group_info_t) pool_alloc (rtx_group_info_pool); |
6fb5fa3c DB |
709 | gi->rtx_base = base; |
710 | gi->id = rtx_group_next_id++; | |
d32f725a | 711 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
6216f94e | 712 | gi->canon_base_addr = canon_rtx (base); |
3f9b14ff SB |
713 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
714 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
715 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
716 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
717 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
718 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
719 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 720 | gi->process_globally = false; |
b8698a0f | 721 | gi->frame_related = |
6fb5fa3c DB |
722 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
723 | gi->offset_map_size_n = 0; | |
724 | gi->offset_map_size_p = 0; | |
725 | gi->offset_map_n = NULL; | |
726 | gi->offset_map_p = NULL; | |
9771b263 | 727 | rtx_group_vec.safe_push (gi); |
6fb5fa3c DB |
728 | } |
729 | ||
730 | return gi; | |
731 | } | |
732 | ||
733 | ||
734 | /* Initialization of data structures. */ | |
735 | ||
736 | static void | |
737 | dse_step0 (void) | |
738 | { | |
739 | locally_deleted = 0; | |
740 | globally_deleted = 0; | |
741 | spill_deleted = 0; | |
742 | ||
3f9b14ff SB |
743 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
744 | gcc_obstack_init (&dse_obstack); | |
745 | ||
746 | scratch = BITMAP_ALLOC (®_obstack); | |
747 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c DB |
748 | |
749 | rtx_store_info_pool | |
b8698a0f | 750 | = create_alloc_pool ("rtx_store_info_pool", |
6fb5fa3c DB |
751 | sizeof (struct store_info), 100); |
752 | read_info_pool | |
b8698a0f | 753 | = create_alloc_pool ("read_info_pool", |
6fb5fa3c DB |
754 | sizeof (struct read_info), 100); |
755 | insn_info_pool | |
b8698a0f | 756 | = create_alloc_pool ("insn_info_pool", |
6fb5fa3c DB |
757 | sizeof (struct insn_info), 100); |
758 | bb_info_pool | |
b8698a0f | 759 | = create_alloc_pool ("bb_info_pool", |
6fb5fa3c DB |
760 | sizeof (struct bb_info), 100); |
761 | rtx_group_info_pool | |
b8698a0f | 762 | = create_alloc_pool ("rtx_group_info_pool", |
6fb5fa3c DB |
763 | sizeof (struct group_info), 100); |
764 | deferred_change_pool | |
b8698a0f | 765 | = create_alloc_pool ("deferred_change_pool", |
6fb5fa3c DB |
766 | sizeof (struct deferred_change), 10); |
767 | ||
c203e8a7 | 768 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
6fb5fa3c | 769 | |
8b1c6fd7 | 770 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
6fb5fa3c DB |
771 | rtx_group_next_id = 0; |
772 | ||
e3b5732b | 773 | stores_off_frame_dead_at_return = !cfun->stdarg; |
6fb5fa3c DB |
774 | |
775 | init_alias_analysis (); | |
b8698a0f | 776 | |
4b943a49 | 777 | clear_alias_group = NULL; |
6fb5fa3c DB |
778 | } |
779 | ||
780 | ||
781 | \f | |
782 | /*---------------------------------------------------------------------------- | |
783 | First step. | |
784 | ||
785 | Scan all of the insns. Any random ordering of the blocks is fine. | |
0d52bcc1 | 786 | Each block is scanned in forward order to accommodate cselib which |
6fb5fa3c DB |
787 | is used to remove stores with non-constant bases. |
788 | ----------------------------------------------------------------------------*/ | |
789 | ||
790 | /* Delete all of the store_info recs from INSN_INFO. */ | |
791 | ||
b8698a0f | 792 | static void |
6fb5fa3c DB |
793 | free_store_info (insn_info_t insn_info) |
794 | { | |
795 | store_info_t store_info = insn_info->store_rec; | |
796 | while (store_info) | |
797 | { | |
798 | store_info_t next = store_info->next; | |
8dd5516b | 799 | if (store_info->is_large) |
dc491a25 | 800 | BITMAP_FREE (store_info->positions_needed.large.bmap); |
6fb5fa3c DB |
801 | if (store_info->cse_base) |
802 | pool_free (cse_store_info_pool, store_info); | |
803 | else | |
804 | pool_free (rtx_store_info_pool, store_info); | |
805 | store_info = next; | |
806 | } | |
807 | ||
808 | insn_info->cannot_delete = true; | |
809 | insn_info->contains_cselib_groups = false; | |
810 | insn_info->store_rec = NULL; | |
811 | } | |
812 | ||
9e582b1d JR |
813 | typedef struct |
814 | { | |
815 | rtx first, current; | |
816 | regset fixed_regs_live; | |
817 | bool failure; | |
818 | } note_add_store_info; | |
819 | ||
820 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
821 | Check if a register is clobbered which is live afterwards. */ | |
822 | ||
823 | static void | |
824 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
825 | { | |
826 | rtx insn; | |
827 | note_add_store_info *info = (note_add_store_info *) data; | |
828 | int r, n; | |
829 | ||
830 | if (!REG_P (loc)) | |
831 | return; | |
832 | ||
833 | /* If this register is referenced by the current or an earlier insn, | |
834 | that's OK. E.g. this applies to the register that is being incremented | |
835 | with this addition. */ | |
836 | for (insn = info->first; | |
837 | insn != NEXT_INSN (info->current); | |
838 | insn = NEXT_INSN (insn)) | |
839 | if (reg_referenced_p (loc, PATTERN (insn))) | |
840 | return; | |
841 | ||
842 | /* If we come here, we have a clobber of a register that's only OK | |
843 | if that register is not live. If we don't have liveness information | |
844 | available, fail now. */ | |
845 | if (!info->fixed_regs_live) | |
846 | { | |
847 | info->failure = true; | |
848 | return; | |
849 | } | |
850 | /* Now check if this is a live fixed register. */ | |
851 | r = REGNO (loc); | |
852 | n = hard_regno_nregs[r][GET_MODE (loc)]; | |
853 | while (--n >= 0) | |
854 | if (REGNO_REG_SET_P (info->fixed_regs_live, r+n)) | |
855 | info->failure = true; | |
856 | } | |
857 | ||
4deef538 AO |
858 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
859 | SRC + SRCOFF before insn ARG. */ | |
6fb5fa3c DB |
860 | |
861 | static int | |
4deef538 AO |
862 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
863 | rtx op ATTRIBUTE_UNUSED, | |
864 | rtx dest, rtx src, rtx srcoff, void *arg) | |
6fb5fa3c | 865 | { |
9e582b1d JR |
866 | insn_info_t insn_info = (insn_info_t) arg; |
867 | rtx insn = insn_info->insn, new_insn, cur; | |
868 | note_add_store_info info; | |
b8698a0f | 869 | |
4deef538 AO |
870 | /* We can reuse all operands without copying, because we are about |
871 | to delete the insn that contained it. */ | |
9e582b1d | 872 | if (srcoff) |
ed079c4b JJ |
873 | { |
874 | start_sequence (); | |
875 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
876 | new_insn = get_insns (); | |
877 | end_sequence (); | |
878 | } | |
9e582b1d JR |
879 | else |
880 | new_insn = gen_move_insn (dest, src); | |
881 | info.first = new_insn; | |
882 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
883 | info.failure = false; | |
884 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
885 | { | |
886 | info.current = cur; | |
887 | note_stores (PATTERN (cur), note_add_store, &info); | |
888 | } | |
6fb5fa3c | 889 | |
9e582b1d JR |
890 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
891 | return it immediately, communicating the failure to its caller. */ | |
892 | if (info.failure) | |
893 | return 1; | |
894 | ||
895 | emit_insn_before (new_insn, insn); | |
6fb5fa3c | 896 | |
8d8e205b | 897 | return 0; |
6fb5fa3c DB |
898 | } |
899 | ||
9e582b1d JR |
900 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
901 | is there, is split into a separate insn. | |
902 | Return true on success (or if there was nothing to do), false on failure. */ | |
6fb5fa3c | 903 | |
9e582b1d JR |
904 | static bool |
905 | check_for_inc_dec_1 (insn_info_t insn_info) | |
6fb5fa3c | 906 | { |
dd60a84c | 907 | rtx_insn *insn = insn_info->insn; |
6fb5fa3c DB |
908 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
909 | if (note) | |
8d8e205b RS |
910 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
911 | insn_info) == 0; | |
9e582b1d | 912 | return true; |
6fb5fa3c DB |
913 | } |
914 | ||
915 | ||
9e582b1d JR |
916 | /* Entry point for postreload. If you work on reload_cse, or you need this |
917 | anywhere else, consider if you can provide register liveness information | |
918 | and add a parameter to this function so that it can be passed down in | |
919 | insn_info.fixed_regs_live. */ | |
920 | bool | |
dd60a84c | 921 | check_for_inc_dec (rtx_insn *insn) |
9e582b1d JR |
922 | { |
923 | struct insn_info insn_info; | |
924 | rtx note; | |
925 | ||
926 | insn_info.insn = insn; | |
927 | insn_info.fixed_regs_live = NULL; | |
928 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
929 | if (note) | |
8d8e205b RS |
930 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
931 | &insn_info) == 0; | |
9e582b1d JR |
932 | return true; |
933 | } | |
934 | ||
b8698a0f | 935 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
6fb5fa3c DB |
936 | |
937 | static void | |
938 | delete_dead_store_insn (insn_info_t insn_info) | |
939 | { | |
940 | read_info_t read_info; | |
941 | ||
942 | if (!dbg_cnt (dse)) | |
943 | return; | |
944 | ||
9e582b1d JR |
945 | if (!check_for_inc_dec_1 (insn_info)) |
946 | return; | |
456610d3 | 947 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 948 | { |
b8698a0f | 949 | fprintf (dump_file, "Locally deleting insn %d ", |
6fb5fa3c DB |
950 | INSN_UID (insn_info->insn)); |
951 | if (insn_info->store_rec->alias_set) | |
b8698a0f | 952 | fprintf (dump_file, "alias set %d\n", |
4862826d | 953 | (int) insn_info->store_rec->alias_set); |
6fb5fa3c DB |
954 | else |
955 | fprintf (dump_file, "\n"); | |
956 | } | |
957 | ||
958 | free_store_info (insn_info); | |
959 | read_info = insn_info->read_rec; | |
b8698a0f | 960 | |
6fb5fa3c DB |
961 | while (read_info) |
962 | { | |
963 | read_info_t next = read_info->next; | |
964 | pool_free (read_info_pool, read_info); | |
965 | read_info = next; | |
966 | } | |
967 | insn_info->read_rec = NULL; | |
968 | ||
969 | delete_insn (insn_info->insn); | |
970 | locally_deleted++; | |
971 | insn_info->insn = NULL; | |
972 | ||
973 | insn_info->wild_read = false; | |
974 | } | |
975 | ||
88d8330d EB |
976 | /* Return whether DECL, a local variable, can possibly escape the current |
977 | function scope. */ | |
978 | ||
979 | static bool | |
980 | local_variable_can_escape (tree decl) | |
981 | { | |
982 | if (TREE_ADDRESSABLE (decl)) | |
983 | return true; | |
984 | ||
985 | /* If this is a partitioned variable, we need to consider all the variables | |
986 | in the partition. This is necessary because a store into one of them can | |
987 | be replaced with a store into another and this may not change the outcome | |
988 | of the escape analysis. */ | |
989 | if (cfun->gimple_df->decls_to_pointers != NULL) | |
990 | { | |
39c8aaa4 | 991 | tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); |
88d8330d | 992 | if (namep) |
39c8aaa4 | 993 | return TREE_ADDRESSABLE (*namep); |
88d8330d EB |
994 | } |
995 | ||
996 | return false; | |
997 | } | |
998 | ||
999 | /* Return whether EXPR can possibly escape the current function scope. */ | |
1000 | ||
d26c7090 ER |
1001 | static bool |
1002 | can_escape (tree expr) | |
1003 | { | |
1004 | tree base; | |
1005 | if (!expr) | |
1006 | return true; | |
1007 | base = get_base_address (expr); | |
1008 | if (DECL_P (base) | |
88d8330d EB |
1009 | && !may_be_aliased (base) |
1010 | && !(TREE_CODE (base) == VAR_DECL | |
1011 | && !DECL_EXTERNAL (base) | |
1012 | && !TREE_STATIC (base) | |
1013 | && local_variable_can_escape (base))) | |
d26c7090 ER |
1014 | return false; |
1015 | return true; | |
1016 | } | |
6fb5fa3c DB |
1017 | |
1018 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
1019 | OFFSET and WIDTH. */ | |
1020 | ||
1021 | static void | |
d26c7090 ER |
1022 | set_usage_bits (group_info_t group, HOST_WIDE_INT offset, HOST_WIDE_INT width, |
1023 | tree expr) | |
6fb5fa3c DB |
1024 | { |
1025 | HOST_WIDE_INT i; | |
d26c7090 | 1026 | bool expr_escapes = can_escape (expr); |
8dd5516b | 1027 | if (offset > -MAX_OFFSET && offset + width < MAX_OFFSET) |
6fb5fa3c DB |
1028 | for (i=offset; i<offset+width; i++) |
1029 | { | |
1030 | bitmap store1; | |
1031 | bitmap store2; | |
d26c7090 | 1032 | bitmap escaped; |
6fb5fa3c DB |
1033 | int ai; |
1034 | if (i < 0) | |
1035 | { | |
1036 | store1 = group->store1_n; | |
1037 | store2 = group->store2_n; | |
d26c7090 | 1038 | escaped = group->escaped_n; |
6fb5fa3c DB |
1039 | ai = -i; |
1040 | } | |
1041 | else | |
1042 | { | |
1043 | store1 = group->store1_p; | |
1044 | store2 = group->store2_p; | |
d26c7090 | 1045 | escaped = group->escaped_p; |
6fb5fa3c DB |
1046 | ai = i; |
1047 | } | |
b8698a0f | 1048 | |
fcaa4ca4 | 1049 | if (!bitmap_set_bit (store1, ai)) |
6fb5fa3c | 1050 | bitmap_set_bit (store2, ai); |
b8698a0f | 1051 | else |
6fb5fa3c | 1052 | { |
6fb5fa3c DB |
1053 | if (i < 0) |
1054 | { | |
1055 | if (group->offset_map_size_n < ai) | |
1056 | group->offset_map_size_n = ai; | |
1057 | } | |
1058 | else | |
1059 | { | |
1060 | if (group->offset_map_size_p < ai) | |
1061 | group->offset_map_size_p = ai; | |
1062 | } | |
1063 | } | |
d26c7090 ER |
1064 | if (expr_escapes) |
1065 | bitmap_set_bit (escaped, ai); | |
6fb5fa3c DB |
1066 | } |
1067 | } | |
1068 | ||
d26c7090 ER |
1069 | static void |
1070 | reset_active_stores (void) | |
1071 | { | |
1072 | active_local_stores = NULL; | |
1073 | active_local_stores_len = 0; | |
1074 | } | |
6fb5fa3c | 1075 | |
d26c7090 | 1076 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
6fb5fa3c DB |
1077 | |
1078 | static void | |
d26c7090 | 1079 | free_read_records (bb_info_t bb_info) |
6fb5fa3c DB |
1080 | { |
1081 | insn_info_t insn_info = bb_info->last_insn; | |
1082 | read_info_t *ptr = &insn_info->read_rec; | |
6fb5fa3c DB |
1083 | while (*ptr) |
1084 | { | |
1085 | read_info_t next = (*ptr)->next; | |
4862826d | 1086 | if ((*ptr)->alias_set == 0) |
6fb5fa3c DB |
1087 | { |
1088 | pool_free (read_info_pool, *ptr); | |
1089 | *ptr = next; | |
d26c7090 | 1090 | } |
b8698a0f | 1091 | else |
d26c7090 | 1092 | ptr = &(*ptr)->next; |
6fb5fa3c | 1093 | } |
d26c7090 ER |
1094 | } |
1095 | ||
1096 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1097 | ||
1098 | static void | |
1099 | add_wild_read (bb_info_t bb_info) | |
1100 | { | |
1101 | insn_info_t insn_info = bb_info->last_insn; | |
6fb5fa3c | 1102 | insn_info->wild_read = true; |
d26c7090 ER |
1103 | free_read_records (bb_info); |
1104 | reset_active_stores (); | |
6fb5fa3c DB |
1105 | } |
1106 | ||
d26c7090 ER |
1107 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1108 | non-frame locations. */ | |
1109 | ||
1110 | static void | |
1111 | add_non_frame_wild_read (bb_info_t bb_info) | |
1112 | { | |
1113 | insn_info_t insn_info = bb_info->last_insn; | |
1114 | insn_info->non_frame_wild_read = true; | |
1115 | free_read_records (bb_info); | |
1116 | reset_active_stores (); | |
1117 | } | |
6fb5fa3c | 1118 | |
50f0f366 EB |
1119 | /* Return true if X is a constant or one of the registers that behave |
1120 | as a constant over the life of a function. This is equivalent to | |
1121 | !rtx_varies_p for memory addresses. */ | |
6fb5fa3c DB |
1122 | |
1123 | static bool | |
1124 | const_or_frame_p (rtx x) | |
1125 | { | |
d8116890 KZ |
1126 | if (CONSTANT_P (x)) |
1127 | return true; | |
1128 | ||
1129 | if (GET_CODE (x) == REG) | |
6fb5fa3c | 1130 | { |
6fb5fa3c DB |
1131 | /* Note that we have to test for the actual rtx used for the frame |
1132 | and arg pointers and not just the register number in case we have | |
1133 | eliminated the frame and/or arg pointer and are using it | |
1134 | for pseudos. */ | |
1135 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1136 | /* The arg pointer varies if it is not a fixed register. */ | |
1137 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1138 | || x == pic_offset_table_rtx) | |
1139 | return true; | |
1140 | return false; | |
6fb5fa3c | 1141 | } |
d8116890 KZ |
1142 | |
1143 | return false; | |
6fb5fa3c DB |
1144 | } |
1145 | ||
b8698a0f L |
1146 | /* Take all reasonable action to put the address of MEM into the form |
1147 | that we can do analysis on. | |
6fb5fa3c DB |
1148 | |
1149 | The gold standard is to get the address into the form: address + | |
1150 | OFFSET where address is something that rtx_varies_p considers a | |
1151 | constant. When we can get the address in this form, we can do | |
1152 | global analysis on it. Note that for constant bases, address is | |
1153 | not actually returned, only the group_id. The address can be | |
1154 | obtained from that. | |
1155 | ||
1156 | If that fails, we try cselib to get a value we can at least use | |
b8698a0f L |
1157 | locally. If that fails we return false. |
1158 | ||
6fb5fa3c DB |
1159 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1160 | group for non_varying bases. | |
1161 | ||
1162 | FOR_READ is true if this is a mem read and false if not. */ | |
1163 | ||
1164 | static bool | |
1165 | canon_address (rtx mem, | |
4862826d | 1166 | alias_set_type *alias_set_out, |
6fb5fa3c | 1167 | int *group_id, |
b8698a0f | 1168 | HOST_WIDE_INT *offset, |
6fb5fa3c DB |
1169 | cselib_val **base) |
1170 | { | |
372d6395 | 1171 | enum machine_mode address_mode = get_address_mode (mem); |
6fb5fa3c DB |
1172 | rtx mem_address = XEXP (mem, 0); |
1173 | rtx expanded_address, address; | |
403c7520 JJ |
1174 | int expanded; |
1175 | ||
6fb5fa3c DB |
1176 | *alias_set_out = 0; |
1177 | ||
4deef538 | 1178 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
6fb5fa3c | 1179 | |
456610d3 | 1180 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1181 | { |
1182 | fprintf (dump_file, " mem: "); | |
1183 | print_inline_rtx (dump_file, mem_address, 0); | |
1184 | fprintf (dump_file, "\n"); | |
1185 | } | |
1186 | ||
403c7520 JJ |
1187 | /* First see if just canon_rtx (mem_address) is const or frame, |
1188 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1189 | address = NULL_RTX; | |
1190 | for (expanded = 0; expanded < 2; expanded++) | |
1191 | { | |
1192 | if (expanded) | |
1193 | { | |
1194 | /* Use cselib to replace all of the reg references with the full | |
b8698a0f | 1195 | expression. This will take care of the case where we have |
6fb5fa3c | 1196 | |
403c7520 JJ |
1197 | r_x = base + offset; |
1198 | val = *r_x; | |
b8698a0f L |
1199 | |
1200 | by making it into | |
6fb5fa3c | 1201 | |
403c7520 | 1202 | val = *(base + offset); */ |
6fb5fa3c | 1203 | |
403c7520 JJ |
1204 | expanded_address = cselib_expand_value_rtx (mem_address, |
1205 | scratch, 5); | |
6fb5fa3c | 1206 | |
403c7520 JJ |
1207 | /* If this fails, just go with the address from first |
1208 | iteration. */ | |
1209 | if (!expanded_address) | |
1210 | break; | |
1211 | } | |
1212 | else | |
1213 | expanded_address = mem_address; | |
6fb5fa3c | 1214 | |
403c7520 JJ |
1215 | /* Split the address into canonical BASE + OFFSET terms. */ |
1216 | address = canon_rtx (expanded_address); | |
6fb5fa3c | 1217 | |
403c7520 | 1218 | *offset = 0; |
6fb5fa3c | 1219 | |
456610d3 | 1220 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1221 | { |
1222 | if (expanded) | |
1223 | { | |
1224 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1225 | print_inline_rtx (dump_file, expanded_address, 0); | |
1226 | fprintf (dump_file, "\n"); | |
1227 | } | |
6fb5fa3c | 1228 | |
403c7520 JJ |
1229 | fprintf (dump_file, "\n after canon_rtx address: "); |
1230 | print_inline_rtx (dump_file, address, 0); | |
1231 | fprintf (dump_file, "\n"); | |
1232 | } | |
6fb5fa3c | 1233 | |
403c7520 JJ |
1234 | if (GET_CODE (address) == CONST) |
1235 | address = XEXP (address, 0); | |
6fb5fa3c | 1236 | |
403c7520 JJ |
1237 | if (GET_CODE (address) == PLUS |
1238 | && CONST_INT_P (XEXP (address, 1))) | |
1239 | { | |
1240 | *offset = INTVAL (XEXP (address, 1)); | |
1241 | address = XEXP (address, 0); | |
1242 | } | |
6fb5fa3c | 1243 | |
09e881c9 BE |
1244 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1245 | && const_or_frame_p (address)) | |
6fb5fa3c | 1246 | { |
403c7520 JJ |
1247 | group_info_t group = get_group_info (address); |
1248 | ||
456610d3 | 1249 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1250 | fprintf (dump_file, " gid=%d offset=%d \n", |
1251 | group->id, (int)*offset); | |
1252 | *base = NULL; | |
1253 | *group_id = group->id; | |
1254 | return true; | |
6fb5fa3c | 1255 | } |
403c7520 JJ |
1256 | } |
1257 | ||
4deef538 | 1258 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
403c7520 JJ |
1259 | *group_id = -1; |
1260 | ||
1261 | if (*base == NULL) | |
1262 | { | |
456610d3 | 1263 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1264 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1265 | return false; | |
6fb5fa3c | 1266 | } |
456610d3 | 1267 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5440c0e7 AO |
1268 | fprintf (dump_file, " varying cselib base=%u:%u offset = %d\n", |
1269 | (*base)->uid, (*base)->hash, (int)*offset); | |
6fb5fa3c DB |
1270 | return true; |
1271 | } | |
1272 | ||
1273 | ||
1274 | /* Clear the rhs field from the active_local_stores array. */ | |
1275 | ||
1276 | static void | |
1277 | clear_rhs_from_active_local_stores (void) | |
1278 | { | |
1279 | insn_info_t ptr = active_local_stores; | |
1280 | ||
1281 | while (ptr) | |
1282 | { | |
1283 | store_info_t store_info = ptr->store_rec; | |
1284 | /* Skip the clobbers. */ | |
1285 | while (!store_info->is_set) | |
1286 | store_info = store_info->next; | |
1287 | ||
1288 | store_info->rhs = NULL; | |
8dd5516b | 1289 | store_info->const_rhs = NULL; |
6fb5fa3c DB |
1290 | |
1291 | ptr = ptr->next_local_store; | |
1292 | } | |
1293 | } | |
1294 | ||
1295 | ||
8dd5516b JJ |
1296 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1297 | ||
1298 | static inline void | |
1299 | set_position_unneeded (store_info_t s_info, int pos) | |
1300 | { | |
1301 | if (__builtin_expect (s_info->is_large, false)) | |
1302 | { | |
fcaa4ca4 NF |
1303 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1304 | s_info->positions_needed.large.count++; | |
8dd5516b JJ |
1305 | } |
1306 | else | |
1307 | s_info->positions_needed.small_bitmask | |
1308 | &= ~(((unsigned HOST_WIDE_INT) 1) << pos); | |
1309 | } | |
1310 | ||
1311 | /* Mark the whole store S_INFO as unneeded. */ | |
1312 | ||
1313 | static inline void | |
1314 | set_all_positions_unneeded (store_info_t s_info) | |
1315 | { | |
1316 | if (__builtin_expect (s_info->is_large, false)) | |
1317 | { | |
1318 | int pos, end = s_info->end - s_info->begin; | |
1319 | for (pos = 0; pos < end; pos++) | |
dc491a25 | 1320 | bitmap_set_bit (s_info->positions_needed.large.bmap, pos); |
8dd5516b JJ |
1321 | s_info->positions_needed.large.count = end; |
1322 | } | |
1323 | else | |
1324 | s_info->positions_needed.small_bitmask = (unsigned HOST_WIDE_INT) 0; | |
1325 | } | |
1326 | ||
1327 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1328 | ||
1329 | static inline bool | |
1330 | any_positions_needed_p (store_info_t s_info) | |
1331 | { | |
1332 | if (__builtin_expect (s_info->is_large, false)) | |
1333 | return (s_info->positions_needed.large.count | |
1334 | < s_info->end - s_info->begin); | |
1335 | else | |
1336 | return (s_info->positions_needed.small_bitmask | |
1337 | != (unsigned HOST_WIDE_INT) 0); | |
1338 | } | |
1339 | ||
1340 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
1341 | store are needed. */ | |
1342 | ||
1343 | static inline bool | |
1344 | all_positions_needed_p (store_info_t s_info, int start, int width) | |
1345 | { | |
1346 | if (__builtin_expect (s_info->is_large, false)) | |
1347 | { | |
1348 | int end = start + width; | |
1349 | while (start < end) | |
dc491a25 | 1350 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, start++)) |
8dd5516b JJ |
1351 | return false; |
1352 | return true; | |
1353 | } | |
1354 | else | |
1355 | { | |
1356 | unsigned HOST_WIDE_INT mask = lowpart_bitmask (width) << start; | |
1357 | return (s_info->positions_needed.small_bitmask & mask) == mask; | |
1358 | } | |
1359 | } | |
1360 | ||
1361 | ||
1362 | static rtx get_stored_val (store_info_t, enum machine_mode, HOST_WIDE_INT, | |
1363 | HOST_WIDE_INT, basic_block, bool); | |
1364 | ||
1365 | ||
6fb5fa3c DB |
1366 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1367 | there is a candidate store, after adding it to the appropriate | |
1368 | local store group if so. */ | |
1369 | ||
1370 | static int | |
1371 | record_store (rtx body, bb_info_t bb_info) | |
1372 | { | |
6216f94e | 1373 | rtx mem, rhs, const_rhs, mem_addr; |
6fb5fa3c DB |
1374 | HOST_WIDE_INT offset = 0; |
1375 | HOST_WIDE_INT width = 0; | |
4862826d | 1376 | alias_set_type spill_alias_set; |
6fb5fa3c DB |
1377 | insn_info_t insn_info = bb_info->last_insn; |
1378 | store_info_t store_info = NULL; | |
1379 | int group_id; | |
1380 | cselib_val *base = NULL; | |
8dd5516b | 1381 | insn_info_t ptr, last, redundant_reason; |
6fb5fa3c DB |
1382 | bool store_is_unused; |
1383 | ||
1384 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1385 | return 0; | |
1386 | ||
8dd5516b JJ |
1387 | mem = SET_DEST (body); |
1388 | ||
6fb5fa3c DB |
1389 | /* If this is not used, then this cannot be used to keep the insn |
1390 | from being deleted. On the other hand, it does provide something | |
1391 | that can be used to prove that another store is dead. */ | |
1392 | store_is_unused | |
8dd5516b | 1393 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
6fb5fa3c DB |
1394 | |
1395 | /* Check whether that value is a suitable memory location. */ | |
6fb5fa3c DB |
1396 | if (!MEM_P (mem)) |
1397 | { | |
1398 | /* If the set or clobber is unused, then it does not effect our | |
1399 | ability to get rid of the entire insn. */ | |
1400 | if (!store_is_unused) | |
1401 | insn_info->cannot_delete = true; | |
1402 | return 0; | |
1403 | } | |
1404 | ||
1405 | /* At this point we know mem is a mem. */ | |
1406 | if (GET_MODE (mem) == BLKmode) | |
1407 | { | |
1408 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
1409 | { | |
456610d3 | 1410 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1411 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1412 | add_wild_read (bb_info); | |
1413 | insn_info->cannot_delete = true; | |
8dd5516b | 1414 | return 0; |
6fb5fa3c | 1415 | } |
8dd5516b JJ |
1416 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1417 | as memset (addr, 0, 36); */ | |
f5541398 RS |
1418 | else if (!MEM_SIZE_KNOWN_P (mem) |
1419 | || MEM_SIZE (mem) <= 0 | |
1420 | || MEM_SIZE (mem) > MAX_OFFSET | |
8dd5516b | 1421 | || GET_CODE (body) != SET |
8dd5516b | 1422 | || !CONST_INT_P (SET_SRC (body))) |
6fb5fa3c | 1423 | { |
8dd5516b JJ |
1424 | if (!store_is_unused) |
1425 | { | |
1426 | /* If the set or clobber is unused, then it does not effect our | |
1427 | ability to get rid of the entire insn. */ | |
1428 | insn_info->cannot_delete = true; | |
1429 | clear_rhs_from_active_local_stores (); | |
1430 | } | |
1431 | return 0; | |
6fb5fa3c | 1432 | } |
6fb5fa3c DB |
1433 | } |
1434 | ||
1435 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1436 | if (MEM_VOLATILE_P (mem)) | |
8dd5516b | 1437 | insn_info->cannot_delete = true; |
6fb5fa3c DB |
1438 | |
1439 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
1440 | { | |
1441 | clear_rhs_from_active_local_stores (); | |
1442 | return 0; | |
1443 | } | |
1444 | ||
8dd5516b | 1445 | if (GET_MODE (mem) == BLKmode) |
f5541398 | 1446 | width = MEM_SIZE (mem); |
8dd5516b | 1447 | else |
6f557e0e | 1448 | width = GET_MODE_SIZE (GET_MODE (mem)); |
6fb5fa3c DB |
1449 | |
1450 | if (spill_alias_set) | |
1451 | { | |
1452 | bitmap store1 = clear_alias_group->store1_p; | |
1453 | bitmap store2 = clear_alias_group->store2_p; | |
8dd5516b JJ |
1454 | |
1455 | gcc_assert (GET_MODE (mem) != BLKmode); | |
b8698a0f | 1456 | |
fcaa4ca4 | 1457 | if (!bitmap_set_bit (store1, spill_alias_set)) |
6fb5fa3c | 1458 | bitmap_set_bit (store2, spill_alias_set); |
b8698a0f | 1459 | |
6fb5fa3c DB |
1460 | if (clear_alias_group->offset_map_size_p < spill_alias_set) |
1461 | clear_alias_group->offset_map_size_p = spill_alias_set; | |
b8698a0f | 1462 | |
f883e0a7 | 1463 | store_info = (store_info_t) pool_alloc (rtx_store_info_pool); |
6fb5fa3c | 1464 | |
456610d3 | 1465 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 1466 | fprintf (dump_file, " processing spill store %d(%s)\n", |
4862826d | 1467 | (int) spill_alias_set, GET_MODE_NAME (GET_MODE (mem))); |
6fb5fa3c DB |
1468 | } |
1469 | else if (group_id >= 0) | |
1470 | { | |
1471 | /* In the restrictive case where the base is a constant or the | |
1472 | frame pointer we can do global analysis. */ | |
b8698a0f L |
1473 | |
1474 | group_info_t group | |
9771b263 | 1475 | = rtx_group_vec[group_id]; |
d26c7090 | 1476 | tree expr = MEM_EXPR (mem); |
b8698a0f | 1477 | |
f883e0a7 | 1478 | store_info = (store_info_t) pool_alloc (rtx_store_info_pool); |
d26c7090 | 1479 | set_usage_bits (group, offset, width, expr); |
6fb5fa3c | 1480 | |
456610d3 | 1481 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1482 | fprintf (dump_file, " processing const base store gid=%d[%d..%d)\n", |
1483 | group_id, (int)offset, (int)(offset+width)); | |
1484 | } | |
1485 | else | |
1486 | { | |
9e412ca3 | 1487 | if (may_be_sp_based_p (XEXP (mem, 0))) |
50f0f366 | 1488 | insn_info->stack_pointer_based = true; |
6fb5fa3c | 1489 | insn_info->contains_cselib_groups = true; |
50f0f366 | 1490 | |
f883e0a7 | 1491 | store_info = (store_info_t) pool_alloc (cse_store_info_pool); |
6fb5fa3c DB |
1492 | group_id = -1; |
1493 | ||
456610d3 | 1494 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1495 | fprintf (dump_file, " processing cselib store [%d..%d)\n", |
1496 | (int)offset, (int)(offset+width)); | |
1497 | } | |
1498 | ||
8dd5516b JJ |
1499 | const_rhs = rhs = NULL_RTX; |
1500 | if (GET_CODE (body) == SET | |
1501 | /* No place to keep the value after ra. */ | |
1502 | && !reload_completed | |
1503 | && (REG_P (SET_SRC (body)) | |
1504 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1505 | || CONSTANT_P (SET_SRC (body))) | |
1506 | && !MEM_VOLATILE_P (mem) | |
1507 | /* Sometimes the store and reload is used for truncation and | |
1508 | rounding. */ | |
1509 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1510 | { | |
1511 | rhs = SET_SRC (body); | |
1512 | if (CONSTANT_P (rhs)) | |
1513 | const_rhs = rhs; | |
1514 | else if (body == PATTERN (insn_info->insn)) | |
1515 | { | |
1516 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1517 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1518 | const_rhs = XEXP (tem, 0); | |
1519 | } | |
1520 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1521 | { | |
1522 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1523 | ||
1524 | if (tem && CONSTANT_P (tem)) | |
1525 | const_rhs = tem; | |
1526 | } | |
1527 | } | |
1528 | ||
6fb5fa3c DB |
1529 | /* Check to see if this stores causes some other stores to be |
1530 | dead. */ | |
1531 | ptr = active_local_stores; | |
1532 | last = NULL; | |
8dd5516b | 1533 | redundant_reason = NULL; |
6216f94e JJ |
1534 | mem = canon_rtx (mem); |
1535 | /* For alias_set != 0 canon_true_dependence should be never called. */ | |
1536 | if (spill_alias_set) | |
1537 | mem_addr = NULL_RTX; | |
1538 | else | |
1539 | { | |
1540 | if (group_id < 0) | |
1541 | mem_addr = base->val_rtx; | |
1542 | else | |
1543 | { | |
1544 | group_info_t group | |
9771b263 | 1545 | = rtx_group_vec[group_id]; |
6216f94e JJ |
1546 | mem_addr = group->canon_base_addr; |
1547 | } | |
1548 | if (offset) | |
372d6395 | 1549 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6216f94e | 1550 | } |
6fb5fa3c DB |
1551 | |
1552 | while (ptr) | |
1553 | { | |
1554 | insn_info_t next = ptr->next_local_store; | |
1555 | store_info_t s_info = ptr->store_rec; | |
60564289 | 1556 | bool del = true; |
6fb5fa3c DB |
1557 | |
1558 | /* Skip the clobbers. We delete the active insn if this insn | |
6ed3da00 | 1559 | shadows the set. To have been put on the active list, it |
6fb5fa3c DB |
1560 | has exactly on set. */ |
1561 | while (!s_info->is_set) | |
1562 | s_info = s_info->next; | |
1563 | ||
1564 | if (s_info->alias_set != spill_alias_set) | |
60564289 | 1565 | del = false; |
6fb5fa3c DB |
1566 | else if (s_info->alias_set) |
1567 | { | |
b8698a0f | 1568 | struct clear_alias_mode_holder *entry |
6fb5fa3c DB |
1569 | = clear_alias_set_lookup (s_info->alias_set); |
1570 | /* Generally, spills cannot be processed if and of the | |
1571 | references to the slot have a different mode. But if | |
1572 | we are in the same block and mode is exactly the same | |
1573 | between this store and one before in the same block, | |
1574 | we can still delete it. */ | |
1575 | if ((GET_MODE (mem) == GET_MODE (s_info->mem)) | |
1576 | && (GET_MODE (mem) == entry->mode)) | |
1577 | { | |
60564289 | 1578 | del = true; |
8dd5516b | 1579 | set_all_positions_unneeded (s_info); |
6fb5fa3c | 1580 | } |
456610d3 | 1581 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 1582 | fprintf (dump_file, " trying spill store in insn=%d alias_set=%d\n", |
4862826d | 1583 | INSN_UID (ptr->insn), (int) s_info->alias_set); |
6fb5fa3c | 1584 | } |
b8698a0f | 1585 | else if ((s_info->group_id == group_id) |
6fb5fa3c DB |
1586 | && (s_info->cse_base == base)) |
1587 | { | |
1588 | HOST_WIDE_INT i; | |
456610d3 | 1589 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 1590 | fprintf (dump_file, " trying store in insn=%d gid=%d[%d..%d)\n", |
b8698a0f | 1591 | INSN_UID (ptr->insn), s_info->group_id, |
6fb5fa3c | 1592 | (int)s_info->begin, (int)s_info->end); |
8dd5516b JJ |
1593 | |
1594 | /* Even if PTR won't be eliminated as unneeded, if both | |
1595 | PTR and this insn store the same constant value, we might | |
1596 | eliminate this insn instead. */ | |
1597 | if (s_info->const_rhs | |
1598 | && const_rhs | |
1599 | && offset >= s_info->begin | |
1600 | && offset + width <= s_info->end | |
1601 | && all_positions_needed_p (s_info, offset - s_info->begin, | |
1602 | width)) | |
1603 | { | |
1604 | if (GET_MODE (mem) == BLKmode) | |
1605 | { | |
1606 | if (GET_MODE (s_info->mem) == BLKmode | |
1607 | && s_info->const_rhs == const_rhs) | |
1608 | redundant_reason = ptr; | |
1609 | } | |
1610 | else if (s_info->const_rhs == const0_rtx | |
1611 | && const_rhs == const0_rtx) | |
1612 | redundant_reason = ptr; | |
1613 | else | |
1614 | { | |
1615 | rtx val; | |
1616 | start_sequence (); | |
1617 | val = get_stored_val (s_info, GET_MODE (mem), | |
1618 | offset, offset + width, | |
1619 | BLOCK_FOR_INSN (insn_info->insn), | |
1620 | true); | |
1621 | if (get_insns () != NULL) | |
1622 | val = NULL_RTX; | |
1623 | end_sequence (); | |
1624 | if (val && rtx_equal_p (val, const_rhs)) | |
1625 | redundant_reason = ptr; | |
1626 | } | |
1627 | } | |
1628 | ||
1629 | for (i = MAX (offset, s_info->begin); | |
1630 | i < offset + width && i < s_info->end; | |
1631 | i++) | |
1632 | set_position_unneeded (s_info, i - s_info->begin); | |
6fb5fa3c DB |
1633 | } |
1634 | else if (s_info->rhs) | |
1635 | /* Need to see if it is possible for this store to overwrite | |
1636 | the value of store_info. If it is, set the rhs to NULL to | |
1637 | keep it from being used to remove a load. */ | |
1638 | { | |
b8698a0f | 1639 | if (canon_true_dependence (s_info->mem, |
6fb5fa3c DB |
1640 | GET_MODE (s_info->mem), |
1641 | s_info->mem_addr, | |
53d9622b | 1642 | mem, mem_addr)) |
8dd5516b JJ |
1643 | { |
1644 | s_info->rhs = NULL; | |
1645 | s_info->const_rhs = NULL; | |
1646 | } | |
6fb5fa3c | 1647 | } |
6216f94e | 1648 | |
6fb5fa3c DB |
1649 | /* An insn can be deleted if every position of every one of |
1650 | its s_infos is zero. */ | |
1b6fa860 | 1651 | if (any_positions_needed_p (s_info)) |
60564289 | 1652 | del = false; |
8dd5516b | 1653 | |
60564289 | 1654 | if (del) |
6fb5fa3c DB |
1655 | { |
1656 | insn_info_t insn_to_delete = ptr; | |
b8698a0f | 1657 | |
dabd47e7 | 1658 | active_local_stores_len--; |
6fb5fa3c DB |
1659 | if (last) |
1660 | last->next_local_store = ptr->next_local_store; | |
1661 | else | |
1662 | active_local_stores = ptr->next_local_store; | |
b8698a0f | 1663 | |
1b6fa860 JJ |
1664 | if (!insn_to_delete->cannot_delete) |
1665 | delete_dead_store_insn (insn_to_delete); | |
6fb5fa3c DB |
1666 | } |
1667 | else | |
1668 | last = ptr; | |
b8698a0f | 1669 | |
6fb5fa3c DB |
1670 | ptr = next; |
1671 | } | |
b8698a0f | 1672 | |
6fb5fa3c DB |
1673 | /* Finish filling in the store_info. */ |
1674 | store_info->next = insn_info->store_rec; | |
1675 | insn_info->store_rec = store_info; | |
6216f94e | 1676 | store_info->mem = mem; |
6fb5fa3c | 1677 | store_info->alias_set = spill_alias_set; |
6216f94e | 1678 | store_info->mem_addr = mem_addr; |
6fb5fa3c | 1679 | store_info->cse_base = base; |
8dd5516b JJ |
1680 | if (width > HOST_BITS_PER_WIDE_INT) |
1681 | { | |
1682 | store_info->is_large = true; | |
1683 | store_info->positions_needed.large.count = 0; | |
3f9b14ff | 1684 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
8dd5516b JJ |
1685 | } |
1686 | else | |
1687 | { | |
1688 | store_info->is_large = false; | |
1689 | store_info->positions_needed.small_bitmask = lowpart_bitmask (width); | |
1690 | } | |
6fb5fa3c DB |
1691 | store_info->group_id = group_id; |
1692 | store_info->begin = offset; | |
1693 | store_info->end = offset + width; | |
1694 | store_info->is_set = GET_CODE (body) == SET; | |
8dd5516b JJ |
1695 | store_info->rhs = rhs; |
1696 | store_info->const_rhs = const_rhs; | |
1697 | store_info->redundant_reason = redundant_reason; | |
6fb5fa3c | 1698 | |
6fb5fa3c DB |
1699 | /* If this is a clobber, we return 0. We will only be able to |
1700 | delete this insn if there is only one store USED store, but we | |
1701 | can use the clobber to delete other stores earlier. */ | |
1702 | return store_info->is_set ? 1 : 0; | |
1703 | } | |
1704 | ||
1705 | ||
1706 | static void | |
1707 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1708 | { | |
b8698a0f | 1709 | fprintf (dump_file, "%s insn=%d %s\n", start, |
6fb5fa3c DB |
1710 | INSN_UID (insn_info->insn), |
1711 | insn_info->store_rec ? "has store" : "naked"); | |
1712 | } | |
1713 | ||
1714 | ||
8660aaae EC |
1715 | /* If the modes are different and the value's source and target do not |
1716 | line up, we need to extract the value from lower part of the rhs of | |
1717 | the store, shift it, and then put it into a form that can be shoved | |
1718 | into the read_insn. This function generates a right SHIFT of a | |
1719 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1720 | shift sequence is returned or NULL if we failed to find a | |
1721 | shift. */ | |
1722 | ||
1723 | static rtx | |
18b526e8 | 1724 | find_shift_sequence (int access_size, |
8660aaae | 1725 | store_info_t store_info, |
8dd5516b JJ |
1726 | enum machine_mode read_mode, |
1727 | int shift, bool speed, bool require_cst) | |
8660aaae EC |
1728 | { |
1729 | enum machine_mode store_mode = GET_MODE (store_info->mem); | |
18b526e8 RS |
1730 | enum machine_mode new_mode; |
1731 | rtx read_reg = NULL; | |
8660aaae EC |
1732 | |
1733 | /* Some machines like the x86 have shift insns for each size of | |
1734 | operand. Other machines like the ppc or the ia-64 may only have | |
1735 | shift insns that shift values within 32 or 64 bit registers. | |
1736 | This loop tries to find the smallest shift insn that will right | |
1737 | justify the value we want to read but is available in one insn on | |
1738 | the machine. */ | |
1739 | ||
18b526e8 RS |
1740 | for (new_mode = smallest_mode_for_size (access_size * BITS_PER_UNIT, |
1741 | MODE_INT); | |
1742 | GET_MODE_BITSIZE (new_mode) <= BITS_PER_WORD; | |
1743 | new_mode = GET_MODE_WIDER_MODE (new_mode)) | |
8660aaae | 1744 | { |
18b526e8 | 1745 | rtx target, new_reg, shift_seq, insn, new_lhs; |
d898d29b | 1746 | int cost; |
348eea5f | 1747 | |
72a2609f JJ |
1748 | /* If a constant was stored into memory, try to simplify it here, |
1749 | otherwise the cost of the shift might preclude this optimization | |
1750 | e.g. at -Os, even when no actual shift will be needed. */ | |
8dd5516b | 1751 | if (store_info->const_rhs) |
72a2609f JJ |
1752 | { |
1753 | unsigned int byte = subreg_lowpart_offset (new_mode, store_mode); | |
8dd5516b JJ |
1754 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1755 | store_mode, byte); | |
72a2609f JJ |
1756 | if (ret && CONSTANT_P (ret)) |
1757 | { | |
1758 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, | |
1759 | ret, GEN_INT (shift)); | |
1760 | if (ret && CONSTANT_P (ret)) | |
1761 | { | |
1762 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1763 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1764 | if (ret && CONSTANT_P (ret) | |
5e8f01f4 | 1765 | && set_src_cost (ret, speed) <= COSTS_N_INSNS (1)) |
72a2609f JJ |
1766 | return ret; |
1767 | } | |
1768 | } | |
1769 | } | |
1770 | ||
8dd5516b JJ |
1771 | if (require_cst) |
1772 | return NULL_RTX; | |
1773 | ||
18b526e8 RS |
1774 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1775 | requires a real instruction. */ | |
1776 | if (GET_MODE_BITSIZE (new_mode) < GET_MODE_BITSIZE (store_mode) | |
d0edd768 | 1777 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
348eea5f RS |
1778 | continue; |
1779 | ||
18b526e8 RS |
1780 | /* Also try a wider mode if the necessary punning is either not |
1781 | desirable or not possible. */ | |
1782 | if (!CONSTANT_P (store_info->rhs) | |
1783 | && !MODES_TIEABLE_P (new_mode, store_mode)) | |
1784 | continue; | |
18b526e8 | 1785 | |
348eea5f | 1786 | new_reg = gen_reg_rtx (new_mode); |
8660aaae EC |
1787 | |
1788 | start_sequence (); | |
1789 | ||
1790 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1791 | of one dsp where the cost of these two was not the same. But | |
1792 | this really is a rare case anyway. */ | |
1793 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
1794 | GEN_INT (shift), new_reg, 1, OPTAB_DIRECT); | |
1795 | ||
c6f3019a RS |
1796 | shift_seq = get_insns (); |
1797 | end_sequence (); | |
8660aaae | 1798 | |
c6f3019a RS |
1799 | if (target != new_reg || shift_seq == NULL) |
1800 | continue; | |
1801 | ||
1802 | cost = 0; | |
1803 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1804 | if (INSN_P (insn)) | |
f40751dd | 1805 | cost += insn_rtx_cost (PATTERN (insn), speed); |
c6f3019a RS |
1806 | |
1807 | /* The computation up to here is essentially independent | |
1808 | of the arguments and could be precomputed. It may | |
1809 | not be worth doing so. We could precompute if | |
1810 | worthwhile or at least cache the results. The result | |
06acf7d0 RS |
1811 | technically depends on both SHIFT and ACCESS_SIZE, |
1812 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
c6f3019a RS |
1813 | |
1814 | if (cost > COSTS_N_INSNS (1)) | |
1815 | continue; | |
1816 | ||
d898d29b JJ |
1817 | new_lhs = extract_low_bits (new_mode, store_mode, |
1818 | copy_rtx (store_info->rhs)); | |
1819 | if (new_lhs == NULL_RTX) | |
1820 | continue; | |
1821 | ||
c6f3019a RS |
1822 | /* We found an acceptable shift. Generate a move to |
1823 | take the value from the store and put it into the | |
1824 | shift pseudo, then shift it, then generate another | |
1825 | move to put in into the target of the read. */ | |
18b526e8 | 1826 | emit_move_insn (new_reg, new_lhs); |
c6f3019a | 1827 | emit_insn (shift_seq); |
18b526e8 | 1828 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
c6f3019a | 1829 | break; |
8660aaae EC |
1830 | } |
1831 | ||
18b526e8 | 1832 | return read_reg; |
8660aaae EC |
1833 | } |
1834 | ||
1835 | ||
02b47899 KZ |
1836 | /* Call back for note_stores to find the hard regs set or clobbered by |
1837 | insn. Data is a bitmap of the hardregs set so far. */ | |
1838 | ||
1839 | static void | |
1840 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1841 | { | |
1842 | bitmap regs_set = (bitmap) data; | |
1843 | ||
1844 | if (REG_P (x) | |
f773c2bd | 1845 | && HARD_REGISTER_P (x)) |
02b47899 | 1846 | { |
f773c2bd AS |
1847 | unsigned int regno = REGNO (x); |
1848 | bitmap_set_range (regs_set, regno, | |
1849 | hard_regno_nregs[regno][GET_MODE (x)]); | |
02b47899 KZ |
1850 | } |
1851 | } | |
1852 | ||
8dd5516b JJ |
1853 | /* Helper function for replace_read and record_store. |
1854 | Attempt to return a value stored in STORE_INFO, from READ_BEGIN | |
1855 | to one before READ_END bytes read in READ_MODE. Return NULL | |
1856 | if not successful. If REQUIRE_CST is true, return always constant. */ | |
1857 | ||
1858 | static rtx | |
1859 | get_stored_val (store_info_t store_info, enum machine_mode read_mode, | |
1860 | HOST_WIDE_INT read_begin, HOST_WIDE_INT read_end, | |
1861 | basic_block bb, bool require_cst) | |
1862 | { | |
1863 | enum machine_mode store_mode = GET_MODE (store_info->mem); | |
1864 | int shift; | |
1865 | int access_size; /* In bytes. */ | |
1866 | rtx read_reg; | |
1867 | ||
1868 | /* To get here the read is within the boundaries of the write so | |
1869 | shift will never be negative. Start out with the shift being in | |
1870 | bytes. */ | |
1871 | if (store_mode == BLKmode) | |
1872 | shift = 0; | |
1873 | else if (BYTES_BIG_ENDIAN) | |
1874 | shift = store_info->end - read_end; | |
1875 | else | |
1876 | shift = read_begin - store_info->begin; | |
1877 | ||
1878 | access_size = shift + GET_MODE_SIZE (read_mode); | |
1879 | ||
1880 | /* From now on it is bits. */ | |
1881 | shift *= BITS_PER_UNIT; | |
1882 | ||
1883 | if (shift) | |
1884 | read_reg = find_shift_sequence (access_size, store_info, read_mode, shift, | |
1885 | optimize_bb_for_speed_p (bb), | |
1886 | require_cst); | |
1887 | else if (store_mode == BLKmode) | |
1888 | { | |
1889 | /* The store is a memset (addr, const_val, const_size). */ | |
1890 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
1891 | store_mode = int_mode_for_mode (read_mode); | |
1892 | if (store_mode == BLKmode) | |
1893 | read_reg = NULL_RTX; | |
1894 | else if (store_info->rhs == const0_rtx) | |
1895 | read_reg = extract_low_bits (read_mode, store_mode, const0_rtx); | |
1896 | else if (GET_MODE_BITSIZE (store_mode) > HOST_BITS_PER_WIDE_INT | |
1897 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) | |
1898 | read_reg = NULL_RTX; | |
1899 | else | |
1900 | { | |
1901 | unsigned HOST_WIDE_INT c | |
1902 | = INTVAL (store_info->rhs) | |
1903 | & (((HOST_WIDE_INT) 1 << BITS_PER_UNIT) - 1); | |
1904 | int shift = BITS_PER_UNIT; | |
1905 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1906 | { | |
1907 | c |= (c << shift); | |
1908 | shift <<= 1; | |
1909 | } | |
6d26322f | 1910 | read_reg = gen_int_mode (c, store_mode); |
8dd5516b JJ |
1911 | read_reg = extract_low_bits (read_mode, store_mode, read_reg); |
1912 | } | |
1913 | } | |
1914 | else if (store_info->const_rhs | |
1915 | && (require_cst | |
1916 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1917 | read_reg = extract_low_bits (read_mode, store_mode, | |
1918 | copy_rtx (store_info->const_rhs)); | |
1919 | else | |
1920 | read_reg = extract_low_bits (read_mode, store_mode, | |
1921 | copy_rtx (store_info->rhs)); | |
1922 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1923 | read_reg = NULL_RTX; | |
1924 | return read_reg; | |
1925 | } | |
02b47899 | 1926 | |
6fb5fa3c DB |
1927 | /* Take a sequence of: |
1928 | A <- r1 | |
1929 | ... | |
1930 | ... <- A | |
1931 | ||
b8698a0f | 1932 | and change it into |
6fb5fa3c DB |
1933 | r2 <- r1 |
1934 | A <- r1 | |
1935 | ... | |
1936 | ... <- r2 | |
1937 | ||
8660aaae EC |
1938 | or |
1939 | ||
1940 | r3 <- extract (r1) | |
1941 | r3 <- r3 >> shift | |
1942 | r2 <- extract (r3) | |
1943 | ... <- r2 | |
1944 | ||
1945 | or | |
1946 | ||
1947 | r2 <- extract (r1) | |
1948 | ... <- r2 | |
1949 | ||
1950 | Depending on the alignment and the mode of the store and | |
1951 | subsequent load. | |
1952 | ||
1953 | ||
1954 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
6fb5fa3c DB |
1955 | and READ_INSN are for the read. Return true if the replacement |
1956 | went ok. */ | |
1957 | ||
1958 | static bool | |
b8698a0f | 1959 | replace_read (store_info_t store_info, insn_info_t store_insn, |
8dd5516b JJ |
1960 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
1961 | bitmap regs_live) | |
6fb5fa3c | 1962 | { |
8660aaae EC |
1963 | enum machine_mode store_mode = GET_MODE (store_info->mem); |
1964 | enum machine_mode read_mode = GET_MODE (read_info->mem); | |
02b47899 | 1965 | rtx insns, this_insn, read_reg; |
8dd5516b | 1966 | basic_block bb; |
8660aaae | 1967 | |
6fb5fa3c DB |
1968 | if (!dbg_cnt (dse)) |
1969 | return false; | |
1970 | ||
18b526e8 RS |
1971 | /* Create a sequence of instructions to set up the read register. |
1972 | This sequence goes immediately before the store and its result | |
1973 | is read by the load. | |
1974 | ||
1975 | We need to keep this in perspective. We are replacing a read | |
8660aaae EC |
1976 | with a sequence of insns, but the read will almost certainly be |
1977 | in cache, so it is not going to be an expensive one. Thus, we | |
1978 | are not willing to do a multi insn shift or worse a subroutine | |
1979 | call to get rid of the read. */ | |
456610d3 | 1980 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
1981 | fprintf (dump_file, "trying to replace %smode load in insn %d" |
1982 | " from %smode store in insn %d\n", | |
1983 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
1984 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
1985 | start_sequence (); | |
8dd5516b JJ |
1986 | bb = BLOCK_FOR_INSN (read_insn->insn); |
1987 | read_reg = get_stored_val (store_info, | |
1988 | read_mode, read_info->begin, read_info->end, | |
1989 | bb, false); | |
18b526e8 | 1990 | if (read_reg == NULL_RTX) |
8660aaae | 1991 | { |
18b526e8 | 1992 | end_sequence (); |
456610d3 | 1993 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
1994 | fprintf (dump_file, " -- could not extract bits of stored value\n"); |
1995 | return false; | |
8660aaae | 1996 | } |
18b526e8 RS |
1997 | /* Force the value into a new register so that it won't be clobbered |
1998 | between the store and the load. */ | |
1999 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
2000 | insns = get_insns (); | |
2001 | end_sequence (); | |
8660aaae | 2002 | |
02b47899 KZ |
2003 | if (insns != NULL_RTX) |
2004 | { | |
2005 | /* Now we have to scan the set of new instructions to see if the | |
2006 | sequence contains and sets of hardregs that happened to be | |
2007 | live at this point. For instance, this can happen if one of | |
2008 | the insns sets the CC and the CC happened to be live at that | |
2009 | point. This does occasionally happen, see PR 37922. */ | |
3f9b14ff | 2010 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
02b47899 KZ |
2011 | |
2012 | for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
2013 | note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
b8698a0f | 2014 | |
02b47899 KZ |
2015 | bitmap_and_into (regs_set, regs_live); |
2016 | if (!bitmap_empty_p (regs_set)) | |
2017 | { | |
456610d3 | 2018 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02b47899 | 2019 | { |
b8698a0f | 2020 | fprintf (dump_file, |
02b47899 KZ |
2021 | "abandoning replacement because sequence clobbers live hardregs:"); |
2022 | df_print_regset (dump_file, regs_set); | |
2023 | } | |
b8698a0f | 2024 | |
02b47899 KZ |
2025 | BITMAP_FREE (regs_set); |
2026 | return false; | |
2027 | } | |
2028 | BITMAP_FREE (regs_set); | |
2029 | } | |
2030 | ||
8660aaae | 2031 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
6fb5fa3c | 2032 | { |
f883e0a7 KG |
2033 | deferred_change_t deferred_change = |
2034 | (deferred_change_t) pool_alloc (deferred_change_pool); | |
b8698a0f | 2035 | |
8660aaae EC |
2036 | /* Insert this right before the store insn where it will be safe |
2037 | from later insns that might change it before the read. */ | |
2038 | emit_insn_before (insns, store_insn->insn); | |
b8698a0f | 2039 | |
8660aaae EC |
2040 | /* And now for the kludge part: cselib croaks if you just |
2041 | return at this point. There are two reasons for this: | |
b8698a0f | 2042 | |
8660aaae EC |
2043 | 1) Cselib has an idea of how many pseudos there are and |
2044 | that does not include the new ones we just added. | |
b8698a0f | 2045 | |
8660aaae EC |
2046 | 2) Cselib does not know about the move insn we added |
2047 | above the store_info, and there is no way to tell it | |
2048 | about it, because it has "moved on". | |
b8698a0f | 2049 | |
8660aaae EC |
2050 | Problem (1) is fixable with a certain amount of engineering. |
2051 | Problem (2) is requires starting the bb from scratch. This | |
2052 | could be expensive. | |
b8698a0f | 2053 | |
8660aaae EC |
2054 | So we are just going to have to lie. The move/extraction |
2055 | insns are not really an issue, cselib did not see them. But | |
2056 | the use of the new pseudo read_insn is a real problem because | |
2057 | cselib has not scanned this insn. The way that we solve this | |
2058 | problem is that we are just going to put the mem back for now | |
2059 | and when we are finished with the block, we undo this. We | |
2060 | keep a table of mems to get rid of. At the end of the basic | |
2061 | block we can put them back. */ | |
b8698a0f | 2062 | |
8660aaae EC |
2063 | *loc = read_info->mem; |
2064 | deferred_change->next = deferred_change_list; | |
2065 | deferred_change_list = deferred_change; | |
2066 | deferred_change->loc = loc; | |
2067 | deferred_change->reg = read_reg; | |
b8698a0f | 2068 | |
8660aaae EC |
2069 | /* Get rid of the read_info, from the point of view of the |
2070 | rest of dse, play like this read never happened. */ | |
2071 | read_insn->read_rec = read_info->next; | |
2072 | pool_free (read_info_pool, read_info); | |
456610d3 | 2073 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2074 | { |
2075 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2076 | print_simple_rtl (dump_file, read_reg); | |
2077 | fprintf (dump_file, "\n"); | |
2078 | } | |
8660aaae | 2079 | return true; |
6fb5fa3c | 2080 | } |
b8698a0f | 2081 | else |
6fb5fa3c | 2082 | { |
456610d3 | 2083 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2084 | { |
2085 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2086 | print_simple_rtl (dump_file, read_reg); | |
2087 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2088 | } | |
6fb5fa3c DB |
2089 | return false; |
2090 | } | |
2091 | } | |
2092 | ||
d7111da8 RS |
2093 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2094 | be active. */ | |
6fb5fa3c | 2095 | |
d7111da8 RS |
2096 | static void |
2097 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info) | |
6fb5fa3c | 2098 | { |
6216f94e | 2099 | rtx mem = *loc, mem_addr; |
6fb5fa3c DB |
2100 | insn_info_t insn_info; |
2101 | HOST_WIDE_INT offset = 0; | |
2102 | HOST_WIDE_INT width = 0; | |
4862826d | 2103 | alias_set_type spill_alias_set = 0; |
b8698a0f | 2104 | cselib_val *base = NULL; |
6fb5fa3c DB |
2105 | int group_id; |
2106 | read_info_t read_info; | |
2107 | ||
6fb5fa3c DB |
2108 | insn_info = bb_info->last_insn; |
2109 | ||
2110 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2111 | || (MEM_VOLATILE_P (mem))) | |
2112 | { | |
456610d3 | 2113 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2114 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
2115 | add_wild_read (bb_info); | |
2116 | insn_info->cannot_delete = true; | |
d7111da8 | 2117 | return; |
6fb5fa3c DB |
2118 | } |
2119 | ||
2120 | /* If it is reading readonly mem, then there can be no conflict with | |
2121 | another write. */ | |
2122 | if (MEM_READONLY_P (mem)) | |
d7111da8 | 2123 | return; |
6fb5fa3c DB |
2124 | |
2125 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
2126 | { | |
456610d3 | 2127 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2128 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
2129 | add_wild_read (bb_info); | |
d7111da8 | 2130 | return; |
6fb5fa3c DB |
2131 | } |
2132 | ||
2133 | if (GET_MODE (mem) == BLKmode) | |
2134 | width = -1; | |
2135 | else | |
2136 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2137 | ||
f883e0a7 | 2138 | read_info = (read_info_t) pool_alloc (read_info_pool); |
6fb5fa3c DB |
2139 | read_info->group_id = group_id; |
2140 | read_info->mem = mem; | |
2141 | read_info->alias_set = spill_alias_set; | |
2142 | read_info->begin = offset; | |
2143 | read_info->end = offset + width; | |
2144 | read_info->next = insn_info->read_rec; | |
2145 | insn_info->read_rec = read_info; | |
6216f94e JJ |
2146 | /* For alias_set != 0 canon_true_dependence should be never called. */ |
2147 | if (spill_alias_set) | |
2148 | mem_addr = NULL_RTX; | |
2149 | else | |
2150 | { | |
2151 | if (group_id < 0) | |
2152 | mem_addr = base->val_rtx; | |
2153 | else | |
2154 | { | |
2155 | group_info_t group | |
9771b263 | 2156 | = rtx_group_vec[group_id]; |
6216f94e JJ |
2157 | mem_addr = group->canon_base_addr; |
2158 | } | |
2159 | if (offset) | |
372d6395 | 2160 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6216f94e | 2161 | } |
6fb5fa3c | 2162 | |
0d52bcc1 | 2163 | /* We ignore the clobbers in store_info. The is mildly aggressive, |
6fb5fa3c DB |
2164 | but there really should not be a clobber followed by a read. */ |
2165 | ||
2166 | if (spill_alias_set) | |
2167 | { | |
2168 | insn_info_t i_ptr = active_local_stores; | |
2169 | insn_info_t last = NULL; | |
2170 | ||
456610d3 | 2171 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2172 | fprintf (dump_file, " processing spill load %d\n", |
4862826d | 2173 | (int) spill_alias_set); |
6fb5fa3c DB |
2174 | |
2175 | while (i_ptr) | |
2176 | { | |
2177 | store_info_t store_info = i_ptr->store_rec; | |
2178 | ||
2179 | /* Skip the clobbers. */ | |
2180 | while (!store_info->is_set) | |
2181 | store_info = store_info->next; | |
b8698a0f | 2182 | |
6fb5fa3c DB |
2183 | if (store_info->alias_set == spill_alias_set) |
2184 | { | |
456610d3 | 2185 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2186 | dump_insn_info ("removing from active", i_ptr); |
2187 | ||
dabd47e7 | 2188 | active_local_stores_len--; |
6fb5fa3c DB |
2189 | if (last) |
2190 | last->next_local_store = i_ptr->next_local_store; | |
2191 | else | |
2192 | active_local_stores = i_ptr->next_local_store; | |
2193 | } | |
2194 | else | |
2195 | last = i_ptr; | |
2196 | i_ptr = i_ptr->next_local_store; | |
2197 | } | |
2198 | } | |
2199 | else if (group_id >= 0) | |
2200 | { | |
2201 | /* This is the restricted case where the base is a constant or | |
2202 | the frame pointer and offset is a constant. */ | |
2203 | insn_info_t i_ptr = active_local_stores; | |
2204 | insn_info_t last = NULL; | |
b8698a0f | 2205 | |
456610d3 | 2206 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2207 | { |
2208 | if (width == -1) | |
2209 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", | |
2210 | group_id); | |
2211 | else | |
2212 | fprintf (dump_file, " processing const load gid=%d[%d..%d)\n", | |
2213 | group_id, (int)offset, (int)(offset+width)); | |
2214 | } | |
2215 | ||
2216 | while (i_ptr) | |
2217 | { | |
2218 | bool remove = false; | |
2219 | store_info_t store_info = i_ptr->store_rec; | |
b8698a0f | 2220 | |
6fb5fa3c DB |
2221 | /* Skip the clobbers. */ |
2222 | while (!store_info->is_set) | |
2223 | store_info = store_info->next; | |
b8698a0f | 2224 | |
6fb5fa3c DB |
2225 | /* There are three cases here. */ |
2226 | if (store_info->group_id < 0) | |
2227 | /* We have a cselib store followed by a read from a | |
2228 | const base. */ | |
b8698a0f L |
2229 | remove |
2230 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2231 | GET_MODE (store_info->mem), |
2232 | store_info->mem_addr, | |
53d9622b | 2233 | mem, mem_addr); |
b8698a0f | 2234 | |
6fb5fa3c DB |
2235 | else if (group_id == store_info->group_id) |
2236 | { | |
2237 | /* This is a block mode load. We may get lucky and | |
2238 | canon_true_dependence may save the day. */ | |
2239 | if (width == -1) | |
b8698a0f L |
2240 | remove |
2241 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2242 | GET_MODE (store_info->mem), |
2243 | store_info->mem_addr, | |
53d9622b | 2244 | mem, mem_addr); |
b8698a0f | 2245 | |
6fb5fa3c DB |
2246 | /* If this read is just reading back something that we just |
2247 | stored, rewrite the read. */ | |
b8698a0f | 2248 | else |
6fb5fa3c DB |
2249 | { |
2250 | if (store_info->rhs | |
8dd5516b JJ |
2251 | && offset >= store_info->begin |
2252 | && offset + width <= store_info->end | |
2253 | && all_positions_needed_p (store_info, | |
2254 | offset - store_info->begin, | |
2255 | width) | |
2256 | && replace_read (store_info, i_ptr, read_info, | |
2257 | insn_info, loc, bb_info->regs_live)) | |
d7111da8 | 2258 | return; |
8dd5516b | 2259 | |
6fb5fa3c DB |
2260 | /* The bases are the same, just see if the offsets |
2261 | overlap. */ | |
b8698a0f | 2262 | if ((offset < store_info->end) |
6fb5fa3c DB |
2263 | && (offset + width > store_info->begin)) |
2264 | remove = true; | |
2265 | } | |
2266 | } | |
b8698a0f L |
2267 | |
2268 | /* else | |
6fb5fa3c DB |
2269 | The else case that is missing here is that the |
2270 | bases are constant but different. There is nothing | |
2271 | to do here because there is no overlap. */ | |
b8698a0f | 2272 | |
6fb5fa3c DB |
2273 | if (remove) |
2274 | { | |
456610d3 | 2275 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2276 | dump_insn_info ("removing from active", i_ptr); |
2277 | ||
dabd47e7 | 2278 | active_local_stores_len--; |
6fb5fa3c DB |
2279 | if (last) |
2280 | last->next_local_store = i_ptr->next_local_store; | |
2281 | else | |
2282 | active_local_stores = i_ptr->next_local_store; | |
2283 | } | |
2284 | else | |
2285 | last = i_ptr; | |
2286 | i_ptr = i_ptr->next_local_store; | |
2287 | } | |
2288 | } | |
b8698a0f | 2289 | else |
6fb5fa3c DB |
2290 | { |
2291 | insn_info_t i_ptr = active_local_stores; | |
2292 | insn_info_t last = NULL; | |
456610d3 | 2293 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2294 | { |
2295 | fprintf (dump_file, " processing cselib load mem:"); | |
2296 | print_inline_rtx (dump_file, mem, 0); | |
2297 | fprintf (dump_file, "\n"); | |
2298 | } | |
2299 | ||
2300 | while (i_ptr) | |
2301 | { | |
2302 | bool remove = false; | |
2303 | store_info_t store_info = i_ptr->store_rec; | |
b8698a0f | 2304 | |
456610d3 | 2305 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2306 | fprintf (dump_file, " processing cselib load against insn %d\n", |
2307 | INSN_UID (i_ptr->insn)); | |
2308 | ||
2309 | /* Skip the clobbers. */ | |
2310 | while (!store_info->is_set) | |
2311 | store_info = store_info->next; | |
2312 | ||
2313 | /* If this read is just reading back something that we just | |
2314 | stored, rewrite the read. */ | |
2315 | if (store_info->rhs | |
2316 | && store_info->group_id == -1 | |
2317 | && store_info->cse_base == base | |
efc3527a | 2318 | && width != -1 |
8dd5516b JJ |
2319 | && offset >= store_info->begin |
2320 | && offset + width <= store_info->end | |
2321 | && all_positions_needed_p (store_info, | |
2322 | offset - store_info->begin, width) | |
2323 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, | |
2324 | bb_info->regs_live)) | |
d7111da8 | 2325 | return; |
6fb5fa3c DB |
2326 | |
2327 | if (!store_info->alias_set) | |
b8698a0f | 2328 | remove = canon_true_dependence (store_info->mem, |
6fb5fa3c DB |
2329 | GET_MODE (store_info->mem), |
2330 | store_info->mem_addr, | |
53d9622b | 2331 | mem, mem_addr); |
b8698a0f | 2332 | |
6fb5fa3c DB |
2333 | if (remove) |
2334 | { | |
456610d3 | 2335 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2336 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2337 | |
dabd47e7 | 2338 | active_local_stores_len--; |
6fb5fa3c DB |
2339 | if (last) |
2340 | last->next_local_store = i_ptr->next_local_store; | |
2341 | else | |
2342 | active_local_stores = i_ptr->next_local_store; | |
2343 | } | |
2344 | else | |
2345 | last = i_ptr; | |
2346 | i_ptr = i_ptr->next_local_store; | |
2347 | } | |
2348 | } | |
6fb5fa3c DB |
2349 | } |
2350 | ||
d7111da8 | 2351 | /* A note_uses callback in which DATA points the INSN_INFO for |
6fb5fa3c DB |
2352 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2353 | true for any part of *LOC. */ | |
2354 | ||
2355 | static void | |
2356 | check_mem_read_use (rtx *loc, void *data) | |
2357 | { | |
d7111da8 RS |
2358 | subrtx_ptr_iterator::array_type array; |
2359 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
2360 | { | |
2361 | rtx *loc = *iter; | |
2362 | if (MEM_P (*loc)) | |
2363 | check_mem_read_rtx (loc, (bb_info_t) data); | |
2364 | } | |
6fb5fa3c DB |
2365 | } |
2366 | ||
8dd5516b JJ |
2367 | |
2368 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2369 | So far it only handles arguments passed in registers. */ | |
2370 | ||
2371 | static bool | |
2372 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2373 | { | |
d5cc9181 JR |
2374 | CUMULATIVE_ARGS args_so_far_v; |
2375 | cumulative_args_t args_so_far; | |
8dd5516b JJ |
2376 | tree arg; |
2377 | int idx; | |
2378 | ||
d5cc9181 JR |
2379 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2380 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
8dd5516b JJ |
2381 | |
2382 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2383 | for (idx = 0; | |
2384 | arg != void_list_node && idx < nargs; | |
2385 | arg = TREE_CHAIN (arg), idx++) | |
2386 | { | |
2387 | enum machine_mode mode = TYPE_MODE (TREE_VALUE (arg)); | |
3c07301f | 2388 | rtx reg, link, tmp; |
d5cc9181 | 2389 | reg = targetm.calls.function_arg (args_so_far, mode, NULL_TREE, true); |
8dd5516b JJ |
2390 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode |
2391 | || GET_MODE_CLASS (mode) != MODE_INT) | |
2392 | return false; | |
2393 | ||
2394 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2395 | link; | |
2396 | link = XEXP (link, 1)) | |
2397 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2398 | { | |
2399 | args[idx] = XEXP (XEXP (link, 0), 0); | |
2400 | if (REG_P (args[idx]) | |
2401 | && REGNO (args[idx]) == REGNO (reg) | |
2402 | && (GET_MODE (args[idx]) == mode | |
2403 | || (GET_MODE_CLASS (GET_MODE (args[idx])) == MODE_INT | |
2404 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2405 | <= UNITS_PER_WORD) | |
2406 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2407 | > GET_MODE_SIZE (mode))))) | |
2408 | break; | |
2409 | } | |
2410 | if (!link) | |
2411 | return false; | |
2412 | ||
2413 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2414 | if (GET_MODE (args[idx]) != mode) | |
2415 | { | |
2416 | if (!tmp || !CONST_INT_P (tmp)) | |
2417 | return false; | |
6d26322f | 2418 | tmp = gen_int_mode (INTVAL (tmp), mode); |
8dd5516b JJ |
2419 | } |
2420 | if (tmp) | |
2421 | args[idx] = tmp; | |
2422 | ||
d5cc9181 | 2423 | targetm.calls.function_arg_advance (args_so_far, mode, NULL_TREE, true); |
8dd5516b JJ |
2424 | } |
2425 | if (arg != void_list_node || idx != nargs) | |
2426 | return false; | |
2427 | return true; | |
2428 | } | |
2429 | ||
9e582b1d JR |
2430 | /* Return a bitmap of the fixed registers contained in IN. */ |
2431 | ||
2432 | static bitmap | |
2433 | copy_fixed_regs (const_bitmap in) | |
2434 | { | |
2435 | bitmap ret; | |
2436 | ||
2437 | ret = ALLOC_REG_SET (NULL); | |
2438 | bitmap_and (ret, in, fixed_reg_set_regset); | |
2439 | return ret; | |
2440 | } | |
8dd5516b | 2441 | |
6fb5fa3c DB |
2442 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2443 | if some part of it is not a candidate store and assigns to a | |
2444 | non-register target. */ | |
2445 | ||
2446 | static void | |
dd60a84c | 2447 | scan_insn (bb_info_t bb_info, rtx_insn *insn) |
6fb5fa3c DB |
2448 | { |
2449 | rtx body; | |
f883e0a7 | 2450 | insn_info_t insn_info = (insn_info_t) pool_alloc (insn_info_pool); |
6fb5fa3c DB |
2451 | int mems_found = 0; |
2452 | memset (insn_info, 0, sizeof (struct insn_info)); | |
2453 | ||
456610d3 | 2454 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2455 | fprintf (dump_file, "\n**scanning insn=%d\n", |
2456 | INSN_UID (insn)); | |
2457 | ||
2458 | insn_info->prev_insn = bb_info->last_insn; | |
2459 | insn_info->insn = insn; | |
2460 | bb_info->last_insn = insn_info; | |
b8698a0f | 2461 | |
b5b8b0ac AO |
2462 | if (DEBUG_INSN_P (insn)) |
2463 | { | |
2464 | insn_info->cannot_delete = true; | |
2465 | return; | |
2466 | } | |
6fb5fa3c | 2467 | |
6fb5fa3c DB |
2468 | /* Look at all of the uses in the insn. */ |
2469 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2470 | ||
2471 | if (CALL_P (insn)) | |
2472 | { | |
8dd5516b JJ |
2473 | bool const_call; |
2474 | tree memset_call = NULL_TREE; | |
2475 | ||
6fb5fa3c | 2476 | insn_info->cannot_delete = true; |
50f0f366 | 2477 | |
6fb5fa3c | 2478 | /* Const functions cannot do anything bad i.e. read memory, |
50f0f366 | 2479 | however, they can read their parameters which may have |
8dd5516b JJ |
2480 | been pushed onto the stack. |
2481 | memset and bzero don't read memory either. */ | |
2482 | const_call = RTL_CONST_CALL_P (insn); | |
2483 | if (!const_call) | |
2484 | { | |
da4fdf2d SB |
2485 | rtx call = get_call_rtx_from (insn); |
2486 | if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF) | |
8dd5516b JJ |
2487 | { |
2488 | rtx symbol = XEXP (XEXP (call, 0), 0); | |
2489 | if (SYMBOL_REF_DECL (symbol) | |
2490 | && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL) | |
2491 | { | |
2492 | if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol)) | |
2493 | == BUILT_IN_NORMAL | |
2494 | && (DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol)) | |
2495 | == BUILT_IN_MEMSET)) | |
2496 | || SYMBOL_REF_DECL (symbol) == block_clear_fn) | |
2497 | memset_call = SYMBOL_REF_DECL (symbol); | |
2498 | } | |
2499 | } | |
2500 | } | |
2501 | if (const_call || memset_call) | |
6fb5fa3c DB |
2502 | { |
2503 | insn_info_t i_ptr = active_local_stores; | |
2504 | insn_info_t last = NULL; | |
2505 | ||
456610d3 | 2506 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2507 | fprintf (dump_file, "%s call %d\n", |
2508 | const_call ? "const" : "memset", INSN_UID (insn)); | |
6fb5fa3c | 2509 | |
64520bdc EB |
2510 | /* See the head comment of the frame_read field. */ |
2511 | if (reload_completed) | |
2512 | insn_info->frame_read = true; | |
2513 | ||
2514 | /* Loop over the active stores and remove those which are | |
2515 | killed by the const function call. */ | |
6fb5fa3c DB |
2516 | while (i_ptr) |
2517 | { | |
64520bdc EB |
2518 | bool remove_store = false; |
2519 | ||
2520 | /* The stack pointer based stores are always killed. */ | |
50f0f366 | 2521 | if (i_ptr->stack_pointer_based) |
64520bdc EB |
2522 | remove_store = true; |
2523 | ||
2524 | /* If the frame is read, the frame related stores are killed. */ | |
2525 | else if (insn_info->frame_read) | |
2526 | { | |
2527 | store_info_t store_info = i_ptr->store_rec; | |
2528 | ||
2529 | /* Skip the clobbers. */ | |
2530 | while (!store_info->is_set) | |
2531 | store_info = store_info->next; | |
2532 | ||
2533 | if (store_info->group_id >= 0 | |
9771b263 | 2534 | && rtx_group_vec[store_info->group_id]->frame_related) |
64520bdc EB |
2535 | remove_store = true; |
2536 | } | |
2537 | ||
2538 | if (remove_store) | |
6fb5fa3c | 2539 | { |
456610d3 | 2540 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2541 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2542 | |
dabd47e7 | 2543 | active_local_stores_len--; |
6fb5fa3c DB |
2544 | if (last) |
2545 | last->next_local_store = i_ptr->next_local_store; | |
2546 | else | |
2547 | active_local_stores = i_ptr->next_local_store; | |
2548 | } | |
2549 | else | |
2550 | last = i_ptr; | |
64520bdc | 2551 | |
6fb5fa3c DB |
2552 | i_ptr = i_ptr->next_local_store; |
2553 | } | |
8dd5516b JJ |
2554 | |
2555 | if (memset_call) | |
2556 | { | |
2557 | rtx args[3]; | |
2558 | if (get_call_args (insn, memset_call, args, 3) | |
2559 | && CONST_INT_P (args[1]) | |
2560 | && CONST_INT_P (args[2]) | |
2561 | && INTVAL (args[2]) > 0) | |
2562 | { | |
2563 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
f5541398 | 2564 | set_mem_size (mem, INTVAL (args[2])); |
8dd5516b JJ |
2565 | body = gen_rtx_SET (VOIDmode, mem, args[1]); |
2566 | mems_found += record_store (body, bb_info); | |
456610d3 | 2567 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2568 | fprintf (dump_file, "handling memset as BLKmode store\n"); |
2569 | if (mems_found == 1) | |
2570 | { | |
dabd47e7 JJ |
2571 | if (active_local_stores_len++ |
2572 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2573 | { | |
2574 | active_local_stores_len = 1; | |
2575 | active_local_stores = NULL; | |
2576 | } | |
9e582b1d JR |
2577 | insn_info->fixed_regs_live |
2578 | = copy_fixed_regs (bb_info->regs_live); | |
8dd5516b JJ |
2579 | insn_info->next_local_store = active_local_stores; |
2580 | active_local_stores = insn_info; | |
2581 | } | |
2582 | } | |
2583 | } | |
6fb5fa3c DB |
2584 | } |
2585 | ||
50f0f366 | 2586 | else |
d26c7090 ER |
2587 | /* Every other call, including pure functions, may read any memory |
2588 | that is not relative to the frame. */ | |
2589 | add_non_frame_wild_read (bb_info); | |
50f0f366 | 2590 | |
6fb5fa3c DB |
2591 | return; |
2592 | } | |
2593 | ||
2594 | /* Assuming that there are sets in these insns, we cannot delete | |
2595 | them. */ | |
2596 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
0a64eeca | 2597 | || volatile_refs_p (PATTERN (insn)) |
2da02156 | 2598 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
6fb5fa3c DB |
2599 | || (RTX_FRAME_RELATED_P (insn)) |
2600 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2601 | insn_info->cannot_delete = true; | |
b8698a0f | 2602 | |
6fb5fa3c DB |
2603 | body = PATTERN (insn); |
2604 | if (GET_CODE (body) == PARALLEL) | |
2605 | { | |
2606 | int i; | |
2607 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2608 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2609 | } | |
2610 | else | |
2611 | mems_found += record_store (body, bb_info); | |
2612 | ||
456610d3 | 2613 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2614 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
6fb5fa3c DB |
2615 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2616 | ||
8dd5516b JJ |
2617 | /* If we found some sets of mems, add it into the active_local_stores so |
2618 | that it can be locally deleted if found dead or used for | |
2619 | replace_read and redundant constant store elimination. Otherwise mark | |
2620 | it as cannot delete. This simplifies the processing later. */ | |
2621 | if (mems_found == 1) | |
6fb5fa3c | 2622 | { |
dabd47e7 JJ |
2623 | if (active_local_stores_len++ |
2624 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2625 | { | |
2626 | active_local_stores_len = 1; | |
2627 | active_local_stores = NULL; | |
2628 | } | |
9e582b1d | 2629 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
6fb5fa3c DB |
2630 | insn_info->next_local_store = active_local_stores; |
2631 | active_local_stores = insn_info; | |
2632 | } | |
2633 | else | |
2634 | insn_info->cannot_delete = true; | |
2635 | } | |
2636 | ||
2637 | ||
2638 | /* Remove BASE from the set of active_local_stores. This is a | |
2639 | callback from cselib that is used to get rid of the stores in | |
2640 | active_local_stores. */ | |
2641 | ||
2642 | static void | |
2643 | remove_useless_values (cselib_val *base) | |
2644 | { | |
2645 | insn_info_t insn_info = active_local_stores; | |
2646 | insn_info_t last = NULL; | |
2647 | ||
2648 | while (insn_info) | |
2649 | { | |
2650 | store_info_t store_info = insn_info->store_rec; | |
60564289 | 2651 | bool del = false; |
6fb5fa3c DB |
2652 | |
2653 | /* If ANY of the store_infos match the cselib group that is | |
2654 | being deleted, then the insn can not be deleted. */ | |
2655 | while (store_info) | |
2656 | { | |
b8698a0f | 2657 | if ((store_info->group_id == -1) |
6fb5fa3c DB |
2658 | && (store_info->cse_base == base)) |
2659 | { | |
60564289 | 2660 | del = true; |
6fb5fa3c DB |
2661 | break; |
2662 | } | |
2663 | store_info = store_info->next; | |
2664 | } | |
2665 | ||
60564289 | 2666 | if (del) |
6fb5fa3c | 2667 | { |
dabd47e7 | 2668 | active_local_stores_len--; |
6fb5fa3c DB |
2669 | if (last) |
2670 | last->next_local_store = insn_info->next_local_store; | |
2671 | else | |
2672 | active_local_stores = insn_info->next_local_store; | |
2673 | free_store_info (insn_info); | |
2674 | } | |
2675 | else | |
2676 | last = insn_info; | |
b8698a0f | 2677 | |
6fb5fa3c DB |
2678 | insn_info = insn_info->next_local_store; |
2679 | } | |
2680 | } | |
2681 | ||
2682 | ||
2683 | /* Do all of step 1. */ | |
2684 | ||
2685 | static void | |
2686 | dse_step1 (void) | |
2687 | { | |
2688 | basic_block bb; | |
3f9b14ff | 2689 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
b8698a0f | 2690 | |
457eeaae | 2691 | cselib_init (0); |
6fb5fa3c DB |
2692 | all_blocks = BITMAP_ALLOC (NULL); |
2693 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2694 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2695 | ||
04a90bec | 2696 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
2697 | { |
2698 | insn_info_t ptr; | |
f883e0a7 | 2699 | bb_info_t bb_info = (bb_info_t) pool_alloc (bb_info_pool); |
6fb5fa3c DB |
2700 | |
2701 | memset (bb_info, 0, sizeof (struct bb_info)); | |
2702 | bitmap_set_bit (all_blocks, bb->index); | |
02b47899 KZ |
2703 | bb_info->regs_live = regs_live; |
2704 | ||
2705 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2706 | df_simulate_initialize_forwards (bb, regs_live); | |
6fb5fa3c DB |
2707 | |
2708 | bb_table[bb->index] = bb_info; | |
2709 | cselib_discard_hook = remove_useless_values; | |
2710 | ||
2711 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2712 | { | |
dd60a84c | 2713 | rtx_insn *insn; |
6fb5fa3c DB |
2714 | |
2715 | cse_store_info_pool | |
b8698a0f | 2716 | = create_alloc_pool ("cse_store_info_pool", |
6fb5fa3c DB |
2717 | sizeof (struct store_info), 100); |
2718 | active_local_stores = NULL; | |
dabd47e7 | 2719 | active_local_stores_len = 0; |
6fb5fa3c | 2720 | cselib_clear_table (); |
b8698a0f | 2721 | |
6fb5fa3c DB |
2722 | /* Scan the insns. */ |
2723 | FOR_BB_INSNS (bb, insn) | |
2724 | { | |
2725 | if (INSN_P (insn)) | |
2726 | scan_insn (bb_info, insn); | |
2727 | cselib_process_insn (insn); | |
02b47899 KZ |
2728 | if (INSN_P (insn)) |
2729 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
6fb5fa3c | 2730 | } |
b8698a0f | 2731 | |
6fb5fa3c DB |
2732 | /* This is something of a hack, because the global algorithm |
2733 | is supposed to take care of the case where stores go dead | |
2734 | at the end of the function. However, the global | |
2735 | algorithm must take a more conservative view of block | |
2736 | mode reads than the local alg does. So to get the case | |
2737 | where you have a store to the frame followed by a non | |
0d52bcc1 | 2738 | overlapping block more read, we look at the active local |
6fb5fa3c DB |
2739 | stores at the end of the function and delete all of the |
2740 | frame and spill based ones. */ | |
2741 | if (stores_off_frame_dead_at_return | |
2742 | && (EDGE_COUNT (bb->succs) == 0 | |
2743 | || (single_succ_p (bb) | |
fefa31b5 | 2744 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
e3b5732b | 2745 | && ! crtl->calls_eh_return))) |
6fb5fa3c DB |
2746 | { |
2747 | insn_info_t i_ptr = active_local_stores; | |
2748 | while (i_ptr) | |
2749 | { | |
2750 | store_info_t store_info = i_ptr->store_rec; | |
2751 | ||
2752 | /* Skip the clobbers. */ | |
2753 | while (!store_info->is_set) | |
2754 | store_info = store_info->next; | |
8dd5516b | 2755 | if (store_info->alias_set && !i_ptr->cannot_delete) |
6fb5fa3c | 2756 | delete_dead_store_insn (i_ptr); |
b8698a0f | 2757 | else |
6fb5fa3c DB |
2758 | if (store_info->group_id >= 0) |
2759 | { | |
b8698a0f | 2760 | group_info_t group |
9771b263 | 2761 | = rtx_group_vec[store_info->group_id]; |
8dd5516b | 2762 | if (group->frame_related && !i_ptr->cannot_delete) |
6fb5fa3c DB |
2763 | delete_dead_store_insn (i_ptr); |
2764 | } | |
2765 | ||
2766 | i_ptr = i_ptr->next_local_store; | |
2767 | } | |
2768 | } | |
2769 | ||
2770 | /* Get rid of the loads that were discovered in | |
2771 | replace_read. Cselib is finished with this block. */ | |
2772 | while (deferred_change_list) | |
2773 | { | |
2774 | deferred_change_t next = deferred_change_list->next; | |
2775 | ||
2776 | /* There is no reason to validate this change. That was | |
2777 | done earlier. */ | |
2778 | *deferred_change_list->loc = deferred_change_list->reg; | |
2779 | pool_free (deferred_change_pool, deferred_change_list); | |
2780 | deferred_change_list = next; | |
2781 | } | |
2782 | ||
2783 | /* Get rid of all of the cselib based store_infos in this | |
2784 | block and mark the containing insns as not being | |
2785 | deletable. */ | |
2786 | ptr = bb_info->last_insn; | |
2787 | while (ptr) | |
2788 | { | |
2789 | if (ptr->contains_cselib_groups) | |
8dd5516b JJ |
2790 | { |
2791 | store_info_t s_info = ptr->store_rec; | |
2792 | while (s_info && !s_info->is_set) | |
2793 | s_info = s_info->next; | |
2794 | if (s_info | |
2795 | && s_info->redundant_reason | |
2796 | && s_info->redundant_reason->insn | |
2797 | && !ptr->cannot_delete) | |
2798 | { | |
456610d3 | 2799 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2800 | fprintf (dump_file, "Locally deleting insn %d " |
2801 | "because insn %d stores the " | |
2802 | "same value and couldn't be " | |
2803 | "eliminated\n", | |
2804 | INSN_UID (ptr->insn), | |
2805 | INSN_UID (s_info->redundant_reason->insn)); | |
2806 | delete_dead_store_insn (ptr); | |
2807 | } | |
8dd5516b JJ |
2808 | free_store_info (ptr); |
2809 | } | |
2810 | else | |
2811 | { | |
2812 | store_info_t s_info; | |
2813 | ||
2814 | /* Free at least positions_needed bitmaps. */ | |
2815 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2816 | if (s_info->is_large) | |
2817 | { | |
dc491a25 | 2818 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
8dd5516b JJ |
2819 | s_info->is_large = false; |
2820 | } | |
2821 | } | |
6fb5fa3c DB |
2822 | ptr = ptr->prev_insn; |
2823 | } | |
2824 | ||
2825 | free_alloc_pool (cse_store_info_pool); | |
2826 | } | |
02b47899 | 2827 | bb_info->regs_live = NULL; |
6fb5fa3c DB |
2828 | } |
2829 | ||
02b47899 | 2830 | BITMAP_FREE (regs_live); |
6fb5fa3c | 2831 | cselib_finish (); |
c203e8a7 | 2832 | rtx_group_table->empty (); |
6fb5fa3c DB |
2833 | } |
2834 | ||
2835 | \f | |
2836 | /*---------------------------------------------------------------------------- | |
2837 | Second step. | |
2838 | ||
2839 | Assign each byte position in the stores that we are going to | |
2840 | analyze globally to a position in the bitmaps. Returns true if | |
6ed3da00 | 2841 | there are any bit positions assigned. |
6fb5fa3c DB |
2842 | ----------------------------------------------------------------------------*/ |
2843 | ||
2844 | static void | |
2845 | dse_step2_init (void) | |
2846 | { | |
2847 | unsigned int i; | |
2848 | group_info_t group; | |
2849 | ||
9771b263 | 2850 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2851 | { |
2852 | /* For all non stack related bases, we only consider a store to | |
2853 | be deletable if there are two or more stores for that | |
2854 | position. This is because it takes one store to make the | |
2855 | other store redundant. However, for the stores that are | |
2856 | stack related, we consider them if there is only one store | |
2857 | for the position. We do this because the stack related | |
2858 | stores can be deleted if their is no read between them and | |
2859 | the end of the function. | |
b8698a0f | 2860 | |
6fb5fa3c DB |
2861 | To make this work in the current framework, we take the stack |
2862 | related bases add all of the bits from store1 into store2. | |
2863 | This has the effect of making the eligible even if there is | |
2864 | only one store. */ | |
2865 | ||
2866 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2867 | { | |
2868 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2869 | bitmap_ior_into (group->store2_p, group->store1_p); | |
456610d3 | 2870 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2871 | fprintf (dump_file, "group %d is frame related ", i); |
6fb5fa3c DB |
2872 | } |
2873 | ||
2874 | group->offset_map_size_n++; | |
3f9b14ff SB |
2875 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2876 | group->offset_map_size_n); | |
6fb5fa3c | 2877 | group->offset_map_size_p++; |
3f9b14ff SB |
2878 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2879 | group->offset_map_size_p); | |
6fb5fa3c | 2880 | group->process_globally = false; |
456610d3 | 2881 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2882 | { |
b8698a0f | 2883 | fprintf (dump_file, "group %d(%d+%d): ", i, |
6fb5fa3c DB |
2884 | (int)bitmap_count_bits (group->store2_n), |
2885 | (int)bitmap_count_bits (group->store2_p)); | |
2886 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2887 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2888 | } | |
2889 | } | |
2890 | } | |
2891 | ||
2892 | ||
2893 | /* Init the offset tables for the normal case. */ | |
2894 | ||
2895 | static bool | |
2896 | dse_step2_nospill (void) | |
2897 | { | |
2898 | unsigned int i; | |
2899 | group_info_t group; | |
2900 | /* Position 0 is unused because 0 is used in the maps to mean | |
2901 | unused. */ | |
2902 | current_position = 1; | |
9771b263 | 2903 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2904 | { |
2905 | bitmap_iterator bi; | |
2906 | unsigned int j; | |
2907 | ||
2908 | if (group == clear_alias_group) | |
2909 | continue; | |
2910 | ||
c3284718 RS |
2911 | memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
2912 | memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
6fb5fa3c DB |
2913 | bitmap_clear (group->group_kill); |
2914 | ||
2915 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2916 | { | |
2917 | bitmap_set_bit (group->group_kill, current_position); | |
d26c7090 ER |
2918 | if (bitmap_bit_p (group->escaped_n, j)) |
2919 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2920 | group->offset_map_n[j] = current_position++; |
2921 | group->process_globally = true; | |
2922 | } | |
2923 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2924 | { | |
b8698a0f | 2925 | bitmap_set_bit (group->group_kill, current_position); |
d26c7090 ER |
2926 | if (bitmap_bit_p (group->escaped_p, j)) |
2927 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2928 | group->offset_map_p[j] = current_position++; |
2929 | group->process_globally = true; | |
2930 | } | |
2931 | } | |
2932 | return current_position != 1; | |
2933 | } | |
2934 | ||
2935 | ||
6fb5fa3c DB |
2936 | \f |
2937 | /*---------------------------------------------------------------------------- | |
2938 | Third step. | |
b8698a0f | 2939 | |
6fb5fa3c DB |
2940 | Build the bit vectors for the transfer functions. |
2941 | ----------------------------------------------------------------------------*/ | |
2942 | ||
2943 | ||
6fb5fa3c DB |
2944 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
2945 | there, return 0. */ | |
2946 | ||
2947 | static int | |
2948 | get_bitmap_index (group_info_t group_info, HOST_WIDE_INT offset) | |
2949 | { | |
2950 | if (offset < 0) | |
2951 | { | |
2952 | HOST_WIDE_INT offset_p = -offset; | |
2953 | if (offset_p >= group_info->offset_map_size_n) | |
2954 | return 0; | |
2955 | return group_info->offset_map_n[offset_p]; | |
2956 | } | |
2957 | else | |
2958 | { | |
2959 | if (offset >= group_info->offset_map_size_p) | |
2960 | return 0; | |
2961 | return group_info->offset_map_p[offset]; | |
2962 | } | |
2963 | } | |
2964 | ||
2965 | ||
2966 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
2967 | may be NULL. */ | |
2968 | ||
b8698a0f | 2969 | static void |
6fb5fa3c DB |
2970 | scan_stores_nospill (store_info_t store_info, bitmap gen, bitmap kill) |
2971 | { | |
2972 | while (store_info) | |
2973 | { | |
2974 | HOST_WIDE_INT i; | |
b8698a0f | 2975 | group_info_t group_info |
9771b263 | 2976 | = rtx_group_vec[store_info->group_id]; |
6fb5fa3c DB |
2977 | if (group_info->process_globally) |
2978 | for (i = store_info->begin; i < store_info->end; i++) | |
2979 | { | |
2980 | int index = get_bitmap_index (group_info, i); | |
2981 | if (index != 0) | |
2982 | { | |
2983 | bitmap_set_bit (gen, index); | |
2984 | if (kill) | |
2985 | bitmap_clear_bit (kill, index); | |
2986 | } | |
2987 | } | |
2988 | store_info = store_info->next; | |
2989 | } | |
2990 | } | |
2991 | ||
2992 | ||
2993 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
2994 | may be NULL. */ | |
2995 | ||
b8698a0f | 2996 | static void |
6fb5fa3c DB |
2997 | scan_stores_spill (store_info_t store_info, bitmap gen, bitmap kill) |
2998 | { | |
2999 | while (store_info) | |
3000 | { | |
3001 | if (store_info->alias_set) | |
3002 | { | |
b8698a0f | 3003 | int index = get_bitmap_index (clear_alias_group, |
6fb5fa3c DB |
3004 | store_info->alias_set); |
3005 | if (index != 0) | |
3006 | { | |
3007 | bitmap_set_bit (gen, index); | |
3008 | if (kill) | |
3009 | bitmap_clear_bit (kill, index); | |
3010 | } | |
3011 | } | |
3012 | store_info = store_info->next; | |
3013 | } | |
3014 | } | |
3015 | ||
3016 | ||
3017 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3018 | may be NULL. */ | |
3019 | ||
3020 | static void | |
3021 | scan_reads_nospill (insn_info_t insn_info, bitmap gen, bitmap kill) | |
3022 | { | |
3023 | read_info_t read_info = insn_info->read_rec; | |
3024 | int i; | |
3025 | group_info_t group; | |
3026 | ||
64520bdc EB |
3027 | /* If this insn reads the frame, kill all the frame related stores. */ |
3028 | if (insn_info->frame_read) | |
3029 | { | |
9771b263 | 3030 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
64520bdc EB |
3031 | if (group->process_globally && group->frame_related) |
3032 | { | |
3033 | if (kill) | |
3034 | bitmap_ior_into (kill, group->group_kill); | |
b8698a0f | 3035 | bitmap_and_compl_into (gen, group->group_kill); |
64520bdc EB |
3036 | } |
3037 | } | |
d26c7090 ER |
3038 | if (insn_info->non_frame_wild_read) |
3039 | { | |
3040 | /* Kill all non-frame related stores. Kill all stores of variables that | |
3041 | escape. */ | |
3042 | if (kill) | |
3043 | bitmap_ior_into (kill, kill_on_calls); | |
3044 | bitmap_and_compl_into (gen, kill_on_calls); | |
9771b263 | 3045 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
d26c7090 ER |
3046 | if (group->process_globally && !group->frame_related) |
3047 | { | |
3048 | if (kill) | |
3049 | bitmap_ior_into (kill, group->group_kill); | |
3050 | bitmap_and_compl_into (gen, group->group_kill); | |
3051 | } | |
3052 | } | |
6fb5fa3c DB |
3053 | while (read_info) |
3054 | { | |
9771b263 | 3055 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3056 | { |
3057 | if (group->process_globally) | |
3058 | { | |
3059 | if (i == read_info->group_id) | |
3060 | { | |
3061 | if (read_info->begin > read_info->end) | |
3062 | { | |
3063 | /* Begin > end for block mode reads. */ | |
3064 | if (kill) | |
3065 | bitmap_ior_into (kill, group->group_kill); | |
3066 | bitmap_and_compl_into (gen, group->group_kill); | |
3067 | } | |
3068 | else | |
3069 | { | |
3070 | /* The groups are the same, just process the | |
3071 | offsets. */ | |
3072 | HOST_WIDE_INT j; | |
3073 | for (j = read_info->begin; j < read_info->end; j++) | |
3074 | { | |
3075 | int index = get_bitmap_index (group, j); | |
3076 | if (index != 0) | |
3077 | { | |
3078 | if (kill) | |
3079 | bitmap_set_bit (kill, index); | |
3080 | bitmap_clear_bit (gen, index); | |
3081 | } | |
3082 | } | |
3083 | } | |
3084 | } | |
3085 | else | |
3086 | { | |
3087 | /* The groups are different, if the alias sets | |
3088 | conflict, clear the entire group. We only need | |
3089 | to apply this test if the read_info is a cselib | |
3090 | read. Anything with a constant base cannot alias | |
3091 | something else with a different constant | |
3092 | base. */ | |
3093 | if ((read_info->group_id < 0) | |
b8698a0f | 3094 | && canon_true_dependence (group->base_mem, |
d32f725a | 3095 | GET_MODE (group->base_mem), |
6216f94e | 3096 | group->canon_base_addr, |
53d9622b | 3097 | read_info->mem, NULL_RTX)) |
6fb5fa3c DB |
3098 | { |
3099 | if (kill) | |
3100 | bitmap_ior_into (kill, group->group_kill); | |
3101 | bitmap_and_compl_into (gen, group->group_kill); | |
3102 | } | |
3103 | } | |
3104 | } | |
3105 | } | |
b8698a0f | 3106 | |
6fb5fa3c DB |
3107 | read_info = read_info->next; |
3108 | } | |
3109 | } | |
3110 | ||
3111 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3112 | may be NULL. */ | |
3113 | ||
3114 | static void | |
3115 | scan_reads_spill (read_info_t read_info, bitmap gen, bitmap kill) | |
3116 | { | |
3117 | while (read_info) | |
3118 | { | |
3119 | if (read_info->alias_set) | |
3120 | { | |
b8698a0f | 3121 | int index = get_bitmap_index (clear_alias_group, |
6fb5fa3c DB |
3122 | read_info->alias_set); |
3123 | if (index != 0) | |
3124 | { | |
3125 | if (kill) | |
3126 | bitmap_set_bit (kill, index); | |
3127 | bitmap_clear_bit (gen, index); | |
3128 | } | |
3129 | } | |
b8698a0f | 3130 | |
6fb5fa3c DB |
3131 | read_info = read_info->next; |
3132 | } | |
3133 | } | |
3134 | ||
3135 | ||
3136 | /* Return the insn in BB_INFO before the first wild read or if there | |
3137 | are no wild reads in the block, return the last insn. */ | |
3138 | ||
3139 | static insn_info_t | |
3140 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3141 | { | |
3142 | insn_info_t insn_info = bb_info->last_insn; | |
3143 | insn_info_t last_wild_read = NULL; | |
3144 | ||
3145 | while (insn_info) | |
3146 | { | |
3147 | if (insn_info->wild_read) | |
3148 | { | |
3149 | last_wild_read = insn_info->prev_insn; | |
3150 | /* Block starts with wild read. */ | |
3151 | if (!last_wild_read) | |
3152 | return NULL; | |
3153 | } | |
3154 | ||
3155 | insn_info = insn_info->prev_insn; | |
3156 | } | |
3157 | ||
3158 | if (last_wild_read) | |
3159 | return last_wild_read; | |
3160 | else | |
3161 | return bb_info->last_insn; | |
3162 | } | |
3163 | ||
3164 | ||
3165 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3166 | the block in order to build the gen and kill sets for the block. | |
3167 | We start at ptr which may be the last insn in the block or may be | |
3168 | the first insn with a wild read. In the latter case we are able to | |
3169 | skip the rest of the block because it just does not matter: | |
3170 | anything that happens is hidden by the wild read. */ | |
3171 | ||
3172 | static void | |
3173 | dse_step3_scan (bool for_spills, basic_block bb) | |
3174 | { | |
3175 | bb_info_t bb_info = bb_table[bb->index]; | |
3176 | insn_info_t insn_info; | |
3177 | ||
3178 | if (for_spills) | |
3179 | /* There are no wild reads in the spill case. */ | |
3180 | insn_info = bb_info->last_insn; | |
3181 | else | |
3182 | insn_info = find_insn_before_first_wild_read (bb_info); | |
b8698a0f | 3183 | |
6fb5fa3c DB |
3184 | /* In the spill case or in the no_spill case if there is no wild |
3185 | read in the block, we will need a kill set. */ | |
3186 | if (insn_info == bb_info->last_insn) | |
3187 | { | |
3188 | if (bb_info->kill) | |
3189 | bitmap_clear (bb_info->kill); | |
3190 | else | |
3f9b14ff | 3191 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c | 3192 | } |
b8698a0f | 3193 | else |
6fb5fa3c DB |
3194 | if (bb_info->kill) |
3195 | BITMAP_FREE (bb_info->kill); | |
3196 | ||
3197 | while (insn_info) | |
3198 | { | |
3199 | /* There may have been code deleted by the dce pass run before | |
3200 | this phase. */ | |
3201 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3202 | { | |
b8698a0f | 3203 | /* Process the read(s) last. */ |
6fb5fa3c DB |
3204 | if (for_spills) |
3205 | { | |
3206 | scan_stores_spill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3207 | scan_reads_spill (insn_info->read_rec, bb_info->gen, bb_info->kill); | |
3208 | } | |
3209 | else | |
3210 | { | |
3211 | scan_stores_nospill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3212 | scan_reads_nospill (insn_info, bb_info->gen, bb_info->kill); | |
3213 | } | |
b8698a0f | 3214 | } |
6fb5fa3c DB |
3215 | |
3216 | insn_info = insn_info->prev_insn; | |
3217 | } | |
3218 | } | |
3219 | ||
3220 | ||
3221 | /* Set the gen set of the exit block, and also any block with no | |
3222 | successors that does not have a wild read. */ | |
3223 | ||
3224 | static void | |
3225 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3226 | { | |
3227 | /* The gen set is all 0's for the exit block except for the | |
3228 | frame_pointer_group. */ | |
b8698a0f | 3229 | |
6fb5fa3c DB |
3230 | if (stores_off_frame_dead_at_return) |
3231 | { | |
3232 | unsigned int i; | |
3233 | group_info_t group; | |
b8698a0f | 3234 | |
9771b263 | 3235 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3236 | { |
3237 | if (group->process_globally && group->frame_related) | |
3238 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3239 | } | |
3240 | } | |
3241 | } | |
3242 | ||
3243 | ||
3244 | /* Find all of the blocks that are not backwards reachable from the | |
3245 | exit block or any block with no successors (BB). These are the | |
3246 | infinite loops or infinite self loops. These blocks will still | |
3247 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3248 | ||
3249 | static void | |
3250 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3251 | { | |
3252 | edge e; | |
3253 | edge_iterator ei; | |
3254 | ||
d7c028c0 | 3255 | if (bitmap_bit_p (unreachable_blocks, bb->index)) |
6fb5fa3c | 3256 | { |
d7c028c0 | 3257 | bitmap_clear_bit (unreachable_blocks, bb->index); |
6fb5fa3c | 3258 | FOR_EACH_EDGE (e, ei, bb->preds) |
b8698a0f | 3259 | { |
6fb5fa3c | 3260 | mark_reachable_blocks (unreachable_blocks, e->src); |
b8698a0f | 3261 | } |
6fb5fa3c DB |
3262 | } |
3263 | } | |
3264 | ||
3265 | /* Build the transfer functions for the function. */ | |
3266 | ||
3267 | static void | |
3268 | dse_step3 (bool for_spills) | |
3269 | { | |
3270 | basic_block bb; | |
8b1c6fd7 | 3271 | sbitmap unreachable_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
6fb5fa3c DB |
3272 | sbitmap_iterator sbi; |
3273 | bitmap all_ones = NULL; | |
3274 | unsigned int i; | |
b8698a0f | 3275 | |
f61e445a | 3276 | bitmap_ones (unreachable_blocks); |
6fb5fa3c | 3277 | |
04a90bec | 3278 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3279 | { |
3280 | bb_info_t bb_info = bb_table[bb->index]; | |
3281 | if (bb_info->gen) | |
3282 | bitmap_clear (bb_info->gen); | |
3283 | else | |
3f9b14ff | 3284 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3285 | |
3286 | if (bb->index == ENTRY_BLOCK) | |
3287 | ; | |
3288 | else if (bb->index == EXIT_BLOCK) | |
3289 | dse_step3_exit_block_scan (bb_info); | |
3290 | else | |
3291 | dse_step3_scan (for_spills, bb); | |
3292 | if (EDGE_COUNT (bb->succs) == 0) | |
3293 | mark_reachable_blocks (unreachable_blocks, bb); | |
3294 | ||
3295 | /* If this is the second time dataflow is run, delete the old | |
3296 | sets. */ | |
3297 | if (bb_info->in) | |
3298 | BITMAP_FREE (bb_info->in); | |
3299 | if (bb_info->out) | |
3300 | BITMAP_FREE (bb_info->out); | |
3301 | } | |
3302 | ||
3303 | /* For any block in an infinite loop, we must initialize the out set | |
3304 | to all ones. This could be expensive, but almost never occurs in | |
3305 | practice. However, it is common in regression tests. */ | |
d4ac4ce2 | 3306 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
6fb5fa3c DB |
3307 | { |
3308 | if (bitmap_bit_p (all_blocks, i)) | |
3309 | { | |
3310 | bb_info_t bb_info = bb_table[i]; | |
3311 | if (!all_ones) | |
3312 | { | |
3313 | unsigned int j; | |
3314 | group_info_t group; | |
3315 | ||
3f9b14ff | 3316 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
9771b263 | 3317 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
6fb5fa3c DB |
3318 | bitmap_ior_into (all_ones, group->group_kill); |
3319 | } | |
3320 | if (!bb_info->out) | |
3321 | { | |
3f9b14ff | 3322 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3323 | bitmap_copy (bb_info->out, all_ones); |
3324 | } | |
3325 | } | |
3326 | } | |
3327 | ||
3328 | if (all_ones) | |
3329 | BITMAP_FREE (all_ones); | |
3330 | sbitmap_free (unreachable_blocks); | |
3331 | } | |
3332 | ||
3333 | ||
3334 | \f | |
3335 | /*---------------------------------------------------------------------------- | |
3336 | Fourth step. | |
3337 | ||
3338 | Solve the bitvector equations. | |
3339 | ----------------------------------------------------------------------------*/ | |
3340 | ||
3341 | ||
3342 | /* Confluence function for blocks with no successors. Create an out | |
3343 | set from the gen set of the exit block. This block logically has | |
3344 | the exit block as a successor. */ | |
3345 | ||
3346 | ||
3347 | ||
3348 | static void | |
3349 | dse_confluence_0 (basic_block bb) | |
3350 | { | |
3351 | bb_info_t bb_info = bb_table[bb->index]; | |
3352 | ||
3353 | if (bb->index == EXIT_BLOCK) | |
3354 | return; | |
3355 | ||
3356 | if (!bb_info->out) | |
3357 | { | |
3f9b14ff | 3358 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3359 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3360 | } | |
3361 | } | |
3362 | ||
3363 | /* Propagate the information from the in set of the dest of E to the | |
3364 | out set of the src of E. If the various in or out sets are not | |
3365 | there, that means they are all ones. */ | |
3366 | ||
1a0f3fa1 | 3367 | static bool |
6fb5fa3c DB |
3368 | dse_confluence_n (edge e) |
3369 | { | |
3370 | bb_info_t src_info = bb_table[e->src->index]; | |
3371 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3372 | ||
3373 | if (dest_info->in) | |
3374 | { | |
3375 | if (src_info->out) | |
3376 | bitmap_and_into (src_info->out, dest_info->in); | |
3377 | else | |
3378 | { | |
3f9b14ff | 3379 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3380 | bitmap_copy (src_info->out, dest_info->in); |
3381 | } | |
3382 | } | |
1a0f3fa1 | 3383 | return true; |
6fb5fa3c DB |
3384 | } |
3385 | ||
3386 | ||
3387 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
b8698a0f | 3388 | block. There are three cases: |
6fb5fa3c DB |
3389 | |
3390 | 1) The block has no kill set. In this case the kill set is all | |
3391 | ones. It does not matter what the out set of the block is, none of | |
3392 | the info can reach the top. The only thing that reaches the top is | |
3393 | the gen set and we just copy the set. | |
3394 | ||
3395 | 2) There is a kill set but no out set and bb has successors. In | |
3396 | this case we just return. Eventually an out set will be created and | |
3397 | it is better to wait than to create a set of ones. | |
3398 | ||
3399 | 3) There is both a kill and out set. We apply the obvious transfer | |
3400 | function. | |
3401 | */ | |
3402 | ||
3403 | static bool | |
3404 | dse_transfer_function (int bb_index) | |
3405 | { | |
3406 | bb_info_t bb_info = bb_table[bb_index]; | |
3407 | ||
3408 | if (bb_info->kill) | |
3409 | { | |
3410 | if (bb_info->out) | |
3411 | { | |
3412 | /* Case 3 above. */ | |
3413 | if (bb_info->in) | |
b8698a0f | 3414 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3415 | bb_info->out, bb_info->kill); |
3416 | else | |
3417 | { | |
3f9b14ff | 3418 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
b8698a0f | 3419 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3420 | bb_info->out, bb_info->kill); |
3421 | return true; | |
3422 | } | |
3423 | } | |
3424 | else | |
3425 | /* Case 2 above. */ | |
3426 | return false; | |
3427 | } | |
3428 | else | |
3429 | { | |
3430 | /* Case 1 above. If there is already an in set, nothing | |
3431 | happens. */ | |
3432 | if (bb_info->in) | |
3433 | return false; | |
3434 | else | |
3435 | { | |
3f9b14ff | 3436 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3437 | bitmap_copy (bb_info->in, bb_info->gen); |
3438 | return true; | |
3439 | } | |
3440 | } | |
3441 | } | |
3442 | ||
3443 | /* Solve the dataflow equations. */ | |
3444 | ||
3445 | static void | |
3446 | dse_step4 (void) | |
3447 | { | |
b8698a0f L |
3448 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3449 | dse_confluence_n, dse_transfer_function, | |
3450 | all_blocks, df_get_postorder (DF_BACKWARD), | |
6fb5fa3c | 3451 | df_get_n_blocks (DF_BACKWARD)); |
456610d3 | 3452 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3453 | { |
3454 | basic_block bb; | |
3455 | ||
3456 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
04a90bec | 3457 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3458 | { |
3459 | bb_info_t bb_info = bb_table[bb->index]; | |
3460 | ||
3461 | df_print_bb_index (bb, dump_file); | |
3462 | if (bb_info->in) | |
3463 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3464 | else | |
3465 | fprintf (dump_file, " in: *MISSING*\n"); | |
3466 | if (bb_info->gen) | |
3467 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3468 | else | |
3469 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3470 | if (bb_info->kill) | |
3471 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3472 | else | |
3473 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3474 | if (bb_info->out) | |
3475 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3476 | else | |
3477 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3478 | } | |
3479 | } | |
3480 | } | |
3481 | ||
3482 | ||
3483 | \f | |
3484 | /*---------------------------------------------------------------------------- | |
3485 | Fifth step. | |
3486 | ||
0d52bcc1 | 3487 | Delete the stores that can only be deleted using the global information. |
6fb5fa3c DB |
3488 | ----------------------------------------------------------------------------*/ |
3489 | ||
3490 | ||
3491 | static void | |
3492 | dse_step5_nospill (void) | |
3493 | { | |
3494 | basic_block bb; | |
11cd3bed | 3495 | FOR_EACH_BB_FN (bb, cfun) |
6fb5fa3c DB |
3496 | { |
3497 | bb_info_t bb_info = bb_table[bb->index]; | |
3498 | insn_info_t insn_info = bb_info->last_insn; | |
3499 | bitmap v = bb_info->out; | |
3500 | ||
3501 | while (insn_info) | |
3502 | { | |
3503 | bool deleted = false; | |
3504 | if (dump_file && insn_info->insn) | |
3505 | { | |
3506 | fprintf (dump_file, "starting to process insn %d\n", | |
3507 | INSN_UID (insn_info->insn)); | |
3508 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3509 | } | |
3510 | ||
3511 | /* There may have been code deleted by the dce pass run before | |
3512 | this phase. */ | |
b8698a0f | 3513 | if (insn_info->insn |
6fb5fa3c DB |
3514 | && INSN_P (insn_info->insn) |
3515 | && (!insn_info->cannot_delete) | |
3516 | && (!bitmap_empty_p (v))) | |
3517 | { | |
3518 | store_info_t store_info = insn_info->store_rec; | |
3519 | ||
3520 | /* Try to delete the current insn. */ | |
3521 | deleted = true; | |
b8698a0f | 3522 | |
6fb5fa3c DB |
3523 | /* Skip the clobbers. */ |
3524 | while (!store_info->is_set) | |
3525 | store_info = store_info->next; | |
3526 | ||
3527 | if (store_info->alias_set) | |
3528 | deleted = false; | |
3529 | else | |
3530 | { | |
3531 | HOST_WIDE_INT i; | |
b8698a0f | 3532 | group_info_t group_info |
9771b263 | 3533 | = rtx_group_vec[store_info->group_id]; |
b8698a0f | 3534 | |
6fb5fa3c DB |
3535 | for (i = store_info->begin; i < store_info->end; i++) |
3536 | { | |
3537 | int index = get_bitmap_index (group_info, i); | |
b8698a0f | 3538 | |
456610d3 | 3539 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 3540 | fprintf (dump_file, "i = %d, index = %d\n", (int)i, index); |
6fb5fa3c DB |
3541 | if (index == 0 || !bitmap_bit_p (v, index)) |
3542 | { | |
456610d3 | 3543 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 3544 | fprintf (dump_file, "failing at i = %d\n", (int)i); |
6fb5fa3c DB |
3545 | deleted = false; |
3546 | break; | |
3547 | } | |
3548 | } | |
3549 | } | |
3550 | if (deleted) | |
3551 | { | |
9e582b1d JR |
3552 | if (dbg_cnt (dse) |
3553 | && check_for_inc_dec_1 (insn_info)) | |
6fb5fa3c | 3554 | { |
6fb5fa3c DB |
3555 | delete_insn (insn_info->insn); |
3556 | insn_info->insn = NULL; | |
3557 | globally_deleted++; | |
3558 | } | |
3559 | } | |
3560 | } | |
3561 | /* We do want to process the local info if the insn was | |
6ed3da00 | 3562 | deleted. For instance, if the insn did a wild read, we |
6fb5fa3c | 3563 | no longer need to trash the info. */ |
b8698a0f | 3564 | if (insn_info->insn |
6fb5fa3c DB |
3565 | && INSN_P (insn_info->insn) |
3566 | && (!deleted)) | |
3567 | { | |
3568 | scan_stores_nospill (insn_info->store_rec, v, NULL); | |
3569 | if (insn_info->wild_read) | |
3570 | { | |
456610d3 | 3571 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3572 | fprintf (dump_file, "wild read\n"); |
3573 | bitmap_clear (v); | |
3574 | } | |
d26c7090 ER |
3575 | else if (insn_info->read_rec |
3576 | || insn_info->non_frame_wild_read) | |
6fb5fa3c | 3577 | { |
d26c7090 | 3578 | if (dump_file && !insn_info->non_frame_wild_read) |
6fb5fa3c | 3579 | fprintf (dump_file, "regular read\n"); |
456610d3 | 3580 | else if (dump_file && (dump_flags & TDF_DETAILS)) |
d26c7090 | 3581 | fprintf (dump_file, "non-frame wild read\n"); |
6fb5fa3c DB |
3582 | scan_reads_nospill (insn_info, v, NULL); |
3583 | } | |
3584 | } | |
b8698a0f | 3585 | |
6fb5fa3c DB |
3586 | insn_info = insn_info->prev_insn; |
3587 | } | |
3588 | } | |
3589 | } | |
3590 | ||
3591 | ||
6fb5fa3c DB |
3592 | \f |
3593 | /*---------------------------------------------------------------------------- | |
3594 | Sixth step. | |
3595 | ||
8dd5516b JJ |
3596 | Delete stores made redundant by earlier stores (which store the same |
3597 | value) that couldn't be eliminated. | |
3598 | ----------------------------------------------------------------------------*/ | |
3599 | ||
3600 | static void | |
3601 | dse_step6 (void) | |
3602 | { | |
3603 | basic_block bb; | |
3604 | ||
04a90bec | 3605 | FOR_ALL_BB_FN (bb, cfun) |
8dd5516b JJ |
3606 | { |
3607 | bb_info_t bb_info = bb_table[bb->index]; | |
3608 | insn_info_t insn_info = bb_info->last_insn; | |
3609 | ||
3610 | while (insn_info) | |
3611 | { | |
3612 | /* There may have been code deleted by the dce pass run before | |
3613 | this phase. */ | |
3614 | if (insn_info->insn | |
3615 | && INSN_P (insn_info->insn) | |
3616 | && !insn_info->cannot_delete) | |
3617 | { | |
3618 | store_info_t s_info = insn_info->store_rec; | |
3619 | ||
3620 | while (s_info && !s_info->is_set) | |
3621 | s_info = s_info->next; | |
3622 | if (s_info | |
3623 | && s_info->redundant_reason | |
3624 | && s_info->redundant_reason->insn | |
3625 | && INSN_P (s_info->redundant_reason->insn)) | |
3626 | { | |
eb92d49a | 3627 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
456610d3 | 3628 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
3629 | fprintf (dump_file, "Locally deleting insn %d " |
3630 | "because insn %d stores the " | |
3631 | "same value and couldn't be " | |
3632 | "eliminated\n", | |
3633 | INSN_UID (insn_info->insn), | |
3634 | INSN_UID (rinsn)); | |
3635 | delete_dead_store_insn (insn_info); | |
3636 | } | |
3637 | } | |
3638 | insn_info = insn_info->prev_insn; | |
3639 | } | |
3640 | } | |
3641 | } | |
3642 | \f | |
3643 | /*---------------------------------------------------------------------------- | |
3644 | Seventh step. | |
3645 | ||
b8698a0f | 3646 | Destroy everything left standing. |
6fb5fa3c DB |
3647 | ----------------------------------------------------------------------------*/ |
3648 | ||
b8698a0f | 3649 | static void |
3f9b14ff | 3650 | dse_step7 (void) |
6fb5fa3c | 3651 | { |
3f9b14ff SB |
3652 | bitmap_obstack_release (&dse_bitmap_obstack); |
3653 | obstack_free (&dse_obstack, NULL); | |
370f38e8 | 3654 | |
6fb5fa3c DB |
3655 | end_alias_analysis (); |
3656 | free (bb_table); | |
c203e8a7 TS |
3657 | delete rtx_group_table; |
3658 | rtx_group_table = NULL; | |
9771b263 | 3659 | rtx_group_vec.release (); |
6fb5fa3c DB |
3660 | BITMAP_FREE (all_blocks); |
3661 | BITMAP_FREE (scratch); | |
3662 | ||
3663 | free_alloc_pool (rtx_store_info_pool); | |
3664 | free_alloc_pool (read_info_pool); | |
3665 | free_alloc_pool (insn_info_pool); | |
3666 | free_alloc_pool (bb_info_pool); | |
3667 | free_alloc_pool (rtx_group_info_pool); | |
3668 | free_alloc_pool (deferred_change_pool); | |
3669 | } | |
3670 | ||
3671 | ||
6fb5fa3c DB |
3672 | /* ------------------------------------------------------------------------- |
3673 | DSE | |
3674 | ------------------------------------------------------------------------- */ | |
3675 | ||
3676 | /* Callback for running pass_rtl_dse. */ | |
3677 | ||
3678 | static unsigned int | |
3679 | rest_of_handle_dse (void) | |
3680 | { | |
6fb5fa3c DB |
3681 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3682 | ||
02b47899 KZ |
3683 | /* Need the notes since we must track live hardregs in the forwards |
3684 | direction. */ | |
3685 | df_note_add_problem (); | |
3686 | df_analyze (); | |
3687 | ||
6fb5fa3c DB |
3688 | dse_step0 (); |
3689 | dse_step1 (); | |
3690 | dse_step2_init (); | |
3691 | if (dse_step2_nospill ()) | |
3692 | { | |
3693 | df_set_flags (DF_LR_RUN_DCE); | |
3694 | df_analyze (); | |
456610d3 | 3695 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3696 | fprintf (dump_file, "doing global processing\n"); |
3697 | dse_step3 (false); | |
3698 | dse_step4 (); | |
3699 | dse_step5_nospill (); | |
3700 | } | |
3701 | ||
8dd5516b | 3702 | dse_step6 (); |
3f9b14ff | 3703 | dse_step7 (); |
6fb5fa3c DB |
3704 | |
3705 | if (dump_file) | |
3706 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d, spill deletions = %d\n", | |
3707 | locally_deleted, globally_deleted, spill_deleted); | |
3708 | return 0; | |
3709 | } | |
3710 | ||
27a4cd48 DM |
3711 | namespace { |
3712 | ||
3713 | const pass_data pass_data_rtl_dse1 = | |
3714 | { | |
3715 | RTL_PASS, /* type */ | |
3716 | "dse1", /* name */ | |
3717 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3718 | TV_DSE1, /* tv_id */ |
3719 | 0, /* properties_required */ | |
3720 | 0, /* properties_provided */ | |
3721 | 0, /* properties_destroyed */ | |
3722 | 0, /* todo_flags_start */ | |
3bea341f | 3723 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c DB |
3724 | }; |
3725 | ||
27a4cd48 DM |
3726 | class pass_rtl_dse1 : public rtl_opt_pass |
3727 | { | |
3728 | public: | |
c3284718 RS |
3729 | pass_rtl_dse1 (gcc::context *ctxt) |
3730 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
27a4cd48 DM |
3731 | {} |
3732 | ||
3733 | /* opt_pass methods: */ | |
1a3d085c TS |
3734 | virtual bool gate (function *) |
3735 | { | |
3736 | return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
3737 | } | |
3738 | ||
be55bfe6 | 3739 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3740 | |
3741 | }; // class pass_rtl_dse1 | |
3742 | ||
3743 | } // anon namespace | |
3744 | ||
3745 | rtl_opt_pass * | |
3746 | make_pass_rtl_dse1 (gcc::context *ctxt) | |
3747 | { | |
3748 | return new pass_rtl_dse1 (ctxt); | |
3749 | } | |
3750 | ||
3751 | namespace { | |
3752 | ||
3753 | const pass_data pass_data_rtl_dse2 = | |
3754 | { | |
3755 | RTL_PASS, /* type */ | |
3756 | "dse2", /* name */ | |
3757 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3758 | TV_DSE2, /* tv_id */ |
3759 | 0, /* properties_required */ | |
3760 | 0, /* properties_provided */ | |
3761 | 0, /* properties_destroyed */ | |
3762 | 0, /* todo_flags_start */ | |
3bea341f | 3763 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c | 3764 | }; |
27a4cd48 DM |
3765 | |
3766 | class pass_rtl_dse2 : public rtl_opt_pass | |
3767 | { | |
3768 | public: | |
c3284718 RS |
3769 | pass_rtl_dse2 (gcc::context *ctxt) |
3770 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
27a4cd48 DM |
3771 | {} |
3772 | ||
3773 | /* opt_pass methods: */ | |
1a3d085c TS |
3774 | virtual bool gate (function *) |
3775 | { | |
3776 | return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
3777 | } | |
3778 | ||
be55bfe6 | 3779 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3780 | |
3781 | }; // class pass_rtl_dse2 | |
3782 | ||
3783 | } // anon namespace | |
3784 | ||
3785 | rtl_opt_pass * | |
3786 | make_pass_rtl_dse2 (gcc::context *ctxt) | |
3787 | { | |
3788 | return new pass_rtl_dse2 (ctxt); | |
3789 | } |