Mercurial > hg > CbC > CbC_gcc
annotate gcc/dse.c @ 111:04ced10e8804
gcc 7
| author | kono |
|---|---|
| date | Fri, 27 Oct 2017 22:46:09 +0900 |
| parents | f6334be47118 |
| children | 84e7813d76e9 |
| rev | line source |
|---|---|
| 0 | 1 /* RTL dead store elimination. |
| 111 | 2 Copyright (C) 2005-2017 Free Software Foundation, Inc. |
| 0 | 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 | |
| 11 Software Foundation; either version 3, or (at your option) any later | |
| 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 | |
| 20 along with GCC; see the file COPYING3. If not see | |
| 21 <http://www.gnu.org/licenses/>. */ | |
| 22 | |
| 23 #undef BASELINE | |
| 24 | |
| 25 #include "config.h" | |
| 26 #include "system.h" | |
| 27 #include "coretypes.h" | |
| 111 | 28 #include "backend.h" |
| 29 #include "target.h" | |
| 0 | 30 #include "rtl.h" |
| 31 #include "tree.h" | |
| 111 | 32 #include "gimple.h" |
| 33 #include "predict.h" | |
| 34 #include "df.h" | |
| 35 #include "memmodel.h" | |
| 0 | 36 #include "tm_p.h" |
| 111 | 37 #include "gimple-ssa.h" |
| 38 #include "expmed.h" | |
| 39 #include "optabs.h" | |
| 40 #include "emit-rtl.h" | |
| 41 #include "recog.h" | |
| 42 #include "alias.h" | |
| 43 #include "stor-layout.h" | |
| 44 #include "cfgrtl.h" | |
| 0 | 45 #include "cselib.h" |
| 46 #include "tree-pass.h" | |
| 111 | 47 #include "explow.h" |
| 0 | 48 #include "expr.h" |
| 49 #include "dbgcnt.h" | |
| 111 | 50 #include "params.h" |
| 51 #include "rtl-iter.h" | |
| 52 #include "cfgcleanup.h" | |
| 0 | 53 |
| 54 /* This file contains three techniques for performing Dead Store | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
55 Elimination (dse). |
| 0 | 56 |
| 57 * The first technique performs dse locally on any base address. It | |
| 58 is based on the cselib which is a local value numbering technique. | |
| 59 This technique is local to a basic block but deals with a fairly | |
| 60 general addresses. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
61 |
| 0 | 62 * The second technique performs dse globally but is restricted to |
| 63 base addresses that are either constant or are relative to the | |
| 64 frame_pointer. | |
| 65 | |
| 66 * The third technique, (which is only done after register allocation) | |
| 111 | 67 processes the spill slots. This differs from the second |
| 0 | 68 technique because it takes advantage of the fact that spilling is |
| 69 completely free from the effects of aliasing. | |
| 70 | |
| 71 Logically, dse is a backwards dataflow problem. A store can be | |
| 72 deleted if it if cannot be reached in the backward direction by any | |
| 73 use of the value being stored. However, the local technique uses a | |
| 74 forwards scan of the basic block because cselib requires that the | |
| 75 block be processed in that order. | |
| 76 | |
| 77 The pass is logically broken into 7 steps: | |
| 78 | |
| 79 0) Initialization. | |
| 80 | |
| 81 1) The local algorithm, as well as scanning the insns for the two | |
| 82 global algorithms. | |
| 83 | |
| 84 2) Analysis to see if the global algs are necessary. In the case | |
| 85 of stores base on a constant address, there must be at least two | |
| 86 stores to that address, to make it possible to delete some of the | |
| 87 stores. In the case of stores off of the frame or spill related | |
| 88 stores, only one store to an address is necessary because those | |
| 89 stores die at the end of the function. | |
| 90 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
91 3) Set up the global dataflow equations based on processing the |
| 0 | 92 info parsed in the first step. |
| 93 | |
| 94 4) Solve the dataflow equations. | |
| 95 | |
| 96 5) Delete the insns that the global analysis has indicated are | |
| 97 unnecessary. | |
| 98 | |
| 111 | 99 6) Delete insns that store the same value as preceding store |
| 0 | 100 where the earlier store couldn't be eliminated. |
| 101 | |
| 102 7) Cleanup. | |
| 103 | |
| 104 This step uses cselib and canon_rtx to build the largest expression | |
| 105 possible for each address. This pass is a forwards pass through | |
| 106 each basic block. From the point of view of the global technique, | |
| 107 the first pass could examine a block in either direction. The | |
| 108 forwards ordering is to accommodate cselib. | |
| 109 | |
| 111 | 110 We make a simplifying assumption: addresses fall into four broad |
| 0 | 111 categories: |
| 112 | |
| 113 1) base has rtx_varies_p == false, offset is constant. | |
| 114 2) base has rtx_varies_p == false, offset variable. | |
| 115 3) base has rtx_varies_p == true, offset constant. | |
| 116 4) base has rtx_varies_p == true, offset variable. | |
| 117 | |
| 118 The local passes are able to process all 4 kinds of addresses. The | |
| 111 | 119 global pass only handles 1). |
| 0 | 120 |
| 121 The global problem is formulated as follows: | |
| 122 | |
| 123 A store, S1, to address A, where A is not relative to the stack | |
| 124 frame, can be eliminated if all paths from S1 to the end of the | |
| 111 | 125 function contain another store to A before a read to A. |
| 0 | 126 |
| 127 If the address A is relative to the stack frame, a store S2 to A | |
| 111 | 128 can be eliminated if there are no paths from S2 that reach the |
| 0 | 129 end of the function that read A before another store to A. In |
| 111 | 130 this case S2 can be deleted if there are paths from S2 to the |
| 0 | 131 end of the function that have no reads or writes to A. This |
| 132 second case allows stores to the stack frame to be deleted that | |
| 133 would otherwise die when the function returns. This cannot be | |
| 134 done if stores_off_frame_dead_at_return is not true. See the doc | |
| 135 for that variable for when this variable is false. | |
| 136 | |
| 137 The global problem is formulated as a backwards set union | |
| 138 dataflow problem where the stores are the gens and reads are the | |
| 139 kills. Set union problems are rare and require some special | |
| 140 handling given our representation of bitmaps. A straightforward | |
| 111 | 141 implementation requires a lot of bitmaps filled with 1s. |
| 0 | 142 These are expensive and cumbersome in our bitmap formulation so |
| 143 care has been taken to avoid large vectors filled with 1s. See | |
| 144 the comments in bb_info and in the dataflow confluence functions | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
145 for details. |
| 0 | 146 |
| 147 There are two places for further enhancements to this algorithm: | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
148 |
| 0 | 149 1) The original dse which was embedded in a pass called flow also |
| 150 did local address forwarding. For example in | |
| 151 | |
| 152 A <- r100 | |
| 153 ... <- A | |
| 154 | |
| 155 flow would replace the right hand side of the second insn with a | |
| 156 reference to r100. Most of the information is available to add this | |
| 157 to this pass. It has not done it because it is a lot of work in | |
| 158 the case that either r100 is assigned to between the first and | |
| 159 second insn and/or the second insn is a load of part of the value | |
| 160 stored by the first insn. | |
| 161 | |
| 162 insn 5 in gcc.c-torture/compile/990203-1.c simple case. | |
| 163 insn 15 in gcc.c-torture/execute/20001017-2.c simple case. | |
| 164 insn 25 in gcc.c-torture/execute/20001026-1.c simple case. | |
| 165 insn 44 in gcc.c-torture/execute/20010910-1.c simple case. | |
| 166 | |
| 167 2) The cleaning up of spill code is quite profitable. It currently | |
| 168 depends on reading tea leaves and chicken entrails left by reload. | |
| 169 This pass depends on reload creating a singleton alias set for each | |
| 170 spill slot and telling the next dse pass which of these alias sets | |
| 171 are the singletons. Rather than analyze the addresses of the | |
| 172 spills, dse's spill processing just does analysis of the loads and | |
| 173 stores that use those alias sets. There are three cases where this | |
| 174 falls short: | |
| 175 | |
| 176 a) Reload sometimes creates the slot for one mode of access, and | |
| 177 then inserts loads and/or stores for a smaller mode. In this | |
| 178 case, the current code just punts on the slot. The proper thing | |
| 179 to do is to back out and use one bit vector position for each | |
| 180 byte of the entity associated with the slot. This depends on | |
| 181 KNOWING that reload always generates the accesses for each of the | |
| 182 bytes in some canonical (read that easy to understand several | |
| 183 passes after reload happens) way. | |
| 184 | |
| 185 b) Reload sometimes decides that spill slot it allocated was not | |
| 186 large enough for the mode and goes back and allocates more slots | |
| 187 with the same mode and alias set. The backout in this case is a | |
| 188 little more graceful than (a). In this case the slot is unmarked | |
| 189 as being a spill slot and if final address comes out to be based | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
190 off the frame pointer, the global algorithm handles this slot. |
| 0 | 191 |
| 192 c) For any pass that may prespill, there is currently no | |
| 193 mechanism to tell the dse pass that the slot being used has the | |
| 194 special properties that reload uses. It may be that all that is | |
| 195 required is to have those passes make the same calls that reload | |
| 196 does, assuming that the alias sets can be manipulated in the same | |
| 197 way. */ | |
| 198 | |
| 199 /* There are limits to the size of constant offsets we model for the | |
| 200 global problem. There are certainly test cases, that exceed this | |
| 201 limit, however, it is unlikely that there are important programs | |
| 202 that really have constant offsets this size. */ | |
| 203 #define MAX_OFFSET (64 * 1024) | |
| 204 | |
| 111 | 205 /* Obstack for the DSE dataflow bitmaps. We don't want to put these |
| 206 on the default obstack because these bitmaps can grow quite large | |
| 207 (~2GB for the small (!) test case of PR54146) and we'll hold on to | |
| 208 all that memory until the end of the compiler run. | |
| 209 As a bonus, delete_tree_live_info can destroy all the bitmaps by just | |
| 210 releasing the whole obstack. */ | |
| 211 static bitmap_obstack dse_bitmap_obstack; | |
| 212 | |
| 213 /* Obstack for other data. As for above: Kinda nice to be able to | |
| 214 throw it all away at the end in one big sweep. */ | |
| 215 static struct obstack dse_obstack; | |
| 216 | |
| 217 /* Scratch bitmap for cselib's cselib_expand_value_rtx. */ | |
| 0 | 218 static bitmap scratch = NULL; |
| 111 | 219 |
| 220 struct insn_info_type; | |
| 0 | 221 |
| 222 /* This structure holds information about a candidate store. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
223 struct store_info |
| 0 | 224 { |
| 225 | |
| 226 /* False means this is a clobber. */ | |
| 227 bool is_set; | |
| 228 | |
| 229 /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ | |
| 230 bool is_large; | |
| 231 | |
| 232 /* The id of the mem group of the base address. If rtx_varies_p is | |
| 233 true, this is -1. Otherwise, it is the index into the group | |
| 234 table. */ | |
| 235 int group_id; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
236 |
| 0 | 237 /* This is the cselib value. */ |
| 238 cselib_val *cse_base; | |
| 239 | |
| 240 /* This canonized mem. */ | |
| 241 rtx mem; | |
| 242 | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
243 /* Canonized MEM address for use by canon_true_dependence. */ |
| 0 | 244 rtx mem_addr; |
| 245 | |
| 246 /* The offset of the first and byte before the last byte associated | |
| 247 with the operation. */ | |
| 248 HOST_WIDE_INT begin, end; | |
| 249 | |
| 250 union | |
| 251 { | |
| 252 /* A bitmask as wide as the number of bytes in the word that | |
| 253 contains a 1 if the byte may be needed. The store is unused if | |
| 254 all of the bits are 0. This is used if IS_LARGE is false. */ | |
| 255 unsigned HOST_WIDE_INT small_bitmask; | |
| 256 | |
| 257 struct | |
| 258 { | |
| 259 /* A bitmap with one bit per byte. Cleared bit means the position | |
| 260 is needed. Used if IS_LARGE is false. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
261 bitmap bmap; |
| 0 | 262 |
| 263 /* Number of set bits (i.e. unneeded bytes) in BITMAP. If it is | |
| 264 equal to END - BEGIN, the whole store is unused. */ | |
| 265 int count; | |
| 266 } large; | |
| 267 } positions_needed; | |
| 268 | |
| 269 /* The next store info for this insn. */ | |
| 270 struct store_info *next; | |
| 271 | |
| 272 /* The right hand side of the store. This is used if there is a | |
| 273 subsequent reload of the mems address somewhere later in the | |
| 274 basic block. */ | |
| 275 rtx rhs; | |
| 276 | |
| 277 /* If rhs is or holds a constant, this contains that constant, | |
| 278 otherwise NULL. */ | |
| 279 rtx const_rhs; | |
| 280 | |
| 281 /* Set if this store stores the same constant value as REDUNDANT_REASON | |
| 282 insn stored. These aren't eliminated early, because doing that | |
| 283 might prevent the earlier larger store to be eliminated. */ | |
| 111 | 284 struct insn_info_type *redundant_reason; |
| 0 | 285 }; |
| 286 | |
| 287 /* Return a bitmask with the first N low bits set. */ | |
| 288 | |
| 289 static unsigned HOST_WIDE_INT | |
| 290 lowpart_bitmask (int n) | |
| 291 { | |
| 111 | 292 unsigned HOST_WIDE_INT mask = HOST_WIDE_INT_M1U; |
| 0 | 293 return mask >> (HOST_BITS_PER_WIDE_INT - n); |
| 294 } | |
| 295 | |
| 111 | 296 static object_allocator<store_info> cse_store_info_pool ("cse_store_info_pool"); |
| 297 | |
| 298 static object_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool"); | |
| 0 | 299 |
| 300 /* This structure holds information about a load. These are only | |
| 301 built for rtx bases. */ | |
| 111 | 302 struct read_info_type |
| 0 | 303 { |
| 304 /* The id of the mem group of the base address. */ | |
| 305 int group_id; | |
| 306 | |
| 307 /* The offset of the first and byte after the last byte associated | |
| 308 with the operation. If begin == end == 0, the read did not have | |
| 309 a constant offset. */ | |
| 310 int begin, end; | |
| 311 | |
| 312 /* The mem being read. */ | |
| 313 rtx mem; | |
| 314 | |
| 315 /* The next read_info for this insn. */ | |
| 111 | 316 struct read_info_type *next; |
| 0 | 317 }; |
| 111 | 318 typedef struct read_info_type *read_info_t; |
| 319 | |
| 320 static object_allocator<read_info_type> read_info_type_pool ("read_info_pool"); | |
| 0 | 321 |
| 322 /* One of these records is created for each insn. */ | |
| 323 | |
| 111 | 324 struct insn_info_type |
| 0 | 325 { |
| 326 /* Set true if the insn contains a store but the insn itself cannot | |
| 327 be deleted. This is set if the insn is a parallel and there is | |
| 328 more than one non dead output or if the insn is in some way | |
| 329 volatile. */ | |
| 330 bool cannot_delete; | |
| 331 | |
| 332 /* This field is only used by the global algorithm. It is set true | |
| 333 if the insn contains any read of mem except for a (1). This is | |
| 334 also set if the insn is a call or has a clobber mem. If the insn | |
| 335 contains a wild read, the use_rec will be null. */ | |
| 336 bool wild_read; | |
| 337 | |
| 111 | 338 /* This is true only for CALL instructions which could potentially read |
| 339 any non-frame memory location. This field is used by the global | |
| 340 algorithm. */ | |
| 341 bool non_frame_wild_read; | |
| 342 | |
| 0 | 343 /* This field is only used for the processing of const functions. |
| 344 These functions cannot read memory, but they can read the stack | |
| 345 because that is where they may get their parms. We need to be | |
| 346 this conservative because, like the store motion pass, we don't | |
| 347 consider CALL_INSN_FUNCTION_USAGE when processing call insns. | |
| 348 Moreover, we need to distinguish two cases: | |
| 349 1. Before reload (register elimination), the stores related to | |
| 350 outgoing arguments are stack pointer based and thus deemed | |
| 351 of non-constant base in this pass. This requires special | |
| 352 handling but also means that the frame pointer based stores | |
| 353 need not be killed upon encountering a const function call. | |
| 354 2. After reload, the stores related to outgoing arguments can be | |
| 355 either stack pointer or hard frame pointer based. This means | |
| 356 that we have no other choice than also killing all the frame | |
| 357 pointer based stores upon encountering a const function call. | |
| 111 | 358 This field is set after reload for const function calls and before |
| 359 reload for const tail function calls on targets where arg pointer | |
| 360 is the frame pointer. Having this set is less severe than a wild | |
| 361 read, it just means that all the frame related stores are killed | |
| 362 rather than all the stores. */ | |
| 0 | 363 bool frame_read; |
| 364 | |
| 365 /* This field is only used for the processing of const functions. | |
| 366 It is set if the insn may contain a stack pointer based store. */ | |
| 367 bool stack_pointer_based; | |
| 368 | |
| 369 /* This is true if any of the sets within the store contains a | |
| 370 cselib base. Such stores can only be deleted by the local | |
| 371 algorithm. */ | |
| 372 bool contains_cselib_groups; | |
| 373 | |
| 374 /* The insn. */ | |
| 111 | 375 rtx_insn *insn; |
| 0 | 376 |
| 377 /* The list of mem sets or mem clobbers that are contained in this | |
| 378 insn. If the insn is deletable, it contains only one mem set. | |
| 379 But it could also contain clobbers. Insns that contain more than | |
| 380 one mem set are not deletable, but each of those mems are here in | |
| 381 order to provide info to delete other insns. */ | |
| 111 | 382 store_info *store_rec; |
| 0 | 383 |
| 384 /* The linked list of mem uses in this insn. Only the reads from | |
| 385 rtx bases are listed here. The reads to cselib bases are | |
| 386 completely processed during the first scan and so are never | |
| 387 created. */ | |
| 388 read_info_t read_rec; | |
| 389 | |
| 111 | 390 /* The live fixed registers. We assume only fixed registers can |
| 391 cause trouble by being clobbered from an expanded pattern; | |
| 392 storing only the live fixed registers (rather than all registers) | |
| 393 means less memory needs to be allocated / copied for the individual | |
| 394 stores. */ | |
| 395 regset fixed_regs_live; | |
| 396 | |
| 0 | 397 /* The prev insn in the basic block. */ |
| 111 | 398 struct insn_info_type * prev_insn; |
| 0 | 399 |
| 400 /* The linked list of insns that are in consideration for removal in | |
| 111 | 401 the forwards pass through the basic block. This pointer may be |
| 0 | 402 trash as it is not cleared when a wild read occurs. The only |
| 403 time it is guaranteed to be correct is when the traversal starts | |
| 404 at active_local_stores. */ | |
| 111 | 405 struct insn_info_type * next_local_store; |
| 0 | 406 }; |
| 111 | 407 typedef struct insn_info_type *insn_info_t; |
| 408 | |
| 409 static object_allocator<insn_info_type> insn_info_type_pool ("insn_info_pool"); | |
| 0 | 410 |
| 411 /* The linked list of stores that are under consideration in this | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
412 basic block. */ |
| 0 | 413 static insn_info_t active_local_stores; |
| 111 | 414 static int active_local_stores_len; |
| 415 | |
| 416 struct dse_bb_info_type | |
| 0 | 417 { |
| 418 /* Pointer to the insn info for the last insn in the block. These | |
| 419 are linked so this is how all of the insns are reached. During | |
| 420 scanning this is the current insn being scanned. */ | |
| 421 insn_info_t last_insn; | |
| 422 | |
| 423 /* The info for the global dataflow problem. */ | |
| 424 | |
| 425 | |
| 426 /* This is set if the transfer function should and in the wild_read | |
| 427 bitmap before applying the kill and gen sets. That vector knocks | |
| 428 out most of the bits in the bitmap and thus speeds up the | |
| 429 operations. */ | |
| 430 bool apply_wild_read; | |
| 431 | |
| 432 /* The following 4 bitvectors hold information about which positions | |
| 433 of which stores are live or dead. They are indexed by | |
| 434 get_bitmap_index. */ | |
| 435 | |
| 436 /* The set of store positions that exist in this block before a wild read. */ | |
| 437 bitmap gen; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
438 |
| 0 | 439 /* The set of load positions that exist in this block above the |
| 440 same position of a store. */ | |
| 441 bitmap kill; | |
| 442 | |
| 443 /* The set of stores that reach the top of the block without being | |
| 444 killed by a read. | |
| 445 | |
| 446 Do not represent the in if it is all ones. Note that this is | |
| 447 what the bitvector should logically be initialized to for a set | |
| 448 intersection problem. However, like the kill set, this is too | |
| 449 expensive. So initially, the in set will only be created for the | |
| 450 exit block and any block that contains a wild read. */ | |
| 451 bitmap in; | |
| 452 | |
| 453 /* The set of stores that reach the bottom of the block from it's | |
| 454 successors. | |
| 455 | |
| 456 Do not represent the in if it is all ones. Note that this is | |
| 457 what the bitvector should logically be initialized to for a set | |
| 458 intersection problem. However, like the kill and in set, this is | |
| 459 too expensive. So what is done is that the confluence operator | |
| 460 just initializes the vector from one of the out sets of the | |
| 461 successors of the block. */ | |
| 462 bitmap out; | |
| 463 | |
| 464 /* The following bitvector is indexed by the reg number. It | |
| 465 contains the set of regs that are live at the current instruction | |
| 466 being processed. While it contains info for all of the | |
| 111 | 467 registers, only the hard registers are actually examined. It is used |
| 468 to assure that shift and/or add sequences that are inserted do not | |
| 469 accidentally clobber live hard regs. */ | |
| 0 | 470 bitmap regs_live; |
| 471 }; | |
| 472 | |
| 111 | 473 typedef struct dse_bb_info_type *bb_info_t; |
| 474 | |
| 475 static object_allocator<dse_bb_info_type> dse_bb_info_type_pool | |
| 476 ("bb_info_pool"); | |
| 0 | 477 |
| 478 /* Table to hold all bb_infos. */ | |
| 479 static bb_info_t *bb_table; | |
| 480 | |
| 481 /* There is a group_info for each rtx base that is used to reference | |
| 482 memory. There are also not many of the rtx bases because they are | |
| 483 very limited in scope. */ | |
| 484 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
485 struct group_info |
| 0 | 486 { |
| 487 /* The actual base of the address. */ | |
| 488 rtx rtx_base; | |
| 489 | |
| 490 /* The sequential id of the base. This allows us to have a | |
| 491 canonical ordering of these that is not based on addresses. */ | |
| 492 int id; | |
| 493 | |
| 494 /* True if there are any positions that are to be processed | |
| 495 globally. */ | |
| 496 bool process_globally; | |
| 497 | |
| 498 /* True if the base of this group is either the frame_pointer or | |
| 499 hard_frame_pointer. */ | |
| 500 bool frame_related; | |
| 501 | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
502 /* A mem wrapped around the base pointer for the group in order to do |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
503 read dependency. It must be given BLKmode in order to encompass all |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
504 the possible offsets from the base. */ |
| 0 | 505 rtx base_mem; |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
506 |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
507 /* Canonized version of base_mem's address. */ |
|
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
508 rtx canon_base_addr; |
| 0 | 509 |
| 510 /* These two sets of two bitmaps are used to keep track of how many | |
| 511 stores are actually referencing that position from this base. We | |
| 512 only do this for rtx bases as this will be used to assign | |
| 513 positions in the bitmaps for the global problem. Bit N is set in | |
| 514 store1 on the first store for offset N. Bit N is set in store2 | |
| 515 for the second store to offset N. This is all we need since we | |
| 516 only care about offsets that have two or more stores for them. | |
| 517 | |
| 518 The "_n" suffix is for offsets less than 0 and the "_p" suffix is | |
| 519 for 0 and greater offsets. | |
| 520 | |
| 521 There is one special case here, for stores into the stack frame, | |
| 522 we will or store1 into store2 before deciding which stores look | |
| 523 at globally. This is because stores to the stack frame that have | |
| 524 no other reads before the end of the function can also be | |
| 525 deleted. */ | |
| 526 bitmap store1_n, store1_p, store2_n, store2_p; | |
| 527 | |
| 111 | 528 /* These bitmaps keep track of offsets in this group escape this function. |
| 529 An offset escapes if it corresponds to a named variable whose | |
| 530 addressable flag is set. */ | |
| 531 bitmap escaped_n, escaped_p; | |
| 532 | |
| 0 | 533 /* The positions in this bitmap have the same assignments as the in, |
| 534 out, gen and kill bitmaps. This bitmap is all zeros except for | |
| 535 the positions that are occupied by stores for this group. */ | |
| 536 bitmap group_kill; | |
| 537 | |
| 538 /* The offset_map is used to map the offsets from this base into | |
| 539 positions in the global bitmaps. It is only created after all of | |
| 540 the all of stores have been scanned and we know which ones we | |
| 541 care about. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
542 int *offset_map_n, *offset_map_p; |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
543 int offset_map_size_n, offset_map_size_p; |
| 0 | 544 }; |
| 111 | 545 |
| 546 static object_allocator<group_info> group_info_pool ("rtx_group_info_pool"); | |
| 0 | 547 |
| 548 /* Index into the rtx_group_vec. */ | |
| 549 static int rtx_group_next_id; | |
| 550 | |
| 111 | 551 |
| 552 static vec<group_info *> rtx_group_vec; | |
| 0 | 553 |
| 554 | |
| 555 /* This structure holds the set of changes that are being deferred | |
| 556 when removing read operation. See replace_read. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
557 struct deferred_change |
| 0 | 558 { |
| 559 | |
| 560 /* The mem that is being replaced. */ | |
| 561 rtx *loc; | |
| 562 | |
| 563 /* The reg it is being replaced with. */ | |
| 564 rtx reg; | |
| 565 | |
| 566 struct deferred_change *next; | |
| 567 }; | |
| 568 | |
| 111 | 569 static object_allocator<deferred_change> deferred_change_pool |
| 570 ("deferred_change_pool"); | |
| 571 | |
| 572 static deferred_change *deferred_change_list = NULL; | |
| 0 | 573 |
| 574 /* This is true except if cfun->stdarg -- i.e. we cannot do | |
| 575 this for vararg functions because they play games with the frame. */ | |
| 576 static bool stores_off_frame_dead_at_return; | |
| 577 | |
| 578 /* Counter for stats. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
579 static int globally_deleted; |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
580 static int locally_deleted; |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
581 |
| 0 | 582 static bitmap all_blocks; |
| 583 | |
| 111 | 584 /* Locations that are killed by calls in the global phase. */ |
| 585 static bitmap kill_on_calls; | |
| 586 | |
| 0 | 587 /* The number of bits used in the global bitmaps. */ |
| 588 static unsigned int current_position; | |
| 589 | |
| 590 /*---------------------------------------------------------------------------- | |
| 591 Zeroth step. | |
| 592 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
593 Initialization. |
| 0 | 594 ----------------------------------------------------------------------------*/ |
| 595 | |
| 596 | |
| 597 /* Hashtable callbacks for maintaining the "bases" field of | |
| 598 store_group_info, given that the addresses are function invariants. */ | |
| 599 | |
| 111 | 600 struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
| 0 | 601 { |
| 111 | 602 static inline hashval_t hash (const group_info *); |
| 603 static inline bool equal (const group_info *, const group_info *); | |
| 604 }; | |
| 605 | |
| 606 inline bool | |
| 607 invariant_group_base_hasher::equal (const group_info *gi1, | |
| 608 const group_info *gi2) | |
| 609 { | |
| 0 | 610 return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
| 611 } | |
| 612 | |
| 111 | 613 inline hashval_t |
| 614 invariant_group_base_hasher::hash (const group_info *gi) | |
| 0 | 615 { |
| 616 int do_not_record; | |
| 617 return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); | |
| 618 } | |
| 619 | |
| 111 | 620 /* Tables of group_info structures, hashed by base value. */ |
| 621 static hash_table<invariant_group_base_hasher> *rtx_group_table; | |
| 622 | |
| 0 | 623 |
| 624 /* Get the GROUP for BASE. Add a new group if it is not there. */ | |
| 625 | |
| 111 | 626 static group_info * |
| 0 | 627 get_group_info (rtx base) |
| 628 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
629 struct group_info tmp_gi; |
| 111 | 630 group_info *gi; |
| 631 group_info **slot; | |
| 632 | |
| 633 gcc_assert (base != NULL_RTX); | |
| 634 | |
| 635 /* Find the store_base_info structure for BASE, creating a new one | |
| 636 if necessary. */ | |
| 637 tmp_gi.rtx_base = base; | |
| 638 slot = rtx_group_table->find_slot (&tmp_gi, INSERT); | |
| 639 gi = *slot; | |
| 0 | 640 |
| 641 if (gi == NULL) | |
| 642 { | |
| 111 | 643 *slot = gi = group_info_pool.allocate (); |
| 0 | 644 gi->rtx_base = base; |
| 645 gi->id = rtx_group_next_id++; | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
646 gi->base_mem = gen_rtx_MEM (BLKmode, base); |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
647 gi->canon_base_addr = canon_rtx (base); |
| 111 | 648 gi->store1_n = BITMAP_ALLOC (&dse_bitmap_obstack); |
| 649 gi->store1_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
| 650 gi->store2_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
| 651 gi->store2_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
| 652 gi->escaped_p = BITMAP_ALLOC (&dse_bitmap_obstack); | |
| 653 gi->escaped_n = BITMAP_ALLOC (&dse_bitmap_obstack); | |
| 654 gi->group_kill = BITMAP_ALLOC (&dse_bitmap_obstack); | |
| 0 | 655 gi->process_globally = false; |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
656 gi->frame_related = |
| 0 | 657 (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx); |
| 658 gi->offset_map_size_n = 0; | |
| 659 gi->offset_map_size_p = 0; | |
| 660 gi->offset_map_n = NULL; | |
| 661 gi->offset_map_p = NULL; | |
| 111 | 662 rtx_group_vec.safe_push (gi); |
| 0 | 663 } |
| 664 | |
| 665 return gi; | |
| 666 } | |
| 667 | |
| 668 | |
| 669 /* Initialization of data structures. */ | |
| 670 | |
| 671 static void | |
| 672 dse_step0 (void) | |
| 673 { | |
| 674 locally_deleted = 0; | |
| 675 globally_deleted = 0; | |
| 111 | 676 |
| 677 bitmap_obstack_initialize (&dse_bitmap_obstack); | |
| 678 gcc_obstack_init (&dse_obstack); | |
| 679 | |
| 680 scratch = BITMAP_ALLOC (®_obstack); | |
| 681 kill_on_calls = BITMAP_ALLOC (&dse_bitmap_obstack); | |
| 682 | |
| 683 | |
| 684 rtx_group_table = new hash_table<invariant_group_base_hasher> (11); | |
| 685 | |
| 686 bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); | |
| 0 | 687 rtx_group_next_id = 0; |
| 688 | |
| 689 stores_off_frame_dead_at_return = !cfun->stdarg; | |
| 690 | |
| 691 init_alias_analysis (); | |
| 692 } | |
| 693 | |
| 694 | |
| 695 | |
| 696 /*---------------------------------------------------------------------------- | |
| 697 First step. | |
| 698 | |
| 699 Scan all of the insns. Any random ordering of the blocks is fine. | |
| 700 Each block is scanned in forward order to accommodate cselib which | |
| 701 is used to remove stores with non-constant bases. | |
| 702 ----------------------------------------------------------------------------*/ | |
| 703 | |
| 704 /* Delete all of the store_info recs from INSN_INFO. */ | |
| 705 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
706 static void |
| 0 | 707 free_store_info (insn_info_t insn_info) |
| 708 { | |
| 111 | 709 store_info *cur = insn_info->store_rec; |
| 710 while (cur) | |
| 0 | 711 { |
| 111 | 712 store_info *next = cur->next; |
| 713 if (cur->is_large) | |
| 714 BITMAP_FREE (cur->positions_needed.large.bmap); | |
| 715 if (cur->cse_base) | |
| 716 cse_store_info_pool.remove (cur); | |
| 0 | 717 else |
| 111 | 718 rtx_store_info_pool.remove (cur); |
| 719 cur = next; | |
| 0 | 720 } |
| 721 | |
| 722 insn_info->cannot_delete = true; | |
| 723 insn_info->contains_cselib_groups = false; | |
| 724 insn_info->store_rec = NULL; | |
| 725 } | |
| 726 | |
| 111 | 727 struct note_add_store_info |
| 728 { | |
| 729 rtx_insn *first, *current; | |
| 730 regset fixed_regs_live; | |
| 731 bool failure; | |
| 732 }; | |
| 733 | |
| 734 /* Callback for emit_inc_dec_insn_before via note_stores. | |
| 735 Check if a register is clobbered which is live afterwards. */ | |
| 736 | |
| 737 static void | |
| 738 note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) | |
| 739 { | |
| 740 rtx_insn *insn; | |
| 741 note_add_store_info *info = (note_add_store_info *) data; | |
| 742 | |
| 743 if (!REG_P (loc)) | |
| 744 return; | |
| 745 | |
| 746 /* If this register is referenced by the current or an earlier insn, | |
| 747 that's OK. E.g. this applies to the register that is being incremented | |
| 748 with this addition. */ | |
| 749 for (insn = info->first; | |
| 750 insn != NEXT_INSN (info->current); | |
| 751 insn = NEXT_INSN (insn)) | |
| 752 if (reg_referenced_p (loc, PATTERN (insn))) | |
| 753 return; | |
| 754 | |
| 755 /* If we come here, we have a clobber of a register that's only OK | |
| 756 if that register is not live. If we don't have liveness information | |
| 757 available, fail now. */ | |
| 758 if (!info->fixed_regs_live) | |
| 759 { | |
| 760 info->failure = true; | |
| 761 return; | |
| 762 } | |
| 763 /* Now check if this is a live fixed register. */ | |
| 764 unsigned int end_regno = END_REGNO (loc); | |
| 765 for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) | |
| 766 if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) | |
| 767 info->failure = true; | |
| 768 } | |
| 769 | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
770 /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
771 SRC + SRCOFF before insn ARG. */ |
| 0 | 772 |
| 773 static int | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
774 emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
775 rtx op ATTRIBUTE_UNUSED, |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
776 rtx dest, rtx src, rtx srcoff, void *arg) |
| 0 | 777 { |
| 111 | 778 insn_info_t insn_info = (insn_info_t) arg; |
| 779 rtx_insn *insn = insn_info->insn, *new_insn, *cur; | |
| 780 note_add_store_info info; | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
781 |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
782 /* We can reuse all operands without copying, because we are about |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
783 to delete the insn that contained it. */ |
| 111 | 784 if (srcoff) |
| 785 { | |
| 786 start_sequence (); | |
| 787 emit_insn (gen_add3_insn (dest, src, srcoff)); | |
| 788 new_insn = get_insns (); | |
| 789 end_sequence (); | |
| 790 } | |
| 791 else | |
| 792 new_insn = gen_move_insn (dest, src); | |
| 793 info.first = new_insn; | |
| 794 info.fixed_regs_live = insn_info->fixed_regs_live; | |
| 795 info.failure = false; | |
| 796 for (cur = new_insn; cur; cur = NEXT_INSN (cur)) | |
| 797 { | |
| 798 info.current = cur; | |
| 799 note_stores (PATTERN (cur), note_add_store, &info); | |
| 800 } | |
| 801 | |
| 802 /* If a failure was flagged above, return 1 so that for_each_inc_dec will | |
| 803 return it immediately, communicating the failure to its caller. */ | |
| 804 if (info.failure) | |
| 805 return 1; | |
| 806 | |
| 807 emit_insn_before (new_insn, insn); | |
| 808 | |
| 809 return 0; | |
| 0 | 810 } |
| 811 | |
| 111 | 812 /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
| 813 is there, is split into a separate insn. | |
| 814 Return true on success (or if there was nothing to do), false on failure. */ | |
| 815 | |
| 816 static bool | |
| 817 check_for_inc_dec_1 (insn_info_t insn_info) | |
| 0 | 818 { |
| 111 | 819 rtx_insn *insn = insn_info->insn; |
| 0 | 820 rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
| 821 if (note) | |
| 111 | 822 return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
| 823 insn_info) == 0; | |
| 824 return true; | |
| 0 | 825 } |
| 826 | |
| 827 | |
| 111 | 828 /* Entry point for postreload. If you work on reload_cse, or you need this |
| 829 anywhere else, consider if you can provide register liveness information | |
| 830 and add a parameter to this function so that it can be passed down in | |
| 831 insn_info.fixed_regs_live. */ | |
| 832 bool | |
| 833 check_for_inc_dec (rtx_insn *insn) | |
| 834 { | |
| 835 insn_info_type insn_info; | |
| 836 rtx note; | |
| 837 | |
| 838 insn_info.insn = insn; | |
| 839 insn_info.fixed_regs_live = NULL; | |
| 840 note = find_reg_note (insn, REG_INC, NULL_RTX); | |
| 841 if (note) | |
| 842 return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, | |
| 843 &insn_info) == 0; | |
| 844 return true; | |
| 845 } | |
| 846 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
847 /* Delete the insn and free all of the fields inside INSN_INFO. */ |
| 0 | 848 |
| 849 static void | |
| 850 delete_dead_store_insn (insn_info_t insn_info) | |
| 851 { | |
| 852 read_info_t read_info; | |
| 853 | |
| 854 if (!dbg_cnt (dse)) | |
| 855 return; | |
| 856 | |
| 111 | 857 if (!check_for_inc_dec_1 (insn_info)) |
| 858 return; | |
| 859 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 860 fprintf (dump_file, "Locally deleting insn %d\n", | |
| 861 INSN_UID (insn_info->insn)); | |
| 0 | 862 |
| 863 free_store_info (insn_info); | |
| 864 read_info = insn_info->read_rec; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
865 |
| 0 | 866 while (read_info) |
| 867 { | |
| 868 read_info_t next = read_info->next; | |
| 111 | 869 read_info_type_pool.remove (read_info); |
| 0 | 870 read_info = next; |
| 871 } | |
| 872 insn_info->read_rec = NULL; | |
| 873 | |
| 874 delete_insn (insn_info->insn); | |
| 875 locally_deleted++; | |
| 876 insn_info->insn = NULL; | |
| 877 | |
| 878 insn_info->wild_read = false; | |
| 879 } | |
| 880 | |
| 111 | 881 /* Return whether DECL, a local variable, can possibly escape the current |
| 882 function scope. */ | |
| 883 | |
| 884 static bool | |
| 885 local_variable_can_escape (tree decl) | |
| 886 { | |
| 887 if (TREE_ADDRESSABLE (decl)) | |
| 888 return true; | |
| 889 | |
| 890 /* If this is a partitioned variable, we need to consider all the variables | |
| 891 in the partition. This is necessary because a store into one of them can | |
| 892 be replaced with a store into another and this may not change the outcome | |
| 893 of the escape analysis. */ | |
| 894 if (cfun->gimple_df->decls_to_pointers != NULL) | |
| 895 { | |
| 896 tree *namep = cfun->gimple_df->decls_to_pointers->get (decl); | |
| 897 if (namep) | |
| 898 return TREE_ADDRESSABLE (*namep); | |
| 899 } | |
| 900 | |
| 901 return false; | |
| 902 } | |
| 903 | |
| 904 /* Return whether EXPR can possibly escape the current function scope. */ | |
| 905 | |
| 906 static bool | |
| 907 can_escape (tree expr) | |
| 908 { | |
| 909 tree base; | |
| 910 if (!expr) | |
| 911 return true; | |
| 912 base = get_base_address (expr); | |
| 913 if (DECL_P (base) | |
| 914 && !may_be_aliased (base) | |
| 915 && !(VAR_P (base) | |
| 916 && !DECL_EXTERNAL (base) | |
| 917 && !TREE_STATIC (base) | |
| 918 && local_variable_can_escape (base))) | |
| 919 return false; | |
| 920 return true; | |
| 921 } | |
| 0 | 922 |
| 923 /* Set the store* bitmaps offset_map_size* fields in GROUP based on | |
| 924 OFFSET and WIDTH. */ | |
| 925 | |
| 926 static void | |
| 111 | 927 set_usage_bits (group_info *group, HOST_WIDE_INT offset, HOST_WIDE_INT width, |
| 928 tree expr) | |
| 0 | 929 { |
| 930 HOST_WIDE_INT i; | |
| 111 | 931 bool expr_escapes = can_escape (expr); |
| 0 | 932 if (offset > -MAX_OFFSET && offset + width < MAX_OFFSET) |
| 933 for (i=offset; i<offset+width; i++) | |
| 934 { | |
| 935 bitmap store1; | |
| 936 bitmap store2; | |
| 111 | 937 bitmap escaped; |
| 0 | 938 int ai; |
| 939 if (i < 0) | |
| 940 { | |
| 941 store1 = group->store1_n; | |
| 942 store2 = group->store2_n; | |
| 111 | 943 escaped = group->escaped_n; |
| 0 | 944 ai = -i; |
| 945 } | |
| 946 else | |
| 947 { | |
| 948 store1 = group->store1_p; | |
| 949 store2 = group->store2_p; | |
| 111 | 950 escaped = group->escaped_p; |
| 0 | 951 ai = i; |
| 952 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
953 |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
954 if (!bitmap_set_bit (store1, ai)) |
| 0 | 955 bitmap_set_bit (store2, ai); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
956 else |
| 0 | 957 { |
| 958 if (i < 0) | |
| 959 { | |
| 960 if (group->offset_map_size_n < ai) | |
| 961 group->offset_map_size_n = ai; | |
| 962 } | |
| 963 else | |
| 964 { | |
| 965 if (group->offset_map_size_p < ai) | |
| 966 group->offset_map_size_p = ai; | |
| 967 } | |
| 968 } | |
| 111 | 969 if (expr_escapes) |
| 970 bitmap_set_bit (escaped, ai); | |
| 0 | 971 } |
| 972 } | |
| 973 | |
| 111 | 974 static void |
| 975 reset_active_stores (void) | |
| 976 { | |
| 977 active_local_stores = NULL; | |
| 978 active_local_stores_len = 0; | |
| 979 } | |
| 980 | |
| 981 /* Free all READ_REC of the LAST_INSN of BB_INFO. */ | |
| 982 | |
| 983 static void | |
| 984 free_read_records (bb_info_t bb_info) | |
| 985 { | |
| 986 insn_info_t insn_info = bb_info->last_insn; | |
| 987 read_info_t *ptr = &insn_info->read_rec; | |
| 988 while (*ptr) | |
| 989 { | |
| 990 read_info_t next = (*ptr)->next; | |
| 991 read_info_type_pool.remove (*ptr); | |
| 992 *ptr = next; | |
| 993 } | |
| 994 } | |
| 0 | 995 |
| 996 /* Set the BB_INFO so that the last insn is marked as a wild read. */ | |
| 997 | |
| 998 static void | |
| 999 add_wild_read (bb_info_t bb_info) | |
| 1000 { | |
| 1001 insn_info_t insn_info = bb_info->last_insn; | |
| 1002 insn_info->wild_read = true; | |
| 111 | 1003 free_read_records (bb_info); |
| 1004 reset_active_stores (); | |
| 0 | 1005 } |
| 1006 | |
| 111 | 1007 /* Set the BB_INFO so that the last insn is marked as a wild read of |
| 1008 non-frame locations. */ | |
| 1009 | |
| 1010 static void | |
| 1011 add_non_frame_wild_read (bb_info_t bb_info) | |
| 1012 { | |
| 1013 insn_info_t insn_info = bb_info->last_insn; | |
| 1014 insn_info->non_frame_wild_read = true; | |
| 1015 free_read_records (bb_info); | |
| 1016 reset_active_stores (); | |
| 1017 } | |
| 0 | 1018 |
| 1019 /* Return true if X is a constant or one of the registers that behave | |
| 1020 as a constant over the life of a function. This is equivalent to | |
| 1021 !rtx_varies_p for memory addresses. */ | |
| 1022 | |
| 1023 static bool | |
| 1024 const_or_frame_p (rtx x) | |
| 1025 { | |
| 111 | 1026 if (CONSTANT_P (x)) |
| 1027 return true; | |
| 1028 | |
| 1029 if (GET_CODE (x) == REG) | |
| 0 | 1030 { |
| 1031 /* Note that we have to test for the actual rtx used for the frame | |
| 1032 and arg pointers and not just the register number in case we have | |
| 1033 eliminated the frame and/or arg pointer and are using it | |
| 1034 for pseudos. */ | |
| 1035 if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx | |
| 1036 /* The arg pointer varies if it is not a fixed register. */ | |
| 1037 || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) | |
| 1038 || x == pic_offset_table_rtx) | |
| 1039 return true; | |
| 1040 return false; | |
| 1041 } | |
| 111 | 1042 |
| 1043 return false; | |
| 0 | 1044 } |
| 1045 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1046 /* Take all reasonable action to put the address of MEM into the form |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1047 that we can do analysis on. |
| 0 | 1048 |
| 1049 The gold standard is to get the address into the form: address + | |
| 1050 OFFSET where address is something that rtx_varies_p considers a | |
| 1051 constant. When we can get the address in this form, we can do | |
| 1052 global analysis on it. Note that for constant bases, address is | |
| 1053 not actually returned, only the group_id. The address can be | |
| 1054 obtained from that. | |
| 1055 | |
| 1056 If that fails, we try cselib to get a value we can at least use | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1057 locally. If that fails we return false. |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1058 |
| 0 | 1059 The GROUP_ID is set to -1 for cselib bases and the index of the |
| 1060 group for non_varying bases. | |
| 1061 | |
| 1062 FOR_READ is true if this is a mem read and false if not. */ | |
| 1063 | |
| 1064 static bool | |
| 1065 canon_address (rtx mem, | |
| 1066 int *group_id, | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1067 HOST_WIDE_INT *offset, |
| 0 | 1068 cselib_val **base) |
| 1069 { | |
| 111 | 1070 machine_mode address_mode = get_address_mode (mem); |
| 0 | 1071 rtx mem_address = XEXP (mem, 0); |
| 1072 rtx expanded_address, address; | |
| 36 | 1073 int expanded; |
| 1074 | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1075 cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
| 0 | 1076 |
| 111 | 1077 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1078 { |
| 1079 fprintf (dump_file, " mem: "); | |
| 1080 print_inline_rtx (dump_file, mem_address, 0); | |
| 1081 fprintf (dump_file, "\n"); | |
| 1082 } | |
| 1083 | |
| 36 | 1084 /* First see if just canon_rtx (mem_address) is const or frame, |
| 1085 if not, try cselib_expand_value_rtx and call canon_rtx on that. */ | |
| 1086 address = NULL_RTX; | |
| 1087 for (expanded = 0; expanded < 2; expanded++) | |
| 1088 { | |
| 1089 if (expanded) | |
| 1090 { | |
| 1091 /* Use cselib to replace all of the reg references with the full | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1092 expression. This will take care of the case where we have |
| 36 | 1093 |
| 1094 r_x = base + offset; | |
| 1095 val = *r_x; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1096 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1097 by making it into |
| 36 | 1098 |
| 1099 val = *(base + offset); */ | |
| 1100 | |
| 1101 expanded_address = cselib_expand_value_rtx (mem_address, | |
| 1102 scratch, 5); | |
| 1103 | |
| 1104 /* If this fails, just go with the address from first | |
| 1105 iteration. */ | |
| 1106 if (!expanded_address) | |
| 1107 break; | |
| 1108 } | |
| 1109 else | |
| 1110 expanded_address = mem_address; | |
| 1111 | |
| 1112 /* Split the address into canonical BASE + OFFSET terms. */ | |
| 1113 address = canon_rtx (expanded_address); | |
| 1114 | |
| 1115 *offset = 0; | |
| 0 | 1116 |
| 111 | 1117 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 36 | 1118 { |
| 1119 if (expanded) | |
| 1120 { | |
| 1121 fprintf (dump_file, "\n after cselib_expand address: "); | |
| 1122 print_inline_rtx (dump_file, expanded_address, 0); | |
| 1123 fprintf (dump_file, "\n"); | |
| 1124 } | |
| 1125 | |
| 1126 fprintf (dump_file, "\n after canon_rtx address: "); | |
| 1127 print_inline_rtx (dump_file, address, 0); | |
| 1128 fprintf (dump_file, "\n"); | |
| 1129 } | |
| 1130 | |
| 1131 if (GET_CODE (address) == CONST) | |
| 1132 address = XEXP (address, 0); | |
| 1133 | |
| 1134 if (GET_CODE (address) == PLUS | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1135 && CONST_INT_P (XEXP (address, 1))) |
| 36 | 1136 { |
| 1137 *offset = INTVAL (XEXP (address, 1)); | |
| 1138 address = XEXP (address, 0); | |
| 1139 } | |
| 1140 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1141 if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1142 && const_or_frame_p (address)) |
| 36 | 1143 { |
| 111 | 1144 group_info *group = get_group_info (address); |
| 1145 | |
| 1146 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 36 | 1147 fprintf (dump_file, " gid=%d offset=%d \n", |
| 1148 group->id, (int)*offset); | |
| 1149 *base = NULL; | |
| 1150 *group_id = group->id; | |
| 1151 return true; | |
| 1152 } | |
| 0 | 1153 } |
| 36 | 1154 |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1155 *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
| 36 | 1156 *group_id = -1; |
| 1157 | |
| 1158 if (*base == NULL) | |
| 0 | 1159 { |
| 111 | 1160 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 36 | 1161 fprintf (dump_file, " no cselib val - should be a wild read.\n"); |
| 1162 return false; | |
| 0 | 1163 } |
| 111 | 1164 if (dump_file && (dump_flags & TDF_DETAILS)) |
|
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1165 fprintf (dump_file, " varying cselib base=%u:%u offset = %d\n", |
|
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1166 (*base)->uid, (*base)->hash, (int)*offset); |
| 0 | 1167 return true; |
| 1168 } | |
| 1169 | |
| 1170 | |
| 1171 /* Clear the rhs field from the active_local_stores array. */ | |
| 1172 | |
| 1173 static void | |
| 1174 clear_rhs_from_active_local_stores (void) | |
| 1175 { | |
| 1176 insn_info_t ptr = active_local_stores; | |
| 1177 | |
| 1178 while (ptr) | |
| 1179 { | |
| 111 | 1180 store_info *store_info = ptr->store_rec; |
| 0 | 1181 /* Skip the clobbers. */ |
| 1182 while (!store_info->is_set) | |
| 1183 store_info = store_info->next; | |
| 1184 | |
| 1185 store_info->rhs = NULL; | |
| 1186 store_info->const_rhs = NULL; | |
| 1187 | |
| 1188 ptr = ptr->next_local_store; | |
| 1189 } | |
| 1190 } | |
| 1191 | |
| 1192 | |
| 1193 /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ | |
| 1194 | |
| 1195 static inline void | |
| 111 | 1196 set_position_unneeded (store_info *s_info, int pos) |
| 0 | 1197 { |
| 1198 if (__builtin_expect (s_info->is_large, false)) | |
| 1199 { | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1200 if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1201 s_info->positions_needed.large.count++; |
| 0 | 1202 } |
| 1203 else | |
| 1204 s_info->positions_needed.small_bitmask | |
| 111 | 1205 &= ~(HOST_WIDE_INT_1U << pos); |
| 0 | 1206 } |
| 1207 | |
| 1208 /* Mark the whole store S_INFO as unneeded. */ | |
| 1209 | |
| 1210 static inline void | |
| 111 | 1211 set_all_positions_unneeded (store_info *s_info) |
| 0 | 1212 { |
| 1213 if (__builtin_expect (s_info->is_large, false)) | |
| 1214 { | |
| 1215 int pos, end = s_info->end - s_info->begin; | |
| 1216 for (pos = 0; pos < end; pos++) | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1217 bitmap_set_bit (s_info->positions_needed.large.bmap, pos); |
| 0 | 1218 s_info->positions_needed.large.count = end; |
| 1219 } | |
| 1220 else | |
| 111 | 1221 s_info->positions_needed.small_bitmask = HOST_WIDE_INT_0U; |
| 0 | 1222 } |
| 1223 | |
| 1224 /* Return TRUE if any bytes from S_INFO store are needed. */ | |
| 1225 | |
| 1226 static inline bool | |
| 111 | 1227 any_positions_needed_p (store_info *s_info) |
| 0 | 1228 { |
| 1229 if (__builtin_expect (s_info->is_large, false)) | |
| 1230 return (s_info->positions_needed.large.count | |
| 1231 < s_info->end - s_info->begin); | |
| 1232 else | |
| 111 | 1233 return (s_info->positions_needed.small_bitmask != HOST_WIDE_INT_0U); |
| 0 | 1234 } |
| 1235 | |
| 1236 /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO | |
| 1237 store are needed. */ | |
| 1238 | |
| 1239 static inline bool | |
| 111 | 1240 all_positions_needed_p (store_info *s_info, int start, int width) |
| 0 | 1241 { |
| 1242 if (__builtin_expect (s_info->is_large, false)) | |
| 1243 { | |
| 1244 int end = start + width; | |
| 1245 while (start < end) | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1246 if (bitmap_bit_p (s_info->positions_needed.large.bmap, start++)) |
| 0 | 1247 return false; |
| 1248 return true; | |
| 1249 } | |
| 1250 else | |
| 1251 { | |
| 1252 unsigned HOST_WIDE_INT mask = lowpart_bitmask (width) << start; | |
| 1253 return (s_info->positions_needed.small_bitmask & mask) == mask; | |
| 1254 } | |
| 1255 } | |
| 1256 | |
| 1257 | |
| 111 | 1258 static rtx get_stored_val (store_info *, machine_mode, HOST_WIDE_INT, |
| 0 | 1259 HOST_WIDE_INT, basic_block, bool); |
| 1260 | |
| 1261 | |
| 1262 /* BODY is an instruction pattern that belongs to INSN. Return 1 if | |
| 1263 there is a candidate store, after adding it to the appropriate | |
| 1264 local store group if so. */ | |
| 1265 | |
| 1266 static int | |
| 1267 record_store (rtx body, bb_info_t bb_info) | |
| 1268 { | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1269 rtx mem, rhs, const_rhs, mem_addr; |
| 0 | 1270 HOST_WIDE_INT offset = 0; |
| 1271 HOST_WIDE_INT width = 0; | |
| 1272 insn_info_t insn_info = bb_info->last_insn; | |
| 111 | 1273 store_info *store_info = NULL; |
| 0 | 1274 int group_id; |
| 1275 cselib_val *base = NULL; | |
| 1276 insn_info_t ptr, last, redundant_reason; | |
| 1277 bool store_is_unused; | |
| 1278 | |
| 1279 if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) | |
| 1280 return 0; | |
| 1281 | |
| 1282 mem = SET_DEST (body); | |
| 1283 | |
| 1284 /* If this is not used, then this cannot be used to keep the insn | |
| 1285 from being deleted. On the other hand, it does provide something | |
| 1286 that can be used to prove that another store is dead. */ | |
| 1287 store_is_unused | |
| 1288 = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); | |
| 1289 | |
| 1290 /* Check whether that value is a suitable memory location. */ | |
| 1291 if (!MEM_P (mem)) | |
| 1292 { | |
| 1293 /* If the set or clobber is unused, then it does not effect our | |
| 1294 ability to get rid of the entire insn. */ | |
| 1295 if (!store_is_unused) | |
| 1296 insn_info->cannot_delete = true; | |
| 1297 return 0; | |
| 1298 } | |
| 1299 | |
| 1300 /* At this point we know mem is a mem. */ | |
| 1301 if (GET_MODE (mem) == BLKmode) | |
| 1302 { | |
| 1303 if (GET_CODE (XEXP (mem, 0)) == SCRATCH) | |
| 1304 { | |
| 111 | 1305 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1306 fprintf (dump_file, " adding wild read for (clobber (mem:BLK (scratch))\n"); |
| 1307 add_wild_read (bb_info); | |
| 1308 insn_info->cannot_delete = true; | |
| 1309 return 0; | |
| 1310 } | |
| 1311 /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) | |
| 1312 as memset (addr, 0, 36); */ | |
| 111 | 1313 else if (!MEM_SIZE_KNOWN_P (mem) |
| 1314 || MEM_SIZE (mem) <= 0 | |
| 1315 || MEM_SIZE (mem) > MAX_OFFSET | |
| 0 | 1316 || GET_CODE (body) != SET |
| 1317 || !CONST_INT_P (SET_SRC (body))) | |
| 1318 { | |
| 1319 if (!store_is_unused) | |
| 1320 { | |
| 1321 /* If the set or clobber is unused, then it does not effect our | |
| 1322 ability to get rid of the entire insn. */ | |
| 1323 insn_info->cannot_delete = true; | |
| 1324 clear_rhs_from_active_local_stores (); | |
| 1325 } | |
| 1326 return 0; | |
| 1327 } | |
| 1328 } | |
| 1329 | |
| 1330 /* We can still process a volatile mem, we just cannot delete it. */ | |
| 1331 if (MEM_VOLATILE_P (mem)) | |
| 1332 insn_info->cannot_delete = true; | |
| 1333 | |
| 111 | 1334 if (!canon_address (mem, &group_id, &offset, &base)) |
| 0 | 1335 { |
| 1336 clear_rhs_from_active_local_stores (); | |
| 1337 return 0; | |
| 1338 } | |
| 1339 | |
| 1340 if (GET_MODE (mem) == BLKmode) | |
| 111 | 1341 width = MEM_SIZE (mem); |
| 0 | 1342 else |
| 111 | 1343 width = GET_MODE_SIZE (GET_MODE (mem)); |
| 1344 | |
| 1345 if (group_id >= 0) | |
| 0 | 1346 { |
| 1347 /* In the restrictive case where the base is a constant or the | |
| 1348 frame pointer we can do global analysis. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1349 |
| 111 | 1350 group_info *group |
| 1351 = rtx_group_vec[group_id]; | |
| 1352 tree expr = MEM_EXPR (mem); | |
| 1353 | |
| 1354 store_info = rtx_store_info_pool.allocate (); | |
| 1355 set_usage_bits (group, offset, width, expr); | |
| 1356 | |
| 1357 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 0 | 1358 fprintf (dump_file, " processing const base store gid=%d[%d..%d)\n", |
| 1359 group_id, (int)offset, (int)(offset+width)); | |
| 1360 } | |
| 1361 else | |
| 1362 { | |
| 111 | 1363 if (may_be_sp_based_p (XEXP (mem, 0))) |
| 0 | 1364 insn_info->stack_pointer_based = true; |
| 1365 insn_info->contains_cselib_groups = true; | |
| 1366 | |
| 111 | 1367 store_info = cse_store_info_pool.allocate (); |
| 0 | 1368 group_id = -1; |
| 1369 | |
| 111 | 1370 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1371 fprintf (dump_file, " processing cselib store [%d..%d)\n", |
| 1372 (int)offset, (int)(offset+width)); | |
| 1373 } | |
| 1374 | |
| 1375 const_rhs = rhs = NULL_RTX; | |
| 1376 if (GET_CODE (body) == SET | |
| 1377 /* No place to keep the value after ra. */ | |
| 1378 && !reload_completed | |
| 1379 && (REG_P (SET_SRC (body)) | |
| 1380 || GET_CODE (SET_SRC (body)) == SUBREG | |
| 1381 || CONSTANT_P (SET_SRC (body))) | |
| 1382 && !MEM_VOLATILE_P (mem) | |
| 1383 /* Sometimes the store and reload is used for truncation and | |
| 1384 rounding. */ | |
| 1385 && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) | |
| 1386 { | |
| 1387 rhs = SET_SRC (body); | |
| 1388 if (CONSTANT_P (rhs)) | |
| 1389 const_rhs = rhs; | |
| 1390 else if (body == PATTERN (insn_info->insn)) | |
| 1391 { | |
| 1392 rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); | |
| 1393 if (tem && CONSTANT_P (XEXP (tem, 0))) | |
| 1394 const_rhs = XEXP (tem, 0); | |
| 1395 } | |
| 1396 if (const_rhs == NULL_RTX && REG_P (rhs)) | |
| 1397 { | |
| 1398 rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); | |
| 1399 | |
| 1400 if (tem && CONSTANT_P (tem)) | |
| 1401 const_rhs = tem; | |
| 1402 } | |
| 1403 } | |
| 1404 | |
| 1405 /* Check to see if this stores causes some other stores to be | |
| 1406 dead. */ | |
| 1407 ptr = active_local_stores; | |
| 1408 last = NULL; | |
| 1409 redundant_reason = NULL; | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1410 mem = canon_rtx (mem); |
| 111 | 1411 |
| 1412 if (group_id < 0) | |
| 1413 mem_addr = base->val_rtx; | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1414 else |
|
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1415 { |
| 111 | 1416 group_info *group = rtx_group_vec[group_id]; |
| 1417 mem_addr = group->canon_base_addr; | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1418 } |
| 111 | 1419 if (offset) |
| 1420 mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); | |
| 0 | 1421 |
| 1422 while (ptr) | |
| 1423 { | |
| 1424 insn_info_t next = ptr->next_local_store; | |
| 111 | 1425 struct store_info *s_info = ptr->store_rec; |
| 0 | 1426 bool del = true; |
| 1427 | |
| 1428 /* Skip the clobbers. We delete the active insn if this insn | |
| 1429 shadows the set. To have been put on the active list, it | |
| 1430 has exactly on set. */ | |
| 1431 while (!s_info->is_set) | |
| 1432 s_info = s_info->next; | |
| 1433 | |
| 111 | 1434 if (s_info->group_id == group_id && s_info->cse_base == base) |
| 0 | 1435 { |
| 1436 HOST_WIDE_INT i; | |
| 111 | 1437 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1438 fprintf (dump_file, " trying store in insn=%d gid=%d[%d..%d)\n", |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1439 INSN_UID (ptr->insn), s_info->group_id, |
| 0 | 1440 (int)s_info->begin, (int)s_info->end); |
| 1441 | |
| 1442 /* Even if PTR won't be eliminated as unneeded, if both | |
| 1443 PTR and this insn store the same constant value, we might | |
| 1444 eliminate this insn instead. */ | |
| 1445 if (s_info->const_rhs | |
| 1446 && const_rhs | |
| 1447 && offset >= s_info->begin | |
| 1448 && offset + width <= s_info->end | |
| 1449 && all_positions_needed_p (s_info, offset - s_info->begin, | |
| 1450 width)) | |
| 1451 { | |
| 1452 if (GET_MODE (mem) == BLKmode) | |
| 1453 { | |
| 1454 if (GET_MODE (s_info->mem) == BLKmode | |
| 1455 && s_info->const_rhs == const_rhs) | |
| 1456 redundant_reason = ptr; | |
| 1457 } | |
| 1458 else if (s_info->const_rhs == const0_rtx | |
| 1459 && const_rhs == const0_rtx) | |
| 1460 redundant_reason = ptr; | |
| 1461 else | |
| 1462 { | |
| 1463 rtx val; | |
| 1464 start_sequence (); | |
| 1465 val = get_stored_val (s_info, GET_MODE (mem), | |
| 1466 offset, offset + width, | |
| 1467 BLOCK_FOR_INSN (insn_info->insn), | |
| 1468 true); | |
| 1469 if (get_insns () != NULL) | |
| 1470 val = NULL_RTX; | |
| 1471 end_sequence (); | |
| 1472 if (val && rtx_equal_p (val, const_rhs)) | |
| 1473 redundant_reason = ptr; | |
| 1474 } | |
| 1475 } | |
| 1476 | |
| 1477 for (i = MAX (offset, s_info->begin); | |
| 1478 i < offset + width && i < s_info->end; | |
| 1479 i++) | |
| 1480 set_position_unneeded (s_info, i - s_info->begin); | |
| 1481 } | |
| 1482 else if (s_info->rhs) | |
| 1483 /* Need to see if it is possible for this store to overwrite | |
| 1484 the value of store_info. If it is, set the rhs to NULL to | |
| 1485 keep it from being used to remove a load. */ | |
| 1486 { | |
| 111 | 1487 if (canon_output_dependence (s_info->mem, true, |
| 1488 mem, GET_MODE (mem), | |
| 1489 mem_addr)) | |
| 0 | 1490 { |
| 1491 s_info->rhs = NULL; | |
| 1492 s_info->const_rhs = NULL; | |
| 1493 } | |
| 1494 } | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1495 |
| 0 | 1496 /* An insn can be deleted if every position of every one of |
| 1497 its s_infos is zero. */ | |
| 111 | 1498 if (any_positions_needed_p (s_info)) |
| 0 | 1499 del = false; |
| 1500 | |
| 1501 if (del) | |
| 1502 { | |
| 1503 insn_info_t insn_to_delete = ptr; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1504 |
| 111 | 1505 active_local_stores_len--; |
| 0 | 1506 if (last) |
| 1507 last->next_local_store = ptr->next_local_store; | |
| 1508 else | |
| 1509 active_local_stores = ptr->next_local_store; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1510 |
| 111 | 1511 if (!insn_to_delete->cannot_delete) |
| 1512 delete_dead_store_insn (insn_to_delete); | |
| 0 | 1513 } |
| 1514 else | |
| 1515 last = ptr; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1516 |
| 0 | 1517 ptr = next; |
| 1518 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1519 |
| 0 | 1520 /* Finish filling in the store_info. */ |
| 1521 store_info->next = insn_info->store_rec; | |
| 1522 insn_info->store_rec = store_info; | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1523 store_info->mem = mem; |
|
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1524 store_info->mem_addr = mem_addr; |
| 0 | 1525 store_info->cse_base = base; |
| 1526 if (width > HOST_BITS_PER_WIDE_INT) | |
| 1527 { | |
| 1528 store_info->is_large = true; | |
| 1529 store_info->positions_needed.large.count = 0; | |
| 111 | 1530 store_info->positions_needed.large.bmap = BITMAP_ALLOC (&dse_bitmap_obstack); |
| 0 | 1531 } |
| 1532 else | |
| 1533 { | |
| 1534 store_info->is_large = false; | |
| 1535 store_info->positions_needed.small_bitmask = lowpart_bitmask (width); | |
| 1536 } | |
| 1537 store_info->group_id = group_id; | |
| 1538 store_info->begin = offset; | |
| 1539 store_info->end = offset + width; | |
| 1540 store_info->is_set = GET_CODE (body) == SET; | |
| 1541 store_info->rhs = rhs; | |
| 1542 store_info->const_rhs = const_rhs; | |
| 1543 store_info->redundant_reason = redundant_reason; | |
| 1544 | |
| 1545 /* If this is a clobber, we return 0. We will only be able to | |
| 1546 delete this insn if there is only one store USED store, but we | |
| 1547 can use the clobber to delete other stores earlier. */ | |
| 1548 return store_info->is_set ? 1 : 0; | |
| 1549 } | |
| 1550 | |
| 1551 | |
| 1552 static void | |
| 1553 dump_insn_info (const char * start, insn_info_t insn_info) | |
| 1554 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1555 fprintf (dump_file, "%s insn=%d %s\n", start, |
| 0 | 1556 INSN_UID (insn_info->insn), |
| 1557 insn_info->store_rec ? "has store" : "naked"); | |
| 1558 } | |
| 1559 | |
| 1560 | |
| 1561 /* If the modes are different and the value's source and target do not | |
| 1562 line up, we need to extract the value from lower part of the rhs of | |
| 1563 the store, shift it, and then put it into a form that can be shoved | |
| 1564 into the read_insn. This function generates a right SHIFT of a | |
| 1565 value that is at least ACCESS_SIZE bytes wide of READ_MODE. The | |
| 1566 shift sequence is returned or NULL if we failed to find a | |
| 1567 shift. */ | |
| 1568 | |
| 1569 static rtx | |
| 1570 find_shift_sequence (int access_size, | |
| 111 | 1571 store_info *store_info, |
| 1572 machine_mode read_mode, | |
| 0 | 1573 int shift, bool speed, bool require_cst) |
| 1574 { | |
| 111 | 1575 machine_mode store_mode = GET_MODE (store_info->mem); |
| 1576 scalar_int_mode new_mode; | |
| 0 | 1577 rtx read_reg = NULL; |
| 1578 | |
| 1579 /* Some machines like the x86 have shift insns for each size of | |
| 1580 operand. Other machines like the ppc or the ia-64 may only have | |
| 1581 shift insns that shift values within 32 or 64 bit registers. | |
| 1582 This loop tries to find the smallest shift insn that will right | |
| 1583 justify the value we want to read but is available in one insn on | |
| 1584 the machine. */ | |
| 1585 | |
| 111 | 1586 opt_scalar_int_mode new_mode_iter; |
| 1587 FOR_EACH_MODE_FROM (new_mode_iter, | |
| 1588 smallest_int_mode_for_size (access_size * BITS_PER_UNIT)) | |
| 0 | 1589 { |
| 111 | 1590 rtx target, new_reg, new_lhs; |
| 1591 rtx_insn *shift_seq, *insn; | |
| 0 | 1592 int cost; |
| 1593 | |
| 111 | 1594 new_mode = new_mode_iter.require (); |
| 1595 if (GET_MODE_BITSIZE (new_mode) > BITS_PER_WORD) | |
| 1596 break; | |
| 1597 | |
| 0 | 1598 /* If a constant was stored into memory, try to simplify it here, |
| 1599 otherwise the cost of the shift might preclude this optimization | |
| 1600 e.g. at -Os, even when no actual shift will be needed. */ | |
| 1601 if (store_info->const_rhs) | |
| 1602 { | |
| 1603 unsigned int byte = subreg_lowpart_offset (new_mode, store_mode); | |
| 1604 rtx ret = simplify_subreg (new_mode, store_info->const_rhs, | |
| 1605 store_mode, byte); | |
| 1606 if (ret && CONSTANT_P (ret)) | |
| 1607 { | |
| 1608 ret = simplify_const_binary_operation (LSHIFTRT, new_mode, | |
| 1609 ret, GEN_INT (shift)); | |
| 1610 if (ret && CONSTANT_P (ret)) | |
| 1611 { | |
| 1612 byte = subreg_lowpart_offset (read_mode, new_mode); | |
| 1613 ret = simplify_subreg (read_mode, ret, new_mode, byte); | |
| 1614 if (ret && CONSTANT_P (ret) | |
| 111 | 1615 && (set_src_cost (ret, read_mode, speed) |
| 1616 <= COSTS_N_INSNS (1))) | |
| 0 | 1617 return ret; |
| 1618 } | |
| 1619 } | |
| 1620 } | |
| 1621 | |
| 1622 if (require_cst) | |
| 1623 return NULL_RTX; | |
| 1624 | |
| 1625 /* Try a wider mode if truncating the store mode to NEW_MODE | |
| 1626 requires a real instruction. */ | |
| 1627 if (GET_MODE_BITSIZE (new_mode) < GET_MODE_BITSIZE (store_mode) | |
| 111 | 1628 && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
| 0 | 1629 continue; |
| 1630 | |
| 1631 /* Also try a wider mode if the necessary punning is either not | |
| 1632 desirable or not possible. */ | |
| 1633 if (!CONSTANT_P (store_info->rhs) | |
| 111 | 1634 && !targetm.modes_tieable_p (new_mode, store_mode)) |
| 0 | 1635 continue; |
| 1636 | |
| 1637 new_reg = gen_reg_rtx (new_mode); | |
| 1638 | |
| 1639 start_sequence (); | |
| 1640 | |
| 1641 /* In theory we could also check for an ashr. Ian Taylor knows | |
| 1642 of one dsp where the cost of these two was not the same. But | |
| 1643 this really is a rare case anyway. */ | |
| 1644 target = expand_binop (new_mode, lshr_optab, new_reg, | |
| 1645 GEN_INT (shift), new_reg, 1, OPTAB_DIRECT); | |
| 1646 | |
| 1647 shift_seq = get_insns (); | |
| 1648 end_sequence (); | |
| 1649 | |
| 1650 if (target != new_reg || shift_seq == NULL) | |
| 1651 continue; | |
| 1652 | |
| 1653 cost = 0; | |
| 1654 for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) | |
| 1655 if (INSN_P (insn)) | |
| 111 | 1656 cost += insn_cost (insn, speed); |
| 0 | 1657 |
| 1658 /* The computation up to here is essentially independent | |
| 1659 of the arguments and could be precomputed. It may | |
| 1660 not be worth doing so. We could precompute if | |
| 1661 worthwhile or at least cache the results. The result | |
| 1662 technically depends on both SHIFT and ACCESS_SIZE, | |
| 1663 but in practice the answer will depend only on ACCESS_SIZE. */ | |
| 1664 | |
| 1665 if (cost > COSTS_N_INSNS (1)) | |
| 1666 continue; | |
| 1667 | |
| 1668 new_lhs = extract_low_bits (new_mode, store_mode, | |
| 1669 copy_rtx (store_info->rhs)); | |
| 1670 if (new_lhs == NULL_RTX) | |
| 1671 continue; | |
| 1672 | |
| 1673 /* We found an acceptable shift. Generate a move to | |
| 1674 take the value from the store and put it into the | |
| 1675 shift pseudo, then shift it, then generate another | |
| 1676 move to put in into the target of the read. */ | |
| 1677 emit_move_insn (new_reg, new_lhs); | |
| 1678 emit_insn (shift_seq); | |
| 1679 read_reg = extract_low_bits (read_mode, new_mode, new_reg); | |
| 1680 break; | |
| 1681 } | |
| 1682 | |
| 1683 return read_reg; | |
| 1684 } | |
| 1685 | |
| 1686 | |
| 1687 /* Call back for note_stores to find the hard regs set or clobbered by | |
| 1688 insn. Data is a bitmap of the hardregs set so far. */ | |
| 1689 | |
| 1690 static void | |
| 1691 look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) | |
| 1692 { | |
| 1693 bitmap regs_set = (bitmap) data; | |
| 1694 | |
| 1695 if (REG_P (x) | |
| 111 | 1696 && HARD_REGISTER_P (x)) |
| 1697 bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); | |
| 0 | 1698 } |
| 1699 | |
| 1700 /* Helper function for replace_read and record_store. | |
| 1701 Attempt to return a value stored in STORE_INFO, from READ_BEGIN | |
| 1702 to one before READ_END bytes read in READ_MODE. Return NULL | |
| 1703 if not successful. If REQUIRE_CST is true, return always constant. */ | |
| 1704 | |
| 1705 static rtx | |
| 111 | 1706 get_stored_val (store_info *store_info, machine_mode read_mode, |
| 0 | 1707 HOST_WIDE_INT read_begin, HOST_WIDE_INT read_end, |
| 1708 basic_block bb, bool require_cst) | |
| 1709 { | |
| 111 | 1710 machine_mode store_mode = GET_MODE (store_info->mem); |
| 0 | 1711 int shift; |
| 1712 int access_size; /* In bytes. */ | |
| 1713 rtx read_reg; | |
| 1714 | |
| 1715 /* To get here the read is within the boundaries of the write so | |
| 1716 shift will never be negative. Start out with the shift being in | |
| 1717 bytes. */ | |
| 1718 if (store_mode == BLKmode) | |
| 1719 shift = 0; | |
| 1720 else if (BYTES_BIG_ENDIAN) | |
| 1721 shift = store_info->end - read_end; | |
| 1722 else | |
| 1723 shift = read_begin - store_info->begin; | |
| 1724 | |
| 1725 access_size = shift + GET_MODE_SIZE (read_mode); | |
| 1726 | |
| 1727 /* From now on it is bits. */ | |
| 1728 shift *= BITS_PER_UNIT; | |
| 1729 | |
| 1730 if (shift) | |
| 1731 read_reg = find_shift_sequence (access_size, store_info, read_mode, shift, | |
| 1732 optimize_bb_for_speed_p (bb), | |
| 1733 require_cst); | |
| 1734 else if (store_mode == BLKmode) | |
| 1735 { | |
| 1736 /* The store is a memset (addr, const_val, const_size). */ | |
| 1737 gcc_assert (CONST_INT_P (store_info->rhs)); | |
| 111 | 1738 scalar_int_mode int_store_mode; |
| 1739 if (!int_mode_for_mode (read_mode).exists (&int_store_mode)) | |
| 0 | 1740 read_reg = NULL_RTX; |
| 1741 else if (store_info->rhs == const0_rtx) | |
| 111 | 1742 read_reg = extract_low_bits (read_mode, int_store_mode, const0_rtx); |
| 1743 else if (GET_MODE_BITSIZE (int_store_mode) > HOST_BITS_PER_WIDE_INT | |
| 0 | 1744 || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) |
| 1745 read_reg = NULL_RTX; | |
| 1746 else | |
| 1747 { | |
| 1748 unsigned HOST_WIDE_INT c | |
| 1749 = INTVAL (store_info->rhs) | |
| 111 | 1750 & ((HOST_WIDE_INT_1 << BITS_PER_UNIT) - 1); |
| 0 | 1751 int shift = BITS_PER_UNIT; |
| 1752 while (shift < HOST_BITS_PER_WIDE_INT) | |
| 1753 { | |
| 1754 c |= (c << shift); | |
| 1755 shift <<= 1; | |
| 1756 } | |
| 111 | 1757 read_reg = gen_int_mode (c, int_store_mode); |
| 1758 read_reg = extract_low_bits (read_mode, int_store_mode, read_reg); | |
| 0 | 1759 } |
| 1760 } | |
| 1761 else if (store_info->const_rhs | |
| 1762 && (require_cst | |
| 1763 || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) | |
| 1764 read_reg = extract_low_bits (read_mode, store_mode, | |
| 1765 copy_rtx (store_info->const_rhs)); | |
| 1766 else | |
| 1767 read_reg = extract_low_bits (read_mode, store_mode, | |
| 1768 copy_rtx (store_info->rhs)); | |
| 1769 if (require_cst && read_reg && !CONSTANT_P (read_reg)) | |
| 1770 read_reg = NULL_RTX; | |
| 1771 return read_reg; | |
| 1772 } | |
| 1773 | |
| 1774 /* Take a sequence of: | |
| 1775 A <- r1 | |
| 1776 ... | |
| 1777 ... <- A | |
| 1778 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1779 and change it into |
| 0 | 1780 r2 <- r1 |
| 1781 A <- r1 | |
| 1782 ... | |
| 1783 ... <- r2 | |
| 1784 | |
| 1785 or | |
| 1786 | |
| 1787 r3 <- extract (r1) | |
| 1788 r3 <- r3 >> shift | |
| 1789 r2 <- extract (r3) | |
| 1790 ... <- r2 | |
| 1791 | |
| 1792 or | |
| 1793 | |
| 1794 r2 <- extract (r1) | |
| 1795 ... <- r2 | |
| 1796 | |
| 1797 Depending on the alignment and the mode of the store and | |
| 1798 subsequent load. | |
| 1799 | |
| 1800 | |
| 1801 The STORE_INFO and STORE_INSN are for the store and READ_INFO | |
| 1802 and READ_INSN are for the read. Return true if the replacement | |
| 1803 went ok. */ | |
| 1804 | |
| 1805 static bool | |
| 111 | 1806 replace_read (store_info *store_info, insn_info_t store_insn, |
| 0 | 1807 read_info_t read_info, insn_info_t read_insn, rtx *loc, |
| 1808 bitmap regs_live) | |
| 1809 { | |
| 111 | 1810 machine_mode store_mode = GET_MODE (store_info->mem); |
| 1811 machine_mode read_mode = GET_MODE (read_info->mem); | |
| 1812 rtx_insn *insns, *this_insn; | |
| 1813 rtx read_reg; | |
| 0 | 1814 basic_block bb; |
| 1815 | |
| 1816 if (!dbg_cnt (dse)) | |
| 1817 return false; | |
| 1818 | |
| 1819 /* Create a sequence of instructions to set up the read register. | |
| 1820 This sequence goes immediately before the store and its result | |
| 1821 is read by the load. | |
| 1822 | |
| 1823 We need to keep this in perspective. We are replacing a read | |
| 1824 with a sequence of insns, but the read will almost certainly be | |
| 1825 in cache, so it is not going to be an expensive one. Thus, we | |
| 1826 are not willing to do a multi insn shift or worse a subroutine | |
| 1827 call to get rid of the read. */ | |
| 111 | 1828 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1829 fprintf (dump_file, "trying to replace %smode load in insn %d" |
| 1830 " from %smode store in insn %d\n", | |
| 1831 GET_MODE_NAME (read_mode), INSN_UID (read_insn->insn), | |
| 1832 GET_MODE_NAME (store_mode), INSN_UID (store_insn->insn)); | |
| 1833 start_sequence (); | |
| 1834 bb = BLOCK_FOR_INSN (read_insn->insn); | |
| 1835 read_reg = get_stored_val (store_info, | |
| 1836 read_mode, read_info->begin, read_info->end, | |
| 1837 bb, false); | |
| 1838 if (read_reg == NULL_RTX) | |
| 1839 { | |
| 1840 end_sequence (); | |
| 111 | 1841 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1842 fprintf (dump_file, " -- could not extract bits of stored value\n"); |
| 1843 return false; | |
| 1844 } | |
| 1845 /* Force the value into a new register so that it won't be clobbered | |
| 1846 between the store and the load. */ | |
| 1847 read_reg = copy_to_mode_reg (read_mode, read_reg); | |
| 1848 insns = get_insns (); | |
| 1849 end_sequence (); | |
| 1850 | |
| 1851 if (insns != NULL_RTX) | |
| 1852 { | |
| 1853 /* Now we have to scan the set of new instructions to see if the | |
| 1854 sequence contains and sets of hardregs that happened to be | |
| 1855 live at this point. For instance, this can happen if one of | |
| 1856 the insns sets the CC and the CC happened to be live at that | |
| 1857 point. This does occasionally happen, see PR 37922. */ | |
| 111 | 1858 bitmap regs_set = BITMAP_ALLOC (®_obstack); |
| 0 | 1859 |
| 1860 for (this_insn = insns; this_insn != NULL_RTX; this_insn = NEXT_INSN (this_insn)) | |
| 1861 note_stores (PATTERN (this_insn), look_for_hardregs, regs_set); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1862 |
| 0 | 1863 bitmap_and_into (regs_set, regs_live); |
| 1864 if (!bitmap_empty_p (regs_set)) | |
| 1865 { | |
| 111 | 1866 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1867 { |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1868 fprintf (dump_file, |
| 0 | 1869 "abandoning replacement because sequence clobbers live hardregs:"); |
| 1870 df_print_regset (dump_file, regs_set); | |
| 1871 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1872 |
| 0 | 1873 BITMAP_FREE (regs_set); |
| 1874 return false; | |
| 1875 } | |
| 1876 BITMAP_FREE (regs_set); | |
| 1877 } | |
| 1878 | |
| 1879 if (validate_change (read_insn->insn, loc, read_reg, 0)) | |
| 1880 { | |
| 111 | 1881 deferred_change *change = deferred_change_pool.allocate (); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1882 |
| 0 | 1883 /* Insert this right before the store insn where it will be safe |
| 1884 from later insns that might change it before the read. */ | |
| 1885 emit_insn_before (insns, store_insn->insn); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1886 |
| 0 | 1887 /* And now for the kludge part: cselib croaks if you just |
| 1888 return at this point. There are two reasons for this: | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1889 |
| 0 | 1890 1) Cselib has an idea of how many pseudos there are and |
| 1891 that does not include the new ones we just added. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1892 |
| 0 | 1893 2) Cselib does not know about the move insn we added |
| 1894 above the store_info, and there is no way to tell it | |
| 1895 about it, because it has "moved on". | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1896 |
| 0 | 1897 Problem (1) is fixable with a certain amount of engineering. |
| 1898 Problem (2) is requires starting the bb from scratch. This | |
| 1899 could be expensive. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1900 |
| 0 | 1901 So we are just going to have to lie. The move/extraction |
| 1902 insns are not really an issue, cselib did not see them. But | |
| 1903 the use of the new pseudo read_insn is a real problem because | |
| 1904 cselib has not scanned this insn. The way that we solve this | |
| 1905 problem is that we are just going to put the mem back for now | |
| 1906 and when we are finished with the block, we undo this. We | |
| 1907 keep a table of mems to get rid of. At the end of the basic | |
| 1908 block we can put them back. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1909 |
| 0 | 1910 *loc = read_info->mem; |
| 111 | 1911 change->next = deferred_change_list; |
| 1912 deferred_change_list = change; | |
| 1913 change->loc = loc; | |
| 1914 change->reg = read_reg; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1915 |
| 0 | 1916 /* Get rid of the read_info, from the point of view of the |
| 1917 rest of dse, play like this read never happened. */ | |
| 1918 read_insn->read_rec = read_info->next; | |
| 111 | 1919 read_info_type_pool.remove (read_info); |
| 1920 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 0 | 1921 { |
| 1922 fprintf (dump_file, " -- replaced the loaded MEM with "); | |
| 1923 print_simple_rtl (dump_file, read_reg); | |
| 1924 fprintf (dump_file, "\n"); | |
| 1925 } | |
| 1926 return true; | |
| 1927 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1928 else |
| 0 | 1929 { |
| 111 | 1930 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1931 { |
| 1932 fprintf (dump_file, " -- replacing the loaded MEM with "); | |
| 1933 print_simple_rtl (dump_file, read_reg); | |
| 1934 fprintf (dump_file, " led to an invalid instruction\n"); | |
| 1935 } | |
| 1936 return false; | |
| 1937 } | |
| 1938 } | |
| 1939 | |
| 111 | 1940 /* Check the address of MEM *LOC and kill any appropriate stores that may |
| 1941 be active. */ | |
| 1942 | |
| 1943 static void | |
| 1944 check_mem_read_rtx (rtx *loc, bb_info_t bb_info) | |
| 0 | 1945 { |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1946 rtx mem = *loc, mem_addr; |
| 0 | 1947 insn_info_t insn_info; |
| 1948 HOST_WIDE_INT offset = 0; | |
| 1949 HOST_WIDE_INT width = 0; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
1950 cselib_val *base = NULL; |
| 0 | 1951 int group_id; |
| 1952 read_info_t read_info; | |
| 1953 | |
| 1954 insn_info = bb_info->last_insn; | |
| 1955 | |
| 1956 if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) | |
| 1957 || (MEM_VOLATILE_P (mem))) | |
| 1958 { | |
| 111 | 1959 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1960 fprintf (dump_file, " adding wild read, volatile or barrier.\n"); |
| 1961 add_wild_read (bb_info); | |
| 1962 insn_info->cannot_delete = true; | |
| 111 | 1963 return; |
| 0 | 1964 } |
| 1965 | |
| 1966 /* If it is reading readonly mem, then there can be no conflict with | |
| 1967 another write. */ | |
| 1968 if (MEM_READONLY_P (mem)) | |
| 111 | 1969 return; |
| 1970 | |
| 1971 if (!canon_address (mem, &group_id, &offset, &base)) | |
| 0 | 1972 { |
| 111 | 1973 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 1974 fprintf (dump_file, " adding wild read, canon_address failure.\n"); |
| 1975 add_wild_read (bb_info); | |
| 111 | 1976 return; |
| 0 | 1977 } |
| 1978 | |
| 1979 if (GET_MODE (mem) == BLKmode) | |
| 1980 width = -1; | |
| 1981 else | |
| 1982 width = GET_MODE_SIZE (GET_MODE (mem)); | |
| 1983 | |
| 111 | 1984 read_info = read_info_type_pool.allocate (); |
| 0 | 1985 read_info->group_id = group_id; |
| 1986 read_info->mem = mem; | |
| 1987 read_info->begin = offset; | |
| 1988 read_info->end = offset + width; | |
| 1989 read_info->next = insn_info->read_rec; | |
| 1990 insn_info->read_rec = read_info; | |
| 111 | 1991 if (group_id < 0) |
| 1992 mem_addr = base->val_rtx; | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1993 else |
|
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1994 { |
| 111 | 1995 group_info *group = rtx_group_vec[group_id]; |
| 1996 mem_addr = group->canon_base_addr; | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
1997 } |
| 111 | 1998 if (offset) |
| 1999 mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); | |
| 2000 | |
| 2001 if (group_id >= 0) | |
| 0 | 2002 { |
| 2003 /* This is the restricted case where the base is a constant or | |
| 2004 the frame pointer and offset is a constant. */ | |
| 2005 insn_info_t i_ptr = active_local_stores; | |
| 2006 insn_info_t last = NULL; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2007 |
| 111 | 2008 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 2009 { |
| 2010 if (width == -1) | |
| 2011 fprintf (dump_file, " processing const load gid=%d[BLK]\n", | |
| 2012 group_id); | |
| 2013 else | |
| 2014 fprintf (dump_file, " processing const load gid=%d[%d..%d)\n", | |
| 2015 group_id, (int)offset, (int)(offset+width)); | |
| 2016 } | |
| 2017 | |
| 2018 while (i_ptr) | |
| 2019 { | |
| 2020 bool remove = false; | |
| 111 | 2021 store_info *store_info = i_ptr->store_rec; |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2022 |
| 0 | 2023 /* Skip the clobbers. */ |
| 2024 while (!store_info->is_set) | |
| 2025 store_info = store_info->next; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2026 |
| 0 | 2027 /* There are three cases here. */ |
| 2028 if (store_info->group_id < 0) | |
| 2029 /* We have a cselib store followed by a read from a | |
| 2030 const base. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2031 remove |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2032 = canon_true_dependence (store_info->mem, |
| 0 | 2033 GET_MODE (store_info->mem), |
| 2034 store_info->mem_addr, | |
| 111 | 2035 mem, mem_addr); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2036 |
| 0 | 2037 else if (group_id == store_info->group_id) |
| 2038 { | |
| 2039 /* This is a block mode load. We may get lucky and | |
| 2040 canon_true_dependence may save the day. */ | |
| 2041 if (width == -1) | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2042 remove |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2043 = canon_true_dependence (store_info->mem, |
| 0 | 2044 GET_MODE (store_info->mem), |
| 2045 store_info->mem_addr, | |
| 111 | 2046 mem, mem_addr); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2047 |
| 0 | 2048 /* If this read is just reading back something that we just |
| 2049 stored, rewrite the read. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2050 else |
| 0 | 2051 { |
| 2052 if (store_info->rhs | |
| 2053 && offset >= store_info->begin | |
| 2054 && offset + width <= store_info->end | |
| 2055 && all_positions_needed_p (store_info, | |
| 2056 offset - store_info->begin, | |
| 2057 width) | |
| 2058 && replace_read (store_info, i_ptr, read_info, | |
| 2059 insn_info, loc, bb_info->regs_live)) | |
| 111 | 2060 return; |
| 0 | 2061 |
| 2062 /* The bases are the same, just see if the offsets | |
| 2063 overlap. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2064 if ((offset < store_info->end) |
| 0 | 2065 && (offset + width > store_info->begin)) |
| 2066 remove = true; | |
| 2067 } | |
| 2068 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2069 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2070 /* else |
| 0 | 2071 The else case that is missing here is that the |
| 2072 bases are constant but different. There is nothing | |
| 2073 to do here because there is no overlap. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2074 |
| 0 | 2075 if (remove) |
| 2076 { | |
| 111 | 2077 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 2078 dump_insn_info ("removing from active", i_ptr); |
| 2079 | |
| 111 | 2080 active_local_stores_len--; |
| 0 | 2081 if (last) |
| 2082 last->next_local_store = i_ptr->next_local_store; | |
| 2083 else | |
| 2084 active_local_stores = i_ptr->next_local_store; | |
| 2085 } | |
| 2086 else | |
| 2087 last = i_ptr; | |
| 2088 i_ptr = i_ptr->next_local_store; | |
| 2089 } | |
| 2090 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2091 else |
| 0 | 2092 { |
| 2093 insn_info_t i_ptr = active_local_stores; | |
| 2094 insn_info_t last = NULL; | |
| 111 | 2095 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 2096 { |
| 2097 fprintf (dump_file, " processing cselib load mem:"); | |
| 2098 print_inline_rtx (dump_file, mem, 0); | |
| 2099 fprintf (dump_file, "\n"); | |
| 2100 } | |
| 2101 | |
| 2102 while (i_ptr) | |
| 2103 { | |
| 2104 bool remove = false; | |
| 111 | 2105 store_info *store_info = i_ptr->store_rec; |
| 2106 | |
| 2107 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 0 | 2108 fprintf (dump_file, " processing cselib load against insn %d\n", |
| 2109 INSN_UID (i_ptr->insn)); | |
| 2110 | |
| 2111 /* Skip the clobbers. */ | |
| 2112 while (!store_info->is_set) | |
| 2113 store_info = store_info->next; | |
| 2114 | |
| 2115 /* If this read is just reading back something that we just | |
| 2116 stored, rewrite the read. */ | |
| 2117 if (store_info->rhs | |
| 2118 && store_info->group_id == -1 | |
| 2119 && store_info->cse_base == base | |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
2120 && width != -1 |
| 0 | 2121 && offset >= store_info->begin |
| 2122 && offset + width <= store_info->end | |
| 2123 && all_positions_needed_p (store_info, | |
| 2124 offset - store_info->begin, width) | |
| 2125 && replace_read (store_info, i_ptr, read_info, insn_info, loc, | |
| 2126 bb_info->regs_live)) | |
| 111 | 2127 return; |
| 2128 | |
| 2129 remove = canon_true_dependence (store_info->mem, | |
| 2130 GET_MODE (store_info->mem), | |
| 2131 store_info->mem_addr, | |
| 2132 mem, mem_addr); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2133 |
| 0 | 2134 if (remove) |
| 2135 { | |
| 111 | 2136 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 2137 dump_insn_info ("removing from active", i_ptr); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2138 |
| 111 | 2139 active_local_stores_len--; |
| 0 | 2140 if (last) |
| 2141 last->next_local_store = i_ptr->next_local_store; | |
| 2142 else | |
| 2143 active_local_stores = i_ptr->next_local_store; | |
| 2144 } | |
| 2145 else | |
| 2146 last = i_ptr; | |
| 2147 i_ptr = i_ptr->next_local_store; | |
| 2148 } | |
| 2149 } | |
| 2150 } | |
| 2151 | |
| 111 | 2152 /* A note_uses callback in which DATA points the INSN_INFO for |
| 0 | 2153 as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
| 2154 true for any part of *LOC. */ | |
| 2155 | |
| 2156 static void | |
| 2157 check_mem_read_use (rtx *loc, void *data) | |
| 2158 { | |
| 111 | 2159 subrtx_ptr_iterator::array_type array; |
| 2160 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) | |
| 2161 { | |
| 2162 rtx *loc = *iter; | |
| 2163 if (MEM_P (*loc)) | |
| 2164 check_mem_read_rtx (loc, (bb_info_t) data); | |
| 2165 } | |
| 0 | 2166 } |
| 2167 | |
| 2168 | |
| 2169 /* Get arguments passed to CALL_INSN. Return TRUE if successful. | |
| 2170 So far it only handles arguments passed in registers. */ | |
| 2171 | |
| 2172 static bool | |
| 2173 get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) | |
| 2174 { | |
| 111 | 2175 CUMULATIVE_ARGS args_so_far_v; |
| 2176 cumulative_args_t args_so_far; | |
| 0 | 2177 tree arg; |
| 2178 int idx; | |
| 2179 | |
| 111 | 2180 INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
| 2181 args_so_far = pack_cumulative_args (&args_so_far_v); | |
| 0 | 2182 |
| 2183 arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); | |
| 2184 for (idx = 0; | |
| 2185 arg != void_list_node && idx < nargs; | |
| 2186 arg = TREE_CHAIN (arg), idx++) | |
| 2187 { | |
| 111 | 2188 scalar_int_mode mode; |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2189 rtx reg, link, tmp; |
| 111 | 2190 |
| 2191 if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg)), &mode)) | |
| 2192 return false; | |
| 2193 | |
| 2194 reg = targetm.calls.function_arg (args_so_far, mode, NULL_TREE, true); | |
| 2195 if (!reg || !REG_P (reg) || GET_MODE (reg) != mode) | |
| 0 | 2196 return false; |
| 2197 | |
| 2198 for (link = CALL_INSN_FUNCTION_USAGE (call_insn); | |
| 2199 link; | |
| 2200 link = XEXP (link, 1)) | |
| 2201 if (GET_CODE (XEXP (link, 0)) == USE) | |
| 2202 { | |
| 111 | 2203 scalar_int_mode arg_mode; |
| 0 | 2204 args[idx] = XEXP (XEXP (link, 0), 0); |
| 2205 if (REG_P (args[idx]) | |
| 2206 && REGNO (args[idx]) == REGNO (reg) | |
| 2207 && (GET_MODE (args[idx]) == mode | |
| 111 | 2208 || (is_int_mode (GET_MODE (args[idx]), &arg_mode) |
| 2209 && (GET_MODE_SIZE (arg_mode) <= UNITS_PER_WORD) | |
| 2210 && (GET_MODE_SIZE (arg_mode) > GET_MODE_SIZE (mode))))) | |
| 0 | 2211 break; |
| 2212 } | |
| 2213 if (!link) | |
| 2214 return false; | |
| 2215 | |
| 2216 tmp = cselib_expand_value_rtx (args[idx], scratch, 5); | |
| 2217 if (GET_MODE (args[idx]) != mode) | |
| 2218 { | |
| 2219 if (!tmp || !CONST_INT_P (tmp)) | |
| 2220 return false; | |
| 111 | 2221 tmp = gen_int_mode (INTVAL (tmp), mode); |
| 0 | 2222 } |
| 2223 if (tmp) | |
| 2224 args[idx] = tmp; | |
| 2225 | |
| 111 | 2226 targetm.calls.function_arg_advance (args_so_far, mode, NULL_TREE, true); |
| 0 | 2227 } |
| 2228 if (arg != void_list_node || idx != nargs) | |
| 2229 return false; | |
| 2230 return true; | |
| 2231 } | |
| 2232 | |
| 111 | 2233 /* Return a bitmap of the fixed registers contained in IN. */ |
| 2234 | |
| 2235 static bitmap | |
| 2236 copy_fixed_regs (const_bitmap in) | |
| 2237 { | |
| 2238 bitmap ret; | |
| 2239 | |
| 2240 ret = ALLOC_REG_SET (NULL); | |
| 2241 bitmap_and (ret, in, fixed_reg_set_regset); | |
| 2242 return ret; | |
| 2243 } | |
| 0 | 2244 |
| 2245 /* Apply record_store to all candidate stores in INSN. Mark INSN | |
| 2246 if some part of it is not a candidate store and assigns to a | |
| 2247 non-register target. */ | |
| 2248 | |
| 2249 static void | |
| 111 | 2250 scan_insn (bb_info_t bb_info, rtx_insn *insn) |
| 0 | 2251 { |
| 2252 rtx body; | |
| 111 | 2253 insn_info_type *insn_info = insn_info_type_pool.allocate (); |
| 0 | 2254 int mems_found = 0; |
| 111 | 2255 memset (insn_info, 0, sizeof (struct insn_info_type)); |
| 2256 | |
| 2257 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 0 | 2258 fprintf (dump_file, "\n**scanning insn=%d\n", |
| 2259 INSN_UID (insn)); | |
| 2260 | |
| 2261 insn_info->prev_insn = bb_info->last_insn; | |
| 2262 insn_info->insn = insn; | |
| 2263 bb_info->last_insn = insn_info; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2264 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2265 if (DEBUG_INSN_P (insn)) |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2266 { |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2267 insn_info->cannot_delete = true; |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2268 return; |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2269 } |
| 0 | 2270 |
| 2271 /* Look at all of the uses in the insn. */ | |
| 2272 note_uses (&PATTERN (insn), check_mem_read_use, bb_info); | |
| 2273 | |
| 2274 if (CALL_P (insn)) | |
| 2275 { | |
| 2276 bool const_call; | |
| 111 | 2277 rtx call, sym; |
| 0 | 2278 tree memset_call = NULL_TREE; |
| 2279 | |
| 2280 insn_info->cannot_delete = true; | |
| 2281 | |
| 2282 /* Const functions cannot do anything bad i.e. read memory, | |
| 2283 however, they can read their parameters which may have | |
| 2284 been pushed onto the stack. | |
| 2285 memset and bzero don't read memory either. */ | |
| 2286 const_call = RTL_CONST_CALL_P (insn); | |
| 111 | 2287 if (!const_call |
| 2288 && (call = get_call_rtx_from (insn)) | |
| 2289 && (sym = XEXP (XEXP (call, 0), 0)) | |
| 2290 && GET_CODE (sym) == SYMBOL_REF | |
| 2291 && SYMBOL_REF_DECL (sym) | |
| 2292 && TREE_CODE (SYMBOL_REF_DECL (sym)) == FUNCTION_DECL | |
| 2293 && DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (sym)) == BUILT_IN_NORMAL | |
| 2294 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (sym)) == BUILT_IN_MEMSET) | |
| 2295 memset_call = SYMBOL_REF_DECL (sym); | |
| 2296 | |
| 0 | 2297 if (const_call || memset_call) |
| 2298 { | |
| 2299 insn_info_t i_ptr = active_local_stores; | |
| 2300 insn_info_t last = NULL; | |
| 2301 | |
| 111 | 2302 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 2303 fprintf (dump_file, "%s call %d\n", |
| 2304 const_call ? "const" : "memset", INSN_UID (insn)); | |
| 2305 | |
| 2306 /* See the head comment of the frame_read field. */ | |
| 111 | 2307 if (reload_completed |
| 2308 /* Tail calls are storing their arguments using | |
| 2309 arg pointer. If it is a frame pointer on the target, | |
| 2310 even before reload we need to kill frame pointer based | |
| 2311 stores. */ | |
| 2312 || (SIBLING_CALL_P (insn) | |
| 2313 && HARD_FRAME_POINTER_IS_ARG_POINTER)) | |
| 0 | 2314 insn_info->frame_read = true; |
| 2315 | |
| 2316 /* Loop over the active stores and remove those which are | |
| 2317 killed by the const function call. */ | |
| 2318 while (i_ptr) | |
| 2319 { | |
| 2320 bool remove_store = false; | |
| 2321 | |
| 2322 /* The stack pointer based stores are always killed. */ | |
| 2323 if (i_ptr->stack_pointer_based) | |
| 2324 remove_store = true; | |
| 2325 | |
| 2326 /* If the frame is read, the frame related stores are killed. */ | |
| 2327 else if (insn_info->frame_read) | |
| 2328 { | |
| 111 | 2329 store_info *store_info = i_ptr->store_rec; |
| 0 | 2330 |
| 2331 /* Skip the clobbers. */ | |
| 2332 while (!store_info->is_set) | |
| 2333 store_info = store_info->next; | |
| 2334 | |
| 2335 if (store_info->group_id >= 0 | |
| 111 | 2336 && rtx_group_vec[store_info->group_id]->frame_related) |
| 0 | 2337 remove_store = true; |
| 2338 } | |
| 2339 | |
| 2340 if (remove_store) | |
| 2341 { | |
| 111 | 2342 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 2343 dump_insn_info ("removing from active", i_ptr); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2344 |
| 111 | 2345 active_local_stores_len--; |
| 0 | 2346 if (last) |
| 2347 last->next_local_store = i_ptr->next_local_store; | |
| 2348 else | |
| 2349 active_local_stores = i_ptr->next_local_store; | |
| 2350 } | |
| 2351 else | |
| 2352 last = i_ptr; | |
| 2353 | |
| 2354 i_ptr = i_ptr->next_local_store; | |
| 2355 } | |
| 2356 | |
| 2357 if (memset_call) | |
| 2358 { | |
| 2359 rtx args[3]; | |
| 2360 if (get_call_args (insn, memset_call, args, 3) | |
| 2361 && CONST_INT_P (args[1]) | |
| 2362 && CONST_INT_P (args[2]) | |
| 2363 && INTVAL (args[2]) > 0) | |
| 2364 { | |
| 2365 rtx mem = gen_rtx_MEM (BLKmode, args[0]); | |
| 111 | 2366 set_mem_size (mem, INTVAL (args[2])); |
| 2367 body = gen_rtx_SET (mem, args[1]); | |
| 0 | 2368 mems_found += record_store (body, bb_info); |
| 111 | 2369 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 2370 fprintf (dump_file, "handling memset as BLKmode store\n"); |
| 2371 if (mems_found == 1) | |
| 2372 { | |
| 111 | 2373 if (active_local_stores_len++ |
| 2374 >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
| 2375 { | |
| 2376 active_local_stores_len = 1; | |
| 2377 active_local_stores = NULL; | |
| 2378 } | |
| 2379 insn_info->fixed_regs_live | |
| 2380 = copy_fixed_regs (bb_info->regs_live); | |
| 0 | 2381 insn_info->next_local_store = active_local_stores; |
| 2382 active_local_stores = insn_info; | |
| 2383 } | |
| 2384 } | |
| 111 | 2385 else |
| 2386 clear_rhs_from_active_local_stores (); | |
| 0 | 2387 } |
| 2388 } | |
| 111 | 2389 else if (SIBLING_CALL_P (insn) && reload_completed) |
| 2390 /* Arguments for a sibling call that are pushed to memory are passed | |
| 2391 using the incoming argument pointer of the current function. After | |
| 2392 reload that might be (and likely is) frame pointer based. */ | |
| 2393 add_wild_read (bb_info); | |
| 0 | 2394 else |
| 111 | 2395 /* Every other call, including pure functions, may read any memory |
| 2396 that is not relative to the frame. */ | |
| 2397 add_non_frame_wild_read (bb_info); | |
| 0 | 2398 |
| 2399 return; | |
| 2400 } | |
| 2401 | |
| 2402 /* Assuming that there are sets in these insns, we cannot delete | |
| 2403 them. */ | |
| 2404 if ((GET_CODE (PATTERN (insn)) == CLOBBER) | |
| 2405 || volatile_refs_p (PATTERN (insn)) | |
| 111 | 2406 || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
| 0 | 2407 || (RTX_FRAME_RELATED_P (insn)) |
| 2408 || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) | |
| 2409 insn_info->cannot_delete = true; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2410 |
| 0 | 2411 body = PATTERN (insn); |
| 2412 if (GET_CODE (body) == PARALLEL) | |
| 2413 { | |
| 2414 int i; | |
| 2415 for (i = 0; i < XVECLEN (body, 0); i++) | |
| 2416 mems_found += record_store (XVECEXP (body, 0, i), bb_info); | |
| 2417 } | |
| 2418 else | |
| 2419 mems_found += record_store (body, bb_info); | |
| 2420 | |
| 111 | 2421 if (dump_file && (dump_flags & TDF_DETAILS)) |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2422 fprintf (dump_file, "mems_found = %d, cannot_delete = %s\n", |
| 0 | 2423 mems_found, insn_info->cannot_delete ? "true" : "false"); |
| 2424 | |
| 2425 /* If we found some sets of mems, add it into the active_local_stores so | |
| 2426 that it can be locally deleted if found dead or used for | |
| 2427 replace_read and redundant constant store elimination. Otherwise mark | |
| 2428 it as cannot delete. This simplifies the processing later. */ | |
| 2429 if (mems_found == 1) | |
| 2430 { | |
| 111 | 2431 if (active_local_stores_len++ |
| 2432 >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES)) | |
| 2433 { | |
| 2434 active_local_stores_len = 1; | |
| 2435 active_local_stores = NULL; | |
| 2436 } | |
| 2437 insn_info->fixed_regs_live = copy_fixed_regs (bb_info->regs_live); | |
| 0 | 2438 insn_info->next_local_store = active_local_stores; |
| 2439 active_local_stores = insn_info; | |
| 2440 } | |
| 2441 else | |
| 2442 insn_info->cannot_delete = true; | |
| 2443 } | |
| 2444 | |
| 2445 | |
| 2446 /* Remove BASE from the set of active_local_stores. This is a | |
| 2447 callback from cselib that is used to get rid of the stores in | |
| 2448 active_local_stores. */ | |
| 2449 | |
| 2450 static void | |
| 2451 remove_useless_values (cselib_val *base) | |
| 2452 { | |
| 2453 insn_info_t insn_info = active_local_stores; | |
| 2454 insn_info_t last = NULL; | |
| 2455 | |
| 2456 while (insn_info) | |
| 2457 { | |
| 111 | 2458 store_info *store_info = insn_info->store_rec; |
| 0 | 2459 bool del = false; |
| 2460 | |
| 2461 /* If ANY of the store_infos match the cselib group that is | |
| 2462 being deleted, then the insn can not be deleted. */ | |
| 2463 while (store_info) | |
| 2464 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2465 if ((store_info->group_id == -1) |
| 0 | 2466 && (store_info->cse_base == base)) |
| 2467 { | |
| 2468 del = true; | |
| 2469 break; | |
| 2470 } | |
| 2471 store_info = store_info->next; | |
| 2472 } | |
| 2473 | |
| 2474 if (del) | |
| 2475 { | |
| 111 | 2476 active_local_stores_len--; |
| 0 | 2477 if (last) |
| 2478 last->next_local_store = insn_info->next_local_store; | |
| 2479 else | |
| 2480 active_local_stores = insn_info->next_local_store; | |
| 2481 free_store_info (insn_info); | |
| 2482 } | |
| 2483 else | |
| 2484 last = insn_info; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2485 |
| 0 | 2486 insn_info = insn_info->next_local_store; |
| 2487 } | |
| 2488 } | |
| 2489 | |
| 2490 | |
| 2491 /* Do all of step 1. */ | |
| 2492 | |
| 2493 static void | |
| 2494 dse_step1 (void) | |
| 2495 { | |
| 2496 basic_block bb; | |
| 111 | 2497 bitmap regs_live = BITMAP_ALLOC (®_obstack); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2498 |
|
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2499 cselib_init (0); |
| 0 | 2500 all_blocks = BITMAP_ALLOC (NULL); |
| 2501 bitmap_set_bit (all_blocks, ENTRY_BLOCK); | |
| 2502 bitmap_set_bit (all_blocks, EXIT_BLOCK); | |
| 2503 | |
| 111 | 2504 FOR_ALL_BB_FN (bb, cfun) |
| 0 | 2505 { |
| 2506 insn_info_t ptr; | |
| 111 | 2507 bb_info_t bb_info = dse_bb_info_type_pool.allocate (); |
| 2508 | |
| 2509 memset (bb_info, 0, sizeof (dse_bb_info_type)); | |
| 0 | 2510 bitmap_set_bit (all_blocks, bb->index); |
| 2511 bb_info->regs_live = regs_live; | |
| 2512 | |
| 2513 bitmap_copy (regs_live, DF_LR_IN (bb)); | |
| 2514 df_simulate_initialize_forwards (bb, regs_live); | |
| 2515 | |
| 2516 bb_table[bb->index] = bb_info; | |
| 2517 cselib_discard_hook = remove_useless_values; | |
| 2518 | |
| 2519 if (bb->index >= NUM_FIXED_BLOCKS) | |
| 2520 { | |
| 111 | 2521 rtx_insn *insn; |
| 2522 | |
| 0 | 2523 active_local_stores = NULL; |
| 111 | 2524 active_local_stores_len = 0; |
| 0 | 2525 cselib_clear_table (); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2526 |
| 0 | 2527 /* Scan the insns. */ |
| 2528 FOR_BB_INSNS (bb, insn) | |
| 2529 { | |
| 2530 if (INSN_P (insn)) | |
| 2531 scan_insn (bb_info, insn); | |
| 2532 cselib_process_insn (insn); | |
| 2533 if (INSN_P (insn)) | |
| 2534 df_simulate_one_insn_forwards (bb, insn, regs_live); | |
| 2535 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2536 |
| 0 | 2537 /* This is something of a hack, because the global algorithm |
| 2538 is supposed to take care of the case where stores go dead | |
| 2539 at the end of the function. However, the global | |
| 2540 algorithm must take a more conservative view of block | |
| 2541 mode reads than the local alg does. So to get the case | |
| 2542 where you have a store to the frame followed by a non | |
| 2543 overlapping block more read, we look at the active local | |
| 2544 stores at the end of the function and delete all of the | |
| 2545 frame and spill based ones. */ | |
| 2546 if (stores_off_frame_dead_at_return | |
| 2547 && (EDGE_COUNT (bb->succs) == 0 | |
| 2548 || (single_succ_p (bb) | |
| 111 | 2549 && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
| 0 | 2550 && ! crtl->calls_eh_return))) |
| 2551 { | |
| 2552 insn_info_t i_ptr = active_local_stores; | |
| 2553 while (i_ptr) | |
| 2554 { | |
| 111 | 2555 store_info *store_info = i_ptr->store_rec; |
| 0 | 2556 |
| 2557 /* Skip the clobbers. */ | |
| 2558 while (!store_info->is_set) | |
| 2559 store_info = store_info->next; | |
| 111 | 2560 if (store_info->group_id >= 0) |
| 2561 { | |
| 2562 group_info *group = rtx_group_vec[store_info->group_id]; | |
| 2563 if (group->frame_related && !i_ptr->cannot_delete) | |
| 2564 delete_dead_store_insn (i_ptr); | |
| 2565 } | |
| 0 | 2566 |
| 2567 i_ptr = i_ptr->next_local_store; | |
| 2568 } | |
| 2569 } | |
| 2570 | |
| 2571 /* Get rid of the loads that were discovered in | |
| 2572 replace_read. Cselib is finished with this block. */ | |
| 2573 while (deferred_change_list) | |
| 2574 { | |
| 111 | 2575 deferred_change *next = deferred_change_list->next; |
| 0 | 2576 |
| 2577 /* There is no reason to validate this change. That was | |
| 2578 done earlier. */ | |
| 2579 *deferred_change_list->loc = deferred_change_list->reg; | |
| 111 | 2580 deferred_change_pool.remove (deferred_change_list); |
| 0 | 2581 deferred_change_list = next; |
| 2582 } | |
| 2583 | |
| 2584 /* Get rid of all of the cselib based store_infos in this | |
| 2585 block and mark the containing insns as not being | |
| 2586 deletable. */ | |
| 2587 ptr = bb_info->last_insn; | |
| 2588 while (ptr) | |
| 2589 { | |
| 2590 if (ptr->contains_cselib_groups) | |
| 2591 { | |
| 111 | 2592 store_info *s_info = ptr->store_rec; |
| 0 | 2593 while (s_info && !s_info->is_set) |
| 2594 s_info = s_info->next; | |
| 2595 if (s_info | |
| 2596 && s_info->redundant_reason | |
| 2597 && s_info->redundant_reason->insn | |
| 2598 && !ptr->cannot_delete) | |
| 2599 { | |
| 111 | 2600 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 2601 fprintf (dump_file, "Locally deleting insn %d " |
| 2602 "because insn %d stores the " | |
| 2603 "same value and couldn't be " | |
| 2604 "eliminated\n", | |
| 2605 INSN_UID (ptr->insn), | |
| 2606 INSN_UID (s_info->redundant_reason->insn)); | |
| 2607 delete_dead_store_insn (ptr); | |
| 2608 } | |
| 2609 free_store_info (ptr); | |
| 2610 } | |
| 2611 else | |
| 2612 { | |
| 111 | 2613 store_info *s_info; |
| 0 | 2614 |
| 2615 /* Free at least positions_needed bitmaps. */ | |
| 2616 for (s_info = ptr->store_rec; s_info; s_info = s_info->next) | |
| 2617 if (s_info->is_large) | |
| 2618 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2619 BITMAP_FREE (s_info->positions_needed.large.bmap); |
| 0 | 2620 s_info->is_large = false; |
| 2621 } | |
| 2622 } | |
| 2623 ptr = ptr->prev_insn; | |
| 2624 } | |
| 2625 | |
| 111 | 2626 cse_store_info_pool.release (); |
| 0 | 2627 } |
| 2628 bb_info->regs_live = NULL; | |
| 2629 } | |
| 2630 | |
| 2631 BITMAP_FREE (regs_live); | |
| 2632 cselib_finish (); | |
| 111 | 2633 rtx_group_table->empty (); |
| 0 | 2634 } |
| 2635 | |
| 2636 | |
| 2637 /*---------------------------------------------------------------------------- | |
| 2638 Second step. | |
| 2639 | |
| 2640 Assign each byte position in the stores that we are going to | |
| 2641 analyze globally to a position in the bitmaps. Returns true if | |
| 2642 there are any bit positions assigned. | |
| 2643 ----------------------------------------------------------------------------*/ | |
| 2644 | |
| 2645 static void | |
| 2646 dse_step2_init (void) | |
| 2647 { | |
| 2648 unsigned int i; | |
| 111 | 2649 group_info *group; |
| 2650 | |
| 2651 FOR_EACH_VEC_ELT (rtx_group_vec, i, group) | |
| 0 | 2652 { |
| 2653 /* For all non stack related bases, we only consider a store to | |
| 2654 be deletable if there are two or more stores for that | |
| 2655 position. This is because it takes one store to make the | |
| 2656 other store redundant. However, for the stores that are | |
| 2657 stack related, we consider them if there is only one store | |
| 2658 for the position. We do this because the stack related | |
| 2659 stores can be deleted if their is no read between them and | |
| 2660 the end of the function. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2661 |
| 0 | 2662 To make this work in the current framework, we take the stack |
| 2663 related bases add all of the bits from store1 into store2. | |
| 2664 This has the effect of making the eligible even if there is | |
| 2665 only one store. */ | |
| 2666 | |
| 2667 if (stores_off_frame_dead_at_return && group->frame_related) | |
| 2668 { | |
| 2669 bitmap_ior_into (group->store2_n, group->store1_n); | |
| 2670 bitmap_ior_into (group->store2_p, group->store1_p); | |
| 111 | 2671 if (dump_file && (dump_flags & TDF_DETAILS)) |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2672 fprintf (dump_file, "group %d is frame related ", i); |
| 0 | 2673 } |
| 2674 | |
| 2675 group->offset_map_size_n++; | |
| 111 | 2676 group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
| 2677 group->offset_map_size_n); | |
| 0 | 2678 group->offset_map_size_p++; |
| 111 | 2679 group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
| 2680 group->offset_map_size_p); | |
| 0 | 2681 group->process_globally = false; |
| 111 | 2682 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 2683 { |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2684 fprintf (dump_file, "group %d(%d+%d): ", i, |
| 0 | 2685 (int)bitmap_count_bits (group->store2_n), |
| 2686 (int)bitmap_count_bits (group->store2_p)); | |
| 2687 bitmap_print (dump_file, group->store2_n, "n ", " "); | |
| 2688 bitmap_print (dump_file, group->store2_p, "p ", "\n"); | |
| 2689 } | |
| 2690 } | |
| 2691 } | |
| 2692 | |
| 2693 | |
| 111 | 2694 /* Init the offset tables. */ |
| 0 | 2695 |
| 2696 static bool | |
| 111 | 2697 dse_step2 (void) |
| 0 | 2698 { |
| 2699 unsigned int i; | |
| 111 | 2700 group_info *group; |
| 0 | 2701 /* Position 0 is unused because 0 is used in the maps to mean |
| 2702 unused. */ | |
| 2703 current_position = 1; | |
| 111 | 2704 FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
| 0 | 2705 { |
| 2706 bitmap_iterator bi; | |
| 2707 unsigned int j; | |
| 2708 | |
| 111 | 2709 memset (group->offset_map_n, 0, sizeof (int) * group->offset_map_size_n); |
| 2710 memset (group->offset_map_p, 0, sizeof (int) * group->offset_map_size_p); | |
| 0 | 2711 bitmap_clear (group->group_kill); |
| 2712 | |
| 2713 EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) | |
| 2714 { | |
| 2715 bitmap_set_bit (group->group_kill, current_position); | |
| 111 | 2716 if (bitmap_bit_p (group->escaped_n, j)) |
| 2717 bitmap_set_bit (kill_on_calls, current_position); | |
| 0 | 2718 group->offset_map_n[j] = current_position++; |
| 2719 group->process_globally = true; | |
| 2720 } | |
| 2721 EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) | |
| 2722 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2723 bitmap_set_bit (group->group_kill, current_position); |
| 111 | 2724 if (bitmap_bit_p (group->escaped_p, j)) |
| 2725 bitmap_set_bit (kill_on_calls, current_position); | |
| 0 | 2726 group->offset_map_p[j] = current_position++; |
| 2727 group->process_globally = true; | |
| 2728 } | |
| 2729 } | |
| 2730 return current_position != 1; | |
| 2731 } | |
| 2732 | |
| 2733 | |
| 2734 | |
| 2735 /*---------------------------------------------------------------------------- | |
| 2736 Third step. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2737 |
| 0 | 2738 Build the bit vectors for the transfer functions. |
| 2739 ----------------------------------------------------------------------------*/ | |
| 2740 | |
| 2741 | |
| 2742 /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not | |
| 2743 there, return 0. */ | |
| 2744 | |
| 2745 static int | |
| 111 | 2746 get_bitmap_index (group_info *group_info, HOST_WIDE_INT offset) |
| 0 | 2747 { |
| 2748 if (offset < 0) | |
| 2749 { | |
| 2750 HOST_WIDE_INT offset_p = -offset; | |
| 2751 if (offset_p >= group_info->offset_map_size_n) | |
| 2752 return 0; | |
| 2753 return group_info->offset_map_n[offset_p]; | |
| 2754 } | |
| 2755 else | |
| 2756 { | |
| 2757 if (offset >= group_info->offset_map_size_p) | |
| 2758 return 0; | |
| 2759 return group_info->offset_map_p[offset]; | |
| 2760 } | |
| 2761 } | |
| 2762 | |
| 2763 | |
| 2764 /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL | |
| 2765 may be NULL. */ | |
| 2766 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2767 static void |
| 111 | 2768 scan_stores (store_info *store_info, bitmap gen, bitmap kill) |
| 0 | 2769 { |
| 2770 while (store_info) | |
| 2771 { | |
| 2772 HOST_WIDE_INT i; | |
| 111 | 2773 group_info *group_info |
| 2774 = rtx_group_vec[store_info->group_id]; | |
| 0 | 2775 if (group_info->process_globally) |
| 2776 for (i = store_info->begin; i < store_info->end; i++) | |
| 2777 { | |
| 2778 int index = get_bitmap_index (group_info, i); | |
| 2779 if (index != 0) | |
| 2780 { | |
| 2781 bitmap_set_bit (gen, index); | |
| 2782 if (kill) | |
| 2783 bitmap_clear_bit (kill, index); | |
| 2784 } | |
| 2785 } | |
| 2786 store_info = store_info->next; | |
| 2787 } | |
| 2788 } | |
| 2789 | |
| 2790 | |
| 2791 /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL | |
| 2792 may be NULL. */ | |
| 2793 | |
| 2794 static void | |
| 111 | 2795 scan_reads (insn_info_t insn_info, bitmap gen, bitmap kill) |
| 0 | 2796 { |
| 2797 read_info_t read_info = insn_info->read_rec; | |
| 2798 int i; | |
| 111 | 2799 group_info *group; |
| 0 | 2800 |
| 2801 /* If this insn reads the frame, kill all the frame related stores. */ | |
| 2802 if (insn_info->frame_read) | |
| 2803 { | |
| 111 | 2804 FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
| 0 | 2805 if (group->process_globally && group->frame_related) |
| 2806 { | |
| 2807 if (kill) | |
| 2808 bitmap_ior_into (kill, group->group_kill); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2809 bitmap_and_compl_into (gen, group->group_kill); |
| 0 | 2810 } |
| 2811 } | |
| 111 | 2812 if (insn_info->non_frame_wild_read) |
| 2813 { | |
| 2814 /* Kill all non-frame related stores. Kill all stores of variables that | |
| 2815 escape. */ | |
| 2816 if (kill) | |
| 2817 bitmap_ior_into (kill, kill_on_calls); | |
| 2818 bitmap_and_compl_into (gen, kill_on_calls); | |
| 2819 FOR_EACH_VEC_ELT (rtx_group_vec, i, group) | |
| 2820 if (group->process_globally && !group->frame_related) | |
| 2821 { | |
| 2822 if (kill) | |
| 2823 bitmap_ior_into (kill, group->group_kill); | |
| 2824 bitmap_and_compl_into (gen, group->group_kill); | |
| 2825 } | |
| 2826 } | |
| 0 | 2827 while (read_info) |
| 2828 { | |
| 111 | 2829 FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
| 0 | 2830 { |
| 2831 if (group->process_globally) | |
| 2832 { | |
| 2833 if (i == read_info->group_id) | |
| 2834 { | |
| 2835 if (read_info->begin > read_info->end) | |
| 2836 { | |
| 2837 /* Begin > end for block mode reads. */ | |
| 2838 if (kill) | |
| 2839 bitmap_ior_into (kill, group->group_kill); | |
| 2840 bitmap_and_compl_into (gen, group->group_kill); | |
| 2841 } | |
| 2842 else | |
| 2843 { | |
| 2844 /* The groups are the same, just process the | |
| 2845 offsets. */ | |
| 2846 HOST_WIDE_INT j; | |
| 2847 for (j = read_info->begin; j < read_info->end; j++) | |
| 2848 { | |
| 2849 int index = get_bitmap_index (group, j); | |
| 2850 if (index != 0) | |
| 2851 { | |
| 2852 if (kill) | |
| 2853 bitmap_set_bit (kill, index); | |
| 2854 bitmap_clear_bit (gen, index); | |
| 2855 } | |
| 2856 } | |
| 2857 } | |
| 2858 } | |
| 2859 else | |
| 2860 { | |
| 2861 /* The groups are different, if the alias sets | |
| 2862 conflict, clear the entire group. We only need | |
| 2863 to apply this test if the read_info is a cselib | |
| 2864 read. Anything with a constant base cannot alias | |
| 2865 something else with a different constant | |
| 2866 base. */ | |
| 2867 if ((read_info->group_id < 0) | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2868 && canon_true_dependence (group->base_mem, |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
2869 GET_MODE (group->base_mem), |
|
19
58ad6c70ea60
update gcc from 4.4.0 to 4.4.1.
kent@firefly.cr.ie.u-ryukyu.ac.jp
parents:
0
diff
changeset
|
2870 group->canon_base_addr, |
| 111 | 2871 read_info->mem, NULL_RTX)) |
| 0 | 2872 { |
| 2873 if (kill) | |
| 2874 bitmap_ior_into (kill, group->group_kill); | |
| 2875 bitmap_and_compl_into (gen, group->group_kill); | |
| 2876 } | |
| 2877 } | |
| 2878 } | |
| 2879 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2880 |
| 0 | 2881 read_info = read_info->next; |
| 2882 } | |
| 2883 } | |
| 2884 | |
| 2885 | |
| 2886 /* Return the insn in BB_INFO before the first wild read or if there | |
| 2887 are no wild reads in the block, return the last insn. */ | |
| 2888 | |
| 2889 static insn_info_t | |
| 2890 find_insn_before_first_wild_read (bb_info_t bb_info) | |
| 2891 { | |
| 2892 insn_info_t insn_info = bb_info->last_insn; | |
| 2893 insn_info_t last_wild_read = NULL; | |
| 2894 | |
| 2895 while (insn_info) | |
| 2896 { | |
| 2897 if (insn_info->wild_read) | |
| 2898 { | |
| 2899 last_wild_read = insn_info->prev_insn; | |
| 2900 /* Block starts with wild read. */ | |
| 2901 if (!last_wild_read) | |
| 2902 return NULL; | |
| 2903 } | |
| 2904 | |
| 2905 insn_info = insn_info->prev_insn; | |
| 2906 } | |
| 2907 | |
| 2908 if (last_wild_read) | |
| 2909 return last_wild_read; | |
| 2910 else | |
| 2911 return bb_info->last_insn; | |
| 2912 } | |
| 2913 | |
| 2914 | |
| 2915 /* Scan the insns in BB_INFO starting at PTR and going to the top of | |
| 2916 the block in order to build the gen and kill sets for the block. | |
| 2917 We start at ptr which may be the last insn in the block or may be | |
| 2918 the first insn with a wild read. In the latter case we are able to | |
| 2919 skip the rest of the block because it just does not matter: | |
| 2920 anything that happens is hidden by the wild read. */ | |
| 2921 | |
| 2922 static void | |
| 111 | 2923 dse_step3_scan (basic_block bb) |
| 0 | 2924 { |
| 2925 bb_info_t bb_info = bb_table[bb->index]; | |
| 2926 insn_info_t insn_info; | |
| 2927 | |
| 111 | 2928 insn_info = find_insn_before_first_wild_read (bb_info); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2929 |
| 0 | 2930 /* In the spill case or in the no_spill case if there is no wild |
| 2931 read in the block, we will need a kill set. */ | |
| 2932 if (insn_info == bb_info->last_insn) | |
| 2933 { | |
| 2934 if (bb_info->kill) | |
| 2935 bitmap_clear (bb_info->kill); | |
| 2936 else | |
| 111 | 2937 bb_info->kill = BITMAP_ALLOC (&dse_bitmap_obstack); |
| 0 | 2938 } |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2939 else |
| 0 | 2940 if (bb_info->kill) |
| 2941 BITMAP_FREE (bb_info->kill); | |
| 2942 | |
| 2943 while (insn_info) | |
| 2944 { | |
| 2945 /* There may have been code deleted by the dce pass run before | |
| 2946 this phase. */ | |
| 2947 if (insn_info->insn && INSN_P (insn_info->insn)) | |
| 2948 { | |
| 111 | 2949 scan_stores (insn_info->store_rec, bb_info->gen, bb_info->kill); |
| 2950 scan_reads (insn_info, bb_info->gen, bb_info->kill); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2951 } |
| 0 | 2952 |
| 2953 insn_info = insn_info->prev_insn; | |
| 2954 } | |
| 2955 } | |
| 2956 | |
| 2957 | |
| 2958 /* Set the gen set of the exit block, and also any block with no | |
| 2959 successors that does not have a wild read. */ | |
| 2960 | |
| 2961 static void | |
| 2962 dse_step3_exit_block_scan (bb_info_t bb_info) | |
| 2963 { | |
| 2964 /* The gen set is all 0's for the exit block except for the | |
| 2965 frame_pointer_group. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2966 |
| 0 | 2967 if (stores_off_frame_dead_at_return) |
| 2968 { | |
| 2969 unsigned int i; | |
| 111 | 2970 group_info *group; |
| 2971 | |
| 2972 FOR_EACH_VEC_ELT (rtx_group_vec, i, group) | |
| 0 | 2973 { |
| 2974 if (group->process_globally && group->frame_related) | |
| 2975 bitmap_ior_into (bb_info->gen, group->group_kill); | |
| 2976 } | |
| 2977 } | |
| 2978 } | |
| 2979 | |
| 2980 | |
| 2981 /* Find all of the blocks that are not backwards reachable from the | |
| 2982 exit block or any block with no successors (BB). These are the | |
| 2983 infinite loops or infinite self loops. These blocks will still | |
| 2984 have their bits set in UNREACHABLE_BLOCKS. */ | |
| 2985 | |
| 2986 static void | |
| 2987 mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) | |
| 2988 { | |
| 2989 edge e; | |
| 2990 edge_iterator ei; | |
| 2991 | |
| 111 | 2992 if (bitmap_bit_p (unreachable_blocks, bb->index)) |
| 0 | 2993 { |
| 111 | 2994 bitmap_clear_bit (unreachable_blocks, bb->index); |
| 0 | 2995 FOR_EACH_EDGE (e, ei, bb->preds) |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2996 { |
| 0 | 2997 mark_reachable_blocks (unreachable_blocks, e->src); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
2998 } |
| 0 | 2999 } |
| 3000 } | |
| 3001 | |
| 3002 /* Build the transfer functions for the function. */ | |
| 3003 | |
| 3004 static void | |
| 111 | 3005 dse_step3 () |
| 0 | 3006 { |
| 3007 basic_block bb; | |
| 3008 sbitmap_iterator sbi; | |
| 3009 bitmap all_ones = NULL; | |
| 3010 unsigned int i; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3011 |
| 111 | 3012 auto_sbitmap unreachable_blocks (last_basic_block_for_fn (cfun)); |
| 3013 bitmap_ones (unreachable_blocks); | |
| 3014 | |
| 3015 FOR_ALL_BB_FN (bb, cfun) | |
| 0 | 3016 { |
| 3017 bb_info_t bb_info = bb_table[bb->index]; | |
| 3018 if (bb_info->gen) | |
| 3019 bitmap_clear (bb_info->gen); | |
| 3020 else | |
| 111 | 3021 bb_info->gen = BITMAP_ALLOC (&dse_bitmap_obstack); |
| 0 | 3022 |
| 3023 if (bb->index == ENTRY_BLOCK) | |
| 3024 ; | |
| 3025 else if (bb->index == EXIT_BLOCK) | |
| 3026 dse_step3_exit_block_scan (bb_info); | |
| 3027 else | |
| 111 | 3028 dse_step3_scan (bb); |
| 0 | 3029 if (EDGE_COUNT (bb->succs) == 0) |
| 3030 mark_reachable_blocks (unreachable_blocks, bb); | |
| 3031 | |
| 3032 /* If this is the second time dataflow is run, delete the old | |
| 3033 sets. */ | |
| 3034 if (bb_info->in) | |
| 3035 BITMAP_FREE (bb_info->in); | |
| 3036 if (bb_info->out) | |
| 3037 BITMAP_FREE (bb_info->out); | |
| 3038 } | |
| 3039 | |
| 3040 /* For any block in an infinite loop, we must initialize the out set | |
| 3041 to all ones. This could be expensive, but almost never occurs in | |
| 3042 practice. However, it is common in regression tests. */ | |
| 111 | 3043 EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
| 0 | 3044 { |
| 3045 if (bitmap_bit_p (all_blocks, i)) | |
| 3046 { | |
| 3047 bb_info_t bb_info = bb_table[i]; | |
| 3048 if (!all_ones) | |
| 3049 { | |
| 3050 unsigned int j; | |
| 111 | 3051 group_info *group; |
| 3052 | |
| 3053 all_ones = BITMAP_ALLOC (&dse_bitmap_obstack); | |
| 3054 FOR_EACH_VEC_ELT (rtx_group_vec, j, group) | |
| 0 | 3055 bitmap_ior_into (all_ones, group->group_kill); |
| 3056 } | |
| 3057 if (!bb_info->out) | |
| 3058 { | |
| 111 | 3059 bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
| 0 | 3060 bitmap_copy (bb_info->out, all_ones); |
| 3061 } | |
| 3062 } | |
| 3063 } | |
| 3064 | |
| 3065 if (all_ones) | |
| 3066 BITMAP_FREE (all_ones); | |
| 3067 } | |
| 3068 | |
| 3069 | |
| 3070 | |
| 3071 /*---------------------------------------------------------------------------- | |
| 3072 Fourth step. | |
| 3073 | |
| 3074 Solve the bitvector equations. | |
| 3075 ----------------------------------------------------------------------------*/ | |
| 3076 | |
| 3077 | |
| 3078 /* Confluence function for blocks with no successors. Create an out | |
| 3079 set from the gen set of the exit block. This block logically has | |
| 3080 the exit block as a successor. */ | |
| 3081 | |
| 3082 | |
| 3083 | |
| 3084 static void | |
| 3085 dse_confluence_0 (basic_block bb) | |
| 3086 { | |
| 3087 bb_info_t bb_info = bb_table[bb->index]; | |
| 3088 | |
| 3089 if (bb->index == EXIT_BLOCK) | |
| 3090 return; | |
| 3091 | |
| 3092 if (!bb_info->out) | |
| 3093 { | |
| 111 | 3094 bb_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
| 0 | 3095 bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
| 3096 } | |
| 3097 } | |
| 3098 | |
| 3099 /* Propagate the information from the in set of the dest of E to the | |
| 3100 out set of the src of E. If the various in or out sets are not | |
| 3101 there, that means they are all ones. */ | |
| 3102 | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
3103 static bool |
| 0 | 3104 dse_confluence_n (edge e) |
| 3105 { | |
| 3106 bb_info_t src_info = bb_table[e->src->index]; | |
| 3107 bb_info_t dest_info = bb_table[e->dest->index]; | |
| 3108 | |
| 3109 if (dest_info->in) | |
| 3110 { | |
| 3111 if (src_info->out) | |
| 3112 bitmap_and_into (src_info->out, dest_info->in); | |
| 3113 else | |
| 3114 { | |
| 111 | 3115 src_info->out = BITMAP_ALLOC (&dse_bitmap_obstack); |
| 0 | 3116 bitmap_copy (src_info->out, dest_info->in); |
| 3117 } | |
| 3118 } | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
3119 return true; |
| 0 | 3120 } |
| 3121 | |
| 3122 | |
| 3123 /* Propagate the info from the out to the in set of BB_INDEX's basic | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3124 block. There are three cases: |
| 0 | 3125 |
| 3126 1) The block has no kill set. In this case the kill set is all | |
| 3127 ones. It does not matter what the out set of the block is, none of | |
| 3128 the info can reach the top. The only thing that reaches the top is | |
| 3129 the gen set and we just copy the set. | |
| 3130 | |
| 3131 2) There is a kill set but no out set and bb has successors. In | |
| 3132 this case we just return. Eventually an out set will be created and | |
| 3133 it is better to wait than to create a set of ones. | |
| 3134 | |
| 3135 3) There is both a kill and out set. We apply the obvious transfer | |
| 3136 function. | |
| 3137 */ | |
| 3138 | |
| 3139 static bool | |
| 3140 dse_transfer_function (int bb_index) | |
| 3141 { | |
| 3142 bb_info_t bb_info = bb_table[bb_index]; | |
| 3143 | |
| 3144 if (bb_info->kill) | |
| 3145 { | |
| 3146 if (bb_info->out) | |
| 3147 { | |
| 3148 /* Case 3 above. */ | |
| 3149 if (bb_info->in) | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3150 return bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
| 0 | 3151 bb_info->out, bb_info->kill); |
| 3152 else | |
| 3153 { | |
| 111 | 3154 bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3155 bitmap_ior_and_compl (bb_info->in, bb_info->gen, |
| 0 | 3156 bb_info->out, bb_info->kill); |
| 3157 return true; | |
| 3158 } | |
| 3159 } | |
| 3160 else | |
| 3161 /* Case 2 above. */ | |
| 3162 return false; | |
| 3163 } | |
| 3164 else | |
| 3165 { | |
| 3166 /* Case 1 above. If there is already an in set, nothing | |
| 3167 happens. */ | |
| 3168 if (bb_info->in) | |
| 3169 return false; | |
| 3170 else | |
| 3171 { | |
| 111 | 3172 bb_info->in = BITMAP_ALLOC (&dse_bitmap_obstack); |
| 0 | 3173 bitmap_copy (bb_info->in, bb_info->gen); |
| 3174 return true; | |
| 3175 } | |
| 3176 } | |
| 3177 } | |
| 3178 | |
| 3179 /* Solve the dataflow equations. */ | |
| 3180 | |
| 3181 static void | |
| 3182 dse_step4 (void) | |
| 3183 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3184 df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3185 dse_confluence_n, dse_transfer_function, |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3186 all_blocks, df_get_postorder (DF_BACKWARD), |
| 0 | 3187 df_get_n_blocks (DF_BACKWARD)); |
| 111 | 3188 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 3189 { |
| 3190 basic_block bb; | |
| 3191 | |
| 3192 fprintf (dump_file, "\n\n*** Global dataflow info after analysis.\n"); | |
| 111 | 3193 FOR_ALL_BB_FN (bb, cfun) |
| 0 | 3194 { |
| 3195 bb_info_t bb_info = bb_table[bb->index]; | |
| 3196 | |
| 3197 df_print_bb_index (bb, dump_file); | |
| 3198 if (bb_info->in) | |
| 3199 bitmap_print (dump_file, bb_info->in, " in: ", "\n"); | |
| 3200 else | |
| 3201 fprintf (dump_file, " in: *MISSING*\n"); | |
| 3202 if (bb_info->gen) | |
| 3203 bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); | |
| 3204 else | |
| 3205 fprintf (dump_file, " gen: *MISSING*\n"); | |
| 3206 if (bb_info->kill) | |
| 3207 bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); | |
| 3208 else | |
| 3209 fprintf (dump_file, " kill: *MISSING*\n"); | |
| 3210 if (bb_info->out) | |
| 3211 bitmap_print (dump_file, bb_info->out, " out: ", "\n"); | |
| 3212 else | |
| 3213 fprintf (dump_file, " out: *MISSING*\n\n"); | |
| 3214 } | |
| 3215 } | |
| 3216 } | |
| 3217 | |
| 3218 | |
| 3219 | |
| 3220 /*---------------------------------------------------------------------------- | |
| 3221 Fifth step. | |
| 3222 | |
| 3223 Delete the stores that can only be deleted using the global information. | |
| 3224 ----------------------------------------------------------------------------*/ | |
| 3225 | |
| 3226 | |
| 3227 static void | |
| 111 | 3228 dse_step5 (void) |
| 0 | 3229 { |
| 3230 basic_block bb; | |
| 111 | 3231 FOR_EACH_BB_FN (bb, cfun) |
| 0 | 3232 { |
| 3233 bb_info_t bb_info = bb_table[bb->index]; | |
| 3234 insn_info_t insn_info = bb_info->last_insn; | |
| 3235 bitmap v = bb_info->out; | |
| 3236 | |
| 3237 while (insn_info) | |
| 3238 { | |
| 3239 bool deleted = false; | |
| 3240 if (dump_file && insn_info->insn) | |
| 3241 { | |
| 3242 fprintf (dump_file, "starting to process insn %d\n", | |
| 3243 INSN_UID (insn_info->insn)); | |
| 3244 bitmap_print (dump_file, v, " v: ", "\n"); | |
| 3245 } | |
| 3246 | |
| 3247 /* There may have been code deleted by the dce pass run before | |
| 3248 this phase. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3249 if (insn_info->insn |
| 0 | 3250 && INSN_P (insn_info->insn) |
| 3251 && (!insn_info->cannot_delete) | |
| 3252 && (!bitmap_empty_p (v))) | |
| 3253 { | |
| 111 | 3254 store_info *store_info = insn_info->store_rec; |
| 0 | 3255 |
| 3256 /* Try to delete the current insn. */ | |
| 3257 deleted = true; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3258 |
| 0 | 3259 /* Skip the clobbers. */ |
| 3260 while (!store_info->is_set) | |
| 3261 store_info = store_info->next; | |
| 3262 | |
| 111 | 3263 HOST_WIDE_INT i; |
| 3264 group_info *group_info = rtx_group_vec[store_info->group_id]; | |
| 3265 | |
| 3266 for (i = store_info->begin; i < store_info->end; i++) | |
| 0 | 3267 { |
| 111 | 3268 int index = get_bitmap_index (group_info, i); |
| 3269 | |
| 3270 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 3271 fprintf (dump_file, "i = %d, index = %d\n", (int)i, index); | |
| 3272 if (index == 0 || !bitmap_bit_p (v, index)) | |
| 0 | 3273 { |
| 111 | 3274 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 3275 fprintf (dump_file, "failing at i = %d\n", (int)i); | |
| 3276 deleted = false; | |
| 3277 break; | |
| 0 | 3278 } |
| 3279 } | |
| 3280 if (deleted) | |
| 3281 { | |
| 111 | 3282 if (dbg_cnt (dse) |
| 3283 && check_for_inc_dec_1 (insn_info)) | |
| 0 | 3284 { |
| 3285 delete_insn (insn_info->insn); | |
| 3286 insn_info->insn = NULL; | |
| 3287 globally_deleted++; | |
| 3288 } | |
| 3289 } | |
| 3290 } | |
| 3291 /* We do want to process the local info if the insn was | |
| 3292 deleted. For instance, if the insn did a wild read, we | |
| 3293 no longer need to trash the info. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3294 if (insn_info->insn |
| 0 | 3295 && INSN_P (insn_info->insn) |
| 3296 && (!deleted)) | |
| 3297 { | |
| 111 | 3298 scan_stores (insn_info->store_rec, v, NULL); |
| 0 | 3299 if (insn_info->wild_read) |
| 3300 { | |
| 111 | 3301 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 3302 fprintf (dump_file, "wild read\n"); |
| 3303 bitmap_clear (v); | |
| 3304 } | |
| 111 | 3305 else if (insn_info->read_rec |
| 3306 || insn_info->non_frame_wild_read | |
| 3307 || insn_info->frame_read) | |
| 0 | 3308 { |
| 111 | 3309 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 3310 { | |
| 3311 if (!insn_info->non_frame_wild_read | |
| 3312 && !insn_info->frame_read) | |
| 3313 fprintf (dump_file, "regular read\n"); | |
| 3314 if (insn_info->non_frame_wild_read) | |
| 3315 fprintf (dump_file, "non-frame wild read\n"); | |
| 3316 if (insn_info->frame_read) | |
| 3317 fprintf (dump_file, "frame read\n"); | |
| 3318 } | |
| 3319 scan_reads (insn_info, v, NULL); | |
| 0 | 3320 } |
| 3321 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3322 |
| 0 | 3323 insn_info = insn_info->prev_insn; |
| 3324 } | |
| 3325 } | |
| 3326 } | |
| 3327 | |
| 3328 | |
| 3329 | |
| 3330 /*---------------------------------------------------------------------------- | |
| 3331 Sixth step. | |
| 3332 | |
| 3333 Delete stores made redundant by earlier stores (which store the same | |
| 3334 value) that couldn't be eliminated. | |
| 3335 ----------------------------------------------------------------------------*/ | |
| 3336 | |
| 3337 static void | |
| 3338 dse_step6 (void) | |
| 3339 { | |
| 3340 basic_block bb; | |
| 3341 | |
| 111 | 3342 FOR_ALL_BB_FN (bb, cfun) |
| 0 | 3343 { |
| 3344 bb_info_t bb_info = bb_table[bb->index]; | |
| 3345 insn_info_t insn_info = bb_info->last_insn; | |
| 3346 | |
| 3347 while (insn_info) | |
| 3348 { | |
| 3349 /* There may have been code deleted by the dce pass run before | |
| 3350 this phase. */ | |
| 3351 if (insn_info->insn | |
| 3352 && INSN_P (insn_info->insn) | |
| 3353 && !insn_info->cannot_delete) | |
| 3354 { | |
| 111 | 3355 store_info *s_info = insn_info->store_rec; |
| 0 | 3356 |
| 3357 while (s_info && !s_info->is_set) | |
| 3358 s_info = s_info->next; | |
| 3359 if (s_info | |
| 3360 && s_info->redundant_reason | |
| 3361 && s_info->redundant_reason->insn | |
| 3362 && INSN_P (s_info->redundant_reason->insn)) | |
| 3363 { | |
| 111 | 3364 rtx_insn *rinsn = s_info->redundant_reason->insn; |
| 3365 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 0 | 3366 fprintf (dump_file, "Locally deleting insn %d " |
| 3367 "because insn %d stores the " | |
| 3368 "same value and couldn't be " | |
| 3369 "eliminated\n", | |
| 3370 INSN_UID (insn_info->insn), | |
| 3371 INSN_UID (rinsn)); | |
| 3372 delete_dead_store_insn (insn_info); | |
| 3373 } | |
| 3374 } | |
| 3375 insn_info = insn_info->prev_insn; | |
| 3376 } | |
| 3377 } | |
| 3378 } | |
| 3379 | |
| 3380 /*---------------------------------------------------------------------------- | |
| 3381 Seventh step. | |
| 3382 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3383 Destroy everything left standing. |
| 0 | 3384 ----------------------------------------------------------------------------*/ |
| 3385 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
36
diff
changeset
|
3386 static void |
| 111 | 3387 dse_step7 (void) |
| 0 | 3388 { |
| 111 | 3389 bitmap_obstack_release (&dse_bitmap_obstack); |
| 3390 obstack_free (&dse_obstack, NULL); | |
| 0 | 3391 |
| 3392 end_alias_analysis (); | |
| 3393 free (bb_table); | |
| 111 | 3394 delete rtx_group_table; |
| 3395 rtx_group_table = NULL; | |
| 3396 rtx_group_vec.release (); | |
| 0 | 3397 BITMAP_FREE (all_blocks); |
| 3398 BITMAP_FREE (scratch); | |
| 3399 | |
| 111 | 3400 rtx_store_info_pool.release (); |
| 3401 read_info_type_pool.release (); | |
| 3402 insn_info_type_pool.release (); | |
| 3403 dse_bb_info_type_pool.release (); | |
| 3404 group_info_pool.release (); | |
| 3405 deferred_change_pool.release (); | |
| 0 | 3406 } |
| 3407 | |
| 3408 | |
| 3409 /* ------------------------------------------------------------------------- | |
| 3410 DSE | |
| 3411 ------------------------------------------------------------------------- */ | |
| 3412 | |
| 3413 /* Callback for running pass_rtl_dse. */ | |
| 3414 | |
| 3415 static unsigned int | |
| 3416 rest_of_handle_dse (void) | |
| 3417 { | |
| 3418 df_set_flags (DF_DEFER_INSN_RESCAN); | |
| 3419 | |
| 3420 /* Need the notes since we must track live hardregs in the forwards | |
| 3421 direction. */ | |
| 3422 df_note_add_problem (); | |
| 3423 df_analyze (); | |
| 3424 | |
| 3425 dse_step0 (); | |
| 3426 dse_step1 (); | |
| 3427 dse_step2_init (); | |
| 111 | 3428 if (dse_step2 ()) |
| 0 | 3429 { |
| 3430 df_set_flags (DF_LR_RUN_DCE); | |
| 3431 df_analyze (); | |
| 111 | 3432 if (dump_file && (dump_flags & TDF_DETAILS)) |
| 0 | 3433 fprintf (dump_file, "doing global processing\n"); |
| 111 | 3434 dse_step3 (); |
| 0 | 3435 dse_step4 (); |
| 111 | 3436 dse_step5 (); |
| 0 | 3437 } |
| 3438 | |
| 3439 dse_step6 (); | |
| 111 | 3440 dse_step7 (); |
| 0 | 3441 |
| 3442 if (dump_file) | |
| 111 | 3443 fprintf (dump_file, "dse: local deletions = %d, global deletions = %d\n", |
| 3444 locally_deleted, globally_deleted); | |
| 3445 | |
| 3446 /* DSE can eliminate potentially-trapping MEMs. | |
| 3447 Remove any EH edges associated with them. */ | |
| 3448 if ((locally_deleted || globally_deleted) | |
| 3449 && cfun->can_throw_non_call_exceptions | |
| 3450 && purge_all_dead_edges ()) | |
| 3451 cleanup_cfg (0); | |
| 3452 | |
| 0 | 3453 return 0; |
| 3454 } | |
| 3455 | |
| 111 | 3456 namespace { |
| 3457 | |
| 3458 const pass_data pass_data_rtl_dse1 = | |
| 0 | 3459 { |
| 111 | 3460 RTL_PASS, /* type */ |
| 3461 "dse1", /* name */ | |
| 3462 OPTGROUP_NONE, /* optinfo_flags */ | |
| 3463 TV_DSE1, /* tv_id */ | |
| 3464 0, /* properties_required */ | |
| 3465 0, /* properties_provided */ | |
| 3466 0, /* properties_destroyed */ | |
| 3467 0, /* todo_flags_start */ | |
| 3468 TODO_df_finish, /* todo_flags_finish */ | |
| 3469 }; | |
| 3470 | |
| 3471 class pass_rtl_dse1 : public rtl_opt_pass | |
| 0 | 3472 { |
| 111 | 3473 public: |
| 3474 pass_rtl_dse1 (gcc::context *ctxt) | |
| 3475 : rtl_opt_pass (pass_data_rtl_dse1, ctxt) | |
| 3476 {} | |
| 3477 | |
| 3478 /* opt_pass methods: */ | |
| 3479 virtual bool gate (function *) | |
| 3480 { | |
| 3481 return optimize > 0 && flag_dse && dbg_cnt (dse1); | |
| 3482 } | |
| 3483 | |
| 3484 virtual unsigned int execute (function *) { return rest_of_handle_dse (); } | |
| 3485 | |
| 3486 }; // class pass_rtl_dse1 | |
| 3487 | |
| 3488 } // anon namespace | |
| 3489 | |
| 3490 rtl_opt_pass * | |
| 3491 make_pass_rtl_dse1 (gcc::context *ctxt) | |
| 0 | 3492 { |
| 111 | 3493 return new pass_rtl_dse1 (ctxt); |
| 3494 } | |
| 3495 | |
| 3496 namespace { | |
| 3497 | |
| 3498 const pass_data pass_data_rtl_dse2 = | |
| 3499 { | |
| 3500 RTL_PASS, /* type */ | |
| 3501 "dse2", /* name */ | |
| 3502 OPTGROUP_NONE, /* optinfo_flags */ | |
| 3503 TV_DSE2, /* tv_id */ | |
| 3504 0, /* properties_required */ | |
| 3505 0, /* properties_provided */ | |
| 3506 0, /* properties_destroyed */ | |
| 3507 0, /* todo_flags_start */ | |
| 3508 TODO_df_finish, /* todo_flags_finish */ | |
| 0 | 3509 }; |
| 3510 | |
| 111 | 3511 class pass_rtl_dse2 : public rtl_opt_pass |
| 0 | 3512 { |
| 111 | 3513 public: |
| 3514 pass_rtl_dse2 (gcc::context *ctxt) | |
| 3515 : rtl_opt_pass (pass_data_rtl_dse2, ctxt) | |
| 3516 {} | |
| 3517 | |
| 3518 /* opt_pass methods: */ | |
| 3519 virtual bool gate (function *) | |
| 3520 { | |
| 3521 return optimize > 0 && flag_dse && dbg_cnt (dse2); | |
| 3522 } | |
| 3523 | |
| 3524 virtual unsigned int execute (function *) { return rest_of_handle_dse (); } | |
| 3525 | |
| 3526 }; // class pass_rtl_dse2 | |
| 3527 | |
| 3528 } // anon namespace | |
| 3529 | |
| 3530 rtl_opt_pass * | |
| 3531 make_pass_rtl_dse2 (gcc::context *ctxt) | |
| 3532 { | |
| 3533 return new pass_rtl_dse2 (ctxt); | |
| 3534 } |