]>
Commit | Line | Data |
---|---|---|
6fb5fa3c | 1 | /* RTL dead store elimination. |
5624e564 | 2 | Copyright (C) 2005-2015 Free Software Foundation, Inc. |
6fb5fa3c DB |
3 | |
4 | Contributed by Richard Sandiford <rsandifor@codesourcery.com> | |
5 | and Kenneth Zadeck <zadeck@naturalbridge.com> | |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 11 | Software Foundation; either version 3, or (at your option) any later |
6fb5fa3c DB |
12 | version. |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ | |
6fb5fa3c DB |
22 | |
23 | #undef BASELINE | |
24 | ||
25 | #include "config.h" | |
26 | #include "system.h" | |
27 | #include "coretypes.h" | |
c7131fb2 AM |
28 | #include "backend.h" |
29 | #include "tree.h" | |
30 | #include "gimple.h" | |
6fb5fa3c | 31 | #include "rtl.h" |
c7131fb2 | 32 | #include "df.h" |
40e23961 | 33 | #include "alias.h" |
40e23961 | 34 | #include "fold-const.h" |
d8a2d370 | 35 | #include "stor-layout.h" |
18b526e8 | 36 | #include "tm_p.h" |
6fb5fa3c | 37 | #include "regs.h" |
9e582b1d | 38 | #include "regset.h" |
6fb5fa3c | 39 | #include "flags.h" |
60393bbc | 40 | #include "cfgrtl.h" |
6fb5fa3c | 41 | #include "cselib.h" |
6fb5fa3c DB |
42 | #include "tree-pass.h" |
43 | #include "alloc-pool.h" | |
6fb5fa3c | 44 | #include "insn-config.h" |
36566b39 PK |
45 | #include "expmed.h" |
46 | #include "dojump.h" | |
47 | #include "explow.h" | |
48 | #include "calls.h" | |
49 | #include "emit-rtl.h" | |
50 | #include "varasm.h" | |
51 | #include "stmt.h" | |
6fb5fa3c DB |
52 | #include "expr.h" |
53 | #include "recog.h" | |
b0710fe1 | 54 | #include "insn-codes.h" |
8660aaae | 55 | #include "optabs.h" |
6fb5fa3c | 56 | #include "dbgcnt.h" |
67d0afe9 | 57 | #include "target.h" |
dabd47e7 | 58 | #include "params.h" |
2fb9a547 | 59 | #include "internal-fn.h" |
442b4905 | 60 | #include "gimple-ssa.h" |
d7111da8 | 61 | #include "rtl-iter.h" |
6c5ed3f1 | 62 | #include "cfgcleanup.h" |
6fb5fa3c DB |
63 | |
64 | /* This file contains three techniques for performing Dead Store | |
b8698a0f | 65 | Elimination (dse). |
6fb5fa3c DB |
66 | |
67 | * The first technique performs dse locally on any base address. It | |
68 | is based on the cselib which is a local value numbering technique. | |
69 | This technique is local to a basic block but deals with a fairly | |
70 | general addresses. | |
b8698a0f | 71 | |
6fb5fa3c DB |
72 | * The second technique performs dse globally but is restricted to |
73 | base addresses that are either constant or are relative to the | |
74 | frame_pointer. | |
75 | ||
76 | * The third technique, (which is only done after register allocation) | |
77 | processes the spill spill slots. This differs from the second | |
78 | technique because it takes advantage of the fact that spilling is | |
79 | completely free from the effects of aliasing. | |
80 | ||
81 | Logically, dse is a backwards dataflow problem. A store can be | |
82 | deleted if it if cannot be reached in the backward direction by any | |
83 | use of the value being stored. However, the local technique uses a | |
84 | forwards scan of the basic block because cselib requires that the | |
85 | block be processed in that order. | |
86 | ||
87 | The pass is logically broken into 7 steps: | |
88 | ||
89 | 0) Initialization. | |
90 | ||
91 | 1) The local algorithm, as well as scanning the insns for the two | |
92 | global algorithms. | |
93 | ||
94 | 2) Analysis to see if the global algs are necessary. In the case | |
95 | of stores base on a constant address, there must be at least two | |
96 | stores to that address, to make it possible to delete some of the | |
97 | stores. In the case of stores off of the frame or spill related | |
98 | stores, only one store to an address is necessary because those | |
99 | stores die at the end of the function. | |
100 | ||
b8698a0f | 101 | 3) Set up the global dataflow equations based on processing the |
6fb5fa3c DB |
102 | info parsed in the first step. |
103 | ||
104 | 4) Solve the dataflow equations. | |
105 | ||
106 | 5) Delete the insns that the global analysis has indicated are | |
107 | unnecessary. | |
108 | ||
073a8998 | 109 | 6) Delete insns that store the same value as preceding store |
8dd5516b JJ |
110 | where the earlier store couldn't be eliminated. |
111 | ||
112 | 7) Cleanup. | |
6fb5fa3c DB |
113 | |
114 | This step uses cselib and canon_rtx to build the largest expression | |
115 | possible for each address. This pass is a forwards pass through | |
116 | each basic block. From the point of view of the global technique, | |
117 | the first pass could examine a block in either direction. The | |
0d52bcc1 | 118 | forwards ordering is to accommodate cselib. |
6fb5fa3c | 119 | |
5ef0b50d | 120 | We make a simplifying assumption: addresses fall into four broad |
6fb5fa3c DB |
121 | categories: |
122 | ||
123 | 1) base has rtx_varies_p == false, offset is constant. | |
124 | 2) base has rtx_varies_p == false, offset variable. | |
125 | 3) base has rtx_varies_p == true, offset constant. | |
126 | 4) base has rtx_varies_p == true, offset variable. | |
127 | ||
128 | The local passes are able to process all 4 kinds of addresses. The | |
5ef0b50d | 129 | global pass only handles 1). |
6fb5fa3c DB |
130 | |
131 | The global problem is formulated as follows: | |
132 | ||
133 | A store, S1, to address A, where A is not relative to the stack | |
134 | frame, can be eliminated if all paths from S1 to the end of the | |
5ef0b50d | 135 | function contain another store to A before a read to A. |
6fb5fa3c DB |
136 | |
137 | If the address A is relative to the stack frame, a store S2 to A | |
5ef0b50d | 138 | can be eliminated if there are no paths from S2 that reach the |
6fb5fa3c | 139 | end of the function that read A before another store to A. In |
5ef0b50d | 140 | this case S2 can be deleted if there are paths from S2 to the |
6fb5fa3c DB |
141 | end of the function that have no reads or writes to A. This |
142 | second case allows stores to the stack frame to be deleted that | |
143 | would otherwise die when the function returns. This cannot be | |
144 | done if stores_off_frame_dead_at_return is not true. See the doc | |
145 | for that variable for when this variable is false. | |
146 | ||
147 | The global problem is formulated as a backwards set union | |
148 | dataflow problem where the stores are the gens and reads are the | |
149 | kills. Set union problems are rare and require some special | |
150 | handling given our representation of bitmaps. A straightforward | |
5ef0b50d | 151 | implementation requires a lot of bitmaps filled with 1s. |
6fb5fa3c DB |
152 | These are expensive and cumbersome in our bitmap formulation so |
153 | care has been taken to avoid large vectors filled with 1s. See | |
154 | the comments in bb_info and in the dataflow confluence functions | |
b8698a0f | 155 | for details. |
6fb5fa3c DB |
156 | |
157 | There are two places for further enhancements to this algorithm: | |
b8698a0f | 158 | |
6fb5fa3c DB |
159 | 1) The original dse which was embedded in a pass called flow also |
160 | did local address forwarding. For example in | |
161 | ||
162 | A <- r100 | |
163 | ... <- A | |
164 | ||
165 | flow would replace the right hand side of the second insn with a | |
6ed3da00 | 166 | reference to r100. Most of the information is available to add this |
6fb5fa3c DB |
167 | to this pass. It has not done it because it is a lot of work in |
168 | the case that either r100 is assigned to between the first and | |
169 | second insn and/or the second insn is a load of part of the value | |
170 | stored by the first insn. | |
171 | ||
172 | insn 5 in gcc.c-torture/compile/990203-1.c simple case. | |
173 | insn 15 in gcc.c-torture/execute/20001017-2.c simple case. | |
174 | insn 25 in gcc.c-torture/execute/20001026-1.c simple case. | |
175 | insn 44 in gcc.c-torture/execute/20010910-1.c simple case. | |
176 | ||
177 | 2) The cleaning up of spill code is quite profitable. It currently | |
178 | depends on reading tea leaves and chicken entrails left by reload. | |
179 | This pass depends on reload creating a singleton alias set for each | |
180 | spill slot and telling the next dse pass which of these alias sets | |
181 | are the singletons. Rather than analyze the addresses of the | |
182 | spills, dse's spill processing just does analysis of the loads and | |
183 | stores that use those alias sets. There are three cases where this | |
184 | falls short: | |
185 | ||
186 | a) Reload sometimes creates the slot for one mode of access, and | |
187 | then inserts loads and/or stores for a smaller mode. In this | |
188 | case, the current code just punts on the slot. The proper thing | |
189 | to do is to back out and use one bit vector position for each | |
190 | byte of the entity associated with the slot. This depends on | |
191 | KNOWING that reload always generates the accesses for each of the | |
192 | bytes in some canonical (read that easy to understand several | |
193 | passes after reload happens) way. | |
194 | ||
195 | b) Reload sometimes decides that spill slot it allocated was not | |
196 | large enough for the mode and goes back and allocates more slots | |
197 | with the same mode and alias set. The backout in this case is a | |
198 | little more graceful than (a). In this case the slot is unmarked | |
199 | as being a spill slot and if final address comes out to be based | |
b8698a0f | 200 | off the frame pointer, the global algorithm handles this slot. |
6fb5fa3c DB |
201 | |
202 | c) For any pass that may prespill, there is currently no | |
203 | mechanism to tell the dse pass that the slot being used has the | |
204 | special properties that reload uses. It may be that all that is | |
0d52bcc1 | 205 | required is to have those passes make the same calls that reload |
6fb5fa3c DB |
206 | does, assuming that the alias sets can be manipulated in the same |
207 | way. */ | |
208 | ||
209 | /* There are limits to the size of constant offsets we model for the | |
210 | global problem. There are certainly test cases, that exceed this | |
211 | limit, however, it is unlikely that there are important programs | |
212 | that really have constant offsets this size. */ | |
213 | #define MAX_OFFSET (64 * 1024) | |
214 | ||
3f9b14ff SB |
215 | /* Obstack for the DSE dataflow bitmaps. We don't want to put these |
216 | on the default obstack because these bitmaps can grow quite large | |
217 | (~2GB for the small (!) test case of PR54146) and we'll hold on to | |
218 | all that memory until the end of the compiler run. | |
219 | As a bonus, delete_tree_live_info can destroy all the bitmaps by just | |
220 | releasing the whole obstack. */ | |
221 | static bitmap_obstack dse_bitmap_obstack; | |
222 | ||
223 | /* Obstack for other data. As for above: Kinda nice to be able to | |
224 | throw it all away at the end in one big sweep. */ | |
225 | static struct obstack dse_obstack; | |
226 | ||
227 | /* Scratch bitmap for cselib's cselib_expand_value_rtx. */ | |
6fb5fa3c | 228 | static bitmap scratch = NULL; |
3f9b14ff | 229 | |
da6603c6 | 230 | struct insn_info_type; |
6fb5fa3c DB |
231 | |
232 | /* This structure holds information about a candidate store. */ | |
b8698a0f | 233 | struct store_info |
6fb5fa3c DB |
234 | { |
235 | ||
236 | /* False means this is a clobber. */ | |
237 | bool is_set; | |
238 | ||
8dd5516b JJ |
239 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
240 | bool is_large; | |
241 | ||
6fb5fa3c DB |
242 | /* The id of the mem group of the base address. If rtx_varies_p is |
243 | true, this is -1. Otherwise, it is the index into the group | |
244 | table. */ | |
245 | int group_id; | |
b8698a0f | 246 | |
6fb5fa3c DB |
247 | /* This is the cselib value. */ |
248 | cselib_val *cse_base; | |
249 | ||
250 | /* This canonized mem. */ | |
251 | rtx mem; | |
252 | ||
6216f94e | 253 | /* Canonized MEM address for use by canon_true_dependence. */ |
6fb5fa3c DB |
254 | rtx mem_addr; |
255 | ||
256 | /* If this is non-zero, it is the alias set of a spill location. */ | |
4862826d | 257 | alias_set_type alias_set; |
6fb5fa3c DB |
258 | |
259 | /* The offset of the first and byte before the last byte associated | |
260 | with the operation. */ | |
8dd5516b JJ |
261 | HOST_WIDE_INT begin, end; |
262 | ||
263 | union | |
264 | { | |
265 | /* A bitmask as wide as the number of bytes in the word that | |
266 | contains a 1 if the byte may be needed. The store is unused if | |
267 | all of the bits are 0. This is used if IS_LARGE is false. */ | |
268 | unsigned HOST_WIDE_INT small_bitmask; | |
269 | ||
270 | struct | |
271 | { | |
272 | /* A bitmap with one bit per byte. Cleared bit means the position | |
273 | is needed. Used if IS_LARGE is false. */ | |
dc491a25 | 274 | bitmap bmap; |
6fb5fa3c | 275 | |
8dd5516b JJ |
276 | /* Number of set bits (i.e. unneeded bytes) in BITMAP. If it is |
277 | equal to END - BEGIN, the whole store is unused. */ | |
278 | int count; | |
279 | } large; | |
280 | } positions_needed; | |
6fb5fa3c DB |
281 | |
282 | /* The next store info for this insn. */ | |
283 | struct store_info *next; | |
284 | ||
285 | /* The right hand side of the store. This is used if there is a | |
286 | subsequent reload of the mems address somewhere later in the | |
287 | basic block. */ | |
8dd5516b JJ |
288 | rtx rhs; |
289 | ||
290 | /* If rhs is or holds a constant, this contains that constant, | |
291 | otherwise NULL. */ | |
292 | rtx const_rhs; | |
293 | ||
294 | /* Set if this store stores the same constant value as REDUNDANT_REASON | |
295 | insn stored. These aren't eliminated early, because doing that | |
296 | might prevent the earlier larger store to be eliminated. */ | |
da6603c6 | 297 | struct insn_info_type *redundant_reason; |
6fb5fa3c DB |
298 | }; |
299 | ||
4fe663b0 L |
300 | /* Return a bitmask with the first N low bits set. */ |
301 | ||
302 | static unsigned HOST_WIDE_INT | |
303 | lowpart_bitmask (int n) | |
304 | { | |
305 | unsigned HOST_WIDE_INT mask = ~(unsigned HOST_WIDE_INT) 0; | |
306 | return mask >> (HOST_BITS_PER_WIDE_INT - n); | |
307 | } | |
308 | ||
6fb5fa3c | 309 | typedef struct store_info *store_info_t; |
da6603c6 ML |
310 | static pool_allocator<store_info> cse_store_info_pool ("cse_store_info_pool", |
311 | 100); | |
312 | ||
313 | static pool_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool", | |
314 | 100); | |
6fb5fa3c DB |
315 | |
316 | /* This structure holds information about a load. These are only | |
317 | built for rtx bases. */ | |
da6603c6 | 318 | struct read_info_type |
6fb5fa3c DB |
319 | { |
320 | /* The id of the mem group of the base address. */ | |
321 | int group_id; | |
322 | ||
323 | /* If this is non-zero, it is the alias set of a spill location. */ | |
4862826d | 324 | alias_set_type alias_set; |
6fb5fa3c DB |
325 | |
326 | /* The offset of the first and byte after the last byte associated | |
327 | with the operation. If begin == end == 0, the read did not have | |
328 | a constant offset. */ | |
329 | int begin, end; | |
330 | ||
331 | /* The mem being read. */ | |
332 | rtx mem; | |
333 | ||
334 | /* The next read_info for this insn. */ | |
da6603c6 ML |
335 | struct read_info_type *next; |
336 | ||
337 | /* Pool allocation new operator. */ | |
338 | inline void *operator new (size_t) | |
339 | { | |
340 | return pool.allocate (); | |
341 | } | |
342 | ||
343 | /* Delete operator utilizing pool allocation. */ | |
344 | inline void operator delete (void *ptr) | |
345 | { | |
346 | pool.remove ((read_info_type *) ptr); | |
347 | } | |
348 | ||
349 | /* Memory allocation pool. */ | |
350 | static pool_allocator<read_info_type> pool; | |
6fb5fa3c | 351 | }; |
da6603c6 | 352 | typedef struct read_info_type *read_info_t; |
6fb5fa3c | 353 | |
da6603c6 | 354 | pool_allocator<read_info_type> read_info_type::pool ("read_info_pool", 100); |
6fb5fa3c DB |
355 | |
356 | /* One of these records is created for each insn. */ | |
357 | ||
da6603c6 | 358 | struct insn_info_type |
6fb5fa3c DB |
359 | { |
360 | /* Set true if the insn contains a store but the insn itself cannot | |
361 | be deleted. This is set if the insn is a parallel and there is | |
362 | more than one non dead output or if the insn is in some way | |
363 | volatile. */ | |
364 | bool cannot_delete; | |
365 | ||
366 | /* This field is only used by the global algorithm. It is set true | |
367 | if the insn contains any read of mem except for a (1). This is | |
368 | also set if the insn is a call or has a clobber mem. If the insn | |
369 | contains a wild read, the use_rec will be null. */ | |
370 | bool wild_read; | |
371 | ||
d26c7090 ER |
372 | /* This is true only for CALL instructions which could potentially read |
373 | any non-frame memory location. This field is used by the global | |
374 | algorithm. */ | |
375 | bool non_frame_wild_read; | |
376 | ||
50f0f366 EB |
377 | /* This field is only used for the processing of const functions. |
378 | These functions cannot read memory, but they can read the stack | |
64520bdc EB |
379 | because that is where they may get their parms. We need to be |
380 | this conservative because, like the store motion pass, we don't | |
381 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
382 | Moreover, we need to distinguish two cases: | |
383 | 1. Before reload (register elimination), the stores related to | |
384 | outgoing arguments are stack pointer based and thus deemed | |
385 | of non-constant base in this pass. This requires special | |
386 | handling but also means that the frame pointer based stores | |
387 | need not be killed upon encountering a const function call. | |
388 | 2. After reload, the stores related to outgoing arguments can be | |
389 | either stack pointer or hard frame pointer based. This means | |
390 | that we have no other choice than also killing all the frame | |
391 | pointer based stores upon encountering a const function call. | |
57534689 JJ |
392 | This field is set after reload for const function calls and before |
393 | reload for const tail function calls on targets where arg pointer | |
394 | is the frame pointer. Having this set is less severe than a wild | |
395 | read, it just means that all the frame related stores are killed | |
396 | rather than all the stores. */ | |
64520bdc EB |
397 | bool frame_read; |
398 | ||
399 | /* This field is only used for the processing of const functions. | |
400 | It is set if the insn may contain a stack pointer based store. */ | |
50f0f366 | 401 | bool stack_pointer_based; |
6fb5fa3c DB |
402 | |
403 | /* This is true if any of the sets within the store contains a | |
404 | cselib base. Such stores can only be deleted by the local | |
405 | algorithm. */ | |
406 | bool contains_cselib_groups; | |
407 | ||
408 | /* The insn. */ | |
dd60a84c | 409 | rtx_insn *insn; |
6fb5fa3c DB |
410 | |
411 | /* The list of mem sets or mem clobbers that are contained in this | |
412 | insn. If the insn is deletable, it contains only one mem set. | |
413 | But it could also contain clobbers. Insns that contain more than | |
414 | one mem set are not deletable, but each of those mems are here in | |
6ed3da00 | 415 | order to provide info to delete other insns. */ |
6fb5fa3c DB |
416 | store_info_t store_rec; |
417 | ||
418 | /* The linked list of mem uses in this insn. Only the reads from | |
419 | rtx bases are listed here. The reads to cselib bases are | |
420 | completely processed during the first scan and so are never | |
421 | created. */ | |
422 | read_info_t read_rec; | |
423 | ||
9e582b1d JR |
424 | /* The live fixed registers. We assume only fixed registers can |
425 | cause trouble by being clobbered from an expanded pattern; | |
426 | storing only the live fixed registers (rather than all registers) | |
427 | means less memory needs to be allocated / copied for the individual | |
428 | stores. */ | |
429 | regset fixed_regs_live; | |
430 | ||
6fb5fa3c | 431 | /* The prev insn in the basic block. */ |
da6603c6 | 432 | struct insn_info_type * prev_insn; |
6fb5fa3c DB |
433 | |
434 | /* The linked list of insns that are in consideration for removal in | |
073a8998 | 435 | the forwards pass through the basic block. This pointer may be |
6fb5fa3c | 436 | trash as it is not cleared when a wild read occurs. The only |
fa10beec | 437 | time it is guaranteed to be correct is when the traversal starts |
6fb5fa3c | 438 | at active_local_stores. */ |
da6603c6 ML |
439 | struct insn_info_type * next_local_store; |
440 | ||
441 | /* Pool allocation new operator. */ | |
442 | inline void *operator new (size_t) | |
443 | { | |
444 | return pool.allocate (); | |
445 | } | |
446 | ||
447 | /* Delete operator utilizing pool allocation. */ | |
448 | inline void operator delete (void *ptr) | |
449 | { | |
450 | pool.remove ((insn_info_type *) ptr); | |
451 | } | |
452 | ||
453 | /* Memory allocation pool. */ | |
454 | static pool_allocator<insn_info_type> pool; | |
6fb5fa3c | 455 | }; |
da6603c6 | 456 | typedef struct insn_info_type *insn_info_t; |
6fb5fa3c | 457 | |
da6603c6 | 458 | pool_allocator<insn_info_type> insn_info_type::pool ("insn_info_pool", 100); |
6fb5fa3c DB |
459 | |
460 | /* The linked list of stores that are under consideration in this | |
b8698a0f | 461 | basic block. */ |
6fb5fa3c | 462 | static insn_info_t active_local_stores; |
dabd47e7 | 463 | static int active_local_stores_len; |
6fb5fa3c | 464 | |
da6603c6 | 465 | struct dse_bb_info_type |
6fb5fa3c | 466 | { |
6fb5fa3c DB |
467 | /* Pointer to the insn info for the last insn in the block. These |
468 | are linked so this is how all of the insns are reached. During | |
469 | scanning this is the current insn being scanned. */ | |
470 | insn_info_t last_insn; | |
471 | ||
472 | /* The info for the global dataflow problem. */ | |
473 | ||
474 | ||
475 | /* This is set if the transfer function should and in the wild_read | |
476 | bitmap before applying the kill and gen sets. That vector knocks | |
477 | out most of the bits in the bitmap and thus speeds up the | |
478 | operations. */ | |
479 | bool apply_wild_read; | |
480 | ||
02b47899 KZ |
481 | /* The following 4 bitvectors hold information about which positions |
482 | of which stores are live or dead. They are indexed by | |
483 | get_bitmap_index. */ | |
484 | ||
6fb5fa3c DB |
485 | /* The set of store positions that exist in this block before a wild read. */ |
486 | bitmap gen; | |
b8698a0f | 487 | |
6fb5fa3c DB |
488 | /* The set of load positions that exist in this block above the |
489 | same position of a store. */ | |
490 | bitmap kill; | |
491 | ||
492 | /* The set of stores that reach the top of the block without being | |
493 | killed by a read. | |
494 | ||
495 | Do not represent the in if it is all ones. Note that this is | |
496 | what the bitvector should logically be initialized to for a set | |
497 | intersection problem. However, like the kill set, this is too | |
498 | expensive. So initially, the in set will only be created for the | |
499 | exit block and any block that contains a wild read. */ | |
500 | bitmap in; | |
501 | ||
502 | /* The set of stores that reach the bottom of the block from it's | |
503 | successors. | |
504 | ||
505 | Do not represent the in if it is all ones. Note that this is | |
506 | what the bitvector should logically be initialized to for a set | |
507 | intersection problem. However, like the kill and in set, this is | |
508 | too expensive. So what is done is that the confluence operator | |
509 | just initializes the vector from one of the out sets of the | |
510 | successors of the block. */ | |
511 | bitmap out; | |
02b47899 KZ |
512 | |
513 | /* The following bitvector is indexed by the reg number. It | |
514 | contains the set of regs that are live at the current instruction | |
515 | being processed. While it contains info for all of the | |
9e582b1d JR |
516 | registers, only the hard registers are actually examined. It is used |
517 | to assure that shift and/or add sequences that are inserted do not | |
073a8998 | 518 | accidentally clobber live hard regs. */ |
02b47899 | 519 | bitmap regs_live; |
da6603c6 ML |
520 | |
521 | /* Pool allocation new operator. */ | |
522 | inline void *operator new (size_t) | |
523 | { | |
524 | return pool.allocate (); | |
525 | } | |
526 | ||
527 | /* Delete operator utilizing pool allocation. */ | |
528 | inline void operator delete (void *ptr) | |
529 | { | |
530 | pool.remove ((dse_bb_info_type *) ptr); | |
531 | } | |
532 | ||
533 | /* Memory allocation pool. */ | |
534 | static pool_allocator<dse_bb_info_type> pool; | |
6fb5fa3c DB |
535 | }; |
536 | ||
da6603c6 ML |
537 | typedef struct dse_bb_info_type *bb_info_t; |
538 | pool_allocator<dse_bb_info_type> dse_bb_info_type::pool ("bb_info_pool", 100); | |
6fb5fa3c DB |
539 | |
540 | /* Table to hold all bb_infos. */ | |
541 | static bb_info_t *bb_table; | |
542 | ||
543 | /* There is a group_info for each rtx base that is used to reference | |
544 | memory. There are also not many of the rtx bases because they are | |
545 | very limited in scope. */ | |
546 | ||
b8698a0f | 547 | struct group_info |
6fb5fa3c DB |
548 | { |
549 | /* The actual base of the address. */ | |
550 | rtx rtx_base; | |
551 | ||
552 | /* The sequential id of the base. This allows us to have a | |
553 | canonical ordering of these that is not based on addresses. */ | |
554 | int id; | |
555 | ||
8f5929e1 JJ |
556 | /* True if there are any positions that are to be processed |
557 | globally. */ | |
558 | bool process_globally; | |
559 | ||
560 | /* True if the base of this group is either the frame_pointer or | |
561 | hard_frame_pointer. */ | |
562 | bool frame_related; | |
563 | ||
d32f725a OH |
564 | /* A mem wrapped around the base pointer for the group in order to do |
565 | read dependency. It must be given BLKmode in order to encompass all | |
566 | the possible offsets from the base. */ | |
6fb5fa3c | 567 | rtx base_mem; |
b8698a0f | 568 | |
6216f94e JJ |
569 | /* Canonized version of base_mem's address. */ |
570 | rtx canon_base_addr; | |
6fb5fa3c DB |
571 | |
572 | /* These two sets of two bitmaps are used to keep track of how many | |
6ed3da00 | 573 | stores are actually referencing that position from this base. We |
6fb5fa3c | 574 | only do this for rtx bases as this will be used to assign |
6ed3da00 | 575 | positions in the bitmaps for the global problem. Bit N is set in |
6fb5fa3c DB |
576 | store1 on the first store for offset N. Bit N is set in store2 |
577 | for the second store to offset N. This is all we need since we | |
578 | only care about offsets that have two or more stores for them. | |
579 | ||
580 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
581 | for 0 and greater offsets. | |
582 | ||
583 | There is one special case here, for stores into the stack frame, | |
584 | we will or store1 into store2 before deciding which stores look | |
585 | at globally. This is because stores to the stack frame that have | |
586 | no other reads before the end of the function can also be | |
587 | deleted. */ | |
588 | bitmap store1_n, store1_p, store2_n, store2_p; | |
589 | ||
d26c7090 ER |
590 | /* These bitmaps keep track of offsets in this group escape this function. |
591 | An offset escapes if it corresponds to a named variable whose | |
592 | addressable flag is set. */ | |
593 | bitmap escaped_n, escaped_p; | |
594 | ||
6ed3da00 | 595 | /* The positions in this bitmap have the same assignments as the in, |
6fb5fa3c | 596 | out, gen and kill bitmaps. This bitmap is all zeros except for |
6ed3da00 | 597 | the positions that are occupied by stores for this group. */ |
6fb5fa3c DB |
598 | bitmap group_kill; |
599 | ||
6fb5fa3c | 600 | /* The offset_map is used to map the offsets from this base into |
6ed3da00 | 601 | positions in the global bitmaps. It is only created after all of |
6fb5fa3c DB |
602 | the all of stores have been scanned and we know which ones we |
603 | care about. */ | |
b8698a0f L |
604 | int *offset_map_n, *offset_map_p; |
605 | int offset_map_size_n, offset_map_size_p; | |
da6603c6 ML |
606 | |
607 | /* Pool allocation new operator. */ | |
608 | inline void *operator new (size_t) | |
609 | { | |
610 | return pool.allocate (); | |
611 | } | |
612 | ||
613 | /* Delete operator utilizing pool allocation. */ | |
614 | inline void operator delete (void *ptr) | |
615 | { | |
616 | pool.remove ((group_info *) ptr); | |
617 | } | |
618 | ||
619 | /* Memory allocation pool. */ | |
620 | static pool_allocator<group_info> pool; | |
6fb5fa3c DB |
621 | }; |
622 | typedef struct group_info *group_info_t; | |
5f754896 | 623 | typedef const struct group_info *const_group_info_t; |
da6603c6 ML |
624 | |
625 | pool_allocator<group_info> group_info::pool ("rtx_group_info_pool", 100); | |
6fb5fa3c | 626 | |
6fb5fa3c DB |
627 | /* Index into the rtx_group_vec. */ |
628 | static int rtx_group_next_id; | |
629 | ||
6fb5fa3c | 630 | |
9771b263 | 631 | static vec<group_info_t> rtx_group_vec; |
6fb5fa3c DB |
632 | |
633 | ||
634 | /* This structure holds the set of changes that are being deferred | |
635 | when removing read operation. See replace_read. */ | |
b8698a0f | 636 | struct deferred_change |
6fb5fa3c DB |
637 | { |
638 | ||
639 | /* The mem that is being replaced. */ | |
640 | rtx *loc; | |
641 | ||
642 | /* The reg it is being replaced with. */ | |
643 | rtx reg; | |
644 | ||
645 | struct deferred_change *next; | |
da6603c6 ML |
646 | |
647 | /* Pool allocation new operator. */ | |
648 | inline void *operator new (size_t) | |
649 | { | |
650 | return pool.allocate (); | |
651 | } | |
652 | ||
653 | /* Delete operator utilizing pool allocation. */ | |
654 | inline void operator delete (void *ptr) | |
655 | { | |
656 | pool.remove ((deferred_change *) ptr); | |
657 | } | |
658 | ||
659 | /* Memory allocation pool. */ | |
660 | static pool_allocator<deferred_change> pool; | |
6fb5fa3c DB |
661 | }; |
662 | ||
663 | typedef struct deferred_change *deferred_change_t; | |
da6603c6 ML |
664 | |
665 | pool_allocator<deferred_change> deferred_change::pool | |
666 | ("deferred_change_pool", 10); | |
6fb5fa3c DB |
667 | |
668 | static deferred_change_t deferred_change_list = NULL; | |
669 | ||
6fb5fa3c DB |
670 | /* The group that holds all of the clear_alias_sets. */ |
671 | static group_info_t clear_alias_group; | |
672 | ||
673 | /* The modes of the clear_alias_sets. */ | |
674 | static htab_t clear_alias_mode_table; | |
675 | ||
676 | /* Hash table element to look up the mode for an alias set. */ | |
677 | struct clear_alias_mode_holder | |
678 | { | |
4862826d | 679 | alias_set_type alias_set; |
ef4bddc2 | 680 | machine_mode mode; |
6fb5fa3c DB |
681 | }; |
682 | ||
e3b5732b | 683 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
9dd9bf80 | 684 | this for vararg functions because they play games with the frame. */ |
6fb5fa3c DB |
685 | static bool stores_off_frame_dead_at_return; |
686 | ||
687 | /* Counter for stats. */ | |
b8698a0f L |
688 | static int globally_deleted; |
689 | static int locally_deleted; | |
690 | static int spill_deleted; | |
691 | ||
6fb5fa3c DB |
692 | static bitmap all_blocks; |
693 | ||
d26c7090 ER |
694 | /* Locations that are killed by calls in the global phase. */ |
695 | static bitmap kill_on_calls; | |
696 | ||
6fb5fa3c DB |
697 | /* The number of bits used in the global bitmaps. */ |
698 | static unsigned int current_position; | |
6fb5fa3c DB |
699 | \f |
700 | /*---------------------------------------------------------------------------- | |
701 | Zeroth step. | |
702 | ||
b8698a0f | 703 | Initialization. |
6fb5fa3c DB |
704 | ----------------------------------------------------------------------------*/ |
705 | ||
6fb5fa3c DB |
706 | |
707 | /* Find the entry associated with ALIAS_SET. */ | |
708 | ||
709 | static struct clear_alias_mode_holder * | |
4862826d | 710 | clear_alias_set_lookup (alias_set_type alias_set) |
6fb5fa3c DB |
711 | { |
712 | struct clear_alias_mode_holder tmp_holder; | |
713 | void **slot; | |
b8698a0f | 714 | |
6fb5fa3c DB |
715 | tmp_holder.alias_set = alias_set; |
716 | slot = htab_find_slot (clear_alias_mode_table, &tmp_holder, NO_INSERT); | |
717 | gcc_assert (*slot); | |
b8698a0f | 718 | |
f883e0a7 | 719 | return (struct clear_alias_mode_holder *) *slot; |
6fb5fa3c DB |
720 | } |
721 | ||
722 | ||
723 | /* Hashtable callbacks for maintaining the "bases" field of | |
724 | store_group_info, given that the addresses are function invariants. */ | |
725 | ||
8d67ee55 | 726 | struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
703c8606 | 727 | { |
67f58944 TS |
728 | static inline hashval_t hash (const group_info *); |
729 | static inline bool equal (const group_info *, const group_info *); | |
703c8606 LC |
730 | }; |
731 | ||
732 | inline bool | |
67f58944 TS |
733 | invariant_group_base_hasher::equal (const group_info *gi1, |
734 | const group_info *gi2) | |
6fb5fa3c | 735 | { |
6fb5fa3c DB |
736 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
737 | } | |
738 | ||
703c8606 | 739 | inline hashval_t |
67f58944 | 740 | invariant_group_base_hasher::hash (const group_info *gi) |
6fb5fa3c | 741 | { |
6fb5fa3c DB |
742 | int do_not_record; |
743 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
744 | } | |
745 | ||
703c8606 | 746 | /* Tables of group_info structures, hashed by base value. */ |
c203e8a7 | 747 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
703c8606 | 748 | |
6fb5fa3c DB |
749 | |
750 | /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
751 | ||
752 | static group_info_t | |
753 | get_group_info (rtx base) | |
754 | { | |
b8698a0f L |
755 | struct group_info tmp_gi; |
756 | group_info_t gi; | |
703c8606 | 757 | group_info **slot; |
6fb5fa3c DB |
758 | |
759 | if (base) | |
760 | { | |
761 | /* Find the store_base_info structure for BASE, creating a new one | |
762 | if necessary. */ | |
763 | tmp_gi.rtx_base = base; | |
c203e8a7 | 764 | slot = rtx_group_table->find_slot (&tmp_gi, INSERT); |
6fb5fa3c DB |
765 | gi = (group_info_t) *slot; |
766 | } | |
767 | else | |
768 | { | |
769 | if (!clear_alias_group) | |
770 | { | |
da6603c6 | 771 | clear_alias_group = gi = new group_info; |
6fb5fa3c DB |
772 | memset (gi, 0, sizeof (struct group_info)); |
773 | gi->id = rtx_group_next_id++; | |
3f9b14ff SB |
774 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
775 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
776 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
777 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
778 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
779 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
780 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c DB |
781 | gi->process_globally = false; |
782 | gi->offset_map_size_n = 0; | |
783 | gi->offset_map_size_p = 0; | |
784 | gi->offset_map_n = NULL; | |
785 | gi->offset_map_p = NULL; | |
9771b263 | 786 | rtx_group_vec.safe_push (gi); |
6fb5fa3c DB |
787 | } |
788 | return clear_alias_group; | |
789 | } | |
790 | ||
791 | if (gi == NULL) | |
792 | { | |
da6603c6 | 793 | *slot = gi = new group_info; |
6fb5fa3c DB |
794 | gi->rtx_base = base; |
795 | gi->id = rtx_group_next_id++; | |
d32f725a | 796 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
6216f94e | 797 | gi->canon_base_addr = canon_rtx (base); |
3f9b14ff SB |
798 | gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
799 | gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
800 | gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
801 | gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
802 | gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
803 | gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
804 | gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 805 | gi->process_globally = false; |
b8698a0f | 806 | gi->frame_related = |
6fb5fa3c DB |
807 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
808 | gi->offset_map_size_n = 0; | |
809 | gi->offset_map_size_p = 0; | |
810 | gi->offset_map_n = NULL; | |
811 | gi->offset_map_p = NULL; | |
9771b263 | 812 | rtx_group_vec.safe_push (gi); |
6fb5fa3c DB |
813 | } |
814 | ||
815 | return gi; | |
816 | } | |
817 | ||
818 | ||
819 | /* Initialization of data structures. */ | |
820 | ||
821 | static void | |
822 | dse_step0 (void) | |
823 | { | |
824 | locally_deleted = 0; | |
825 | globally_deleted = 0; | |
826 | spill_deleted = 0; | |
827 | ||
3f9b14ff SB |
828 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
829 | gcc_obstack_init (&dse_obstack); | |
830 | ||
831 | scratch = BITMAP_ALLOC (®_obstack); | |
832 | kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
6fb5fa3c | 833 | |
6fb5fa3c | 834 | |
c203e8a7 | 835 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
6fb5fa3c | 836 | |
8b1c6fd7 | 837 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
6fb5fa3c DB |
838 | rtx_group_next_id = 0; |
839 | ||
e3b5732b | 840 | stores_off_frame_dead_at_return = !cfun->stdarg; |
6fb5fa3c DB |
841 | |
842 | init_alias_analysis (); | |
b8698a0f | 843 | |
4b943a49 | 844 | clear_alias_group = NULL; |
6fb5fa3c DB |
845 | } |
846 | ||
847 | ||
848 | \f | |
849 | /*---------------------------------------------------------------------------- | |
850 | First step. | |
851 | ||
852 | Scan all of the insns. Any random ordering of the blocks is fine. | |
0d52bcc1 | 853 | Each block is scanned in forward order to accommodate cselib which |
6fb5fa3c DB |
854 | is used to remove stores with non-constant bases. |
855 | ----------------------------------------------------------------------------*/ | |
856 | ||
857 | /* Delete all of the store_info recs from INSN_INFO. */ | |
858 | ||
b8698a0f | 859 | static void |
6fb5fa3c DB |
860 | free_store_info (insn_info_t insn_info) |
861 | { | |
862 | store_info_t store_info = insn_info->store_rec; | |
863 | while (store_info) | |
864 | { | |
865 | store_info_t next = store_info->next; | |
8dd5516b | 866 | if (store_info->is_large) |
dc491a25 | 867 | BITMAP_FREE (store_info->positions_needed.large.bmap); |
6fb5fa3c | 868 | if (store_info->cse_base) |
da6603c6 | 869 | cse_store_info_pool.remove (store_info); |
6fb5fa3c | 870 | else |
da6603c6 | 871 | rtx_store_info_pool.remove (store_info); |
6fb5fa3c DB |
872 | store_info = next; |
873 | } | |
874 | ||
875 | insn_info->cannot_delete = true; | |
876 | insn_info->contains_cselib_groups = false; | |
877 | insn_info->store_rec = NULL; | |
878 | } | |
879 | ||
9e582b1d JR |
880 | typedef struct |
881 | { | |
dc01c3d1 | 882 | rtx_insn *first, *current; |
9e582b1d JR |
883 | regset fixed_regs_live; |
884 | bool failure; | |
885 | } note_add_store_info; | |
886 | ||
887 | /* Callback for emit_inc_dec_insn_before via note_stores. | |
888 | Check if a register is clobbered which is live afterwards. */ | |
889 | ||
890 | static void | |
891 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
892 | { | |
dc01c3d1 | 893 | rtx_insn *insn; |
9e582b1d | 894 | note_add_store_info *info = (note_add_store_info *) data; |
9e582b1d JR |
895 | |
896 | if (!REG_P (loc)) | |
897 | return; | |
898 | ||
899 | /* If this register is referenced by the current or an earlier insn, | |
900 | that's OK. E.g. this applies to the register that is being incremented | |
901 | with this addition. */ | |
902 | for (insn = info->first; | |
903 | insn != NEXT_INSN (info->current); | |
904 | insn = NEXT_INSN (insn)) | |
905 | if (reg_referenced_p (loc, PATTERN (insn))) | |
906 | return; | |
907 | ||
908 | /* If we come here, we have a clobber of a register that's only OK | |
909 | if that register is not live. If we don't have liveness information | |
910 | available, fail now. */ | |
911 | if (!info->fixed_regs_live) | |
912 | { | |
53d1bae9 | 913 | info->failure = true; |
9e582b1d JR |
914 | return; |
915 | } | |
916 | /* Now check if this is a live fixed register. */ | |
53d1bae9 RS |
917 | unsigned int end_regno = END_REGNO (loc); |
918 | for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) | |
919 | if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) | |
920 | info->failure = true; | |
9e582b1d JR |
921 | } |
922 | ||
4deef538 AO |
923 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
924 | SRC + SRCOFF before insn ARG. */ | |
6fb5fa3c DB |
925 | |
926 | static int | |
4deef538 AO |
927 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
928 | rtx op ATTRIBUTE_UNUSED, | |
929 | rtx dest, rtx src, rtx srcoff, void *arg) | |
6fb5fa3c | 930 | { |
9e582b1d | 931 | insn_info_t insn_info = (insn_info_t) arg; |
dc01c3d1 | 932 | rtx_insn *insn = insn_info->insn, *new_insn, *cur; |
9e582b1d | 933 | note_add_store_info info; |
b8698a0f | 934 | |
4deef538 AO |
935 | /* We can reuse all operands without copying, because we are about |
936 | to delete the insn that contained it. */ | |
9e582b1d | 937 | if (srcoff) |
ed079c4b JJ |
938 | { |
939 | start_sequence (); | |
940 | emit_insn (gen_add3_insn (dest, src, srcoff)); | |
941 | new_insn = get_insns (); | |
942 | end_sequence (); | |
943 | } | |
9e582b1d | 944 | else |
1476d1bd | 945 | new_insn = gen_move_insn (dest, src); |
9e582b1d JR |
946 | info.first = new_insn; |
947 | info.fixed_regs_live = insn_info->fixed_regs_live; | |
948 | info.failure = false; | |
949 | for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
950 | { | |
951 | info.current = cur; | |
952 | note_stores (PATTERN (cur), note_add_store, &info); | |
953 | } | |
6fb5fa3c | 954 | |
9e582b1d JR |
955 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
956 | return it immediately, communicating the failure to its caller. */ | |
957 | if (info.failure) | |
958 | return 1; | |
959 | ||
960 | emit_insn_before (new_insn, insn); | |
6fb5fa3c | 961 | |
8d8e205b | 962 | return 0; |
6fb5fa3c DB |
963 | } |
964 | ||
9e582b1d JR |
965 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
966 | is there, is split into a separate insn. | |
967 | Return true on success (or if there was nothing to do), false on failure. */ | |
6fb5fa3c | 968 | |
9e582b1d JR |
969 | static bool |
970 | check_for_inc_dec_1 (insn_info_t insn_info) | |
6fb5fa3c | 971 | { |
dd60a84c | 972 | rtx_insn *insn = insn_info->insn; |
6fb5fa3c DB |
973 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
974 | if (note) | |
8d8e205b RS |
975 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
976 | insn_info) == 0; | |
9e582b1d | 977 | return true; |
6fb5fa3c DB |
978 | } |
979 | ||
980 | ||
9e582b1d JR |
981 | /* Entry point for postreload. If you work on reload_cse, or you need this |
982 | anywhere else, consider if you can provide register liveness information | |
983 | and add a parameter to this function so that it can be passed down in | |
984 | insn_info.fixed_regs_live. */ | |
985 | bool | |
dd60a84c | 986 | check_for_inc_dec (rtx_insn *insn) |
9e582b1d | 987 | { |
da6603c6 | 988 | insn_info_type insn_info; |
9e582b1d JR |
989 | rtx note; |
990 | ||
991 | insn_info.insn = insn; | |
992 | insn_info.fixed_regs_live = NULL; | |
993 | note = find_reg_note (insn, REG_INC, NULL_RTX); | |
994 | if (note) | |
8d8e205b RS |
995 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
996 | &insn_info) == 0; | |
9e582b1d JR |
997 | return true; |
998 | } | |
999 | ||
b8698a0f | 1000 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
6fb5fa3c DB |
1001 | |
1002 | static void | |
1003 | delete_dead_store_insn (insn_info_t insn_info) | |
1004 | { | |
1005 | read_info_t read_info; | |
1006 | ||
1007 | if (!dbg_cnt (dse)) | |
1008 | return; | |
1009 | ||
9e582b1d JR |
1010 | if (!check_for_inc_dec_1 (insn_info)) |
1011 | return; | |
456610d3 | 1012 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 1013 | { |
b8698a0f | 1014 | fprintf (dump_file, "Locally deleting insn %d ", |
6fb5fa3c DB |
1015 | INSN_UID (insn_info->insn)); |
1016 | if (insn_info->store_rec->alias_set) | |
b8698a0f | 1017 | fprintf (dump_file, "alias set %d\n", |
4862826d | 1018 | (int) insn_info->store_rec->alias_set); |
6fb5fa3c DB |
1019 | else |
1020 | fprintf (dump_file, "\n"); | |
1021 | } | |
1022 | ||
1023 | free_store_info (insn_info); | |
1024 | read_info = insn_info->read_rec; | |
b8698a0f | 1025 | |
6fb5fa3c DB |
1026 | while (read_info) |
1027 | { | |
1028 | read_info_t next = read_info->next; | |
da6603c6 | 1029 | delete read_info; |
6fb5fa3c DB |
1030 | read_info = next; |
1031 | } | |
1032 | insn_info->read_rec = NULL; | |
1033 | ||
1034 | delete_insn (insn_info->insn); | |
1035 | locally_deleted++; | |
1036 | insn_info->insn = NULL; | |
1037 | ||
1038 | insn_info->wild_read = false; | |
1039 | } | |
1040 | ||
88d8330d EB |
1041 | /* Return whether DECL, a local variable, can possibly escape the current |
1042 | function scope. */ | |
1043 | ||
1044 | static bool | |
1045 | local_variable_can_escape (tree decl) | |
1046 | { | |
1047 | if (TREE_ADDRESSABLE (decl)) | |
1048 | return true; | |
1049 | ||
1050 | /* If this is a partitioned variable, we need to consider all the variables | |
1051 | in the partition. This is necessary because a store into one of them can | |
1052 | be replaced with a store into another and this may not change the outcome | |
1053 | of the escape analysis. */ | |
1054 | if (cfun->gimple_df->decls_to_pointers != NULL) | |
1055 | { | |
39c8aaa4 | 1056 | tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); |
88d8330d | 1057 | if (namep) |
39c8aaa4 | 1058 | return TREE_ADDRESSABLE (*namep); |
88d8330d EB |
1059 | } |
1060 | ||
1061 | return false; | |
1062 | } | |
1063 | ||
1064 | /* Return whether EXPR can possibly escape the current function scope. */ | |
1065 | ||
d26c7090 ER |
1066 | static bool |
1067 | can_escape (tree expr) | |
1068 | { | |
1069 | tree base; | |
1070 | if (!expr) | |
1071 | return true; | |
1072 | base = get_base_address (expr); | |
1073 | if (DECL_P (base) | |
88d8330d EB |
1074 | && !may_be_aliased (base) |
1075 | && !(TREE_CODE (base) == VAR_DECL | |
1076 | && !DECL_EXTERNAL (base) | |
1077 | && !TREE_STATIC (base) | |
1078 | && local_variable_can_escape (base))) | |
d26c7090 ER |
1079 | return false; |
1080 | return true; | |
1081 | } | |
6fb5fa3c DB |
1082 | |
1083 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
1084 | OFFSET and WIDTH. */ | |
1085 | ||
1086 | static void | |
d26c7090 ER |
1087 | set_usage_bits (group_info_t group, HOST_WIDE_INT offset, HOST_WIDE_INT width, |
1088 | tree expr) | |
6fb5fa3c DB |
1089 | { |
1090 | HOST_WIDE_INT i; | |
d26c7090 | 1091 | bool expr_escapes = can_escape (expr); |
8dd5516b | 1092 | if (offset > -MAX_OFFSET && offset + width < MAX_OFFSET) |
6fb5fa3c DB |
1093 | for (i=offset; i<offset+width; i++) |
1094 | { | |
1095 | bitmap store1; | |
1096 | bitmap store2; | |
d26c7090 | 1097 | bitmap escaped; |
6fb5fa3c DB |
1098 | int ai; |
1099 | if (i < 0) | |
1100 | { | |
1101 | store1 = group->store1_n; | |
1102 | store2 = group->store2_n; | |
d26c7090 | 1103 | escaped = group->escaped_n; |
6fb5fa3c DB |
1104 | ai = -i; |
1105 | } | |
1106 | else | |
1107 | { | |
1108 | store1 = group->store1_p; | |
1109 | store2 = group->store2_p; | |
d26c7090 | 1110 | escaped = group->escaped_p; |
6fb5fa3c DB |
1111 | ai = i; |
1112 | } | |
b8698a0f | 1113 | |
fcaa4ca4 | 1114 | if (!bitmap_set_bit (store1, ai)) |
6fb5fa3c | 1115 | bitmap_set_bit (store2, ai); |
b8698a0f | 1116 | else |
6fb5fa3c | 1117 | { |
6fb5fa3c DB |
1118 | if (i < 0) |
1119 | { | |
1120 | if (group->offset_map_size_n < ai) | |
1121 | group->offset_map_size_n = ai; | |
1122 | } | |
1123 | else | |
1124 | { | |
1125 | if (group->offset_map_size_p < ai) | |
1126 | group->offset_map_size_p = ai; | |
1127 | } | |
1128 | } | |
d26c7090 ER |
1129 | if (expr_escapes) |
1130 | bitmap_set_bit (escaped, ai); | |
6fb5fa3c DB |
1131 | } |
1132 | } | |
1133 | ||
d26c7090 ER |
1134 | static void |
1135 | reset_active_stores (void) | |
1136 | { | |
1137 | active_local_stores = NULL; | |
1138 | active_local_stores_len = 0; | |
1139 | } | |
6fb5fa3c | 1140 | |
d26c7090 | 1141 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
6fb5fa3c DB |
1142 | |
1143 | static void | |
d26c7090 | 1144 | free_read_records (bb_info_t bb_info) |
6fb5fa3c DB |
1145 | { |
1146 | insn_info_t insn_info = bb_info->last_insn; | |
1147 | read_info_t *ptr = &insn_info->read_rec; | |
6fb5fa3c DB |
1148 | while (*ptr) |
1149 | { | |
1150 | read_info_t next = (*ptr)->next; | |
4862826d | 1151 | if ((*ptr)->alias_set == 0) |
6fb5fa3c | 1152 | { |
da6603c6 | 1153 | delete *ptr; |
6fb5fa3c | 1154 | *ptr = next; |
d26c7090 | 1155 | } |
b8698a0f | 1156 | else |
d26c7090 | 1157 | ptr = &(*ptr)->next; |
6fb5fa3c | 1158 | } |
d26c7090 ER |
1159 | } |
1160 | ||
1161 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
1162 | ||
1163 | static void | |
1164 | add_wild_read (bb_info_t bb_info) | |
1165 | { | |
1166 | insn_info_t insn_info = bb_info->last_insn; | |
6fb5fa3c | 1167 | insn_info->wild_read = true; |
d26c7090 ER |
1168 | free_read_records (bb_info); |
1169 | reset_active_stores (); | |
6fb5fa3c DB |
1170 | } |
1171 | ||
d26c7090 ER |
1172 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1173 | non-frame locations. */ | |
1174 | ||
1175 | static void | |
1176 | add_non_frame_wild_read (bb_info_t bb_info) | |
1177 | { | |
1178 | insn_info_t insn_info = bb_info->last_insn; | |
1179 | insn_info->non_frame_wild_read = true; | |
1180 | free_read_records (bb_info); | |
1181 | reset_active_stores (); | |
1182 | } | |
6fb5fa3c | 1183 | |
50f0f366 EB |
1184 | /* Return true if X is a constant or one of the registers that behave |
1185 | as a constant over the life of a function. This is equivalent to | |
1186 | !rtx_varies_p for memory addresses. */ | |
6fb5fa3c DB |
1187 | |
1188 | static bool | |
1189 | const_or_frame_p (rtx x) | |
1190 | { | |
d8116890 KZ |
1191 | if (CONSTANT_P (x)) |
1192 | return true; | |
1193 | ||
1194 | if (GET_CODE (x) == REG) | |
6fb5fa3c | 1195 | { |
6fb5fa3c DB |
1196 | /* Note that we have to test for the actual rtx used for the frame |
1197 | and arg pointers and not just the register number in case we have | |
1198 | eliminated the frame and/or arg pointer and are using it | |
1199 | for pseudos. */ | |
1200 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
1201 | /* The arg pointer varies if it is not a fixed register. */ | |
1202 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
1203 | || x == pic_offset_table_rtx) | |
1204 | return true; | |
1205 | return false; | |
6fb5fa3c | 1206 | } |
da6603c6 | 1207 | |
d8116890 | 1208 | return false; |
6fb5fa3c DB |
1209 | } |
1210 | ||
b8698a0f L |
1211 | /* Take all reasonable action to put the address of MEM into the form |
1212 | that we can do analysis on. | |
6fb5fa3c DB |
1213 | |
1214 | The gold standard is to get the address into the form: address + | |
1215 | OFFSET where address is something that rtx_varies_p considers a | |
1216 | constant. When we can get the address in this form, we can do | |
1217 | global analysis on it. Note that for constant bases, address is | |
1218 | not actually returned, only the group_id. The address can be | |
1219 | obtained from that. | |
1220 | ||
1221 | If that fails, we try cselib to get a value we can at least use | |
b8698a0f L |
1222 | locally. If that fails we return false. |
1223 | ||
6fb5fa3c DB |
1224 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1225 | group for non_varying bases. | |
1226 | ||
1227 | FOR_READ is true if this is a mem read and false if not. */ | |
1228 | ||
1229 | static bool | |
1230 | canon_address (rtx mem, | |
4862826d | 1231 | alias_set_type *alias_set_out, |
6fb5fa3c | 1232 | int *group_id, |
b8698a0f | 1233 | HOST_WIDE_INT *offset, |
6fb5fa3c DB |
1234 | cselib_val **base) |
1235 | { | |
ef4bddc2 | 1236 | machine_mode address_mode = get_address_mode (mem); |
6fb5fa3c DB |
1237 | rtx mem_address = XEXP (mem, 0); |
1238 | rtx expanded_address, address; | |
403c7520 JJ |
1239 | int expanded; |
1240 | ||
6fb5fa3c DB |
1241 | *alias_set_out = 0; |
1242 | ||
4deef538 | 1243 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
6fb5fa3c | 1244 | |
456610d3 | 1245 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1246 | { |
1247 | fprintf (dump_file, " mem: "); | |
1248 | print_inline_rtx (dump_file, mem_address, 0); | |
1249 | fprintf (dump_file, "\n"); | |
1250 | } | |
1251 | ||
403c7520 JJ |
1252 | /* First see if just canon_rtx (mem_address) is const or frame, |
1253 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
1254 | address = NULL_RTX; | |
1255 | for (expanded = 0; expanded < 2; expanded++) | |
1256 | { | |
1257 | if (expanded) | |
1258 | { | |
1259 | /* Use cselib to replace all of the reg references with the full | |
b8698a0f | 1260 | expression. This will take care of the case where we have |
6fb5fa3c | 1261 | |
403c7520 JJ |
1262 | r_x = base + offset; |
1263 | val = *r_x; | |
b8698a0f L |
1264 | |
1265 | by making it into | |
6fb5fa3c | 1266 | |
403c7520 | 1267 | val = *(base + offset); */ |
6fb5fa3c | 1268 | |
403c7520 JJ |
1269 | expanded_address = cselib_expand_value_rtx (mem_address, |
1270 | scratch, 5); | |
6fb5fa3c | 1271 | |
403c7520 JJ |
1272 | /* If this fails, just go with the address from first |
1273 | iteration. */ | |
1274 | if (!expanded_address) | |
1275 | break; | |
1276 | } | |
1277 | else | |
1278 | expanded_address = mem_address; | |
6fb5fa3c | 1279 | |
403c7520 JJ |
1280 | /* Split the address into canonical BASE + OFFSET terms. */ |
1281 | address = canon_rtx (expanded_address); | |
6fb5fa3c | 1282 | |
403c7520 | 1283 | *offset = 0; |
6fb5fa3c | 1284 | |
456610d3 | 1285 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1286 | { |
1287 | if (expanded) | |
1288 | { | |
1289 | fprintf (dump_file, "\n after cselib_expand address: "); | |
1290 | print_inline_rtx (dump_file, expanded_address, 0); | |
1291 | fprintf (dump_file, "\n"); | |
1292 | } | |
6fb5fa3c | 1293 | |
403c7520 JJ |
1294 | fprintf (dump_file, "\n after canon_rtx address: "); |
1295 | print_inline_rtx (dump_file, address, 0); | |
1296 | fprintf (dump_file, "\n"); | |
1297 | } | |
6fb5fa3c | 1298 | |
403c7520 JJ |
1299 | if (GET_CODE (address) == CONST) |
1300 | address = XEXP (address, 0); | |
6fb5fa3c | 1301 | |
403c7520 JJ |
1302 | if (GET_CODE (address) == PLUS |
1303 | && CONST_INT_P (XEXP (address, 1))) | |
1304 | { | |
1305 | *offset = INTVAL (XEXP (address, 1)); | |
1306 | address = XEXP (address, 0); | |
1307 | } | |
6fb5fa3c | 1308 | |
09e881c9 BE |
1309 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1310 | && const_or_frame_p (address)) | |
6fb5fa3c | 1311 | { |
403c7520 JJ |
1312 | group_info_t group = get_group_info (address); |
1313 | ||
456610d3 | 1314 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1315 | fprintf (dump_file, " gid=%d offset=%d \n", |
1316 | group->id, (int)*offset); | |
1317 | *base = NULL; | |
1318 | *group_id = group->id; | |
1319 | return true; | |
6fb5fa3c | 1320 | } |
403c7520 JJ |
1321 | } |
1322 | ||
4deef538 | 1323 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
403c7520 JJ |
1324 | *group_id = -1; |
1325 | ||
1326 | if (*base == NULL) | |
1327 | { | |
456610d3 | 1328 | if (dump_file && (dump_flags & TDF_DETAILS)) |
403c7520 JJ |
1329 | fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
1330 | return false; | |
6fb5fa3c | 1331 | } |
456610d3 | 1332 | if (dump_file && (dump_flags & TDF_DETAILS)) |
5440c0e7 AO |
1333 | fprintf (dump_file, " varying cselib base=%u:%u offset = %d\n", |
1334 | (*base)->uid, (*base)->hash, (int)*offset); | |
6fb5fa3c DB |
1335 | return true; |
1336 | } | |
1337 | ||
1338 | ||
1339 | /* Clear the rhs field from the active_local_stores array. */ | |
1340 | ||
1341 | static void | |
1342 | clear_rhs_from_active_local_stores (void) | |
1343 | { | |
1344 | insn_info_t ptr = active_local_stores; | |
1345 | ||
1346 | while (ptr) | |
1347 | { | |
1348 | store_info_t store_info = ptr->store_rec; | |
1349 | /* Skip the clobbers. */ | |
1350 | while (!store_info->is_set) | |
1351 | store_info = store_info->next; | |
1352 | ||
1353 | store_info->rhs = NULL; | |
8dd5516b | 1354 | store_info->const_rhs = NULL; |
6fb5fa3c DB |
1355 | |
1356 | ptr = ptr->next_local_store; | |
1357 | } | |
1358 | } | |
1359 | ||
1360 | ||
8dd5516b JJ |
1361 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1362 | ||
1363 | static inline void | |
1364 | set_position_unneeded (store_info_t s_info, int pos) | |
1365 | { | |
1366 | if (__builtin_expect (s_info->is_large, false)) | |
1367 | { | |
fcaa4ca4 NF |
1368 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1369 | s_info->positions_needed.large.count++; | |
8dd5516b JJ |
1370 | } |
1371 | else | |
1372 | s_info->positions_needed.small_bitmask | |
1373 | &= ~(((unsigned HOST_WIDE_INT) 1) << pos); | |
1374 | } | |
1375 | ||
1376 | /* Mark the whole store S_INFO as unneeded. */ | |
1377 | ||
1378 | static inline void | |
1379 | set_all_positions_unneeded (store_info_t s_info) | |
1380 | { | |
1381 | if (__builtin_expect (s_info->is_large, false)) | |
1382 | { | |
1383 | int pos, end = s_info->end - s_info->begin; | |
1384 | for (pos = 0; pos < end; pos++) | |
dc491a25 | 1385 | bitmap_set_bit (s_info->positions_needed.large.bmap, pos); |
8dd5516b JJ |
1386 | s_info->positions_needed.large.count = end; |
1387 | } | |
1388 | else | |
1389 | s_info->positions_needed.small_bitmask = (unsigned HOST_WIDE_INT) 0; | |
1390 | } | |
1391 | ||
1392 | /* Return TRUE if any bytes from S_INFO store are needed. */ | |
1393 | ||
1394 | static inline bool | |
1395 | any_positions_needed_p (store_info_t s_info) | |
1396 | { | |
1397 | if (__builtin_expect (s_info->is_large, false)) | |
1398 | return (s_info->positions_needed.large.count | |
1399 | < s_info->end - s_info->begin); | |
1400 | else | |
1401 | return (s_info->positions_needed.small_bitmask | |
1402 | != (unsigned HOST_WIDE_INT) 0); | |
1403 | } | |
1404 | ||
1405 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
1406 | store are needed. */ | |
1407 | ||
1408 | static inline bool | |
1409 | all_positions_needed_p (store_info_t s_info, int start, int width) | |
1410 | { | |
1411 | if (__builtin_expect (s_info->is_large, false)) | |
1412 | { | |
1413 | int end = start + width; | |
1414 | while (start < end) | |
dc491a25 | 1415 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, start++)) |
8dd5516b JJ |
1416 | return false; |
1417 | return true; | |
1418 | } | |
1419 | else | |
1420 | { | |
1421 | unsigned HOST_WIDE_INT mask = lowpart_bitmask (width) << start; | |
1422 | return (s_info->positions_needed.small_bitmask & mask) == mask; | |
1423 | } | |
1424 | } | |
1425 | ||
1426 | ||
ef4bddc2 | 1427 | static rtx get_stored_val (store_info_t, machine_mode, HOST_WIDE_INT, |
8dd5516b JJ |
1428 | HOST_WIDE_INT, basic_block, bool); |
1429 | ||
1430 | ||
6fb5fa3c DB |
1431 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1432 | there is a candidate store, after adding it to the appropriate | |
1433 | local store group if so. */ | |
1434 | ||
1435 | static int | |
1436 | record_store (rtx body, bb_info_t bb_info) | |
1437 | { | |
6216f94e | 1438 | rtx mem, rhs, const_rhs, mem_addr; |
6fb5fa3c DB |
1439 | HOST_WIDE_INT offset = 0; |
1440 | HOST_WIDE_INT width = 0; | |
4862826d | 1441 | alias_set_type spill_alias_set; |
6fb5fa3c DB |
1442 | insn_info_t insn_info = bb_info->last_insn; |
1443 | store_info_t store_info = NULL; | |
1444 | int group_id; | |
1445 | cselib_val *base = NULL; | |
8dd5516b | 1446 | insn_info_t ptr, last, redundant_reason; |
6fb5fa3c DB |
1447 | bool store_is_unused; |
1448 | ||
1449 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
1450 | return 0; | |
1451 | ||
8dd5516b JJ |
1452 | mem = SET_DEST (body); |
1453 | ||
6fb5fa3c DB |
1454 | /* If this is not used, then this cannot be used to keep the insn |
1455 | from being deleted. On the other hand, it does provide something | |
1456 | that can be used to prove that another store is dead. */ | |
1457 | store_is_unused | |
8dd5516b | 1458 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
6fb5fa3c DB |
1459 | |
1460 | /* Check whether that value is a suitable memory location. */ | |
6fb5fa3c DB |
1461 | if (!MEM_P (mem)) |
1462 | { | |
1463 | /* If the set or clobber is unused, then it does not effect our | |
1464 | ability to get rid of the entire insn. */ | |
1465 | if (!store_is_unused) | |
1466 | insn_info->cannot_delete = true; | |
1467 | return 0; | |
1468 | } | |
1469 | ||
1470 | /* At this point we know mem is a mem. */ | |
1471 | if (GET_MODE (mem) == BLKmode) | |
1472 | { | |
1473 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
1474 | { | |
456610d3 | 1475 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1476 | fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1477 | add_wild_read (bb_info); | |
1478 | insn_info->cannot_delete = true; | |
8dd5516b | 1479 | return 0; |
6fb5fa3c | 1480 | } |
8dd5516b JJ |
1481 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1482 | as memset (addr, 0, 36); */ | |
f5541398 RS |
1483 | else if (!MEM_SIZE_KNOWN_P (mem) |
1484 | || MEM_SIZE (mem) <= 0 | |
1485 | || MEM_SIZE (mem) > MAX_OFFSET | |
8dd5516b | 1486 | || GET_CODE (body) != SET |
8dd5516b | 1487 | || !CONST_INT_P (SET_SRC (body))) |
6fb5fa3c | 1488 | { |
8dd5516b JJ |
1489 | if (!store_is_unused) |
1490 | { | |
1491 | /* If the set or clobber is unused, then it does not effect our | |
1492 | ability to get rid of the entire insn. */ | |
1493 | insn_info->cannot_delete = true; | |
1494 | clear_rhs_from_active_local_stores (); | |
1495 | } | |
1496 | return 0; | |
6fb5fa3c | 1497 | } |
6fb5fa3c DB |
1498 | } |
1499 | ||
1500 | /* We can still process a volatile mem, we just cannot delete it. */ | |
1501 | if (MEM_VOLATILE_P (mem)) | |
8dd5516b | 1502 | insn_info->cannot_delete = true; |
6fb5fa3c DB |
1503 | |
1504 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
1505 | { | |
1506 | clear_rhs_from_active_local_stores (); | |
1507 | return 0; | |
1508 | } | |
1509 | ||
8dd5516b | 1510 | if (GET_MODE (mem) == BLKmode) |
f5541398 | 1511 | width = MEM_SIZE (mem); |
8dd5516b | 1512 | else |
6f557e0e | 1513 | width = GET_MODE_SIZE (GET_MODE (mem)); |
6fb5fa3c DB |
1514 | |
1515 | if (spill_alias_set) | |
1516 | { | |
1517 | bitmap store1 = clear_alias_group->store1_p; | |
1518 | bitmap store2 = clear_alias_group->store2_p; | |
8dd5516b JJ |
1519 | |
1520 | gcc_assert (GET_MODE (mem) != BLKmode); | |
b8698a0f | 1521 | |
fcaa4ca4 | 1522 | if (!bitmap_set_bit (store1, spill_alias_set)) |
6fb5fa3c | 1523 | bitmap_set_bit (store2, spill_alias_set); |
b8698a0f | 1524 | |
6fb5fa3c DB |
1525 | if (clear_alias_group->offset_map_size_p < spill_alias_set) |
1526 | clear_alias_group->offset_map_size_p = spill_alias_set; | |
b8698a0f | 1527 | |
da6603c6 | 1528 | store_info = rtx_store_info_pool.allocate (); |
6fb5fa3c | 1529 | |
456610d3 | 1530 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 1531 | fprintf (dump_file, " processing spill store %d(%s)\n", |
4862826d | 1532 | (int) spill_alias_set, GET_MODE_NAME (GET_MODE (mem))); |
6fb5fa3c DB |
1533 | } |
1534 | else if (group_id >= 0) | |
1535 | { | |
1536 | /* In the restrictive case where the base is a constant or the | |
1537 | frame pointer we can do global analysis. */ | |
b8698a0f L |
1538 | |
1539 | group_info_t group | |
9771b263 | 1540 | = rtx_group_vec[group_id]; |
d26c7090 | 1541 | tree expr = MEM_EXPR (mem); |
b8698a0f | 1542 | |
da6603c6 | 1543 | store_info = rtx_store_info_pool.allocate (); |
d26c7090 | 1544 | set_usage_bits (group, offset, width, expr); |
6fb5fa3c | 1545 | |
456610d3 | 1546 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1547 | fprintf (dump_file, " processing const base store gid=%d[%d..%d)\n", |
1548 | group_id, (int)offset, (int)(offset+width)); | |
1549 | } | |
1550 | else | |
1551 | { | |
9e412ca3 | 1552 | if (may_be_sp_based_p (XEXP (mem, 0))) |
50f0f366 | 1553 | insn_info->stack_pointer_based = true; |
6fb5fa3c | 1554 | insn_info->contains_cselib_groups = true; |
50f0f366 | 1555 | |
da6603c6 | 1556 | store_info = cse_store_info_pool.allocate (); |
6fb5fa3c DB |
1557 | group_id = -1; |
1558 | ||
456610d3 | 1559 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
1560 | fprintf (dump_file, " processing cselib store [%d..%d)\n", |
1561 | (int)offset, (int)(offset+width)); | |
1562 | } | |
1563 | ||
8dd5516b JJ |
1564 | const_rhs = rhs = NULL_RTX; |
1565 | if (GET_CODE (body) == SET | |
1566 | /* No place to keep the value after ra. */ | |
1567 | && !reload_completed | |
1568 | && (REG_P (SET_SRC (body)) | |
1569 | || GET_CODE (SET_SRC (body)) == SUBREG | |
1570 | || CONSTANT_P (SET_SRC (body))) | |
1571 | && !MEM_VOLATILE_P (mem) | |
1572 | /* Sometimes the store and reload is used for truncation and | |
1573 | rounding. */ | |
1574 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
1575 | { | |
1576 | rhs = SET_SRC (body); | |
1577 | if (CONSTANT_P (rhs)) | |
1578 | const_rhs = rhs; | |
1579 | else if (body == PATTERN (insn_info->insn)) | |
1580 | { | |
1581 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
1582 | if (tem && CONSTANT_P (XEXP (tem, 0))) | |
1583 | const_rhs = XEXP (tem, 0); | |
1584 | } | |
1585 | if (const_rhs == NULL_RTX && REG_P (rhs)) | |
1586 | { | |
1587 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
1588 | ||
1589 | if (tem && CONSTANT_P (tem)) | |
1590 | const_rhs = tem; | |
1591 | } | |
1592 | } | |
1593 | ||
6fb5fa3c DB |
1594 | /* Check to see if this stores causes some other stores to be |
1595 | dead. */ | |
1596 | ptr = active_local_stores; | |
1597 | last = NULL; | |
8dd5516b | 1598 | redundant_reason = NULL; |
6216f94e JJ |
1599 | mem = canon_rtx (mem); |
1600 | /* For alias_set != 0 canon_true_dependence should be never called. */ | |
1601 | if (spill_alias_set) | |
1602 | mem_addr = NULL_RTX; | |
1603 | else | |
1604 | { | |
1605 | if (group_id < 0) | |
1606 | mem_addr = base->val_rtx; | |
1607 | else | |
1608 | { | |
1609 | group_info_t group | |
9771b263 | 1610 | = rtx_group_vec[group_id]; |
6216f94e JJ |
1611 | mem_addr = group->canon_base_addr; |
1612 | } | |
ff1803c1 WM |
1613 | /* get_addr can only handle VALUE but cannot handle expr like: |
1614 | VALUE + OFFSET, so call get_addr to get original addr for | |
1615 | mem_addr before plus_constant. */ | |
1616 | mem_addr = get_addr (mem_addr); | |
6216f94e | 1617 | if (offset) |
372d6395 | 1618 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6216f94e | 1619 | } |
6fb5fa3c DB |
1620 | |
1621 | while (ptr) | |
1622 | { | |
1623 | insn_info_t next = ptr->next_local_store; | |
1624 | store_info_t s_info = ptr->store_rec; | |
60564289 | 1625 | bool del = true; |
6fb5fa3c DB |
1626 | |
1627 | /* Skip the clobbers. We delete the active insn if this insn | |
6ed3da00 | 1628 | shadows the set. To have been put on the active list, it |
6fb5fa3c DB |
1629 | has exactly on set. */ |
1630 | while (!s_info->is_set) | |
1631 | s_info = s_info->next; | |
1632 | ||
1633 | if (s_info->alias_set != spill_alias_set) | |
60564289 | 1634 | del = false; |
6fb5fa3c DB |
1635 | else if (s_info->alias_set) |
1636 | { | |
b8698a0f | 1637 | struct clear_alias_mode_holder *entry |
6fb5fa3c DB |
1638 | = clear_alias_set_lookup (s_info->alias_set); |
1639 | /* Generally, spills cannot be processed if and of the | |
1640 | references to the slot have a different mode. But if | |
1641 | we are in the same block and mode is exactly the same | |
1642 | between this store and one before in the same block, | |
1643 | we can still delete it. */ | |
1644 | if ((GET_MODE (mem) == GET_MODE (s_info->mem)) | |
1645 | && (GET_MODE (mem) == entry->mode)) | |
1646 | { | |
60564289 | 1647 | del = true; |
8dd5516b | 1648 | set_all_positions_unneeded (s_info); |
6fb5fa3c | 1649 | } |
456610d3 | 1650 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 1651 | fprintf (dump_file, " trying spill store in insn=%d alias_set=%d\n", |
4862826d | 1652 | INSN_UID (ptr->insn), (int) s_info->alias_set); |
6fb5fa3c | 1653 | } |
b8698a0f | 1654 | else if ((s_info->group_id == group_id) |
6fb5fa3c DB |
1655 | && (s_info->cse_base == base)) |
1656 | { | |
1657 | HOST_WIDE_INT i; | |
456610d3 | 1658 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 1659 | fprintf (dump_file, " trying store in insn=%d gid=%d[%d..%d)\n", |
b8698a0f | 1660 | INSN_UID (ptr->insn), s_info->group_id, |
6fb5fa3c | 1661 | (int)s_info->begin, (int)s_info->end); |
8dd5516b JJ |
1662 | |
1663 | /* Even if PTR won't be eliminated as unneeded, if both | |
1664 | PTR and this insn store the same constant value, we might | |
1665 | eliminate this insn instead. */ | |
1666 | if (s_info->const_rhs | |
1667 | && const_rhs | |
1668 | && offset >= s_info->begin | |
1669 | && offset + width <= s_info->end | |
1670 | && all_positions_needed_p (s_info, offset - s_info->begin, | |
1671 | width)) | |
1672 | { | |
1673 | if (GET_MODE (mem) == BLKmode) | |
1674 | { | |
1675 | if (GET_MODE (s_info->mem) == BLKmode | |
1676 | && s_info->const_rhs == const_rhs) | |
1677 | redundant_reason = ptr; | |
1678 | } | |
1679 | else if (s_info->const_rhs == const0_rtx | |
1680 | && const_rhs == const0_rtx) | |
1681 | redundant_reason = ptr; | |
1682 | else | |
1683 | { | |
1684 | rtx val; | |
1685 | start_sequence (); | |
1686 | val = get_stored_val (s_info, GET_MODE (mem), | |
1687 | offset, offset + width, | |
1688 | BLOCK_FOR_INSN (insn_info->insn), | |
1689 | true); | |
1690 | if (get_insns () != NULL) | |
1691 | val = NULL_RTX; | |
1692 | end_sequence (); | |
1693 | if (val && rtx_equal_p (val, const_rhs)) | |
1694 | redundant_reason = ptr; | |
1695 | } | |
1696 | } | |
1697 | ||
1698 | for (i = MAX (offset, s_info->begin); | |
1699 | i < offset + width && i < s_info->end; | |
1700 | i++) | |
1701 | set_position_unneeded (s_info, i - s_info->begin); | |
6fb5fa3c DB |
1702 | } |
1703 | else if (s_info->rhs) | |
1704 | /* Need to see if it is possible for this store to overwrite | |
1705 | the value of store_info. If it is, set the rhs to NULL to | |
1706 | keep it from being used to remove a load. */ | |
1707 | { | |
b8698a0f | 1708 | if (canon_true_dependence (s_info->mem, |
6fb5fa3c DB |
1709 | GET_MODE (s_info->mem), |
1710 | s_info->mem_addr, | |
53d9622b | 1711 | mem, mem_addr)) |
8dd5516b JJ |
1712 | { |
1713 | s_info->rhs = NULL; | |
1714 | s_info->const_rhs = NULL; | |
1715 | } | |
6fb5fa3c | 1716 | } |
6216f94e | 1717 | |
6fb5fa3c DB |
1718 | /* An insn can be deleted if every position of every one of |
1719 | its s_infos is zero. */ | |
1b6fa860 | 1720 | if (any_positions_needed_p (s_info)) |
60564289 | 1721 | del = false; |
8dd5516b | 1722 | |
60564289 | 1723 | if (del) |
6fb5fa3c DB |
1724 | { |
1725 | insn_info_t insn_to_delete = ptr; | |
b8698a0f | 1726 | |
dabd47e7 | 1727 | active_local_stores_len--; |
6fb5fa3c DB |
1728 | if (last) |
1729 | last->next_local_store = ptr->next_local_store; | |
1730 | else | |
1731 | active_local_stores = ptr->next_local_store; | |
b8698a0f | 1732 | |
1b6fa860 JJ |
1733 | if (!insn_to_delete->cannot_delete) |
1734 | delete_dead_store_insn (insn_to_delete); | |
6fb5fa3c DB |
1735 | } |
1736 | else | |
1737 | last = ptr; | |
b8698a0f | 1738 | |
6fb5fa3c DB |
1739 | ptr = next; |
1740 | } | |
b8698a0f | 1741 | |
6fb5fa3c DB |
1742 | /* Finish filling in the store_info. */ |
1743 | store_info->next = insn_info->store_rec; | |
1744 | insn_info->store_rec = store_info; | |
6216f94e | 1745 | store_info->mem = mem; |
6fb5fa3c | 1746 | store_info->alias_set = spill_alias_set; |
6216f94e | 1747 | store_info->mem_addr = mem_addr; |
6fb5fa3c | 1748 | store_info->cse_base = base; |
8dd5516b JJ |
1749 | if (width > HOST_BITS_PER_WIDE_INT) |
1750 | { | |
1751 | store_info->is_large = true; | |
1752 | store_info->positions_needed.large.count = 0; | |
3f9b14ff | 1753 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
8dd5516b JJ |
1754 | } |
1755 | else | |
1756 | { | |
1757 | store_info->is_large = false; | |
1758 | store_info->positions_needed.small_bitmask = lowpart_bitmask (width); | |
1759 | } | |
6fb5fa3c DB |
1760 | store_info->group_id = group_id; |
1761 | store_info->begin = offset; | |
1762 | store_info->end = offset + width; | |
1763 | store_info->is_set = GET_CODE (body) == SET; | |
8dd5516b JJ |
1764 | store_info->rhs = rhs; |
1765 | store_info->const_rhs = const_rhs; | |
1766 | store_info->redundant_reason = redundant_reason; | |
6fb5fa3c | 1767 | |
6fb5fa3c DB |
1768 | /* If this is a clobber, we return 0. We will only be able to |
1769 | delete this insn if there is only one store USED store, but we | |
1770 | can use the clobber to delete other stores earlier. */ | |
1771 | return store_info->is_set ? 1 : 0; | |
1772 | } | |
1773 | ||
1774 | ||
1775 | static void | |
1776 | dump_insn_info (const char * start, insn_info_t insn_info) | |
1777 | { | |
b8698a0f | 1778 | fprintf (dump_file, "%s insn=%d %s\n", start, |
6fb5fa3c DB |
1779 | INSN_UID (insn_info->insn), |
1780 | insn_info->store_rec ? "has store" : "naked"); | |
1781 | } | |
1782 | ||
1783 | ||
8660aaae EC |
1784 | /* If the modes are different and the value's source and target do not |
1785 | line up, we need to extract the value from lower part of the rhs of | |
1786 | the store, shift it, and then put it into a form that can be shoved | |
1787 | into the read_insn. This function generates a right SHIFT of a | |
1788 | value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
1789 | shift sequence is returned or NULL if we failed to find a | |
1790 | shift. */ | |
1791 | ||
1792 | static rtx | |
18b526e8 | 1793 | find_shift_sequence (int access_size, |
8660aaae | 1794 | store_info_t store_info, |
ef4bddc2 | 1795 | machine_mode read_mode, |
8dd5516b | 1796 | int shift, bool speed, bool require_cst) |
8660aaae | 1797 | { |
ef4bddc2 RS |
1798 | machine_mode store_mode = GET_MODE (store_info->mem); |
1799 | machine_mode new_mode; | |
18b526e8 | 1800 | rtx read_reg = NULL; |
8660aaae EC |
1801 | |
1802 | /* Some machines like the x86 have shift insns for each size of | |
1803 | operand. Other machines like the ppc or the ia-64 may only have | |
1804 | shift insns that shift values within 32 or 64 bit registers. | |
1805 | This loop tries to find the smallest shift insn that will right | |
1806 | justify the value we want to read but is available in one insn on | |
1807 | the machine. */ | |
1808 | ||
18b526e8 RS |
1809 | for (new_mode = smallest_mode_for_size (access_size * BITS_PER_UNIT, |
1810 | MODE_INT); | |
1811 | GET_MODE_BITSIZE (new_mode) <= BITS_PER_WORD; | |
1812 | new_mode = GET_MODE_WIDER_MODE (new_mode)) | |
8660aaae | 1813 | { |
dc01c3d1 DM |
1814 | rtx target, new_reg, new_lhs; |
1815 | rtx_insn *shift_seq, *insn; | |
d898d29b | 1816 | int cost; |
348eea5f | 1817 | |
72a2609f JJ |
1818 | /* If a constant was stored into memory, try to simplify it here, |
1819 | otherwise the cost of the shift might preclude this optimization | |
1820 | e.g. at -Os, even when no actual shift will be needed. */ | |
8dd5516b | 1821 | if (store_info->const_rhs) |
72a2609f JJ |
1822 | { |
1823 | unsigned int byte = subreg_lowpart_offset (new_mode, store_mode); | |
8dd5516b JJ |
1824 | rtx ret = simplify_subreg (new_mode, store_info->const_rhs, |
1825 | store_mode, byte); | |
72a2609f JJ |
1826 | if (ret && CONSTANT_P (ret)) |
1827 | { | |
1828 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, | |
1829 | ret, GEN_INT (shift)); | |
1830 | if (ret && CONSTANT_P (ret)) | |
1831 | { | |
1832 | byte = subreg_lowpart_offset (read_mode, new_mode); | |
1833 | ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
1834 | if (ret && CONSTANT_P (ret) | |
e548c9df AM |
1835 | && (set_src_cost (ret, read_mode, speed) |
1836 | <= COSTS_N_INSNS (1))) | |
72a2609f JJ |
1837 | return ret; |
1838 | } | |
1839 | } | |
1840 | } | |
1841 | ||
8dd5516b JJ |
1842 | if (require_cst) |
1843 | return NULL_RTX; | |
1844 | ||
18b526e8 RS |
1845 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1846 | requires a real instruction. */ | |
1847 | if (GET_MODE_BITSIZE (new_mode) < GET_MODE_BITSIZE (store_mode) | |
d0edd768 | 1848 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
348eea5f RS |
1849 | continue; |
1850 | ||
18b526e8 RS |
1851 | /* Also try a wider mode if the necessary punning is either not |
1852 | desirable or not possible. */ | |
1853 | if (!CONSTANT_P (store_info->rhs) | |
1854 | && !MODES_TIEABLE_P (new_mode, store_mode)) | |
1855 | continue; | |
18b526e8 | 1856 | |
348eea5f | 1857 | new_reg = gen_reg_rtx (new_mode); |
8660aaae EC |
1858 | |
1859 | start_sequence (); | |
1860 | ||
1861 | /* In theory we could also check for an ashr. Ian Taylor knows | |
1862 | of one dsp where the cost of these two was not the same. But | |
1863 | this really is a rare case anyway. */ | |
1864 | target = expand_binop (new_mode, lshr_optab, new_reg, | |
1865 | GEN_INT (shift), new_reg, 1, OPTAB_DIRECT); | |
1866 | ||
c6f3019a RS |
1867 | shift_seq = get_insns (); |
1868 | end_sequence (); | |
8660aaae | 1869 | |
c6f3019a RS |
1870 | if (target != new_reg || shift_seq == NULL) |
1871 | continue; | |
1872 | ||
1873 | cost = 0; | |
1874 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
1875 | if (INSN_P (insn)) | |
f40751dd | 1876 | cost += insn_rtx_cost (PATTERN (insn), speed); |
c6f3019a RS |
1877 | |
1878 | /* The computation up to here is essentially independent | |
1879 | of the arguments and could be precomputed. It may | |
1880 | not be worth doing so. We could precompute if | |
1881 | worthwhile or at least cache the results. The result | |
06acf7d0 RS |
1882 | technically depends on both SHIFT and ACCESS_SIZE, |
1883 | but in practice the answer will depend only on ACCESS_SIZE. */ | |
c6f3019a RS |
1884 | |
1885 | if (cost > COSTS_N_INSNS (1)) | |
1886 | continue; | |
1887 | ||
d898d29b JJ |
1888 | new_lhs = extract_low_bits (new_mode, store_mode, |
1889 | copy_rtx (store_info->rhs)); | |
1890 | if (new_lhs == NULL_RTX) | |
1891 | continue; | |
1892 | ||
c6f3019a RS |
1893 | /* We found an acceptable shift. Generate a move to |
1894 | take the value from the store and put it into the | |
1895 | shift pseudo, then shift it, then generate another | |
1896 | move to put in into the target of the read. */ | |
18b526e8 | 1897 | emit_move_insn (new_reg, new_lhs); |
c6f3019a | 1898 | emit_insn (shift_seq); |
18b526e8 | 1899 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
c6f3019a | 1900 | break; |
8660aaae EC |
1901 | } |
1902 | ||
18b526e8 | 1903 | return read_reg; |
8660aaae EC |
1904 | } |
1905 | ||
1906 | ||
02b47899 KZ |
1907 | /* Call back for note_stores to find the hard regs set or clobbered by |
1908 | insn. Data is a bitmap of the hardregs set so far. */ | |
1909 | ||
1910 | static void | |
1911 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
1912 | { | |
1913 | bitmap regs_set = (bitmap) data; | |
1914 | ||
1915 | if (REG_P (x) | |
f773c2bd | 1916 | && HARD_REGISTER_P (x)) |
dc8afb70 | 1917 | bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); |
02b47899 KZ |
1918 | } |
1919 | ||
8dd5516b JJ |
1920 | /* Helper function for replace_read and record_store. |
1921 | Attempt to return a value stored in STORE_INFO, from READ_BEGIN | |
1922 | to one before READ_END bytes read in READ_MODE. Return NULL | |
1923 | if not successful. If REQUIRE_CST is true, return always constant. */ | |
1924 | ||
1925 | static rtx | |
ef4bddc2 | 1926 | get_stored_val (store_info_t store_info, machine_mode read_mode, |
8dd5516b JJ |
1927 | HOST_WIDE_INT read_begin, HOST_WIDE_INT read_end, |
1928 | basic_block bb, bool require_cst) | |
1929 | { | |
ef4bddc2 | 1930 | machine_mode store_mode = GET_MODE (store_info->mem); |
8dd5516b JJ |
1931 | int shift; |
1932 | int access_size; /* In bytes. */ | |
1933 | rtx read_reg; | |
1934 | ||
1935 | /* To get here the read is within the boundaries of the write so | |
1936 | shift will never be negative. Start out with the shift being in | |
1937 | bytes. */ | |
1938 | if (store_mode == BLKmode) | |
1939 | shift = 0; | |
1940 | else if (BYTES_BIG_ENDIAN) | |
1941 | shift = store_info->end - read_end; | |
1942 | else | |
1943 | shift = read_begin - store_info->begin; | |
1944 | ||
1945 | access_size = shift + GET_MODE_SIZE (read_mode); | |
1946 | ||
1947 | /* From now on it is bits. */ | |
1948 | shift *= BITS_PER_UNIT; | |
1949 | ||
1950 | if (shift) | |
1951 | read_reg = find_shift_sequence (access_size, store_info, read_mode, shift, | |
1952 | optimize_bb_for_speed_p (bb), | |
1953 | require_cst); | |
1954 | else if (store_mode == BLKmode) | |
1955 | { | |
1956 | /* The store is a memset (addr, const_val, const_size). */ | |
1957 | gcc_assert (CONST_INT_P (store_info->rhs)); | |
1958 | store_mode = int_mode_for_mode (read_mode); | |
1959 | if (store_mode == BLKmode) | |
1960 | read_reg = NULL_RTX; | |
1961 | else if (store_info->rhs == const0_rtx) | |
1962 | read_reg = extract_low_bits (read_mode, store_mode, const0_rtx); | |
1963 | else if (GET_MODE_BITSIZE (store_mode) > HOST_BITS_PER_WIDE_INT | |
1964 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) | |
1965 | read_reg = NULL_RTX; | |
1966 | else | |
1967 | { | |
1968 | unsigned HOST_WIDE_INT c | |
1969 | = INTVAL (store_info->rhs) | |
1970 | & (((HOST_WIDE_INT) 1 << BITS_PER_UNIT) - 1); | |
1971 | int shift = BITS_PER_UNIT; | |
1972 | while (shift < HOST_BITS_PER_WIDE_INT) | |
1973 | { | |
1974 | c |= (c << shift); | |
1975 | shift <<= 1; | |
1976 | } | |
6d26322f | 1977 | read_reg = gen_int_mode (c, store_mode); |
8dd5516b JJ |
1978 | read_reg = extract_low_bits (read_mode, store_mode, read_reg); |
1979 | } | |
1980 | } | |
1981 | else if (store_info->const_rhs | |
1982 | && (require_cst | |
1983 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
1984 | read_reg = extract_low_bits (read_mode, store_mode, | |
1985 | copy_rtx (store_info->const_rhs)); | |
1986 | else | |
1987 | read_reg = extract_low_bits (read_mode, store_mode, | |
1988 | copy_rtx (store_info->rhs)); | |
1989 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
1990 | read_reg = NULL_RTX; | |
1991 | return read_reg; | |
1992 | } | |
02b47899 | 1993 | |
6fb5fa3c DB |
1994 | /* Take a sequence of: |
1995 | A <- r1 | |
1996 | ... | |
1997 | ... <- A | |
1998 | ||
b8698a0f | 1999 | and change it into |
6fb5fa3c DB |
2000 | r2 <- r1 |
2001 | A <- r1 | |
2002 | ... | |
2003 | ... <- r2 | |
2004 | ||
8660aaae EC |
2005 | or |
2006 | ||
2007 | r3 <- extract (r1) | |
2008 | r3 <- r3 >> shift | |
2009 | r2 <- extract (r3) | |
2010 | ... <- r2 | |
2011 | ||
2012 | or | |
2013 | ||
2014 | r2 <- extract (r1) | |
2015 | ... <- r2 | |
2016 | ||
2017 | Depending on the alignment and the mode of the store and | |
2018 | subsequent load. | |
2019 | ||
2020 | ||
2021 | The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
6fb5fa3c DB |
2022 | and READ_INSN are for the read. Return true if the replacement |
2023 | went ok. */ | |
2024 | ||
2025 | static bool | |
b8698a0f | 2026 | replace_read (store_info_t store_info, insn_info_t store_insn, |
8dd5516b JJ |
2027 | read_info_t read_info, insn_info_t read_insn, rtx *loc, |
2028 | bitmap regs_live) | |
6fb5fa3c | 2029 | { |
ef4bddc2 RS |
2030 | machine_mode store_mode = GET_MODE (store_info->mem); |
2031 | machine_mode read_mode = GET_MODE (read_info->mem); | |
dc01c3d1 DM |
2032 | rtx_insn *insns, *this_insn; |
2033 | rtx read_reg; | |
8dd5516b | 2034 | basic_block bb; |
8660aaae | 2035 | |
6fb5fa3c DB |
2036 | if (!dbg_cnt (dse)) |
2037 | return false; | |
2038 | ||
18b526e8 RS |
2039 | /* Create a sequence of instructions to set up the read register. |
2040 | This sequence goes immediately before the store and its result | |
2041 | is read by the load. | |
2042 | ||
2043 | We need to keep this in perspective. We are replacing a read | |
8660aaae EC |
2044 | with a sequence of insns, but the read will almost certainly be |
2045 | in cache, so it is not going to be an expensive one. Thus, we | |
2046 | are not willing to do a multi insn shift or worse a subroutine | |
2047 | call to get rid of the read. */ | |
456610d3 | 2048 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2049 | fprintf (dump_file, "trying to replace %smode load in insn %d" |
2050 | " from %smode store in insn %d\n", | |
2051 | GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
2052 | GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
2053 | start_sequence (); | |
8dd5516b JJ |
2054 | bb = BLOCK_FOR_INSN (read_insn->insn); |
2055 | read_reg = get_stored_val (store_info, | |
2056 | read_mode, read_info->begin, read_info->end, | |
2057 | bb, false); | |
18b526e8 | 2058 | if (read_reg == NULL_RTX) |
8660aaae | 2059 | { |
18b526e8 | 2060 | end_sequence (); |
456610d3 | 2061 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2062 | fprintf (dump_file, " -- could not extract bits of stored value\n"); |
2063 | return false; | |
8660aaae | 2064 | } |
18b526e8 RS |
2065 | /* Force the value into a new register so that it won't be clobbered |
2066 | between the store and the load. */ | |
2067 | read_reg = copy_to_mode_reg (read_mode, read_reg); | |
2068 | insns = get_insns (); | |
2069 | end_sequence (); | |
8660aaae | 2070 | |
02b47899 KZ |
2071 | if (insns != NULL_RTX) |
2072 | { | |
2073 | /* Now we have to scan the set of new instructions to see if the | |
2074 | sequence contains and sets of hardregs that happened to be | |
2075 | live at this point. For instance, this can happen if one of | |
2076 | the insns sets the CC and the CC happened to be live at that | |
2077 | point. This does occasionally happen, see PR 37922. */ | |
3f9b14ff | 2078 | bitmap regs_set = BITMAP_ALLOC (®_obstack); |
02b47899 KZ |
2079 | |
2080 | for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
2081 | note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
b8698a0f | 2082 | |
02b47899 KZ |
2083 | bitmap_and_into (regs_set, regs_live); |
2084 | if (!bitmap_empty_p (regs_set)) | |
2085 | { | |
456610d3 | 2086 | if (dump_file && (dump_flags & TDF_DETAILS)) |
02b47899 | 2087 | { |
b8698a0f | 2088 | fprintf (dump_file, |
02b47899 KZ |
2089 | "abandoning replacement because sequence clobbers live hardregs:"); |
2090 | df_print_regset (dump_file, regs_set); | |
2091 | } | |
b8698a0f | 2092 | |
02b47899 KZ |
2093 | BITMAP_FREE (regs_set); |
2094 | return false; | |
2095 | } | |
2096 | BITMAP_FREE (regs_set); | |
2097 | } | |
2098 | ||
8660aaae | 2099 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
6fb5fa3c | 2100 | { |
da6603c6 | 2101 | deferred_change_t change = new deferred_change; |
b8698a0f | 2102 | |
8660aaae EC |
2103 | /* Insert this right before the store insn where it will be safe |
2104 | from later insns that might change it before the read. */ | |
2105 | emit_insn_before (insns, store_insn->insn); | |
b8698a0f | 2106 | |
8660aaae EC |
2107 | /* And now for the kludge part: cselib croaks if you just |
2108 | return at this point. There are two reasons for this: | |
b8698a0f | 2109 | |
8660aaae EC |
2110 | 1) Cselib has an idea of how many pseudos there are and |
2111 | that does not include the new ones we just added. | |
b8698a0f | 2112 | |
8660aaae EC |
2113 | 2) Cselib does not know about the move insn we added |
2114 | above the store_info, and there is no way to tell it | |
2115 | about it, because it has "moved on". | |
b8698a0f | 2116 | |
8660aaae EC |
2117 | Problem (1) is fixable with a certain amount of engineering. |
2118 | Problem (2) is requires starting the bb from scratch. This | |
2119 | could be expensive. | |
b8698a0f | 2120 | |
8660aaae EC |
2121 | So we are just going to have to lie. The move/extraction |
2122 | insns are not really an issue, cselib did not see them. But | |
2123 | the use of the new pseudo read_insn is a real problem because | |
2124 | cselib has not scanned this insn. The way that we solve this | |
2125 | problem is that we are just going to put the mem back for now | |
2126 | and when we are finished with the block, we undo this. We | |
2127 | keep a table of mems to get rid of. At the end of the basic | |
2128 | block we can put them back. */ | |
b8698a0f | 2129 | |
8660aaae | 2130 | *loc = read_info->mem; |
da6603c6 ML |
2131 | change->next = deferred_change_list; |
2132 | deferred_change_list = change; | |
2133 | change->loc = loc; | |
2134 | change->reg = read_reg; | |
b8698a0f | 2135 | |
8660aaae EC |
2136 | /* Get rid of the read_info, from the point of view of the |
2137 | rest of dse, play like this read never happened. */ | |
2138 | read_insn->read_rec = read_info->next; | |
da6603c6 | 2139 | delete read_info; |
456610d3 | 2140 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2141 | { |
2142 | fprintf (dump_file, " -- replaced the loaded MEM with "); | |
2143 | print_simple_rtl (dump_file, read_reg); | |
2144 | fprintf (dump_file, "\n"); | |
2145 | } | |
8660aaae | 2146 | return true; |
6fb5fa3c | 2147 | } |
b8698a0f | 2148 | else |
6fb5fa3c | 2149 | { |
456610d3 | 2150 | if (dump_file && (dump_flags & TDF_DETAILS)) |
18b526e8 RS |
2151 | { |
2152 | fprintf (dump_file, " -- replacing the loaded MEM with "); | |
2153 | print_simple_rtl (dump_file, read_reg); | |
2154 | fprintf (dump_file, " led to an invalid instruction\n"); | |
2155 | } | |
6fb5fa3c DB |
2156 | return false; |
2157 | } | |
2158 | } | |
2159 | ||
d7111da8 RS |
2160 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2161 | be active. */ | |
6fb5fa3c | 2162 | |
d7111da8 RS |
2163 | static void |
2164 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info) | |
6fb5fa3c | 2165 | { |
6216f94e | 2166 | rtx mem = *loc, mem_addr; |
6fb5fa3c DB |
2167 | insn_info_t insn_info; |
2168 | HOST_WIDE_INT offset = 0; | |
2169 | HOST_WIDE_INT width = 0; | |
4862826d | 2170 | alias_set_type spill_alias_set = 0; |
b8698a0f | 2171 | cselib_val *base = NULL; |
6fb5fa3c DB |
2172 | int group_id; |
2173 | read_info_t read_info; | |
2174 | ||
6fb5fa3c DB |
2175 | insn_info = bb_info->last_insn; |
2176 | ||
2177 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
2178 | || (MEM_VOLATILE_P (mem))) | |
2179 | { | |
456610d3 | 2180 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2181 | fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
2182 | add_wild_read (bb_info); | |
2183 | insn_info->cannot_delete = true; | |
d7111da8 | 2184 | return; |
6fb5fa3c DB |
2185 | } |
2186 | ||
2187 | /* If it is reading readonly mem, then there can be no conflict with | |
2188 | another write. */ | |
2189 | if (MEM_READONLY_P (mem)) | |
d7111da8 | 2190 | return; |
6fb5fa3c DB |
2191 | |
2192 | if (!canon_address (mem, &spill_alias_set, &group_id, &offset, &base)) | |
2193 | { | |
456610d3 | 2194 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2195 | fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
2196 | add_wild_read (bb_info); | |
d7111da8 | 2197 | return; |
6fb5fa3c DB |
2198 | } |
2199 | ||
2200 | if (GET_MODE (mem) == BLKmode) | |
2201 | width = -1; | |
2202 | else | |
2203 | width = GET_MODE_SIZE (GET_MODE (mem)); | |
2204 | ||
da6603c6 | 2205 | read_info = new read_info_type; |
6fb5fa3c DB |
2206 | read_info->group_id = group_id; |
2207 | read_info->mem = mem; | |
2208 | read_info->alias_set = spill_alias_set; | |
2209 | read_info->begin = offset; | |
2210 | read_info->end = offset + width; | |
2211 | read_info->next = insn_info->read_rec; | |
2212 | insn_info->read_rec = read_info; | |
6216f94e JJ |
2213 | /* For alias_set != 0 canon_true_dependence should be never called. */ |
2214 | if (spill_alias_set) | |
2215 | mem_addr = NULL_RTX; | |
2216 | else | |
2217 | { | |
2218 | if (group_id < 0) | |
2219 | mem_addr = base->val_rtx; | |
2220 | else | |
2221 | { | |
2222 | group_info_t group | |
9771b263 | 2223 | = rtx_group_vec[group_id]; |
6216f94e JJ |
2224 | mem_addr = group->canon_base_addr; |
2225 | } | |
ff1803c1 WM |
2226 | /* get_addr can only handle VALUE but cannot handle expr like: |
2227 | VALUE + OFFSET, so call get_addr to get original addr for | |
2228 | mem_addr before plus_constant. */ | |
2229 | mem_addr = get_addr (mem_addr); | |
6216f94e | 2230 | if (offset) |
372d6395 | 2231 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
6216f94e | 2232 | } |
6fb5fa3c | 2233 | |
0d52bcc1 | 2234 | /* We ignore the clobbers in store_info. The is mildly aggressive, |
6fb5fa3c DB |
2235 | but there really should not be a clobber followed by a read. */ |
2236 | ||
2237 | if (spill_alias_set) | |
2238 | { | |
2239 | insn_info_t i_ptr = active_local_stores; | |
2240 | insn_info_t last = NULL; | |
2241 | ||
456610d3 | 2242 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2243 | fprintf (dump_file, " processing spill load %d\n", |
4862826d | 2244 | (int) spill_alias_set); |
6fb5fa3c DB |
2245 | |
2246 | while (i_ptr) | |
2247 | { | |
2248 | store_info_t store_info = i_ptr->store_rec; | |
2249 | ||
2250 | /* Skip the clobbers. */ | |
2251 | while (!store_info->is_set) | |
2252 | store_info = store_info->next; | |
b8698a0f | 2253 | |
6fb5fa3c DB |
2254 | if (store_info->alias_set == spill_alias_set) |
2255 | { | |
456610d3 | 2256 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2257 | dump_insn_info ("removing from active", i_ptr); |
2258 | ||
dabd47e7 | 2259 | active_local_stores_len--; |
6fb5fa3c DB |
2260 | if (last) |
2261 | last->next_local_store = i_ptr->next_local_store; | |
2262 | else | |
2263 | active_local_stores = i_ptr->next_local_store; | |
2264 | } | |
2265 | else | |
2266 | last = i_ptr; | |
2267 | i_ptr = i_ptr->next_local_store; | |
2268 | } | |
2269 | } | |
2270 | else if (group_id >= 0) | |
2271 | { | |
2272 | /* This is the restricted case where the base is a constant or | |
2273 | the frame pointer and offset is a constant. */ | |
2274 | insn_info_t i_ptr = active_local_stores; | |
2275 | insn_info_t last = NULL; | |
b8698a0f | 2276 | |
456610d3 | 2277 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2278 | { |
2279 | if (width == -1) | |
2280 | fprintf (dump_file, " processing const load gid=%d[BLK]\n", | |
2281 | group_id); | |
2282 | else | |
2283 | fprintf (dump_file, " processing const load gid=%d[%d..%d)\n", | |
2284 | group_id, (int)offset, (int)(offset+width)); | |
2285 | } | |
2286 | ||
2287 | while (i_ptr) | |
2288 | { | |
2289 | bool remove = false; | |
2290 | store_info_t store_info = i_ptr->store_rec; | |
b8698a0f | 2291 | |
6fb5fa3c DB |
2292 | /* Skip the clobbers. */ |
2293 | while (!store_info->is_set) | |
2294 | store_info = store_info->next; | |
b8698a0f | 2295 | |
6fb5fa3c DB |
2296 | /* There are three cases here. */ |
2297 | if (store_info->group_id < 0) | |
2298 | /* We have a cselib store followed by a read from a | |
2299 | const base. */ | |
b8698a0f L |
2300 | remove |
2301 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2302 | GET_MODE (store_info->mem), |
2303 | store_info->mem_addr, | |
53d9622b | 2304 | mem, mem_addr); |
b8698a0f | 2305 | |
6fb5fa3c DB |
2306 | else if (group_id == store_info->group_id) |
2307 | { | |
2308 | /* This is a block mode load. We may get lucky and | |
2309 | canon_true_dependence may save the day. */ | |
2310 | if (width == -1) | |
b8698a0f L |
2311 | remove |
2312 | = canon_true_dependence (store_info->mem, | |
6fb5fa3c DB |
2313 | GET_MODE (store_info->mem), |
2314 | store_info->mem_addr, | |
53d9622b | 2315 | mem, mem_addr); |
b8698a0f | 2316 | |
6fb5fa3c DB |
2317 | /* If this read is just reading back something that we just |
2318 | stored, rewrite the read. */ | |
b8698a0f | 2319 | else |
6fb5fa3c DB |
2320 | { |
2321 | if (store_info->rhs | |
8dd5516b JJ |
2322 | && offset >= store_info->begin |
2323 | && offset + width <= store_info->end | |
2324 | && all_positions_needed_p (store_info, | |
2325 | offset - store_info->begin, | |
2326 | width) | |
2327 | && replace_read (store_info, i_ptr, read_info, | |
2328 | insn_info, loc, bb_info->regs_live)) | |
d7111da8 | 2329 | return; |
8dd5516b | 2330 | |
6fb5fa3c DB |
2331 | /* The bases are the same, just see if the offsets |
2332 | overlap. */ | |
b8698a0f | 2333 | if ((offset < store_info->end) |
6fb5fa3c DB |
2334 | && (offset + width > store_info->begin)) |
2335 | remove = true; | |
2336 | } | |
2337 | } | |
b8698a0f L |
2338 | |
2339 | /* else | |
6fb5fa3c DB |
2340 | The else case that is missing here is that the |
2341 | bases are constant but different. There is nothing | |
2342 | to do here because there is no overlap. */ | |
b8698a0f | 2343 | |
6fb5fa3c DB |
2344 | if (remove) |
2345 | { | |
456610d3 | 2346 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2347 | dump_insn_info ("removing from active", i_ptr); |
2348 | ||
dabd47e7 | 2349 | active_local_stores_len--; |
6fb5fa3c DB |
2350 | if (last) |
2351 | last->next_local_store = i_ptr->next_local_store; | |
2352 | else | |
2353 | active_local_stores = i_ptr->next_local_store; | |
2354 | } | |
2355 | else | |
2356 | last = i_ptr; | |
2357 | i_ptr = i_ptr->next_local_store; | |
2358 | } | |
2359 | } | |
b8698a0f | 2360 | else |
6fb5fa3c DB |
2361 | { |
2362 | insn_info_t i_ptr = active_local_stores; | |
2363 | insn_info_t last = NULL; | |
456610d3 | 2364 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2365 | { |
2366 | fprintf (dump_file, " processing cselib load mem:"); | |
2367 | print_inline_rtx (dump_file, mem, 0); | |
2368 | fprintf (dump_file, "\n"); | |
2369 | } | |
2370 | ||
2371 | while (i_ptr) | |
2372 | { | |
2373 | bool remove = false; | |
2374 | store_info_t store_info = i_ptr->store_rec; | |
b8698a0f | 2375 | |
456610d3 | 2376 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2377 | fprintf (dump_file, " processing cselib load against insn %d\n", |
2378 | INSN_UID (i_ptr->insn)); | |
2379 | ||
2380 | /* Skip the clobbers. */ | |
2381 | while (!store_info->is_set) | |
2382 | store_info = store_info->next; | |
2383 | ||
2384 | /* If this read is just reading back something that we just | |
2385 | stored, rewrite the read. */ | |
2386 | if (store_info->rhs | |
2387 | && store_info->group_id == -1 | |
2388 | && store_info->cse_base == base | |
efc3527a | 2389 | && width != -1 |
8dd5516b JJ |
2390 | && offset >= store_info->begin |
2391 | && offset + width <= store_info->end | |
2392 | && all_positions_needed_p (store_info, | |
2393 | offset - store_info->begin, width) | |
2394 | && replace_read (store_info, i_ptr, read_info, insn_info, loc, | |
2395 | bb_info->regs_live)) | |
d7111da8 | 2396 | return; |
6fb5fa3c DB |
2397 | |
2398 | if (!store_info->alias_set) | |
b8698a0f | 2399 | remove = canon_true_dependence (store_info->mem, |
6fb5fa3c DB |
2400 | GET_MODE (store_info->mem), |
2401 | store_info->mem_addr, | |
53d9622b | 2402 | mem, mem_addr); |
b8698a0f | 2403 | |
6fb5fa3c DB |
2404 | if (remove) |
2405 | { | |
456610d3 | 2406 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2407 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2408 | |
dabd47e7 | 2409 | active_local_stores_len--; |
6fb5fa3c DB |
2410 | if (last) |
2411 | last->next_local_store = i_ptr->next_local_store; | |
2412 | else | |
2413 | active_local_stores = i_ptr->next_local_store; | |
2414 | } | |
2415 | else | |
2416 | last = i_ptr; | |
2417 | i_ptr = i_ptr->next_local_store; | |
2418 | } | |
2419 | } | |
6fb5fa3c DB |
2420 | } |
2421 | ||
d7111da8 | 2422 | /* A note_uses callback in which DATA points the INSN_INFO for |
6fb5fa3c DB |
2423 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2424 | true for any part of *LOC. */ | |
2425 | ||
2426 | static void | |
2427 | check_mem_read_use (rtx *loc, void *data) | |
2428 | { | |
d7111da8 RS |
2429 | subrtx_ptr_iterator::array_type array; |
2430 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
2431 | { | |
2432 | rtx *loc = *iter; | |
2433 | if (MEM_P (*loc)) | |
2434 | check_mem_read_rtx (loc, (bb_info_t) data); | |
2435 | } | |
6fb5fa3c DB |
2436 | } |
2437 | ||
8dd5516b JJ |
2438 | |
2439 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
2440 | So far it only handles arguments passed in registers. */ | |
2441 | ||
2442 | static bool | |
2443 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
2444 | { | |
d5cc9181 JR |
2445 | CUMULATIVE_ARGS args_so_far_v; |
2446 | cumulative_args_t args_so_far; | |
8dd5516b JJ |
2447 | tree arg; |
2448 | int idx; | |
2449 | ||
d5cc9181 JR |
2450 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2451 | args_so_far = pack_cumulative_args (&args_so_far_v); | |
8dd5516b JJ |
2452 | |
2453 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
2454 | for (idx = 0; | |
2455 | arg != void_list_node && idx < nargs; | |
2456 | arg = TREE_CHAIN (arg), idx++) | |
2457 | { | |
ef4bddc2 | 2458 | machine_mode mode = TYPE_MODE (TREE_VALUE (arg)); |
3c07301f | 2459 | rtx reg, link, tmp; |
d5cc9181 | 2460 | reg = targetm.calls.function_arg (args_so_far, mode, NULL_TREE, true); |
8dd5516b JJ |
2461 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode |
2462 | || GET_MODE_CLASS (mode) != MODE_INT) | |
2463 | return false; | |
2464 | ||
2465 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
2466 | link; | |
2467 | link = XEXP (link, 1)) | |
2468 | if (GET_CODE (XEXP (link, 0)) == USE) | |
2469 | { | |
2470 | args[idx] = XEXP (XEXP (link, 0), 0); | |
2471 | if (REG_P (args[idx]) | |
2472 | && REGNO (args[idx]) == REGNO (reg) | |
2473 | && (GET_MODE (args[idx]) == mode | |
2474 | || (GET_MODE_CLASS (GET_MODE (args[idx])) == MODE_INT | |
2475 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2476 | <= UNITS_PER_WORD) | |
2477 | && (GET_MODE_SIZE (GET_MODE (args[idx])) | |
2478 | > GET_MODE_SIZE (mode))))) | |
2479 | break; | |
2480 | } | |
2481 | if (!link) | |
2482 | return false; | |
2483 | ||
2484 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
2485 | if (GET_MODE (args[idx]) != mode) | |
2486 | { | |
2487 | if (!tmp || !CONST_INT_P (tmp)) | |
2488 | return false; | |
6d26322f | 2489 | tmp = gen_int_mode (INTVAL (tmp), mode); |
8dd5516b JJ |
2490 | } |
2491 | if (tmp) | |
2492 | args[idx] = tmp; | |
2493 | ||
d5cc9181 | 2494 | targetm.calls.function_arg_advance (args_so_far, mode, NULL_TREE, true); |
8dd5516b JJ |
2495 | } |
2496 | if (arg != void_list_node || idx != nargs) | |
2497 | return false; | |
2498 | return true; | |
2499 | } | |
2500 | ||
9e582b1d JR |
2501 | /* Return a bitmap of the fixed registers contained in IN. */ |
2502 | ||
2503 | static bitmap | |
2504 | copy_fixed_regs (const_bitmap in) | |
2505 | { | |
2506 | bitmap ret; | |
2507 | ||
2508 | ret = ALLOC_REG_SET (NULL); | |
2509 | bitmap_and (ret, in, fixed_reg_set_regset); | |
2510 | return ret; | |
2511 | } | |
8dd5516b | 2512 | |
6fb5fa3c DB |
2513 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2514 | if some part of it is not a candidate store and assigns to a | |
2515 | non-register target. */ | |
2516 | ||
2517 | static void | |
dd60a84c | 2518 | scan_insn (bb_info_t bb_info, rtx_insn *insn) |
6fb5fa3c DB |
2519 | { |
2520 | rtx body; | |
da6603c6 | 2521 | insn_info_type *insn_info = new insn_info_type; |
6fb5fa3c | 2522 | int mems_found = 0; |
da6603c6 | 2523 | memset (insn_info, 0, sizeof (struct insn_info_type)); |
6fb5fa3c | 2524 | |
456610d3 | 2525 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
2526 | fprintf (dump_file, "\n**scanning insn=%d\n", |
2527 | INSN_UID (insn)); | |
2528 | ||
2529 | insn_info->prev_insn = bb_info->last_insn; | |
2530 | insn_info->insn = insn; | |
2531 | bb_info->last_insn = insn_info; | |
b8698a0f | 2532 | |
b5b8b0ac AO |
2533 | if (DEBUG_INSN_P (insn)) |
2534 | { | |
2535 | insn_info->cannot_delete = true; | |
2536 | return; | |
2537 | } | |
6fb5fa3c | 2538 | |
6fb5fa3c DB |
2539 | /* Look at all of the uses in the insn. */ |
2540 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
2541 | ||
2542 | if (CALL_P (insn)) | |
2543 | { | |
8dd5516b JJ |
2544 | bool const_call; |
2545 | tree memset_call = NULL_TREE; | |
2546 | ||
6fb5fa3c | 2547 | insn_info->cannot_delete = true; |
50f0f366 | 2548 | |
6fb5fa3c | 2549 | /* Const functions cannot do anything bad i.e. read memory, |
50f0f366 | 2550 | however, they can read their parameters which may have |
8dd5516b JJ |
2551 | been pushed onto the stack. |
2552 | memset and bzero don't read memory either. */ | |
2553 | const_call = RTL_CONST_CALL_P (insn); | |
2554 | if (!const_call) | |
2555 | { | |
da4fdf2d SB |
2556 | rtx call = get_call_rtx_from (insn); |
2557 | if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF) | |
8dd5516b JJ |
2558 | { |
2559 | rtx symbol = XEXP (XEXP (call, 0), 0); | |
2560 | if (SYMBOL_REF_DECL (symbol) | |
2561 | && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL) | |
2562 | { | |
2563 | if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol)) | |
2564 | == BUILT_IN_NORMAL | |
2565 | && (DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol)) | |
2566 | == BUILT_IN_MEMSET)) | |
2567 | || SYMBOL_REF_DECL (symbol) == block_clear_fn) | |
2568 | memset_call = SYMBOL_REF_DECL (symbol); | |
2569 | } | |
2570 | } | |
2571 | } | |
2572 | if (const_call || memset_call) | |
6fb5fa3c DB |
2573 | { |
2574 | insn_info_t i_ptr = active_local_stores; | |
2575 | insn_info_t last = NULL; | |
2576 | ||
456610d3 | 2577 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2578 | fprintf (dump_file, "%s call %d\n", |
2579 | const_call ? "const" : "memset", INSN_UID (insn)); | |
6fb5fa3c | 2580 | |
64520bdc | 2581 | /* See the head comment of the frame_read field. */ |
57534689 JJ |
2582 | if (reload_completed |
2583 | /* Tail calls are storing their arguments using | |
2584 | arg pointer. If it is a frame pointer on the target, | |
2585 | even before reload we need to kill frame pointer based | |
2586 | stores. */ | |
2587 | || (SIBLING_CALL_P (insn) | |
2588 | && HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
64520bdc EB |
2589 | insn_info->frame_read = true; |
2590 | ||
2591 | /* Loop over the active stores and remove those which are | |
2592 | killed by the const function call. */ | |
6fb5fa3c DB |
2593 | while (i_ptr) |
2594 | { | |
64520bdc EB |
2595 | bool remove_store = false; |
2596 | ||
2597 | /* The stack pointer based stores are always killed. */ | |
50f0f366 | 2598 | if (i_ptr->stack_pointer_based) |
64520bdc EB |
2599 | remove_store = true; |
2600 | ||
2601 | /* If the frame is read, the frame related stores are killed. */ | |
2602 | else if (insn_info->frame_read) | |
2603 | { | |
2604 | store_info_t store_info = i_ptr->store_rec; | |
2605 | ||
2606 | /* Skip the clobbers. */ | |
2607 | while (!store_info->is_set) | |
2608 | store_info = store_info->next; | |
2609 | ||
2610 | if (store_info->group_id >= 0 | |
9771b263 | 2611 | && rtx_group_vec[store_info->group_id]->frame_related) |
64520bdc EB |
2612 | remove_store = true; |
2613 | } | |
2614 | ||
2615 | if (remove_store) | |
6fb5fa3c | 2616 | { |
456610d3 | 2617 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2618 | dump_insn_info ("removing from active", i_ptr); |
b8698a0f | 2619 | |
dabd47e7 | 2620 | active_local_stores_len--; |
6fb5fa3c DB |
2621 | if (last) |
2622 | last->next_local_store = i_ptr->next_local_store; | |
2623 | else | |
2624 | active_local_stores = i_ptr->next_local_store; | |
2625 | } | |
2626 | else | |
2627 | last = i_ptr; | |
64520bdc | 2628 | |
6fb5fa3c DB |
2629 | i_ptr = i_ptr->next_local_store; |
2630 | } | |
8dd5516b JJ |
2631 | |
2632 | if (memset_call) | |
2633 | { | |
2634 | rtx args[3]; | |
2635 | if (get_call_args (insn, memset_call, args, 3) | |
2636 | && CONST_INT_P (args[1]) | |
2637 | && CONST_INT_P (args[2]) | |
2638 | && INTVAL (args[2]) > 0) | |
2639 | { | |
2640 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
f5541398 | 2641 | set_mem_size (mem, INTVAL (args[2])); |
f7df4a84 | 2642 | body = gen_rtx_SET (mem, args[1]); |
8dd5516b | 2643 | mems_found += record_store (body, bb_info); |
456610d3 | 2644 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2645 | fprintf (dump_file, "handling memset as BLKmode store\n"); |
2646 | if (mems_found == 1) | |
2647 | { | |
dabd47e7 JJ |
2648 | if (active_local_stores_len++ |
2649 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2650 | { | |
2651 | active_local_stores_len = 1; | |
2652 | active_local_stores = NULL; | |
2653 | } | |
9e582b1d JR |
2654 | insn_info->fixed_regs_live |
2655 | = copy_fixed_regs (bb_info->regs_live); | |
8dd5516b JJ |
2656 | insn_info->next_local_store = active_local_stores; |
2657 | active_local_stores = insn_info; | |
2658 | } | |
2659 | } | |
2660 | } | |
6fb5fa3c | 2661 | } |
57534689 JJ |
2662 | else if (SIBLING_CALL_P (insn) && reload_completed) |
2663 | /* Arguments for a sibling call that are pushed to memory are passed | |
2664 | using the incoming argument pointer of the current function. After | |
2665 | reload that might be (and likely is) frame pointer based. */ | |
2666 | add_wild_read (bb_info); | |
50f0f366 | 2667 | else |
d26c7090 ER |
2668 | /* Every other call, including pure functions, may read any memory |
2669 | that is not relative to the frame. */ | |
2670 | add_non_frame_wild_read (bb_info); | |
50f0f366 | 2671 | |
6fb5fa3c DB |
2672 | return; |
2673 | } | |
2674 | ||
2675 | /* Assuming that there are sets in these insns, we cannot delete | |
2676 | them. */ | |
2677 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
0a64eeca | 2678 | || volatile_refs_p (PATTERN (insn)) |
2da02156 | 2679 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
6fb5fa3c DB |
2680 | || (RTX_FRAME_RELATED_P (insn)) |
2681 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
2682 | insn_info->cannot_delete = true; | |
b8698a0f | 2683 | |
6fb5fa3c DB |
2684 | body = PATTERN (insn); |
2685 | if (GET_CODE (body) == PARALLEL) | |
2686 | { | |
2687 | int i; | |
2688 | for (i = 0; i < XVECLEN (body, 0); i++) | |
2689 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
2690 | } | |
2691 | else | |
2692 | mems_found += record_store (body, bb_info); | |
2693 | ||
456610d3 | 2694 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2695 | fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
6fb5fa3c DB |
2696 | mems_found, insn_info->cannot_delete ? "true" : "false"); |
2697 | ||
8dd5516b JJ |
2698 | /* If we found some sets of mems, add it into the active_local_stores so |
2699 | that it can be locally deleted if found dead or used for | |
2700 | replace_read and redundant constant store elimination. Otherwise mark | |
2701 | it as cannot delete. This simplifies the processing later. */ | |
2702 | if (mems_found == 1) | |
6fb5fa3c | 2703 | { |
dabd47e7 JJ |
2704 | if (active_local_stores_len++ |
2705 | >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
2706 | { | |
2707 | active_local_stores_len = 1; | |
2708 | active_local_stores = NULL; | |
2709 | } | |
9e582b1d | 2710 | insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); |
6fb5fa3c DB |
2711 | insn_info->next_local_store = active_local_stores; |
2712 | active_local_stores = insn_info; | |
2713 | } | |
2714 | else | |
2715 | insn_info->cannot_delete = true; | |
2716 | } | |
2717 | ||
2718 | ||
2719 | /* Remove BASE from the set of active_local_stores. This is a | |
2720 | callback from cselib that is used to get rid of the stores in | |
2721 | active_local_stores. */ | |
2722 | ||
2723 | static void | |
2724 | remove_useless_values (cselib_val *base) | |
2725 | { | |
2726 | insn_info_t insn_info = active_local_stores; | |
2727 | insn_info_t last = NULL; | |
2728 | ||
2729 | while (insn_info) | |
2730 | { | |
2731 | store_info_t store_info = insn_info->store_rec; | |
60564289 | 2732 | bool del = false; |
6fb5fa3c DB |
2733 | |
2734 | /* If ANY of the store_infos match the cselib group that is | |
2735 | being deleted, then the insn can not be deleted. */ | |
2736 | while (store_info) | |
2737 | { | |
b8698a0f | 2738 | if ((store_info->group_id == -1) |
6fb5fa3c DB |
2739 | && (store_info->cse_base == base)) |
2740 | { | |
60564289 | 2741 | del = true; |
6fb5fa3c DB |
2742 | break; |
2743 | } | |
2744 | store_info = store_info->next; | |
2745 | } | |
2746 | ||
60564289 | 2747 | if (del) |
6fb5fa3c | 2748 | { |
dabd47e7 | 2749 | active_local_stores_len--; |
6fb5fa3c DB |
2750 | if (last) |
2751 | last->next_local_store = insn_info->next_local_store; | |
2752 | else | |
2753 | active_local_stores = insn_info->next_local_store; | |
2754 | free_store_info (insn_info); | |
2755 | } | |
2756 | else | |
2757 | last = insn_info; | |
b8698a0f | 2758 | |
6fb5fa3c DB |
2759 | insn_info = insn_info->next_local_store; |
2760 | } | |
2761 | } | |
2762 | ||
2763 | ||
2764 | /* Do all of step 1. */ | |
2765 | ||
2766 | static void | |
2767 | dse_step1 (void) | |
2768 | { | |
2769 | basic_block bb; | |
3f9b14ff | 2770 | bitmap regs_live = BITMAP_ALLOC (®_obstack); |
b8698a0f | 2771 | |
457eeaae | 2772 | cselib_init (0); |
6fb5fa3c DB |
2773 | all_blocks = BITMAP_ALLOC (NULL); |
2774 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
2775 | bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
2776 | ||
04a90bec | 2777 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
2778 | { |
2779 | insn_info_t ptr; | |
da6603c6 | 2780 | bb_info_t bb_info = new dse_bb_info_type; |
6fb5fa3c | 2781 | |
da6603c6 | 2782 | memset (bb_info, 0, sizeof (dse_bb_info_type)); |
6fb5fa3c | 2783 | bitmap_set_bit (all_blocks, bb->index); |
02b47899 KZ |
2784 | bb_info->regs_live = regs_live; |
2785 | ||
2786 | bitmap_copy (regs_live, DF_LR_IN (bb)); | |
2787 | df_simulate_initialize_forwards (bb, regs_live); | |
6fb5fa3c DB |
2788 | |
2789 | bb_table[bb->index] = bb_info; | |
2790 | cselib_discard_hook = remove_useless_values; | |
2791 | ||
2792 | if (bb->index >= NUM_FIXED_BLOCKS) | |
2793 | { | |
dd60a84c | 2794 | rtx_insn *insn; |
6fb5fa3c | 2795 | |
6fb5fa3c | 2796 | active_local_stores = NULL; |
dabd47e7 | 2797 | active_local_stores_len = 0; |
6fb5fa3c | 2798 | cselib_clear_table (); |
b8698a0f | 2799 | |
6fb5fa3c DB |
2800 | /* Scan the insns. */ |
2801 | FOR_BB_INSNS (bb, insn) | |
2802 | { | |
2803 | if (INSN_P (insn)) | |
2804 | scan_insn (bb_info, insn); | |
2805 | cselib_process_insn (insn); | |
02b47899 KZ |
2806 | if (INSN_P (insn)) |
2807 | df_simulate_one_insn_forwards (bb, insn, regs_live); | |
6fb5fa3c | 2808 | } |
b8698a0f | 2809 | |
6fb5fa3c DB |
2810 | /* This is something of a hack, because the global algorithm |
2811 | is supposed to take care of the case where stores go dead | |
2812 | at the end of the function. However, the global | |
2813 | algorithm must take a more conservative view of block | |
2814 | mode reads than the local alg does. So to get the case | |
2815 | where you have a store to the frame followed by a non | |
0d52bcc1 | 2816 | overlapping block more read, we look at the active local |
6fb5fa3c DB |
2817 | stores at the end of the function and delete all of the |
2818 | frame and spill based ones. */ | |
2819 | if (stores_off_frame_dead_at_return | |
2820 | && (EDGE_COUNT (bb->succs) == 0 | |
2821 | || (single_succ_p (bb) | |
fefa31b5 | 2822 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
e3b5732b | 2823 | && ! crtl->calls_eh_return))) |
6fb5fa3c DB |
2824 | { |
2825 | insn_info_t i_ptr = active_local_stores; | |
2826 | while (i_ptr) | |
2827 | { | |
2828 | store_info_t store_info = i_ptr->store_rec; | |
2829 | ||
2830 | /* Skip the clobbers. */ | |
2831 | while (!store_info->is_set) | |
2832 | store_info = store_info->next; | |
8dd5516b | 2833 | if (store_info->alias_set && !i_ptr->cannot_delete) |
6fb5fa3c | 2834 | delete_dead_store_insn (i_ptr); |
b8698a0f | 2835 | else |
6fb5fa3c DB |
2836 | if (store_info->group_id >= 0) |
2837 | { | |
b8698a0f | 2838 | group_info_t group |
9771b263 | 2839 | = rtx_group_vec[store_info->group_id]; |
8dd5516b | 2840 | if (group->frame_related && !i_ptr->cannot_delete) |
6fb5fa3c DB |
2841 | delete_dead_store_insn (i_ptr); |
2842 | } | |
2843 | ||
2844 | i_ptr = i_ptr->next_local_store; | |
2845 | } | |
2846 | } | |
2847 | ||
2848 | /* Get rid of the loads that were discovered in | |
2849 | replace_read. Cselib is finished with this block. */ | |
2850 | while (deferred_change_list) | |
2851 | { | |
2852 | deferred_change_t next = deferred_change_list->next; | |
2853 | ||
2854 | /* There is no reason to validate this change. That was | |
2855 | done earlier. */ | |
2856 | *deferred_change_list->loc = deferred_change_list->reg; | |
da6603c6 | 2857 | delete deferred_change_list; |
6fb5fa3c DB |
2858 | deferred_change_list = next; |
2859 | } | |
2860 | ||
2861 | /* Get rid of all of the cselib based store_infos in this | |
2862 | block and mark the containing insns as not being | |
2863 | deletable. */ | |
2864 | ptr = bb_info->last_insn; | |
2865 | while (ptr) | |
2866 | { | |
2867 | if (ptr->contains_cselib_groups) | |
8dd5516b JJ |
2868 | { |
2869 | store_info_t s_info = ptr->store_rec; | |
2870 | while (s_info && !s_info->is_set) | |
2871 | s_info = s_info->next; | |
2872 | if (s_info | |
2873 | && s_info->redundant_reason | |
2874 | && s_info->redundant_reason->insn | |
2875 | && !ptr->cannot_delete) | |
2876 | { | |
456610d3 | 2877 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
2878 | fprintf (dump_file, "Locally deleting insn %d " |
2879 | "because insn %d stores the " | |
2880 | "same value and couldn't be " | |
2881 | "eliminated\n", | |
2882 | INSN_UID (ptr->insn), | |
2883 | INSN_UID (s_info->redundant_reason->insn)); | |
2884 | delete_dead_store_insn (ptr); | |
2885 | } | |
8dd5516b JJ |
2886 | free_store_info (ptr); |
2887 | } | |
2888 | else | |
2889 | { | |
2890 | store_info_t s_info; | |
2891 | ||
2892 | /* Free at least positions_needed bitmaps. */ | |
2893 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
2894 | if (s_info->is_large) | |
2895 | { | |
dc491a25 | 2896 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
8dd5516b JJ |
2897 | s_info->is_large = false; |
2898 | } | |
2899 | } | |
6fb5fa3c DB |
2900 | ptr = ptr->prev_insn; |
2901 | } | |
2902 | ||
da6603c6 | 2903 | cse_store_info_pool.release (); |
6fb5fa3c | 2904 | } |
02b47899 | 2905 | bb_info->regs_live = NULL; |
6fb5fa3c DB |
2906 | } |
2907 | ||
02b47899 | 2908 | BITMAP_FREE (regs_live); |
6fb5fa3c | 2909 | cselib_finish (); |
c203e8a7 | 2910 | rtx_group_table->empty (); |
6fb5fa3c DB |
2911 | } |
2912 | ||
2913 | \f | |
2914 | /*---------------------------------------------------------------------------- | |
2915 | Second step. | |
2916 | ||
2917 | Assign each byte position in the stores that we are going to | |
2918 | analyze globally to a position in the bitmaps. Returns true if | |
6ed3da00 | 2919 | there are any bit positions assigned. |
6fb5fa3c DB |
2920 | ----------------------------------------------------------------------------*/ |
2921 | ||
2922 | static void | |
2923 | dse_step2_init (void) | |
2924 | { | |
2925 | unsigned int i; | |
2926 | group_info_t group; | |
2927 | ||
9771b263 | 2928 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2929 | { |
2930 | /* For all non stack related bases, we only consider a store to | |
2931 | be deletable if there are two or more stores for that | |
2932 | position. This is because it takes one store to make the | |
2933 | other store redundant. However, for the stores that are | |
2934 | stack related, we consider them if there is only one store | |
2935 | for the position. We do this because the stack related | |
2936 | stores can be deleted if their is no read between them and | |
2937 | the end of the function. | |
b8698a0f | 2938 | |
6fb5fa3c DB |
2939 | To make this work in the current framework, we take the stack |
2940 | related bases add all of the bits from store1 into store2. | |
2941 | This has the effect of making the eligible even if there is | |
2942 | only one store. */ | |
2943 | ||
2944 | if (stores_off_frame_dead_at_return && group->frame_related) | |
2945 | { | |
2946 | bitmap_ior_into (group->store2_n, group->store1_n); | |
2947 | bitmap_ior_into (group->store2_p, group->store1_p); | |
456610d3 | 2948 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 2949 | fprintf (dump_file, "group %d is frame related ", i); |
6fb5fa3c DB |
2950 | } |
2951 | ||
2952 | group->offset_map_size_n++; | |
3f9b14ff SB |
2953 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2954 | group->offset_map_size_n); | |
6fb5fa3c | 2955 | group->offset_map_size_p++; |
3f9b14ff SB |
2956 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2957 | group->offset_map_size_p); | |
6fb5fa3c | 2958 | group->process_globally = false; |
456610d3 | 2959 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c | 2960 | { |
b8698a0f | 2961 | fprintf (dump_file, "group %d(%d+%d): ", i, |
6fb5fa3c DB |
2962 | (int)bitmap_count_bits (group->store2_n), |
2963 | (int)bitmap_count_bits (group->store2_p)); | |
2964 | bitmap_print (dump_file, group->store2_n, "n ", " "); | |
2965 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
2966 | } | |
2967 | } | |
2968 | } | |
2969 | ||
2970 | ||
2971 | /* Init the offset tables for the normal case. */ | |
2972 | ||
2973 | static bool | |
2974 | dse_step2_nospill (void) | |
2975 | { | |
2976 | unsigned int i; | |
2977 | group_info_t group; | |
2978 | /* Position 0 is unused because 0 is used in the maps to mean | |
2979 | unused. */ | |
2980 | current_position = 1; | |
9771b263 | 2981 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
2982 | { |
2983 | bitmap_iterator bi; | |
2984 | unsigned int j; | |
2985 | ||
2986 | if (group == clear_alias_group) | |
2987 | continue; | |
2988 | ||
c3284718 RS |
2989 | memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
2990 | memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
6fb5fa3c DB |
2991 | bitmap_clear (group->group_kill); |
2992 | ||
2993 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
2994 | { | |
2995 | bitmap_set_bit (group->group_kill, current_position); | |
d26c7090 ER |
2996 | if (bitmap_bit_p (group->escaped_n, j)) |
2997 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
2998 | group->offset_map_n[j] = current_position++; |
2999 | group->process_globally = true; | |
3000 | } | |
3001 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
3002 | { | |
b8698a0f | 3003 | bitmap_set_bit (group->group_kill, current_position); |
d26c7090 ER |
3004 | if (bitmap_bit_p (group->escaped_p, j)) |
3005 | bitmap_set_bit (kill_on_calls, current_position); | |
6fb5fa3c DB |
3006 | group->offset_map_p[j] = current_position++; |
3007 | group->process_globally = true; | |
3008 | } | |
3009 | } | |
3010 | return current_position != 1; | |
3011 | } | |
3012 | ||
3013 | ||
6fb5fa3c DB |
3014 | \f |
3015 | /*---------------------------------------------------------------------------- | |
3016 | Third step. | |
b8698a0f | 3017 | |
6fb5fa3c DB |
3018 | Build the bit vectors for the transfer functions. |
3019 | ----------------------------------------------------------------------------*/ | |
3020 | ||
3021 | ||
6fb5fa3c DB |
3022 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
3023 | there, return 0. */ | |
3024 | ||
3025 | static int | |
3026 | get_bitmap_index (group_info_t group_info, HOST_WIDE_INT offset) | |
3027 | { | |
3028 | if (offset < 0) | |
3029 | { | |
3030 | HOST_WIDE_INT offset_p = -offset; | |
3031 | if (offset_p >= group_info->offset_map_size_n) | |
3032 | return 0; | |
3033 | return group_info->offset_map_n[offset_p]; | |
3034 | } | |
3035 | else | |
3036 | { | |
3037 | if (offset >= group_info->offset_map_size_p) | |
3038 | return 0; | |
3039 | return group_info->offset_map_p[offset]; | |
3040 | } | |
3041 | } | |
3042 | ||
3043 | ||
3044 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
3045 | may be NULL. */ | |
3046 | ||
b8698a0f | 3047 | static void |
6fb5fa3c DB |
3048 | scan_stores_nospill (store_info_t store_info, bitmap gen, bitmap kill) |
3049 | { | |
3050 | while (store_info) | |
3051 | { | |
3052 | HOST_WIDE_INT i; | |
b8698a0f | 3053 | group_info_t group_info |
9771b263 | 3054 | = rtx_group_vec[store_info->group_id]; |
6fb5fa3c DB |
3055 | if (group_info->process_globally) |
3056 | for (i = store_info->begin; i < store_info->end; i++) | |
3057 | { | |
3058 | int index = get_bitmap_index (group_info, i); | |
3059 | if (index != 0) | |
3060 | { | |
3061 | bitmap_set_bit (gen, index); | |
3062 | if (kill) | |
3063 | bitmap_clear_bit (kill, index); | |
3064 | } | |
3065 | } | |
3066 | store_info = store_info->next; | |
3067 | } | |
3068 | } | |
3069 | ||
3070 | ||
3071 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
3072 | may be NULL. */ | |
3073 | ||
b8698a0f | 3074 | static void |
6fb5fa3c DB |
3075 | scan_stores_spill (store_info_t store_info, bitmap gen, bitmap kill) |
3076 | { | |
3077 | while (store_info) | |
3078 | { | |
3079 | if (store_info->alias_set) | |
3080 | { | |
b8698a0f | 3081 | int index = get_bitmap_index (clear_alias_group, |
6fb5fa3c DB |
3082 | store_info->alias_set); |
3083 | if (index != 0) | |
3084 | { | |
3085 | bitmap_set_bit (gen, index); | |
3086 | if (kill) | |
3087 | bitmap_clear_bit (kill, index); | |
3088 | } | |
3089 | } | |
3090 | store_info = store_info->next; | |
3091 | } | |
3092 | } | |
3093 | ||
3094 | ||
3095 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3096 | may be NULL. */ | |
3097 | ||
3098 | static void | |
3099 | scan_reads_nospill (insn_info_t insn_info, bitmap gen, bitmap kill) | |
3100 | { | |
3101 | read_info_t read_info = insn_info->read_rec; | |
3102 | int i; | |
3103 | group_info_t group; | |
3104 | ||
64520bdc EB |
3105 | /* If this insn reads the frame, kill all the frame related stores. */ |
3106 | if (insn_info->frame_read) | |
3107 | { | |
9771b263 | 3108 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
64520bdc EB |
3109 | if (group->process_globally && group->frame_related) |
3110 | { | |
3111 | if (kill) | |
3112 | bitmap_ior_into (kill, group->group_kill); | |
b8698a0f | 3113 | bitmap_and_compl_into (gen, group->group_kill); |
64520bdc EB |
3114 | } |
3115 | } | |
d26c7090 ER |
3116 | if (insn_info->non_frame_wild_read) |
3117 | { | |
3118 | /* Kill all non-frame related stores. Kill all stores of variables that | |
3119 | escape. */ | |
3120 | if (kill) | |
3121 | bitmap_ior_into (kill, kill_on_calls); | |
3122 | bitmap_and_compl_into (gen, kill_on_calls); | |
9771b263 | 3123 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
d26c7090 ER |
3124 | if (group->process_globally && !group->frame_related) |
3125 | { | |
3126 | if (kill) | |
3127 | bitmap_ior_into (kill, group->group_kill); | |
3128 | bitmap_and_compl_into (gen, group->group_kill); | |
3129 | } | |
3130 | } | |
6fb5fa3c DB |
3131 | while (read_info) |
3132 | { | |
9771b263 | 3133 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3134 | { |
3135 | if (group->process_globally) | |
3136 | { | |
3137 | if (i == read_info->group_id) | |
3138 | { | |
3139 | if (read_info->begin > read_info->end) | |
3140 | { | |
3141 | /* Begin > end for block mode reads. */ | |
3142 | if (kill) | |
3143 | bitmap_ior_into (kill, group->group_kill); | |
3144 | bitmap_and_compl_into (gen, group->group_kill); | |
3145 | } | |
3146 | else | |
3147 | { | |
3148 | /* The groups are the same, just process the | |
3149 | offsets. */ | |
3150 | HOST_WIDE_INT j; | |
3151 | for (j = read_info->begin; j < read_info->end; j++) | |
3152 | { | |
3153 | int index = get_bitmap_index (group, j); | |
3154 | if (index != 0) | |
3155 | { | |
3156 | if (kill) | |
3157 | bitmap_set_bit (kill, index); | |
3158 | bitmap_clear_bit (gen, index); | |
3159 | } | |
3160 | } | |
3161 | } | |
3162 | } | |
3163 | else | |
3164 | { | |
3165 | /* The groups are different, if the alias sets | |
3166 | conflict, clear the entire group. We only need | |
3167 | to apply this test if the read_info is a cselib | |
3168 | read. Anything with a constant base cannot alias | |
3169 | something else with a different constant | |
3170 | base. */ | |
3171 | if ((read_info->group_id < 0) | |
b8698a0f | 3172 | && canon_true_dependence (group->base_mem, |
d32f725a | 3173 | GET_MODE (group->base_mem), |
6216f94e | 3174 | group->canon_base_addr, |
53d9622b | 3175 | read_info->mem, NULL_RTX)) |
6fb5fa3c DB |
3176 | { |
3177 | if (kill) | |
3178 | bitmap_ior_into (kill, group->group_kill); | |
3179 | bitmap_and_compl_into (gen, group->group_kill); | |
3180 | } | |
3181 | } | |
3182 | } | |
3183 | } | |
b8698a0f | 3184 | |
6fb5fa3c DB |
3185 | read_info = read_info->next; |
3186 | } | |
3187 | } | |
3188 | ||
3189 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
3190 | may be NULL. */ | |
3191 | ||
3192 | static void | |
3193 | scan_reads_spill (read_info_t read_info, bitmap gen, bitmap kill) | |
3194 | { | |
3195 | while (read_info) | |
3196 | { | |
3197 | if (read_info->alias_set) | |
3198 | { | |
b8698a0f | 3199 | int index = get_bitmap_index (clear_alias_group, |
6fb5fa3c DB |
3200 | read_info->alias_set); |
3201 | if (index != 0) | |
3202 | { | |
3203 | if (kill) | |
3204 | bitmap_set_bit (kill, index); | |
3205 | bitmap_clear_bit (gen, index); | |
3206 | } | |
3207 | } | |
b8698a0f | 3208 | |
6fb5fa3c DB |
3209 | read_info = read_info->next; |
3210 | } | |
3211 | } | |
3212 | ||
3213 | ||
3214 | /* Return the insn in BB_INFO before the first wild read or if there | |
3215 | are no wild reads in the block, return the last insn. */ | |
3216 | ||
3217 | static insn_info_t | |
3218 | find_insn_before_first_wild_read (bb_info_t bb_info) | |
3219 | { | |
3220 | insn_info_t insn_info = bb_info->last_insn; | |
3221 | insn_info_t last_wild_read = NULL; | |
3222 | ||
3223 | while (insn_info) | |
3224 | { | |
3225 | if (insn_info->wild_read) | |
3226 | { | |
3227 | last_wild_read = insn_info->prev_insn; | |
3228 | /* Block starts with wild read. */ | |
3229 | if (!last_wild_read) | |
3230 | return NULL; | |
3231 | } | |
3232 | ||
3233 | insn_info = insn_info->prev_insn; | |
3234 | } | |
3235 | ||
3236 | if (last_wild_read) | |
3237 | return last_wild_read; | |
3238 | else | |
3239 | return bb_info->last_insn; | |
3240 | } | |
3241 | ||
3242 | ||
3243 | /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
3244 | the block in order to build the gen and kill sets for the block. | |
3245 | We start at ptr which may be the last insn in the block or may be | |
3246 | the first insn with a wild read. In the latter case we are able to | |
3247 | skip the rest of the block because it just does not matter: | |
3248 | anything that happens is hidden by the wild read. */ | |
3249 | ||
3250 | static void | |
3251 | dse_step3_scan (bool for_spills, basic_block bb) | |
3252 | { | |
3253 | bb_info_t bb_info = bb_table[bb->index]; | |
3254 | insn_info_t insn_info; | |
3255 | ||
3256 | if (for_spills) | |
3257 | /* There are no wild reads in the spill case. */ | |
3258 | insn_info = bb_info->last_insn; | |
3259 | else | |
3260 | insn_info = find_insn_before_first_wild_read (bb_info); | |
b8698a0f | 3261 | |
6fb5fa3c DB |
3262 | /* In the spill case or in the no_spill case if there is no wild |
3263 | read in the block, we will need a kill set. */ | |
3264 | if (insn_info == bb_info->last_insn) | |
3265 | { | |
3266 | if (bb_info->kill) | |
3267 | bitmap_clear (bb_info->kill); | |
3268 | else | |
3f9b14ff | 3269 | bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c | 3270 | } |
b8698a0f | 3271 | else |
6fb5fa3c DB |
3272 | if (bb_info->kill) |
3273 | BITMAP_FREE (bb_info->kill); | |
3274 | ||
3275 | while (insn_info) | |
3276 | { | |
3277 | /* There may have been code deleted by the dce pass run before | |
3278 | this phase. */ | |
3279 | if (insn_info->insn && INSN_P (insn_info->insn)) | |
3280 | { | |
b8698a0f | 3281 | /* Process the read(s) last. */ |
6fb5fa3c DB |
3282 | if (for_spills) |
3283 | { | |
3284 | scan_stores_spill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3285 | scan_reads_spill (insn_info->read_rec, bb_info->gen, bb_info->kill); | |
3286 | } | |
3287 | else | |
3288 | { | |
3289 | scan_stores_nospill (insn_info->store_rec, bb_info->gen, bb_info->kill); | |
3290 | scan_reads_nospill (insn_info, bb_info->gen, bb_info->kill); | |
3291 | } | |
b8698a0f | 3292 | } |
6fb5fa3c DB |
3293 | |
3294 | insn_info = insn_info->prev_insn; | |
3295 | } | |
3296 | } | |
3297 | ||
3298 | ||
3299 | /* Set the gen set of the exit block, and also any block with no | |
3300 | successors that does not have a wild read. */ | |
3301 | ||
3302 | static void | |
3303 | dse_step3_exit_block_scan (bb_info_t bb_info) | |
3304 | { | |
3305 | /* The gen set is all 0's for the exit block except for the | |
3306 | frame_pointer_group. */ | |
b8698a0f | 3307 | |
6fb5fa3c DB |
3308 | if (stores_off_frame_dead_at_return) |
3309 | { | |
3310 | unsigned int i; | |
3311 | group_info_t group; | |
b8698a0f | 3312 | |
9771b263 | 3313 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
6fb5fa3c DB |
3314 | { |
3315 | if (group->process_globally && group->frame_related) | |
3316 | bitmap_ior_into (bb_info->gen, group->group_kill); | |
3317 | } | |
3318 | } | |
3319 | } | |
3320 | ||
3321 | ||
3322 | /* Find all of the blocks that are not backwards reachable from the | |
3323 | exit block or any block with no successors (BB). These are the | |
3324 | infinite loops or infinite self loops. These blocks will still | |
3325 | have their bits set in UNREACHABLE_BLOCKS. */ | |
3326 | ||
3327 | static void | |
3328 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
3329 | { | |
3330 | edge e; | |
3331 | edge_iterator ei; | |
3332 | ||
d7c028c0 | 3333 | if (bitmap_bit_p (unreachable_blocks, bb->index)) |
6fb5fa3c | 3334 | { |
d7c028c0 | 3335 | bitmap_clear_bit (unreachable_blocks, bb->index); |
6fb5fa3c | 3336 | FOR_EACH_EDGE (e, ei, bb->preds) |
b8698a0f | 3337 | { |
6fb5fa3c | 3338 | mark_reachable_blocks (unreachable_blocks, e->src); |
b8698a0f | 3339 | } |
6fb5fa3c DB |
3340 | } |
3341 | } | |
3342 | ||
3343 | /* Build the transfer functions for the function. */ | |
3344 | ||
3345 | static void | |
3346 | dse_step3 (bool for_spills) | |
3347 | { | |
3348 | basic_block bb; | |
8b1c6fd7 | 3349 | sbitmap unreachable_blocks = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
6fb5fa3c DB |
3350 | sbitmap_iterator sbi; |
3351 | bitmap all_ones = NULL; | |
3352 | unsigned int i; | |
b8698a0f | 3353 | |
f61e445a | 3354 | bitmap_ones (unreachable_blocks); |
6fb5fa3c | 3355 | |
04a90bec | 3356 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3357 | { |
3358 | bb_info_t bb_info = bb_table[bb->index]; | |
3359 | if (bb_info->gen) | |
3360 | bitmap_clear (bb_info->gen); | |
3361 | else | |
3f9b14ff | 3362 | bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3363 | |
3364 | if (bb->index == ENTRY_BLOCK) | |
3365 | ; | |
3366 | else if (bb->index == EXIT_BLOCK) | |
3367 | dse_step3_exit_block_scan (bb_info); | |
3368 | else | |
3369 | dse_step3_scan (for_spills, bb); | |
3370 | if (EDGE_COUNT (bb->succs) == 0) | |
3371 | mark_reachable_blocks (unreachable_blocks, bb); | |
3372 | ||
3373 | /* If this is the second time dataflow is run, delete the old | |
3374 | sets. */ | |
3375 | if (bb_info->in) | |
3376 | BITMAP_FREE (bb_info->in); | |
3377 | if (bb_info->out) | |
3378 | BITMAP_FREE (bb_info->out); | |
3379 | } | |
3380 | ||
3381 | /* For any block in an infinite loop, we must initialize the out set | |
3382 | to all ones. This could be expensive, but almost never occurs in | |
3383 | practice. However, it is common in regression tests. */ | |
d4ac4ce2 | 3384 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
6fb5fa3c DB |
3385 | { |
3386 | if (bitmap_bit_p (all_blocks, i)) | |
3387 | { | |
3388 | bb_info_t bb_info = bb_table[i]; | |
3389 | if (!all_ones) | |
3390 | { | |
3391 | unsigned int j; | |
3392 | group_info_t group; | |
3393 | ||
3f9b14ff | 3394 | all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); |
9771b263 | 3395 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
6fb5fa3c DB |
3396 | bitmap_ior_into (all_ones, group->group_kill); |
3397 | } | |
3398 | if (!bb_info->out) | |
3399 | { | |
3f9b14ff | 3400 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3401 | bitmap_copy (bb_info->out, all_ones); |
3402 | } | |
3403 | } | |
3404 | } | |
3405 | ||
3406 | if (all_ones) | |
3407 | BITMAP_FREE (all_ones); | |
3408 | sbitmap_free (unreachable_blocks); | |
3409 | } | |
3410 | ||
3411 | ||
3412 | \f | |
3413 | /*---------------------------------------------------------------------------- | |
3414 | Fourth step. | |
3415 | ||
3416 | Solve the bitvector equations. | |
3417 | ----------------------------------------------------------------------------*/ | |
3418 | ||
3419 | ||
3420 | /* Confluence function for blocks with no successors. Create an out | |
3421 | set from the gen set of the exit block. This block logically has | |
3422 | the exit block as a successor. */ | |
3423 | ||
3424 | ||
3425 | ||
3426 | static void | |
3427 | dse_confluence_0 (basic_block bb) | |
3428 | { | |
3429 | bb_info_t bb_info = bb_table[bb->index]; | |
3430 | ||
3431 | if (bb->index == EXIT_BLOCK) | |
3432 | return; | |
3433 | ||
3434 | if (!bb_info->out) | |
3435 | { | |
3f9b14ff | 3436 | bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3437 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3438 | } | |
3439 | } | |
3440 | ||
3441 | /* Propagate the information from the in set of the dest of E to the | |
3442 | out set of the src of E. If the various in or out sets are not | |
3443 | there, that means they are all ones. */ | |
3444 | ||
1a0f3fa1 | 3445 | static bool |
6fb5fa3c DB |
3446 | dse_confluence_n (edge e) |
3447 | { | |
3448 | bb_info_t src_info = bb_table[e->src->index]; | |
3449 | bb_info_t dest_info = bb_table[e->dest->index]; | |
3450 | ||
3451 | if (dest_info->in) | |
3452 | { | |
3453 | if (src_info->out) | |
3454 | bitmap_and_into (src_info->out, dest_info->in); | |
3455 | else | |
3456 | { | |
3f9b14ff | 3457 | src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3458 | bitmap_copy (src_info->out, dest_info->in); |
3459 | } | |
3460 | } | |
1a0f3fa1 | 3461 | return true; |
6fb5fa3c DB |
3462 | } |
3463 | ||
3464 | ||
3465 | /* Propagate the info from the out to the in set of BB_INDEX's basic | |
b8698a0f | 3466 | block. There are three cases: |
6fb5fa3c DB |
3467 | |
3468 | 1) The block has no kill set. In this case the kill set is all | |
3469 | ones. It does not matter what the out set of the block is, none of | |
3470 | the info can reach the top. The only thing that reaches the top is | |
3471 | the gen set and we just copy the set. | |
3472 | ||
3473 | 2) There is a kill set but no out set and bb has successors. In | |
3474 | this case we just return. Eventually an out set will be created and | |
3475 | it is better to wait than to create a set of ones. | |
3476 | ||
3477 | 3) There is both a kill and out set. We apply the obvious transfer | |
3478 | function. | |
3479 | */ | |
3480 | ||
3481 | static bool | |
3482 | dse_transfer_function (int bb_index) | |
3483 | { | |
3484 | bb_info_t bb_info = bb_table[bb_index]; | |
3485 | ||
3486 | if (bb_info->kill) | |
3487 | { | |
3488 | if (bb_info->out) | |
3489 | { | |
3490 | /* Case 3 above. */ | |
3491 | if (bb_info->in) | |
b8698a0f | 3492 | return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3493 | bb_info->out, bb_info->kill); |
3494 | else | |
3495 | { | |
3f9b14ff | 3496 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
b8698a0f | 3497 | bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
6fb5fa3c DB |
3498 | bb_info->out, bb_info->kill); |
3499 | return true; | |
3500 | } | |
3501 | } | |
3502 | else | |
3503 | /* Case 2 above. */ | |
3504 | return false; | |
3505 | } | |
3506 | else | |
3507 | { | |
3508 | /* Case 1 above. If there is already an in set, nothing | |
3509 | happens. */ | |
3510 | if (bb_info->in) | |
3511 | return false; | |
3512 | else | |
3513 | { | |
3f9b14ff | 3514 | bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
6fb5fa3c DB |
3515 | bitmap_copy (bb_info->in, bb_info->gen); |
3516 | return true; | |
3517 | } | |
3518 | } | |
3519 | } | |
3520 | ||
3521 | /* Solve the dataflow equations. */ | |
3522 | ||
3523 | static void | |
3524 | dse_step4 (void) | |
3525 | { | |
b8698a0f L |
3526 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3527 | dse_confluence_n, dse_transfer_function, | |
3528 | all_blocks, df_get_postorder (DF_BACKWARD), | |
6fb5fa3c | 3529 | df_get_n_blocks (DF_BACKWARD)); |
456610d3 | 3530 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3531 | { |
3532 | basic_block bb; | |
3533 | ||
3534 | fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
04a90bec | 3535 | FOR_ALL_BB_FN (bb, cfun) |
6fb5fa3c DB |
3536 | { |
3537 | bb_info_t bb_info = bb_table[bb->index]; | |
3538 | ||
3539 | df_print_bb_index (bb, dump_file); | |
3540 | if (bb_info->in) | |
3541 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
3542 | else | |
3543 | fprintf (dump_file, " in: *MISSING*\n"); | |
3544 | if (bb_info->gen) | |
3545 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
3546 | else | |
3547 | fprintf (dump_file, " gen: *MISSING*\n"); | |
3548 | if (bb_info->kill) | |
3549 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
3550 | else | |
3551 | fprintf (dump_file, " kill: *MISSING*\n"); | |
3552 | if (bb_info->out) | |
3553 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
3554 | else | |
3555 | fprintf (dump_file, " out: *MISSING*\n\n"); | |
3556 | } | |
3557 | } | |
3558 | } | |
3559 | ||
3560 | ||
3561 | \f | |
3562 | /*---------------------------------------------------------------------------- | |
3563 | Fifth step. | |
3564 | ||
0d52bcc1 | 3565 | Delete the stores that can only be deleted using the global information. |
6fb5fa3c DB |
3566 | ----------------------------------------------------------------------------*/ |
3567 | ||
3568 | ||
3569 | static void | |
3570 | dse_step5_nospill (void) | |
3571 | { | |
3572 | basic_block bb; | |
11cd3bed | 3573 | FOR_EACH_BB_FN (bb, cfun) |
6fb5fa3c DB |
3574 | { |
3575 | bb_info_t bb_info = bb_table[bb->index]; | |
3576 | insn_info_t insn_info = bb_info->last_insn; | |
3577 | bitmap v = bb_info->out; | |
3578 | ||
3579 | while (insn_info) | |
3580 | { | |
3581 | bool deleted = false; | |
3582 | if (dump_file && insn_info->insn) | |
3583 | { | |
3584 | fprintf (dump_file, "starting to process insn %d\n", | |
3585 | INSN_UID (insn_info->insn)); | |
3586 | bitmap_print (dump_file, v, " v: ", "\n"); | |
3587 | } | |
3588 | ||
3589 | /* There may have been code deleted by the dce pass run before | |
3590 | this phase. */ | |
b8698a0f | 3591 | if (insn_info->insn |
6fb5fa3c DB |
3592 | && INSN_P (insn_info->insn) |
3593 | && (!insn_info->cannot_delete) | |
3594 | && (!bitmap_empty_p (v))) | |
3595 | { | |
3596 | store_info_t store_info = insn_info->store_rec; | |
3597 | ||
3598 | /* Try to delete the current insn. */ | |
3599 | deleted = true; | |
b8698a0f | 3600 | |
6fb5fa3c DB |
3601 | /* Skip the clobbers. */ |
3602 | while (!store_info->is_set) | |
3603 | store_info = store_info->next; | |
3604 | ||
3605 | if (store_info->alias_set) | |
3606 | deleted = false; | |
3607 | else | |
3608 | { | |
3609 | HOST_WIDE_INT i; | |
b8698a0f | 3610 | group_info_t group_info |
9771b263 | 3611 | = rtx_group_vec[store_info->group_id]; |
b8698a0f | 3612 | |
6fb5fa3c DB |
3613 | for (i = store_info->begin; i < store_info->end; i++) |
3614 | { | |
3615 | int index = get_bitmap_index (group_info, i); | |
b8698a0f | 3616 | |
456610d3 | 3617 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 3618 | fprintf (dump_file, "i = %d, index = %d\n", (int)i, index); |
6fb5fa3c DB |
3619 | if (index == 0 || !bitmap_bit_p (v, index)) |
3620 | { | |
456610d3 | 3621 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 3622 | fprintf (dump_file, "failing at i = %d\n", (int)i); |
6fb5fa3c DB |
3623 | deleted = false; |
3624 | break; | |
3625 | } | |
3626 | } | |
3627 | } | |
3628 | if (deleted) | |
3629 | { | |
9e582b1d JR |
3630 | if (dbg_cnt (dse) |
3631 | && check_for_inc_dec_1 (insn_info)) | |
6fb5fa3c | 3632 | { |
6fb5fa3c DB |
3633 | delete_insn (insn_info->insn); |
3634 | insn_info->insn = NULL; | |
3635 | globally_deleted++; | |
3636 | } | |
3637 | } | |
3638 | } | |
3639 | /* We do want to process the local info if the insn was | |
6ed3da00 | 3640 | deleted. For instance, if the insn did a wild read, we |
6fb5fa3c | 3641 | no longer need to trash the info. */ |
b8698a0f | 3642 | if (insn_info->insn |
6fb5fa3c DB |
3643 | && INSN_P (insn_info->insn) |
3644 | && (!deleted)) | |
3645 | { | |
3646 | scan_stores_nospill (insn_info->store_rec, v, NULL); | |
3647 | if (insn_info->wild_read) | |
3648 | { | |
456610d3 | 3649 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3650 | fprintf (dump_file, "wild read\n"); |
3651 | bitmap_clear (v); | |
3652 | } | |
d26c7090 ER |
3653 | else if (insn_info->read_rec |
3654 | || insn_info->non_frame_wild_read) | |
6fb5fa3c | 3655 | { |
d26c7090 | 3656 | if (dump_file && !insn_info->non_frame_wild_read) |
6fb5fa3c | 3657 | fprintf (dump_file, "regular read\n"); |
456610d3 | 3658 | else if (dump_file && (dump_flags & TDF_DETAILS)) |
d26c7090 | 3659 | fprintf (dump_file, "non-frame wild read\n"); |
6fb5fa3c DB |
3660 | scan_reads_nospill (insn_info, v, NULL); |
3661 | } | |
3662 | } | |
b8698a0f | 3663 | |
6fb5fa3c DB |
3664 | insn_info = insn_info->prev_insn; |
3665 | } | |
3666 | } | |
3667 | } | |
3668 | ||
3669 | ||
6fb5fa3c DB |
3670 | \f |
3671 | /*---------------------------------------------------------------------------- | |
3672 | Sixth step. | |
3673 | ||
8dd5516b JJ |
3674 | Delete stores made redundant by earlier stores (which store the same |
3675 | value) that couldn't be eliminated. | |
3676 | ----------------------------------------------------------------------------*/ | |
3677 | ||
3678 | static void | |
3679 | dse_step6 (void) | |
3680 | { | |
3681 | basic_block bb; | |
3682 | ||
04a90bec | 3683 | FOR_ALL_BB_FN (bb, cfun) |
8dd5516b JJ |
3684 | { |
3685 | bb_info_t bb_info = bb_table[bb->index]; | |
3686 | insn_info_t insn_info = bb_info->last_insn; | |
3687 | ||
3688 | while (insn_info) | |
3689 | { | |
3690 | /* There may have been code deleted by the dce pass run before | |
3691 | this phase. */ | |
3692 | if (insn_info->insn | |
3693 | && INSN_P (insn_info->insn) | |
3694 | && !insn_info->cannot_delete) | |
3695 | { | |
3696 | store_info_t s_info = insn_info->store_rec; | |
3697 | ||
3698 | while (s_info && !s_info->is_set) | |
3699 | s_info = s_info->next; | |
3700 | if (s_info | |
3701 | && s_info->redundant_reason | |
3702 | && s_info->redundant_reason->insn | |
3703 | && INSN_P (s_info->redundant_reason->insn)) | |
3704 | { | |
eb92d49a | 3705 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
456610d3 | 3706 | if (dump_file && (dump_flags & TDF_DETAILS)) |
8dd5516b JJ |
3707 | fprintf (dump_file, "Locally deleting insn %d " |
3708 | "because insn %d stores the " | |
3709 | "same value and couldn't be " | |
3710 | "eliminated\n", | |
3711 | INSN_UID (insn_info->insn), | |
3712 | INSN_UID (rinsn)); | |
3713 | delete_dead_store_insn (insn_info); | |
3714 | } | |
3715 | } | |
3716 | insn_info = insn_info->prev_insn; | |
3717 | } | |
3718 | } | |
3719 | } | |
3720 | \f | |
3721 | /*---------------------------------------------------------------------------- | |
3722 | Seventh step. | |
3723 | ||
b8698a0f | 3724 | Destroy everything left standing. |
6fb5fa3c DB |
3725 | ----------------------------------------------------------------------------*/ |
3726 | ||
b8698a0f | 3727 | static void |
3f9b14ff | 3728 | dse_step7 (void) |
6fb5fa3c | 3729 | { |
3f9b14ff SB |
3730 | bitmap_obstack_release (&dse_bitmap_obstack); |
3731 | obstack_free (&dse_obstack, NULL); | |
370f38e8 | 3732 | |
6fb5fa3c DB |
3733 | end_alias_analysis (); |
3734 | free (bb_table); | |
c203e8a7 TS |
3735 | delete rtx_group_table; |
3736 | rtx_group_table = NULL; | |
9771b263 | 3737 | rtx_group_vec.release (); |
6fb5fa3c DB |
3738 | BITMAP_FREE (all_blocks); |
3739 | BITMAP_FREE (scratch); | |
3740 | ||
da6603c6 ML |
3741 | rtx_store_info_pool.release (); |
3742 | read_info_type::pool.release (); | |
3743 | insn_info_type::pool.release (); | |
3744 | dse_bb_info_type::pool.release (); | |
3745 | group_info::pool.release (); | |
3746 | deferred_change::pool.release (); | |
6fb5fa3c DB |
3747 | } |
3748 | ||
3749 | ||
6fb5fa3c DB |
3750 | /* ------------------------------------------------------------------------- |
3751 | DSE | |
3752 | ------------------------------------------------------------------------- */ | |
3753 | ||
3754 | /* Callback for running pass_rtl_dse. */ | |
3755 | ||
3756 | static unsigned int | |
3757 | rest_of_handle_dse (void) | |
3758 | { | |
6fb5fa3c DB |
3759 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3760 | ||
02b47899 KZ |
3761 | /* Need the notes since we must track live hardregs in the forwards |
3762 | direction. */ | |
3763 | df_note_add_problem (); | |
3764 | df_analyze (); | |
3765 | ||
6fb5fa3c DB |
3766 | dse_step0 (); |
3767 | dse_step1 (); | |
3768 | dse_step2_init (); | |
3769 | if (dse_step2_nospill ()) | |
3770 | { | |
3771 | df_set_flags (DF_LR_RUN_DCE); | |
3772 | df_analyze (); | |
456610d3 | 3773 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6fb5fa3c DB |
3774 | fprintf (dump_file, "doing global processing\n"); |
3775 | dse_step3 (false); | |
3776 | dse_step4 (); | |
3777 | dse_step5_nospill (); | |
3778 | } | |
3779 | ||
8dd5516b | 3780 | dse_step6 (); |
3f9b14ff | 3781 | dse_step7 (); |
6fb5fa3c DB |
3782 | |
3783 | if (dump_file) | |
3784 | fprintf (dump_file, "dse: local deletions = %d, global deletions = %d, spill deletions = %d\n", | |
3785 | locally_deleted, globally_deleted, spill_deleted); | |
6c5ed3f1 JJ |
3786 | |
3787 | /* DSE can eliminate potentially-trapping MEMs. | |
3788 | Remove any EH edges associated with them. */ | |
3789 | if ((locally_deleted || globally_deleted) | |
3790 | && cfun->can_throw_non_call_exceptions | |
3791 | && purge_all_dead_edges ()) | |
3792 | cleanup_cfg (0); | |
3793 | ||
6fb5fa3c DB |
3794 | return 0; |
3795 | } | |
3796 | ||
27a4cd48 DM |
3797 | namespace { |
3798 | ||
3799 | const pass_data pass_data_rtl_dse1 = | |
3800 | { | |
3801 | RTL_PASS, /* type */ | |
3802 | "dse1", /* name */ | |
3803 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3804 | TV_DSE1, /* tv_id */ |
3805 | 0, /* properties_required */ | |
3806 | 0, /* properties_provided */ | |
3807 | 0, /* properties_destroyed */ | |
3808 | 0, /* todo_flags_start */ | |
3bea341f | 3809 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c DB |
3810 | }; |
3811 | ||
27a4cd48 DM |
3812 | class pass_rtl_dse1 : public rtl_opt_pass |
3813 | { | |
3814 | public: | |
c3284718 RS |
3815 | pass_rtl_dse1 (gcc::context *ctxt) |
3816 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
27a4cd48 DM |
3817 | {} |
3818 | ||
3819 | /* opt_pass methods: */ | |
1a3d085c TS |
3820 | virtual bool gate (function *) |
3821 | { | |
3822 | return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
3823 | } | |
3824 | ||
be55bfe6 | 3825 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3826 | |
3827 | }; // class pass_rtl_dse1 | |
3828 | ||
3829 | } // anon namespace | |
3830 | ||
3831 | rtl_opt_pass * | |
3832 | make_pass_rtl_dse1 (gcc::context *ctxt) | |
3833 | { | |
3834 | return new pass_rtl_dse1 (ctxt); | |
3835 | } | |
3836 | ||
3837 | namespace { | |
3838 | ||
3839 | const pass_data pass_data_rtl_dse2 = | |
3840 | { | |
3841 | RTL_PASS, /* type */ | |
3842 | "dse2", /* name */ | |
3843 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3844 | TV_DSE2, /* tv_id */ |
3845 | 0, /* properties_required */ | |
3846 | 0, /* properties_provided */ | |
3847 | 0, /* properties_destroyed */ | |
3848 | 0, /* todo_flags_start */ | |
3bea341f | 3849 | TODO_df_finish, /* todo_flags_finish */ |
6fb5fa3c | 3850 | }; |
27a4cd48 DM |
3851 | |
3852 | class pass_rtl_dse2 : public rtl_opt_pass | |
3853 | { | |
3854 | public: | |
c3284718 RS |
3855 | pass_rtl_dse2 (gcc::context *ctxt) |
3856 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
27a4cd48 DM |
3857 | {} |
3858 | ||
3859 | /* opt_pass methods: */ | |
1a3d085c TS |
3860 | virtual bool gate (function *) |
3861 | { | |
3862 | return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
3863 | } | |
3864 | ||
be55bfe6 | 3865 | virtual unsigned int execute (function *) { return rest_of_handle_dse (); } |
27a4cd48 DM |
3866 | |
3867 | }; // class pass_rtl_dse2 | |
3868 | ||
3869 | } // anon namespace | |
3870 | ||
3871 | rtl_opt_pass * | |
3872 | make_pass_rtl_dse2 (gcc::context *ctxt) | |
3873 | { | |
3874 | return new pass_rtl_dse2 (ctxt); | |
3875 | } |