Mercurial > hg > CbC > GCC_original
annotate gcc/tree-predcom.c @ 16:04ced10e8804
gcc 7
| author | kono |
|---|---|
| date | Fri, 27 Oct 2017 22:46:09 +0900 |
| parents | f6334be47118 |
| children | 84e7813d76e9 |
| rev | line source |
|---|---|
| 0 | 1 /* Predictive commoning. |
| 16 | 2 Copyright (C) 2005-2017 Free Software Foundation, Inc. |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3 |
| 0 | 4 This file is part of GCC. |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
5 |
| 0 | 6 GCC is free software; you can redistribute it and/or modify it |
| 7 under the terms of the GNU General Public License as published by the | |
| 8 Free Software Foundation; either version 3, or (at your option) any | |
| 9 later version. | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
10 |
| 0 | 11 GCC is distributed in the hope that it will be useful, but WITHOUT |
| 12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 14 for more details. | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
15 |
| 0 | 16 You should have received a copy of the GNU General Public License |
| 17 along with GCC; see the file COPYING3. If not see | |
| 18 <http://www.gnu.org/licenses/>. */ | |
| 19 | |
| 20 /* This file implements the predictive commoning optimization. Predictive | |
| 21 commoning can be viewed as CSE around a loop, and with some improvements, | |
| 22 as generalized strength reduction-- i.e., reusing values computed in | |
| 23 earlier iterations of a loop in the later ones. So far, the pass only | |
| 24 handles the most useful case, that is, reusing values of memory references. | |
| 25 If you think this is all just a special case of PRE, you are sort of right; | |
| 26 however, concentrating on loops is simpler, and makes it possible to | |
| 27 incorporate data dependence analysis to detect the opportunities, perform | |
| 28 loop unrolling to avoid copies together with renaming immediately, | |
| 29 and if needed, we could also take register pressure into account. | |
| 30 | |
| 31 Let us demonstrate what is done on an example: | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
32 |
| 0 | 33 for (i = 0; i < 100; i++) |
| 34 { | |
| 35 a[i+2] = a[i] + a[i+1]; | |
| 36 b[10] = b[10] + i; | |
| 37 c[i] = c[99 - i]; | |
| 38 d[i] = d[i + 1]; | |
| 39 } | |
| 40 | |
| 41 1) We find data references in the loop, and split them to mutually | |
| 42 independent groups (i.e., we find components of a data dependence | |
| 43 graph). We ignore read-read dependences whose distance is not constant. | |
| 44 (TODO -- we could also ignore antidependences). In this example, we | |
| 45 find the following groups: | |
| 46 | |
| 47 a[i]{read}, a[i+1]{read}, a[i+2]{write} | |
| 48 b[10]{read}, b[10]{write} | |
| 49 c[99 - i]{read}, c[i]{write} | |
| 50 d[i + 1]{read}, d[i]{write} | |
| 51 | |
| 52 2) Inside each of the group, we verify several conditions: | |
| 53 a) all the references must differ in indices only, and the indices | |
| 54 must all have the same step | |
| 55 b) the references must dominate loop latch (and thus, they must be | |
| 56 ordered by dominance relation). | |
| 57 c) the distance of the indices must be a small multiple of the step | |
| 58 We are then able to compute the difference of the references (# of | |
| 59 iterations before they point to the same place as the first of them). | |
| 60 Also, in case there are writes in the loop, we split the groups into | |
| 61 chains whose head is the write whose values are used by the reads in | |
| 62 the same chain. The chains are then processed independently, | |
| 63 making the further transformations simpler. Also, the shorter chains | |
| 64 need the same number of registers, but may require lower unrolling | |
| 65 factor in order to get rid of the copies on the loop latch. | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
66 |
| 0 | 67 In our example, we get the following chains (the chain for c is invalid). |
| 68 | |
| 69 a[i]{read,+0}, a[i+1]{read,-1}, a[i+2]{write,-2} | |
| 70 b[10]{read,+0}, b[10]{write,+0} | |
| 71 d[i + 1]{read,+0}, d[i]{write,+1} | |
| 72 | |
| 73 3) For each read, we determine the read or write whose value it reuses, | |
| 74 together with the distance of this reuse. I.e. we take the last | |
| 75 reference before it with distance 0, or the last of the references | |
| 76 with the smallest positive distance to the read. Then, we remove | |
| 77 the references that are not used in any of these chains, discard the | |
| 78 empty groups, and propagate all the links so that they point to the | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
79 single root reference of the chain (adjusting their distance |
| 0 | 80 appropriately). Some extra care needs to be taken for references with |
| 81 step 0. In our example (the numbers indicate the distance of the | |
| 82 reuse), | |
| 83 | |
| 84 a[i] --> (*) 2, a[i+1] --> (*) 1, a[i+2] (*) | |
| 85 b[10] --> (*) 1, b[10] (*) | |
| 86 | |
| 87 4) The chains are combined together if possible. If the corresponding | |
| 88 elements of two chains are always combined together with the same | |
| 89 operator, we remember just the result of this combination, instead | |
| 90 of remembering the values separately. We may need to perform | |
| 91 reassociation to enable combining, for example | |
| 92 | |
| 93 e[i] + f[i+1] + e[i+1] + f[i] | |
| 94 | |
| 95 can be reassociated as | |
| 96 | |
| 97 (e[i] + f[i]) + (e[i+1] + f[i+1]) | |
| 98 | |
| 99 and we can combine the chains for e and f into one chain. | |
| 100 | |
| 101 5) For each root reference (end of the chain) R, let N be maximum distance | |
| 16 | 102 of a reference reusing its value. Variables R0 up to RN are created, |
| 0 | 103 together with phi nodes that transfer values from R1 .. RN to |
| 104 R0 .. R(N-1). | |
| 105 Initial values are loaded to R0..R(N-1) (in case not all references | |
| 106 must necessarily be accessed and they may trap, we may fail here; | |
| 107 TODO sometimes, the loads could be guarded by a check for the number | |
| 108 of iterations). Values loaded/stored in roots are also copied to | |
| 109 RN. Other reads are replaced with the appropriate variable Ri. | |
| 110 Everything is put to SSA form. | |
| 111 | |
| 112 As a small improvement, if R0 is dead after the root (i.e., all uses of | |
| 113 the value with the maximum distance dominate the root), we can avoid | |
| 114 creating RN and use R0 instead of it. | |
| 115 | |
| 116 In our example, we get (only the parts concerning a and b are shown): | |
| 117 for (i = 0; i < 100; i++) | |
| 118 { | |
| 119 f = phi (a[0], s); | |
| 120 s = phi (a[1], f); | |
| 121 x = phi (b[10], x); | |
| 122 | |
| 123 f = f + s; | |
| 124 a[i+2] = f; | |
| 125 x = x + i; | |
| 126 b[10] = x; | |
| 127 } | |
| 128 | |
| 129 6) Factor F for unrolling is determined as the smallest common multiple of | |
| 130 (N + 1) for each root reference (N for references for that we avoided | |
| 131 creating RN). If F and the loop is small enough, loop is unrolled F | |
| 132 times. The stores to RN (R0) in the copies of the loop body are | |
| 133 periodically replaced with R0, R1, ... (R1, R2, ...), so that they can | |
| 134 be coalesced and the copies can be eliminated. | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
135 |
| 0 | 136 TODO -- copy propagation and other optimizations may change the live |
| 137 ranges of the temporary registers and prevent them from being coalesced; | |
| 138 this may increase the register pressure. | |
| 139 | |
| 140 In our case, F = 2 and the (main loop of the) result is | |
| 141 | |
| 142 for (i = 0; i < ...; i += 2) | |
| 143 { | |
| 144 f = phi (a[0], f); | |
| 145 s = phi (a[1], s); | |
| 146 x = phi (b[10], x); | |
| 147 | |
| 148 f = f + s; | |
| 149 a[i+2] = f; | |
| 150 x = x + i; | |
| 151 b[10] = x; | |
| 152 | |
| 153 s = s + f; | |
| 154 a[i+3] = s; | |
| 155 x = x + i; | |
| 156 b[10] = x; | |
| 157 } | |
| 158 | |
| 16 | 159 Apart from predictive commoning on Load-Load and Store-Load chains, we |
| 160 also support Store-Store chains -- stores killed by other store can be | |
| 161 eliminated. Given below example: | |
| 162 | |
| 163 for (i = 0; i < n; i++) | |
| 164 { | |
| 165 a[i] = 1; | |
| 166 a[i+2] = 2; | |
| 167 } | |
| 168 | |
| 169 It can be replaced with: | |
| 170 | |
| 171 t0 = a[0]; | |
| 172 t1 = a[1]; | |
| 173 for (i = 0; i < n; i++) | |
| 174 { | |
| 175 a[i] = 1; | |
| 176 t2 = 2; | |
| 177 t0 = t1; | |
| 178 t1 = t2; | |
| 179 } | |
| 180 a[n] = t0; | |
| 181 a[n+1] = t1; | |
| 182 | |
| 183 If the loop runs more than 1 iterations, it can be further simplified into: | |
| 184 | |
| 185 for (i = 0; i < n; i++) | |
| 186 { | |
| 187 a[i] = 1; | |
| 188 } | |
| 189 a[n] = 2; | |
| 190 a[n+1] = 2; | |
| 191 | |
| 192 The interesting part is this can be viewed either as general store motion | |
| 193 or general dead store elimination in either intra/inter-iterations way. | |
| 194 | |
| 195 TODO: For now, we don't support store-store chains in multi-exit loops. We | |
| 196 force to not unroll in case of store-store chain even if other chains might | |
| 197 ask for unroll. | |
| 0 | 198 |
| 199 Predictive commoning can be generalized for arbitrary computations (not | |
| 200 just memory loads), and also nontrivial transfer functions (e.g., replacing | |
| 201 i * i with ii_last + 2 * i + 1), to generalize strength reduction. */ | |
| 202 | |
| 203 #include "config.h" | |
| 204 #include "system.h" | |
| 205 #include "coretypes.h" | |
| 16 | 206 #include "backend.h" |
| 207 #include "rtl.h" | |
| 0 | 208 #include "tree.h" |
| 16 | 209 #include "gimple.h" |
| 210 #include "predict.h" | |
| 211 #include "tree-pass.h" | |
| 212 #include "ssa.h" | |
| 213 #include "gimple-pretty-print.h" | |
| 214 #include "alias.h" | |
| 215 #include "fold-const.h" | |
| 0 | 216 #include "cfgloop.h" |
| 16 | 217 #include "tree-eh.h" |
| 218 #include "gimplify.h" | |
| 219 #include "gimple-iterator.h" | |
| 220 #include "gimplify-me.h" | |
| 221 #include "tree-ssa-loop-ivopts.h" | |
| 222 #include "tree-ssa-loop-manip.h" | |
| 223 #include "tree-ssa-loop-niter.h" | |
| 224 #include "tree-ssa-loop.h" | |
| 225 #include "tree-into-ssa.h" | |
| 226 #include "tree-dfa.h" | |
| 227 #include "tree-ssa.h" | |
| 0 | 228 #include "tree-data-ref.h" |
| 229 #include "tree-scalar-evolution.h" | |
| 230 #include "params.h" | |
| 231 #include "tree-affine.h" | |
| 16 | 232 #include "builtins.h" |
| 0 | 233 |
| 234 /* The maximum number of iterations between the considered memory | |
| 235 references. */ | |
| 236 | |
| 237 #define MAX_DISTANCE (target_avail_regs < 16 ? 4 : 8) | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
238 |
| 0 | 239 /* Data references (or phi nodes that carry data reference values across |
| 240 loop iterations). */ | |
| 241 | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
242 typedef struct dref_d |
| 0 | 243 { |
| 244 /* The reference itself. */ | |
| 245 struct data_reference *ref; | |
| 246 | |
| 247 /* The statement in that the reference appears. */ | |
| 16 | 248 gimple *stmt; |
| 0 | 249 |
| 250 /* In case that STMT is a phi node, this field is set to the SSA name | |
| 251 defined by it in replace_phis_by_defined_names (in order to avoid | |
| 252 pointing to phi node that got reallocated in the meantime). */ | |
| 253 tree name_defined_by_phi; | |
| 254 | |
| 255 /* Distance of the reference from the root of the chain (in number of | |
| 256 iterations of the loop). */ | |
| 257 unsigned distance; | |
| 258 | |
| 259 /* Number of iterations offset from the first reference in the component. */ | |
| 16 | 260 widest_int offset; |
| 0 | 261 |
| 262 /* Number of the reference in a component, in dominance ordering. */ | |
| 263 unsigned pos; | |
| 264 | |
| 265 /* True if the memory reference is always accessed when the loop is | |
| 266 entered. */ | |
| 267 unsigned always_accessed : 1; | |
| 268 } *dref; | |
| 269 | |
| 270 | |
| 271 /* Type of the chain of the references. */ | |
| 272 | |
| 273 enum chain_type | |
| 274 { | |
| 275 /* The addresses of the references in the chain are constant. */ | |
| 276 CT_INVARIANT, | |
| 277 | |
| 278 /* There are only loads in the chain. */ | |
| 279 CT_LOAD, | |
| 280 | |
| 281 /* Root of the chain is store, the rest are loads. */ | |
| 282 CT_STORE_LOAD, | |
| 283 | |
| 16 | 284 /* There are only stores in the chain. */ |
| 285 CT_STORE_STORE, | |
| 286 | |
| 0 | 287 /* A combination of two chains. */ |
| 288 CT_COMBINATION | |
| 289 }; | |
| 290 | |
| 291 /* Chains of data references. */ | |
| 292 | |
| 293 typedef struct chain | |
| 294 { | |
| 295 /* Type of the chain. */ | |
| 296 enum chain_type type; | |
| 297 | |
| 298 /* For combination chains, the operator and the two chains that are | |
| 299 combined, and the type of the result. */ | |
| 300 enum tree_code op; | |
| 301 tree rslt_type; | |
| 302 struct chain *ch1, *ch2; | |
| 303 | |
| 304 /* The references in the chain. */ | |
| 16 | 305 vec<dref> refs; |
| 0 | 306 |
| 307 /* The maximum distance of the reference in the chain from the root. */ | |
| 308 unsigned length; | |
| 309 | |
| 310 /* The variables used to copy the value throughout iterations. */ | |
| 16 | 311 vec<tree> vars; |
| 0 | 312 |
| 313 /* Initializers for the variables. */ | |
| 16 | 314 vec<tree> inits; |
| 315 | |
| 316 /* Finalizers for the eliminated stores. */ | |
| 317 vec<tree> finis; | |
| 318 | |
| 319 /* gimple stmts intializing the initial variables of the chain. */ | |
| 320 gimple_seq init_seq; | |
| 321 | |
| 322 /* gimple stmts finalizing the eliminated stores of the chain. */ | |
| 323 gimple_seq fini_seq; | |
| 0 | 324 |
| 325 /* True if there is a use of a variable with the maximal distance | |
| 326 that comes after the root in the loop. */ | |
| 327 unsigned has_max_use_after : 1; | |
| 328 | |
| 329 /* True if all the memory references in the chain are always accessed. */ | |
| 330 unsigned all_always_accessed : 1; | |
| 331 | |
| 332 /* True if this chain was combined together with some other chain. */ | |
| 333 unsigned combined : 1; | |
| 16 | 334 |
| 335 /* True if this is store elimination chain and eliminated stores store | |
| 336 loop invariant value into memory. */ | |
| 337 unsigned inv_store_elimination : 1; | |
| 0 | 338 } *chain_p; |
| 339 | |
| 340 | |
| 341 /* Describes the knowledge about the step of the memory references in | |
| 342 the component. */ | |
| 343 | |
| 344 enum ref_step_type | |
| 345 { | |
| 346 /* The step is zero. */ | |
| 347 RS_INVARIANT, | |
| 348 | |
| 349 /* The step is nonzero. */ | |
| 350 RS_NONZERO, | |
| 351 | |
| 352 /* The step may or may not be nonzero. */ | |
| 353 RS_ANY | |
| 354 }; | |
| 355 | |
| 356 /* Components of the data dependence graph. */ | |
| 357 | |
| 358 struct component | |
| 359 { | |
| 360 /* The references in the component. */ | |
| 16 | 361 vec<dref> refs; |
| 0 | 362 |
| 363 /* What we know about the step of the references in the component. */ | |
| 364 enum ref_step_type comp_step; | |
| 365 | |
| 16 | 366 /* True if all references in component are stores and we try to do |
| 367 intra/inter loop iteration dead store elimination. */ | |
| 368 bool eliminate_store_p; | |
| 369 | |
| 0 | 370 /* Next component in the list. */ |
| 371 struct component *next; | |
| 372 }; | |
| 373 | |
| 374 /* Bitmap of ssa names defined by looparound phi nodes covered by chains. */ | |
| 375 | |
| 376 static bitmap looparound_phis; | |
| 377 | |
| 378 /* Cache used by tree_to_aff_combination_expand. */ | |
| 379 | |
| 16 | 380 static hash_map<tree, name_expansion *> *name_expansions; |
| 0 | 381 |
| 382 /* Dumps data reference REF to FILE. */ | |
| 383 | |
| 384 extern void dump_dref (FILE *, dref); | |
| 385 void | |
| 386 dump_dref (FILE *file, dref ref) | |
| 387 { | |
| 388 if (ref->ref) | |
| 389 { | |
| 390 fprintf (file, " "); | |
| 391 print_generic_expr (file, DR_REF (ref->ref), TDF_SLIM); | |
| 392 fprintf (file, " (id %u%s)\n", ref->pos, | |
| 393 DR_IS_READ (ref->ref) ? "" : ", write"); | |
| 394 | |
| 395 fprintf (file, " offset "); | |
| 16 | 396 print_decs (ref->offset, file); |
| 0 | 397 fprintf (file, "\n"); |
| 398 | |
| 399 fprintf (file, " distance %u\n", ref->distance); | |
| 400 } | |
| 401 else | |
| 402 { | |
| 403 if (gimple_code (ref->stmt) == GIMPLE_PHI) | |
| 404 fprintf (file, " looparound ref\n"); | |
| 405 else | |
| 406 fprintf (file, " combination ref\n"); | |
| 407 fprintf (file, " in statement "); | |
| 408 print_gimple_stmt (file, ref->stmt, 0, TDF_SLIM); | |
| 409 fprintf (file, "\n"); | |
| 410 fprintf (file, " distance %u\n", ref->distance); | |
| 411 } | |
| 412 | |
| 413 } | |
| 414 | |
| 415 /* Dumps CHAIN to FILE. */ | |
| 416 | |
| 417 extern void dump_chain (FILE *, chain_p); | |
| 418 void | |
| 419 dump_chain (FILE *file, chain_p chain) | |
| 420 { | |
| 421 dref a; | |
| 422 const char *chain_type; | |
| 423 unsigned i; | |
| 424 tree var; | |
| 425 | |
| 426 switch (chain->type) | |
| 427 { | |
| 428 case CT_INVARIANT: | |
| 429 chain_type = "Load motion"; | |
| 430 break; | |
| 431 | |
| 432 case CT_LOAD: | |
| 433 chain_type = "Loads-only"; | |
| 434 break; | |
| 435 | |
| 436 case CT_STORE_LOAD: | |
| 437 chain_type = "Store-loads"; | |
| 438 break; | |
| 439 | |
| 16 | 440 case CT_STORE_STORE: |
| 441 chain_type = "Store-stores"; | |
| 442 break; | |
| 443 | |
| 0 | 444 case CT_COMBINATION: |
| 445 chain_type = "Combination"; | |
| 446 break; | |
| 447 | |
| 448 default: | |
| 449 gcc_unreachable (); | |
| 450 } | |
| 451 | |
| 452 fprintf (file, "%s chain %p%s\n", chain_type, (void *) chain, | |
| 453 chain->combined ? " (combined)" : ""); | |
| 454 if (chain->type != CT_INVARIANT) | |
| 455 fprintf (file, " max distance %u%s\n", chain->length, | |
| 456 chain->has_max_use_after ? "" : ", may reuse first"); | |
| 457 | |
| 458 if (chain->type == CT_COMBINATION) | |
| 459 { | |
| 460 fprintf (file, " equal to %p %s %p in type ", | |
| 461 (void *) chain->ch1, op_symbol_code (chain->op), | |
| 462 (void *) chain->ch2); | |
| 463 print_generic_expr (file, chain->rslt_type, TDF_SLIM); | |
| 464 fprintf (file, "\n"); | |
| 465 } | |
| 466 | |
| 16 | 467 if (chain->vars.exists ()) |
| 0 | 468 { |
| 469 fprintf (file, " vars"); | |
| 16 | 470 FOR_EACH_VEC_ELT (chain->vars, i, var) |
| 0 | 471 { |
| 472 fprintf (file, " "); | |
| 473 print_generic_expr (file, var, TDF_SLIM); | |
| 474 } | |
| 475 fprintf (file, "\n"); | |
| 476 } | |
| 477 | |
| 16 | 478 if (chain->inits.exists ()) |
| 0 | 479 { |
| 480 fprintf (file, " inits"); | |
| 16 | 481 FOR_EACH_VEC_ELT (chain->inits, i, var) |
| 0 | 482 { |
| 483 fprintf (file, " "); | |
| 484 print_generic_expr (file, var, TDF_SLIM); | |
| 485 } | |
| 486 fprintf (file, "\n"); | |
| 487 } | |
| 488 | |
| 489 fprintf (file, " references:\n"); | |
| 16 | 490 FOR_EACH_VEC_ELT (chain->refs, i, a) |
| 0 | 491 dump_dref (file, a); |
| 492 | |
| 493 fprintf (file, "\n"); | |
| 494 } | |
| 495 | |
| 496 /* Dumps CHAINS to FILE. */ | |
| 497 | |
| 16 | 498 extern void dump_chains (FILE *, vec<chain_p> ); |
| 0 | 499 void |
| 16 | 500 dump_chains (FILE *file, vec<chain_p> chains) |
| 0 | 501 { |
| 502 chain_p chain; | |
| 503 unsigned i; | |
| 504 | |
| 16 | 505 FOR_EACH_VEC_ELT (chains, i, chain) |
| 0 | 506 dump_chain (file, chain); |
| 507 } | |
| 508 | |
| 509 /* Dumps COMP to FILE. */ | |
| 510 | |
| 511 extern void dump_component (FILE *, struct component *); | |
| 512 void | |
| 513 dump_component (FILE *file, struct component *comp) | |
| 514 { | |
| 515 dref a; | |
| 516 unsigned i; | |
| 517 | |
| 518 fprintf (file, "Component%s:\n", | |
| 519 comp->comp_step == RS_INVARIANT ? " (invariant)" : ""); | |
| 16 | 520 FOR_EACH_VEC_ELT (comp->refs, i, a) |
| 0 | 521 dump_dref (file, a); |
| 522 fprintf (file, "\n"); | |
| 523 } | |
| 524 | |
| 525 /* Dumps COMPS to FILE. */ | |
| 526 | |
| 527 extern void dump_components (FILE *, struct component *); | |
| 528 void | |
| 529 dump_components (FILE *file, struct component *comps) | |
| 530 { | |
| 531 struct component *comp; | |
| 532 | |
| 533 for (comp = comps; comp; comp = comp->next) | |
| 534 dump_component (file, comp); | |
| 535 } | |
| 536 | |
| 537 /* Frees a chain CHAIN. */ | |
| 538 | |
| 539 static void | |
| 540 release_chain (chain_p chain) | |
| 541 { | |
| 542 dref ref; | |
| 543 unsigned i; | |
| 544 | |
| 545 if (chain == NULL) | |
| 546 return; | |
| 547 | |
| 16 | 548 FOR_EACH_VEC_ELT (chain->refs, i, ref) |
| 0 | 549 free (ref); |
| 550 | |
| 16 | 551 chain->refs.release (); |
| 552 chain->vars.release (); | |
| 553 chain->inits.release (); | |
| 554 if (chain->init_seq) | |
| 555 gimple_seq_discard (chain->init_seq); | |
| 556 | |
| 557 chain->finis.release (); | |
| 558 if (chain->fini_seq) | |
| 559 gimple_seq_discard (chain->fini_seq); | |
| 0 | 560 |
| 561 free (chain); | |
| 562 } | |
| 563 | |
| 564 /* Frees CHAINS. */ | |
| 565 | |
| 566 static void | |
| 16 | 567 release_chains (vec<chain_p> chains) |
| 0 | 568 { |
| 569 unsigned i; | |
| 570 chain_p chain; | |
| 571 | |
| 16 | 572 FOR_EACH_VEC_ELT (chains, i, chain) |
| 0 | 573 release_chain (chain); |
| 16 | 574 chains.release (); |
| 0 | 575 } |
| 576 | |
| 577 /* Frees a component COMP. */ | |
| 578 | |
| 579 static void | |
| 580 release_component (struct component *comp) | |
| 581 { | |
| 16 | 582 comp->refs.release (); |
| 0 | 583 free (comp); |
| 584 } | |
| 585 | |
| 586 /* Frees list of components COMPS. */ | |
| 587 | |
| 588 static void | |
| 589 release_components (struct component *comps) | |
| 590 { | |
| 591 struct component *act, *next; | |
| 592 | |
| 593 for (act = comps; act; act = next) | |
| 594 { | |
| 595 next = act->next; | |
| 596 release_component (act); | |
| 597 } | |
| 598 } | |
| 599 | |
| 600 /* Finds a root of tree given by FATHERS containing A, and performs path | |
| 601 shortening. */ | |
| 602 | |
| 603 static unsigned | |
| 604 component_of (unsigned fathers[], unsigned a) | |
| 605 { | |
| 606 unsigned root, n; | |
| 607 | |
| 608 for (root = a; root != fathers[root]; root = fathers[root]) | |
| 609 continue; | |
| 610 | |
| 611 for (; a != root; a = n) | |
| 612 { | |
| 613 n = fathers[a]; | |
| 614 fathers[a] = root; | |
| 615 } | |
| 616 | |
| 617 return root; | |
| 618 } | |
| 619 | |
| 620 /* Join operation for DFU. FATHERS gives the tree, SIZES are sizes of the | |
| 621 components, A and B are components to merge. */ | |
| 622 | |
| 623 static void | |
| 624 merge_comps (unsigned fathers[], unsigned sizes[], unsigned a, unsigned b) | |
| 625 { | |
| 626 unsigned ca = component_of (fathers, a); | |
| 627 unsigned cb = component_of (fathers, b); | |
| 628 | |
| 629 if (ca == cb) | |
| 630 return; | |
| 631 | |
| 632 if (sizes[ca] < sizes[cb]) | |
| 633 { | |
| 634 sizes[cb] += sizes[ca]; | |
| 635 fathers[ca] = cb; | |
| 636 } | |
| 637 else | |
| 638 { | |
| 639 sizes[ca] += sizes[cb]; | |
| 640 fathers[cb] = ca; | |
| 641 } | |
| 642 } | |
| 643 | |
| 644 /* Returns true if A is a reference that is suitable for predictive commoning | |
| 645 in the innermost loop that contains it. REF_STEP is set according to the | |
| 646 step of the reference A. */ | |
| 647 | |
| 648 static bool | |
| 649 suitable_reference_p (struct data_reference *a, enum ref_step_type *ref_step) | |
| 650 { | |
| 651 tree ref = DR_REF (a), step = DR_STEP (a); | |
| 652 | |
| 653 if (!step | |
| 16 | 654 || TREE_THIS_VOLATILE (ref) |
| 0 | 655 || !is_gimple_reg_type (TREE_TYPE (ref)) |
| 656 || tree_could_throw_p (ref)) | |
| 657 return false; | |
| 658 | |
| 659 if (integer_zerop (step)) | |
| 660 *ref_step = RS_INVARIANT; | |
| 661 else if (integer_nonzerop (step)) | |
| 662 *ref_step = RS_NONZERO; | |
| 663 else | |
| 664 *ref_step = RS_ANY; | |
| 665 | |
| 666 return true; | |
| 667 } | |
| 668 | |
| 669 /* Stores DR_OFFSET (DR) + DR_INIT (DR) to OFFSET. */ | |
| 670 | |
| 671 static void | |
| 672 aff_combination_dr_offset (struct data_reference *dr, aff_tree *offset) | |
| 673 { | |
| 16 | 674 tree type = TREE_TYPE (DR_OFFSET (dr)); |
| 0 | 675 aff_tree delta; |
| 676 | |
| 16 | 677 tree_to_aff_combination_expand (DR_OFFSET (dr), type, offset, |
| 0 | 678 &name_expansions); |
| 16 | 679 aff_combination_const (&delta, type, wi::to_widest (DR_INIT (dr))); |
| 0 | 680 aff_combination_add (offset, &delta); |
| 681 } | |
| 682 | |
| 683 /* Determines number of iterations of the innermost enclosing loop before B | |
| 684 refers to exactly the same location as A and stores it to OFF. If A and | |
| 685 B do not have the same step, they never meet, or anything else fails, | |
| 686 returns false, otherwise returns true. Both A and B are assumed to | |
| 687 satisfy suitable_reference_p. */ | |
| 688 | |
| 689 static bool | |
| 690 determine_offset (struct data_reference *a, struct data_reference *b, | |
| 16 | 691 widest_int *off) |
| 0 | 692 { |
| 693 aff_tree diff, baseb, step; | |
| 694 tree typea, typeb; | |
| 695 | |
| 696 /* Check that both the references access the location in the same type. */ | |
| 697 typea = TREE_TYPE (DR_REF (a)); | |
| 698 typeb = TREE_TYPE (DR_REF (b)); | |
| 699 if (!useless_type_conversion_p (typeb, typea)) | |
| 700 return false; | |
| 701 | |
| 702 /* Check whether the base address and the step of both references is the | |
| 703 same. */ | |
| 704 if (!operand_equal_p (DR_STEP (a), DR_STEP (b), 0) | |
| 705 || !operand_equal_p (DR_BASE_ADDRESS (a), DR_BASE_ADDRESS (b), 0)) | |
| 706 return false; | |
| 707 | |
| 708 if (integer_zerop (DR_STEP (a))) | |
| 709 { | |
| 710 /* If the references have loop invariant address, check that they access | |
| 711 exactly the same location. */ | |
| 16 | 712 *off = 0; |
| 0 | 713 return (operand_equal_p (DR_OFFSET (a), DR_OFFSET (b), 0) |
| 714 && operand_equal_p (DR_INIT (a), DR_INIT (b), 0)); | |
| 715 } | |
| 716 | |
| 717 /* Compare the offsets of the addresses, and check whether the difference | |
| 718 is a multiple of step. */ | |
| 719 aff_combination_dr_offset (a, &diff); | |
| 720 aff_combination_dr_offset (b, &baseb); | |
| 16 | 721 aff_combination_scale (&baseb, -1); |
| 0 | 722 aff_combination_add (&diff, &baseb); |
| 723 | |
| 16 | 724 tree_to_aff_combination_expand (DR_STEP (a), TREE_TYPE (DR_STEP (a)), |
| 0 | 725 &step, &name_expansions); |
| 726 return aff_combination_constant_multiple_p (&diff, &step, off); | |
| 727 } | |
| 728 | |
| 729 /* Returns the last basic block in LOOP for that we are sure that | |
| 730 it is executed whenever the loop is entered. */ | |
| 731 | |
| 732 static basic_block | |
| 733 last_always_executed_block (struct loop *loop) | |
| 734 { | |
| 735 unsigned i; | |
| 16 | 736 vec<edge> exits = get_loop_exit_edges (loop); |
| 0 | 737 edge ex; |
| 738 basic_block last = loop->latch; | |
| 739 | |
| 16 | 740 FOR_EACH_VEC_ELT (exits, i, ex) |
| 0 | 741 last = nearest_common_dominator (CDI_DOMINATORS, last, ex->src); |
| 16 | 742 exits.release (); |
| 0 | 743 |
| 744 return last; | |
| 745 } | |
| 746 | |
| 747 /* Splits dependence graph on DATAREFS described by DEPENDS to components. */ | |
| 748 | |
| 749 static struct component * | |
| 750 split_data_refs_to_components (struct loop *loop, | |
| 16 | 751 vec<data_reference_p> datarefs, |
| 752 vec<ddr_p> depends) | |
| 0 | 753 { |
| 16 | 754 unsigned i, n = datarefs.length (); |
| 0 | 755 unsigned ca, ia, ib, bad; |
| 756 unsigned *comp_father = XNEWVEC (unsigned, n + 1); | |
| 757 unsigned *comp_size = XNEWVEC (unsigned, n + 1); | |
| 758 struct component **comps; | |
| 759 struct data_reference *dr, *dra, *drb; | |
| 760 struct data_dependence_relation *ddr; | |
| 761 struct component *comp_list = NULL, *comp; | |
| 762 dref dataref; | |
| 16 | 763 /* Don't do store elimination if loop has multiple exit edges. */ |
| 764 bool eliminate_store_p = single_exit (loop) != NULL; | |
| 0 | 765 basic_block last_always_executed = last_always_executed_block (loop); |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
766 |
| 16 | 767 FOR_EACH_VEC_ELT (datarefs, i, dr) |
| 0 | 768 { |
| 769 if (!DR_REF (dr)) | |
| 770 { | |
| 771 /* A fake reference for call or asm_expr that may clobber memory; | |
| 772 just fail. */ | |
| 773 goto end; | |
| 774 } | |
| 16 | 775 /* predcom pass isn't prepared to handle calls with data references. */ |
| 776 if (is_gimple_call (DR_STMT (dr))) | |
| 777 goto end; | |
| 0 | 778 dr->aux = (void *) (size_t) i; |
| 779 comp_father[i] = i; | |
| 780 comp_size[i] = 1; | |
| 781 } | |
| 782 | |
| 783 /* A component reserved for the "bad" data references. */ | |
| 784 comp_father[n] = n; | |
| 785 comp_size[n] = 1; | |
| 786 | |
| 16 | 787 FOR_EACH_VEC_ELT (datarefs, i, dr) |
| 0 | 788 { |
| 789 enum ref_step_type dummy; | |
| 790 | |
| 791 if (!suitable_reference_p (dr, &dummy)) | |
| 792 { | |
| 793 ia = (unsigned) (size_t) dr->aux; | |
| 794 merge_comps (comp_father, comp_size, n, ia); | |
| 795 } | |
| 796 } | |
| 797 | |
| 16 | 798 FOR_EACH_VEC_ELT (depends, i, ddr) |
| 0 | 799 { |
| 16 | 800 widest_int dummy_off; |
| 0 | 801 |
| 802 if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
| 803 continue; | |
| 804 | |
| 805 dra = DDR_A (ddr); | |
| 806 drb = DDR_B (ddr); | |
| 16 | 807 |
| 808 /* Don't do store elimination if there is any unknown dependence for | |
| 809 any store data reference. */ | |
| 810 if ((DR_IS_WRITE (dra) || DR_IS_WRITE (drb)) | |
| 811 && (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
| 812 || DDR_NUM_DIST_VECTS (ddr) == 0)) | |
| 813 eliminate_store_p = false; | |
| 814 | |
| 0 | 815 ia = component_of (comp_father, (unsigned) (size_t) dra->aux); |
| 816 ib = component_of (comp_father, (unsigned) (size_t) drb->aux); | |
| 817 if (ia == ib) | |
| 818 continue; | |
| 819 | |
| 820 bad = component_of (comp_father, n); | |
| 821 | |
| 822 /* If both A and B are reads, we may ignore unsuitable dependences. */ | |
| 16 | 823 if (DR_IS_READ (dra) && DR_IS_READ (drb)) |
| 824 { | |
| 825 if (ia == bad || ib == bad | |
| 826 || !determine_offset (dra, drb, &dummy_off)) | |
| 827 continue; | |
| 828 } | |
| 829 /* If A is read and B write or vice versa and there is unsuitable | |
| 830 dependence, instead of merging both components into a component | |
| 831 that will certainly not pass suitable_component_p, just put the | |
| 832 read into bad component, perhaps at least the write together with | |
| 833 all the other data refs in it's component will be optimizable. */ | |
| 834 else if (DR_IS_READ (dra) && ib != bad) | |
| 835 { | |
| 836 if (ia == bad) | |
| 837 continue; | |
| 838 else if (!determine_offset (dra, drb, &dummy_off)) | |
| 839 { | |
| 840 merge_comps (comp_father, comp_size, bad, ia); | |
| 841 continue; | |
| 842 } | |
| 843 } | |
| 844 else if (DR_IS_READ (drb) && ia != bad) | |
| 845 { | |
| 846 if (ib == bad) | |
| 847 continue; | |
| 848 else if (!determine_offset (dra, drb, &dummy_off)) | |
| 849 { | |
| 850 merge_comps (comp_father, comp_size, bad, ib); | |
| 851 continue; | |
| 852 } | |
| 853 } | |
| 854 else if (DR_IS_WRITE (dra) && DR_IS_WRITE (drb) | |
| 855 && ia != bad && ib != bad | |
| 856 && !determine_offset (dra, drb, &dummy_off)) | |
| 857 { | |
| 858 merge_comps (comp_father, comp_size, bad, ia); | |
| 859 merge_comps (comp_father, comp_size, bad, ib); | |
| 860 continue; | |
| 861 } | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
862 |
| 0 | 863 merge_comps (comp_father, comp_size, ia, ib); |
| 864 } | |
| 865 | |
| 16 | 866 if (eliminate_store_p) |
| 867 { | |
| 868 tree niters = number_of_latch_executions (loop); | |
| 869 | |
| 870 /* Don't do store elimination if niters info is unknown because stores | |
| 871 in the last iteration can't be eliminated and we need to recover it | |
| 872 after loop. */ | |
| 873 eliminate_store_p = (niters != NULL_TREE && niters != chrec_dont_know); | |
| 874 } | |
| 875 | |
| 0 | 876 comps = XCNEWVEC (struct component *, n); |
| 877 bad = component_of (comp_father, n); | |
| 16 | 878 FOR_EACH_VEC_ELT (datarefs, i, dr) |
| 0 | 879 { |
| 880 ia = (unsigned) (size_t) dr->aux; | |
| 881 ca = component_of (comp_father, ia); | |
| 882 if (ca == bad) | |
| 883 continue; | |
| 884 | |
| 885 comp = comps[ca]; | |
| 886 if (!comp) | |
| 887 { | |
| 888 comp = XCNEW (struct component); | |
| 16 | 889 comp->refs.create (comp_size[ca]); |
| 890 comp->eliminate_store_p = eliminate_store_p; | |
| 0 | 891 comps[ca] = comp; |
| 892 } | |
| 893 | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
894 dataref = XCNEW (struct dref_d); |
| 0 | 895 dataref->ref = dr; |
| 896 dataref->stmt = DR_STMT (dr); | |
| 16 | 897 dataref->offset = 0; |
| 0 | 898 dataref->distance = 0; |
| 899 | |
| 900 dataref->always_accessed | |
| 901 = dominated_by_p (CDI_DOMINATORS, last_always_executed, | |
| 902 gimple_bb (dataref->stmt)); | |
| 16 | 903 dataref->pos = comp->refs.length (); |
| 904 comp->refs.quick_push (dataref); | |
| 905 if (DR_IS_READ (dr)) | |
| 906 comp->eliminate_store_p = false; | |
| 0 | 907 } |
| 908 | |
| 909 for (i = 0; i < n; i++) | |
| 910 { | |
| 911 comp = comps[i]; | |
| 912 if (comp) | |
| 913 { | |
| 914 comp->next = comp_list; | |
| 915 comp_list = comp; | |
| 916 } | |
| 917 } | |
| 918 free (comps); | |
| 919 | |
| 920 end: | |
| 921 free (comp_father); | |
| 922 free (comp_size); | |
| 923 return comp_list; | |
| 924 } | |
| 925 | |
| 926 /* Returns true if the component COMP satisfies the conditions | |
| 927 described in 2) at the beginning of this file. LOOP is the current | |
| 928 loop. */ | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
929 |
| 0 | 930 static bool |
| 931 suitable_component_p (struct loop *loop, struct component *comp) | |
| 932 { | |
| 933 unsigned i; | |
| 934 dref a, first; | |
| 935 basic_block ba, bp = loop->header; | |
| 936 bool ok, has_write = false; | |
| 937 | |
| 16 | 938 FOR_EACH_VEC_ELT (comp->refs, i, a) |
| 0 | 939 { |
| 940 ba = gimple_bb (a->stmt); | |
| 941 | |
| 942 if (!just_once_each_iteration_p (loop, ba)) | |
| 943 return false; | |
| 944 | |
| 945 gcc_assert (dominated_by_p (CDI_DOMINATORS, ba, bp)); | |
| 946 bp = ba; | |
| 947 | |
|
14
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
11
diff
changeset
|
948 if (DR_IS_WRITE (a->ref)) |
| 0 | 949 has_write = true; |
| 950 } | |
| 951 | |
| 16 | 952 first = comp->refs[0]; |
| 0 | 953 ok = suitable_reference_p (first->ref, &comp->comp_step); |
| 954 gcc_assert (ok); | |
| 16 | 955 first->offset = 0; |
| 956 | |
| 957 for (i = 1; comp->refs.iterate (i, &a); i++) | |
| 0 | 958 { |
| 959 if (!determine_offset (first->ref, a->ref, &a->offset)) | |
| 960 return false; | |
| 961 | |
| 16 | 962 enum ref_step_type a_step; |
| 963 gcc_checking_assert (suitable_reference_p (a->ref, &a_step) | |
| 964 && a_step == comp->comp_step); | |
| 0 | 965 } |
| 966 | |
| 967 /* If there is a write inside the component, we must know whether the | |
| 968 step is nonzero or not -- we would not otherwise be able to recognize | |
| 969 whether the value accessed by reads comes from the OFFSET-th iteration | |
| 970 or the previous one. */ | |
| 971 if (has_write && comp->comp_step == RS_ANY) | |
| 972 return false; | |
| 973 | |
| 974 return true; | |
| 975 } | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
976 |
| 0 | 977 /* Check the conditions on references inside each of components COMPS, |
| 978 and remove the unsuitable components from the list. The new list | |
| 979 of components is returned. The conditions are described in 2) at | |
| 980 the beginning of this file. LOOP is the current loop. */ | |
| 981 | |
| 982 static struct component * | |
| 983 filter_suitable_components (struct loop *loop, struct component *comps) | |
| 984 { | |
| 985 struct component **comp, *act; | |
| 986 | |
| 987 for (comp = &comps; *comp; ) | |
| 988 { | |
| 989 act = *comp; | |
| 990 if (suitable_component_p (loop, act)) | |
| 991 comp = &act->next; | |
| 992 else | |
| 993 { | |
| 994 dref ref; | |
| 995 unsigned i; | |
| 996 | |
| 997 *comp = act->next; | |
| 16 | 998 FOR_EACH_VEC_ELT (act->refs, i, ref) |
| 0 | 999 free (ref); |
| 1000 release_component (act); | |
| 1001 } | |
| 1002 } | |
| 1003 | |
| 1004 return comps; | |
| 1005 } | |
| 1006 | |
| 1007 /* Compares two drefs A and B by their offset and position. Callback for | |
| 1008 qsort. */ | |
| 1009 | |
| 1010 static int | |
| 1011 order_drefs (const void *a, const void *b) | |
| 1012 { | |
| 1013 const dref *const da = (const dref *) a; | |
| 1014 const dref *const db = (const dref *) b; | |
| 16 | 1015 int offcmp = wi::cmps ((*da)->offset, (*db)->offset); |
| 0 | 1016 |
| 1017 if (offcmp != 0) | |
| 1018 return offcmp; | |
| 1019 | |
| 1020 return (*da)->pos - (*db)->pos; | |
| 1021 } | |
| 1022 | |
| 1023 /* Returns root of the CHAIN. */ | |
| 1024 | |
| 1025 static inline dref | |
| 1026 get_chain_root (chain_p chain) | |
| 1027 { | |
| 16 | 1028 return chain->refs[0]; |
| 1029 } | |
| 1030 | |
| 1031 /* Given CHAIN, returns the last ref at DISTANCE, or NULL if it doesn't | |
| 1032 exist. */ | |
| 1033 | |
| 1034 static inline dref | |
| 1035 get_chain_last_ref_at (chain_p chain, unsigned distance) | |
| 1036 { | |
| 1037 unsigned i; | |
| 1038 | |
| 1039 for (i = chain->refs.length (); i > 0; i--) | |
| 1040 if (distance == chain->refs[i - 1]->distance) | |
| 1041 break; | |
| 1042 | |
| 1043 return (i > 0) ? chain->refs[i - 1] : NULL; | |
| 0 | 1044 } |
| 1045 | |
| 1046 /* Adds REF to the chain CHAIN. */ | |
| 1047 | |
| 1048 static void | |
| 1049 add_ref_to_chain (chain_p chain, dref ref) | |
| 1050 { | |
| 1051 dref root = get_chain_root (chain); | |
| 16 | 1052 |
| 1053 gcc_assert (wi::les_p (root->offset, ref->offset)); | |
| 1054 widest_int dist = ref->offset - root->offset; | |
| 1055 if (wi::leu_p (MAX_DISTANCE, dist)) | |
| 0 | 1056 { |
| 1057 free (ref); | |
| 1058 return; | |
| 1059 } | |
| 16 | 1060 gcc_assert (wi::fits_uhwi_p (dist)); |
| 1061 | |
| 1062 chain->refs.safe_push (ref); | |
| 1063 | |
| 1064 ref->distance = dist.to_uhwi (); | |
| 0 | 1065 |
| 1066 if (ref->distance >= chain->length) | |
| 1067 { | |
| 1068 chain->length = ref->distance; | |
| 1069 chain->has_max_use_after = false; | |
| 1070 } | |
| 1071 | |
| 16 | 1072 /* Don't set the flag for store-store chain since there is no use. */ |
| 1073 if (chain->type != CT_STORE_STORE | |
| 1074 && ref->distance == chain->length | |
| 0 | 1075 && ref->pos > root->pos) |
| 1076 chain->has_max_use_after = true; | |
| 1077 | |
| 1078 chain->all_always_accessed &= ref->always_accessed; | |
| 1079 } | |
| 1080 | |
| 1081 /* Returns the chain for invariant component COMP. */ | |
| 1082 | |
| 1083 static chain_p | |
| 1084 make_invariant_chain (struct component *comp) | |
| 1085 { | |
| 1086 chain_p chain = XCNEW (struct chain); | |
| 1087 unsigned i; | |
| 1088 dref ref; | |
| 1089 | |
| 1090 chain->type = CT_INVARIANT; | |
| 1091 | |
| 1092 chain->all_always_accessed = true; | |
| 1093 | |
| 16 | 1094 FOR_EACH_VEC_ELT (comp->refs, i, ref) |
| 0 | 1095 { |
| 16 | 1096 chain->refs.safe_push (ref); |
| 0 | 1097 chain->all_always_accessed &= ref->always_accessed; |
| 1098 } | |
| 1099 | |
| 16 | 1100 chain->inits = vNULL; |
| 1101 chain->finis = vNULL; | |
| 1102 | |
| 0 | 1103 return chain; |
| 1104 } | |
| 1105 | |
| 16 | 1106 /* Make a new chain of type TYPE rooted at REF. */ |
| 0 | 1107 |
| 1108 static chain_p | |
| 16 | 1109 make_rooted_chain (dref ref, enum chain_type type) |
| 0 | 1110 { |
| 1111 chain_p chain = XCNEW (struct chain); | |
| 1112 | |
| 16 | 1113 chain->type = type; |
| 1114 chain->refs.safe_push (ref); | |
| 0 | 1115 chain->all_always_accessed = ref->always_accessed; |
| 1116 ref->distance = 0; | |
| 1117 | |
| 16 | 1118 chain->inits = vNULL; |
| 1119 chain->finis = vNULL; | |
| 1120 | |
| 0 | 1121 return chain; |
| 1122 } | |
| 1123 | |
| 1124 /* Returns true if CHAIN is not trivial. */ | |
| 1125 | |
| 1126 static bool | |
| 1127 nontrivial_chain_p (chain_p chain) | |
| 1128 { | |
| 16 | 1129 return chain != NULL && chain->refs.length () > 1; |
| 0 | 1130 } |
| 1131 | |
| 1132 /* Returns the ssa name that contains the value of REF, or NULL_TREE if there | |
| 1133 is no such name. */ | |
| 1134 | |
| 1135 static tree | |
| 1136 name_for_ref (dref ref) | |
| 1137 { | |
| 1138 tree name; | |
| 1139 | |
| 1140 if (is_gimple_assign (ref->stmt)) | |
| 1141 { | |
| 1142 if (!ref->ref || DR_IS_READ (ref->ref)) | |
| 1143 name = gimple_assign_lhs (ref->stmt); | |
| 1144 else | |
| 1145 name = gimple_assign_rhs1 (ref->stmt); | |
| 1146 } | |
| 1147 else | |
| 1148 name = PHI_RESULT (ref->stmt); | |
| 1149 | |
| 1150 return (TREE_CODE (name) == SSA_NAME ? name : NULL_TREE); | |
| 1151 } | |
| 1152 | |
| 1153 /* Returns true if REF is a valid initializer for ROOT with given DISTANCE (in | |
| 1154 iterations of the innermost enclosing loop). */ | |
| 1155 | |
| 1156 static bool | |
| 1157 valid_initializer_p (struct data_reference *ref, | |
| 1158 unsigned distance, struct data_reference *root) | |
| 1159 { | |
| 1160 aff_tree diff, base, step; | |
| 16 | 1161 widest_int off; |
| 0 | 1162 |
| 1163 /* Both REF and ROOT must be accessing the same object. */ | |
| 1164 if (!operand_equal_p (DR_BASE_ADDRESS (ref), DR_BASE_ADDRESS (root), 0)) | |
| 1165 return false; | |
| 1166 | |
| 1167 /* The initializer is defined outside of loop, hence its address must be | |
| 1168 invariant inside the loop. */ | |
| 1169 gcc_assert (integer_zerop (DR_STEP (ref))); | |
| 1170 | |
| 1171 /* If the address of the reference is invariant, initializer must access | |
| 1172 exactly the same location. */ | |
| 1173 if (integer_zerop (DR_STEP (root))) | |
| 1174 return (operand_equal_p (DR_OFFSET (ref), DR_OFFSET (root), 0) | |
| 1175 && operand_equal_p (DR_INIT (ref), DR_INIT (root), 0)); | |
| 1176 | |
| 1177 /* Verify that this index of REF is equal to the root's index at | |
| 1178 -DISTANCE-th iteration. */ | |
| 1179 aff_combination_dr_offset (root, &diff); | |
| 1180 aff_combination_dr_offset (ref, &base); | |
| 16 | 1181 aff_combination_scale (&base, -1); |
| 0 | 1182 aff_combination_add (&diff, &base); |
| 1183 | |
| 16 | 1184 tree_to_aff_combination_expand (DR_STEP (root), TREE_TYPE (DR_STEP (root)), |
| 1185 &step, &name_expansions); | |
| 0 | 1186 if (!aff_combination_constant_multiple_p (&diff, &step, &off)) |
| 1187 return false; | |
| 1188 | |
| 16 | 1189 if (off != distance) |
| 0 | 1190 return false; |
| 1191 | |
| 1192 return true; | |
| 1193 } | |
| 1194 | |
| 1195 /* Finds looparound phi node of LOOP that copies the value of REF, and if its | |
| 1196 initial value is correct (equal to initial value of REF shifted by one | |
| 1197 iteration), returns the phi node. Otherwise, NULL_TREE is returned. ROOT | |
| 1198 is the root of the current chain. */ | |
| 1199 | |
| 16 | 1200 static gphi * |
| 0 | 1201 find_looparound_phi (struct loop *loop, dref ref, dref root) |
| 1202 { | |
| 1203 tree name, init, init_ref; | |
| 16 | 1204 gphi *phi = NULL; |
| 1205 gimple *init_stmt; | |
| 0 | 1206 edge latch = loop_latch_edge (loop); |
| 1207 struct data_reference init_dr; | |
| 16 | 1208 gphi_iterator psi; |
| 0 | 1209 |
| 1210 if (is_gimple_assign (ref->stmt)) | |
| 1211 { | |
| 1212 if (DR_IS_READ (ref->ref)) | |
| 1213 name = gimple_assign_lhs (ref->stmt); | |
| 1214 else | |
| 1215 name = gimple_assign_rhs1 (ref->stmt); | |
| 1216 } | |
| 1217 else | |
| 1218 name = PHI_RESULT (ref->stmt); | |
| 1219 if (!name) | |
| 1220 return NULL; | |
| 1221 | |
| 1222 for (psi = gsi_start_phis (loop->header); !gsi_end_p (psi); gsi_next (&psi)) | |
| 1223 { | |
| 16 | 1224 phi = psi.phi (); |
| 0 | 1225 if (PHI_ARG_DEF_FROM_EDGE (phi, latch) == name) |
| 1226 break; | |
| 1227 } | |
| 1228 | |
| 1229 if (gsi_end_p (psi)) | |
| 1230 return NULL; | |
| 1231 | |
| 1232 init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
| 1233 if (TREE_CODE (init) != SSA_NAME) | |
| 1234 return NULL; | |
| 1235 init_stmt = SSA_NAME_DEF_STMT (init); | |
| 1236 if (gimple_code (init_stmt) != GIMPLE_ASSIGN) | |
| 1237 return NULL; | |
| 1238 gcc_assert (gimple_assign_lhs (init_stmt) == init); | |
| 1239 | |
| 1240 init_ref = gimple_assign_rhs1 (init_stmt); | |
| 1241 if (!REFERENCE_CLASS_P (init_ref) | |
| 1242 && !DECL_P (init_ref)) | |
| 1243 return NULL; | |
| 1244 | |
| 1245 /* Analyze the behavior of INIT_REF with respect to LOOP (innermost | |
| 1246 loop enclosing PHI). */ | |
| 1247 memset (&init_dr, 0, sizeof (struct data_reference)); | |
| 1248 DR_REF (&init_dr) = init_ref; | |
| 1249 DR_STMT (&init_dr) = phi; | |
| 16 | 1250 if (!dr_analyze_innermost (&DR_INNERMOST (&init_dr), init_ref, loop)) |
| 0 | 1251 return NULL; |
| 1252 | |
| 1253 if (!valid_initializer_p (&init_dr, ref->distance + 1, root->ref)) | |
| 1254 return NULL; | |
| 1255 | |
| 1256 return phi; | |
| 1257 } | |
| 1258 | |
| 1259 /* Adds a reference for the looparound copy of REF in PHI to CHAIN. */ | |
| 1260 | |
| 1261 static void | |
| 16 | 1262 insert_looparound_copy (chain_p chain, dref ref, gphi *phi) |
| 0 | 1263 { |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1264 dref nw = XCNEW (struct dref_d), aref; |
| 0 | 1265 unsigned i; |
| 1266 | |
| 1267 nw->stmt = phi; | |
| 1268 nw->distance = ref->distance + 1; | |
| 1269 nw->always_accessed = 1; | |
| 1270 | |
| 16 | 1271 FOR_EACH_VEC_ELT (chain->refs, i, aref) |
| 0 | 1272 if (aref->distance >= nw->distance) |
| 1273 break; | |
| 16 | 1274 chain->refs.safe_insert (i, nw); |
| 0 | 1275 |
| 1276 if (nw->distance > chain->length) | |
| 1277 { | |
| 1278 chain->length = nw->distance; | |
| 1279 chain->has_max_use_after = false; | |
| 1280 } | |
| 1281 } | |
| 1282 | |
| 1283 /* For references in CHAIN that are copied around the LOOP (created previously | |
| 1284 by PRE, or by user), add the results of such copies to the chain. This | |
| 1285 enables us to remove the copies by unrolling, and may need less registers | |
| 1286 (also, it may allow us to combine chains together). */ | |
| 1287 | |
| 1288 static void | |
| 1289 add_looparound_copies (struct loop *loop, chain_p chain) | |
| 1290 { | |
| 1291 unsigned i; | |
| 1292 dref ref, root = get_chain_root (chain); | |
| 16 | 1293 gphi *phi; |
| 1294 | |
| 1295 if (chain->type == CT_STORE_STORE) | |
| 1296 return; | |
| 1297 | |
| 1298 FOR_EACH_VEC_ELT (chain->refs, i, ref) | |
| 0 | 1299 { |
| 1300 phi = find_looparound_phi (loop, ref, root); | |
| 1301 if (!phi) | |
| 1302 continue; | |
| 1303 | |
| 1304 bitmap_set_bit (looparound_phis, SSA_NAME_VERSION (PHI_RESULT (phi))); | |
| 1305 insert_looparound_copy (chain, ref, phi); | |
| 1306 } | |
| 1307 } | |
| 1308 | |
| 1309 /* Find roots of the values and determine distances in the component COMP. | |
| 1310 The references are redistributed into CHAINS. LOOP is the current | |
| 1311 loop. */ | |
| 1312 | |
| 1313 static void | |
| 1314 determine_roots_comp (struct loop *loop, | |
| 1315 struct component *comp, | |
| 16 | 1316 vec<chain_p> *chains) |
| 0 | 1317 { |
| 1318 unsigned i; | |
| 1319 dref a; | |
| 1320 chain_p chain = NULL; | |
| 16 | 1321 widest_int last_ofs = 0; |
| 1322 enum chain_type type; | |
| 0 | 1323 |
| 1324 /* Invariants are handled specially. */ | |
| 1325 if (comp->comp_step == RS_INVARIANT) | |
| 1326 { | |
| 1327 chain = make_invariant_chain (comp); | |
| 16 | 1328 chains->safe_push (chain); |
| 0 | 1329 return; |
| 1330 } | |
| 1331 | |
| 16 | 1332 /* Trivial component. */ |
| 1333 if (comp->refs.length () <= 1) | |
| 1334 return; | |
| 1335 | |
| 1336 comp->refs.qsort (order_drefs); | |
| 1337 FOR_EACH_VEC_ELT (comp->refs, i, a) | |
| 0 | 1338 { |
| 16 | 1339 if (!chain |
| 1340 || (!comp->eliminate_store_p && DR_IS_WRITE (a->ref)) | |
| 1341 || wi::leu_p (MAX_DISTANCE, a->offset - last_ofs)) | |
| 0 | 1342 { |
| 1343 if (nontrivial_chain_p (chain)) | |
|
11
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
9
diff
changeset
|
1344 { |
|
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
9
diff
changeset
|
1345 add_looparound_copies (loop, chain); |
| 16 | 1346 chains->safe_push (chain); |
|
11
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
9
diff
changeset
|
1347 } |
| 0 | 1348 else |
| 1349 release_chain (chain); | |
| 16 | 1350 |
| 1351 if (DR_IS_READ (a->ref)) | |
| 1352 type = CT_LOAD; | |
| 1353 else | |
| 1354 type = comp->eliminate_store_p ? CT_STORE_STORE : CT_STORE_LOAD; | |
| 1355 | |
| 1356 chain = make_rooted_chain (a, type); | |
|
11
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
9
diff
changeset
|
1357 last_ofs = a->offset; |
| 0 | 1358 continue; |
| 1359 } | |
| 1360 | |
| 1361 add_ref_to_chain (chain, a); | |
| 1362 } | |
| 1363 | |
| 1364 if (nontrivial_chain_p (chain)) | |
| 1365 { | |
| 1366 add_looparound_copies (loop, chain); | |
| 16 | 1367 chains->safe_push (chain); |
| 0 | 1368 } |
| 1369 else | |
| 1370 release_chain (chain); | |
| 1371 } | |
| 1372 | |
| 1373 /* Find roots of the values and determine distances in components COMPS, and | |
| 1374 separates the references to CHAINS. LOOP is the current loop. */ | |
| 1375 | |
| 1376 static void | |
| 1377 determine_roots (struct loop *loop, | |
| 16 | 1378 struct component *comps, vec<chain_p> *chains) |
| 0 | 1379 { |
| 1380 struct component *comp; | |
| 1381 | |
| 1382 for (comp = comps; comp; comp = comp->next) | |
| 1383 determine_roots_comp (loop, comp, chains); | |
| 1384 } | |
| 1385 | |
| 1386 /* Replace the reference in statement STMT with temporary variable | |
| 1387 NEW_TREE. If SET is true, NEW_TREE is instead initialized to the value of | |
| 1388 the reference in the statement. IN_LHS is true if the reference | |
| 1389 is in the lhs of STMT, false if it is in rhs. */ | |
| 1390 | |
| 1391 static void | |
| 16 | 1392 replace_ref_with (gimple *stmt, tree new_tree, bool set, bool in_lhs) |
| 0 | 1393 { |
| 1394 tree val; | |
| 16 | 1395 gassign *new_stmt; |
| 0 | 1396 gimple_stmt_iterator bsi, psi; |
| 1397 | |
| 1398 if (gimple_code (stmt) == GIMPLE_PHI) | |
| 1399 { | |
| 1400 gcc_assert (!in_lhs && !set); | |
| 1401 | |
| 1402 val = PHI_RESULT (stmt); | |
| 1403 bsi = gsi_after_labels (gimple_bb (stmt)); | |
| 1404 psi = gsi_for_stmt (stmt); | |
| 1405 remove_phi_node (&psi, false); | |
| 1406 | |
| 1407 /* Turn the phi node into GIMPLE_ASSIGN. */ | |
| 1408 new_stmt = gimple_build_assign (val, new_tree); | |
| 1409 gsi_insert_before (&bsi, new_stmt, GSI_NEW_STMT); | |
| 1410 return; | |
| 1411 } | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1412 |
| 0 | 1413 /* Since the reference is of gimple_reg type, it should only |
| 1414 appear as lhs or rhs of modify statement. */ | |
| 1415 gcc_assert (is_gimple_assign (stmt)); | |
| 1416 | |
| 1417 bsi = gsi_for_stmt (stmt); | |
| 1418 | |
| 1419 /* If we do not need to initialize NEW_TREE, just replace the use of OLD. */ | |
| 1420 if (!set) | |
| 1421 { | |
| 1422 gcc_assert (!in_lhs); | |
| 1423 gimple_assign_set_rhs_from_tree (&bsi, new_tree); | |
| 1424 stmt = gsi_stmt (bsi); | |
| 1425 update_stmt (stmt); | |
| 1426 return; | |
| 1427 } | |
| 1428 | |
| 1429 if (in_lhs) | |
| 1430 { | |
| 1431 /* We have statement | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1432 |
| 0 | 1433 OLD = VAL |
| 1434 | |
| 1435 If OLD is a memory reference, then VAL is gimple_val, and we transform | |
| 1436 this to | |
| 1437 | |
| 1438 OLD = VAL | |
| 1439 NEW = VAL | |
| 1440 | |
|
9
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1441 Otherwise, we are replacing a combination chain, |
| 0 | 1442 VAL is the expression that performs the combination, and OLD is an |
| 1443 SSA name. In this case, we transform the assignment to | |
| 1444 | |
| 1445 OLD = VAL | |
| 1446 NEW = OLD | |
| 1447 | |
| 1448 */ | |
| 1449 | |
| 1450 val = gimple_assign_lhs (stmt); | |
| 1451 if (TREE_CODE (val) != SSA_NAME) | |
| 1452 { | |
| 1453 val = gimple_assign_rhs1 (stmt); | |
| 16 | 1454 gcc_assert (gimple_assign_single_p (stmt)); |
| 1455 if (TREE_CLOBBER_P (val)) | |
| 1456 val = get_or_create_ssa_default_def (cfun, SSA_NAME_VAR (new_tree)); | |
| 1457 else | |
| 1458 gcc_assert (gimple_assign_copy_p (stmt)); | |
| 0 | 1459 } |
| 1460 } | |
| 1461 else | |
| 1462 { | |
| 1463 /* VAL = OLD | |
| 1464 | |
| 1465 is transformed to | |
| 1466 | |
| 1467 VAL = OLD | |
| 1468 NEW = VAL */ | |
| 1469 | |
| 1470 val = gimple_assign_lhs (stmt); | |
| 1471 } | |
| 1472 | |
| 1473 new_stmt = gimple_build_assign (new_tree, unshare_expr (val)); | |
| 1474 gsi_insert_after (&bsi, new_stmt, GSI_NEW_STMT); | |
| 1475 } | |
| 1476 | |
| 16 | 1477 /* Returns a memory reference to DR in the (NITERS + ITER)-th iteration |
| 1478 of the loop it was analyzed in. Append init stmts to STMTS. */ | |
| 0 | 1479 |
| 1480 static tree | |
| 16 | 1481 ref_at_iteration (data_reference_p dr, int iter, |
| 1482 gimple_seq *stmts, tree niters = NULL_TREE) | |
| 0 | 1483 { |
| 16 | 1484 tree off = DR_OFFSET (dr); |
| 1485 tree coff = DR_INIT (dr); | |
| 1486 tree ref = DR_REF (dr); | |
| 1487 enum tree_code ref_code = ERROR_MARK; | |
| 1488 tree ref_type = NULL_TREE; | |
| 1489 tree ref_op1 = NULL_TREE; | |
| 1490 tree ref_op2 = NULL_TREE; | |
| 1491 tree new_offset; | |
| 1492 | |
| 1493 if (iter != 0) | |
| 0 | 1494 { |
| 16 | 1495 new_offset = size_binop (MULT_EXPR, DR_STEP (dr), ssize_int (iter)); |
| 1496 if (TREE_CODE (new_offset) == INTEGER_CST) | |
| 1497 coff = size_binop (PLUS_EXPR, coff, new_offset); | |
| 1498 else | |
| 1499 off = size_binop (PLUS_EXPR, off, new_offset); | |
| 0 | 1500 } |
| 16 | 1501 |
| 1502 if (niters != NULL_TREE) | |
| 0 | 1503 { |
| 16 | 1504 niters = fold_convert (ssizetype, niters); |
| 1505 new_offset = size_binop (MULT_EXPR, DR_STEP (dr), niters); | |
| 1506 if (TREE_CODE (niters) == INTEGER_CST) | |
| 1507 coff = size_binop (PLUS_EXPR, coff, new_offset); | |
| 1508 else | |
| 1509 off = size_binop (PLUS_EXPR, off, new_offset); | |
| 0 | 1510 } |
| 16 | 1511 |
| 1512 /* While data-ref analysis punts on bit offsets it still handles | |
| 1513 bitfield accesses at byte boundaries. Cope with that. Note that | |
| 1514 if the bitfield object also starts at a byte-boundary we can simply | |
| 1515 replicate the COMPONENT_REF, but we have to subtract the component's | |
| 1516 byte-offset from the MEM_REF address first. | |
| 1517 Otherwise we simply build a BIT_FIELD_REF knowing that the bits | |
| 1518 start at offset zero. */ | |
| 1519 if (TREE_CODE (ref) == COMPONENT_REF | |
| 1520 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))) | |
| 0 | 1521 { |
| 16 | 1522 unsigned HOST_WIDE_INT boff; |
| 1523 tree field = TREE_OPERAND (ref, 1); | |
| 1524 tree offset = component_ref_field_offset (ref); | |
| 1525 ref_type = TREE_TYPE (ref); | |
| 1526 boff = tree_to_uhwi (DECL_FIELD_BIT_OFFSET (field)); | |
| 1527 /* This can occur in Ada. See the comment in get_bit_range. */ | |
| 1528 if (boff % BITS_PER_UNIT != 0 | |
| 1529 || !tree_fits_uhwi_p (offset)) | |
| 0 | 1530 { |
| 16 | 1531 ref_code = BIT_FIELD_REF; |
| 1532 ref_op1 = DECL_SIZE (field); | |
| 1533 ref_op2 = bitsize_zero_node; | |
| 0 | 1534 } |
| 1535 else | |
| 1536 { | |
| 16 | 1537 boff >>= LOG2_BITS_PER_UNIT; |