]>
Commit | Line | Data |
---|---|---|
3072d30e | 1 | /* RTL dead store elimination. |
0ac758f7 | 2 | Copyright (C) 2005, 2006, 2007, 2008, 2009 Free Software Foundation, Inc. |
3072d30e | 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 | |
8c4c00c1 | 11 | Software Foundation; either version 3, or (at your option) any later |
3072d30e | 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 | |
8c4c00c1 | 20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ | |
3072d30e | 22 | |
23 | #undef BASELINE | |
24 | ||
25 | #include "config.h" | |
26 | #include "system.h" | |
27 | #include "coretypes.h" | |
28 | #include "hashtab.h" | |
29 | #include "tm.h" | |
30 | #include "rtl.h" | |
31 | #include "tree.h" | |
10d4de0e | 32 | #include "tm_p.h" |
3072d30e | 33 | #include "regs.h" |
34 | #include "hard-reg-set.h" | |
35 | #include "flags.h" | |
36 | #include "df.h" | |
37 | #include "cselib.h" | |
38 | #include "timevar.h" | |
39 | #include "tree-pass.h" | |
40 | #include "alloc-pool.h" | |
41 | #include "alias.h" | |
42 | #include "insn-config.h" | |
43 | #include "expr.h" | |
44 | #include "recog.h" | |
45 | #include "dse.h" | |
5c9051a4 | 46 | #include "optabs.h" |
3072d30e | 47 | #include "dbgcnt.h" |
9311ed8a | 48 | #include "target.h" |
3072d30e | 49 | |
50 | /* This file contains three techniques for performing Dead Store | |
48e1416a | 51 | Elimination (dse). |
3072d30e | 52 | |
53 | * The first technique performs dse locally on any base address. It | |
54 | is based on the cselib which is a local value numbering technique. | |
55 | This technique is local to a basic block but deals with a fairly | |
56 | general addresses. | |
48e1416a | 57 | |
3072d30e | 58 | * The second technique performs dse globally but is restricted to |
59 | base addresses that are either constant or are relative to the | |
60 | frame_pointer. | |
61 | ||
62 | * The third technique, (which is only done after register allocation) | |
63 | processes the spill spill slots. This differs from the second | |
64 | technique because it takes advantage of the fact that spilling is | |
65 | completely free from the effects of aliasing. | |
66 | ||
67 | Logically, dse is a backwards dataflow problem. A store can be | |
68 | deleted if it if cannot be reached in the backward direction by any | |
69 | use of the value being stored. However, the local technique uses a | |
70 | forwards scan of the basic block because cselib requires that the | |
71 | block be processed in that order. | |
72 | ||
73 | The pass is logically broken into 7 steps: | |
74 | ||
75 | 0) Initialization. | |
76 | ||
77 | 1) The local algorithm, as well as scanning the insns for the two | |
78 | global algorithms. | |
79 | ||
80 | 2) Analysis to see if the global algs are necessary. In the case | |
81 | of stores base on a constant address, there must be at least two | |
82 | stores to that address, to make it possible to delete some of the | |
83 | stores. In the case of stores off of the frame or spill related | |
84 | stores, only one store to an address is necessary because those | |
85 | stores die at the end of the function. | |
86 | ||
48e1416a | 87 | 3) Set up the global dataflow equations based on processing the |
3072d30e | 88 | info parsed in the first step. |
89 | ||
90 | 4) Solve the dataflow equations. | |
91 | ||
92 | 5) Delete the insns that the global analysis has indicated are | |
93 | unnecessary. | |
94 | ||
aa140b76 | 95 | 6) Delete insns that store the same value as preceeding store |
96 | where the earlier store couldn't be eliminated. | |
97 | ||
98 | 7) Cleanup. | |
3072d30e | 99 | |
100 | This step uses cselib and canon_rtx to build the largest expression | |
101 | possible for each address. This pass is a forwards pass through | |
102 | each basic block. From the point of view of the global technique, | |
103 | the first pass could examine a block in either direction. The | |
bef304b8 | 104 | forwards ordering is to accommodate cselib. |
3072d30e | 105 | |
106 | We a simplifying assumption: addresses fall into four broad | |
107 | categories: | |
108 | ||
109 | 1) base has rtx_varies_p == false, offset is constant. | |
110 | 2) base has rtx_varies_p == false, offset variable. | |
111 | 3) base has rtx_varies_p == true, offset constant. | |
112 | 4) base has rtx_varies_p == true, offset variable. | |
113 | ||
114 | The local passes are able to process all 4 kinds of addresses. The | |
115 | global pass only handles (1). | |
116 | ||
117 | The global problem is formulated as follows: | |
118 | ||
119 | A store, S1, to address A, where A is not relative to the stack | |
120 | frame, can be eliminated if all paths from S1 to the end of the | |
121 | of the function contain another store to A before a read to A. | |
122 | ||
123 | If the address A is relative to the stack frame, a store S2 to A | |
124 | can be eliminated if there are no paths from S1 that reach the | |
125 | end of the function that read A before another store to A. In | |
126 | this case S2 can be deleted if there are paths to from S2 to the | |
127 | end of the function that have no reads or writes to A. This | |
128 | second case allows stores to the stack frame to be deleted that | |
129 | would otherwise die when the function returns. This cannot be | |
130 | done if stores_off_frame_dead_at_return is not true. See the doc | |
131 | for that variable for when this variable is false. | |
132 | ||
133 | The global problem is formulated as a backwards set union | |
134 | dataflow problem where the stores are the gens and reads are the | |
135 | kills. Set union problems are rare and require some special | |
136 | handling given our representation of bitmaps. A straightforward | |
137 | implementation of requires a lot of bitmaps filled with 1s. | |
138 | These are expensive and cumbersome in our bitmap formulation so | |
139 | care has been taken to avoid large vectors filled with 1s. See | |
140 | the comments in bb_info and in the dataflow confluence functions | |
48e1416a | 141 | for details. |
3072d30e | 142 | |
143 | There are two places for further enhancements to this algorithm: | |
48e1416a | 144 | |
3072d30e | 145 | 1) The original dse which was embedded in a pass called flow also |
146 | did local address forwarding. For example in | |
147 | ||
148 | A <- r100 | |
149 | ... <- A | |
150 | ||
151 | flow would replace the right hand side of the second insn with a | |
6dfdc153 | 152 | reference to r100. Most of the information is available to add this |
3072d30e | 153 | to this pass. It has not done it because it is a lot of work in |
154 | the case that either r100 is assigned to between the first and | |
155 | second insn and/or the second insn is a load of part of the value | |
156 | stored by the first insn. | |
157 | ||
158 | insn 5 in gcc.c-torture/compile/990203-1.c simple case. | |
159 | insn 15 in gcc.c-torture/execute/20001017-2.c simple case. | |
160 | insn 25 in gcc.c-torture/execute/20001026-1.c simple case. | |
161 | insn 44 in gcc.c-torture/execute/20010910-1.c simple case. | |
162 | ||
163 | 2) The cleaning up of spill code is quite profitable. It currently | |
164 | depends on reading tea leaves and chicken entrails left by reload. | |
165 | This pass depends on reload creating a singleton alias set for each | |
166 | spill slot and telling the next dse pass which of these alias sets | |
167 | are the singletons. Rather than analyze the addresses of the | |
168 | spills, dse's spill processing just does analysis of the loads and | |
169 | stores that use those alias sets. There are three cases where this | |
170 | falls short: | |
171 | ||
172 | a) Reload sometimes creates the slot for one mode of access, and | |
173 | then inserts loads and/or stores for a smaller mode. In this | |
174 | case, the current code just punts on the slot. The proper thing | |
175 | to do is to back out and use one bit vector position for each | |
176 | byte of the entity associated with the slot. This depends on | |
177 | KNOWING that reload always generates the accesses for each of the | |
178 | bytes in some canonical (read that easy to understand several | |
179 | passes after reload happens) way. | |
180 | ||
181 | b) Reload sometimes decides that spill slot it allocated was not | |
182 | large enough for the mode and goes back and allocates more slots | |
183 | with the same mode and alias set. The backout in this case is a | |
184 | little more graceful than (a). In this case the slot is unmarked | |
185 | as being a spill slot and if final address comes out to be based | |
48e1416a | 186 | off the frame pointer, the global algorithm handles this slot. |
3072d30e | 187 | |
188 | c) For any pass that may prespill, there is currently no | |
189 | mechanism to tell the dse pass that the slot being used has the | |
190 | special properties that reload uses. It may be that all that is | |
bef304b8 | 191 | required is to have those passes make the same calls that reload |
3072d30e | 192 | does, assuming that the alias sets can be manipulated in the same |
193 | way. */ | |
194 | ||
195 | /* There are limits to the size of constant offsets we model for the | |
196 | global problem. There are certainly test cases, that exceed this | |
197 | limit, however, it is unlikely that there are important programs | |
198 | that really have constant offsets this size. */ | |
199 | #define MAX_OFFSET (64 * 1024) | |
200 | ||
201 | ||
202 | static bitmap scratch = NULL; | |
203 | struct insn_info; | |
204 | ||
205 | /* This structure holds information about a candidate store. */ | |
48e1416a | 206 | struct store_info |
3072d30e | 207 | { |
208 | ||
209 | /* False means this is a clobber. */ | |
210 | bool is_set; | |
211 | ||
aa140b76 | 212 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
213 | bool is_large; | |
214 | ||
3072d30e | 215 | /* The id of the mem group of the base address. If rtx_varies_p is |
216 | true, this is -1. Otherwise, it is the index into the group | |
217 | table. */ | |
218 | int group_id; | |
48e1416a | 219 | |
3072d30e | 220 | /* This is the cselib value. */ |
221 | cselib_val *cse_base; | |
222 | ||
223 | /* This canonized mem. */ | |
224 | rtx mem; | |
225 | ||
82d2c88b | 226 | /* Canonized MEM address for use by canon_true_dependence. */ |
3072d30e | 227 | rtx mem_addr; |
228 | ||
229 | /* If this is non-zero, it is the alias set of a spill location. */ | |
32c2fdea | 230 | alias_set_type alias_set; |
3072d30e | 231 | |
232 | /* The offset of the first and byte before the last byte associated | |
233 | with the operation. */ | |
aa140b76 | 234 | HOST_WIDE_INT begin, end; |
235 | ||
236 | union | |
237 | { | |
238 | /* A bitmask as wide as the number of bytes in the word that | |
239 | contains a 1 if the byte may be needed. The store is unused if | |
240 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
241 | unsigned HOST_WIDE_INT small_bitmask; | |
242 | ||
243 | struct | |
244 | { | |
245 | /* A bitmap with one bit per byte. Cleared bit means the position | |
246 | is needed. Used if IS_LARGE is false. */ | |
843bd2fa | 247 | bitmap bmap; |
3072d30e | 248 | |
aa140b76 | 249 | /* Number of set bits (i.e. unneeded bytes) in BITMAP. If it is |
250 | equal to END - BEGIN, the whole store is unused. */ | |
251 | int count; | |
252 | } large; | |
253 | } positions_needed; | |
3072d30e | 254 | |
255 | /* The next store info for this insn. */ | |
256 | struct store_info *next; | |
257 | ||
258 | /* The right hand side of the store. This is used if there is a | |
259 | subsequent reload of the mems address somewhere later in the | |
260 | basic block. */ | |
aa140b76 | 261 | rtx rhs; |
262 | ||
263 | /* If rhs is or holds a constant, this contains that constant, | |
264 | otherwise NULL. */ | |
265 | rtx const_rhs; | |
266 | ||
267 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
268 | insn stored. These aren't eliminated early, because doing that | |
269 | might prevent the earlier larger store to be eliminated. */ | |
270 | struct insn_info *redundant_reason; | |
3072d30e | 271 | }; |
272 | ||
4e43e20a | 273 | /* Return a bitmask with the first N low bits set. */ |
274 | ||
275 | static unsigned HOST_WIDE_INT | |
276 | lowpart_bitmask (int n) | |
277 | { | |
278 | unsigned HOST_WIDE_INT mask = ~(unsigned HOST_WIDE_INT) 0; | |
279 | return mask >> (HOST_BITS_PER_WIDE_INT - n); | |
280 | } | |
281 | ||
3072d30e | 282 | typedef struct store_info *store_info_t; |
283 | static alloc_pool cse_store_info_pool; | |
284 | static alloc_pool rtx_store_info_pool; | |
285 | ||
286 | /* This structure holds information about a load. These are only | |
287 | built for rtx bases. */ | |
48e1416a | 288 | struct read_info |
3072d30e | 289 | { |
290 | /* The id of the mem group of the base address. */ | |
291 | int group_id; | |
292 | ||
293 | /* If this is non-zero, it is the alias set of a spill location. */ | |
32c2fdea | 294 | alias_set_type alias_set; |
3072d30e | 295 | |
296 | /* The offset of the first and byte after the last byte associated | |
297 | with the operation. If begin == end == 0, the read did not have | |
298 | a constant offset. */ | |
299 | int begin, end; | |
300 | ||
301 | /* The mem being read. */ | |
302 | rtx mem; | |
303 | ||
304 | /* The next read_info for this insn. */ | |
305 | struct read_info *next; | |
306 | }; | |
307 | typedef struct read_info *read_info_t; | |
308 | static alloc_pool read_info_pool; | |
309 | ||
310 | ||
311 | /* One of these records is created for each insn. */ | |
312 | ||
48e1416a | 313 | struct insn_info |
3072d30e | 314 | { |
315 | /* Set true if the insn contains a store but the insn itself cannot | |
316 | be deleted. This is set if the insn is a parallel and there is | |
317 | more than one non dead output or if the insn is in some way | |
318 | volatile. */ | |
319 | bool cannot_delete; | |
320 | ||
321 | /* This field is only used by the global algorithm. It is set true | |
322 | if the insn contains any read of mem except for a (1). This is | |
323 | also set if the insn is a call or has a clobber mem. If the insn | |
324 | contains a wild read, the use_rec will be null. */ | |
325 | bool wild_read; | |
326 | ||
17e1318c | 327 | /* This field is only used for the processing of const functions. |
328 | These functions cannot read memory, but they can read the stack | |
16bf64db | 329 | because that is where they may get their parms. We need to be |
330 | this conservative because, like the store motion pass, we don't | |
331 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
332 | Moreover, we need to distinguish two cases: | |
333 | 1. Before reload (register elimination), the stores related to | |
334 | outgoing arguments are stack pointer based and thus deemed | |
335 | of non-constant base in this pass. This requires special | |
336 | handling but also means that the frame pointer based stores | |
337 | need not be killed upon encountering a const function call. | |
338 | 2. After reload, the stores related to outgoing arguments can be | |
339 | either stack pointer or hard frame pointer based. This means | |
340 | that we have no other choice than also killing all the frame | |
341 | pointer based stores upon encountering a const function call. | |
342 | This field is set after reload for const function calls. Having | |
343 | this set is less severe than a wild read, it just means that all | |
344 | the frame related stores are killed rather than all the stores. */ | |
345 | bool frame_read; | |
346 | ||
347 | /* This field is only used for the processing of const functions. | |
348 | It is set if the insn may contain a stack pointer based store. */ | |
17e1318c | 349 | bool stack_pointer_based; |
3072d30e | 350 | |
351 | /* This is true if any of the sets within the store contains a | |
352 | cselib base. Such stores can only be deleted by the local | |
353 | algorithm. */ | |
354 | bool contains_cselib_groups; | |
355 | ||
356 | /* The insn. */ | |
357 | rtx insn; | |
358 | ||
359 | /* The list of mem sets or mem clobbers that are contained in this | |
360 | insn. If the insn is deletable, it contains only one mem set. | |
361 | But it could also contain clobbers. Insns that contain more than | |
362 | one mem set are not deletable, but each of those mems are here in | |
6dfdc153 | 363 | order to provide info to delete other insns. */ |
3072d30e | 364 | store_info_t store_rec; |
365 | ||
366 | /* The linked list of mem uses in this insn. Only the reads from | |
367 | rtx bases are listed here. The reads to cselib bases are | |
368 | completely processed during the first scan and so are never | |
369 | created. */ | |
370 | read_info_t read_rec; | |
371 | ||
372 | /* The prev insn in the basic block. */ | |
373 | struct insn_info * prev_insn; | |
374 | ||
375 | /* The linked list of insns that are in consideration for removal in | |
376 | the forwards pass thru the basic block. This pointer may be | |
377 | trash as it is not cleared when a wild read occurs. The only | |
f0b5f617 | 378 | time it is guaranteed to be correct is when the traversal starts |
3072d30e | 379 | at active_local_stores. */ |
380 | struct insn_info * next_local_store; | |
381 | }; | |
382 | ||
383 | typedef struct insn_info *insn_info_t; | |
384 | static alloc_pool insn_info_pool; | |
385 | ||
386 | /* The linked list of stores that are under consideration in this | |
48e1416a | 387 | basic block. */ |
3072d30e | 388 | static insn_info_t active_local_stores; |
389 | ||
48e1416a | 390 | struct bb_info |
3072d30e | 391 | { |
392 | ||
393 | /* Pointer to the insn info for the last insn in the block. These | |
394 | are linked so this is how all of the insns are reached. During | |
395 | scanning this is the current insn being scanned. */ | |
396 | insn_info_t last_insn; | |
397 | ||
398 | /* The info for the global dataflow problem. */ | |
399 | ||
400 | ||
401 | /* This is set if the transfer function should and in the wild_read | |
402 | bitmap before applying the kill and gen sets. That vector knocks | |
403 | out most of the bits in the bitmap and thus speeds up the | |
404 | operations. */ | |
405 | bool apply_wild_read; | |
406 | ||
a1b0a968 | 407 | /* The following 4 bitvectors hold information about which positions |
408 | of which stores are live or dead. They are indexed by | |
409 | get_bitmap_index. */ | |
410 | ||
3072d30e | 411 | /* The set of store positions that exist in this block before a wild read. */ |
412 | bitmap gen; | |
48e1416a | 413 | |
3072d30e | 414 | /* The set of load positions that exist in this block above the |
415 | same position of a store. */ | |
416 | bitmap kill; | |
417 | ||
418 | /* The set of stores that reach the top of the block without being | |
419 | killed by a read. | |
420 | ||
421 | Do not represent the in if it is all ones. Note that this is | |
422 | what the bitvector should logically be initialized to for a set | |
423 | intersection problem. However, like the kill set, this is too | |
424 | expensive. So initially, the in set will only be created for the | |
425 | exit block and any block that contains a wild read. */ | |
426 | bitmap in; | |
427 | ||
428 | /* The set of stores that reach the bottom of the block from it's | |
429 | successors. | |
430 | ||
431 | Do not represent the in if it is all ones. Note that this is | |
432 | what the bitvector should logically be initialized to for a set | |
433 | intersection problem. However, like the kill and in set, this is | |
434 | too expensive. So what is done is that the confluence operator | |
435 | just initializes the vector from one of the out sets of the | |
436 | successors of the block. */ | |
437 | bitmap out; | |
a1b0a968 | 438 | |
439 | /* The following bitvector is indexed by the reg number. It | |
440 | contains the set of regs that are live at the current instruction | |
441 | being processed. While it contains info for all of the | |
442 | registers, only the pseudos are actually examined. It is used to | |
443 | assure that shift sequences that are inserted do not accidently | |
444 | clobber live hard regs. */ | |
445 | bitmap regs_live; | |
3072d30e | 446 | }; |
447 | ||
448 | typedef struct bb_info *bb_info_t; | |
449 | static alloc_pool bb_info_pool; | |
450 | ||
451 | /* Table to hold all bb_infos. */ | |
452 | static bb_info_t *bb_table; | |
453 | ||
454 | /* There is a group_info for each rtx base that is used to reference | |
455 | memory. There are also not many of the rtx bases because they are | |
456 | very limited in scope. */ | |
457 | ||
48e1416a | 458 | struct group_info |
3072d30e | 459 | { |
460 | /* The actual base of the address. */ | |
461 | rtx rtx_base; | |
462 | ||
463 | /* The sequential id of the base. This allows us to have a | |
464 | canonical ordering of these that is not based on addresses. */ | |
465 | int id; | |
466 | ||
0ac758f7 | 467 | /* True if there are any positions that are to be processed |
468 | globally. */ | |
469 | bool process_globally; | |
470 | ||
471 | /* True if the base of this group is either the frame_pointer or | |
472 | hard_frame_pointer. */ | |
473 | bool frame_related; | |
474 | ||
3072d30e | 475 | /* A mem wrapped around the base pointer for the group in order to |
476 | do read dependency. */ | |
477 | rtx base_mem; | |
48e1416a | 478 | |
82d2c88b | 479 | /* Canonized version of base_mem's address. */ |
480 | rtx canon_base_addr; | |
3072d30e | 481 | |
482 | /* These two sets of two bitmaps are used to keep track of how many | |
6dfdc153 | 483 | stores are actually referencing that position from this base. We |
3072d30e | 484 | only do this for rtx bases as this will be used to assign |
6dfdc153 | 485 | positions in the bitmaps for the global problem. Bit N is set in |
3072d30e | 486 | store1 on the first store for offset N. Bit N is set in store2 |
487 | for the second store to offset N. This is all we need since we | |
488 | only care about offsets that have two or more stores for them. | |
489 | ||
490 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
491 | for 0 and greater offsets. | |
492 | ||
493 | There is one special case here, for stores into the stack frame, | |
494 | we will or store1 into store2 before deciding which stores look | |
495 | at globally. This is because stores to the stack frame that have | |
496 | no other reads before the end of the function can also be | |
497 | deleted. */ | |
498 | bitmap store1_n, store1_p, store2_n, store2_p; | |
499 | ||
6dfdc153 | 500 | /* The positions in this bitmap have the same assignments as the in, |
3072d30e | 501 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6dfdc153 | 502 | the positions that are occupied by stores for this group. */ |
3072d30e | 503 | bitmap group_kill; |
504 | ||
3072d30e | 505 | /* The offset_map is used to map the offsets from this base into |
6dfdc153 | 506 | positions in the global bitmaps. It is only created after all of |
3072d30e | 507 | the all of stores have been scanned and we know which ones we |
508 | care about. */ | |
48e1416a | 509 | int *offset_map_n, *offset_map_p; |
510 | int offset_map_size_n, offset_map_size_p; | |
3072d30e | 511 | }; |
512 | typedef struct group_info *group_info_t; | |
c1fdef8e | 513 | typedef const struct group_info *const_group_info_t; |
3072d30e | 514 | static alloc_pool rtx_group_info_pool; |
515 | ||
516 | /* Tables of group_info structures, hashed by base value. */ | |
517 | static htab_t rtx_group_table; | |
518 | ||
519 | /* Index into the rtx_group_vec. */ | |
520 | static int rtx_group_next_id; | |
521 | ||
522 | DEF_VEC_P(group_info_t); | |
523 | DEF_VEC_ALLOC_P(group_info_t,heap); | |
524 | ||
525 | static VEC(group_info_t,heap) *rtx_group_vec; | |
526 | ||
527 | ||
528 | /* This structure holds the set of changes that are being deferred | |
529 | when removing read operation. See replace_read. */ | |
48e1416a | 530 | struct deferred_change |
3072d30e | 531 | { |
532 | ||
533 | /* The mem that is being replaced. */ | |
534 | rtx *loc; | |
535 | ||
536 | /* The reg it is being replaced with. */ | |
537 | rtx reg; | |
538 | ||
539 | struct deferred_change *next; | |
540 | }; | |
541 | ||
542 | typedef struct deferred_change *deferred_change_t; | |
543 | static alloc_pool deferred_change_pool; | |
544 | ||
545 | static deferred_change_t deferred_change_list = NULL; | |
546 | ||
547 | /* This are used to hold the alias sets of spill variables. Since | |
548 | these are never aliased and there may be a lot of them, it makes | |
549 | sense to treat them specially. This bitvector is only allocated in | |
550 | calls from dse_record_singleton_alias_set which currently is only | |
551 | made during reload1. So when dse is called before reload this | |
552 | mechanism does nothing. */ | |
553 | ||
554 | static bitmap clear_alias_sets = NULL; | |
555 | ||
556 | /* The set of clear_alias_sets that have been disqualified because | |
557 | there are loads or stores using a different mode than the alias set | |
48e1416a | 558 | was registered with. */ |
3072d30e | 559 | static bitmap disqualified_clear_alias_sets = NULL; |
560 | ||
561 | /* The group that holds all of the clear_alias_sets. */ | |
562 | static group_info_t clear_alias_group; | |
563 | ||
564 | /* The modes of the clear_alias_sets. */ | |
565 | static htab_t clear_alias_mode_table; | |
566 | ||
567 | /* Hash table element to look up the mode for an alias set. */ | |
568 | struct clear_alias_mode_holder | |
569 | { | |
32c2fdea | 570 | alias_set_type alias_set; |
3072d30e | 571 | enum machine_mode mode; |
572 | }; | |
573 | ||
574 | static alloc_pool clear_alias_mode_pool; | |
575 | ||
18d50ae6 | 576 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
ff3ae375 | 577 | this for vararg functions because they play games with the frame. */ |
3072d30e | 578 | static bool stores_off_frame_dead_at_return; |
579 | ||
580 | /* Counter for stats. */ | |
48e1416a | 581 | static int globally_deleted; |
582 | static int locally_deleted; | |
583 | static int spill_deleted; | |
584 | ||
3072d30e | 585 | static bitmap all_blocks; |
586 | ||
587 | /* The number of bits used in the global bitmaps. */ | |
588 | static unsigned int current_position; | |
589 | ||
590 | ||
591 | static bool gate_dse (void); | |
4ff06051 | 592 | static bool gate_dse1 (void); |
593 | static bool gate_dse2 (void); | |
3072d30e | 594 | |
595 | \f | |
596 | /*---------------------------------------------------------------------------- | |
597 | Zeroth step. | |
598 | ||
48e1416a | 599 | Initialization. |
3072d30e | 600 | ----------------------------------------------------------------------------*/ |
601 | ||
602 | /* Hashtable callbacks for maintaining the "bases" field of | |
603 | store_group_info, given that the addresses are function invariants. */ | |
604 | ||
605 | static int | |
606 | clear_alias_mode_eq (const void *p1, const void *p2) | |
607 | { | |
48e1416a | 608 | const struct clear_alias_mode_holder * h1 |
3072d30e | 609 | = (const struct clear_alias_mode_holder *) p1; |
48e1416a | 610 | const struct clear_alias_mode_holder * h2 |
3072d30e | 611 | = (const struct clear_alias_mode_holder *) p2; |
612 | return h1->alias_set == h2->alias_set; | |
613 | } | |
614 | ||
615 | ||
616 | static hashval_t | |
617 | clear_alias_mode_hash (const void *p) | |
618 | { | |
48e1416a | 619 | const struct clear_alias_mode_holder *holder |
3072d30e | 620 | = (const struct clear_alias_mode_holder *) p; |
621 | return holder->alias_set; | |
622 | } | |
623 | ||
624 | ||
625 | /* Find the entry associated with ALIAS_SET. */ | |
626 | ||
627 | static struct clear_alias_mode_holder * | |
32c2fdea | 628 | clear_alias_set_lookup (alias_set_type alias_set) |
3072d30e | 629 | { |
630 | struct clear_alias_mode_holder tmp_holder; | |
631 | void **slot; | |
48e1416a | 632 | |
3072d30e | 633 | tmp_holder.alias_set = alias_set; |
634 | slot = htab_find_slot (clear_alias_mode_table, &tmp_holder, NO_INSERT); | |
635 | gcc_assert (*slot); | |
48e1416a | 636 | |
364c0c59 | 637 | return (struct clear_alias_mode_holder *) *slot; |
3072d30e | 638 | } |
639 | ||
640 | ||
641 | /* Hashtable callbacks for maintaining the "bases" field of | |
642 | store_group_info, given that the addresses are function invariants. */ | |
643 | ||
644 | static int | |
645 | invariant_group_base_eq (const void *p1, const void *p2) | |
646 | { | |
c1fdef8e | 647 | const_group_info_t gi1 = (const_group_info_t) p1; |
648 | const_group_info_t gi2 = (const_group_info_t) p2; | |
3072d30e | 649 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
650 | } | |
651 | ||
652 | ||
653 | static hashval_t | |
654 | invariant_group_base_hash (const void *p) | |
655 | { | |
c1fdef8e | 656 | const_group_info_t gi = (const_group_info_t) p; |
3072d30e | 657 | int do_not_record; |
658 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
659 | } | |
660 | ||
661 | ||
662 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
663 | ||
664 | static group_info_t | |
665 | get_group_info (rtx base) | |
666 | { | |
48e1416a | 667 | struct group_info tmp_gi; |
668 | group_info_t gi; | |
3072d30e | 669 | void **slot; |
670 | ||
671 | if (base) | |
672 | { | |
673 | /* Find the store_base_info structure for BASE, creating a new one | |
674 | if necessary. */ | |
675 | tmp_gi.rtx_base = base; | |
676 | slot = htab_find_slot (rtx_group_table, &tmp_gi, INSERT); | |
677 | gi = (group_info_t) *slot; | |
678 | } | |
679 | else | |
680 | { | |
681 | if (!clear_alias_group) | |
682 | { | |
364c0c59 | 683 | clear_alias_group = gi = |
684 | (group_info_t) pool_alloc (rtx_group_info_pool); | |
3072d30e | 685 | memset (gi, 0, sizeof (struct group_info)); |
686 | gi->id = rtx_group_next_id++; | |
687 | gi->store1_n = BITMAP_ALLOC (NULL); | |
688 | gi->store1_p = BITMAP_ALLOC (NULL); | |
689 | gi->store2_n = BITMAP_ALLOC (NULL); | |
690 | gi->store2_p = BITMAP_ALLOC (NULL); | |
691 | gi->group_kill = BITMAP_ALLOC (NULL); | |
692 | gi->process_globally = false; | |
693 | gi->offset_map_size_n = 0; | |
694 | gi->offset_map_size_p = 0; | |
695 | gi->offset_map_n = NULL; | |
696 | gi->offset_map_p = NULL; | |
697 | VEC_safe_push (group_info_t, heap, rtx_group_vec, gi); | |
698 | } | |
699 | return clear_alias_group; | |
700 | } | |
701 | ||
702 | if (gi == NULL) | |
703 | { | |
364c0c59 | 704 | *slot = gi = (group_info_t) pool_alloc (rtx_group_info_pool); |
3072d30e | 705 | gi->rtx_base = base; |
706 | gi->id = rtx_group_next_id++; | |
707 | gi->base_mem = gen_rtx_MEM (QImode, base); | |
82d2c88b | 708 | gi->canon_base_addr = canon_rtx (base); |
3072d30e | 709 | gi->store1_n = BITMAP_ALLOC (NULL); |
710 | gi->store1_p = BITMAP_ALLOC (NULL); | |
711 | gi->store2_n = BITMAP_ALLOC (NULL); | |
712 | gi->store2_p = BITMAP_ALLOC (NULL); | |
713 | gi->group_kill = BITMAP_ALLOC (NULL); | |
714 | gi->process_globally = false; | |
48e1416a | 715 | gi->frame_related = |
3072d30e | 716 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
717 | gi->offset_map_size_n = 0; | |
718 | gi->offset_map_size_p = 0; | |
719 | gi->offset_map_n = NULL; | |
720 | gi->offset_map_p = NULL; | |
721 | VEC_safe_push (group_info_t, heap, rtx_group_vec, gi); | |
722 | } | |
723 | ||
724 | return gi; | |
725 | } | |
726 | ||
727 | ||
728 | /* Initialization of data structures. */ | |
729 | ||
730 | static void | |
731 | dse_step0 (void) | |
732 | { | |
733 | locally_deleted = 0; | |
734 | globally_deleted = 0; | |
735 | spill_deleted = 0; | |
736 | ||
737 | scratch = BITMAP_ALLOC (NULL); | |
738 | ||
739 | rtx_store_info_pool | |
48e1416a | 740 | = create_alloc_pool ("rtx_store_info_pool", |
3072d30e | 741 | sizeof (struct store_info), 100); |
742 | read_info_pool | |
48e1416a | 743 | = create_alloc_pool ("read_info_pool", |
3072d30e | 744 | sizeof (struct read_info), 100); |
745 | insn_info_pool | |
48e1416a | 746 | = create_alloc_pool ("insn_info_pool", |
3072d30e | 747 | sizeof (struct insn_info), 100); |
748 | bb_info_pool | |
48e1416a | 749 | = create_alloc_pool ("bb_info_pool", |
3072d30e | 750 | sizeof (struct bb_info), 100); |
751 | rtx_group_info_pool | |
48e1416a | 752 | = create_alloc_pool ("rtx_group_info_pool", |
3072d30e | 753 | sizeof (struct group_info), 100); |
754 | deferred_change_pool | |
48e1416a | 755 | = create_alloc_pool ("deferred_change_pool", |
3072d30e | 756 | sizeof (struct deferred_change), 10); |
757 | ||
758 | rtx_group_table = htab_create (11, invariant_group_base_hash, | |
759 | invariant_group_base_eq, NULL); | |
760 | ||
761 | bb_table = XCNEWVEC (bb_info_t, last_basic_block); | |
762 | rtx_group_next_id = 0; | |
763 | ||
18d50ae6 | 764 | stores_off_frame_dead_at_return = !cfun->stdarg; |
3072d30e | 765 | |
766 | init_alias_analysis (); | |
48e1416a | 767 | |
3072d30e | 768 | if (clear_alias_sets) |
769 | clear_alias_group = get_group_info (NULL); | |
770 | else | |
771 | clear_alias_group = NULL; | |
772 | } | |
773 | ||
774 | ||
775 | \f | |
776 | /*---------------------------------------------------------------------------- | |
777 | First step. | |
778 | ||
779 | Scan all of the insns. Any random ordering of the blocks is fine. | |
bef304b8 | 780 | Each block is scanned in forward order to accommodate cselib which |
3072d30e | 781 | is used to remove stores with non-constant bases. |
782 | ----------------------------------------------------------------------------*/ | |
783 | ||
784 | /* Delete all of the store_info recs from INSN_INFO. */ | |
785 | ||
48e1416a | 786 | static void |
3072d30e | 787 | free_store_info (insn_info_t insn_info) |
788 | { | |
789 | store_info_t store_info = insn_info->store_rec; | |
790 | while (store_info) | |
791 | { | |
792 | store_info_t next = store_info->next; | |
aa140b76 | 793 | if (store_info->is_large) |
843bd2fa | 794 | BITMAP_FREE (store_info->positions_needed.large.bmap); |
3072d30e | 795 | if (store_info->cse_base) |
796 | pool_free (cse_store_info_pool, store_info); | |
797 | else | |
798 | pool_free (rtx_store_info_pool, store_info); | |
799 | store_info = next; | |
800 | } | |
801 | ||
802 | insn_info->cannot_delete = true; | |
803 | insn_info->contains_cselib_groups = false; | |
804 | insn_info->store_rec = NULL; | |
805 | } | |
806 | ||
807 | ||
808 | struct insn_size { | |
809 | int size; | |
810 | rtx insn; | |
811 | }; | |
812 | ||
813 | ||
814 | /* Add an insn to do the add inside a x if it is a | |
815 | PRE/POST-INC/DEC/MODIFY. D is an structure containing the insn and | |
816 | the size of the mode of the MEM that this is inside of. */ | |
817 | ||
818 | static int | |
819 | replace_inc_dec (rtx *r, void *d) | |
820 | { | |
821 | rtx x = *r; | |
822 | struct insn_size *data = (struct insn_size *)d; | |
823 | switch (GET_CODE (x)) | |
824 | { | |
825 | case PRE_INC: | |
826 | case POST_INC: | |
827 | { | |
828 | rtx r1 = XEXP (x, 0); | |
98155838 | 829 | rtx c = gen_int_mode (data->size, GET_MODE (r1)); |
830 | emit_insn_before (gen_rtx_SET (VOIDmode, r1, | |
831 | gen_rtx_PLUS (GET_MODE (r1), r1, c)), | |
14eddddd | 832 | data->insn); |
3072d30e | 833 | return -1; |
834 | } | |
48e1416a | 835 | |
3072d30e | 836 | case PRE_DEC: |
837 | case POST_DEC: | |
838 | { | |
839 | rtx r1 = XEXP (x, 0); | |
98155838 | 840 | rtx c = gen_int_mode (-data->size, GET_MODE (r1)); |
841 | emit_insn_before (gen_rtx_SET (VOIDmode, r1, | |
842 | gen_rtx_PLUS (GET_MODE (r1), r1, c)), | |
14eddddd | 843 | data->insn); |
3072d30e | 844 | return -1; |
845 | } | |
48e1416a | 846 | |
3072d30e | 847 | case PRE_MODIFY: |
848 | case POST_MODIFY: | |
849 | { | |
bef304b8 | 850 | /* We can reuse the add because we are about to delete the |
3072d30e | 851 | insn that contained it. */ |
852 | rtx add = XEXP (x, 0); | |
853 | rtx r1 = XEXP (add, 0); | |
98155838 | 854 | emit_insn_before (gen_rtx_SET (VOIDmode, r1, add), data->insn); |
3072d30e | 855 | return -1; |
856 | } | |
857 | ||
858 | default: | |
859 | return 0; | |
860 | } | |
861 | } | |
48e1416a | 862 | |
3072d30e | 863 | |
864 | /* If X is a MEM, check the address to see if it is PRE/POST-INC/DEC/MODIFY | |
865 | and generate an add to replace that. */ | |
866 | ||
867 | static int | |
868 | replace_inc_dec_mem (rtx *r, void *d) | |
869 | { | |
870 | rtx x = *r; | |
14eddddd | 871 | if (x != NULL_RTX && MEM_P (x)) |
3072d30e | 872 | { |
873 | struct insn_size data; | |
874 | ||
875 | data.size = GET_MODE_SIZE (GET_MODE (x)); | |
14eddddd | 876 | data.insn = (rtx) d; |
3072d30e | 877 | |
878 | for_each_rtx (&XEXP (x, 0), replace_inc_dec, &data); | |
48e1416a | 879 | |
3072d30e | 880 | return -1; |
881 | } | |
882 | return 0; | |
883 | } | |
884 | ||
885 | /* Before we delete INSN, make sure that the auto inc/dec, if it is | |
886 | there, is split into a separate insn. */ | |
887 | ||
888 | static void | |
889 | check_for_inc_dec (rtx insn) | |
890 | { | |
891 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); | |
892 | if (note) | |
893 | for_each_rtx (&insn, replace_inc_dec_mem, insn); | |
894 | } | |
895 | ||
896 | ||
48e1416a | 897 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
3072d30e | 898 | |
899 | static void | |
900 | delete_dead_store_insn (insn_info_t insn_info) | |
901 | { | |
902 | read_info_t read_info; | |
903 | ||
904 | if (!dbg_cnt (dse)) | |
905 | return; | |
906 | ||
907 | check_for_inc_dec (insn_info->insn); | |
908 | if (dump_file) | |
909 | { | |
48e1416a | 910 | fprintf (dump_file, "Locally deleting insn %d ", |
3072d30e | 911 | INSN_UID (insn_info->insn)); |
912 | if (insn_info->store_rec->alias_set) | |
48e1416a | 913 | fprintf (dump_file, "alias set %d\n", |
32c2fdea | 914 | (int) insn_info->store_rec->alias_set); |
3072d30e | 915 | else |
916 | fprintf (dump_file, "\n"); | |
917 | } | |
918 | ||
919 | free_store_info (insn_info); | |
920 | read_info = insn_info->read_rec; | |
48e1416a | 921 | |
3072d30e | 922 | while (read_info) |
923 | { | |
924 | read_info_t next = read_info->next; | |
925 | pool_free (read_info_pool, read_info); | |
926 | read_info = next; | |
927 | } | |
928 | insn_info->read_rec = NULL; | |
929 | ||
930 | delete_insn (insn_info->insn); | |
931 | locally_deleted++; | |
932 | insn_info->insn = NULL; | |
933 | ||
934 | insn_info->wild_read = false; | |
935 | } | |
936 | ||
937 | ||
938 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
939 | OFFSET and WIDTH. */ | |
940 | ||
941 | static void | |
942 | set_usage_bits (group_info_t group, HOST_WIDE_INT offset, HOST_WIDE_INT width) | |
943 | { | |
944 | HOST_WIDE_INT i; | |
945 | ||
aa140b76 | 946 | if (offset > -MAX_OFFSET && offset + width < MAX_OFFSET) |
3072d30e | 947 | for (i=offset; i<offset+width; i++) |
948 | { | |
949 | bitmap store1; | |
950 | bitmap store2; | |
951 | int ai; | |
952 | if (i < 0) | |
953 | { | |
954 | store1 = group->store1_n; | |
955 | store2 = group->store2_n; | |
956 | ai = -i; | |
957 | } | |
958 | else | |
959 | { | |
960 | store1 = group->store1_p; | |
961 | store2 = group->store2_p; | |
962 | ai = i; | |
963 | } | |
48e1416a | 964 | |
3072d30e | 965 | if (bitmap_bit_p (store1, ai)) |
966 | bitmap_set_bit (store2, ai); | |
48e1416a | 967 | else |
3072d30e | 968 | { |
969 | bitmap_set_bit (store1, ai); | |
970 | if (i < 0) | |
971 | { | |
972 | if (group->offset_map_size_n < ai) | |
973 | group->offset_map_size_n = ai; | |
974 | } | |
975 | else | |
976 | { | |
977 | if (group->offset_map_size_p < ai) | |
978 | group->offset_map_size_p = ai; | |
979 | } | |
980 | } | |
981 | } | |
982 | } | |
983 | ||
984 | ||
985 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
986 | ||
987 | static void | |
988 | add_wild_read (bb_info_t bb_info) | |
989 | { | |
990 | insn_info_t insn_info = bb_info->last_insn; | |
991 | read_info_t *ptr = &insn_info->read_rec; | |
992 | ||
993 | while (*ptr) | |
994 | { | |
995 | read_info_t next = (*ptr)->next; | |
32c2fdea | 996 | if ((*ptr)->alias_set == 0) |
3072d30e | 997 | { |
998 | pool_free (read_info_pool, *ptr); | |
999 | *ptr = next; | |
1000 | } | |
48e1416a | 1001 | else |
3072d30e | 1002 | ptr = &(*ptr)->next; |
1003 | } | |
1004 | insn_info->wild_read = true; | |
1005 | active_local_stores = NULL; | |
1006 | } | |
1007 | ||
1008 | ||
17e1318c | 1009 | /* Return true if X is a constant or one of the registers that behave |
1010 | as a constant over the life of a function. This is equivalent to | |
1011 | !rtx_varies_p for memory addresses. */ | |
3072d30e | 1012 | |
1013 | static bool | |
1014 | const_or_frame_p (rtx x) | |
1015 | { | |
1016 | switch (GET_CODE (x)) | |
1017 | { | |
1018 | case MEM: | |
1019 | return MEM_READONLY_P (x); | |
1020 | ||
1021 | case CONST: | |
1022 | case CONST_INT: | |
1023 | case CONST_DOUBLE: | |
1024 | case CONST_VECTOR: | |
1025 | case SYMBOL_REF: | |
1026 | case LABEL_REF: | |
1027 | return true; | |
1028 | ||
1029 | case REG: | |
1030 | /* Note that we have to test for the actual rtx used for the frame | |
1031 | and arg pointers and not just the register number in case we have | |
1032 | eliminated the frame and/or arg pointer and are using it | |
1033 | for pseudos. */ | |
1034 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1035 | /* The arg pointer varies if it is not a fixed register. */ | |
1036 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1037 | || x == pic_offset_table_rtx) | |
1038 | return true; | |
1039 | return false; | |
1040 | ||
1041 | default: | |
1042 | return false; | |
1043 | } | |
1044 | } | |
1045 | ||
48e1416a | 1046 | /* Take all reasonable action to put the address of MEM into the form |
1047 | that we can do analysis on. | |
3072d30e | 1048 | |
1049 | The gold standard is to get the address into the form: address + | |
1050 | OFFSET where address is something that rtx_varies_p considers a | |
1051 | constant. When we can get the address in this form, we can do | |
1052 | global analysis on it. Note that for constant bases, address is | |
1053 | not actually returned, only the group_id. The address can be | |
1054 | obtained from that. | |
1055 | ||
1056 | If that fails, we try cselib to get a value we can at least use | |
48e1416a | 1057 | locally. If that fails we return false. |
1058 | ||
3072d30e | 1059 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1060 | group for non_varying bases. | |
1061 | ||
1062 | FOR_READ is true if this is a mem read and false if not. */ | |
1063 | ||
1064 | static bool | |
1065 | canon_address (rtx mem, | |
32c2fdea | 1066 | alias_set_type *alias_set_out, |
3072d30e | 1067 | int *group_id, |
48e1416a | 1068 | HOST_WIDE_INT *offset, |
3072d30e | 1069 | cselib_val **base) |
1070 | { | |
98155838 | 1071 | enum machine_mode address_mode |
1072 | = targetm.addr_space.address_mode (MEM_ADDR_SPACE (mem)); | |
3072d30e | 1073 | rtx mem_address = XEXP (mem, 0); |
1074 | rtx expanded_address, address; | |
627540ce | 1075 | int expanded; |
1076 | ||
3072d30e | 1077 | /* Make sure that cselib is has initialized all of the operands of |
1078 | the address before asking it to do the subst. */ | |
1079 | ||
1080 | if (clear_alias_sets) | |
1081 | { | |
1082 | /* If this is a spill, do not do any further processing. */ | |
32c2fdea | 1083 | alias_set_type alias_set = MEM_ALIAS_SET (mem); |
3072d30e | 1084 | if (dump_file) |
32c2fdea | 1085 | fprintf (dump_file, "found alias set %d\n", (int) alias_set); |
3072d30e | 1086 | if (bitmap_bit_p (clear_alias_sets, alias_set)) |
1087 | { | |
48e1416a | 1088 | struct clear_alias_mode_holder *entry |
3072d30e | 1089 | = clear_alias_set_lookup (alias_set); |
1090 | ||
1091 | /* If the modes do not match, we cannot process this set. */ | |
1092 | if (entry->mode != GET_MODE (mem)) | |
1093 | { | |
1094 | if (dump_file) | |
48e1416a | 1095 | fprintf (dump_file, |
1096 | "disqualifying alias set %d, (%s) != (%s)\n", | |
1097 | (int) alias_set, GET_MODE_NAME (entry->mode), | |
3072d30e | 1098 | GET_MODE_NAME (GET_MODE (mem))); |
48e1416a | 1099 | |
3072d30e | 1100 | bitmap_set_bit (disqualified_clear_alias_sets, alias_set); |
1101 | return false; | |
1102 | } | |
1103 | ||
1104 | *alias_set_out = alias_set; | |
1105 | *group_id = clear_alias_group->id; | |
1106 | return true; | |
1107 | } | |
1108 | } | |
1109 | ||
1110 | *alias_set_out = 0; | |
1111 | ||
98155838 | 1112 | cselib_lookup (mem_address, address_mode, 1); |
3072d30e | 1113 | |
1114 | if (dump_file) | |
1115 | { | |
1116 | fprintf (dump_file, " mem: "); | |
1117 | print_inline_rtx (dump_file, mem_address, 0); | |
1118 | fprintf (dump_file, "\n"); | |
1119 | } | |
1120 | ||
627540ce | 1121 | /* First see if just canon_rtx (mem_address) is const or frame, |
1122 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1123 | address = NULL_RTX; | |
1124 | for (expanded = 0; expanded < 2; expanded++) | |
1125 | { | |
1126 | if (expanded) | |
1127 | { | |
1128 | /* Use cselib to replace all of the reg references with the full | |
48e1416a | 1129 | expression. This will take care of the case where we have |
3072d30e | 1130 | |
627540ce | 1131 | r_x = base + offset; |
1132 | val = *r_x; | |
48e1416a | 1133 | |
1134 | by making it into | |
3072d30e | 1135 | |
627540ce | 1136 | val = *(base + offset); */ |
3072d30e | 1137 | |
627540ce | 1138 | expanded_address = cselib_expand_value_rtx (mem_address, |
1139 | scratch, 5); | |
3072d30e | 1140 | |
627540ce | 1141 | /* If this fails, just go with the address from first |
1142 | iteration. */ | |
1143 | if (!expanded_address) | |
1144 | break; | |
1145 | } | |
1146 | else | |
1147 | expanded_address = mem_address; | |
3072d30e | 1148 | |
627540ce | 1149 | /* Split the address into canonical BASE + OFFSET terms. */ |
1150 | address = canon_rtx (expanded_address); | |
3072d30e | 1151 | |
627540ce | 1152 | *offset = 0; |
3072d30e | 1153 | |
627540ce | 1154 | if (dump_file) |
1155 | { | |
1156 | if (expanded) | |
1157 | { | |
1158 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1159 | print_inline_rtx (dump_file, expanded_address, 0); | |
1160 | fprintf (dump_file, "\n"); | |
1161 | } | |
3072d30e | 1162 | |
627540ce | 1163 | fprintf (dump_file, "\n after canon_rtx address: "); |
1164 | print_inline_rtx (dump_file, address, 0); | |
1165 | fprintf (dump_file, "\n"); | |
1166 | } | |
3072d30e | 1167 | |
627540ce | 1168 | if (GET_CODE (address) == CONST) |
1169 | address = XEXP (address, 0); | |
3072d30e | 1170 | |
627540ce | 1171 | if (GET_CODE (address) == PLUS |
1172 | && CONST_INT_P (XEXP (address, 1))) | |
1173 | { | |
1174 | *offset = INTVAL (XEXP (address, 1)); | |
1175 | address = XEXP (address, 0); | |
1176 | } | |
3072d30e | 1177 | |
bd1a81f7 | 1178 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1179 | && const_or_frame_p (address)) | |
3072d30e | 1180 | { |
627540ce | 1181 | group_info_t group = get_group_info (address); |
1182 | ||
3072d30e | 1183 | if (dump_file) |
627540ce | 1184 | fprintf (dump_file, " gid=%d offset=%d \n", |
1185 | group->id, (int)*offset); | |
1186 | *base = NULL; | |
1187 | *group_id = group->id; | |
1188 | return true; | |
3072d30e | 1189 | } |
627540ce | 1190 | } |
1191 | ||
98155838 | 1192 | *base = cselib_lookup (address, address_mode, true); |
627540ce | 1193 | *group_id = -1; |
1194 | ||
1195 | if (*base == NULL) | |
1196 | { | |
3072d30e | 1197 | if (dump_file) |
627540ce | 1198 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1199 | return false; | |
3072d30e | 1200 | } |
627540ce | 1201 | if (dump_file) |
48e1416a | 1202 | fprintf (dump_file, " varying cselib base=%d offset = %d\n", |
627540ce | 1203 | (*base)->value, (int)*offset); |
3072d30e | 1204 | return true; |
1205 | } | |
1206 | ||
1207 | ||
1208 | /* Clear the rhs field from the active_local_stores array. */ | |
1209 | ||
1210 | static void | |
1211 | clear_rhs_from_active_local_stores (void) | |
1212 | { | |
1213 | insn_info_t ptr = active_local_stores; | |
1214 | ||
1215 | while (ptr) | |
1216 | { | |
1217 | store_info_t store_info = ptr->store_rec; | |
1218 | /* Skip the clobbers. */ | |
1219 | while (!store_info->is_set) | |
1220 | store_info = store_info->next; | |
1221 | ||
1222 | store_info->rhs = NULL; | |
aa140b76 | 1223 | store_info->const_rhs = NULL; |
3072d30e | 1224 | |
1225 | ptr = ptr->next_local_store; | |
1226 | } | |
1227 | } | |
1228 | ||
1229 | ||
aa140b76 | 1230 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1231 | ||
1232 | static inline void | |
1233 | set_position_unneeded (store_info_t s_info, int pos) | |
1234 | { | |
1235 | if (__builtin_expect (s_info->is_large, false)) | |
1236 | { | |
843bd2fa | 1237 | if (!bitmap_bit_p (s_info->positions_needed.large.bmap, pos)) |
aa140b76 | 1238 | { |
1239 | s_info->positions_needed.large.count++; | |
843bd2fa | 1240 | bitmap_set_bit (s_info->positions_needed.large.bmap, pos); |
aa140b76 | 1241 | } |
1242 | } | |
1243 | else | |
1244 | s_info->positions_needed.small_bitmask | |
1245 | &= ~(((unsigned HOST_WIDE_INT) 1) << pos); | |
1246 | } | |
1247 | ||
1248 | /* Mark the whole store S_INFO as unneeded. */ | |
1249 | ||
1250 | static inline void | |
1251 | set_all_positions_unneeded (store_info_t s_info) | |
1252 | { | |
1253 | if (__builtin_expect (s_info->is_large, false)) | |
1254 | { | |
1255 | int pos, end = s_info->end - s_info->begin; | |
1256 | for (pos = 0; pos < end; pos++) | |
843bd2fa | 1257 | bitmap_set_bit (s_info->positions_needed.large.bmap, pos); |
aa140b76 | 1258 | s_info->positions_needed.large.count = end; |
1259 | } | |
1260 | else | |
1261 | s_info->positions_needed.small_bitmask = (unsigned HOST_WIDE_INT) 0; | |
1262 | } | |
1263 | ||
1264 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1265 | ||
1266 | static inline bool | |
1267 | any_positions_needed_p (store_info_t s_info) | |
1268 | { | |
1269 | if (__builtin_expect (s_info->is_large, false)) | |
1270 | return (s_info->positions_needed.large.count | |
1271 | < s_info->end - s_info->begin); | |
1272 | else | |
1273 | return (s_info->positions_needed.small_bitmask | |
1274 | != (unsigned HOST_WIDE_INT) 0); | |
1275 | } | |
1276 | ||
1277 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
1278 | store are needed. */ | |
1279 | ||
1280 | static inline bool | |
1281 | all_positions_needed_p (store_info_t s_info, int start, int width) | |
1282 | { | |
1283 | if (__builtin_expect (s_info->is_large, false)) | |
1284 | { | |
1285 | int end = start + width; | |
1286 | while (start < end) | |
843bd2fa | 1287 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, start++)) |
aa140b76 | 1288 | return false; |
1289 | return true; | |
1290 | } | |
1291 | else | |
1292 | { | |
1293 | unsigned HOST_WIDE_INT mask = lowpart_bitmask (width) << start; | |
1294 | return (s_info->positions_needed.small_bitmask & mask) == mask; | |
1295 | } | |
1296 | } | |
1297 | ||
1298 | ||
1299 | static rtx get_stored_val (store_info_t, enum machine_mode, HOST_WIDE_INT, | |
1300 | HOST_WIDE_INT, basic_block, bool); | |
1301 | ||
1302 | ||
3072d30e | 1303 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1304 | there is a candidate store, after adding it to the appropriate | |
1305 | local store group if so. */ | |
1306 | ||
1307 | static int | |
1308 | record_store (rtx body, bb_info_t bb_info) | |
1309 | { | |
82d2c88b | 1310 | rtx mem, rhs, const_rhs, mem_addr; |
3072d30e | 1311 | HOST_WIDE_INT offset = 0; |
1312 | HOST_WIDE_INT width = 0; | |
32c2fdea | 1313 | alias_set_type spill_alias_set; |
3072d30e | 1314 | insn_info_t insn_info = bb_info->last_insn; |
1315 | store_info_t store_info = NULL; | |
1316 | int group_id; | |
1317 | cselib_val *base = NULL; | |
aa140b76 | 1318 | insn_info_t ptr, last, redundant_reason; |
3072d30e | 1319 | bool store_is_unused; |
1320 | ||
1321 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1322 | return 0; | |
1323 | ||
aa140b76 | 1324 | mem = SET_DEST (body); |
1325 | ||
3072d30e | 1326 | /* If this is not used, then this cannot be used to keep the insn |
1327 | from being deleted. On the other hand, it does provide something | |
1328 | that can be used to prove that another store is dead. */ | |
1329 | store_is_unused | |
aa140b76 | 1330 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
3072d30e | 1331 | |
1332 | /* Check whether that value is a suitable memory location. */ | |
3072d30e | 1333 | if (!MEM_P (mem)) |
1334 | { | |
1335 | /* If the set or clobber is unused, then it does not effect our | |
1336 | ability to get rid of the entire insn. */ | |
1337 | if (!store_is_unused) | |
1338 | insn_info->cannot_delete = true; | |
1339 | return 0; | |
1340 | } | |
1341 | ||
1342 | /* At this point we know mem is a mem. */ | |
1343 | if (GET_MODE (mem) == BLKmode) | |
1344 | { | |
1345 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
1346 | { | |
48e1416a | 1347 | if (dump_file) |
3072d30e | 1348 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1349 | add_wild_read (bb_info); | |
1350 | insn_info->cannot_delete = true; | |
aa140b76 | 1351 | return 0; |
3072d30e | 1352 | } |
aa140b76 | 1353 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1354 | as memset (addr, 0, 36); */ | |
1355 | else if (!MEM_SIZE (mem) | |
1356 | || !CONST_INT_P (MEM_SIZE (mem)) | |
1357 | || GET_CODE (body) != SET | |
1358 | || INTVAL (MEM_SIZE (mem)) <= 0 | |
1359 | || INTVAL (MEM_SIZE (mem)) > MAX_OFFSET | |
1360 | || !CONST_INT_P (SET_SRC (body))) | |
3072d30e | 1361 | { |
aa140b76 | 1362 | if (!store_is_unused) |
1363 | { | |
1364 | /* If the set or clobber is unused, then it does not effect our | |
1365 | ability to get rid of the entire insn. */ | |
1366 | insn_info->cannot_delete = true; | |
1367 | clear_rhs_from_active_local_stores (); | |
1368 | } | |
1369 | return 0; | |
3072d30e | 1370 | } |
3072d30e | 1371 | } |
1372 | ||
1373 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1374 | if (MEM_VOLATILE_P (mem)) | |
aa140b76 | 1375 | insn_info->cannot_delete = true; |
3072d30e | 1376 | |
1377 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
1378 | { | |
1379 | clear_rhs_from_active_local_stores (); | |
1380 | return 0; | |
1381 | } | |
1382 | ||
aa140b76 | 1383 | if (GET_MODE (mem) == BLKmode) |
1384 | width = INTVAL (MEM_SIZE (mem)); | |
1385 | else | |
1386 | { | |
1387 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
1388 | gcc_assert ((unsigned) width <= HOST_BITS_PER_WIDE_INT); | |
1389 | } | |
3072d30e | 1390 | |
1391 | if (spill_alias_set) | |
1392 | { | |
1393 | bitmap store1 = clear_alias_group->store1_p; | |
1394 | bitmap store2 = clear_alias_group->store2_p; | |
aa140b76 | 1395 | |
1396 | gcc_assert (GET_MODE (mem) != BLKmode); | |
48e1416a | 1397 | |
3072d30e | 1398 | if (bitmap_bit_p (store1, spill_alias_set)) |
1399 | bitmap_set_bit (store2, spill_alias_set); | |
48e1416a | 1400 | else |
3072d30e | 1401 | bitmap_set_bit (store1, spill_alias_set); |
48e1416a | 1402 | |
3072d30e | 1403 | if (clear_alias_group->offset_map_size_p < spill_alias_set) |
1404 | clear_alias_group->offset_map_size_p = spill_alias_set; | |
48e1416a | 1405 | |
364c0c59 | 1406 | store_info = (store_info_t) pool_alloc (rtx_store_info_pool); |
3072d30e | 1407 | |
1408 | if (dump_file) | |
1409 | fprintf (dump_file, " processing spill store %d(%s)\n", | |
32c2fdea | 1410 | (int) spill_alias_set, GET_MODE_NAME (GET_MODE (mem))); |
3072d30e | 1411 | } |
1412 | else if (group_id >= 0) | |
1413 | { | |
1414 | /* In the restrictive case where the base is a constant or the | |
1415 | frame pointer we can do global analysis. */ | |
48e1416a | 1416 | |
1417 | group_info_t group | |
3072d30e | 1418 | = VEC_index (group_info_t, rtx_group_vec, group_id); |
48e1416a | 1419 | |
364c0c59 | 1420 | store_info = (store_info_t) pool_alloc (rtx_store_info_pool); |
3072d30e | 1421 | set_usage_bits (group, offset, width); |
1422 | ||
1423 | if (dump_file) | |
1424 | fprintf (dump_file, " processing const base store gid=%d[%d..%d)\n", | |
1425 | group_id, (int)offset, (int)(offset+width)); | |
1426 | } | |
1427 | else | |
1428 | { | |
17e1318c | 1429 | rtx base_term = find_base_term (XEXP (mem, 0)); |
1430 | if (!base_term | |
1431 | || (GET_CODE (base_term) == ADDRESS | |
1432 | && GET_MODE (base_term) == Pmode | |
1433 | && XEXP (base_term, 0) == stack_pointer_rtx)) | |
1434 | insn_info->stack_pointer_based = true; | |
3072d30e | 1435 | insn_info->contains_cselib_groups = true; |
17e1318c | 1436 | |
364c0c59 | 1437 | store_info = (store_info_t) pool_alloc (cse_store_info_pool); |
3072d30e | 1438 | group_id = -1; |
1439 | ||
1440 | if (dump_file) | |
1441 | fprintf (dump_file, " processing cselib store [%d..%d)\n", | |
1442 | (int)offset, (int)(offset+width)); | |
1443 | } | |
1444 | ||
aa140b76 | 1445 | const_rhs = rhs = NULL_RTX; |
1446 | if (GET_CODE (body) == SET | |
1447 | /* No place to keep the value after ra. */ | |
1448 | && !reload_completed | |
1449 | && (REG_P (SET_SRC (body)) | |
1450 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1451 | || CONSTANT_P (SET_SRC (body))) | |
1452 | && !MEM_VOLATILE_P (mem) | |
1453 | /* Sometimes the store and reload is used for truncation and | |
1454 | rounding. */ | |
1455 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1456 | { | |
1457 | rhs = SET_SRC (body); | |
1458 | if (CONSTANT_P (rhs)) | |
1459 | const_rhs = rhs; | |
1460 | else if (body == PATTERN (insn_info->insn)) | |
1461 | { | |
1462 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1463 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1464 | const_rhs = XEXP (tem, 0); | |
1465 | } | |
1466 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1467 | { | |
1468 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1469 | ||
1470 | if (tem && CONSTANT_P (tem)) | |
1471 | const_rhs = tem; | |
1472 | } | |
1473 | } | |
1474 | ||
3072d30e | 1475 | /* Check to see if this stores causes some other stores to be |
1476 | dead. */ | |
1477 | ptr = active_local_stores; | |
1478 | last = NULL; | |
aa140b76 | 1479 | redundant_reason = NULL; |
82d2c88b | 1480 | mem = canon_rtx (mem); |
1481 | /* For alias_set != 0 canon_true_dependence should be never called. */ | |
1482 | if (spill_alias_set) | |
1483 | mem_addr = NULL_RTX; | |
1484 | else | |
1485 | { | |
1486 | if (group_id < 0) | |
1487 | mem_addr = base->val_rtx; | |
1488 | else | |
1489 | { | |
1490 | group_info_t group | |
1491 | = VEC_index (group_info_t, rtx_group_vec, group_id); | |
1492 | mem_addr = group->canon_base_addr; | |
1493 | } | |
1494 | if (offset) | |
1495 | mem_addr = plus_constant (mem_addr, offset); | |
1496 | } | |
3072d30e | 1497 | |
1498 | while (ptr) | |
1499 | { | |
1500 | insn_info_t next = ptr->next_local_store; | |
1501 | store_info_t s_info = ptr->store_rec; | |
9ce37fa7 | 1502 | bool del = true; |
3072d30e | 1503 | |
1504 | /* Skip the clobbers. We delete the active insn if this insn | |
6dfdc153 | 1505 | shadows the set. To have been put on the active list, it |
3072d30e | 1506 | has exactly on set. */ |
1507 | while (!s_info->is_set) | |
1508 | s_info = s_info->next; | |
1509 | ||
1510 | if (s_info->alias_set != spill_alias_set) | |
9ce37fa7 | 1511 | del = false; |
3072d30e | 1512 | else if (s_info->alias_set) |
1513 | { | |
48e1416a | 1514 | struct clear_alias_mode_holder *entry |
3072d30e | 1515 | = clear_alias_set_lookup (s_info->alias_set); |
1516 | /* Generally, spills cannot be processed if and of the | |
1517 | references to the slot have a different mode. But if | |
1518 | we are in the same block and mode is exactly the same | |
1519 | between this store and one before in the same block, | |
1520 | we can still delete it. */ | |
1521 | if ((GET_MODE (mem) == GET_MODE (s_info->mem)) | |
1522 | && (GET_MODE (mem) == entry->mode)) | |
1523 | { | |
9ce37fa7 | 1524 | del = true; |
aa140b76 | 1525 | set_all_positions_unneeded (s_info); |
3072d30e | 1526 | } |
1527 | if (dump_file) | |
1528 | fprintf (dump_file, " trying spill store in insn=%d alias_set=%d\n", | |
32c2fdea | 1529 | INSN_UID (ptr->insn), (int) s_info->alias_set); |
3072d30e | 1530 | } |
48e1416a | 1531 | else if ((s_info->group_id == group_id) |
3072d30e | 1532 | && (s_info->cse_base == base)) |
1533 | { | |
1534 | HOST_WIDE_INT i; | |
1535 | if (dump_file) | |
1536 | fprintf (dump_file, " trying store in insn=%d gid=%d[%d..%d)\n", | |
48e1416a | 1537 | INSN_UID (ptr->insn), s_info->group_id, |
3072d30e | 1538 | (int)s_info->begin, (int)s_info->end); |
aa140b76 | 1539 | |
1540 | /* Even if PTR won't be eliminated as unneeded, if both | |
1541 | PTR and this insn store the same constant value, we might | |
1542 | eliminate this insn instead. */ | |
1543 | if (s_info->const_rhs | |
1544 | && const_rhs | |
1545 | && offset >= s_info->begin | |
1546 | && offset + width <= s_info->end | |
1547 | && all_positions_needed_p (s_info, offset - s_info->begin, | |
1548 | width)) | |
1549 | { | |
1550 | if (GET_MODE (mem) == BLKmode) | |
1551 | { | |
1552 | if (GET_MODE (s_info->mem) == BLKmode | |
1553 | && s_info->const_rhs == const_rhs) | |
1554 | redundant_reason = ptr; | |
1555 | } | |
1556 | else if (s_info->const_rhs == const0_rtx | |
1557 | && const_rhs == const0_rtx) | |
1558 | redundant_reason = ptr; | |
1559 | else | |
1560 | { | |
1561 | rtx val; | |
1562 | start_sequence (); | |
1563 | val = get_stored_val (s_info, GET_MODE (mem), | |
1564 | offset, offset + width, | |
1565 | BLOCK_FOR_INSN (insn_info->insn), | |
1566 | true); | |
1567 | if (get_insns () != NULL) | |
1568 | val = NULL_RTX; | |
1569 | end_sequence (); | |
1570 | if (val && rtx_equal_p (val, const_rhs)) | |
1571 | redundant_reason = ptr; | |
1572 | } | |
1573 | } | |
1574 | ||
1575 | for (i = MAX (offset, s_info->begin); | |
1576 | i < offset + width && i < s_info->end; | |
1577 | i++) | |
1578 | set_position_unneeded (s_info, i - s_info->begin); | |
3072d30e | 1579 | } |
1580 | else if (s_info->rhs) | |
1581 | /* Need to see if it is possible for this store to overwrite | |
1582 | the value of store_info. If it is, set the rhs to NULL to | |
1583 | keep it from being used to remove a load. */ | |
1584 | { | |
48e1416a | 1585 | if (canon_true_dependence (s_info->mem, |
3072d30e | 1586 | GET_MODE (s_info->mem), |
1587 | s_info->mem_addr, | |
82d2c88b | 1588 | mem, mem_addr, rtx_varies_p)) |
aa140b76 | 1589 | { |
1590 | s_info->rhs = NULL; | |
1591 | s_info->const_rhs = NULL; | |
1592 | } | |
3072d30e | 1593 | } |
82d2c88b | 1594 | |
3072d30e | 1595 | /* An insn can be deleted if every position of every one of |
1596 | its s_infos is zero. */ | |
aa140b76 | 1597 | if (any_positions_needed_p (s_info) |
1598 | || ptr->cannot_delete) | |
9ce37fa7 | 1599 | del = false; |
aa140b76 | 1600 | |
9ce37fa7 | 1601 | if (del) |
3072d30e | 1602 | { |
1603 | insn_info_t insn_to_delete = ptr; | |
48e1416a | 1604 | |
3072d30e | 1605 | if (last) |
1606 | last->next_local_store = ptr->next_local_store; | |
1607 | else | |
1608 | active_local_stores = ptr->next_local_store; | |
48e1416a | 1609 | |
3072d30e | 1610 | delete_dead_store_insn (insn_to_delete); |
1611 | } | |
1612 | else | |
1613 | last = ptr; | |
48e1416a | 1614 | |
3072d30e | 1615 | ptr = next; |
1616 | } | |
48e1416a | 1617 | |
3072d30e | 1618 | /* Finish filling in the store_info. */ |
1619 | store_info->next = insn_info->store_rec; | |
1620 | insn_info->store_rec = store_info; | |
82d2c88b | 1621 | store_info->mem = mem; |
3072d30e | 1622 | store_info->alias_set = spill_alias_set; |
82d2c88b | 1623 | store_info->mem_addr = mem_addr; |
3072d30e | 1624 | store_info->cse_base = base; |
aa140b76 | 1625 | if (width > HOST_BITS_PER_WIDE_INT) |
1626 | { | |
1627 | store_info->is_large = true; | |
1628 | store_info->positions_needed.large.count = 0; | |
843bd2fa | 1629 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (NULL); |
aa140b76 | 1630 | } |
1631 | else | |
1632 | { | |
1633 | store_info->is_large = false; | |
1634 | store_info->positions_needed.small_bitmask = lowpart_bitmask (width); | |
1635 | } | |
3072d30e | 1636 | store_info->group_id = group_id; |
1637 | store_info->begin = offset; | |
1638 | store_info->end = offset + width; | |
1639 | store_info->is_set = GET_CODE (body) == SET; | |
aa140b76 | 1640 | store_info->rhs = rhs; |
1641 | store_info->const_rhs = const_rhs; | |
1642 | store_info->redundant_reason = redundant_reason; | |
3072d30e | 1643 | |
3072d30e | 1644 | /* If this is a clobber, we return 0. We will only be able to |
1645 | delete this insn if there is only one store USED store, but we | |
1646 | can use the clobber to delete other stores earlier. */ | |
1647 | return store_info->is_set ? 1 : 0; | |
1648 | } | |
1649 | ||
1650 | ||
1651 | static void | |
1652 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1653 | { | |
48e1416a | 1654 | fprintf (dump_file, "%s insn=%d %s\n", start, |
3072d30e | 1655 | INSN_UID (insn_info->insn), |
1656 | insn_info->store_rec ? "has store" : "naked"); | |
1657 | } | |
1658 | ||
1659 | ||
5c9051a4 | 1660 | /* If the modes are different and the value's source and target do not |
1661 | line up, we need to extract the value from lower part of the rhs of | |
1662 | the store, shift it, and then put it into a form that can be shoved | |
1663 | into the read_insn. This function generates a right SHIFT of a | |
1664 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1665 | shift sequence is returned or NULL if we failed to find a | |
1666 | shift. */ | |
1667 | ||
1668 | static rtx | |
10d4de0e | 1669 | find_shift_sequence (int access_size, |
5c9051a4 | 1670 | store_info_t store_info, |
aa140b76 | 1671 | enum machine_mode read_mode, |
1672 | int shift, bool speed, bool require_cst) | |
5c9051a4 | 1673 | { |
1674 | enum machine_mode store_mode = GET_MODE (store_info->mem); | |
10d4de0e | 1675 | enum machine_mode new_mode; |
1676 | rtx read_reg = NULL; | |
5c9051a4 | 1677 | |
1678 | /* Some machines like the x86 have shift insns for each size of | |
1679 | operand. Other machines like the ppc or the ia-64 may only have | |
1680 | shift insns that shift values within 32 or 64 bit registers. | |
1681 | This loop tries to find the smallest shift insn that will right | |
1682 | justify the value we want to read but is available in one insn on | |
1683 | the machine. */ | |
1684 | ||
10d4de0e | 1685 | for (new_mode = smallest_mode_for_size (access_size * BITS_PER_UNIT, |
1686 | MODE_INT); | |
1687 | GET_MODE_BITSIZE (new_mode) <= BITS_PER_WORD; | |
1688 | new_mode = GET_MODE_WIDER_MODE (new_mode)) | |
5c9051a4 | 1689 | { |
10d4de0e | 1690 | rtx target, new_reg, shift_seq, insn, new_lhs; |
171557e8 | 1691 | int cost; |
af97461e | 1692 | |
4ed4afb9 | 1693 | /* If a constant was stored into memory, try to simplify it here, |
1694 | otherwise the cost of the shift might preclude this optimization | |
1695 | e.g. at -Os, even when no actual shift will be needed. */ | |
aa140b76 | 1696 | if (store_info->const_rhs) |
4ed4afb9 | 1697 | { |
1698 | unsigned int byte = subreg_lowpart_offset (new_mode, store_mode); | |
aa140b76 | 1699 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1700 | store_mode, byte); | |
4ed4afb9 | 1701 | if (ret && CONSTANT_P (ret)) |
1702 | { | |
1703 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, | |
1704 | ret, GEN_INT (shift)); | |
1705 | if (ret && CONSTANT_P (ret)) | |
1706 | { | |
1707 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1708 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1709 | if (ret && CONSTANT_P (ret) | |
1710 | && rtx_cost (ret, SET, speed) <= COSTS_N_INSNS (1)) | |
1711 | return ret; | |
1712 | } | |
1713 | } | |
1714 | } | |
1715 | ||
aa140b76 | 1716 | if (require_cst) |
1717 | return NULL_RTX; | |
1718 | ||
10d4de0e | 1719 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1720 | requires a real instruction. */ | |
1721 | if (GET_MODE_BITSIZE (new_mode) < GET_MODE_BITSIZE (store_mode) | |
1722 | && !TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (new_mode), | |
af97461e | 1723 | GET_MODE_BITSIZE (store_mode))) |
1724 | continue; | |
1725 | ||
10d4de0e | 1726 | /* Also try a wider mode if the necessary punning is either not |
1727 | desirable or not possible. */ | |
1728 | if (!CONSTANT_P (store_info->rhs) | |
1729 | && !MODES_TIEABLE_P (new_mode, store_mode)) | |
1730 | continue; | |
10d4de0e | 1731 | |
af97461e | 1732 | new_reg = gen_reg_rtx (new_mode); |
5c9051a4 | 1733 | |
1734 | start_sequence (); | |
1735 | ||
1736 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1737 | of one dsp where the cost of these two was not the same. But | |
1738 | this really is a rare case anyway. */ | |
1739 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
1740 | GEN_INT (shift), new_reg, 1, OPTAB_DIRECT); | |
1741 | ||
597d5470 | 1742 | shift_seq = get_insns (); |
1743 | end_sequence (); | |
5c9051a4 | 1744 | |
597d5470 | 1745 | if (target != new_reg || shift_seq == NULL) |
1746 | continue; | |
1747 | ||
1748 | cost = 0; | |
1749 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1750 | if (INSN_P (insn)) | |
f529eb25 | 1751 | cost += insn_rtx_cost (PATTERN (insn), speed); |
597d5470 | 1752 | |
1753 | /* The computation up to here is essentially independent | |
1754 | of the arguments and could be precomputed. It may | |
1755 | not be worth doing so. We could precompute if | |
1756 | worthwhile or at least cache the results. The result | |
5133fd21 | 1757 | technically depends on both SHIFT and ACCESS_SIZE, |
1758 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
597d5470 | 1759 | |
1760 | if (cost > COSTS_N_INSNS (1)) | |
1761 | continue; | |
1762 | ||
171557e8 | 1763 | new_lhs = extract_low_bits (new_mode, store_mode, |
1764 | copy_rtx (store_info->rhs)); | |
1765 | if (new_lhs == NULL_RTX) | |
1766 | continue; | |
1767 | ||
597d5470 | 1768 | /* We found an acceptable shift. Generate a move to |
1769 | take the value from the store and put it into the | |
1770 | shift pseudo, then shift it, then generate another | |
1771 | move to put in into the target of the read. */ | |
10d4de0e | 1772 | emit_move_insn (new_reg, new_lhs); |
597d5470 | 1773 | emit_insn (shift_seq); |
10d4de0e | 1774 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
597d5470 | 1775 | break; |
5c9051a4 | 1776 | } |
1777 | ||
10d4de0e | 1778 | return read_reg; |
5c9051a4 | 1779 | } |
1780 | ||
1781 | ||
a1b0a968 | 1782 | /* Call back for note_stores to find the hard regs set or clobbered by |
1783 | insn. Data is a bitmap of the hardregs set so far. */ | |
1784 | ||
1785 | static void | |
1786 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1787 | { | |
1788 | bitmap regs_set = (bitmap) data; | |
1789 | ||
1790 | if (REG_P (x) | |
1791 | && REGNO (x) < FIRST_PSEUDO_REGISTER) | |
1792 | { | |
1793 | int regno = REGNO (x); | |
1794 | int n = hard_regno_nregs[regno][GET_MODE (x)]; | |
1795 | while (--n >= 0) | |
1796 | bitmap_set_bit (regs_set, regno + n); | |
1797 | } | |
1798 | } | |
1799 | ||
aa140b76 | 1800 | /* Helper function for replace_read and record_store. |
1801 | Attempt to return a value stored in STORE_INFO, from READ_BEGIN | |
1802 | to one before READ_END bytes read in READ_MODE. Return NULL | |
1803 | if not successful. If REQUIRE_CST is true, return always constant. */ | |
1804 | ||
1805 | static rtx | |
1806 | get_stored_val (store_info_t store_info, enum machine_mode read_mode, | |
1807 | HOST_WIDE_INT read_begin, HOST_WIDE_INT read_end, | |
1808 | basic_block bb, bool require_cst) | |
1809 | { | |
1810 | enum machine_mode store_mode = GET_MODE (store_info->mem); | |
1811 | int shift; | |
1812 | int access_size; /* In bytes. */ | |
1813 | rtx read_reg; | |
1814 | ||
1815 | /* To get here the read is within the boundaries of the write so | |
1816 | shift will never be negative. Start out with the shift being in | |
1817 | bytes. */ | |
1818 | if (store_mode == BLKmode) | |
1819 | shift = 0; | |
1820 | else if (BYTES_BIG_ENDIAN) | |
1821 | shift = store_info->end - read_end; | |
1822 | else | |
1823 | shift = read_begin - store_info->begin; | |
1824 | ||
1825 | access_size = shift + GET_MODE_SIZE (read_mode); | |
1826 | ||
1827 | /* From now on it is bits. */ | |
1828 | shift *= BITS_PER_UNIT; | |
1829 | ||
1830 | if (shift) | |
1831 | read_reg = find_shift_sequence (access_size, store_info, read_mode, shift, | |
1832 | optimize_bb_for_speed_p (bb), | |
1833 | require_cst); | |
1834 | else if (store_mode == BLKmode) | |
1835 | { | |
1836 | /* The store is a memset (addr, const_val, const_size). */ | |
1837 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
1838 | store_mode = int_mode_for_mode (read_mode); | |
1839 | if (store_mode == BLKmode) | |
1840 | read_reg = NULL_RTX; | |
1841 | else if (store_info->rhs == const0_rtx) | |
1842 | read_reg = extract_low_bits (read_mode, store_mode, const0_rtx); | |
1843 | else if (GET_MODE_BITSIZE (store_mode) > HOST_BITS_PER_WIDE_INT | |
1844 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) | |
1845 | read_reg = NULL_RTX; | |
1846 | else | |
1847 | { | |
1848 | unsigned HOST_WIDE_INT c | |
1849 | = INTVAL (store_info->rhs) | |
1850 | & (((HOST_WIDE_INT) 1 << BITS_PER_UNIT) - 1); | |
1851 | int shift = BITS_PER_UNIT; | |
1852 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1853 | { | |
1854 | c |= (c << shift); | |
1855 | shift <<= 1; | |
1856 | } | |
1857 | read_reg = GEN_INT (trunc_int_for_mode (c, store_mode)); | |
1858 | read_reg = extract_low_bits (read_mode, store_mode, read_reg); | |
1859 | } | |
1860 | } | |
1861 | else if (store_info->const_rhs | |
1862 | && (require_cst | |
1863 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1864 | read_reg = extract_low_bits (read_mode, store_mode, | |
1865 | copy_rtx (store_info->const_rhs)); | |
1866 | else | |
1867 | read_reg = extract_low_bits (read_mode, store_mode, | |
1868 | copy_rtx (store_info->rhs)); | |
1869 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1870 | read_reg = NULL_RTX; | |
1871 | return read_reg; | |
1872 | } | |
a1b0a968 | 1873 | |
3072d30e | 1874 | /* Take a sequence of: |
1875 | A <- r1 | |
1876 | ... | |
1877 | ... <- A | |
1878 | ||
48e1416a | 1879 | and change it into |
3072d30e | 1880 | r2 <- r1 |
1881 | A <- r1 | |
1882 | ... | |
1883 | ... <- r2 | |
1884 | ||
5c9051a4 | 1885 | or |
1886 | ||
1887 | r3 <- extract (r1) | |
1888 | r3 <- r3 >> shift | |
1889 | r2 <- extract (r3) | |
1890 | ... <- r2 | |
1891 | ||
1892 | or | |
1893 | ||
1894 | r2 <- extract (r1) | |
1895 | ... <- r2 | |
1896 | ||
1897 | Depending on the alignment and the mode of the store and | |
1898 | subsequent load. | |
1899 | ||
1900 | ||
1901 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
3072d30e | 1902 | and READ_INSN are for the read. Return true if the replacement |
1903 | went ok. */ | |
1904 | ||
1905 | static bool | |
48e1416a | 1906 | replace_read (store_info_t store_info, insn_info_t store_insn, |
aa140b76 | 1907 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
1908 | bitmap regs_live) | |
3072d30e | 1909 | { |
5c9051a4 | 1910 | enum machine_mode store_mode = GET_MODE (store_info->mem); |
1911 | enum machine_mode read_mode = GET_MODE (read_info->mem); | |
a1b0a968 | 1912 | rtx insns, this_insn, read_reg; |
aa140b76 | 1913 | basic_block bb; |
5c9051a4 | 1914 | |
3072d30e | 1915 | if (!dbg_cnt (dse)) |
1916 | return false; | |
1917 | ||
10d4de0e | 1918 | /* Create a sequence of instructions to set up the read register. |
1919 | This sequence goes immediately before the store and its result | |
1920 | is read by the load. | |
1921 | ||
1922 | We need to keep this in perspective. We are replacing a read | |
5c9051a4 | 1923 | with a sequence of insns, but the read will almost certainly be |
1924 | in cache, so it is not going to be an expensive one. Thus, we | |
1925 | are not willing to do a multi insn shift or worse a subroutine | |
1926 | call to get rid of the read. */ | |
10d4de0e | 1927 | if (dump_file) |
1928 | fprintf (dump_file, "trying to replace %smode load in insn %d" | |
1929 | " from %smode store in insn %d\n", | |
1930 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
1931 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
1932 | start_sequence (); | |
aa140b76 | 1933 | bb = BLOCK_FOR_INSN (read_insn->insn); |
1934 | read_reg = get_stored_val (store_info, | |
1935 | read_mode, read_info->begin, read_info->end, | |
1936 | bb, false); | |
10d4de0e | 1937 | if (read_reg == NULL_RTX) |
5c9051a4 | 1938 | { |
10d4de0e | 1939 | end_sequence (); |
1940 | if (dump_file) | |
1941 | fprintf (dump_file, " -- could not extract bits of stored value\n"); | |
1942 | return false; | |
5c9051a4 | 1943 | } |
10d4de0e | 1944 | /* Force the value into a new register so that it won't be clobbered |
1945 | between the store and the load. */ | |
1946 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
1947 | insns = get_insns (); | |
1948 | end_sequence (); | |
5c9051a4 | 1949 | |
a1b0a968 | 1950 | if (insns != NULL_RTX) |
1951 | { | |
1952 | /* Now we have to scan the set of new instructions to see if the | |
1953 | sequence contains and sets of hardregs that happened to be | |
1954 | live at this point. For instance, this can happen if one of | |
1955 | the insns sets the CC and the CC happened to be live at that | |
1956 | point. This does occasionally happen, see PR 37922. */ | |
1957 | bitmap regs_set = BITMAP_ALLOC (NULL); | |
1958 | ||
1959 | for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
1960 | note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
48e1416a | 1961 | |
a1b0a968 | 1962 | bitmap_and_into (regs_set, regs_live); |
1963 | if (!bitmap_empty_p (regs_set)) | |
1964 | { | |
1965 | if (dump_file) | |
1966 | { | |
48e1416a | 1967 | fprintf (dump_file, |
a1b0a968 | 1968 | "abandoning replacement because sequence clobbers live hardregs:"); |
1969 | df_print_regset (dump_file, regs_set); | |
1970 | } | |
48e1416a | 1971 | |
a1b0a968 | 1972 | BITMAP_FREE (regs_set); |
1973 | return false; | |
1974 | } | |
1975 | BITMAP_FREE (regs_set); | |
1976 | } | |
1977 | ||
5c9051a4 | 1978 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
3072d30e | 1979 | { |
364c0c59 | 1980 | deferred_change_t deferred_change = |
1981 | (deferred_change_t) pool_alloc (deferred_change_pool); | |
48e1416a | 1982 | |
5c9051a4 | 1983 | /* Insert this right before the store insn where it will be safe |
1984 | from later insns that might change it before the read. */ | |
1985 | emit_insn_before (insns, store_insn->insn); | |
48e1416a | 1986 | |
5c9051a4 | 1987 | /* And now for the kludge part: cselib croaks if you just |
1988 | return at this point. There are two reasons for this: | |
48e1416a | 1989 | |
5c9051a4 | 1990 | 1) Cselib has an idea of how many pseudos there are and |
1991 | that does not include the new ones we just added. | |
48e1416a | 1992 | |
5c9051a4 | 1993 | 2) Cselib does not know about the move insn we added |
1994 | above the store_info, and there is no way to tell it | |
1995 | about it, because it has "moved on". | |
48e1416a | 1996 | |
5c9051a4 | 1997 | Problem (1) is fixable with a certain amount of engineering. |
1998 | Problem (2) is requires starting the bb from scratch. This | |
1999 | could be expensive. | |
48e1416a | 2000 | |
5c9051a4 | 2001 | So we are just going to have to lie. The move/extraction |
2002 | insns are not really an issue, cselib did not see them. But | |
2003 | the use of the new pseudo read_insn is a real problem because | |
2004 | cselib has not scanned this insn. The way that we solve this | |
2005 | problem is that we are just going to put the mem back for now | |
2006 | and when we are finished with the block, we undo this. We | |
2007 | keep a table of mems to get rid of. At the end of the basic | |
2008 | block we can put them back. */ | |
48e1416a | 2009 | |
5c9051a4 | 2010 | *loc = read_info->mem; |
2011 | deferred_change->next = deferred_change_list; | |
2012 | deferred_change_list = deferred_change; | |
2013 | deferred_change->loc = loc; | |
2014 | deferred_change->reg = read_reg; | |
48e1416a | 2015 | |
5c9051a4 | 2016 | /* Get rid of the read_info, from the point of view of the |
2017 | rest of dse, play like this read never happened. */ | |
2018 | read_insn->read_rec = read_info->next; | |
2019 | pool_free (read_info_pool, read_info); | |
10d4de0e | 2020 | if (dump_file) |
2021 | { | |
2022 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2023 | print_simple_rtl (dump_file, read_reg); | |
2024 | fprintf (dump_file, "\n"); | |
2025 | } | |
5c9051a4 | 2026 | return true; |
3072d30e | 2027 | } |
48e1416a | 2028 | else |
3072d30e | 2029 | { |
3072d30e | 2030 | if (dump_file) |
10d4de0e | 2031 | { |
2032 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2033 | print_simple_rtl (dump_file, read_reg); | |
2034 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2035 | } | |
3072d30e | 2036 | return false; |
2037 | } | |
2038 | } | |
2039 | ||
3072d30e | 2040 | /* A for_each_rtx callback in which DATA is the bb_info. Check to see |
2041 | if LOC is a mem and if it is look at the address and kill any | |
2042 | appropriate stores that may be active. */ | |
2043 | ||
2044 | static int | |
2045 | check_mem_read_rtx (rtx *loc, void *data) | |
2046 | { | |
82d2c88b | 2047 | rtx mem = *loc, mem_addr; |
3072d30e | 2048 | bb_info_t bb_info; |
2049 | insn_info_t insn_info; | |
2050 | HOST_WIDE_INT offset = 0; | |
2051 | HOST_WIDE_INT width = 0; | |
32c2fdea | 2052 | alias_set_type spill_alias_set = 0; |
48e1416a | 2053 | cselib_val *base = NULL; |
3072d30e | 2054 | int group_id; |
2055 | read_info_t read_info; | |
2056 | ||
2057 | if (!mem || !MEM_P (mem)) | |
2058 | return 0; | |
2059 | ||
2060 | bb_info = (bb_info_t) data; | |
2061 | insn_info = bb_info->last_insn; | |
2062 | ||
2063 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2064 | || (MEM_VOLATILE_P (mem))) | |
2065 | { | |
2066 | if (dump_file) | |
2067 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); | |
2068 | add_wild_read (bb_info); | |
2069 | insn_info->cannot_delete = true; | |
2070 | return 0; | |
2071 | } | |
2072 | ||
2073 | /* If it is reading readonly mem, then there can be no conflict with | |
2074 | another write. */ | |
2075 | if (MEM_READONLY_P (mem)) | |
2076 | return 0; | |
2077 | ||
2078 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
2079 | { | |
2080 | if (dump_file) | |
2081 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); | |
2082 | add_wild_read (bb_info); | |
2083 | return 0; | |
2084 | } | |
2085 | ||
2086 | if (GET_MODE (mem) == BLKmode) | |
2087 | width = -1; | |
2088 | else | |
2089 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2090 | ||
364c0c59 | 2091 | read_info = (read_info_t) pool_alloc (read_info_pool); |
3072d30e | 2092 | read_info->group_id = group_id; |
2093 | read_info->mem = mem; | |
2094 | read_info->alias_set = spill_alias_set; | |
2095 | read_info->begin = offset; | |
2096 | read_info->end = offset + width; | |
2097 | read_info->next = insn_info->read_rec; | |
2098 | insn_info->read_rec = read_info; | |
82d2c88b | 2099 | /* For alias_set != 0 canon_true_dependence should be never called. */ |
2100 | if (spill_alias_set) | |
2101 | mem_addr = NULL_RTX; | |
2102 | else | |
2103 | { | |
2104 | if (group_id < 0) | |
2105 | mem_addr = base->val_rtx; | |
2106 | else | |
2107 | { | |
2108 | group_info_t group | |
2109 | = VEC_index (group_info_t, rtx_group_vec, group_id); | |
2110 | mem_addr = group->canon_base_addr; | |
2111 | } | |
2112 | if (offset) | |
2113 | mem_addr = plus_constant (mem_addr, offset); | |
2114 | } | |
3072d30e | 2115 | |
bef304b8 | 2116 | /* We ignore the clobbers in store_info. The is mildly aggressive, |
3072d30e | 2117 | but there really should not be a clobber followed by a read. */ |
2118 | ||
2119 | if (spill_alias_set) | |
2120 | { | |
2121 | insn_info_t i_ptr = active_local_stores; | |
2122 | insn_info_t last = NULL; | |
2123 | ||
2124 | if (dump_file) | |
2125 | fprintf (dump_file, " processing spill load %d\n", | |
32c2fdea | 2126 | (int) spill_alias_set); |
3072d30e | 2127 | |
2128 | while (i_ptr) | |
2129 | { | |
2130 | store_info_t store_info = i_ptr->store_rec; | |
2131 | ||
2132 | /* Skip the clobbers. */ | |
2133 | while (!store_info->is_set) | |
2134 | store_info = store_info->next; | |
48e1416a | 2135 | |
3072d30e | 2136 | if (store_info->alias_set == spill_alias_set) |
2137 | { | |
2138 | if (dump_file) | |
2139 | dump_insn_info ("removing from active", i_ptr); | |
2140 | ||
2141 | if (last) | |
2142 | last->next_local_store = i_ptr->next_local_store; | |
2143 | else | |
2144 | active_local_stores = i_ptr->next_local_store; | |
2145 | } | |
2146 | else | |
2147 | last = i_ptr; | |
2148 | i_ptr = i_ptr->next_local_store; | |
2149 | } | |
2150 | } | |
2151 | else if (group_id >= 0) | |
2152 | { | |
2153 | /* This is the restricted case where the base is a constant or | |
2154 | the frame pointer and offset is a constant. */ | |
2155 | insn_info_t i_ptr = active_local_stores; | |
2156 | insn_info_t last = NULL; | |
48e1416a | 2157 | |
3072d30e | 2158 | if (dump_file) |
2159 | { | |
2160 | if (width == -1) | |
2161 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", | |
2162 | group_id); | |
2163 | else | |
2164 | fprintf (dump_file, " processing const load gid=%d[%d..%d)\n", | |
2165 | group_id, (int)offset, (int)(offset+width)); | |
2166 | } | |
2167 | ||
2168 | while (i_ptr) | |
2169 | { | |
2170 | bool remove = false; | |
2171 | store_info_t store_info = i_ptr->store_rec; | |
48e1416a | 2172 | |
3072d30e | 2173 | /* Skip the clobbers. */ |
2174 | while (!store_info->is_set) | |
2175 | store_info = store_info->next; | |
48e1416a | 2176 | |
3072d30e | 2177 | /* There are three cases here. */ |
2178 | if (store_info->group_id < 0) | |
2179 | /* We have a cselib store followed by a read from a | |
2180 | const base. */ | |
48e1416a | 2181 | remove |
2182 | = canon_true_dependence (store_info->mem, | |
3072d30e | 2183 | GET_MODE (store_info->mem), |
2184 | store_info->mem_addr, | |
82d2c88b | 2185 | mem, mem_addr, rtx_varies_p); |
48e1416a | 2186 | |
3072d30e | 2187 | else if (group_id == store_info->group_id) |
2188 | { | |
2189 | /* This is a block mode load. We may get lucky and | |
2190 | canon_true_dependence may save the day. */ | |
2191 | if (width == -1) | |
48e1416a | 2192 | remove |
2193 | = canon_true_dependence (store_info->mem, | |
3072d30e | 2194 | GET_MODE (store_info->mem), |
2195 | store_info->mem_addr, | |
82d2c88b | 2196 | mem, mem_addr, rtx_varies_p); |
48e1416a | 2197 | |
3072d30e | 2198 | /* If this read is just reading back something that we just |
2199 | stored, rewrite the read. */ | |
48e1416a | 2200 | else |
3072d30e | 2201 | { |
2202 | if (store_info->rhs | |
aa140b76 | 2203 | && offset >= store_info->begin |
2204 | && offset + width <= store_info->end | |
2205 | && all_positions_needed_p (store_info, | |
2206 | offset - store_info->begin, | |
2207 | width) | |
2208 | && replace_read (store_info, i_ptr, read_info, | |
2209 | insn_info, loc, bb_info->regs_live)) | |
2210 | return 0; | |
2211 | ||
3072d30e | 2212 | /* The bases are the same, just see if the offsets |
2213 | overlap. */ | |
48e1416a | 2214 | if ((offset < store_info->end) |
3072d30e | 2215 | && (offset + width > store_info->begin)) |
2216 | remove = true; | |
2217 | } | |
2218 | } | |
48e1416a | 2219 | |
2220 | /* else | |
3072d30e | 2221 | The else case that is missing here is that the |
2222 | bases are constant but different. There is nothing | |
2223 | to do here because there is no overlap. */ | |
48e1416a | 2224 | |
3072d30e | 2225 | if (remove) |
2226 | { | |
2227 | if (dump_file) | |
2228 | dump_insn_info ("removing from active", i_ptr); | |
2229 | ||
2230 | if (last) | |
2231 | last->next_local_store = i_ptr->next_local_store; | |
2232 | else | |
2233 | active_local_stores = i_ptr->next_local_store; | |
2234 | } | |
2235 | else | |
2236 | last = i_ptr; | |
2237 | i_ptr = i_ptr->next_local_store; | |
2238 | } | |
2239 | } | |
48e1416a | 2240 | else |
3072d30e | 2241 | { |
2242 | insn_info_t i_ptr = active_local_stores; | |
2243 | insn_info_t last = NULL; | |
2244 | if (dump_file) | |
2245 | { | |
2246 | fprintf (dump_file, " processing cselib load mem:"); | |
2247 | print_inline_rtx (dump_file, mem, 0); | |
2248 | fprintf (dump_file, "\n"); | |
2249 | } | |
2250 | ||
2251 | while (i_ptr) | |
2252 | { | |
2253 | bool remove = false; | |
2254 | store_info_t store_info = i_ptr->store_rec; | |
48e1416a | 2255 | |
3072d30e | 2256 | if (dump_file) |
2257 | fprintf (dump_file, " processing cselib load against insn %d\n", | |
2258 | INSN_UID (i_ptr->insn)); | |
2259 | ||
2260 | /* Skip the clobbers. */ | |
2261 | while (!store_info->is_set) | |
2262 | store_info = store_info->next; | |
2263 | ||
2264 | /* If this read is just reading back something that we just | |
2265 | stored, rewrite the read. */ | |
2266 | if (store_info->rhs | |
2267 | && store_info->group_id == -1 | |
2268 | && store_info->cse_base == base | |
2ffe5515 | 2269 | && width != -1 |
aa140b76 | 2270 | && offset >= store_info->begin |
2271 | && offset + width <= store_info->end | |
2272 | && all_positions_needed_p (store_info, | |
2273 | offset - store_info->begin, width) | |
2274 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, | |
2275 | bb_info->regs_live)) | |
2276 | return 0; | |
3072d30e | 2277 | |
2278 | if (!store_info->alias_set) | |
48e1416a | 2279 | remove = canon_true_dependence (store_info->mem, |
3072d30e | 2280 | GET_MODE (store_info->mem), |
2281 | store_info->mem_addr, | |
82d2c88b | 2282 | mem, mem_addr, rtx_varies_p); |
48e1416a | 2283 | |
3072d30e | 2284 | if (remove) |
2285 | { | |
2286 | if (dump_file) | |
2287 | dump_insn_info ("removing from active", i_ptr); | |
48e1416a | 2288 | |
3072d30e | 2289 | if (last) |
2290 | last->next_local_store = i_ptr->next_local_store; | |
2291 | else | |
2292 | active_local_stores = i_ptr->next_local_store; | |
2293 | } | |
2294 | else | |
2295 | last = i_ptr; | |
2296 | i_ptr = i_ptr->next_local_store; | |
2297 | } | |
2298 | } | |
2299 | return 0; | |
2300 | } | |
2301 | ||
48e1416a | 2302 | /* A for_each_rtx callback in which DATA points the INSN_INFO for |
3072d30e | 2303 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2304 | true for any part of *LOC. */ | |
2305 | ||
2306 | static void | |
2307 | check_mem_read_use (rtx *loc, void *data) | |
2308 | { | |
2309 | for_each_rtx (loc, check_mem_read_rtx, data); | |
2310 | } | |
2311 | ||
aa140b76 | 2312 | |
2313 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2314 | So far it only handles arguments passed in registers. */ | |
2315 | ||
2316 | static bool | |
2317 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2318 | { | |
2319 | CUMULATIVE_ARGS args_so_far; | |
2320 | tree arg; | |
2321 | int idx; | |
2322 | ||
2323 | INIT_CUMULATIVE_ARGS (args_so_far, TREE_TYPE (fn), NULL_RTX, 0, 3); | |
2324 | ||
2325 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2326 | for (idx = 0; | |
2327 | arg != void_list_node && idx < nargs; | |
2328 | arg = TREE_CHAIN (arg), idx++) | |
2329 | { | |
2330 | enum machine_mode mode = TYPE_MODE (TREE_VALUE (arg)); | |
2331 | rtx reg = FUNCTION_ARG (args_so_far, mode, NULL_TREE, 1), link, tmp; | |
2332 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode | |
2333 | || GET_MODE_CLASS (mode) != MODE_INT) | |
2334 | return false; | |
2335 | ||
2336 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2337 | link; | |
2338 | link = XEXP (link, 1)) | |
2339 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2340 | { | |
2341 | args[idx] = XEXP (XEXP (link, 0), 0); | |
2342 | if (REG_P (args[idx]) | |
2343 | && REGNO (args[idx]) == REGNO (reg) | |
2344 | && (GET_MODE (args[idx]) == mode | |
2345 | || (GET_MODE_CLASS (GET_MODE (args[idx])) == MODE_INT | |
2346 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2347 | <= UNITS_PER_WORD) | |
2348 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2349 | > GET_MODE_SIZE (mode))))) | |
2350 | break; | |
2351 | } | |
2352 | if (!link) | |
2353 | return false; | |
2354 | ||
2355 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2356 | if (GET_MODE (args[idx]) != mode) | |
2357 | { | |
2358 | if (!tmp || !CONST_INT_P (tmp)) | |
2359 | return false; | |
2360 | tmp = GEN_INT (trunc_int_for_mode (INTVAL (tmp), mode)); | |
2361 | } | |
2362 | if (tmp) | |
2363 | args[idx] = tmp; | |
2364 | ||
2365 | FUNCTION_ARG_ADVANCE (args_so_far, mode, NULL_TREE, 1); | |
2366 | } | |
2367 | if (arg != void_list_node || idx != nargs) | |
2368 | return false; | |
2369 | return true; | |
2370 | } | |
2371 | ||
2372 | ||
3072d30e | 2373 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2374 | if some part of it is not a candidate store and assigns to a | |
2375 | non-register target. */ | |
2376 | ||
2377 | static void | |
2378 | scan_insn (bb_info_t bb_info, rtx insn) | |
2379 | { | |
2380 | rtx body; | |
364c0c59 | 2381 | insn_info_t insn_info = (insn_info_t) pool_alloc (insn_info_pool); |
3072d30e | 2382 | int mems_found = 0; |
2383 | memset (insn_info, 0, sizeof (struct insn_info)); | |
2384 | ||
2385 | if (dump_file) | |
2386 | fprintf (dump_file, "\n**scanning insn=%d\n", | |
2387 | INSN_UID (insn)); | |
2388 | ||
2389 | insn_info->prev_insn = bb_info->last_insn; | |
2390 | insn_info->insn = insn; | |
2391 | bb_info->last_insn = insn_info; | |
48e1416a | 2392 | |
9845d120 | 2393 | if (DEBUG_INSN_P (insn)) |
2394 | { | |
2395 | insn_info->cannot_delete = true; | |
2396 | return; | |
2397 | } | |
3072d30e | 2398 | |
bef304b8 | 2399 | /* Cselib clears the table for this case, so we have to essentially |
3072d30e | 2400 | do the same. */ |
2401 | if (NONJUMP_INSN_P (insn) | |
2402 | && GET_CODE (PATTERN (insn)) == ASM_OPERANDS | |
2403 | && MEM_VOLATILE_P (PATTERN (insn))) | |
2404 | { | |
2405 | add_wild_read (bb_info); | |
2406 | insn_info->cannot_delete = true; | |
2407 | return; | |
2408 | } | |
2409 | ||
2410 | /* Look at all of the uses in the insn. */ | |
2411 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2412 | ||
2413 | if (CALL_P (insn)) | |
2414 | { | |
aa140b76 | 2415 | bool const_call; |
2416 | tree memset_call = NULL_TREE; | |
2417 | ||
3072d30e | 2418 | insn_info->cannot_delete = true; |
17e1318c | 2419 | |
3072d30e | 2420 | /* Const functions cannot do anything bad i.e. read memory, |
17e1318c | 2421 | however, they can read their parameters which may have |
aa140b76 | 2422 | been pushed onto the stack. |
2423 | memset and bzero don't read memory either. */ | |
2424 | const_call = RTL_CONST_CALL_P (insn); | |
2425 | if (!const_call) | |
2426 | { | |
2427 | rtx call = PATTERN (insn); | |
2428 | if (GET_CODE (call) == PARALLEL) | |
2429 | call = XVECEXP (call, 0, 0); | |
2430 | if (GET_CODE (call) == SET) | |
2431 | call = SET_SRC (call); | |
2432 | if (GET_CODE (call) == CALL | |
2433 | && MEM_P (XEXP (call, 0)) | |
2434 | && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF) | |
2435 | { | |
2436 | rtx symbol = XEXP (XEXP (call, 0), 0); | |
2437 | if (SYMBOL_REF_DECL (symbol) | |
2438 | && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL) | |
2439 | { | |
2440 | if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol)) | |
2441 | == BUILT_IN_NORMAL | |
2442 | && (DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol)) | |
2443 | == BUILT_IN_MEMSET)) | |
2444 | || SYMBOL_REF_DECL (symbol) == block_clear_fn) | |
2445 | memset_call = SYMBOL_REF_DECL (symbol); | |
2446 | } | |
2447 | } | |
2448 | } | |
2449 | if (const_call || memset_call) | |
3072d30e | 2450 | { |
2451 | insn_info_t i_ptr = active_local_stores; | |
2452 | insn_info_t last = NULL; | |
2453 | ||
2454 | if (dump_file) | |
aa140b76 | 2455 | fprintf (dump_file, "%s call %d\n", |
2456 | const_call ? "const" : "memset", INSN_UID (insn)); | |
3072d30e | 2457 | |
16bf64db | 2458 | /* See the head comment of the frame_read field. */ |
2459 | if (reload_completed) | |
2460 | insn_info->frame_read = true; | |
2461 | ||
2462 | /* Loop over the active stores and remove those which are | |
2463 | killed by the const function call. */ | |
3072d30e | 2464 | while (i_ptr) |
2465 | { | |
16bf64db | 2466 | bool remove_store = false; |
2467 | ||
2468 | /* The stack pointer based stores are always killed. */ | |
17e1318c | 2469 | if (i_ptr->stack_pointer_based) |
16bf64db | 2470 | remove_store = true; |
2471 | ||
2472 | /* If the frame is read, the frame related stores are killed. */ | |
2473 | else if (insn_info->frame_read) | |
2474 | { | |
2475 | store_info_t store_info = i_ptr->store_rec; | |
2476 | ||
2477 | /* Skip the clobbers. */ | |
2478 | while (!store_info->is_set) | |
2479 | store_info = store_info->next; | |
2480 | ||
2481 | if (store_info->group_id >= 0 | |
2482 | && VEC_index (group_info_t, rtx_group_vec, | |
2483 | store_info->group_id)->frame_related) | |
2484 | remove_store = true; | |
2485 | } | |
2486 | ||
2487 | if (remove_store) | |
3072d30e | 2488 | { |
2489 | if (dump_file) | |
2490 | dump_insn_info ("removing from active", i_ptr); | |
48e1416a | 2491 | |
3072d30e | 2492 | if (last) |
2493 | last->next_local_store = i_ptr->next_local_store; | |
2494 | else | |
2495 | active_local_stores = i_ptr->next_local_store; | |
2496 | } | |
2497 | else | |
2498 | last = i_ptr; | |
16bf64db | 2499 | |
3072d30e | 2500 | i_ptr = i_ptr->next_local_store; |
2501 | } | |
aa140b76 | 2502 | |
2503 | if (memset_call) | |
2504 | { | |
2505 | rtx args[3]; | |
2506 | if (get_call_args (insn, memset_call, args, 3) | |
2507 | && CONST_INT_P (args[1]) | |
2508 | && CONST_INT_P (args[2]) | |
2509 | && INTVAL (args[2]) > 0) | |
2510 | { | |
2511 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
2512 | set_mem_size (mem, args[2]); | |
2513 | body = gen_rtx_SET (VOIDmode, mem, args[1]); | |
2514 | mems_found += record_store (body, bb_info); | |
2515 | if (dump_file) | |
2516 | fprintf (dump_file, "handling memset as BLKmode store\n"); | |
2517 | if (mems_found == 1) | |
2518 | { | |
2519 | insn_info->next_local_store = active_local_stores; | |
2520 | active_local_stores = insn_info; | |
2521 | } | |
2522 | } | |
2523 | } | |
3072d30e | 2524 | } |
2525 | ||
17e1318c | 2526 | else |
2527 | /* Every other call, including pure functions, may read memory. */ | |
2528 | add_wild_read (bb_info); | |
2529 | ||
3072d30e | 2530 | return; |
2531 | } | |
2532 | ||
2533 | /* Assuming that there are sets in these insns, we cannot delete | |
2534 | them. */ | |
2535 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
4aafe72f | 2536 | || volatile_refs_p (PATTERN (insn)) |
e38def9c | 2537 | || insn_could_throw_p (insn) |
3072d30e | 2538 | || (RTX_FRAME_RELATED_P (insn)) |
2539 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2540 | insn_info->cannot_delete = true; | |
48e1416a | 2541 | |
3072d30e | 2542 | body = PATTERN (insn); |
2543 | if (GET_CODE (body) == PARALLEL) | |
2544 | { | |
2545 | int i; | |
2546 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2547 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2548 | } | |
2549 | else | |
2550 | mems_found += record_store (body, bb_info); | |
2551 | ||
2552 | if (dump_file) | |
48e1416a | 2553 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
3072d30e | 2554 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2555 | ||
aa140b76 | 2556 | /* If we found some sets of mems, add it into the active_local_stores so |
2557 | that it can be locally deleted if found dead or used for | |
2558 | replace_read and redundant constant store elimination. Otherwise mark | |
2559 | it as cannot delete. This simplifies the processing later. */ | |
2560 | if (mems_found == 1) | |
3072d30e | 2561 | { |
2562 | insn_info->next_local_store = active_local_stores; | |
2563 | active_local_stores = insn_info; | |
2564 | } | |
2565 | else | |
2566 | insn_info->cannot_delete = true; | |
2567 | } | |
2568 | ||
2569 | ||
2570 | /* Remove BASE from the set of active_local_stores. This is a | |
2571 | callback from cselib that is used to get rid of the stores in | |
2572 | active_local_stores. */ | |
2573 | ||
2574 | static void | |
2575 | remove_useless_values (cselib_val *base) | |
2576 | { | |
2577 | insn_info_t insn_info = active_local_stores; | |
2578 | insn_info_t last = NULL; | |
2579 | ||
2580 | while (insn_info) | |
2581 | { | |
2582 | store_info_t store_info = insn_info->store_rec; | |
9ce37fa7 | 2583 | bool del = false; |
3072d30e | 2584 | |
2585 | /* If ANY of the store_infos match the cselib group that is | |
2586 | being deleted, then the insn can not be deleted. */ | |
2587 | while (store_info) | |
2588 | { | |
48e1416a | 2589 | if ((store_info->group_id == -1) |
3072d30e | 2590 | && (store_info->cse_base == base)) |
2591 | { | |
9ce37fa7 | 2592 | del = true; |
3072d30e | 2593 | break; |
2594 | } | |
2595 | store_info = store_info->next; | |
2596 | } | |
2597 | ||
9ce37fa7 | 2598 | if (del) |
3072d30e | 2599 | { |
2600 | if (last) | |
2601 | last->next_local_store = insn_info->next_local_store; | |
2602 | else | |
2603 | active_local_stores = insn_info->next_local_store; | |
2604 | free_store_info (insn_info); | |
2605 | } | |
2606 | else | |
2607 | last = insn_info; | |
48e1416a | 2608 | |
3072d30e | 2609 | insn_info = insn_info->next_local_store; |
2610 | } | |
2611 | } | |
2612 | ||
2613 | ||
2614 | /* Do all of step 1. */ | |
2615 | ||
2616 | static void | |
2617 | dse_step1 (void) | |
2618 | { | |
2619 | basic_block bb; | |
a1b0a968 | 2620 | bitmap regs_live = BITMAP_ALLOC (NULL); |
48e1416a | 2621 | |
3072d30e | 2622 | cselib_init (false); |
2623 | all_blocks = BITMAP_ALLOC (NULL); | |
2624 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2625 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2626 | ||
2627 | FOR_ALL_BB (bb) | |
2628 | { | |
2629 | insn_info_t ptr; | |
364c0c59 | 2630 | bb_info_t bb_info = (bb_info_t) pool_alloc (bb_info_pool); |
3072d30e | 2631 | |
2632 | memset (bb_info, 0, sizeof (struct bb_info)); | |
2633 | bitmap_set_bit (all_blocks, bb->index); | |
a1b0a968 | 2634 | bb_info->regs_live = regs_live; |
2635 | ||
2636 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2637 | df_simulate_initialize_forwards (bb, regs_live); | |
3072d30e | 2638 | |
2639 | bb_table[bb->index] = bb_info; | |
2640 | cselib_discard_hook = remove_useless_values; | |
2641 | ||
2642 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2643 | { | |
2644 | rtx insn; | |
2645 | ||
2646 | cse_store_info_pool | |
48e1416a | 2647 | = create_alloc_pool ("cse_store_info_pool", |
3072d30e | 2648 | sizeof (struct store_info), 100); |
2649 | active_local_stores = NULL; | |
2650 | cselib_clear_table (); | |
48e1416a | 2651 | |
3072d30e | 2652 | /* Scan the insns. */ |
2653 | FOR_BB_INSNS (bb, insn) | |
2654 | { | |
2655 | if (INSN_P (insn)) | |
2656 | scan_insn (bb_info, insn); | |
2657 | cselib_process_insn (insn); | |
a1b0a968 | 2658 | if (INSN_P (insn)) |
2659 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
3072d30e | 2660 | } |
48e1416a | 2661 | |
3072d30e | 2662 | /* This is something of a hack, because the global algorithm |
2663 | is supposed to take care of the case where stores go dead | |
2664 | at the end of the function. However, the global | |
2665 | algorithm must take a more conservative view of block | |
2666 | mode reads than the local alg does. So to get the case | |
2667 | where you have a store to the frame followed by a non | |
bef304b8 | 2668 | overlapping block more read, we look at the active local |
3072d30e | 2669 | stores at the end of the function and delete all of the |
2670 | frame and spill based ones. */ | |
2671 | if (stores_off_frame_dead_at_return | |
2672 | && (EDGE_COUNT (bb->succs) == 0 | |
2673 | || (single_succ_p (bb) | |
2674 | && single_succ (bb) == EXIT_BLOCK_PTR | |
18d50ae6 | 2675 | && ! crtl->calls_eh_return))) |
3072d30e | 2676 | { |
2677 | insn_info_t i_ptr = active_local_stores; | |
2678 | while (i_ptr) | |
2679 | { | |
2680 | store_info_t store_info = i_ptr->store_rec; | |
2681 | ||
2682 | /* Skip the clobbers. */ | |
2683 | while (!store_info->is_set) | |
2684 | store_info = store_info->next; | |
aa140b76 | 2685 | if (store_info->alias_set && !i_ptr->cannot_delete) |
3072d30e | 2686 | delete_dead_store_insn (i_ptr); |
48e1416a | 2687 | else |
3072d30e | 2688 | if (store_info->group_id >= 0) |
2689 | { | |
48e1416a | 2690 | group_info_t group |
3072d30e | 2691 | = VEC_index (group_info_t, rtx_group_vec, store_info->group_id); |
aa140b76 | 2692 | if (group->frame_related && !i_ptr->cannot_delete) |
3072d30e | 2693 | delete_dead_store_insn (i_ptr); |
2694 | } | |
2695 | ||
2696 | i_ptr = i_ptr->next_local_store; | |
2697 | } | |
2698 | } | |
2699 | ||
2700 | /* Get rid of the loads that were discovered in | |
2701 | replace_read. Cselib is finished with this block. */ | |
2702 | while (deferred_change_list) | |
2703 | { | |
2704 | deferred_change_t next = deferred_change_list->next; | |
2705 | ||
2706 | /* There is no reason to validate this change. That was | |
2707 | done earlier. */ | |
2708 | *deferred_change_list->loc = deferred_change_list->reg; | |
2709 | pool_free (deferred_change_pool, deferred_change_list); | |
2710 | deferred_change_list = next; | |
2711 | } | |
2712 | ||
2713 | /* Get rid of all of the cselib based store_infos in this | |
2714 | block and mark the containing insns as not being | |
2715 | deletable. */ | |
2716 | ptr = bb_info->last_insn; | |
2717 | while (ptr) | |
2718 | { | |
2719 | if (ptr->contains_cselib_groups) | |
aa140b76 | 2720 | { |
2721 | store_info_t s_info = ptr->store_rec; | |
2722 | while (s_info && !s_info->is_set) | |
2723 | s_info = s_info->next; | |
2724 | if (s_info | |
2725 | && s_info->redundant_reason | |
2726 | && s_info->redundant_reason->insn | |
2727 | && !ptr->cannot_delete) | |
2728 | { | |
2729 | if (dump_file) | |
2730 | fprintf (dump_file, "Locally deleting insn %d " | |
2731 | "because insn %d stores the " | |
2732 | "same value and couldn't be " | |
2733 | "eliminated\n", | |
2734 | INSN_UID (ptr->insn), | |
2735 | INSN_UID (s_info->redundant_reason->insn)); | |
2736 | delete_dead_store_insn (ptr); | |
2737 | } | |
2738 | if (s_info) | |
2739 | s_info->redundant_reason = NULL; | |
2740 | free_store_info (ptr); | |
2741 | } | |
2742 | else | |
2743 | { | |
2744 | store_info_t s_info; | |
2745 | ||
2746 | /* Free at least positions_needed bitmaps. */ | |
2747 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2748 | if (s_info->is_large) | |
2749 | { | |
843bd2fa | 2750 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
aa140b76 | 2751 | s_info->is_large = false; |
2752 | } | |
2753 | } | |
3072d30e | 2754 | ptr = ptr->prev_insn; |
2755 | } | |
2756 | ||
2757 | free_alloc_pool (cse_store_info_pool); | |
2758 | } | |
a1b0a968 | 2759 | bb_info->regs_live = NULL; |
3072d30e | 2760 | } |
2761 | ||
a1b0a968 | 2762 | BITMAP_FREE (regs_live); |
3072d30e | 2763 | cselib_finish (); |
2764 | htab_empty (rtx_group_table); | |
2765 | } | |
2766 | ||
2767 | \f | |
2768 | /*---------------------------------------------------------------------------- | |
2769 | Second step. | |
2770 | ||
2771 | Assign each byte position in the stores that we are going to | |
2772 | analyze globally to a position in the bitmaps. Returns true if | |
6dfdc153 | 2773 | there are any bit positions assigned. |
3072d30e | 2774 | ----------------------------------------------------------------------------*/ |
2775 | ||
2776 | static void | |
2777 | dse_step2_init (void) | |
2778 | { | |
2779 | unsigned int i; | |
2780 | group_info_t group; | |
2781 | ||
2782 | for (i = 0; VEC_iterate (group_info_t, rtx_group_vec, i, group); i++) | |
2783 | { | |
2784 | /* For all non stack related bases, we only consider a store to | |
2785 | be deletable if there are two or more stores for that | |
2786 | position. This is because it takes one store to make the | |
2787 | other store redundant. However, for the stores that are | |
2788 | stack related, we consider them if there is only one store | |
2789 | for the position. We do this because the stack related | |
2790 | stores can be deleted if their is no read between them and | |
2791 | the end of the function. | |
48e1416a | 2792 | |
3072d30e | 2793 | To make this work in the current framework, we take the stack |
2794 | related bases add all of the bits from store1 into store2. | |
2795 | This has the effect of making the eligible even if there is | |
2796 | only one store. */ | |
2797 | ||
2798 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2799 | { | |
2800 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2801 | bitmap_ior_into (group->store2_p, group->store1_p); | |
2802 | if (dump_file) | |
48e1416a | 2803 | fprintf (dump_file, "group %d is frame related ", i); |
3072d30e | 2804 | } |
2805 | ||
2806 | group->offset_map_size_n++; | |
2807 | group->offset_map_n = XNEWVEC (int, group->offset_map_size_n); | |
2808 | group->offset_map_size_p++; | |
2809 | group->offset_map_p = XNEWVEC (int, group->offset_map_size_p); | |
2810 | group->process_globally = false; | |
2811 | if (dump_file) | |
2812 | { | |
48e1416a | 2813 | fprintf (dump_file, "group %d(%d+%d): ", i, |
3072d30e | 2814 | (int)bitmap_count_bits (group->store2_n), |
2815 | (int)bitmap_count_bits (group->store2_p)); | |
2816 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2817 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2818 | } | |
2819 | } | |
2820 | } | |
2821 | ||
2822 | ||
2823 | /* Init the offset tables for the normal case. */ | |
2824 | ||
2825 | static bool | |
2826 | dse_step2_nospill (void) | |
2827 | { | |
2828 | unsigned int i; | |
2829 | group_info_t group; | |
2830 | /* Position 0 is unused because 0 is used in the maps to mean | |
2831 | unused. */ | |
2832 | current_position = 1; | |
2833 | ||
2834 | for (i = 0; VEC_iterate (group_info_t, rtx_group_vec, i, group); i++) | |
2835 | { | |
2836 | bitmap_iterator bi; | |
2837 | unsigned int j; | |
2838 | ||
2839 | if (group == clear_alias_group) | |
2840 | continue; | |
2841 | ||
2842 | memset (group->offset_map_n, 0, sizeof(int) * group->offset_map_size_n); | |
2843 | memset (group->offset_map_p, 0, sizeof(int) * group->offset_map_size_p); | |
2844 | bitmap_clear (group->group_kill); | |
2845 | ||
2846 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2847 | { | |
2848 | bitmap_set_bit (group->group_kill, current_position); | |
2849 | group->offset_map_n[j] = current_position++; | |
2850 | group->process_globally = true; | |
2851 | } | |
2852 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2853 | { | |
48e1416a | 2854 | bitmap_set_bit (group->group_kill, current_position); |
3072d30e | 2855 | group->offset_map_p[j] = current_position++; |
2856 | group->process_globally = true; | |
2857 | } | |
2858 | } | |
2859 | return current_position != 1; | |
2860 | } | |
2861 | ||
2862 | ||
2863 | /* Init the offset tables for the spill case. */ | |
2864 | ||
2865 | static bool | |
2866 | dse_step2_spill (void) | |
2867 | { | |
2868 | unsigned int j; | |
2869 | group_info_t group = clear_alias_group; | |
2870 | bitmap_iterator bi; | |
2871 | ||
2872 | /* Position 0 is unused because 0 is used in the maps to mean | |
2873 | unused. */ | |
2874 | current_position = 1; | |
2875 | ||
2876 | if (dump_file) | |
2877 | { | |
48e1416a | 2878 | bitmap_print (dump_file, clear_alias_sets, |
3072d30e | 2879 | "clear alias sets ", "\n"); |
48e1416a | 2880 | bitmap_print (dump_file, disqualified_clear_alias_sets, |
3072d30e | 2881 | "disqualified clear alias sets ", "\n"); |
2882 | } | |
2883 | ||
2884 | memset (group->offset_map_n, 0, sizeof(int) * group->offset_map_size_n); | |
2885 | memset (group->offset_map_p, 0, sizeof(int) * group->offset_map_size_p); | |
2886 | bitmap_clear (group->group_kill); | |
48e1416a | 2887 | |
3072d30e | 2888 | /* Remove the disqualified positions from the store2_p set. */ |
2889 | bitmap_and_compl_into (group->store2_p, disqualified_clear_alias_sets); | |
48e1416a | 2890 | |
3072d30e | 2891 | /* We do not need to process the store2_n set because |
2892 | alias_sets are always positive. */ | |
2893 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
2894 | { | |
48e1416a | 2895 | bitmap_set_bit (group->group_kill, current_position); |
3072d30e | 2896 | group->offset_map_p[j] = current_position++; |
2897 | group->process_globally = true; | |
2898 | } | |
2899 | ||
2900 | return current_position != 1; | |
2901 | } | |
2902 | ||
2903 | ||
2904 | \f | |
2905 | /*---------------------------------------------------------------------------- | |
2906 | Third step. | |
48e1416a | 2907 | |
3072d30e | 2908 | Build the bit vectors for the transfer functions. |
2909 | ----------------------------------------------------------------------------*/ | |
2910 | ||
2911 | ||
2912 | /* Note that this is NOT a general purpose function. Any mem that has | |
2913 | an alias set registered here expected to be COMPLETELY unaliased: | |
48e1416a | 2914 | i.e it's addresses are not and need not be examined. |
3072d30e | 2915 | |
2916 | It is known that all references to this address will have this | |
2917 | alias set and there are NO other references to this address in the | |
48e1416a | 2918 | function. |
3072d30e | 2919 | |
2920 | Currently the only place that is known to be clean enough to use | |
48e1416a | 2921 | this interface is the code that assigns the spill locations. |
3072d30e | 2922 | |
2923 | All of the mems that have alias_sets registered are subjected to a | |
2924 | very powerful form of dse where function calls, volatile reads and | |
48e1416a | 2925 | writes, and reads from random location are not taken into account. |
3072d30e | 2926 | |
2927 | It is also assumed that these locations go dead when the function | |
2928 | returns. This assumption could be relaxed if there were found to | |
2929 | be places that this assumption was not correct. | |
2930 | ||
2931 | The MODE is passed in and saved. The mode of each load or store to | |
2932 | a mem with ALIAS_SET is checked against MEM. If the size of that | |
2933 | load or store is different from MODE, processing is halted on this | |
2934 | alias set. For the vast majority of aliases sets, all of the loads | |
2935 | and stores will use the same mode. But vectors are treated | |
2936 | differently: the alias set is established for the entire vector, | |
2937 | but reload will insert loads and stores for individual elements and | |
2938 | we do not necessarily have the information to track those separate | |
2939 | elements. So when we see a mode mismatch, we just bail. */ | |
2940 | ||
2941 | ||
48e1416a | 2942 | void |
2943 | dse_record_singleton_alias_set (alias_set_type alias_set, | |
3072d30e | 2944 | enum machine_mode mode) |
2945 | { | |
2946 | struct clear_alias_mode_holder tmp_holder; | |
2947 | struct clear_alias_mode_holder *entry; | |
2948 | void **slot; | |
2949 | ||
2950 | /* If we are not going to run dse, we need to return now or there | |
2951 | will be problems with allocating the bitmaps. */ | |
2952 | if ((!gate_dse()) || !alias_set) | |
2953 | return; | |
2954 | ||
2955 | if (!clear_alias_sets) | |
2956 | { | |
2957 | clear_alias_sets = BITMAP_ALLOC (NULL); | |
2958 | disqualified_clear_alias_sets = BITMAP_ALLOC (NULL); | |
2959 | clear_alias_mode_table = htab_create (11, clear_alias_mode_hash, | |
2960 | clear_alias_mode_eq, NULL); | |
48e1416a | 2961 | clear_alias_mode_pool = create_alloc_pool ("clear_alias_mode_pool", |
3072d30e | 2962 | sizeof (struct clear_alias_mode_holder), 100); |
2963 | } | |
2964 | ||
2965 | bitmap_set_bit (clear_alias_sets, alias_set); | |
2966 | ||
2967 | tmp_holder.alias_set = alias_set; | |
2968 | ||
2969 | slot = htab_find_slot (clear_alias_mode_table, &tmp_holder, INSERT); | |
2970 | gcc_assert (*slot == NULL); | |
2971 | ||
364c0c59 | 2972 | *slot = entry = |
2973 | (struct clear_alias_mode_holder *) pool_alloc (clear_alias_mode_pool); | |
3072d30e | 2974 | entry->alias_set = alias_set; |
2975 | entry->mode = mode; | |
2976 | } | |
2977 | ||
2978 | ||
2979 | /* Remove ALIAS_SET from the sets of stack slots being considered. */ | |
2980 | ||
48e1416a | 2981 | void |
32c2fdea | 2982 | dse_invalidate_singleton_alias_set (alias_set_type alias_set) |
3072d30e | 2983 | { |
2984 | if ((!gate_dse()) || !alias_set) | |
2985 | return; | |
2986 | ||
2987 | bitmap_clear_bit (clear_alias_sets, alias_set); | |
2988 | } | |
2989 | ||
2990 | ||
2991 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not | |
2992 | there, return 0. */ | |
2993 | ||
2994 | static int | |
2995 | get_bitmap_index (group_info_t group_info, HOST_WIDE_INT offset) | |
2996 | { | |
2997 | if (offset < 0) | |
2998 | { | |
2999 | HOST_WIDE_INT offset_p = -offset; | |
3000 | if (offset_p >= group_info->offset_map_size_n) | |
3001 | return 0; | |
3002 | return group_info->offset_map_n[offset_p]; | |
3003 | } | |
3004 | else | |
3005 | { | |
3006 | if (offset >= group_info->offset_map_size_p) | |
3007 | return 0; | |
3008 | return group_info->offset_map_p[offset]; | |
3009 | } | |
3010 | } | |
3011 | ||
3012 | ||
3013 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
3014 | may be NULL. */ | |
3015 | ||
48e1416a | 3016 | static void |
3072d30e | 3017 | scan_stores_nospill (store_info_t store_info, bitmap gen, bitmap kill) |
3018 | { | |
3019 | while (store_info) | |
3020 | { | |
3021 | HOST_WIDE_INT i; | |
48e1416a | 3022 | group_info_t group_info |
3072d30e | 3023 | = VEC_index (group_info_t, rtx_group_vec, store_info->group_id); |
3024 | if (group_info->process_globally) | |
3025 | for (i = store_info->begin; i < store_info->end; i++) | |
3026 | { | |
3027 | int index = get_bitmap_index (group_info, i); | |
3028 | if (index != 0) | |
3029 | { | |
3030 | bitmap_set_bit (gen, index); | |
3031 | if (kill) | |
3032 | bitmap_clear_bit (kill, index); | |
3033 | } | |
3034 | } | |
3035 | store_info = store_info->next; | |
3036 | } | |
3037 | } | |
3038 | ||
3039 | ||
3040 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
3041 | may be NULL. */ | |
3042 | ||
48e1416a | 3043 | static void |
3072d30e | 3044 | scan_stores_spill (store_info_t store_info, bitmap gen, bitmap kill) |
3045 | { | |
3046 | while (store_info) | |
3047 | { | |
3048 | if (store_info->alias_set) | |
3049 | { | |
48e1416a | 3050 | int index = get_bitmap_index (clear_alias_group, |
3072d30e | 3051 | store_info->alias_set); |
3052 | if (index != 0) | |
3053 | { | |
3054 | bitmap_set_bit (gen, index); | |
3055 | if (kill) | |
3056 | bitmap_clear_bit (kill, index); | |
3057 | } | |
3058 | } | |
3059 | store_info = store_info->next; | |
3060 | } | |
3061 | } | |
3062 | ||
3063 | ||
3064 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3065 | may be NULL. */ | |
3066 | ||
3067 | static void | |
3068 | scan_reads_nospill (insn_info_t insn_info, bitmap gen, bitmap kill) | |
3069 | { | |
3070 | read_info_t read_info = insn_info->read_rec; | |
3071 | int i; | |
3072 | group_info_t group; | |
3073 | ||
16bf64db | 3074 | /* If this insn reads the frame, kill all the frame related stores. */ |
3075 | if (insn_info->frame_read) | |
3076 | { | |
3077 | for (i = 0; VEC_iterate (group_info_t, rtx_group_vec, i, group); i++) | |
3078 | if (group->process_globally && group->frame_related) | |
3079 | { | |
3080 | if (kill) | |
3081 | bitmap_ior_into (kill, group->group_kill); | |
48e1416a | 3082 | bitmap_and_compl_into (gen, group->group_kill); |
16bf64db | 3083 | } |
3084 | } | |
3085 | ||
3072d30e | 3086 | while (read_info) |
3087 | { | |
3088 | for (i = 0; VEC_iterate (group_info_t, rtx_group_vec, i, group); i++) | |
3089 | { | |
3090 | if (group->process_globally) | |
3091 | { | |
3092 | if (i == read_info->group_id) | |
3093 | { | |
3094 | if (read_info->begin > read_info->end) | |
3095 | { | |
3096 | /* Begin > end for block mode reads. */ | |
3097 | if (kill) | |
3098 | bitmap_ior_into (kill, group->group_kill); | |
3099 | bitmap_and_compl_into (gen, group->group_kill); | |
3100 | } | |
3101 | else | |
3102 | { | |
3103 | /* The groups are the same, just process the | |
3104 | offsets. */ | |
3105 | HOST_WIDE_INT j; | |
3106 | for (j = read_info->begin; j < read_info->end; j++) | |
3107 | { | |
3108 | int index = get_bitmap_index (group, j); | |
3109 | if (index != 0) | |
3110 | { | |
3111 | if (kill) | |
3112 | bitmap_set_bit (kill, index); | |
3113 | bitmap_clear_bit (gen, index); | |
3114 | } | |
3115 | } | |
3116 | } | |
3117 | } | |
3118 | else | |
3119 | { | |
3120 | /* The groups are different, if the alias sets | |
3121 | conflict, clear the entire group. We only need | |
3122 | to apply this test if the read_info is a cselib | |
3123 | read. Anything with a constant base cannot alias | |
3124 | something else with a different constant | |
3125 | base. */ | |
3126 | if ((read_info->group_id < 0) | |
48e1416a | 3127 | && canon_true_dependence (group->base_mem, |
3072d30e | 3128 | QImode, |
82d2c88b | 3129 | group->canon_base_addr, |
3130 | read_info->mem, NULL_RTX, | |
3131 | rtx_varies_p)) | |
3072d30e | 3132 | { |
3133 | if (kill) | |
3134 | bitmap_ior_into (kill, group->group_kill); | |
3135 | bitmap_and_compl_into (gen, group->group_kill); | |
3136 | } | |
3137 | } | |
3138 | } | |
3139 | } | |
48e1416a | 3140 | |
3072d30e | 3141 | read_info = read_info->next; |
3142 | } | |
3143 | } | |
3144 | ||
3145 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3146 | may be NULL. */ | |
3147 | ||
3148 | static void | |
3149 | scan_reads_spill (read_info_t read_info, bitmap gen, bitmap kill) | |
3150 | { | |
3151 | while (read_info) | |
3152 | { | |
3153 | if (read_info->alias_set) | |
3154 | { | |
48e1416a | 3155 | int index = get_bitmap_index (clear_alias_group, |
3072d30e | 3156 | read_info->alias_set); |
3157 | if (index != 0) | |
3158 | { | |
3159 | if (kill) | |
3160 | bitmap_set_bit (kill, index); | |
3161 | bitmap_clear_bit (gen, index); | |
3162 | } | |
3163 | } | |
48e1416a | 3164 | |
3072d30e | 3165 | read_info = read_info->next; |
3166 | } | |
3167 | } | |
3168 | ||
3169 | ||
3170 | /* Return the insn in BB_INFO before the first wild read or if there | |
3171 | are no wild reads in the block, return the last insn. */ | |
3172 | ||
3173 | static insn_info_t | |
3174 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3175 | { | |
3176 | insn_info_t insn_info = bb_info->last_insn; | |
3177 | insn_info_t last_wild_read = NULL; | |
3178 | ||
3179 | while (insn_info) | |
3180 | { | |
3181 | if (insn_info->wild_read) | |
3182 | { | |
3183 | last_wild_read = insn_info->prev_insn; | |
3184 | /* Block starts with wild read. */ | |
3185 | if (!last_wild_read) | |
3186 | return NULL; | |
3187 | } | |
3188 | ||
3189 | insn_info = insn_info->prev_insn; | |
3190 | } | |
3191 | ||
3192 | if (last_wild_read) | |
3193 | return last_wild_read; | |
3194 | else | |
3195 | return bb_info->last_insn; | |
3196 | } | |
3197 | ||
3198 | ||
3199 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3200 | the block in order to build the gen and kill sets for the block. | |
3201 | We start at ptr which may be the last insn in the block or may be | |
3202 | the first insn with a wild read. In the latter case we are able to | |
3203 | skip the rest of the block because it just does not matter: | |
3204 | anything that happens is hidden by the wild read. */ | |
3205 | ||
3206 | static void | |
3207 | dse_step3_scan (bool for_spills, basic_block bb) | |
3208 | { | |
3209 | bb_info_t bb_info = bb_table[bb->index]; | |
3210 | insn_info_t insn_info; | |
3211 | ||
3212 | if (for_spills) | |
3213 | /* There are no wild reads in the spill case. */ | |
3214 | insn_info = bb_info->last_insn; | |
3215 | else | |
3216 | insn_info = find_insn_before_first_wild_read (bb_info); | |
48e1416a | 3217 | |
3072d30e | 3218 | /* In the spill case or in the no_spill case if there is no wild |
3219 | read in the block, we will need a kill set. */ | |
3220 | if (insn_info == bb_info->last_insn) | |
3221 | { | |
3222 | if (bb_info->kill) | |
3223 | bitmap_clear (bb_info->kill); | |
3224 | else | |
3225 | bb_info->kill = BITMAP_ALLOC (NULL); | |
3226 | } | |
48e1416a | 3227 | else |
3072d30e | 3228 | if (bb_info->kill) |
3229 | BITMAP_FREE (bb_info->kill); | |
3230 | ||
3231 | while (insn_info) | |
3232 | { | |
3233 | /* There may have been code deleted by the dce pass run before | |
3234 | this phase. */ | |
3235 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3236 | { | |
48e1416a | 3237 | /* Process the read(s) last. */ |
3072d30e | 3238 | if (for_spills) |
3239 | { | |
3240 | scan_stores_spill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3241 | scan_reads_spill (insn_info->read_rec, bb_info->gen, bb_info->kill); | |
3242 | } | |
3243 | else | |
3244 | { | |
3245 | scan_stores_nospill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3246 | scan_reads_nospill (insn_info, bb_info->gen, bb_info->kill); | |
3247 | } | |
48e1416a | 3248 | } |
3072d30e | 3249 | |
3250 | insn_info = insn_info->prev_insn; | |
3251 | } | |
3252 | } | |
3253 | ||
3254 | ||
3255 | /* Set the gen set of the exit block, and also any block with no | |
3256 | successors that does not have a wild read. */ | |
3257 | ||
3258 | static void | |
3259 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3260 | { | |
3261 | /* The gen set is all 0's for the exit block except for the | |
3262 | frame_pointer_group. */ | |
48e1416a | 3263 | |
3072d30e | 3264 | if (stores_off_frame_dead_at_return) |
3265 | { | |
3266 | unsigned int i; | |
3267 | group_info_t group; | |
48e1416a | 3268 | |
3072d30e | 3269 | for (i = 0; VEC_iterate (group_info_t, rtx_group_vec, i, group); i++) |
3270 | { | |
3271 | if (group->process_globally && group->frame_related) | |
3272 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3273 | } | |
3274 | } | |
3275 | } | |
3276 | ||
3277 | ||
3278 | /* Find all of the blocks that are not backwards reachable from the | |
3279 | exit block or any block with no successors (BB). These are the | |
3280 | infinite loops or infinite self loops. These blocks will still | |
3281 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3282 | ||
3283 | static void | |
3284 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3285 | { | |
3286 | edge e; | |
3287 | edge_iterator ei; | |
3288 | ||
3289 | if (TEST_BIT (unreachable_blocks, bb->index)) | |
3290 | { | |
3291 | RESET_BIT (unreachable_blocks, bb->index); | |
3292 | FOR_EACH_EDGE (e, ei, bb->preds) | |
48e1416a | 3293 | { |
3072d30e | 3294 | mark_reachable_blocks (unreachable_blocks, e->src); |
48e1416a | 3295 | } |
3072d30e | 3296 | } |
3297 | } | |
3298 | ||
3299 | /* Build the transfer functions for the function. */ | |
3300 | ||
3301 | static void | |
3302 | dse_step3 (bool for_spills) | |
3303 | { | |
3304 | basic_block bb; | |
3305 | sbitmap unreachable_blocks = sbitmap_alloc (last_basic_block); | |
3306 | sbitmap_iterator sbi; | |
3307 | bitmap all_ones = NULL; | |
3308 | unsigned int i; | |
48e1416a | 3309 | |
3072d30e | 3310 | sbitmap_ones (unreachable_blocks); |
3311 | ||
3312 | FOR_ALL_BB (bb) | |
3313 | { | |
3314 | bb_info_t bb_info = bb_table[bb->index]; | |
3315 | if (bb_info->gen) | |
3316 | bitmap_clear (bb_info->gen); | |
3317 | else | |
3318 | bb_info->gen = BITMAP_ALLOC (NULL); | |
3319 | ||
3320 | if (bb->index == ENTRY_BLOCK) | |
3321 | ; | |
3322 | else if (bb->index == EXIT_BLOCK) | |
3323 | dse_step3_exit_block_scan (bb_info); | |
3324 | else | |
3325 | dse_step3_scan (for_spills, bb); | |
3326 | if (EDGE_COUNT (bb->succs) == 0) | |
3327 | mark_reachable_blocks (unreachable_blocks, bb); | |
3328 | ||
3329 | /* If this is the second time dataflow is run, delete the old | |
3330 | sets. */ | |
3331 | if (bb_info->in) | |
3332 | BITMAP_FREE (bb_info->in); | |
3333 | if (bb_info->out) | |
3334 | BITMAP_FREE (bb_info->out); | |
3335 | } | |
3336 | ||
3337 | /* For any block in an infinite loop, we must initialize the out set | |
3338 | to all ones. This could be expensive, but almost never occurs in | |
3339 | practice. However, it is common in regression tests. */ | |
3340 | EXECUTE_IF_SET_IN_SBITMAP (unreachable_blocks, 0, i, sbi) | |
3341 | { | |
3342 | if (bitmap_bit_p (all_blocks, i)) | |
3343 | { | |
3344 | bb_info_t bb_info = bb_table[i]; | |
3345 | if (!all_ones) | |
3346 | { | |
3347 | unsigned int j; | |
3348 | group_info_t group; | |
3349 | ||
3350 | all_ones = BITMAP_ALLOC (NULL); | |
3351 | for (j = 0; VEC_iterate (group_info_t, rtx_group_vec, j, group); j++) | |
3352 | bitmap_ior_into (all_ones, group->group_kill); | |
3353 | } | |
3354 | if (!bb_info->out) | |
3355 | { | |
3356 | bb_info->out = BITMAP_ALLOC (NULL); | |
3357 | bitmap_copy (bb_info->out, all_ones); | |
3358 | } | |
3359 | } | |
3360 | } | |
3361 | ||
3362 | if (all_ones) | |
3363 | BITMAP_FREE (all_ones); | |
3364 | sbitmap_free (unreachable_blocks); | |
3365 | } | |
3366 | ||
3367 | ||
3368 | \f | |
3369 | /*---------------------------------------------------------------------------- | |
3370 | Fourth step. | |
3371 | ||
3372 | Solve the bitvector equations. | |
3373 | ----------------------------------------------------------------------------*/ | |
3374 | ||
3375 | ||
3376 | /* Confluence function for blocks with no successors. Create an out | |
3377 | set from the gen set of the exit block. This block logically has | |
3378 | the exit block as a successor. */ | |
3379 | ||
3380 | ||
3381 | ||
3382 | static void | |
3383 | dse_confluence_0 (basic_block bb) | |
3384 | { | |
3385 | bb_info_t bb_info = bb_table[bb->index]; | |
3386 | ||
3387 | if (bb->index == EXIT_BLOCK) | |
3388 | return; | |
3389 | ||
3390 | if (!bb_info->out) | |
3391 | { | |
3392 | bb_info->out = BITMAP_ALLOC (NULL); | |
3393 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); | |
3394 | } | |
3395 | } | |
3396 | ||
3397 | /* Propagate the information from the in set of the dest of E to the | |
3398 | out set of the src of E. If the various in or out sets are not | |
3399 | there, that means they are all ones. */ | |
3400 | ||
3401 | static void | |
3402 | dse_confluence_n (edge e) | |
3403 | { | |
3404 | bb_info_t src_info = bb_table[e->src->index]; | |
3405 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3406 | ||
3407 | if (dest_info->in) | |
3408 | { | |
3409 | if (src_info->out) | |
3410 | bitmap_and_into (src_info->out, dest_info->in); | |
3411 | else | |
3412 | { | |
3413 | src_info->out = BITMAP_ALLOC (NULL); | |
3414 | bitmap_copy (src_info->out, dest_info->in); | |
3415 | } | |
3416 | } | |
3417 | } | |
3418 | ||
3419 | ||
3420 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
48e1416a | 3421 | block. There are three cases: |
3072d30e | 3422 | |
3423 | 1) The block has no kill set. In this case the kill set is all | |
3424 | ones. It does not matter what the out set of the block is, none of | |
3425 | the info can reach the top. The only thing that reaches the top is | |
3426 | the gen set and we just copy the set. | |
3427 | ||
3428 | 2) There is a kill set but no out set and bb has successors. In | |
3429 | this case we just return. Eventually an out set will be created and | |
3430 | it is better to wait than to create a set of ones. | |
3431 | ||
3432 | 3) There is both a kill and out set. We apply the obvious transfer | |
3433 | function. | |
3434 | */ | |
3435 | ||
3436 | static bool | |
3437 | dse_transfer_function (int bb_index) | |
3438 | { | |
3439 | bb_info_t bb_info = bb_table[bb_index]; | |
3440 | ||
3441 | if (bb_info->kill) | |
3442 | { | |
3443 | if (bb_info->out) | |
3444 | { | |
3445 | /* Case 3 above. */ | |
3446 | if (bb_info->in) | |
48e1416a | 3447 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
3072d30e | 3448 | bb_info->out, bb_info->kill); |
3449 | else | |
3450 | { | |
3451 | bb_info->in = BITMAP_ALLOC (NULL); | |
48e1416a | 3452 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
3072d30e | 3453 | bb_info->out, bb_info->kill); |
3454 | return true; | |
3455 | } | |
3456 | } | |
3457 | else | |
3458 | /* Case 2 above. */ | |
3459 | return false; | |
3460 | } | |
3461 | else | |
3462 | { | |
3463 | /* Case 1 above. If there is already an in set, nothing | |
3464 | happens. */ | |
3465 | if (bb_info->in) | |
3466 | return false; | |
3467 | else | |
3468 | { | |
3469 | bb_info->in = BITMAP_ALLOC (NULL); | |
3470 | bitmap_copy (bb_info->in, bb_info->gen); | |
3471 | return true; | |
3472 | } | |
3473 | } | |
3474 | } | |
3475 | ||
3476 | /* Solve the dataflow equations. */ | |
3477 | ||
3478 | static void | |
3479 | dse_step4 (void) | |
3480 | { | |
48e1416a | 3481 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3482 | dse_confluence_n, dse_transfer_function, | |
3483 | all_blocks, df_get_postorder (DF_BACKWARD), | |
3072d30e | 3484 | df_get_n_blocks (DF_BACKWARD)); |
3485 | if (dump_file) | |
3486 | { | |
3487 | basic_block bb; | |
3488 | ||
3489 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
3490 | FOR_ALL_BB (bb) | |
3491 | { | |
3492 | bb_info_t bb_info = bb_table[bb->index]; | |
3493 | ||
3494 | df_print_bb_index (bb, dump_file); | |
3495 | if (bb_info->in) | |
3496 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3497 | else | |
3498 | fprintf (dump_file, " in: *MISSING*\n"); | |
3499 | if (bb_info->gen) | |
3500 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3501 | else | |
3502 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3503 | if (bb_info->kill) | |
3504 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3505 | else | |
3506 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3507 | if (bb_info->out) | |
3508 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3509 | else | |
3510 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3511 | } | |
3512 | } | |
3513 | } | |
3514 | ||
3515 | ||
3516 | \f | |
3517 | /*---------------------------------------------------------------------------- | |
3518 | Fifth step. | |
3519 | ||
bef304b8 | 3520 | Delete the stores that can only be deleted using the global information. |
3072d30e | 3521 | ----------------------------------------------------------------------------*/ |
3522 | ||
3523 | ||
3524 | static void | |
3525 | dse_step5_nospill (void) | |
3526 | { | |
3527 | basic_block bb; | |
3528 | FOR_EACH_BB (bb) | |
3529 | { | |
3530 | bb_info_t bb_info = bb_table[bb->index]; | |
3531 | insn_info_t insn_info = bb_info->last_insn; | |
3532 | bitmap v = bb_info->out; | |
3533 | ||
3534 | while (insn_info) | |
3535 | { | |
3536 | bool deleted = false; | |
3537 | if (dump_file && insn_info->insn) | |
3538 | { | |
3539 | fprintf (dump_file, "starting to process insn %d\n", | |
3540 | INSN_UID (insn_info->insn)); | |
3541 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3542 | } | |
3543 | ||
3544 | /* There may have been code deleted by the dce pass run before | |
3545 | this phase. */ | |
48e1416a | 3546 | if (insn_info->insn |
3072d30e | 3547 | && INSN_P (insn_info->insn) |
3548 | && (!insn_info->cannot_delete) | |
3549 | && (!bitmap_empty_p (v))) | |
3550 | { | |
3551 | store_info_t store_info = insn_info->store_rec; | |
3552 | ||
3553 | /* Try to delete the current insn. */ | |
3554 | deleted = true; | |
48e1416a | 3555 | |
3072d30e | 3556 | /* Skip the clobbers. */ |
3557 | while (!store_info->is_set) | |
3558 | store_info = store_info->next; | |
3559 | ||
3560 | if (store_info->alias_set) | |
3561 | deleted = false; | |
3562 | else | |
3563 | { | |
3564 | HOST_WIDE_INT i; | |
48e1416a | 3565 | group_info_t group_info |
3072d30e | 3566 | = VEC_index (group_info_t, rtx_group_vec, store_info->group_id); |
48e1416a | 3567 | |
3072d30e | 3568 | for (i = store_info->begin; i < store_info->end; i++) |
3569 | { | |
3570 | int index = get_bitmap_index (group_info, i); | |
48e1416a | 3571 | |
3072d30e | 3572 | if (dump_file) |
48e1416a | 3573 | fprintf (dump_file, "i = %d, index = %d\n", (int)i, index); |
3072d30e | 3574 | if (index == 0 || !bitmap_bit_p (v, index)) |
3575 | { | |
3576 | if (dump_file) | |
48e1416a | 3577 | fprintf (dump_file, "failing at i = %d\n", (int)i); |
3072d30e | 3578 | deleted = false; |
3579 | break; | |
3580 | } | |
3581 | } | |
3582 | } | |
3583 | if (deleted) | |
3584 | { | |
3585 | if (dbg_cnt (dse)) | |
3586 | { | |
3587 | check_for_inc_dec (insn_info->insn); | |
3588 | delete_insn (insn_info->insn); | |
3589 | insn_info->insn = NULL; | |
3590 | globally_deleted++; | |
3591 | } | |
3592 | } | |
3593 | } | |
3594 | /* We do want to process the local info if the insn was | |
6dfdc153 | 3595 | deleted. For instance, if the insn did a wild read, we |
3072d30e | 3596 | no longer need to trash the info. */ |
48e1416a | 3597 | if (insn_info->insn |
3072d30e | 3598 | && INSN_P (insn_info->insn) |
3599 | && (!deleted)) | |
3600 | { | |
3601 | scan_stores_nospill (insn_info->store_rec, v, NULL); | |
3602 | if (insn_info->wild_read) | |
3603 | { | |
3604 | if (dump_file) | |
3605 | fprintf (dump_file, "wild read\n"); | |
3606 | bitmap_clear (v); | |
3607 | } | |
3608 | else if (insn_info->read_rec) | |
3609 | { | |
3610 | if (dump_file) | |
3611 | fprintf (dump_file, "regular read\n"); | |
3612 | scan_reads_nospill (insn_info, v, NULL); | |
3613 | } | |
3614 | } | |
48e1416a | 3615 | |
3072d30e | 3616 | insn_info = insn_info->prev_insn; |
3617 | } | |
3618 | } | |
3619 | } | |
3620 | ||
3621 | ||
3622 | static void | |
3623 | dse_step5_spill (void) | |
3624 | { | |
3625 | basic_block bb; | |
3626 | FOR_EACH_BB (bb) | |
3627 | { | |
3628 | bb_info_t bb_info = bb_table[bb->index]; | |
3629 | insn_info_t insn_info = bb_info->last_insn; | |
3630 | bitmap v = bb_info->out; | |
3631 | ||
3632 | while (insn_info) | |
3633 | { | |
3634 | bool deleted = false; | |
3635 | /* There may have been code deleted by the dce pass run before | |
3636 | this phase. */ | |
48e1416a | 3637 | if (insn_info->insn |
3072d30e | 3638 | && INSN_P (insn_info->insn) |
3639 | && (!insn_info->cannot_delete) | |
3640 | && (!bitmap_empty_p (v))) | |
3641 | { | |
3642 | /* Try to delete the current insn. */ | |
3643 | store_info_t store_info = insn_info->store_rec; | |
3644 | deleted = true; | |
48e1416a | 3645 | |
3072d30e | 3646 | while (store_info) |
3647 | { | |
3648 | if (store_info->alias_set) | |
3649 | { | |
48e1416a | 3650 | int index = get_bitmap_index (clear_alias_group, |
3072d30e | 3651 | store_info->alias_set); |
3652 | if (index == 0 || !bitmap_bit_p (v, index)) | |
3653 | { | |
3654 | deleted = false; | |
3655 | break; | |
3656 | } | |
3657 | } | |
48e1416a | 3658 | else |
3072d30e | 3659 | deleted = false; |
3660 | store_info = store_info->next; | |
3661 | } | |
3662 | if (deleted && dbg_cnt (dse)) | |
3663 | { | |
3664 | if (dump_file) | |
48e1416a | 3665 | fprintf (dump_file, "Spill deleting insn %d\n", |
3072d30e | 3666 | INSN_UID (insn_info->insn)); |
3667 | check_for_inc_dec (insn_info->insn); | |
3668 | delete_insn (insn_info->insn); | |
3669 | spill_deleted++; | |
3670 | insn_info->insn = NULL; | |
3671 | } | |
3672 | } | |
48e1416a | 3673 | |
3674 | if (insn_info->insn | |
3072d30e | 3675 | && INSN_P (insn_info->insn) |
3676 | && (!deleted)) | |
3677 | { | |
3678 | scan_stores_spill (insn_info->store_rec, v, NULL); | |
3679 | scan_reads_spill (insn_info->read_rec, v, NULL); | |
3680 | } | |
48e1416a | 3681 | |
3072d30e | 3682 | insn_info = insn_info->prev_insn; |
3683 | } | |
3684 | } | |
3685 | } | |
3686 | ||
3687 | ||
3688 | \f | |
3689 | /*---------------------------------------------------------------------------- | |
3690 | Sixth step. | |
3691 | ||
aa140b76 | 3692 | Delete stores made redundant by earlier stores (which store the same |
3693 | value) that couldn't be eliminated. | |
3694 | ----------------------------------------------------------------------------*/ | |
3695 | ||
3696 | static void | |
3697 | dse_step6 (void) | |
3698 | { | |
3699 | basic_block bb; | |
3700 | ||
3701 | FOR_ALL_BB (bb) | |
3702 | { | |
3703 | bb_info_t bb_info = bb_table[bb->index]; | |
3704 | insn_info_t insn_info = bb_info->last_insn; | |
3705 | ||
3706 | while (insn_info) | |
3707 | { | |
3708 | /* There may have been code deleted by the dce pass run before | |
3709 | this phase. */ | |
3710 | if (insn_info->insn | |
3711 | && INSN_P (insn_info->insn) | |
3712 | && !insn_info->cannot_delete) | |
3713 | { | |
3714 | store_info_t s_info = insn_info->store_rec; | |
3715 | ||
3716 | while (s_info && !s_info->is_set) | |
3717 | s_info = s_info->next; | |
3718 | if (s_info | |
3719 | && s_info->redundant_reason | |
3720 | && s_info->redundant_reason->insn | |
3721 | && INSN_P (s_info->redundant_reason->insn)) | |
3722 | { | |
3723 | rtx rinsn = s_info->redundant_reason->insn; | |
3724 | if (dump_file) | |
3725 | fprintf (dump_file, "Locally deleting insn %d " | |
3726 | "because insn %d stores the " | |
3727 | "same value and couldn't be " | |
3728 | "eliminated\n", | |
3729 | INSN_UID (insn_info->insn), | |
3730 | INSN_UID (rinsn)); | |
3731 | delete_dead_store_insn (insn_info); | |
3732 | } | |
3733 | } | |
3734 | insn_info = insn_info->prev_insn; | |
3735 | } | |
3736 | } | |
3737 | } | |
3738 | \f | |
3739 | /*---------------------------------------------------------------------------- | |
3740 | Seventh step. | |
3741 | ||
48e1416a | 3742 | Destroy everything left standing. |
3072d30e | 3743 | ----------------------------------------------------------------------------*/ |
3744 | ||
48e1416a | 3745 | static void |
aa140b76 | 3746 | dse_step7 (bool global_done) |
3072d30e | 3747 | { |
3748 | unsigned int i; | |
3749 | group_info_t group; | |
3750 | basic_block bb; | |
48e1416a | 3751 | |
ce299759 | 3752 | for (i = 0; VEC_iterate (group_info_t, rtx_group_vec, i, group); i++) |
3072d30e | 3753 | { |
ce299759 | 3754 | free (group->offset_map_n); |
3755 | free (group->offset_map_p); | |
3756 | BITMAP_FREE (group->store1_n); | |
3757 | BITMAP_FREE (group->store1_p); | |
3758 | BITMAP_FREE (group->store2_n); | |
3759 | BITMAP_FREE (group->store2_p); | |
3760 | BITMAP_FREE (group->group_kill); | |
3072d30e | 3761 | } |
3762 | ||
ce299759 | 3763 | if (global_done) |
3764 | FOR_ALL_BB (bb) | |
3765 | { | |
3766 | bb_info_t bb_info = bb_table[bb->index]; | |
3767 | BITMAP_FREE (bb_info->gen); | |
3768 | if (bb_info->kill) | |
3769 | BITMAP_FREE (bb_info->kill); | |
3770 | if (bb_info->in) | |
3771 | BITMAP_FREE (bb_info->in); | |
3772 | if (bb_info->out) | |
3773 | BITMAP_FREE (bb_info->out); | |
3774 | } | |
3775 | ||
3072d30e | 3776 | if (clear_alias_sets) |
3777 | { | |
3778 | BITMAP_FREE (clear_alias_sets); | |
3779 | BITMAP_FREE (disqualified_clear_alias_sets); | |
3780 | free_alloc_pool (clear_alias_mode_pool); | |
3781 | htab_delete (clear_alias_mode_table); | |
3782 | } | |
3783 | ||
3784 | end_alias_analysis (); | |
3785 | free (bb_table); | |
3786 | htab_delete (rtx_group_table); | |
3787 | VEC_free (group_info_t, heap, rtx_group_vec); | |
3788 | BITMAP_FREE (all_blocks); | |
3789 | BITMAP_FREE (scratch); | |
3790 | ||
3791 | free_alloc_pool (rtx_store_info_pool); | |
3792 | free_alloc_pool (read_info_pool); | |
3793 | free_alloc_pool (insn_info_pool); | |
3794 | free_alloc_pool (bb_info_pool); | |
3795 | free_alloc_pool (rtx_group_info_pool); | |
3796 | free_alloc_pool (deferred_change_pool); | |
3797 | } | |
3798 | ||
3799 | ||
3072d30e | 3800 | /* ------------------------------------------------------------------------- |
3801 | DSE | |
3802 | ------------------------------------------------------------------------- */ | |
3803 | ||
3804 | /* Callback for running pass_rtl_dse. */ | |
3805 | ||
3806 | static unsigned int | |
3807 | rest_of_handle_dse (void) | |
3808 | { | |
3809 | bool did_global = false; | |
3810 | ||
3811 | df_set_flags (DF_DEFER_INSN_RESCAN); | |
3812 | ||
a1b0a968 | 3813 | /* Need the notes since we must track live hardregs in the forwards |
3814 | direction. */ | |
3815 | df_note_add_problem (); | |
3816 | df_analyze (); | |
3817 | ||
3072d30e | 3818 | dse_step0 (); |
3819 | dse_step1 (); | |
3820 | dse_step2_init (); | |
3821 | if (dse_step2_nospill ()) | |
3822 | { | |
3823 | df_set_flags (DF_LR_RUN_DCE); | |
3824 | df_analyze (); | |
3825 | did_global = true; | |
3826 | if (dump_file) | |
3827 | fprintf (dump_file, "doing global processing\n"); | |
3828 | dse_step3 (false); | |
3829 | dse_step4 (); | |
3830 | dse_step5_nospill (); | |
3831 | } | |
3832 | ||
3833 | /* For the instance of dse that runs after reload, we make a special | |
3834 | pass to process the spills. These are special in that they are | |
3835 | totally transparent, i.e, there is no aliasing issues that need | |
3836 | to be considered. This means that the wild reads that kill | |
48e1416a | 3837 | everything else do not apply here. */ |
3072d30e | 3838 | if (clear_alias_sets && dse_step2_spill ()) |
3839 | { | |
3840 | if (!did_global) | |
3841 | { | |
3842 | df_set_flags (DF_LR_RUN_DCE); | |
3843 | df_analyze (); | |
3844 | } | |
3845 | did_global = true; | |
3846 | if (dump_file) | |
3847 | fprintf (dump_file, "doing global spill processing\n"); | |
3848 | dse_step3 (true); | |
3849 | dse_step4 (); | |
3850 | dse_step5_spill (); | |
3851 | } | |
aa140b76 | 3852 | |
3853 | dse_step6 (); | |
3854 | dse_step7 (did_global); | |
3072d30e | 3855 | |
3856 | if (dump_file) | |
3857 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d, spill deletions = %d\n", | |
3858 | locally_deleted, globally_deleted, spill_deleted); | |
3859 | return 0; | |
3860 | } | |
3861 | ||
3862 | static bool | |
3863 | gate_dse (void) | |
3864 | { | |
4ff06051 | 3865 | return gate_dse1 () || gate_dse2 (); |
3866 | } | |
3867 | ||
3868 | static bool | |
3869 | gate_dse1 (void) | |
3870 | { | |
3871 | return optimize > 0 && flag_dse | |
3872 | && dbg_cnt (dse1); | |
3873 | } | |
3874 | ||
3875 | static bool | |
3876 | gate_dse2 (void) | |
3877 | { | |
3878 | return optimize > 0 && flag_dse | |
3879 | && dbg_cnt (dse2); | |
3072d30e | 3880 | } |
3881 | ||
20099e35 | 3882 | struct rtl_opt_pass pass_rtl_dse1 = |
3072d30e | 3883 | { |
20099e35 | 3884 | { |
3885 | RTL_PASS, | |
3072d30e | 3886 | "dse1", /* name */ |
4ff06051 | 3887 | gate_dse1, /* gate */ |
3072d30e | 3888 | rest_of_handle_dse, /* execute */ |
3889 | NULL, /* sub */ | |
3890 | NULL, /* next */ | |
3891 | 0, /* static_pass_number */ | |
3892 | TV_DSE1, /* tv_id */ | |
3893 | 0, /* properties_required */ | |
3894 | 0, /* properties_provided */ | |
3895 | 0, /* properties_destroyed */ | |
3896 | 0, /* todo_flags_start */ | |
3897 | TODO_dump_func | | |
0806b508 | 3898 | TODO_df_finish | TODO_verify_rtl_sharing | |
20099e35 | 3899 | TODO_ggc_collect /* todo_flags_finish */ |
3900 | } | |
3072d30e | 3901 | }; |
3902 | ||
20099e35 | 3903 | struct rtl_opt_pass pass_rtl_dse2 = |
3072d30e | 3904 | { |
20099e35 | 3905 | { |
3906 | RTL_PASS, | |
3072d30e | 3907 | "dse2", /* name */ |
4ff06051 | 3908 | gate_dse2, /* gate */ |
3072d30e | 3909 | rest_of_handle_dse, /* execute */ |
3910 | NULL, /* sub */ | |
3911 | NULL, /* next */ | |
3912 | 0, /* static_pass_number */ | |
3913 | TV_DSE2, /* tv_id */ | |
3914 | 0, /* properties_required */ | |
3915 | 0, /* properties_provided */ | |
3916 | 0, /* properties_destroyed */ | |
3917 | 0, /* todo_flags_start */ | |
3918 | TODO_dump_func | | |
0806b508 | 3919 | TODO_df_finish | TODO_verify_rtl_sharing | |
20099e35 | 3920 | TODO_ggc_collect /* todo_flags_finish */ |
3921 | } | |
3072d30e | 3922 | }; |