| 1 | /* Induction variable optimizations. |
| 2 | Copyright (C) 2003-2026 Free Software Foundation, Inc. |
| 3 | |
| 4 | This file is part of GCC. |
| 5 | |
| 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. |
| 10 | |
| 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. |
| 15 | |
| 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 pass tries to find the optimal set of induction variables for the loop. |
| 21 | It optimizes just the basic linear induction variables (although adding |
| 22 | support for other types should not be too hard). It includes the |
| 23 | optimizations commonly known as strength reduction, induction variable |
| 24 | coalescing and induction variable elimination. It does it in the |
| 25 | following steps: |
| 26 | |
| 27 | 1) The interesting uses of induction variables are found. This includes |
| 28 | |
| 29 | -- uses of induction variables in non-linear expressions |
| 30 | -- addresses of arrays |
| 31 | -- comparisons of induction variables |
| 32 | |
| 33 | Note the interesting uses are categorized and handled in group. |
| 34 | Generally, address type uses are grouped together if their iv bases |
| 35 | are different in constant offset. |
| 36 | |
| 37 | 2) Candidates for the induction variables are found. This includes |
| 38 | |
| 39 | -- old induction variables |
| 40 | -- the variables defined by expressions derived from the "interesting |
| 41 | groups/uses" above |
| 42 | |
| 43 | 3) The optimal (w.r. to a cost function) set of variables is chosen. The |
| 44 | cost function assigns a cost to sets of induction variables and consists |
| 45 | of three parts: |
| 46 | |
| 47 | -- The group/use costs. Each of the interesting groups/uses chooses |
| 48 | the best induction variable in the set and adds its cost to the sum. |
| 49 | The cost reflects the time spent on modifying the induction variables |
| 50 | value to be usable for the given purpose (adding base and offset for |
| 51 | arrays, etc.). |
| 52 | -- The variable costs. Each of the variables has a cost assigned that |
| 53 | reflects the costs associated with incrementing the value of the |
| 54 | variable. The original variables are somewhat preferred. |
| 55 | -- The set cost. Depending on the size of the set, extra cost may be |
| 56 | added to reflect register pressure. |
| 57 | |
| 58 | All the costs are defined in a machine-specific way, using the target |
| 59 | hooks and machine descriptions to determine them. |
| 60 | |
| 61 | 4) The trees are transformed to use the new variables, the dead code is |
| 62 | removed. |
| 63 | |
| 64 | All of this is done loop by loop. Doing it globally is theoretically |
| 65 | possible, it might give a better performance and it might enable us |
| 66 | to decide costs more precisely, but getting all the interactions right |
| 67 | would be complicated. |
| 68 | |
| 69 | For the targets supporting low-overhead loops, IVOPTs has to take care of |
| 70 | the loops which will probably be transformed in RTL doloop optimization, |
| 71 | to try to make selected IV candidate set optimal. The process of doloop |
| 72 | support includes: |
| 73 | |
| 74 | 1) Analyze the current loop will be transformed to doloop or not, find and |
| 75 | mark its compare type IV use as doloop use (iv_group field doloop_p), and |
| 76 | set flag doloop_use_p of ivopts_data to notify subsequent processings on |
| 77 | doloop. See analyze_and_mark_doloop_use and its callees for the details. |
| 78 | The target hook predict_doloop_p can be used for target specific checks. |
| 79 | |
| 80 | 2) Add one doloop dedicated IV cand {(may_be_zero ? 1 : (niter + 1)), +, -1}, |
| 81 | set flag doloop_p of iv_cand, step cost is set as zero and no extra cost |
| 82 | like biv. For cost determination between doloop IV cand and IV use, the |
| 83 | target hooks doloop_cost_for_generic and doloop_cost_for_address are |
| 84 | provided to add on extra costs for generic type and address type IV use. |
| 85 | Zero cost is assigned to the pair between doloop IV cand and doloop IV |
| 86 | use, and bound zero is set for IV elimination. |
| 87 | |
| 88 | 3) With the cost setting in step 2), the current cost model based IV |
| 89 | selection algorithm will process as usual, pick up doloop dedicated IV if |
| 90 | profitable. */ |
| 91 | |
| 92 | #include "config.h" |
| 93 | #include "system.h" |
| 94 | #include "coretypes.h" |
| 95 | #include "backend.h" |
| 96 | #include "rtl.h" |
| 97 | #include "tree.h" |
| 98 | #include "gimple.h" |
| 99 | #include "cfghooks.h" |
| 100 | #include "tree-pass.h" |
| 101 | #include "memmodel.h" |
| 102 | #include "tm_p.h" |
| 103 | #include "ssa.h" |
| 104 | #include "expmed.h" |
| 105 | #include "insn-config.h" |
| 106 | #include "emit-rtl.h" |
| 107 | #include "recog.h" |
| 108 | #include "cgraph.h" |
| 109 | #include "gimple-pretty-print.h" |
| 110 | #include "alias.h" |
| 111 | #include "fold-const.h" |
| 112 | #include "stor-layout.h" |
| 113 | #include "tree-eh.h" |
| 114 | #include "gimplify.h" |
| 115 | #include "gimple-iterator.h" |
| 116 | #include "gimplify-me.h" |
| 117 | #include "tree-cfg.h" |
| 118 | #include "tree-ssa-loop-ivopts.h" |
| 119 | #include "tree-ssa-loop-manip.h" |
| 120 | #include "tree-ssa-loop-niter.h" |
| 121 | #include "tree-ssa-loop.h" |
| 122 | #include "explow.h" |
| 123 | #include "expr.h" |
| 124 | #include "tree-dfa.h" |
| 125 | #include "tree-ssa.h" |
| 126 | #include "cfgloop.h" |
| 127 | #include "tree-scalar-evolution.h" |
| 128 | #include "tree-affine.h" |
| 129 | #include "tree-ssa-propagate.h" |
| 130 | #include "tree-ssa-address.h" |
| 131 | #include "builtins.h" |
| 132 | #include "tree-vectorizer.h" |
| 133 | #include "dbgcnt.h" |
| 134 | #include "cfganal.h" |
| 135 | #include "gimple-fold.h" |
| 136 | |
| 137 | /* For lang_hooks.types.type_for_mode. */ |
| 138 | #include "langhooks.h" |
| 139 | |
| 140 | /* FIXME: Expressions are expanded to RTL in this pass to determine the |
| 141 | cost of different addressing modes. This should be moved to a TBD |
| 142 | interface between the GIMPLE and RTL worlds. */ |
| 143 | |
| 144 | /* The infinite cost. */ |
| 145 | #define INFTY 1000000000 |
| 146 | |
| 147 | /* Returns the expected number of loop iterations for LOOP. |
| 148 | The average trip count is computed from profile data if it |
| 149 | exists. */ |
| 150 | |
| 151 | static inline unsigned HOST_WIDE_INT |
| 152 | avg_loop_niter (class loop *loop) |
| 153 | { |
| 154 | HOST_WIDE_INT niter = estimated_stmt_executions_int (loop); |
| 155 | if (niter == -1) |
| 156 | { |
| 157 | niter = likely_max_stmt_executions_int (loop); |
| 158 | |
| 159 | if (niter == -1 || niter > param_avg_loop_niter) |
| 160 | return param_avg_loop_niter; |
| 161 | } |
| 162 | |
| 163 | return niter; |
| 164 | } |
| 165 | |
| 166 | struct iv_use; |
| 167 | |
| 168 | /* Representation of the induction variable. */ |
| 169 | struct iv |
| 170 | { |
| 171 | tree base; /* Initial value of the iv. */ |
| 172 | tree base_object; /* A memory object to that the induction variable points. */ |
| 173 | tree step; /* Step of the iv (constant only). */ |
| 174 | tree ssa_name; /* The ssa name with the value. */ |
| 175 | struct iv_use *nonlin_use; /* The identifier in the use if it is the case. */ |
| 176 | bool biv_p; /* Is it a biv? */ |
| 177 | bool no_overflow; /* True if the iv doesn't overflow. */ |
| 178 | bool have_address_use;/* For biv, indicate if it's used in any address |
| 179 | type use. */ |
| 180 | }; |
| 181 | |
| 182 | /* Per-ssa version information (induction variable descriptions, etc.). */ |
| 183 | struct version_info |
| 184 | { |
| 185 | tree name; /* The ssa name. */ |
| 186 | struct iv *iv; /* Induction variable description. */ |
| 187 | bool has_nonlin_use; /* For a loop-level invariant, whether it is used in |
| 188 | an expression that is not an induction variable. */ |
| 189 | bool preserve_biv; /* For the original biv, whether to preserve it. */ |
| 190 | unsigned inv_id; /* Id of an invariant. */ |
| 191 | }; |
| 192 | |
| 193 | /* Types of uses. */ |
| 194 | enum use_type |
| 195 | { |
| 196 | USE_NONLINEAR_EXPR, /* Use in a nonlinear expression. */ |
| 197 | USE_REF_ADDRESS, /* Use is an address for an explicit memory |
| 198 | reference. */ |
| 199 | USE_PTR_ADDRESS, /* Use is a pointer argument to a function in |
| 200 | cases where the expansion of the function |
| 201 | will turn the argument into a normal address. */ |
| 202 | USE_COMPARE /* Use is a compare. */ |
| 203 | }; |
| 204 | |
| 205 | /* Cost of a computation. */ |
| 206 | class comp_cost |
| 207 | { |
| 208 | public: |
| 209 | comp_cost (): cost (0), complexity (0), scratch (0) |
| 210 | {} |
| 211 | |
| 212 | comp_cost (int64_t cost, unsigned complexity, int64_t scratch = 0) |
| 213 | : cost (cost), complexity (complexity), scratch (scratch) |
| 214 | {} |
| 215 | |
| 216 | /* Returns true if COST is infinite. */ |
| 217 | bool infinite_cost_p (); |
| 218 | |
| 219 | /* Adds costs COST1 and COST2. */ |
| 220 | friend comp_cost operator+ (comp_cost cost1, comp_cost cost2); |
| 221 | |
| 222 | /* Adds COST to the comp_cost. */ |
| 223 | comp_cost operator+= (comp_cost cost); |
| 224 | |
| 225 | /* Adds constant C to this comp_cost. */ |
| 226 | comp_cost operator+= (HOST_WIDE_INT c); |
| 227 | |
| 228 | /* Subtracts constant C to this comp_cost. */ |
| 229 | comp_cost operator-= (HOST_WIDE_INT c); |
| 230 | |
| 231 | /* Divide the comp_cost by constant C. */ |
| 232 | comp_cost operator/= (HOST_WIDE_INT c); |
| 233 | |
| 234 | /* Multiply the comp_cost by constant C. */ |
| 235 | comp_cost operator*= (HOST_WIDE_INT c); |
| 236 | |
| 237 | /* Subtracts costs COST1 and COST2. */ |
| 238 | friend comp_cost operator- (comp_cost cost1, comp_cost cost2); |
| 239 | |
| 240 | /* Subtracts COST from this comp_cost. */ |
| 241 | comp_cost operator-= (comp_cost cost); |
| 242 | |
| 243 | /* Returns true if COST1 is smaller than COST2. */ |
| 244 | friend bool operator< (comp_cost cost1, comp_cost cost2); |
| 245 | |
| 246 | /* Returns true if COST1 and COST2 are equal. */ |
| 247 | friend bool operator== (comp_cost cost1, comp_cost cost2); |
| 248 | |
| 249 | /* Returns true if COST1 is smaller or equal than COST2. */ |
| 250 | friend bool operator<= (comp_cost cost1, comp_cost cost2); |
| 251 | |
| 252 | int64_t cost; /* The runtime cost. */ |
| 253 | unsigned complexity; /* The estimate of the complexity of the code for |
| 254 | the computation (in no concrete units -- |
| 255 | complexity field should be larger for more |
| 256 | complex expressions and addressing modes). */ |
| 257 | int64_t scratch; /* Scratch used during cost computation. */ |
| 258 | }; |
| 259 | |
| 260 | static const comp_cost no_cost; |
| 261 | static const comp_cost infinite_cost (INFTY, 0, INFTY); |
| 262 | |
| 263 | bool |
| 264 | comp_cost::infinite_cost_p () |
| 265 | { |
| 266 | return cost == INFTY; |
| 267 | } |
| 268 | |
| 269 | comp_cost |
| 270 | operator+ (comp_cost cost1, comp_cost cost2) |
| 271 | { |
| 272 | if (cost1.infinite_cost_p () || cost2.infinite_cost_p ()) |
| 273 | return infinite_cost; |
| 274 | |
| 275 | gcc_assert (cost1.cost + cost2.cost < infinite_cost.cost); |
| 276 | cost1.cost += cost2.cost; |
| 277 | cost1.complexity += cost2.complexity; |
| 278 | |
| 279 | return cost1; |
| 280 | } |
| 281 | |
| 282 | comp_cost |
| 283 | operator- (comp_cost cost1, comp_cost cost2) |
| 284 | { |
| 285 | if (cost1.infinite_cost_p ()) |
| 286 | return infinite_cost; |
| 287 | |
| 288 | gcc_assert (!cost2.infinite_cost_p ()); |
| 289 | gcc_assert (cost1.cost - cost2.cost < infinite_cost.cost); |
| 290 | |
| 291 | cost1.cost -= cost2.cost; |
| 292 | cost1.complexity -= cost2.complexity; |
| 293 | |
| 294 | return cost1; |
| 295 | } |
| 296 | |
| 297 | comp_cost |
| 298 | comp_cost::operator+= (comp_cost cost) |
| 299 | { |
| 300 | *this = *this + cost; |
| 301 | return *this; |
| 302 | } |
| 303 | |
| 304 | comp_cost |
| 305 | comp_cost::operator+= (HOST_WIDE_INT c) |
| 306 | { |
| 307 | if (c >= INFTY) |
| 308 | this->cost = INFTY; |
| 309 | |
| 310 | if (infinite_cost_p ()) |
| 311 | return *this; |
| 312 | |
| 313 | gcc_assert (this->cost + c < infinite_cost.cost); |
| 314 | this->cost += c; |
| 315 | |
| 316 | return *this; |
| 317 | } |
| 318 | |
| 319 | comp_cost |
| 320 | comp_cost::operator-= (HOST_WIDE_INT c) |
| 321 | { |
| 322 | if (infinite_cost_p ()) |
| 323 | return *this; |
| 324 | |
| 325 | gcc_assert (this->cost - c < infinite_cost.cost); |
| 326 | this->cost -= c; |
| 327 | |
| 328 | return *this; |
| 329 | } |
| 330 | |
| 331 | comp_cost |
| 332 | comp_cost::operator/= (HOST_WIDE_INT c) |
| 333 | { |
| 334 | gcc_assert (c != 0); |
| 335 | if (infinite_cost_p ()) |
| 336 | return *this; |
| 337 | |
| 338 | this->cost /= c; |
| 339 | |
| 340 | return *this; |
| 341 | } |
| 342 | |
| 343 | comp_cost |
| 344 | comp_cost::operator*= (HOST_WIDE_INT c) |
| 345 | { |
| 346 | if (infinite_cost_p ()) |
| 347 | return *this; |
| 348 | |
| 349 | gcc_assert (this->cost * c < infinite_cost.cost); |
| 350 | this->cost *= c; |
| 351 | |
| 352 | return *this; |
| 353 | } |
| 354 | |
| 355 | comp_cost |
| 356 | comp_cost::operator-= (comp_cost cost) |
| 357 | { |
| 358 | *this = *this - cost; |
| 359 | return *this; |
| 360 | } |
| 361 | |
| 362 | bool |
| 363 | operator< (comp_cost cost1, comp_cost cost2) |
| 364 | { |
| 365 | if (cost1.cost == cost2.cost) |
| 366 | return cost1.complexity < cost2.complexity; |
| 367 | |
| 368 | return cost1.cost < cost2.cost; |
| 369 | } |
| 370 | |
| 371 | bool |
| 372 | operator== (comp_cost cost1, comp_cost cost2) |
| 373 | { |
| 374 | return cost1.cost == cost2.cost |
| 375 | && cost1.complexity == cost2.complexity; |
| 376 | } |
| 377 | |
| 378 | bool |
| 379 | operator<= (comp_cost cost1, comp_cost cost2) |
| 380 | { |
| 381 | return cost1 < cost2 || cost1 == cost2; |
| 382 | } |
| 383 | |
| 384 | struct iv_inv_expr_ent; |
| 385 | |
| 386 | /* The candidate - cost pair. */ |
| 387 | class cost_pair |
| 388 | { |
| 389 | public: |
| 390 | struct iv_cand *cand; /* The candidate. */ |
| 391 | comp_cost cost; /* The cost. */ |
| 392 | enum tree_code comp; /* For iv elimination, the comparison. */ |
| 393 | bitmap inv_vars; /* The list of invariant ssa_vars that have to be |
| 394 | preserved when representing iv_use with iv_cand. */ |
| 395 | bitmap inv_exprs; /* The list of newly created invariant expressions |
| 396 | when representing iv_use with iv_cand. */ |
| 397 | tree value; /* For final value elimination, the expression for |
| 398 | the final value of the iv. For iv elimination, |
| 399 | the new bound to compare with. */ |
| 400 | }; |
| 401 | |
| 402 | /* Use. */ |
| 403 | struct iv_use |
| 404 | { |
| 405 | unsigned id; /* The id of the use. */ |
| 406 | unsigned group_id; /* The group id the use belongs to. */ |
| 407 | enum use_type type; /* Type of the use. */ |
| 408 | tree mem_type; /* The memory type to use when testing whether an |
| 409 | address is legitimate, and what the address's |
| 410 | cost is. */ |
| 411 | struct iv *iv; /* The induction variable it is based on. */ |
| 412 | gimple *stmt; /* Statement in that it occurs. */ |
| 413 | tree *op_p; /* The place where it occurs. */ |
| 414 | |
| 415 | tree addr_base; /* Base address with const offset stripped. */ |
| 416 | poly_uint64 addr_offset; |
| 417 | /* Const offset stripped from base address. */ |
| 418 | }; |
| 419 | |
| 420 | /* Group of uses. */ |
| 421 | struct iv_group |
| 422 | { |
| 423 | /* The id of the group. */ |
| 424 | unsigned id; |
| 425 | /* Uses of the group are of the same type. */ |
| 426 | enum use_type type; |
| 427 | /* The set of "related" IV candidates, plus the important ones. */ |
| 428 | bitmap related_cands; |
| 429 | /* Number of IV candidates in the cost_map. */ |
| 430 | unsigned n_map_members; |
| 431 | /* The costs wrto the iv candidates. */ |
| 432 | class cost_pair *cost_map; |
| 433 | /* The selected candidate for the group. */ |
| 434 | struct iv_cand *selected; |
| 435 | /* To indicate this is a doloop use group. */ |
| 436 | bool doloop_p; |
| 437 | /* Uses in the group. */ |
| 438 | vec<struct iv_use *> vuses; |
| 439 | }; |
| 440 | |
| 441 | /* The position where the iv is computed. */ |
| 442 | enum iv_position |
| 443 | { |
| 444 | IP_NORMAL, /* At the end, just before the exit condition. */ |
| 445 | IP_END, /* At the end of the latch block. */ |
| 446 | IP_BEFORE_USE, /* Immediately before a specific use. */ |
| 447 | IP_AFTER_USE, /* Immediately after a specific use. */ |
| 448 | IP_ORIGINAL /* The original biv. */ |
| 449 | }; |
| 450 | |
| 451 | /* The induction variable candidate. */ |
| 452 | struct iv_cand |
| 453 | { |
| 454 | unsigned id; /* The number of the candidate. */ |
| 455 | bool important; /* Whether this is an "important" candidate, i.e. such |
| 456 | that it should be considered by all uses. */ |
| 457 | bool involves_undefs; /* Whether the IV involves undefined values. */ |
| 458 | ENUM_BITFIELD(iv_position) pos : 8; /* Where it is computed. */ |
| 459 | gimple *incremented_at;/* For original biv, the statement where it is |
| 460 | incremented. */ |
| 461 | tree var_before; /* The variable used for it before increment. */ |
| 462 | tree var_after; /* The variable used for it after increment. */ |
| 463 | struct iv *iv; /* The value of the candidate. NULL for |
| 464 | "pseudocandidate" used to indicate the possibility |
| 465 | to replace the final value of an iv by direct |
| 466 | computation of the value. */ |
| 467 | unsigned cost; /* Cost of the candidate. */ |
| 468 | unsigned cost_step; /* Cost of the candidate's increment operation. */ |
| 469 | struct iv_use *ainc_use; /* For IP_{BEFORE,AFTER}_USE candidates, the place |
| 470 | where it is incremented. */ |
| 471 | bitmap inv_vars; /* The list of invariant ssa_vars used in step of the |
| 472 | iv_cand. */ |
| 473 | bitmap inv_exprs; /* If step is more complicated than a single ssa_var, |
| 474 | handle it as a new invariant expression which will |
| 475 | be hoisted out of loop. */ |
| 476 | struct iv *orig_iv; /* The original iv if this cand is added from biv with |
| 477 | smaller type. */ |
| 478 | bool doloop_p; /* Whether this is a doloop candidate. */ |
| 479 | }; |
| 480 | |
| 481 | /* Hashtable entry for common candidate derived from iv uses. */ |
| 482 | class iv_common_cand |
| 483 | { |
| 484 | public: |
| 485 | tree base; |
| 486 | tree step; |
| 487 | /* IV uses from which this common candidate is derived. */ |
| 488 | auto_vec<struct iv_use *> uses; |
| 489 | hashval_t hash; |
| 490 | }; |
| 491 | |
| 492 | /* Hashtable helpers. */ |
| 493 | |
| 494 | struct iv_common_cand_hasher : delete_ptr_hash <iv_common_cand> |
| 495 | { |
| 496 | static inline hashval_t hash (const iv_common_cand *); |
| 497 | static inline bool equal (const iv_common_cand *, const iv_common_cand *); |
| 498 | }; |
| 499 | |
| 500 | /* Hash function for possible common candidates. */ |
| 501 | |
| 502 | inline hashval_t |
| 503 | iv_common_cand_hasher::hash (const iv_common_cand *ccand) |
| 504 | { |
| 505 | return ccand->hash; |
| 506 | } |
| 507 | |
| 508 | /* Hash table equality function for common candidates. */ |
| 509 | |
| 510 | inline bool |
| 511 | iv_common_cand_hasher::equal (const iv_common_cand *ccand1, |
| 512 | const iv_common_cand *ccand2) |
| 513 | { |
| 514 | return (ccand1->hash == ccand2->hash |
| 515 | && operand_equal_p (ccand1->base, ccand2->base, flags: 0) |
| 516 | && operand_equal_p (ccand1->step, ccand2->step, flags: 0) |
| 517 | && (TYPE_PRECISION (TREE_TYPE (ccand1->base)) |
| 518 | == TYPE_PRECISION (TREE_TYPE (ccand2->base)))); |
| 519 | } |
| 520 | |
| 521 | /* Loop invariant expression hashtable entry. */ |
| 522 | |
| 523 | struct iv_inv_expr_ent |
| 524 | { |
| 525 | /* Tree expression of the entry. */ |
| 526 | tree expr; |
| 527 | /* Unique indentifier. */ |
| 528 | int id; |
| 529 | /* Hash value. */ |
| 530 | hashval_t hash; |
| 531 | }; |
| 532 | |
| 533 | /* Sort iv_inv_expr_ent pair A and B by id field. */ |
| 534 | |
| 535 | static int |
| 536 | sort_iv_inv_expr_ent (const void *a, const void *b) |
| 537 | { |
| 538 | const iv_inv_expr_ent * const *e1 = (const iv_inv_expr_ent * const *) (a); |
| 539 | const iv_inv_expr_ent * const *e2 = (const iv_inv_expr_ent * const *) (b); |
| 540 | |
| 541 | unsigned id1 = (*e1)->id; |
| 542 | unsigned id2 = (*e2)->id; |
| 543 | |
| 544 | if (id1 < id2) |
| 545 | return -1; |
| 546 | else if (id1 > id2) |
| 547 | return 1; |
| 548 | else |
| 549 | return 0; |
| 550 | } |
| 551 | |
| 552 | /* Hashtable helpers. */ |
| 553 | |
| 554 | struct iv_inv_expr_hasher : free_ptr_hash <iv_inv_expr_ent> |
| 555 | { |
| 556 | static inline hashval_t hash (const iv_inv_expr_ent *); |
| 557 | static inline bool equal (const iv_inv_expr_ent *, const iv_inv_expr_ent *); |
| 558 | }; |
| 559 | |
| 560 | /* Return true if uses of type TYPE represent some form of address. */ |
| 561 | |
| 562 | inline bool |
| 563 | address_p (use_type type) |
| 564 | { |
| 565 | return type == USE_REF_ADDRESS || type == USE_PTR_ADDRESS; |
| 566 | } |
| 567 | |
| 568 | /* Hash function for loop invariant expressions. */ |
| 569 | |
| 570 | inline hashval_t |
| 571 | iv_inv_expr_hasher::hash (const iv_inv_expr_ent *expr) |
| 572 | { |
| 573 | return expr->hash; |
| 574 | } |
| 575 | |
| 576 | /* Hash table equality function for expressions. */ |
| 577 | |
| 578 | inline bool |
| 579 | iv_inv_expr_hasher::equal (const iv_inv_expr_ent *expr1, |
| 580 | const iv_inv_expr_ent *expr2) |
| 581 | { |
| 582 | return expr1->hash == expr2->hash |
| 583 | && operand_equal_p (expr1->expr, expr2->expr, flags: 0); |
| 584 | } |
| 585 | |
| 586 | struct ivopts_data |
| 587 | { |
| 588 | /* The currently optimized loop. */ |
| 589 | class loop *current_loop; |
| 590 | location_t loop_loc; |
| 591 | |
| 592 | /* Numbers of iterations for all exits of the current loop. */ |
| 593 | hash_map<edge, tree_niter_desc *> *niters; |
| 594 | |
| 595 | /* Number of registers used in it. */ |
| 596 | unsigned regs_used; |
| 597 | |
| 598 | /* The size of version_info array allocated. */ |
| 599 | unsigned version_info_size; |
| 600 | |
| 601 | /* The array of information for the ssa names. */ |
| 602 | struct version_info *version_info; |
| 603 | |
| 604 | /* The hashtable of loop invariant expressions created |
| 605 | by ivopt. */ |
| 606 | hash_table<iv_inv_expr_hasher> *inv_expr_tab; |
| 607 | |
| 608 | /* The bitmap of indices in version_info whose value was changed. */ |
| 609 | bitmap relevant; |
| 610 | |
| 611 | /* The uses of induction variables. */ |
| 612 | vec<iv_group *> vgroups; |
| 613 | |
| 614 | /* The candidates. */ |
| 615 | vec<iv_cand *> vcands; |
| 616 | |
| 617 | /* A bitmap of important candidates. */ |
| 618 | bitmap important_candidates; |
| 619 | |
| 620 | /* Cache used by tree_to_aff_combination_expand. */ |
| 621 | hash_map<tree, name_expansion *> *name_expansion_cache; |
| 622 | |
| 623 | /* The hashtable of common candidates derived from iv uses. */ |
| 624 | hash_table<iv_common_cand_hasher> *iv_common_cand_tab; |
| 625 | |
| 626 | /* The common candidates. */ |
| 627 | vec<iv_common_cand *> iv_common_cands; |
| 628 | |
| 629 | /* Hash map recording base object information of tree exp. */ |
| 630 | hash_map<tree, tree> *base_object_map; |
| 631 | |
| 632 | /* The maximum invariant variable id. */ |
| 633 | unsigned max_inv_var_id; |
| 634 | |
| 635 | /* The maximum invariant expression id. */ |
| 636 | unsigned max_inv_expr_id; |
| 637 | |
| 638 | /* Number of no_overflow BIVs which are not used in memory address. */ |
| 639 | unsigned bivs_not_used_in_addr; |
| 640 | |
| 641 | /* Obstack for iv structure. */ |
| 642 | struct obstack iv_obstack; |
| 643 | |
| 644 | /* Whether to consider just related and important candidates when replacing a |
| 645 | use. */ |
| 646 | bool consider_all_candidates; |
| 647 | |
| 648 | /* Are we optimizing for speed? */ |
| 649 | bool speed; |
| 650 | |
| 651 | /* Whether the loop body includes any function calls. */ |
| 652 | bool body_includes_call; |
| 653 | |
| 654 | /* Whether the loop body can only be exited via single exit. */ |
| 655 | bool loop_single_exit_p; |
| 656 | |
| 657 | /* Whether the loop has doloop comparison use. */ |
| 658 | bool doloop_use_p; |
| 659 | }; |
| 660 | |
| 661 | /* An assignment of iv candidates to uses. */ |
| 662 | |
| 663 | class iv_ca |
| 664 | { |
| 665 | public: |
| 666 | /* The number of uses covered by the assignment. */ |
| 667 | unsigned upto; |
| 668 | |
| 669 | /* Number of uses that cannot be expressed by the candidates in the set. */ |
| 670 | unsigned bad_groups; |
| 671 | |
| 672 | /* Candidate assigned to a use, together with the related costs. */ |
| 673 | class cost_pair **cand_for_group; |
| 674 | |
| 675 | /* Number of times each candidate is used. */ |
| 676 | unsigned *n_cand_uses; |
| 677 | |
| 678 | /* The candidates used. */ |
| 679 | bitmap cands; |
| 680 | |
| 681 | /* The number of candidates in the set. */ |
| 682 | unsigned n_cands; |
| 683 | |
| 684 | /* The number of invariants needed, including both invariant variants and |
| 685 | invariant expressions. */ |
| 686 | unsigned n_invs; |
| 687 | |
| 688 | /* Total cost of expressing uses. */ |
| 689 | comp_cost cand_use_cost; |
| 690 | |
| 691 | /* Total cost of candidates. */ |
| 692 | int64_t cand_cost; |
| 693 | |
| 694 | /* Number of times each invariant variable is used. */ |
| 695 | unsigned *n_inv_var_uses; |
| 696 | |
| 697 | /* Number of times each invariant expression is used. */ |
| 698 | unsigned *n_inv_expr_uses; |
| 699 | |
| 700 | /* Total cost of the assignment. */ |
| 701 | comp_cost cost; |
| 702 | }; |
| 703 | |
| 704 | /* Difference of two iv candidate assignments. */ |
| 705 | |
| 706 | struct iv_ca_delta |
| 707 | { |
| 708 | /* Changed group. */ |
| 709 | struct iv_group *group; |
| 710 | |
| 711 | /* An old assignment (for rollback purposes). */ |
| 712 | class cost_pair *old_cp; |
| 713 | |
| 714 | /* A new assignment. */ |
| 715 | class cost_pair *new_cp; |
| 716 | |
| 717 | /* Next change in the list. */ |
| 718 | struct iv_ca_delta *next; |
| 719 | }; |
| 720 | |
| 721 | /* Bound on number of candidates below that all candidates are considered. */ |
| 722 | |
| 723 | #define CONSIDER_ALL_CANDIDATES_BOUND \ |
| 724 | ((unsigned) param_iv_consider_all_candidates_bound) |
| 725 | |
| 726 | /* If there are more iv occurrences, we just give up (it is quite unlikely that |
| 727 | optimizing such a loop would help, and it would take ages). */ |
| 728 | |
| 729 | #define MAX_CONSIDERED_GROUPS \ |
| 730 | ((unsigned) param_iv_max_considered_uses) |
| 731 | |
| 732 | /* If there are at most this number of ivs in the set, try removing unnecessary |
| 733 | ivs from the set always. */ |
| 734 | |
| 735 | #define ALWAYS_PRUNE_CAND_SET_BOUND \ |
| 736 | ((unsigned) param_iv_always_prune_cand_set_bound) |
| 737 | |
| 738 | /* The list of trees for that the decl_rtl field must be reset is stored |
| 739 | here. */ |
| 740 | |
| 741 | static vec<tree> decl_rtl_to_reset; |
| 742 | |
| 743 | static comp_cost force_expr_to_var_cost (tree, bool); |
| 744 | |
| 745 | /* The single loop exit if it dominates the latch, NULL otherwise. */ |
| 746 | |
| 747 | edge |
| 748 | single_dom_exit (class loop *loop) |
| 749 | { |
| 750 | edge exit = single_exit (loop); |
| 751 | |
| 752 | if (!exit) |
| 753 | return NULL; |
| 754 | |
| 755 | if (!just_once_each_iteration_p (loop, exit->src)) |
| 756 | return NULL; |
| 757 | |
| 758 | return exit; |
| 759 | } |
| 760 | |
| 761 | /* Dumps information about the induction variable IV to FILE. Don't dump |
| 762 | variable's name if DUMP_NAME is FALSE. The information is dumped with |
| 763 | preceding spaces indicated by INDENT_LEVEL. */ |
| 764 | |
| 765 | void |
| 766 | dump_iv (FILE *file, struct iv *iv, bool dump_name, unsigned indent_level) |
| 767 | { |
| 768 | const char *p; |
| 769 | const char spaces[9] = {' ', ' ', ' ', ' ', ' ', ' ', ' ', ' ', '\0'}; |
| 770 | |
| 771 | if (indent_level > 4) |
| 772 | indent_level = 4; |
| 773 | p = spaces + 8 - (indent_level << 1); |
| 774 | |
| 775 | fprintf (stream: file, format: "%sIV struct:\n" , p); |
| 776 | if (iv->ssa_name && dump_name) |
| 777 | { |
| 778 | fprintf (stream: file, format: "%s SSA_NAME:\t" , p); |
| 779 | print_generic_expr (file, iv->ssa_name, TDF_SLIM); |
| 780 | fprintf (stream: file, format: "\n" ); |
| 781 | } |
| 782 | |
| 783 | fprintf (stream: file, format: "%s Type:\t" , p); |
| 784 | print_generic_expr (file, TREE_TYPE (iv->base), TDF_SLIM); |
| 785 | fprintf (stream: file, format: "\n" ); |
| 786 | |
| 787 | fprintf (stream: file, format: "%s Base:\t" , p); |
| 788 | print_generic_expr (file, iv->base, TDF_SLIM); |
| 789 | fprintf (stream: file, format: "\n" ); |
| 790 | |
| 791 | fprintf (stream: file, format: "%s Step:\t" , p); |
| 792 | print_generic_expr (file, iv->step, TDF_SLIM); |
| 793 | fprintf (stream: file, format: "\n" ); |
| 794 | |
| 795 | if (iv->base_object) |
| 796 | { |
| 797 | fprintf (stream: file, format: "%s Object:\t" , p); |
| 798 | print_generic_expr (file, iv->base_object, TDF_SLIM); |
| 799 | fprintf (stream: file, format: "\n" ); |
| 800 | } |
| 801 | |
| 802 | fprintf (stream: file, format: "%s Biv:\t%c\n" , p, iv->biv_p ? 'Y' : 'N'); |
| 803 | |
| 804 | fprintf (stream: file, format: "%s Overflowness wrto loop niter:\t%s\n" , |
| 805 | p, iv->no_overflow ? "No-overflow" : "Overflow" ); |
| 806 | } |
| 807 | |
| 808 | /* Dumps information about the USE to FILE. */ |
| 809 | |
| 810 | void |
| 811 | dump_use (FILE *file, struct iv_use *use) |
| 812 | { |
| 813 | fprintf (stream: file, format: " Use %d.%d:\n" , use->group_id, use->id); |
| 814 | fprintf (stream: file, format: " At stmt:\t" ); |
| 815 | print_gimple_stmt (file, use->stmt, 0); |
| 816 | fprintf (stream: file, format: " At pos:\t" ); |
| 817 | if (use->op_p) |
| 818 | print_generic_expr (file, *use->op_p, TDF_SLIM); |
| 819 | fprintf (stream: file, format: "\n" ); |
| 820 | dump_iv (file, iv: use->iv, dump_name: false, indent_level: 2); |
| 821 | } |
| 822 | |
| 823 | /* Dumps information about the uses to FILE. */ |
| 824 | |
| 825 | void |
| 826 | dump_groups (FILE *file, struct ivopts_data *data) |
| 827 | { |
| 828 | unsigned i, j; |
| 829 | struct iv_group *group; |
| 830 | |
| 831 | for (i = 0; i < data->vgroups.length (); i++) |
| 832 | { |
| 833 | group = data->vgroups[i]; |
| 834 | fprintf (stream: file, format: "Group %d:\n" , group->id); |
| 835 | if (group->type == USE_NONLINEAR_EXPR) |
| 836 | fprintf (stream: file, format: " Type:\tGENERIC\n" ); |
| 837 | else if (group->type == USE_REF_ADDRESS) |
| 838 | fprintf (stream: file, format: " Type:\tREFERENCE ADDRESS\n" ); |
| 839 | else if (group->type == USE_PTR_ADDRESS) |
| 840 | fprintf (stream: file, format: " Type:\tPOINTER ARGUMENT ADDRESS\n" ); |
| 841 | else |
| 842 | { |
| 843 | gcc_assert (group->type == USE_COMPARE); |
| 844 | fprintf (stream: file, format: " Type:\tCOMPARE\n" ); |
| 845 | } |
| 846 | for (j = 0; j < group->vuses.length (); j++) |
| 847 | dump_use (file, use: group->vuses[j]); |
| 848 | } |
| 849 | } |
| 850 | |
| 851 | /* Dumps information about induction variable candidate CAND to FILE. */ |
| 852 | |
| 853 | void |
| 854 | dump_cand (FILE *file, struct iv_cand *cand) |
| 855 | { |
| 856 | struct iv *iv = cand->iv; |
| 857 | |
| 858 | fprintf (stream: file, format: "Candidate %d:\n" , cand->id); |
| 859 | if (cand->inv_vars) |
| 860 | { |
| 861 | fprintf (stream: file, format: " Depend on inv.vars: " ); |
| 862 | dump_bitmap (file, map: cand->inv_vars); |
| 863 | } |
| 864 | if (cand->inv_exprs) |
| 865 | { |
| 866 | fprintf (stream: file, format: " Depend on inv.exprs: " ); |
| 867 | dump_bitmap (file, map: cand->inv_exprs); |
| 868 | } |
| 869 | |
| 870 | if (cand->var_before) |
| 871 | { |
| 872 | fprintf (stream: file, format: " Var befor: " ); |
| 873 | print_generic_expr (file, cand->var_before, TDF_SLIM); |
| 874 | fprintf (stream: file, format: "\n" ); |
| 875 | } |
| 876 | if (cand->var_after) |
| 877 | { |
| 878 | fprintf (stream: file, format: " Var after: " ); |
| 879 | print_generic_expr (file, cand->var_after, TDF_SLIM); |
| 880 | fprintf (stream: file, format: "\n" ); |
| 881 | } |
| 882 | |
| 883 | switch (cand->pos) |
| 884 | { |
| 885 | case IP_NORMAL: |
| 886 | fprintf (stream: file, format: " Incr POS: before exit test\n" ); |
| 887 | break; |
| 888 | |
| 889 | case IP_BEFORE_USE: |
| 890 | fprintf (stream: file, format: " Incr POS: before use %d\n" , cand->ainc_use->id); |
| 891 | break; |
| 892 | |
| 893 | case IP_AFTER_USE: |
| 894 | fprintf (stream: file, format: " Incr POS: after use %d\n" , cand->ainc_use->id); |
| 895 | break; |
| 896 | |
| 897 | case IP_END: |
| 898 | fprintf (stream: file, format: " Incr POS: at end\n" ); |
| 899 | break; |
| 900 | |
| 901 | case IP_ORIGINAL: |
| 902 | fprintf (stream: file, format: " Incr POS: orig biv\n" ); |
| 903 | break; |
| 904 | } |
| 905 | |
| 906 | dump_iv (file, iv, dump_name: false, indent_level: 1); |
| 907 | } |
| 908 | |
| 909 | /* Returns the info for ssa version VER. */ |
| 910 | |
| 911 | static inline struct version_info * |
| 912 | ver_info (struct ivopts_data *data, unsigned ver) |
| 913 | { |
| 914 | return data->version_info + ver; |
| 915 | } |
| 916 | |
| 917 | /* Returns the info for ssa name NAME. */ |
| 918 | |
| 919 | static inline struct version_info * |
| 920 | name_info (struct ivopts_data *data, tree name) |
| 921 | { |
| 922 | return ver_info (data, SSA_NAME_VERSION (name)); |
| 923 | } |
| 924 | |
| 925 | /* Returns true if STMT is after the place where the IP_NORMAL ivs will be |
| 926 | emitted in LOOP. */ |
| 927 | |
| 928 | static bool |
| 929 | stmt_after_ip_normal_pos (class loop *loop, gimple *stmt) |
| 930 | { |
| 931 | basic_block bb = ip_normal_pos (loop), sbb = gimple_bb (g: stmt); |
| 932 | |
| 933 | gcc_assert (bb); |
| 934 | |
| 935 | if (sbb == loop->latch) |
| 936 | return true; |
| 937 | |
| 938 | if (sbb != bb) |
| 939 | return false; |
| 940 | |
| 941 | return stmt == last_nondebug_stmt (bb); |
| 942 | } |
| 943 | |
| 944 | /* Returns true if STMT if after the place where the original induction |
| 945 | variable CAND is incremented. If TRUE_IF_EQUAL is set, we return true |
| 946 | if the positions are identical. */ |
| 947 | |
| 948 | static bool |
| 949 | stmt_after_inc_pos (struct iv_cand *cand, gimple *stmt, bool true_if_equal) |
| 950 | { |
| 951 | basic_block cand_bb = gimple_bb (g: cand->incremented_at); |
| 952 | basic_block stmt_bb = gimple_bb (g: stmt); |
| 953 | |
| 954 | if (!dominated_by_p (CDI_DOMINATORS, stmt_bb, cand_bb)) |
| 955 | return false; |
| 956 | |
| 957 | if (stmt_bb != cand_bb) |
| 958 | return true; |
| 959 | |
| 960 | if (true_if_equal |
| 961 | && gimple_uid (g: stmt) == gimple_uid (g: cand->incremented_at)) |
| 962 | return true; |
| 963 | return gimple_uid (g: stmt) > gimple_uid (g: cand->incremented_at); |
| 964 | } |
| 965 | |
| 966 | /* Returns true if STMT if after the place where the induction variable |
| 967 | CAND is incremented in LOOP. */ |
| 968 | |
| 969 | static bool |
| 970 | stmt_after_increment (class loop *loop, struct iv_cand *cand, gimple *stmt) |
| 971 | { |
| 972 | switch (cand->pos) |
| 973 | { |
| 974 | case IP_END: |
| 975 | return false; |
| 976 | |
| 977 | case IP_NORMAL: |
| 978 | return stmt_after_ip_normal_pos (loop, stmt); |
| 979 | |
| 980 | case IP_ORIGINAL: |
| 981 | case IP_AFTER_USE: |
| 982 | return stmt_after_inc_pos (cand, stmt, true_if_equal: false); |
| 983 | |
| 984 | case IP_BEFORE_USE: |
| 985 | return stmt_after_inc_pos (cand, stmt, true_if_equal: true); |
| 986 | |
| 987 | default: |
| 988 | gcc_unreachable (); |
| 989 | } |
| 990 | } |
| 991 | |
| 992 | /* walk_tree callback for contains_abnormal_ssa_name_p. */ |
| 993 | |
| 994 | static tree |
| 995 | contains_abnormal_ssa_name_p_1 (tree *tp, int *walk_subtrees, void *) |
| 996 | { |
| 997 | if (TREE_CODE (*tp) == SSA_NAME |
| 998 | && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (*tp)) |
| 999 | return *tp; |
| 1000 | |
| 1001 | if (!EXPR_P (*tp)) |
| 1002 | *walk_subtrees = 0; |
| 1003 | |
| 1004 | return NULL_TREE; |
| 1005 | } |
| 1006 | |
| 1007 | /* Returns true if EXPR contains a ssa name that occurs in an |
| 1008 | abnormal phi node. */ |
| 1009 | |
| 1010 | bool |
| 1011 | contains_abnormal_ssa_name_p (tree expr) |
| 1012 | { |
| 1013 | return walk_tree_without_duplicates |
| 1014 | (&expr, contains_abnormal_ssa_name_p_1, NULL) != NULL_TREE; |
| 1015 | } |
| 1016 | |
| 1017 | /* Returns the structure describing number of iterations determined from |
| 1018 | EXIT of DATA->current_loop, or NULL if something goes wrong. */ |
| 1019 | |
| 1020 | static class tree_niter_desc * |
| 1021 | niter_for_exit (struct ivopts_data *data, edge exit) |
| 1022 | { |
| 1023 | class tree_niter_desc *desc; |
| 1024 | tree_niter_desc **slot; |
| 1025 | |
| 1026 | if (!data->niters) |
| 1027 | { |
| 1028 | data->niters = new hash_map<edge, tree_niter_desc *>; |
| 1029 | slot = NULL; |
| 1030 | } |
| 1031 | else |
| 1032 | slot = data->niters->get (k: exit); |
| 1033 | |
| 1034 | if (!slot) |
| 1035 | { |
| 1036 | /* Try to determine number of iterations. We cannot safely work with ssa |
| 1037 | names that appear in phi nodes on abnormal edges, so that we do not |
| 1038 | create overlapping life ranges for them (PR 27283). */ |
| 1039 | desc = XNEW (class tree_niter_desc); |
| 1040 | ::new (static_cast<void*> (desc)) tree_niter_desc (); |
| 1041 | if (!number_of_iterations_exit (data->current_loop, |
| 1042 | exit, niter: desc, true) |
| 1043 | || contains_abnormal_ssa_name_p (expr: desc->niter)) |
| 1044 | { |
| 1045 | desc->~tree_niter_desc (); |
| 1046 | XDELETE (desc); |
| 1047 | desc = NULL; |
| 1048 | } |
| 1049 | data->niters->put (k: exit, v: desc); |
| 1050 | } |
| 1051 | else |
| 1052 | desc = *slot; |
| 1053 | |
| 1054 | return desc; |
| 1055 | } |
| 1056 | |
| 1057 | /* Returns the structure describing number of iterations determined from |
| 1058 | single dominating exit of DATA->current_loop, or NULL if something |
| 1059 | goes wrong. */ |
| 1060 | |
| 1061 | static class tree_niter_desc * |
| 1062 | niter_for_single_dom_exit (struct ivopts_data *data) |
| 1063 | { |
| 1064 | edge exit = single_dom_exit (loop: data->current_loop); |
| 1065 | |
| 1066 | if (!exit) |
| 1067 | return NULL; |
| 1068 | |
| 1069 | return niter_for_exit (data, exit); |
| 1070 | } |
| 1071 | |
| 1072 | /* Initializes data structures used by the iv optimization pass, stored |
| 1073 | in DATA. */ |
| 1074 | |
| 1075 | static void |
| 1076 | tree_ssa_iv_optimize_init (struct ivopts_data *data) |
| 1077 | { |
| 1078 | data->version_info_size = 2 * num_ssa_names; |
| 1079 | data->version_info = XCNEWVEC (struct version_info, data->version_info_size); |
| 1080 | data->relevant = BITMAP_ALLOC (NULL); |
| 1081 | data->important_candidates = BITMAP_ALLOC (NULL); |
| 1082 | data->max_inv_var_id = 0; |
| 1083 | data->max_inv_expr_id = 0; |
| 1084 | data->niters = NULL; |
| 1085 | data->vgroups.create (nelems: 20); |
| 1086 | data->vcands.create (nelems: 20); |
| 1087 | data->inv_expr_tab = new hash_table<iv_inv_expr_hasher> (10); |
| 1088 | data->name_expansion_cache = NULL; |
| 1089 | data->base_object_map = NULL; |
| 1090 | data->iv_common_cand_tab = new hash_table<iv_common_cand_hasher> (10); |
| 1091 | data->iv_common_cands.create (nelems: 20); |
| 1092 | decl_rtl_to_reset.create (nelems: 20); |
| 1093 | gcc_obstack_init (&data->iv_obstack); |
| 1094 | } |
| 1095 | |
| 1096 | /* walk_tree callback for determine_base_object. */ |
| 1097 | |
| 1098 | static tree |
| 1099 | determine_base_object_1 (tree *tp, int *walk_subtrees, void *wdata) |
| 1100 | { |
| 1101 | tree_code code = TREE_CODE (*tp); |
| 1102 | tree obj = NULL_TREE; |
| 1103 | if (code == ADDR_EXPR) |
| 1104 | { |
| 1105 | tree base = get_base_address (TREE_OPERAND (*tp, 0)); |
| 1106 | if (!base) |
| 1107 | obj = *tp; |
| 1108 | else if (TREE_CODE (base) != MEM_REF) |
| 1109 | obj = fold_convert (ptr_type_node, build_fold_addr_expr (base)); |
| 1110 | } |
| 1111 | else if (code == SSA_NAME && POINTER_TYPE_P (TREE_TYPE (*tp))) |
| 1112 | obj = fold_convert (ptr_type_node, *tp); |
| 1113 | |
| 1114 | if (!obj) |
| 1115 | { |
| 1116 | if (!EXPR_P (*tp)) |
| 1117 | *walk_subtrees = 0; |
| 1118 | |
| 1119 | return NULL_TREE; |
| 1120 | } |
| 1121 | /* Record special node for multiple base objects and stop. */ |
| 1122 | if (*static_cast<tree *> (wdata)) |
| 1123 | { |
| 1124 | *static_cast<tree *> (wdata) = integer_zero_node; |
| 1125 | return integer_zero_node; |
| 1126 | } |
| 1127 | /* Record the base object and continue looking. */ |
| 1128 | *static_cast<tree *> (wdata) = obj; |
| 1129 | return NULL_TREE; |
| 1130 | } |
| 1131 | |
| 1132 | /* Returns a memory object to that EXPR points with caching. Return NULL if we |
| 1133 | are able to determine that it does not point to any such object; specially |
| 1134 | return integer_zero_node if EXPR contains multiple base objects. */ |
| 1135 | |
| 1136 | static tree |
| 1137 | determine_base_object (struct ivopts_data *data, tree expr) |
| 1138 | { |
| 1139 | tree *slot, obj = NULL_TREE; |
| 1140 | if (data->base_object_map) |
| 1141 | { |
| 1142 | if ((slot = data->base_object_map->get(k: expr)) != NULL) |
| 1143 | return *slot; |
| 1144 | } |
| 1145 | else |
| 1146 | data->base_object_map = new hash_map<tree, tree>; |
| 1147 | |
| 1148 | (void) walk_tree_without_duplicates (&expr, determine_base_object_1, &obj); |
| 1149 | data->base_object_map->put (k: expr, v: obj); |
| 1150 | return obj; |
| 1151 | } |
| 1152 | |
| 1153 | /* Allocates an induction variable with given initial value BASE and step STEP |
| 1154 | for loop LOOP. NO_OVERFLOW implies the iv doesn't overflow. */ |
| 1155 | |
| 1156 | static struct iv * |
| 1157 | alloc_iv (struct ivopts_data *data, tree base, tree step, |
| 1158 | bool no_overflow = false) |
| 1159 | { |
| 1160 | tree expr = base; |
| 1161 | struct iv *iv = (struct iv*) obstack_alloc (&data->iv_obstack, |
| 1162 | sizeof (struct iv)); |
| 1163 | gcc_assert (step != NULL_TREE); |
| 1164 | |
| 1165 | /* Canonicalize the address expression in base if it were an unsigned |
| 1166 | computation. That leads to more equalities being detected and results in: |
| 1167 | |
| 1168 | 1) More accurate cost can be computed for address expressions; |
| 1169 | 2) Duplicate candidates won't be created for bases in different |
| 1170 | forms, like &a[0] and &a. |
| 1171 | 3) Duplicate candidates won't be created for IV expressions that differ |
| 1172 | only in their sign. */ |
| 1173 | aff_tree comb; |
| 1174 | STRIP_NOPS (expr); |
| 1175 | expr = fold_convert (unsigned_type_for (TREE_TYPE (expr)), expr); |
| 1176 | tree_to_aff_combination (expr, TREE_TYPE (expr), &comb); |
| 1177 | base = fold_convert (TREE_TYPE (base), aff_combination_to_tree (&comb)); |
| 1178 | |
| 1179 | iv->base = base; |
| 1180 | iv->base_object = determine_base_object (data, expr: base); |
| 1181 | iv->step = step; |
| 1182 | iv->biv_p = false; |
| 1183 | iv->nonlin_use = NULL; |
| 1184 | iv->ssa_name = NULL_TREE; |
| 1185 | if (!no_overflow |
| 1186 | && !iv_can_overflow_p (data->current_loop, TREE_TYPE (base), |
| 1187 | base, step)) |
| 1188 | no_overflow = true; |
| 1189 | iv->no_overflow = no_overflow; |
| 1190 | iv->have_address_use = false; |
| 1191 | |
| 1192 | return iv; |
| 1193 | } |
| 1194 | |
| 1195 | /* Sets STEP and BASE for induction variable IV. NO_OVERFLOW implies the IV |
| 1196 | doesn't overflow. */ |
| 1197 | |
| 1198 | static void |
| 1199 | set_iv (struct ivopts_data *data, tree iv, tree base, tree step, |
| 1200 | bool no_overflow) |
| 1201 | { |
| 1202 | struct version_info *info = name_info (data, name: iv); |
| 1203 | |
| 1204 | gcc_assert (!info->iv); |
| 1205 | |
| 1206 | bitmap_set_bit (data->relevant, SSA_NAME_VERSION (iv)); |
| 1207 | info->iv = alloc_iv (data, base, step, no_overflow); |
| 1208 | info->iv->ssa_name = iv; |
| 1209 | } |
| 1210 | |
| 1211 | /* Finds induction variable declaration for VAR. */ |
| 1212 | |
| 1213 | static struct iv * |
| 1214 | get_iv (struct ivopts_data *data, tree var) |
| 1215 | { |
| 1216 | basic_block bb; |
| 1217 | tree type = TREE_TYPE (var); |
| 1218 | |
| 1219 | if (!POINTER_TYPE_P (type) |
| 1220 | && !INTEGRAL_TYPE_P (type)) |
| 1221 | return NULL; |
| 1222 | |
| 1223 | if (!name_info (data, name: var)->iv) |
| 1224 | { |
| 1225 | bb = gimple_bb (SSA_NAME_DEF_STMT (var)); |
| 1226 | |
| 1227 | if (!bb |
| 1228 | || !flow_bb_inside_loop_p (data->current_loop, bb)) |
| 1229 | { |
| 1230 | if (POINTER_TYPE_P (type)) |
| 1231 | type = sizetype; |
| 1232 | set_iv (data, iv: var, base: var, step: build_int_cst (type, 0), no_overflow: true); |
| 1233 | } |
| 1234 | } |
| 1235 | |
| 1236 | return name_info (data, name: var)->iv; |
| 1237 | } |
| 1238 | |
| 1239 | /* Return the first non-invariant ssa var found in EXPR. */ |
| 1240 | |
| 1241 | static tree |
| 1242 | (tree expr) |
| 1243 | { |
| 1244 | int i, n; |
| 1245 | tree tmp; |
| 1246 | enum tree_code code; |
| 1247 | |
| 1248 | if (!expr || is_gimple_min_invariant (expr)) |
| 1249 | return NULL; |
| 1250 | |
| 1251 | code = TREE_CODE (expr); |
| 1252 | if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code))) |
| 1253 | { |
| 1254 | n = TREE_OPERAND_LENGTH (expr); |
| 1255 | for (i = 0; i < n; i++) |
| 1256 | { |
| 1257 | tmp = extract_single_var_from_expr (TREE_OPERAND (expr, i)); |
| 1258 | |
| 1259 | if (tmp) |
| 1260 | return tmp; |
| 1261 | } |
| 1262 | } |
| 1263 | return (TREE_CODE (expr) == SSA_NAME) ? expr : NULL; |
| 1264 | } |
| 1265 | |
| 1266 | /* Finds basic ivs. */ |
| 1267 | |
| 1268 | static bool |
| 1269 | find_bivs (struct ivopts_data *data) |
| 1270 | { |
| 1271 | gphi *phi; |
| 1272 | affine_iv iv; |
| 1273 | tree step, type, base, stop; |
| 1274 | bool found = false; |
| 1275 | class loop *loop = data->current_loop; |
| 1276 | gphi_iterator psi; |
| 1277 | |
| 1278 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
| 1279 | { |
| 1280 | phi = psi.phi (); |
| 1281 | |
| 1282 | if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (phi))) |
| 1283 | continue; |
| 1284 | |
| 1285 | if (virtual_operand_p (PHI_RESULT (phi))) |
| 1286 | continue; |
| 1287 | |
| 1288 | if (!simple_iv (loop, loop, PHI_RESULT (phi), &iv, true)) |
| 1289 | continue; |
| 1290 | |
| 1291 | if (integer_zerop (iv.step)) |
| 1292 | continue; |
| 1293 | |
| 1294 | step = iv.step; |
| 1295 | base = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
| 1296 | /* Stop expanding iv base at the first ssa var referred by iv step. |
| 1297 | Ideally we should stop at any ssa var, because that's expensive |
| 1298 | and unusual to happen, we just do it on the first one. |
| 1299 | |
| 1300 | See PR64705 for the rationale. */ |
| 1301 | stop = extract_single_var_from_expr (expr: step); |
| 1302 | base = expand_simple_operations (base, stop); |
| 1303 | if (contains_abnormal_ssa_name_p (expr: base) |
| 1304 | || contains_abnormal_ssa_name_p (expr: step)) |
| 1305 | continue; |
| 1306 | |
| 1307 | type = TREE_TYPE (PHI_RESULT (phi)); |
| 1308 | base = fold_convert (type, base); |
| 1309 | if (step) |
| 1310 | { |
| 1311 | if (POINTER_TYPE_P (type)) |
| 1312 | step = convert_to_ptrofftype (step); |
| 1313 | else |
| 1314 | step = fold_convert (type, step); |
| 1315 | } |
| 1316 | |
| 1317 | set_iv (data, PHI_RESULT (phi), base, step, no_overflow: iv.no_overflow); |
| 1318 | found = true; |
| 1319 | } |
| 1320 | |
| 1321 | return found; |
| 1322 | } |
| 1323 | |
| 1324 | /* Marks basic ivs. */ |
| 1325 | |
| 1326 | static void |
| 1327 | mark_bivs (struct ivopts_data *data) |
| 1328 | { |
| 1329 | gphi *phi; |
| 1330 | gimple *def; |
| 1331 | tree var; |
| 1332 | struct iv *iv, *incr_iv; |
| 1333 | class loop *loop = data->current_loop; |
| 1334 | basic_block incr_bb; |
| 1335 | gphi_iterator psi; |
| 1336 | |
| 1337 | data->bivs_not_used_in_addr = 0; |
| 1338 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
| 1339 | { |
| 1340 | phi = psi.phi (); |
| 1341 | |
| 1342 | iv = get_iv (data, PHI_RESULT (phi)); |
| 1343 | if (!iv) |
| 1344 | continue; |
| 1345 | |
| 1346 | var = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); |
| 1347 | def = SSA_NAME_DEF_STMT (var); |
| 1348 | /* Don't mark iv peeled from other one as biv. */ |
| 1349 | if (def |
| 1350 | && gimple_code (g: def) == GIMPLE_PHI |
| 1351 | && gimple_bb (g: def) == loop->header) |
| 1352 | continue; |
| 1353 | |
| 1354 | incr_iv = get_iv (data, var); |
| 1355 | if (!incr_iv) |
| 1356 | continue; |
| 1357 | |
| 1358 | /* If the increment is in the subloop, ignore it. */ |
| 1359 | incr_bb = gimple_bb (SSA_NAME_DEF_STMT (var)); |
| 1360 | if (incr_bb->loop_father != data->current_loop |
| 1361 | || (incr_bb->flags & BB_IRREDUCIBLE_LOOP)) |
| 1362 | continue; |
| 1363 | |
| 1364 | iv->biv_p = true; |
| 1365 | incr_iv->biv_p = true; |
| 1366 | if (iv->no_overflow) |
| 1367 | data->bivs_not_used_in_addr++; |
| 1368 | if (incr_iv->no_overflow) |
| 1369 | data->bivs_not_used_in_addr++; |
| 1370 | } |
| 1371 | } |
| 1372 | |
| 1373 | /* Checks whether STMT defines a linear induction variable and stores its |
| 1374 | parameters to IV. */ |
| 1375 | |
| 1376 | static bool |
| 1377 | find_givs_in_stmt_scev (struct ivopts_data *data, gimple *stmt, affine_iv *iv) |
| 1378 | { |
| 1379 | tree lhs, stop; |
| 1380 | class loop *loop = data->current_loop; |
| 1381 | |
| 1382 | iv->base = NULL_TREE; |
| 1383 | iv->step = NULL_TREE; |
| 1384 | |
| 1385 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
| 1386 | return false; |
| 1387 | |
| 1388 | lhs = gimple_assign_lhs (gs: stmt); |
| 1389 | if (TREE_CODE (lhs) != SSA_NAME) |
| 1390 | return false; |
| 1391 | |
| 1392 | if (!simple_iv (loop, loop_containing_stmt (stmt), lhs, iv, true)) |
| 1393 | return false; |
| 1394 | |
| 1395 | /* Stop expanding iv base at the first ssa var referred by iv step. |
| 1396 | Ideally we should stop at any ssa var, because that's expensive |
| 1397 | and unusual to happen, we just do it on the first one. |
| 1398 | |
| 1399 | See PR64705 for the rationale. */ |
| 1400 | stop = extract_single_var_from_expr (expr: iv->step); |
| 1401 | iv->base = expand_simple_operations (iv->base, stop); |
| 1402 | if (contains_abnormal_ssa_name_p (expr: iv->base) |
| 1403 | || contains_abnormal_ssa_name_p (expr: iv->step)) |
| 1404 | return false; |
| 1405 | |
| 1406 | /* If STMT could throw, then do not consider STMT as defining a GIV. |
| 1407 | While this will suppress optimizations, we cannot safely delete this |
| 1408 | GIV and associated statements, even if it appears it is not used. */ |
| 1409 | if (stmt_could_throw_p (cfun, stmt)) |
| 1410 | return false; |
| 1411 | |
| 1412 | return true; |
| 1413 | } |
| 1414 | |
| 1415 | /* Finds general ivs in statement STMT. */ |
| 1416 | |
| 1417 | static void |
| 1418 | find_givs_in_stmt (struct ivopts_data *data, gimple *stmt) |
| 1419 | { |
| 1420 | affine_iv iv; |
| 1421 | |
| 1422 | if (!find_givs_in_stmt_scev (data, stmt, iv: &iv)) |
| 1423 | return; |
| 1424 | |
| 1425 | set_iv (data, iv: gimple_assign_lhs (gs: stmt), base: iv.base, step: iv.step, no_overflow: iv.no_overflow); |
| 1426 | } |
| 1427 | |
| 1428 | /* Finds general ivs in basic block BB. */ |
| 1429 | |
| 1430 | static void |
| 1431 | find_givs_in_bb (struct ivopts_data *data, basic_block bb) |
| 1432 | { |
| 1433 | gimple_stmt_iterator bsi; |
| 1434 | |
| 1435 | for (bsi = gsi_start_bb (bb); !gsi_end_p (i: bsi); gsi_next (i: &bsi)) |
| 1436 | if (!is_gimple_debug (gs: gsi_stmt (i: bsi))) |
| 1437 | find_givs_in_stmt (data, stmt: gsi_stmt (i: bsi)); |
| 1438 | } |
| 1439 | |
| 1440 | /* Finds general ivs. */ |
| 1441 | |
| 1442 | static void |
| 1443 | find_givs (struct ivopts_data *data, basic_block *body) |
| 1444 | { |
| 1445 | class loop *loop = data->current_loop; |
| 1446 | unsigned i; |
| 1447 | |
| 1448 | for (i = 0; i < loop->num_nodes; i++) |
| 1449 | find_givs_in_bb (data, bb: body[i]); |
| 1450 | } |
| 1451 | |
| 1452 | /* For each ssa name defined in LOOP determines whether it is an induction |
| 1453 | variable and if so, its initial value and step. */ |
| 1454 | |
| 1455 | static bool |
| 1456 | find_induction_variables (struct ivopts_data *data, basic_block *body) |
| 1457 | { |
| 1458 | unsigned i; |
| 1459 | bitmap_iterator bi; |
| 1460 | |
| 1461 | if (!find_bivs (data)) |
| 1462 | return false; |
| 1463 | |
| 1464 | find_givs (data, body); |
| 1465 | mark_bivs (data); |
| 1466 | |
| 1467 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 1468 | { |
| 1469 | class tree_niter_desc *niter = niter_for_single_dom_exit (data); |
| 1470 | |
| 1471 | if (niter) |
| 1472 | { |
| 1473 | fprintf (stream: dump_file, format: " number of iterations " ); |
| 1474 | print_generic_expr (dump_file, niter->niter, TDF_SLIM); |
| 1475 | if (!integer_zerop (niter->may_be_zero)) |
| 1476 | { |
| 1477 | fprintf (stream: dump_file, format: "; zero if " ); |
| 1478 | print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM); |
| 1479 | } |
| 1480 | fprintf (stream: dump_file, format: "\n" ); |
| 1481 | }; |
| 1482 | |
| 1483 | fprintf (stream: dump_file, format: "\n<Induction Vars>:\n" ); |
| 1484 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
| 1485 | { |
| 1486 | struct version_info *info = ver_info (data, ver: i); |
| 1487 | if (info->iv && info->iv->step && !integer_zerop (info->iv->step)) |
| 1488 | dump_iv (file: dump_file, iv: ver_info (data, ver: i)->iv, dump_name: true, indent_level: 0); |
| 1489 | } |
| 1490 | } |
| 1491 | |
| 1492 | return true; |
| 1493 | } |
| 1494 | |
| 1495 | /* Records a use of TYPE at *USE_P in STMT whose value is IV in GROUP. |
| 1496 | For address type use, ADDR_BASE is the stripped IV base, ADDR_OFFSET |
| 1497 | is the const offset stripped from IV base and MEM_TYPE is the type |
| 1498 | of the memory being addressed. For uses of other types, ADDR_BASE |
| 1499 | and ADDR_OFFSET are zero by default and MEM_TYPE is NULL_TREE. */ |
| 1500 | |
| 1501 | static struct iv_use * |
| 1502 | record_use (struct iv_group *group, tree *use_p, struct iv *iv, |
| 1503 | gimple *stmt, enum use_type type, tree mem_type, |
| 1504 | tree addr_base, poly_uint64 addr_offset) |
| 1505 | { |
| 1506 | struct iv_use *use = XCNEW (struct iv_use); |
| 1507 | |
| 1508 | use->id = group->vuses.length (); |
| 1509 | use->group_id = group->id; |
| 1510 | use->type = type; |
| 1511 | use->mem_type = mem_type; |
| 1512 | use->iv = iv; |
| 1513 | use->stmt = stmt; |
| 1514 | use->op_p = use_p; |
| 1515 | use->addr_base = addr_base; |
| 1516 | use->addr_offset = addr_offset; |
| 1517 | |
| 1518 | group->vuses.safe_push (obj: use); |
| 1519 | return use; |
| 1520 | } |
| 1521 | |
| 1522 | /* Checks whether OP is a loop-level invariant and if so, records it. |
| 1523 | NONLINEAR_USE is true if the invariant is used in a way we do not |
| 1524 | handle specially. */ |
| 1525 | |
| 1526 | static void |
| 1527 | record_invariant (struct ivopts_data *data, tree op, bool nonlinear_use) |
| 1528 | { |
| 1529 | basic_block bb; |
| 1530 | struct version_info *info; |
| 1531 | |
| 1532 | if (TREE_CODE (op) != SSA_NAME |
| 1533 | || virtual_operand_p (op)) |
| 1534 | return; |
| 1535 | |
| 1536 | bb = gimple_bb (SSA_NAME_DEF_STMT (op)); |
| 1537 | if (bb |
| 1538 | && flow_bb_inside_loop_p (data->current_loop, bb)) |
| 1539 | return; |
| 1540 | |
| 1541 | info = name_info (data, name: op); |
| 1542 | info->name = op; |
| 1543 | info->has_nonlin_use |= nonlinear_use; |
| 1544 | if (!info->inv_id) |
| 1545 | info->inv_id = ++data->max_inv_var_id; |
| 1546 | bitmap_set_bit (data->relevant, SSA_NAME_VERSION (op)); |
| 1547 | } |
| 1548 | |
| 1549 | /* Record a group of TYPE. */ |
| 1550 | |
| 1551 | static struct iv_group * |
| 1552 | record_group (struct ivopts_data *data, enum use_type type) |
| 1553 | { |
| 1554 | struct iv_group *group = XCNEW (struct iv_group); |
| 1555 | |
| 1556 | group->id = data->vgroups.length (); |
| 1557 | group->type = type; |
| 1558 | group->related_cands = BITMAP_ALLOC (NULL); |
| 1559 | group->vuses.create (nelems: 1); |
| 1560 | group->doloop_p = false; |
| 1561 | |
| 1562 | data->vgroups.safe_push (obj: group); |
| 1563 | return group; |
| 1564 | } |
| 1565 | |
| 1566 | /* Record a use of TYPE at *USE_P in STMT whose value is IV in a group. |
| 1567 | New group will be created if there is no existing group for the use. |
| 1568 | MEM_TYPE is the type of memory being addressed, or NULL if this |
| 1569 | isn't an address reference. */ |
| 1570 | |
| 1571 | static struct iv_use * |
| 1572 | record_group_use (struct ivopts_data *data, tree *use_p, |
| 1573 | struct iv *iv, gimple *stmt, enum use_type type, |
| 1574 | tree mem_type) |
| 1575 | { |
| 1576 | tree addr_base = NULL; |
| 1577 | struct iv_group *group = NULL; |
| 1578 | poly_uint64 addr_offset = 0; |
| 1579 | |
| 1580 | /* Record non address type use in a new group. */ |
| 1581 | if (address_p (type)) |
| 1582 | { |
| 1583 | unsigned int i; |
| 1584 | |
| 1585 | gcc_assert (POINTER_TYPE_P (TREE_TYPE (iv->base))); |
| 1586 | tree addr_toffset; |
| 1587 | split_constant_offset (iv->base, &addr_base, &addr_toffset); |
| 1588 | addr_offset = int_cst_value (addr_toffset); |
| 1589 | for (i = 0; i < data->vgroups.length (); i++) |
| 1590 | { |
| 1591 | struct iv_use *use; |
| 1592 | |
| 1593 | group = data->vgroups[i]; |
| 1594 | use = group->vuses[0]; |
| 1595 | if (!address_p (type: use->type)) |
| 1596 | continue; |
| 1597 | |
| 1598 | /* Check if it has the same stripped base and step. */ |
| 1599 | if (operand_equal_p (iv->base_object, use->iv->base_object, flags: 0) |
| 1600 | && operand_equal_p (iv->step, use->iv->step, flags: OEP_ASSUME_WRAPV) |
| 1601 | && operand_equal_p (addr_base, use->addr_base, flags: OEP_ASSUME_WRAPV)) |
| 1602 | break; |
| 1603 | } |
| 1604 | if (i == data->vgroups.length ()) |
| 1605 | group = NULL; |
| 1606 | } |
| 1607 | |
| 1608 | if (!group) |
| 1609 | group = record_group (data, type); |
| 1610 | |
| 1611 | return record_use (group, use_p, iv, stmt, type, mem_type, |
| 1612 | addr_base, addr_offset); |
| 1613 | } |
| 1614 | |
| 1615 | /* Checks whether the use OP is interesting and if so, records it. */ |
| 1616 | |
| 1617 | static struct iv_use * |
| 1618 | find_interesting_uses_op (struct ivopts_data *data, tree op) |
| 1619 | { |
| 1620 | struct iv *iv; |
| 1621 | gimple *stmt; |
| 1622 | struct iv_use *use; |
| 1623 | |
| 1624 | if (TREE_CODE (op) != SSA_NAME) |
| 1625 | return NULL; |
| 1626 | |
| 1627 | iv = get_iv (data, var: op); |
| 1628 | if (!iv) |
| 1629 | return NULL; |
| 1630 | |
| 1631 | if (iv->nonlin_use) |
| 1632 | { |
| 1633 | gcc_assert (iv->nonlin_use->type == USE_NONLINEAR_EXPR); |
| 1634 | return iv->nonlin_use; |
| 1635 | } |
| 1636 | |
| 1637 | if (integer_zerop (iv->step)) |
| 1638 | { |
| 1639 | record_invariant (data, op, nonlinear_use: true); |
| 1640 | return NULL; |
| 1641 | } |
| 1642 | |
| 1643 | stmt = SSA_NAME_DEF_STMT (op); |
| 1644 | gcc_assert (gimple_code (stmt) == GIMPLE_PHI || is_gimple_assign (stmt)); |
| 1645 | |
| 1646 | use = record_group_use (data, NULL, iv, stmt, type: USE_NONLINEAR_EXPR, NULL_TREE); |
| 1647 | iv->nonlin_use = use; |
| 1648 | return use; |
| 1649 | } |
| 1650 | |
| 1651 | /* Indicate how compare type iv_use can be handled. */ |
| 1652 | enum comp_iv_rewrite |
| 1653 | { |
| 1654 | COMP_IV_NA, |
| 1655 | /* We may rewrite compare type iv_use by expressing value of the iv_use. */ |
| 1656 | COMP_IV_EXPR, |
| 1657 | /* We may rewrite compare type iv_uses on both sides of comparison by |
| 1658 | expressing value of each iv_use. */ |
| 1659 | COMP_IV_EXPR_2, |
| 1660 | /* We may rewrite compare type iv_use by expressing value of the iv_use |
| 1661 | or by eliminating it with other iv_cand. */ |
| 1662 | COMP_IV_ELIM |
| 1663 | }; |
| 1664 | |
| 1665 | /* Given a condition in statement STMT, checks whether it is a compare |
| 1666 | of an induction variable and an invariant. If this is the case, |
| 1667 | CONTROL_VAR is set to location of the iv, BOUND to the location of |
| 1668 | the invariant, IV_VAR and IV_BOUND are set to the corresponding |
| 1669 | induction variable descriptions, and true is returned. If this is not |
| 1670 | the case, CONTROL_VAR and BOUND are set to the arguments of the |
| 1671 | condition and false is returned. */ |
| 1672 | |
| 1673 | static enum comp_iv_rewrite |
| 1674 | extract_cond_operands (struct ivopts_data *data, gimple *stmt, |
| 1675 | tree **control_var, tree **bound, |
| 1676 | struct iv **iv_var, struct iv **iv_bound) |
| 1677 | { |
| 1678 | /* The objects returned when COND has constant operands. */ |
| 1679 | static struct iv const_iv; |
| 1680 | static tree zero; |
| 1681 | tree *op0 = &zero, *op1 = &zero; |
| 1682 | struct iv *iv0 = &const_iv, *iv1 = &const_iv; |
| 1683 | enum comp_iv_rewrite rewrite_type = COMP_IV_NA; |
| 1684 | |
| 1685 | if (gimple_code (g: stmt) == GIMPLE_COND) |
| 1686 | { |
| 1687 | gcond *cond_stmt = as_a <gcond *> (p: stmt); |
| 1688 | op0 = gimple_cond_lhs_ptr (gs: cond_stmt); |
| 1689 | op1 = gimple_cond_rhs_ptr (gs: cond_stmt); |
| 1690 | } |
| 1691 | else |
| 1692 | { |
| 1693 | op0 = gimple_assign_rhs1_ptr (gs: stmt); |
| 1694 | op1 = gimple_assign_rhs2_ptr (gs: stmt); |
| 1695 | } |
| 1696 | |
| 1697 | zero = integer_zero_node; |
| 1698 | const_iv.step = integer_zero_node; |
| 1699 | |
| 1700 | if (TREE_CODE (*op0) == SSA_NAME) |
| 1701 | iv0 = get_iv (data, var: *op0); |
| 1702 | if (TREE_CODE (*op1) == SSA_NAME) |
| 1703 | iv1 = get_iv (data, var: *op1); |
| 1704 | |
| 1705 | /* If both sides of comparison are IVs. We can express ivs on both end. */ |
| 1706 | if (iv0 && iv1 && !integer_zerop (iv0->step) && !integer_zerop (iv1->step)) |
| 1707 | { |
| 1708 | rewrite_type = COMP_IV_EXPR_2; |
| 1709 | goto end; |
| 1710 | } |
| 1711 | |
| 1712 | /* If none side of comparison is IV. */ |
| 1713 | if ((!iv0 || integer_zerop (iv0->step)) |
| 1714 | && (!iv1 || integer_zerop (iv1->step))) |
| 1715 | goto end; |
| 1716 | |
| 1717 | /* Control variable may be on the other side. */ |
| 1718 | if (!iv0 || integer_zerop (iv0->step)) |
| 1719 | { |
| 1720 | std::swap (a&: op0, b&: op1); |
| 1721 | std::swap (a&: iv0, b&: iv1); |
| 1722 | } |
| 1723 | /* If one side is IV and the other side isn't loop invariant. */ |
| 1724 | if (!iv1) |
| 1725 | rewrite_type = COMP_IV_EXPR; |
| 1726 | /* If one side is IV and the other side is loop invariant. */ |
| 1727 | else if (!integer_zerop (iv0->step) && integer_zerop (iv1->step)) |
| 1728 | rewrite_type = COMP_IV_ELIM; |
| 1729 | |
| 1730 | end: |
| 1731 | if (control_var) |
| 1732 | *control_var = op0; |
| 1733 | if (iv_var) |
| 1734 | *iv_var = iv0; |
| 1735 | if (bound) |
| 1736 | *bound = op1; |
| 1737 | if (iv_bound) |
| 1738 | *iv_bound = iv1; |
| 1739 | |
| 1740 | return rewrite_type; |
| 1741 | } |
| 1742 | |
| 1743 | /* Checks whether the condition in STMT is interesting and if so, |
| 1744 | records it. */ |
| 1745 | |
| 1746 | static void |
| 1747 | find_interesting_uses_cond (struct ivopts_data *data, gimple *stmt) |
| 1748 | { |
| 1749 | tree *var_p, *bound_p; |
| 1750 | struct iv *var_iv, *bound_iv; |
| 1751 | enum comp_iv_rewrite ret; |
| 1752 | |
| 1753 | ret = extract_cond_operands (data, stmt, |
| 1754 | control_var: &var_p, bound: &bound_p, iv_var: &var_iv, iv_bound: &bound_iv); |
| 1755 | if (ret == COMP_IV_NA) |
| 1756 | { |
| 1757 | find_interesting_uses_op (data, op: *var_p); |
| 1758 | find_interesting_uses_op (data, op: *bound_p); |
| 1759 | return; |
| 1760 | } |
| 1761 | |
| 1762 | record_group_use (data, use_p: var_p, iv: var_iv, stmt, type: USE_COMPARE, NULL_TREE); |
| 1763 | /* Record compare type iv_use for iv on the other side of comparison. */ |
| 1764 | if (ret == COMP_IV_EXPR_2) |
| 1765 | record_group_use (data, use_p: bound_p, iv: bound_iv, stmt, type: USE_COMPARE, NULL_TREE); |
| 1766 | } |
| 1767 | |
| 1768 | /* Returns the outermost loop EXPR is obviously invariant in |
| 1769 | relative to the loop LOOP, i.e. if all its operands are defined |
| 1770 | outside of the returned loop. Returns NULL if EXPR is not |
| 1771 | even obviously invariant in LOOP. */ |
| 1772 | |
| 1773 | class loop * |
| 1774 | outermost_invariant_loop_for_expr (class loop *loop, tree expr) |
| 1775 | { |
| 1776 | basic_block def_bb; |
| 1777 | unsigned i, len; |
| 1778 | |
| 1779 | if (is_gimple_min_invariant (expr)) |
| 1780 | return current_loops->tree_root; |
| 1781 | |
| 1782 | if (TREE_CODE (expr) == SSA_NAME) |
| 1783 | { |
| 1784 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); |
| 1785 | if (def_bb) |
| 1786 | { |
| 1787 | if (flow_bb_inside_loop_p (loop, def_bb)) |
| 1788 | return NULL; |
| 1789 | return superloop_at_depth (loop, |
| 1790 | loop_depth (loop: def_bb->loop_father) + 1); |
| 1791 | } |
| 1792 | |
| 1793 | return current_loops->tree_root; |
| 1794 | } |
| 1795 | |
| 1796 | if (!EXPR_P (expr)) |
| 1797 | return NULL; |
| 1798 | |
| 1799 | unsigned maxdepth = 0; |
| 1800 | len = TREE_OPERAND_LENGTH (expr); |
| 1801 | for (i = 0; i < len; i++) |
| 1802 | { |
| 1803 | class loop *ivloop; |
| 1804 | if (!TREE_OPERAND (expr, i)) |
| 1805 | continue; |
| 1806 | |
| 1807 | ivloop = outermost_invariant_loop_for_expr (loop, TREE_OPERAND (expr, i)); |
| 1808 | if (!ivloop) |
| 1809 | return NULL; |
| 1810 | maxdepth = MAX (maxdepth, loop_depth (ivloop)); |
| 1811 | } |
| 1812 | |
| 1813 | return superloop_at_depth (loop, maxdepth); |
| 1814 | } |
| 1815 | |
| 1816 | /* Returns true if expression EXPR is obviously invariant in LOOP, |
| 1817 | i.e. if all its operands are defined outside of the LOOP. LOOP |
| 1818 | should not be the function body. */ |
| 1819 | |
| 1820 | bool |
| 1821 | expr_invariant_in_loop_p (class loop *loop, tree expr) |
| 1822 | { |
| 1823 | basic_block def_bb; |
| 1824 | unsigned i, len; |
| 1825 | |
| 1826 | gcc_assert (loop_depth (loop) > 0); |
| 1827 | |
| 1828 | if (is_gimple_min_invariant (expr)) |
| 1829 | return true; |
| 1830 | |
| 1831 | if (TREE_CODE (expr) == SSA_NAME) |
| 1832 | { |
| 1833 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); |
| 1834 | if (def_bb |
| 1835 | && flow_bb_inside_loop_p (loop, def_bb)) |
| 1836 | return false; |
| 1837 | |
| 1838 | return true; |
| 1839 | } |
| 1840 | |
| 1841 | if (!EXPR_P (expr)) |
| 1842 | return false; |
| 1843 | |
| 1844 | len = TREE_OPERAND_LENGTH (expr); |
| 1845 | for (i = 0; i < len; i++) |
| 1846 | if (TREE_OPERAND (expr, i) |
| 1847 | && !expr_invariant_in_loop_p (loop, TREE_OPERAND (expr, i))) |
| 1848 | return false; |
| 1849 | |
| 1850 | return true; |
| 1851 | } |
| 1852 | |
| 1853 | /* Given expression EXPR which computes inductive values with respect |
| 1854 | to loop recorded in DATA, this function returns biv from which EXPR |
| 1855 | is derived by tracing definition chains of ssa variables in EXPR. */ |
| 1856 | |
| 1857 | static struct iv* |
| 1858 | find_deriving_biv_for_expr (struct ivopts_data *data, tree expr) |
| 1859 | { |
| 1860 | struct iv *iv; |
| 1861 | unsigned i, n; |
| 1862 | tree e2, e1; |
| 1863 | enum tree_code code; |
| 1864 | gimple *stmt; |
| 1865 | |
| 1866 | if (expr == NULL_TREE) |
| 1867 | return NULL; |
| 1868 | |
| 1869 | if (is_gimple_min_invariant (expr)) |
| 1870 | return NULL; |
| 1871 | |
| 1872 | code = TREE_CODE (expr); |
| 1873 | if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code))) |
| 1874 | { |
| 1875 | n = TREE_OPERAND_LENGTH (expr); |
| 1876 | for (i = 0; i < n; i++) |
| 1877 | { |
| 1878 | iv = find_deriving_biv_for_expr (data, TREE_OPERAND (expr, i)); |
| 1879 | if (iv) |
| 1880 | return iv; |
| 1881 | } |
| 1882 | } |
| 1883 | |
| 1884 | /* Stop if it's not ssa name. */ |
| 1885 | if (code != SSA_NAME) |
| 1886 | return NULL; |
| 1887 | |
| 1888 | iv = get_iv (data, var: expr); |
| 1889 | if (!iv || integer_zerop (iv->step)) |
| 1890 | return NULL; |
| 1891 | else if (iv->biv_p) |
| 1892 | return iv; |
| 1893 | |
| 1894 | stmt = SSA_NAME_DEF_STMT (expr); |
| 1895 | if (gphi *phi = dyn_cast <gphi *> (p: stmt)) |
| 1896 | { |
| 1897 | ssa_op_iter iter; |
| 1898 | use_operand_p use_p; |
| 1899 | basic_block phi_bb = gimple_bb (g: phi); |
| 1900 | |
| 1901 | /* Skip loop header PHI that doesn't define biv. */ |
| 1902 | if (phi_bb->loop_father == data->current_loop) |
| 1903 | return NULL; |
| 1904 | |
| 1905 | if (virtual_operand_p (op: gimple_phi_result (gs: phi))) |
| 1906 | return NULL; |
| 1907 | |
| 1908 | FOR_EACH_PHI_ARG (use_p, phi, iter, SSA_OP_USE) |
| 1909 | { |
| 1910 | tree use = USE_FROM_PTR (use_p); |
| 1911 | iv = find_deriving_biv_for_expr (data, expr: use); |
| 1912 | if (iv) |
| 1913 | return iv; |
| 1914 | } |
| 1915 | return NULL; |
| 1916 | } |
| 1917 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
| 1918 | return NULL; |
| 1919 | |
| 1920 | e1 = gimple_assign_rhs1 (gs: stmt); |
| 1921 | code = gimple_assign_rhs_code (gs: stmt); |
| 1922 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) |
| 1923 | return find_deriving_biv_for_expr (data, expr: e1); |
| 1924 | |
| 1925 | switch (code) |
| 1926 | { |
| 1927 | case MULT_EXPR: |
| 1928 | case PLUS_EXPR: |
| 1929 | case MINUS_EXPR: |
| 1930 | case POINTER_PLUS_EXPR: |
| 1931 | /* Increments, decrements and multiplications by a constant |
| 1932 | are simple. */ |
| 1933 | e2 = gimple_assign_rhs2 (gs: stmt); |
| 1934 | iv = find_deriving_biv_for_expr (data, expr: e2); |
| 1935 | if (iv) |
| 1936 | return iv; |
| 1937 | gcc_fallthrough (); |
| 1938 | |
| 1939 | CASE_CONVERT: |
| 1940 | /* Casts are simple. */ |
| 1941 | return find_deriving_biv_for_expr (data, expr: e1); |
| 1942 | |
| 1943 | default: |
| 1944 | break; |
| 1945 | } |
| 1946 | |
| 1947 | return NULL; |
| 1948 | } |
| 1949 | |
| 1950 | /* Record BIV, its predecessor and successor that they are used in |
| 1951 | address type uses. */ |
| 1952 | |
| 1953 | static void |
| 1954 | record_biv_for_address_use (struct ivopts_data *data, struct iv *biv) |
| 1955 | { |
| 1956 | unsigned i; |
| 1957 | tree type, base_1, base_2; |
| 1958 | bitmap_iterator bi; |
| 1959 | |
| 1960 | if (!biv || !biv->biv_p || integer_zerop (biv->step) |
| 1961 | || biv->have_address_use || !biv->no_overflow) |
| 1962 | return; |
| 1963 | |
| 1964 | type = TREE_TYPE (biv->base); |
| 1965 | if (!INTEGRAL_TYPE_P (type)) |
| 1966 | return; |
| 1967 | |
| 1968 | biv->have_address_use = true; |
| 1969 | data->bivs_not_used_in_addr--; |
| 1970 | base_1 = fold_build2 (PLUS_EXPR, type, biv->base, biv->step); |
| 1971 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
| 1972 | { |
| 1973 | struct iv *iv = ver_info (data, ver: i)->iv; |
| 1974 | |
| 1975 | if (!iv || !iv->biv_p || integer_zerop (iv->step) |
| 1976 | || iv->have_address_use || !iv->no_overflow) |
| 1977 | continue; |
| 1978 | |
| 1979 | if (type != TREE_TYPE (iv->base) |
| 1980 | || !INTEGRAL_TYPE_P (TREE_TYPE (iv->base))) |
| 1981 | continue; |
| 1982 | |
| 1983 | if (!operand_equal_p (biv->step, iv->step, flags: 0)) |
| 1984 | continue; |
| 1985 | |
| 1986 | base_2 = fold_build2 (PLUS_EXPR, type, iv->base, iv->step); |
| 1987 | if (operand_equal_p (base_1, iv->base, flags: 0) |
| 1988 | || operand_equal_p (base_2, biv->base, flags: 0)) |
| 1989 | { |
| 1990 | iv->have_address_use = true; |
| 1991 | data->bivs_not_used_in_addr--; |
| 1992 | } |
| 1993 | } |
| 1994 | } |
| 1995 | |
| 1996 | /* Cumulates the steps of indices into DATA and replaces their values with the |
| 1997 | initial ones. Returns false when the value of the index cannot be determined. |
| 1998 | Callback for for_each_index. */ |
| 1999 | |
| 2000 | struct ifs_ivopts_data |
| 2001 | { |
| 2002 | struct ivopts_data *ivopts_data; |
| 2003 | gimple *stmt; |
| 2004 | tree step; |
| 2005 | }; |
| 2006 | |
| 2007 | static bool |
| 2008 | idx_find_step (tree base, tree *idx, void *data) |
| 2009 | { |
| 2010 | struct ifs_ivopts_data *dta = (struct ifs_ivopts_data *) data; |
| 2011 | struct iv *iv; |
| 2012 | bool use_overflow_semantics = false; |
| 2013 | tree step, iv_base, iv_step, lbound, off; |
| 2014 | class loop *loop = dta->ivopts_data->current_loop; |
| 2015 | |
| 2016 | /* If base is a component ref, require that the offset of the reference |
| 2017 | be invariant. */ |
| 2018 | if (TREE_CODE (base) == COMPONENT_REF) |
| 2019 | { |
| 2020 | off = component_ref_field_offset (base); |
| 2021 | return expr_invariant_in_loop_p (loop, expr: off); |
| 2022 | } |
| 2023 | |
| 2024 | /* If base is array, first check whether we will be able to move the |
| 2025 | reference out of the loop (in order to take its address in strength |
| 2026 | reduction). In order for this to work we need both lower bound |
| 2027 | and step to be loop invariants. */ |
| 2028 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
| 2029 | { |
| 2030 | /* Moreover, for a range, the size needs to be invariant as well. */ |
| 2031 | if (TREE_CODE (base) == ARRAY_RANGE_REF |
| 2032 | && !expr_invariant_in_loop_p (loop, TYPE_SIZE (TREE_TYPE (base)))) |
| 2033 | return false; |
| 2034 | |
| 2035 | step = array_ref_element_size (base); |
| 2036 | lbound = array_ref_low_bound (base); |
| 2037 | |
| 2038 | if (!expr_invariant_in_loop_p (loop, expr: step) |
| 2039 | || !expr_invariant_in_loop_p (loop, expr: lbound)) |
| 2040 | return false; |
| 2041 | } |
| 2042 | |
| 2043 | if (TREE_CODE (*idx) != SSA_NAME) |
| 2044 | return true; |
| 2045 | |
| 2046 | iv = get_iv (data: dta->ivopts_data, var: *idx); |
| 2047 | if (!iv) |
| 2048 | return false; |
| 2049 | |
| 2050 | /* XXX We produce for a base of *D42 with iv->base being &x[0] |
| 2051 | *&x[0], which is not folded and does not trigger the |
| 2052 | ARRAY_REF path below. */ |
| 2053 | *idx = iv->base; |
| 2054 | |
| 2055 | if (integer_zerop (iv->step)) |
| 2056 | return true; |
| 2057 | |
| 2058 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
| 2059 | { |
| 2060 | step = array_ref_element_size (base); |
| 2061 | |
| 2062 | /* We only handle addresses whose step is an integer constant. */ |
| 2063 | if (TREE_CODE (step) != INTEGER_CST) |
| 2064 | return false; |
| 2065 | } |
| 2066 | else |
| 2067 | /* The step for pointer arithmetics already is 1 byte. */ |
| 2068 | step = size_one_node; |
| 2069 | |
| 2070 | iv_base = iv->base; |
| 2071 | iv_step = iv->step; |
| 2072 | if (iv->no_overflow && nowrap_type_p (TREE_TYPE (iv_step))) |
| 2073 | use_overflow_semantics = true; |
| 2074 | |
| 2075 | if (!convert_affine_scev (dta->ivopts_data->current_loop, |
| 2076 | sizetype, &iv_base, &iv_step, dta->stmt, |
| 2077 | use_overflow_semantics)) |
| 2078 | { |
| 2079 | /* The index might wrap. */ |
| 2080 | return false; |
| 2081 | } |
| 2082 | |
| 2083 | step = fold_build2 (MULT_EXPR, sizetype, step, iv_step); |
| 2084 | dta->step = fold_build2 (PLUS_EXPR, sizetype, dta->step, step); |
| 2085 | |
| 2086 | if (dta->ivopts_data->bivs_not_used_in_addr) |
| 2087 | { |
| 2088 | if (!iv->biv_p) |
| 2089 | iv = find_deriving_biv_for_expr (data: dta->ivopts_data, expr: iv->ssa_name); |
| 2090 | |
| 2091 | record_biv_for_address_use (data: dta->ivopts_data, biv: iv); |
| 2092 | } |
| 2093 | return true; |
| 2094 | } |
| 2095 | |
| 2096 | /* Records use in index IDX. Callback for for_each_index. Ivopts data |
| 2097 | object is passed to it in DATA. */ |
| 2098 | |
| 2099 | static bool |
| 2100 | idx_record_use (tree base, tree *idx, |
| 2101 | void *vdata) |
| 2102 | { |
| 2103 | struct ivopts_data *data = (struct ivopts_data *) vdata; |
| 2104 | find_interesting_uses_op (data, op: *idx); |
| 2105 | if (TREE_CODE (base) == ARRAY_REF || TREE_CODE (base) == ARRAY_RANGE_REF) |
| 2106 | { |
| 2107 | if (TREE_OPERAND (base, 2)) |
| 2108 | find_interesting_uses_op (data, TREE_OPERAND (base, 2)); |
| 2109 | if (TREE_OPERAND (base, 3)) |
| 2110 | find_interesting_uses_op (data, TREE_OPERAND (base, 3)); |
| 2111 | } |
| 2112 | return true; |
| 2113 | } |
| 2114 | |
| 2115 | /* If we can prove that TOP = cst * BOT for some constant cst, |
| 2116 | store cst to MUL and return true. Otherwise return false. |
| 2117 | The returned value is always sign-extended, regardless of the |
| 2118 | signedness of TOP and BOT. */ |
| 2119 | |
| 2120 | static bool |
| 2121 | constant_multiple_of (tree top, tree bot, widest_int *mul, |
| 2122 | struct ivopts_data *data) |
| 2123 | { |
| 2124 | aff_tree aff_top, aff_bot; |
| 2125 | tree_to_aff_combination_expand (top, TREE_TYPE (top), &aff_top, |
| 2126 | &data->name_expansion_cache); |
| 2127 | tree_to_aff_combination_expand (bot, TREE_TYPE (bot), &aff_bot, |
| 2128 | &data->name_expansion_cache); |
| 2129 | |
| 2130 | poly_widest_int poly_mul; |
| 2131 | if (aff_combination_constant_multiple_p (&aff_top, &aff_bot, &poly_mul) |
| 2132 | && poly_mul.is_constant (const_value: mul)) |
| 2133 | return true; |
| 2134 | |
| 2135 | return false; |
| 2136 | } |
| 2137 | |
| 2138 | /* Return true if memory reference REF with step STEP may be unaligned. */ |
| 2139 | |
| 2140 | static bool |
| 2141 | may_be_unaligned_p (tree ref, tree step) |
| 2142 | { |
| 2143 | /* TARGET_MEM_REFs are translated directly to valid MEMs on the target, |
| 2144 | thus they are not misaligned. */ |
| 2145 | if (TREE_CODE (ref) == TARGET_MEM_REF) |
| 2146 | return false; |
| 2147 | |
| 2148 | unsigned int align = TYPE_ALIGN (TREE_TYPE (ref)); |
| 2149 | if (GET_MODE_ALIGNMENT (TYPE_MODE (TREE_TYPE (ref))) > align) |
| 2150 | align = GET_MODE_ALIGNMENT (TYPE_MODE (TREE_TYPE (ref))); |
| 2151 | |
| 2152 | unsigned HOST_WIDE_INT bitpos; |
| 2153 | unsigned int ref_align; |
| 2154 | get_object_alignment_1 (ref, &ref_align, &bitpos); |
| 2155 | if (ref_align < align |
| 2156 | || (bitpos % align) != 0 |
| 2157 | || (bitpos % BITS_PER_UNIT) != 0) |
| 2158 | return true; |
| 2159 | |
| 2160 | unsigned int trailing_zeros = tree_ctz (step); |
| 2161 | if (trailing_zeros < HOST_BITS_PER_INT |
| 2162 | && (1U << trailing_zeros) * BITS_PER_UNIT < align) |
| 2163 | return true; |
| 2164 | |
| 2165 | return false; |
| 2166 | } |
| 2167 | |
| 2168 | /* Return true if EXPR may be non-addressable. */ |
| 2169 | |
| 2170 | bool |
| 2171 | may_be_nonaddressable_p (tree expr) |
| 2172 | { |
| 2173 | switch (TREE_CODE (expr)) |
| 2174 | { |
| 2175 | case VAR_DECL: |
| 2176 | /* Check if it's a register variable. */ |
| 2177 | return DECL_HARD_REGISTER (expr); |
| 2178 | |
| 2179 | case TARGET_MEM_REF: |
| 2180 | /* TARGET_MEM_REFs are translated directly to valid MEMs on the |
| 2181 | target, thus they are always addressable. */ |
| 2182 | return false; |
| 2183 | |
| 2184 | case MEM_REF: |
| 2185 | /* Likewise for MEM_REFs, modulo the storage order. */ |
| 2186 | return REF_REVERSE_STORAGE_ORDER (expr); |
| 2187 | |
| 2188 | case BIT_FIELD_REF: |
| 2189 | if (REF_REVERSE_STORAGE_ORDER (expr)) |
| 2190 | return true; |
| 2191 | return may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); |
| 2192 | |
| 2193 | case COMPONENT_REF: |
| 2194 | if (TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_OPERAND (expr, 0)))) |
| 2195 | return true; |
| 2196 | return DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)) |
| 2197 | || may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); |
| 2198 | |
| 2199 | case ARRAY_REF: |
| 2200 | case ARRAY_RANGE_REF: |
| 2201 | if (TYPE_REVERSE_STORAGE_ORDER (TREE_TYPE (TREE_OPERAND (expr, 0)))) |
| 2202 | return true; |
| 2203 | return may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); |
| 2204 | |
| 2205 | case VIEW_CONVERT_EXPR: |
| 2206 | /* This kind of view-conversions may wrap non-addressable objects |
| 2207 | and make them look addressable. After some processing the |
| 2208 | non-addressability may be uncovered again, causing ADDR_EXPRs |
| 2209 | of inappropriate objects to be built. */ |
| 2210 | if (is_gimple_reg (TREE_OPERAND (expr, 0)) |
| 2211 | || !is_gimple_addressable (TREE_OPERAND (expr, 0))) |
| 2212 | return true; |
| 2213 | return may_be_nonaddressable_p (TREE_OPERAND (expr, 0)); |
| 2214 | |
| 2215 | CASE_CONVERT: |
| 2216 | return true; |
| 2217 | |
| 2218 | default: |
| 2219 | break; |
| 2220 | } |
| 2221 | |
| 2222 | return false; |
| 2223 | } |
| 2224 | |
| 2225 | /* Finds addresses in *OP_P inside STMT. */ |
| 2226 | |
| 2227 | static void |
| 2228 | find_interesting_uses_address (struct ivopts_data *data, gimple *stmt, |
| 2229 | tree *op_p) |
| 2230 | { |
| 2231 | tree base = *op_p, step = size_zero_node; |
| 2232 | struct iv *civ; |
| 2233 | struct ifs_ivopts_data ifs_ivopts_data; |
| 2234 | |
| 2235 | /* Do not play with volatile memory references. A bit too conservative, |
| 2236 | perhaps, but safe. */ |
| 2237 | if (gimple_has_volatile_ops (stmt)) |
| 2238 | goto fail; |
| 2239 | |
| 2240 | /* Ignore bitfields for now. Not really something terribly complicated |
| 2241 | to handle. TODO. */ |
| 2242 | if (TREE_CODE (base) == BIT_FIELD_REF) |
| 2243 | goto fail; |
| 2244 | |
| 2245 | base = unshare_expr (base); |
| 2246 | |
| 2247 | if (TREE_CODE (base) == TARGET_MEM_REF) |
| 2248 | { |
| 2249 | tree type = build_pointer_type (TREE_TYPE (base)); |
| 2250 | tree astep; |
| 2251 | |
| 2252 | if (TMR_BASE (base) |
| 2253 | && TREE_CODE (TMR_BASE (base)) == SSA_NAME) |
| 2254 | { |
| 2255 | civ = get_iv (data, TMR_BASE (base)); |
| 2256 | if (!civ) |
| 2257 | goto fail; |
| 2258 | |
| 2259 | TMR_BASE (base) = civ->base; |
| 2260 | step = civ->step; |
| 2261 | } |
| 2262 | if (TMR_INDEX2 (base) |
| 2263 | && TREE_CODE (TMR_INDEX2 (base)) == SSA_NAME) |
| 2264 | { |
| 2265 | civ = get_iv (data, TMR_INDEX2 (base)); |
| 2266 | if (!civ) |
| 2267 | goto fail; |
| 2268 | |
| 2269 | TMR_INDEX2 (base) = civ->base; |
| 2270 | step = civ->step; |
| 2271 | } |
| 2272 | if (TMR_INDEX (base) |
| 2273 | && TREE_CODE (TMR_INDEX (base)) == SSA_NAME) |
| 2274 | { |
| 2275 | civ = get_iv (data, TMR_INDEX (base)); |
| 2276 | if (!civ) |
| 2277 | goto fail; |
| 2278 | |
| 2279 | TMR_INDEX (base) = civ->base; |
| 2280 | astep = civ->step; |
| 2281 | |
| 2282 | if (astep) |
| 2283 | { |
| 2284 | if (TMR_STEP (base)) |
| 2285 | astep = fold_build2 (MULT_EXPR, type, TMR_STEP (base), astep); |
| 2286 | |
| 2287 | step = fold_build2 (PLUS_EXPR, type, step, astep); |
| 2288 | } |
| 2289 | } |
| 2290 | |
| 2291 | if (integer_zerop (step)) |
| 2292 | goto fail; |
| 2293 | base = tree_mem_ref_addr (type, base); |
| 2294 | } |
| 2295 | else |
| 2296 | { |
| 2297 | ifs_ivopts_data.ivopts_data = data; |
| 2298 | ifs_ivopts_data.stmt = stmt; |
| 2299 | ifs_ivopts_data.step = size_zero_node; |
| 2300 | if (!for_each_index (&base, idx_find_step, &ifs_ivopts_data) |
| 2301 | || integer_zerop (ifs_ivopts_data.step)) |
| 2302 | goto fail; |
| 2303 | step = ifs_ivopts_data.step; |
| 2304 | |
| 2305 | /* Check that the base expression is addressable. This needs |
| 2306 | to be done after substituting bases of IVs into it. */ |
| 2307 | if (may_be_nonaddressable_p (expr: base)) |
| 2308 | goto fail; |
| 2309 | |
| 2310 | /* Moreover, on strict alignment platforms, check that it is |
| 2311 | sufficiently aligned. */ |
| 2312 | if (STRICT_ALIGNMENT && may_be_unaligned_p (ref: base, step)) |
| 2313 | goto fail; |
| 2314 | |
| 2315 | base = build_fold_addr_expr (base); |
| 2316 | |
| 2317 | /* Substituting bases of IVs into the base expression might |
| 2318 | have caused folding opportunities. */ |
| 2319 | if (TREE_CODE (base) == ADDR_EXPR) |
| 2320 | { |
| 2321 | tree *ref = &TREE_OPERAND (base, 0); |
| 2322 | while (handled_component_p (t: *ref)) |
| 2323 | ref = &TREE_OPERAND (*ref, 0); |
| 2324 | if (TREE_CODE (*ref) == MEM_REF) |
| 2325 | { |
| 2326 | tree tem = fold_binary (MEM_REF, TREE_TYPE (*ref), |
| 2327 | TREE_OPERAND (*ref, 0), |
| 2328 | TREE_OPERAND (*ref, 1)); |
| 2329 | if (tem) |
| 2330 | *ref = tem; |
| 2331 | } |
| 2332 | } |
| 2333 | } |
| 2334 | |
| 2335 | civ = alloc_iv (data, base, step); |
| 2336 | /* Fail if base object of this memory reference is unknown. */ |
| 2337 | if (civ->base_object == NULL_TREE) |
| 2338 | goto fail; |
| 2339 | |
| 2340 | record_group_use (data, use_p: op_p, iv: civ, stmt, type: USE_REF_ADDRESS, TREE_TYPE (*op_p)); |
| 2341 | return; |
| 2342 | |
| 2343 | fail: |
| 2344 | for_each_index (op_p, idx_record_use, data); |
| 2345 | } |
| 2346 | |
| 2347 | /* Finds and records invariants used in STMT. */ |
| 2348 | |
| 2349 | static void |
| 2350 | find_invariants_stmt (struct ivopts_data *data, gimple *stmt) |
| 2351 | { |
| 2352 | ssa_op_iter iter; |
| 2353 | use_operand_p use_p; |
| 2354 | tree op; |
| 2355 | |
| 2356 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
| 2357 | { |
| 2358 | op = USE_FROM_PTR (use_p); |
| 2359 | record_invariant (data, op, nonlinear_use: false); |
| 2360 | } |
| 2361 | } |
| 2362 | |
| 2363 | /* CALL calls an internal function. If operand *OP_P will become an |
| 2364 | address when the call is expanded, return the type of the memory |
| 2365 | being addressed, otherwise return null. */ |
| 2366 | |
| 2367 | static tree |
| 2368 | get_mem_type_for_internal_fn (gcall *call, tree *op_p) |
| 2369 | { |
| 2370 | switch (gimple_call_internal_fn (gs: call)) |
| 2371 | { |
| 2372 | case IFN_MASK_LOAD: |
| 2373 | case IFN_MASK_LOAD_LANES: |
| 2374 | case IFN_MASK_LEN_LOAD_LANES: |
| 2375 | case IFN_LEN_LOAD: |
| 2376 | case IFN_MASK_LEN_LOAD: |
| 2377 | if (op_p == gimple_call_arg_ptr (gs: call, index: 0)) |
| 2378 | return TREE_TYPE (gimple_call_lhs (call)); |
| 2379 | return NULL_TREE; |
| 2380 | |
| 2381 | case IFN_MASK_STORE: |
| 2382 | case IFN_MASK_STORE_LANES: |
| 2383 | case IFN_MASK_LEN_STORE_LANES: |
| 2384 | case IFN_LEN_STORE: |
| 2385 | case IFN_MASK_LEN_STORE: |
| 2386 | { |
| 2387 | if (op_p == gimple_call_arg_ptr (gs: call, index: 0)) |
| 2388 | { |
| 2389 | internal_fn ifn = gimple_call_internal_fn (gs: call); |
| 2390 | int index = internal_fn_stored_value_index (ifn); |
| 2391 | return TREE_TYPE (gimple_call_arg (call, index)); |
| 2392 | } |
| 2393 | return NULL_TREE; |
| 2394 | } |
| 2395 | |
| 2396 | default: |
| 2397 | return NULL_TREE; |
| 2398 | } |
| 2399 | } |
| 2400 | |
| 2401 | /* IV is a (non-address) iv that describes operand *OP_P of STMT. |
| 2402 | Return true if the operand will become an address when STMT |
| 2403 | is expanded and record the associated address use if so. */ |
| 2404 | |
| 2405 | static bool |
| 2406 | find_address_like_use (struct ivopts_data *data, gimple *stmt, tree *op_p, |
| 2407 | struct iv *iv) |
| 2408 | { |
| 2409 | /* Fail if base object of this memory reference is unknown. */ |
| 2410 | if (iv->base_object == NULL_TREE) |
| 2411 | return false; |
| 2412 | |
| 2413 | tree mem_type = NULL_TREE; |
| 2414 | if (gcall *call = dyn_cast <gcall *> (p: stmt)) |
| 2415 | if (gimple_call_internal_p (gs: call)) |
| 2416 | mem_type = get_mem_type_for_internal_fn (call, op_p); |
| 2417 | if (mem_type) |
| 2418 | { |
| 2419 | iv = alloc_iv (data, base: iv->base, step: iv->step); |
| 2420 | record_group_use (data, use_p: op_p, iv, stmt, type: USE_PTR_ADDRESS, mem_type); |
| 2421 | return true; |
| 2422 | } |
| 2423 | return false; |
| 2424 | } |
| 2425 | |
| 2426 | /* Finds interesting uses of induction variables in the statement STMT. */ |
| 2427 | |
| 2428 | static void |
| 2429 | find_interesting_uses_stmt (struct ivopts_data *data, gimple *stmt) |
| 2430 | { |
| 2431 | struct iv *iv; |
| 2432 | tree op, *lhs, *rhs; |
| 2433 | ssa_op_iter iter; |
| 2434 | use_operand_p use_p; |
| 2435 | enum tree_code code; |
| 2436 | |
| 2437 | find_invariants_stmt (data, stmt); |
| 2438 | |
| 2439 | if (gimple_code (g: stmt) == GIMPLE_COND) |
| 2440 | { |
| 2441 | find_interesting_uses_cond (data, stmt); |
| 2442 | return; |
| 2443 | } |
| 2444 | |
| 2445 | if (is_gimple_assign (gs: stmt)) |
| 2446 | { |
| 2447 | lhs = gimple_assign_lhs_ptr (gs: stmt); |
| 2448 | rhs = gimple_assign_rhs1_ptr (gs: stmt); |
| 2449 | |
| 2450 | if (TREE_CODE (*lhs) == SSA_NAME) |
| 2451 | { |
| 2452 | /* If the statement defines an induction variable, the uses are not |
| 2453 | interesting by themselves. */ |
| 2454 | |
| 2455 | iv = get_iv (data, var: *lhs); |
| 2456 | |
| 2457 | if (iv && !integer_zerop (iv->step)) |
| 2458 | return; |
| 2459 | } |
| 2460 | |
| 2461 | code = gimple_assign_rhs_code (gs: stmt); |
| 2462 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS |
| 2463 | && (REFERENCE_CLASS_P (*rhs) |
| 2464 | || is_gimple_val (*rhs))) |
| 2465 | { |
| 2466 | if (REFERENCE_CLASS_P (*rhs)) |
| 2467 | find_interesting_uses_address (data, stmt, op_p: rhs); |
| 2468 | else |
| 2469 | find_interesting_uses_op (data, op: *rhs); |
| 2470 | |
| 2471 | if (REFERENCE_CLASS_P (*lhs)) |
| 2472 | find_interesting_uses_address (data, stmt, op_p: lhs); |
| 2473 | return; |
| 2474 | } |
| 2475 | else if (TREE_CODE_CLASS (code) == tcc_comparison) |
| 2476 | { |
| 2477 | find_interesting_uses_cond (data, stmt); |
| 2478 | return; |
| 2479 | } |
| 2480 | |
| 2481 | /* TODO -- we should also handle address uses of type |
| 2482 | |
| 2483 | memory = call (whatever); |
| 2484 | |
| 2485 | and |
| 2486 | |
| 2487 | call (memory). */ |
| 2488 | } |
| 2489 | |
| 2490 | if (gimple_code (g: stmt) == GIMPLE_PHI |
| 2491 | && gimple_bb (g: stmt) == data->current_loop->header) |
| 2492 | { |
| 2493 | iv = get_iv (data, PHI_RESULT (stmt)); |
| 2494 | |
| 2495 | if (iv && !integer_zerop (iv->step)) |
| 2496 | return; |
| 2497 | } |
| 2498 | |
| 2499 | FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE) |
| 2500 | { |
| 2501 | op = USE_FROM_PTR (use_p); |
| 2502 | |
| 2503 | if (TREE_CODE (op) != SSA_NAME) |
| 2504 | continue; |
| 2505 | |
| 2506 | iv = get_iv (data, var: op); |
| 2507 | if (!iv) |
| 2508 | continue; |
| 2509 | |
| 2510 | if (!find_address_like_use (data, stmt, op_p: use_p->use, iv)) |
| 2511 | find_interesting_uses_op (data, op); |
| 2512 | } |
| 2513 | } |
| 2514 | |
| 2515 | /* Finds interesting uses of induction variables outside of loops |
| 2516 | on loop exit edge EXIT. */ |
| 2517 | |
| 2518 | static void |
| 2519 | find_interesting_uses_outside (struct ivopts_data *data, edge exit) |
| 2520 | { |
| 2521 | gphi *phi; |
| 2522 | gphi_iterator psi; |
| 2523 | tree def; |
| 2524 | |
| 2525 | for (psi = gsi_start_phis (exit->dest); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
| 2526 | { |
| 2527 | phi = psi.phi (); |
| 2528 | def = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
| 2529 | if (!virtual_operand_p (op: def)) |
| 2530 | find_interesting_uses_op (data, op: def); |
| 2531 | } |
| 2532 | } |
| 2533 | |
| 2534 | /* Return TRUE if OFFSET is within the range of [base + offset] addressing |
| 2535 | mode for memory reference represented by USE. */ |
| 2536 | |
| 2537 | static GTY (()) vec<rtx, va_gc> *addr_list; |
| 2538 | |
| 2539 | static bool |
| 2540 | addr_offset_valid_p (struct iv_use *use, poly_int64 offset) |
| 2541 | { |
| 2542 | rtx reg, addr; |
| 2543 | unsigned list_index; |
| 2544 | addr_space_t as = TYPE_ADDR_SPACE (TREE_TYPE (use->iv->base)); |
| 2545 | machine_mode addr_mode, mem_mode = TYPE_MODE (use->mem_type); |
| 2546 | |
| 2547 | list_index = (unsigned) as * MAX_MACHINE_MODE + (unsigned) mem_mode; |
| 2548 | if (list_index >= vec_safe_length (v: addr_list)) |
| 2549 | vec_safe_grow_cleared (v&: addr_list, len: list_index + MAX_MACHINE_MODE, exact: true); |
| 2550 | |
| 2551 | addr = (*addr_list)[list_index]; |
| 2552 | if (!addr) |
| 2553 | { |
| 2554 | addr_mode = targetm.addr_space.address_mode (as); |
| 2555 | reg = gen_raw_REG (addr_mode, LAST_VIRTUAL_REGISTER + 1); |
| 2556 | addr = gen_rtx_fmt_ee (PLUS, addr_mode, reg, NULL_RTX); |
| 2557 | (*addr_list)[list_index] = addr; |
| 2558 | } |
| 2559 | else |
| 2560 | addr_mode = GET_MODE (addr); |
| 2561 | |
| 2562 | XEXP (addr, 1) = gen_int_mode (offset, addr_mode); |
| 2563 | return (memory_address_addr_space_p (mem_mode, addr, as)); |
| 2564 | } |
| 2565 | |
| 2566 | /* Comparison function to sort group in ascending order of addr_offset. */ |
| 2567 | |
| 2568 | static int |
| 2569 | group_compare_offset (const void *a, const void *b) |
| 2570 | { |
| 2571 | const struct iv_use *const *u1 = (const struct iv_use *const *) a; |
| 2572 | const struct iv_use *const *u2 = (const struct iv_use *const *) b; |
| 2573 | |
| 2574 | return compare_sizes_for_sort (a: (*u1)->addr_offset, b: (*u2)->addr_offset); |
| 2575 | } |
| 2576 | |
| 2577 | /* Check if small groups should be split. Return true if no group |
| 2578 | contains more than two uses with distinct addr_offsets. Return |
| 2579 | false otherwise. We want to split such groups because: |
| 2580 | |
| 2581 | 1) Small groups don't have much benefit and may interfer with |
| 2582 | general candidate selection. |
| 2583 | 2) Size for problem with only small groups is usually small and |
| 2584 | general algorithm can handle it well. |
| 2585 | |
| 2586 | TODO -- Above claim may not hold when we want to merge memory |
| 2587 | accesses with conseuctive addresses. */ |
| 2588 | |
| 2589 | static bool |
| 2590 | split_small_address_groups_p (struct ivopts_data *data) |
| 2591 | { |
| 2592 | unsigned int i, j, distinct = 1; |
| 2593 | struct iv_use *pre; |
| 2594 | struct iv_group *group; |
| 2595 | |
| 2596 | for (i = 0; i < data->vgroups.length (); i++) |
| 2597 | { |
| 2598 | group = data->vgroups[i]; |
| 2599 | if (group->vuses.length () == 1) |
| 2600 | continue; |
| 2601 | |
| 2602 | gcc_assert (address_p (group->type)); |
| 2603 | if (group->vuses.length () == 2) |
| 2604 | { |
| 2605 | if (compare_sizes_for_sort (a: group->vuses[0]->addr_offset, |
| 2606 | b: group->vuses[1]->addr_offset) > 0) |
| 2607 | std::swap (a&: group->vuses[0], b&: group->vuses[1]); |
| 2608 | } |
| 2609 | else |
| 2610 | group->vuses.qsort (group_compare_offset); |
| 2611 | |
| 2612 | if (distinct > 2) |
| 2613 | continue; |
| 2614 | |
| 2615 | distinct = 1; |
| 2616 | for (pre = group->vuses[0], j = 1; j < group->vuses.length (); j++) |
| 2617 | { |
| 2618 | if (maybe_ne (a: group->vuses[j]->addr_offset, b: pre->addr_offset)) |
| 2619 | { |
| 2620 | pre = group->vuses[j]; |
| 2621 | distinct++; |
| 2622 | } |
| 2623 | |
| 2624 | if (distinct > 2) |
| 2625 | break; |
| 2626 | } |
| 2627 | } |
| 2628 | |
| 2629 | return (distinct <= 2); |
| 2630 | } |
| 2631 | |
| 2632 | /* For each group of address type uses, this function further groups |
| 2633 | these uses according to the maximum offset supported by target's |
| 2634 | [base + offset] addressing mode. */ |
| 2635 | |
| 2636 | static void |
| 2637 | split_address_groups (struct ivopts_data *data) |
| 2638 | { |
| 2639 | unsigned int i, j; |
| 2640 | /* Always split group. */ |
| 2641 | bool split_p = split_small_address_groups_p (data); |
| 2642 | |
| 2643 | for (i = 0; i < data->vgroups.length (); i++) |
| 2644 | { |
| 2645 | struct iv_group *new_group = NULL; |
| 2646 | struct iv_group *group = data->vgroups[i]; |
| 2647 | struct iv_use *use = group->vuses[0]; |
| 2648 | |
| 2649 | use->id = 0; |
| 2650 | use->group_id = group->id; |
| 2651 | if (group->vuses.length () == 1) |
| 2652 | continue; |
| 2653 | |
| 2654 | gcc_assert (address_p (use->type)); |
| 2655 | |
| 2656 | for (j = 1; j < group->vuses.length ();) |
| 2657 | { |
| 2658 | struct iv_use *next = group->vuses[j]; |
| 2659 | poly_int64 offset = next->addr_offset - use->addr_offset; |
| 2660 | |
| 2661 | /* Split group if aksed to, or the offset against the first |
| 2662 | use can't fit in offset part of addressing mode. IV uses |
| 2663 | having the same offset are still kept in one group. */ |
| 2664 | if (maybe_ne (a: offset, b: 0) |
| 2665 | && (split_p || !addr_offset_valid_p (use, offset))) |
| 2666 | { |
| 2667 | if (!new_group) |
| 2668 | new_group = record_group (data, type: group->type); |
| 2669 | group->vuses.ordered_remove (ix: j); |
| 2670 | new_group->vuses.safe_push (obj: next); |
| 2671 | continue; |
| 2672 | } |
| 2673 | |
| 2674 | next->id = j; |
| 2675 | next->group_id = group->id; |
| 2676 | j++; |
| 2677 | } |
| 2678 | } |
| 2679 | } |
| 2680 | |
| 2681 | /* Finds uses of the induction variables that are interesting. */ |
| 2682 | |
| 2683 | static void |
| 2684 | find_interesting_uses (struct ivopts_data *data, basic_block *body) |
| 2685 | { |
| 2686 | basic_block bb; |
| 2687 | gimple_stmt_iterator bsi; |
| 2688 | unsigned i; |
| 2689 | edge e; |
| 2690 | |
| 2691 | for (i = 0; i < data->current_loop->num_nodes; i++) |
| 2692 | { |
| 2693 | edge_iterator ei; |
| 2694 | bb = body[i]; |
| 2695 | |
| 2696 | FOR_EACH_EDGE (e, ei, bb->succs) |
| 2697 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
| 2698 | && !flow_bb_inside_loop_p (data->current_loop, e->dest)) |
| 2699 | find_interesting_uses_outside (data, exit: e); |
| 2700 | |
| 2701 | for (bsi = gsi_start_phis (bb); !gsi_end_p (i: bsi); gsi_next (i: &bsi)) |
| 2702 | find_interesting_uses_stmt (data, stmt: gsi_stmt (i: bsi)); |
| 2703 | for (bsi = gsi_start_bb (bb); !gsi_end_p (i: bsi); gsi_next (i: &bsi)) |
| 2704 | if (!is_gimple_debug (gs: gsi_stmt (i: bsi))) |
| 2705 | find_interesting_uses_stmt (data, stmt: gsi_stmt (i: bsi)); |
| 2706 | } |
| 2707 | |
| 2708 | split_address_groups (data); |
| 2709 | |
| 2710 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 2711 | { |
| 2712 | fprintf (stream: dump_file, format: "\n<IV Groups>:\n" ); |
| 2713 | dump_groups (file: dump_file, data); |
| 2714 | fprintf (stream: dump_file, format: "\n" ); |
| 2715 | } |
| 2716 | } |
| 2717 | |
| 2718 | /* Strips constant offsets from EXPR and stores them to OFFSET. If INSIDE_ADDR |
| 2719 | is true, assume we are inside an address. If TOP_COMPREF is true, assume |
| 2720 | we are at the top-level of the processed address. */ |
| 2721 | |
| 2722 | static tree |
| 2723 | strip_offset_1 (tree expr, bool inside_addr, bool top_compref, |
| 2724 | poly_int64 *offset) |
| 2725 | { |
| 2726 | tree op0 = NULL_TREE, op1 = NULL_TREE, tmp, step; |
| 2727 | enum tree_code code; |
| 2728 | tree type, orig_type = TREE_TYPE (expr); |
| 2729 | poly_int64 off0, off1; |
| 2730 | HOST_WIDE_INT st; |
| 2731 | tree orig_expr = expr; |
| 2732 | |
| 2733 | STRIP_NOPS (expr); |
| 2734 | |
| 2735 | type = TREE_TYPE (expr); |
| 2736 | code = TREE_CODE (expr); |
| 2737 | *offset = 0; |
| 2738 | |
| 2739 | switch (code) |
| 2740 | { |
| 2741 | case POINTER_PLUS_EXPR: |
| 2742 | case PLUS_EXPR: |
| 2743 | case MINUS_EXPR: |
| 2744 | op0 = TREE_OPERAND (expr, 0); |
| 2745 | op1 = TREE_OPERAND (expr, 1); |
| 2746 | |
| 2747 | op0 = strip_offset_1 (expr: op0, inside_addr: false, top_compref: false, offset: &off0); |
| 2748 | op1 = strip_offset_1 (expr: op1, inside_addr: false, top_compref: false, offset: &off1); |
| 2749 | |
| 2750 | *offset = (code == MINUS_EXPR ? off0 - off1 : off0 + off1); |
| 2751 | if (op0 == TREE_OPERAND (expr, 0) |
| 2752 | && op1 == TREE_OPERAND (expr, 1)) |
| 2753 | return orig_expr; |
| 2754 | |
| 2755 | if (integer_zerop (op1)) |
| 2756 | expr = op0; |
| 2757 | else if (integer_zerop (op0)) |
| 2758 | { |
| 2759 | if (code == MINUS_EXPR) |
| 2760 | { |
| 2761 | if (TYPE_OVERFLOW_UNDEFINED (type)) |
| 2762 | { |
| 2763 | type = unsigned_type_for (type); |
| 2764 | op1 = fold_convert (type, op1); |
| 2765 | } |
| 2766 | expr = fold_build1 (NEGATE_EXPR, type, op1); |
| 2767 | } |
| 2768 | else |
| 2769 | expr = op1; |
| 2770 | } |
| 2771 | else |
| 2772 | { |
| 2773 | if (TYPE_OVERFLOW_UNDEFINED (type)) |
| 2774 | { |
| 2775 | type = unsigned_type_for (type); |
| 2776 | if (code == POINTER_PLUS_EXPR) |
| 2777 | code = PLUS_EXPR; |
| 2778 | op0 = fold_convert (type, op0); |
| 2779 | op1 = fold_convert (type, op1); |
| 2780 | } |
| 2781 | expr = fold_build2 (code, type, op0, op1); |
| 2782 | } |
| 2783 | |
| 2784 | return fold_convert (orig_type, expr); |
| 2785 | |
| 2786 | case MULT_EXPR: |
| 2787 | op1 = TREE_OPERAND (expr, 1); |
| 2788 | if (!cst_and_fits_in_hwi (op1)) |
| 2789 | return orig_expr; |
| 2790 | |
| 2791 | op0 = TREE_OPERAND (expr, 0); |
| 2792 | op0 = strip_offset_1 (expr: op0, inside_addr: false, top_compref: false, offset: &off0); |
| 2793 | if (op0 == TREE_OPERAND (expr, 0)) |
| 2794 | return orig_expr; |
| 2795 | |
| 2796 | *offset = off0 * int_cst_value (op1); |
| 2797 | if (integer_zerop (op0)) |
| 2798 | expr = op0; |
| 2799 | else |
| 2800 | { |
| 2801 | if (TYPE_OVERFLOW_UNDEFINED (type)) |
| 2802 | { |
| 2803 | type = unsigned_type_for (type); |
| 2804 | op0 = fold_convert (type, op0); |
| 2805 | op1 = fold_convert (type, op1); |
| 2806 | } |
| 2807 | expr = fold_build2 (MULT_EXPR, type, op0, op1); |
| 2808 | } |
| 2809 | |
| 2810 | return fold_convert (orig_type, expr); |
| 2811 | |
| 2812 | case ARRAY_REF: |
| 2813 | case ARRAY_RANGE_REF: |
| 2814 | if (!inside_addr) |
| 2815 | return orig_expr; |
| 2816 | |
| 2817 | step = array_ref_element_size (expr); |
| 2818 | if (!cst_and_fits_in_hwi (step)) |
| 2819 | break; |
| 2820 | |
| 2821 | st = int_cst_value (step); |
| 2822 | op1 = TREE_OPERAND (expr, 1); |
| 2823 | op1 = strip_offset_1 (expr: op1, inside_addr: false, top_compref: false, offset: &off1); |
| 2824 | *offset = off1 * st; |
| 2825 | |
| 2826 | if (top_compref |
| 2827 | && integer_zerop (op1)) |
| 2828 | { |
| 2829 | /* Strip the component reference completely. */ |
| 2830 | op0 = TREE_OPERAND (expr, 0); |
| 2831 | op0 = strip_offset_1 (expr: op0, inside_addr, top_compref, offset: &off0); |
| 2832 | *offset += off0; |
| 2833 | return op0; |
| 2834 | } |
| 2835 | break; |
| 2836 | |
| 2837 | case COMPONENT_REF: |
| 2838 | { |
| 2839 | tree field; |
| 2840 | |
| 2841 | if (!inside_addr) |
| 2842 | return orig_expr; |
| 2843 | |
| 2844 | tmp = component_ref_field_offset (expr); |
| 2845 | field = TREE_OPERAND (expr, 1); |
| 2846 | if (top_compref |
| 2847 | && cst_and_fits_in_hwi (tmp) |
| 2848 | && cst_and_fits_in_hwi (DECL_FIELD_BIT_OFFSET (field))) |
| 2849 | { |
| 2850 | HOST_WIDE_INT boffset, abs_off; |
| 2851 | |
| 2852 | /* Strip the component reference completely. */ |
| 2853 | op0 = TREE_OPERAND (expr, 0); |
| 2854 | op0 = strip_offset_1 (expr: op0, inside_addr, top_compref, offset: &off0); |
| 2855 | boffset = int_cst_value (DECL_FIELD_BIT_OFFSET (field)); |
| 2856 | abs_off = abs_hwi (x: boffset) / BITS_PER_UNIT; |
| 2857 | if (boffset < 0) |
| 2858 | abs_off = -abs_off; |
| 2859 | |
| 2860 | *offset = off0 + int_cst_value (tmp) + abs_off; |
| 2861 | return op0; |
| 2862 | } |
| 2863 | } |
| 2864 | break; |
| 2865 | |
| 2866 | case ADDR_EXPR: |
| 2867 | op0 = TREE_OPERAND (expr, 0); |
| 2868 | op0 = strip_offset_1 (expr: op0, inside_addr: true, top_compref: true, offset: &off0); |
| 2869 | *offset += off0; |
| 2870 | |
| 2871 | if (op0 == TREE_OPERAND (expr, 0)) |
| 2872 | return orig_expr; |
| 2873 | |
| 2874 | expr = build_fold_addr_expr (op0); |
| 2875 | return fold_convert (orig_type, expr); |
| 2876 | |
| 2877 | case MEM_REF: |
| 2878 | /* ??? Offset operand? */ |
| 2879 | inside_addr = false; |
| 2880 | break; |
| 2881 | |
| 2882 | default: |
| 2883 | if (ptrdiff_tree_p (expr, offset) && maybe_ne (a: *offset, b: 0)) |
| 2884 | return build_int_cst (orig_type, 0); |
| 2885 | return orig_expr; |
| 2886 | } |
| 2887 | |
| 2888 | /* Default handling of expressions for that we want to recurse into |
| 2889 | the first operand. */ |
| 2890 | op0 = TREE_OPERAND (expr, 0); |
| 2891 | op0 = strip_offset_1 (expr: op0, inside_addr, top_compref: false, offset: &off0); |
| 2892 | *offset += off0; |
| 2893 | |
| 2894 | if (op0 == TREE_OPERAND (expr, 0) |
| 2895 | && (!op1 || op1 == TREE_OPERAND (expr, 1))) |
| 2896 | return orig_expr; |
| 2897 | |
| 2898 | expr = copy_node (expr); |
| 2899 | TREE_OPERAND (expr, 0) = op0; |
| 2900 | if (op1) |
| 2901 | TREE_OPERAND (expr, 1) = op1; |
| 2902 | |
| 2903 | /* Inside address, we might strip the top level component references, |
| 2904 | thus changing type of the expression. Handling of ADDR_EXPR |
| 2905 | will fix that. */ |
| 2906 | expr = fold_convert (orig_type, expr); |
| 2907 | |
| 2908 | return expr; |
| 2909 | } |
| 2910 | |
| 2911 | /* Strips constant offsets from EXPR and stores them to OFFSET. */ |
| 2912 | |
| 2913 | static tree |
| 2914 | strip_offset (tree expr, poly_uint64 *offset) |
| 2915 | { |
| 2916 | poly_int64 off; |
| 2917 | tree core = strip_offset_1 (expr, inside_addr: false, top_compref: false, offset: &off); |
| 2918 | *offset = off; |
| 2919 | return core; |
| 2920 | } |
| 2921 | |
| 2922 | /* Returns variant of TYPE that can be used as base for different uses. |
| 2923 | We return unsigned type with the same precision, which avoids problems |
| 2924 | with overflows. */ |
| 2925 | |
| 2926 | static tree |
| 2927 | generic_type_for (tree type) |
| 2928 | { |
| 2929 | if (POINTER_TYPE_P (type)) |
| 2930 | return unsigned_type_for (type); |
| 2931 | |
| 2932 | if (TYPE_UNSIGNED (type)) |
| 2933 | return type; |
| 2934 | |
| 2935 | return unsigned_type_for (type); |
| 2936 | } |
| 2937 | |
| 2938 | /* Private data for walk_tree. */ |
| 2939 | |
| 2940 | struct walk_tree_data |
| 2941 | { |
| 2942 | bitmap *inv_vars; |
| 2943 | struct ivopts_data *idata; |
| 2944 | }; |
| 2945 | |
| 2946 | /* Callback function for walk_tree, it records invariants and symbol |
| 2947 | reference in *EXPR_P. DATA is the structure storing result info. */ |
| 2948 | |
| 2949 | static tree |
| 2950 | find_inv_vars_cb (tree *expr_p, int *ws ATTRIBUTE_UNUSED, void *data) |
| 2951 | { |
| 2952 | tree op = *expr_p; |
| 2953 | struct version_info *info; |
| 2954 | struct walk_tree_data *wdata = (struct walk_tree_data*) data; |
| 2955 | |
| 2956 | if (TREE_CODE (op) != SSA_NAME) |
| 2957 | return NULL_TREE; |
| 2958 | |
| 2959 | info = name_info (data: wdata->idata, name: op); |
| 2960 | /* Because we expand simple operations when finding IVs, loop invariant |
| 2961 | variable that isn't referred by the original loop could be used now. |
| 2962 | Record such invariant variables here. */ |
| 2963 | if (!info->iv) |
| 2964 | { |
| 2965 | struct ivopts_data *idata = wdata->idata; |
| 2966 | basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (op)); |
| 2967 | |
| 2968 | if (!bb || !flow_bb_inside_loop_p (idata->current_loop, bb)) |
| 2969 | { |
| 2970 | tree steptype = TREE_TYPE (op); |
| 2971 | if (POINTER_TYPE_P (steptype)) |
| 2972 | steptype = sizetype; |
| 2973 | set_iv (data: idata, iv: op, base: op, step: build_int_cst (steptype, 0), no_overflow: true); |
| 2974 | record_invariant (data: idata, op, nonlinear_use: false); |
| 2975 | } |
| 2976 | } |
| 2977 | if (!info->inv_id || info->has_nonlin_use) |
| 2978 | return NULL_TREE; |
| 2979 | |
| 2980 | if (!*wdata->inv_vars) |
| 2981 | *wdata->inv_vars = BITMAP_ALLOC (NULL); |
| 2982 | bitmap_set_bit (*wdata->inv_vars, info->inv_id); |
| 2983 | |
| 2984 | return NULL_TREE; |
| 2985 | } |
| 2986 | |
| 2987 | /* Records invariants in *EXPR_P. INV_VARS is the bitmap to that we should |
| 2988 | store it. */ |
| 2989 | |
| 2990 | static inline void |
| 2991 | find_inv_vars (struct ivopts_data *data, tree *expr_p, bitmap *inv_vars) |
| 2992 | { |
| 2993 | struct walk_tree_data wdata; |
| 2994 | |
| 2995 | if (!inv_vars) |
| 2996 | return; |
| 2997 | |
| 2998 | wdata.idata = data; |
| 2999 | wdata.inv_vars = inv_vars; |
| 3000 | walk_tree (expr_p, find_inv_vars_cb, &wdata, NULL); |
| 3001 | } |
| 3002 | |
| 3003 | /* Get entry from invariant expr hash table for INV_EXPR. New entry |
| 3004 | will be recorded if it doesn't exist yet. Given below two exprs: |
| 3005 | inv_expr + cst1, inv_expr + cst2 |
| 3006 | It's hard to make decision whether constant part should be stripped |
| 3007 | or not. We choose to not strip based on below facts: |
| 3008 | 1) We need to count ADD cost for constant part if it's stripped, |
| 3009 | which isn't always trivial where this functions is called. |
| 3010 | 2) Stripping constant away may be conflict with following loop |
| 3011 | invariant hoisting pass. |
| 3012 | 3) Not stripping constant away results in more invariant exprs, |
| 3013 | which usually leads to decision preferring lower reg pressure. */ |
| 3014 | |
| 3015 | static iv_inv_expr_ent * |
| 3016 | get_loop_invariant_expr (struct ivopts_data *data, tree inv_expr) |
| 3017 | { |
| 3018 | STRIP_NOPS (inv_expr); |
| 3019 | |
| 3020 | if (poly_int_tree_p (t: inv_expr) |
| 3021 | || TREE_CODE (inv_expr) == SSA_NAME) |
| 3022 | return NULL; |
| 3023 | |
| 3024 | /* Don't strip constant part away as we used to. */ |
| 3025 | |
| 3026 | /* Stores EXPR in DATA->inv_expr_tab, return pointer to iv_inv_expr_ent. */ |
| 3027 | struct iv_inv_expr_ent ent; |
| 3028 | ent.expr = inv_expr; |
| 3029 | ent.hash = iterative_hash_expr (tree: inv_expr, seed: 0); |
| 3030 | struct iv_inv_expr_ent **slot = data->inv_expr_tab->find_slot (value: &ent, insert: INSERT); |
| 3031 | |
| 3032 | if (!*slot) |
| 3033 | { |
| 3034 | *slot = XNEW (struct iv_inv_expr_ent); |
| 3035 | (*slot)->expr = inv_expr; |
| 3036 | (*slot)->hash = ent.hash; |
| 3037 | (*slot)->id = ++data->max_inv_expr_id; |
| 3038 | } |
| 3039 | |
| 3040 | return *slot; |
| 3041 | } |
| 3042 | |
| 3043 | |
| 3044 | /* Return *TP if it is an SSA_NAME marked with TREE_VISITED, i.e., as |
| 3045 | unsuitable as ivopts candidates for potentially involving undefined |
| 3046 | behavior. */ |
| 3047 | |
| 3048 | static tree |
| 3049 | find_ssa_undef (tree *tp, int *walk_subtrees, void *bb_) |
| 3050 | { |
| 3051 | basic_block bb = (basic_block) bb_; |
| 3052 | if (TREE_CODE (*tp) == SSA_NAME |
| 3053 | && ssa_name_maybe_undef_p (var: *tp) |
| 3054 | && !ssa_name_any_use_dominates_bb_p (var: *tp, bb)) |
| 3055 | return *tp; |
| 3056 | if (!EXPR_P (*tp)) |
| 3057 | *walk_subtrees = 0; |
| 3058 | return NULL; |
| 3059 | } |
| 3060 | |
| 3061 | /* Adds a candidate BASE + STEP * i. Important field is set to IMPORTANT and |
| 3062 | position to POS. If USE is not NULL, the candidate is set as related to |
| 3063 | it. If both BASE and STEP are NULL, we add a pseudocandidate for the |
| 3064 | replacement of the final value of the iv by a direct computation. */ |
| 3065 | |
| 3066 | static struct iv_cand * |
| 3067 | add_candidate_1 (struct ivopts_data *data, tree base, tree step, bool important, |
| 3068 | enum iv_position pos, struct iv_use *use, |
| 3069 | gimple *incremented_at, struct iv *orig_iv = NULL, |
| 3070 | bool doloop = false) |
| 3071 | { |
| 3072 | unsigned i; |
| 3073 | struct iv_cand *cand = NULL; |
| 3074 | tree type, orig_type; |
| 3075 | |
| 3076 | gcc_assert (base && step); |
| 3077 | |
| 3078 | /* -fkeep-gc-roots-live means that we have to keep a real pointer |
| 3079 | live, but the ivopts code may replace a real pointer with one |
| 3080 | pointing before or after the memory block that is then adjusted |
| 3081 | into the memory block during the loop. FIXME: It would likely be |
| 3082 | better to actually force the pointer live and still use ivopts; |
| 3083 | for example, it would be enough to write the pointer into memory |
| 3084 | and keep it there until after the loop. */ |
| 3085 | if (flag_keep_gc_roots_live && POINTER_TYPE_P (TREE_TYPE (base))) |
| 3086 | return NULL; |
| 3087 | |
| 3088 | /* If BASE contains undefined SSA names make sure we only record |
| 3089 | the original IV. */ |
| 3090 | bool involves_undefs = false; |
| 3091 | if (walk_tree (&base, find_ssa_undef, data->current_loop->header, NULL)) |
| 3092 | { |
| 3093 | if (pos != IP_ORIGINAL) |
| 3094 | return NULL; |
| 3095 | important = false; |
| 3096 | involves_undefs = true; |
| 3097 | } |
| 3098 | |
| 3099 | /* For non-original variables, make sure their values are computed in a type |
| 3100 | that does not invoke undefined behavior on overflows (since in general, |
| 3101 | we cannot prove that these induction variables are non-wrapping). */ |
| 3102 | if (pos != IP_ORIGINAL) |
| 3103 | { |
| 3104 | orig_type = TREE_TYPE (base); |
| 3105 | type = generic_type_for (type: orig_type); |
| 3106 | if (type != orig_type) |
| 3107 | { |
| 3108 | base = fold_convert (type, base); |
| 3109 | step = fold_convert (type, step); |
| 3110 | } |
| 3111 | } |
| 3112 | |
| 3113 | for (i = 0; i < data->vcands.length (); i++) |
| 3114 | { |
| 3115 | cand = data->vcands[i]; |
| 3116 | |
| 3117 | if (cand->pos != pos) |
| 3118 | continue; |
| 3119 | |
| 3120 | if (cand->incremented_at != incremented_at |
| 3121 | || ((pos == IP_AFTER_USE || pos == IP_BEFORE_USE) |
| 3122 | && cand->ainc_use != use)) |
| 3123 | continue; |
| 3124 | |
| 3125 | if (operand_equal_p (base, cand->iv->base, flags: 0) |
| 3126 | && operand_equal_p (step, cand->iv->step, flags: 0) |
| 3127 | && (TYPE_PRECISION (TREE_TYPE (base)) |
| 3128 | == TYPE_PRECISION (TREE_TYPE (cand->iv->base)))) |
| 3129 | break; |
| 3130 | } |
| 3131 | |
| 3132 | if (i == data->vcands.length ()) |
| 3133 | { |
| 3134 | cand = XCNEW (struct iv_cand); |
| 3135 | cand->id = i; |
| 3136 | cand->iv = alloc_iv (data, base, step); |
| 3137 | cand->pos = pos; |
| 3138 | if (pos != IP_ORIGINAL) |
| 3139 | { |
| 3140 | if (doloop) |
| 3141 | cand->var_before = create_tmp_var_raw (TREE_TYPE (base), "doloop" ); |
| 3142 | else |
| 3143 | cand->var_before = create_tmp_var_raw (TREE_TYPE (base), "ivtmp" ); |
| 3144 | cand->var_after = cand->var_before; |
| 3145 | } |
| 3146 | cand->important = important; |
| 3147 | cand->involves_undefs = involves_undefs; |
| 3148 | cand->incremented_at = incremented_at; |
| 3149 | cand->doloop_p = doloop; |
| 3150 | data->vcands.safe_push (obj: cand); |
| 3151 | |
| 3152 | if (!poly_int_tree_p (t: step)) |
| 3153 | { |
| 3154 | find_inv_vars (data, expr_p: &step, inv_vars: &cand->inv_vars); |
| 3155 | |
| 3156 | iv_inv_expr_ent *inv_expr = get_loop_invariant_expr (data, inv_expr: step); |
| 3157 | /* Share bitmap between inv_vars and inv_exprs for cand. */ |
| 3158 | if (inv_expr != NULL) |
| 3159 | { |
| 3160 | cand->inv_exprs = cand->inv_vars; |
| 3161 | cand->inv_vars = NULL; |
| 3162 | if (cand->inv_exprs) |
| 3163 | bitmap_clear (cand->inv_exprs); |
| 3164 | else |
| 3165 | cand->inv_exprs = BITMAP_ALLOC (NULL); |
| 3166 | |
| 3167 | bitmap_set_bit (cand->inv_exprs, inv_expr->id); |
| 3168 | } |
| 3169 | } |
| 3170 | |
| 3171 | if (pos == IP_AFTER_USE || pos == IP_BEFORE_USE) |
| 3172 | cand->ainc_use = use; |
| 3173 | else |
| 3174 | cand->ainc_use = NULL; |
| 3175 | |
| 3176 | cand->orig_iv = orig_iv; |
| 3177 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 3178 | dump_cand (file: dump_file, cand); |
| 3179 | } |
| 3180 | |
| 3181 | cand->important |= important; |
| 3182 | cand->doloop_p |= doloop; |
| 3183 | |
| 3184 | /* Relate candidate to the group for which it is added. */ |
| 3185 | if (use) |
| 3186 | bitmap_set_bit (data->vgroups[use->group_id]->related_cands, i); |
| 3187 | |
| 3188 | return cand; |
| 3189 | } |
| 3190 | |
| 3191 | /* Returns true if incrementing the induction variable at the end of the LOOP |
| 3192 | is allowed. |
| 3193 | |
| 3194 | The purpose is to avoid splitting latch edge with a biv increment, thus |
| 3195 | creating a jump, possibly confusing other optimization passes and leaving |
| 3196 | less freedom to scheduler. So we allow IP_END only if IP_NORMAL is not |
| 3197 | available (so we do not have a better alternative), or if the latch edge |
| 3198 | is already nonempty. */ |
| 3199 | |
| 3200 | static bool |
| 3201 | allow_ip_end_pos_p (class loop *loop) |
| 3202 | { |
| 3203 | /* Do not allow IP_END when creating the IV would need to split the |
| 3204 | latch edge as that makes all IP_NORMAL invalid. */ |
| 3205 | auto pos = gsi_last_bb (bb: ip_end_pos (loop)); |
| 3206 | if (!gsi_end_p (i: pos) && stmt_ends_bb_p (*pos)) |
| 3207 | return false; |
| 3208 | |
| 3209 | if (!ip_normal_pos (loop)) |
| 3210 | return true; |
| 3211 | |
| 3212 | if (!empty_block_p (ip_end_pos (loop))) |
| 3213 | return true; |
| 3214 | |
| 3215 | return false; |
| 3216 | } |
| 3217 | |
| 3218 | /* If possible, adds autoincrement candidates BASE + STEP * i based on use USE. |
| 3219 | Important field is set to IMPORTANT. */ |
| 3220 | |
| 3221 | static void |
| 3222 | add_autoinc_candidates (struct ivopts_data *data, tree base, tree step, |
| 3223 | bool important, struct iv_use *use) |
| 3224 | { |
| 3225 | basic_block use_bb = gimple_bb (g: use->stmt); |
| 3226 | machine_mode mem_mode; |
| 3227 | unsigned HOST_WIDE_INT cstepi; |
| 3228 | |
| 3229 | /* If we insert the increment in any position other than the standard |
| 3230 | ones, we must ensure that it is incremented once per iteration. |
| 3231 | It must not be in an inner nested loop, or one side of an if |
| 3232 | statement. */ |
| 3233 | if (use_bb->loop_father != data->current_loop |
| 3234 | || !dominated_by_p (CDI_DOMINATORS, data->current_loop->latch, use_bb) |
| 3235 | || stmt_can_throw_internal (cfun, use->stmt) |
| 3236 | || !cst_and_fits_in_hwi (step)) |
| 3237 | return; |
| 3238 | |
| 3239 | cstepi = int_cst_value (step); |
| 3240 | |
| 3241 | mem_mode = TYPE_MODE (use->mem_type); |
| 3242 | if (((USE_LOAD_PRE_INCREMENT (mem_mode) |
| 3243 | || USE_STORE_PRE_INCREMENT (mem_mode)) |
| 3244 | && known_eq (GET_MODE_SIZE (mem_mode), cstepi)) |
| 3245 | || ((USE_LOAD_PRE_DECREMENT (mem_mode) |
| 3246 | || USE_STORE_PRE_DECREMENT (mem_mode)) |
| 3247 | && known_eq (GET_MODE_SIZE (mem_mode), -cstepi))) |
| 3248 | { |
| 3249 | enum tree_code code = MINUS_EXPR; |
| 3250 | tree new_base; |
| 3251 | tree new_step = step; |
| 3252 | |
| 3253 | if (POINTER_TYPE_P (TREE_TYPE (base))) |
| 3254 | { |
| 3255 | new_step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step); |
| 3256 | code = POINTER_PLUS_EXPR; |
| 3257 | } |
| 3258 | else |
| 3259 | new_step = fold_convert (TREE_TYPE (base), new_step); |
| 3260 | new_base = fold_build2 (code, TREE_TYPE (base), base, new_step); |
| 3261 | add_candidate_1 (data, base: new_base, step, important, pos: IP_BEFORE_USE, use, |
| 3262 | incremented_at: use->stmt); |
| 3263 | } |
| 3264 | if (((USE_LOAD_POST_INCREMENT (mem_mode) |
| 3265 | || USE_STORE_POST_INCREMENT (mem_mode)) |
| 3266 | && known_eq (GET_MODE_SIZE (mem_mode), cstepi)) |
| 3267 | || ((USE_LOAD_POST_DECREMENT (mem_mode) |
| 3268 | || USE_STORE_POST_DECREMENT (mem_mode)) |
| 3269 | && known_eq (GET_MODE_SIZE (mem_mode), -cstepi))) |
| 3270 | { |
| 3271 | add_candidate_1 (data, base, step, important, pos: IP_AFTER_USE, use, |
| 3272 | incremented_at: use->stmt); |
| 3273 | } |
| 3274 | } |
| 3275 | |
| 3276 | /* Adds a candidate BASE + STEP * i. Important field is set to IMPORTANT and |
| 3277 | position to POS. If USE is not NULL, the candidate is set as related to |
| 3278 | it. The candidate computation is scheduled before exit condition and at |
| 3279 | the end of loop. */ |
| 3280 | |
| 3281 | static void |
| 3282 | add_candidate (struct ivopts_data *data, tree base, tree step, bool important, |
| 3283 | struct iv_use *use, struct iv *orig_iv = NULL, |
| 3284 | bool doloop = false) |
| 3285 | { |
| 3286 | if (ip_normal_pos (data->current_loop)) |
| 3287 | add_candidate_1 (data, base, step, important, pos: IP_NORMAL, use, NULL, orig_iv, |
| 3288 | doloop); |
| 3289 | /* Exclude doloop candidate here since it requires decrement then comparison |
| 3290 | and jump, the IP_END position doesn't match. */ |
| 3291 | if (!doloop && ip_end_pos (data->current_loop) |
| 3292 | && allow_ip_end_pos_p (loop: data->current_loop)) |
| 3293 | add_candidate_1 (data, base, step, important, pos: IP_END, use, NULL, orig_iv); |
| 3294 | } |
| 3295 | |
| 3296 | /* Adds standard iv candidates. */ |
| 3297 | |
| 3298 | static void |
| 3299 | add_standard_iv_candidates (struct ivopts_data *data) |
| 3300 | { |
| 3301 | add_candidate (data, integer_zero_node, integer_one_node, important: true, NULL); |
| 3302 | |
| 3303 | /* The same for a double-integer type if it is still fast enough. */ |
| 3304 | if (TYPE_PRECISION |
| 3305 | (long_integer_type_node) > TYPE_PRECISION (integer_type_node) |
| 3306 | && TYPE_PRECISION (long_integer_type_node) <= BITS_PER_WORD) |
| 3307 | add_candidate (data, base: build_int_cst (long_integer_type_node, 0), |
| 3308 | step: build_int_cst (long_integer_type_node, 1), important: true, NULL); |
| 3309 | |
| 3310 | /* The same for a double-integer type if it is still fast enough. */ |
| 3311 | if (TYPE_PRECISION |
| 3312 | (long_long_integer_type_node) > TYPE_PRECISION (long_integer_type_node) |
| 3313 | && TYPE_PRECISION (long_long_integer_type_node) <= BITS_PER_WORD) |
| 3314 | add_candidate (data, base: build_int_cst (long_long_integer_type_node, 0), |
| 3315 | step: build_int_cst (long_long_integer_type_node, 1), important: true, NULL); |
| 3316 | } |
| 3317 | |
| 3318 | |
| 3319 | /* Adds candidates bases on the old induction variable IV. */ |
| 3320 | |
| 3321 | static void |
| 3322 | add_iv_candidate_for_biv (struct ivopts_data *data, struct iv *iv) |
| 3323 | { |
| 3324 | gimple *phi; |
| 3325 | tree def; |
| 3326 | struct iv_cand *cand; |
| 3327 | |
| 3328 | /* Check if this biv is used in address type use. */ |
| 3329 | if (iv->no_overflow && iv->have_address_use |
| 3330 | && INTEGRAL_TYPE_P (TREE_TYPE (iv->base)) |
| 3331 | && TYPE_PRECISION (TREE_TYPE (iv->base)) < TYPE_PRECISION (sizetype)) |
| 3332 | { |
| 3333 | tree base = fold_convert (sizetype, iv->base); |
| 3334 | tree step = fold_convert (sizetype, iv->step); |
| 3335 | |
| 3336 | /* Add iv cand of same precision as index part in TARGET_MEM_REF. */ |
| 3337 | add_candidate (data, base, step, important: true, NULL, orig_iv: iv); |
| 3338 | /* Add iv cand of the original type only if it has nonlinear use. */ |
| 3339 | if (iv->nonlin_use) |
| 3340 | add_candidate (data, base: iv->base, step: iv->step, important: true, NULL); |
| 3341 | } |
| 3342 | else |
| 3343 | add_candidate (data, base: iv->base, step: iv->step, important: true, NULL); |
| 3344 | |
| 3345 | /* The same, but with initial value zero. */ |
| 3346 | if (POINTER_TYPE_P (TREE_TYPE (iv->base))) |
| 3347 | add_candidate (data, size_int (0), step: iv->step, important: true, NULL); |
| 3348 | else |
| 3349 | add_candidate (data, base: build_int_cst (TREE_TYPE (iv->base), 0), |
| 3350 | step: iv->step, important: true, NULL); |
| 3351 | |
| 3352 | phi = SSA_NAME_DEF_STMT (iv->ssa_name); |
| 3353 | if (gimple_code (g: phi) == GIMPLE_PHI) |
| 3354 | { |
| 3355 | /* Additionally record the possibility of leaving the original iv |
| 3356 | untouched. */ |
| 3357 | def = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (data->current_loop)); |
| 3358 | /* Don't add candidate if it's from another PHI node because |
| 3359 | it's an affine iv appearing in the form of PEELED_CHREC. */ |
| 3360 | phi = SSA_NAME_DEF_STMT (def); |
| 3361 | if (gimple_code (g: phi) != GIMPLE_PHI) |
| 3362 | { |
| 3363 | cand = add_candidate_1 (data, |
| 3364 | base: iv->base, step: iv->step, important: true, pos: IP_ORIGINAL, NULL, |
| 3365 | SSA_NAME_DEF_STMT (def)); |
| 3366 | if (cand) |
| 3367 | { |
| 3368 | cand->var_before = iv->ssa_name; |
| 3369 | cand->var_after = def; |
| 3370 | } |
| 3371 | } |
| 3372 | else |
| 3373 | gcc_assert (gimple_bb (phi) == data->current_loop->header); |
| 3374 | } |
| 3375 | } |
| 3376 | |
| 3377 | /* Adds candidates based on the old induction variables. */ |
| 3378 | |
| 3379 | static void |
| 3380 | add_iv_candidate_for_bivs (struct ivopts_data *data) |
| 3381 | { |
| 3382 | unsigned i; |
| 3383 | struct iv *iv; |
| 3384 | bitmap_iterator bi; |
| 3385 | |
| 3386 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
| 3387 | { |
| 3388 | iv = ver_info (data, ver: i)->iv; |
| 3389 | if (iv && iv->biv_p && !integer_zerop (iv->step)) |
| 3390 | add_iv_candidate_for_biv (data, iv); |
| 3391 | } |
| 3392 | } |
| 3393 | |
| 3394 | /* Record common candidate {BASE, STEP} derived from USE in hashtable. */ |
| 3395 | |
| 3396 | static void |
| 3397 | record_common_cand (struct ivopts_data *data, tree base, |
| 3398 | tree step, struct iv_use *use) |
| 3399 | { |
| 3400 | class iv_common_cand ent; |
| 3401 | class iv_common_cand **slot; |
| 3402 | |
| 3403 | ent.base = base; |
| 3404 | ent.step = step; |
| 3405 | ent.hash = iterative_hash_expr (tree: base, seed: 0); |
| 3406 | ent.hash = iterative_hash_expr (tree: step, seed: ent.hash); |
| 3407 | |
| 3408 | slot = data->iv_common_cand_tab->find_slot (value: &ent, insert: INSERT); |
| 3409 | if (*slot == NULL) |
| 3410 | { |
| 3411 | *slot = new iv_common_cand (); |
| 3412 | (*slot)->base = base; |
| 3413 | (*slot)->step = step; |
| 3414 | (*slot)->uses.create (nelems: 8); |
| 3415 | (*slot)->hash = ent.hash; |
| 3416 | data->iv_common_cands.safe_push (obj: (*slot)); |
| 3417 | } |
| 3418 | |
| 3419 | gcc_assert (use != NULL); |
| 3420 | (*slot)->uses.safe_push (obj: use); |
| 3421 | return; |
| 3422 | } |
| 3423 | |
| 3424 | /* Comparison function used to sort common candidates. */ |
| 3425 | |
| 3426 | static int |
| 3427 | common_cand_cmp (const void *p1, const void *p2) |
| 3428 | { |
| 3429 | unsigned n1, n2; |
| 3430 | const class iv_common_cand *const *const ccand1 |
| 3431 | = (const class iv_common_cand *const *)p1; |
| 3432 | const class iv_common_cand *const *const ccand2 |
| 3433 | = (const class iv_common_cand *const *)p2; |
| 3434 | |
| 3435 | n1 = (*ccand1)->uses.length (); |
| 3436 | n2 = (*ccand2)->uses.length (); |
| 3437 | return n2 - n1; |
| 3438 | } |
| 3439 | |
| 3440 | /* Adds IV candidates based on common candidated recorded. */ |
| 3441 | |
| 3442 | static void |
| 3443 | add_iv_candidate_derived_from_uses (struct ivopts_data *data) |
| 3444 | { |
| 3445 | unsigned i, j; |
| 3446 | struct iv_cand *cand_1, *cand_2; |
| 3447 | |
| 3448 | data->iv_common_cands.qsort (common_cand_cmp); |
| 3449 | for (i = 0; i < data->iv_common_cands.length (); i++) |
| 3450 | { |
| 3451 | class iv_common_cand *ptr = data->iv_common_cands[i]; |
| 3452 | |
| 3453 | /* Only add IV candidate if it's derived from multiple uses. */ |
| 3454 | if (ptr->uses.length () <= 1) |
| 3455 | break; |
| 3456 | |
| 3457 | cand_1 = NULL; |
| 3458 | cand_2 = NULL; |
| 3459 | if (ip_normal_pos (data->current_loop)) |
| 3460 | cand_1 = add_candidate_1 (data, base: ptr->base, step: ptr->step, |
| 3461 | important: false, pos: IP_NORMAL, NULL, NULL); |
| 3462 | |
| 3463 | if (ip_end_pos (data->current_loop) |
| 3464 | && allow_ip_end_pos_p (loop: data->current_loop)) |
| 3465 | cand_2 = add_candidate_1 (data, base: ptr->base, step: ptr->step, |
| 3466 | important: false, pos: IP_END, NULL, NULL); |
| 3467 | |
| 3468 | /* Bind deriving uses and the new candidates. */ |
| 3469 | for (j = 0; j < ptr->uses.length (); j++) |
| 3470 | { |
| 3471 | struct iv_group *group = data->vgroups[ptr->uses[j]->group_id]; |
| 3472 | if (cand_1) |
| 3473 | bitmap_set_bit (group->related_cands, cand_1->id); |
| 3474 | if (cand_2) |
| 3475 | bitmap_set_bit (group->related_cands, cand_2->id); |
| 3476 | } |
| 3477 | } |
| 3478 | |
| 3479 | /* Release data since it is useless from this point. */ |
| 3480 | data->iv_common_cand_tab->empty (); |
| 3481 | data->iv_common_cands.truncate (size: 0); |
| 3482 | } |
| 3483 | |
| 3484 | /* Adds candidates based on the value of USE's iv. */ |
| 3485 | |
| 3486 | static void |
| 3487 | add_iv_candidate_for_use (struct ivopts_data *data, struct iv_use *use) |
| 3488 | { |
| 3489 | poly_uint64 offset; |
| 3490 | tree base; |
| 3491 | struct iv *iv = use->iv; |
| 3492 | tree basetype = TREE_TYPE (iv->base); |
| 3493 | |
| 3494 | /* Don't add candidate for iv_use with non integer, pointer or non-mode |
| 3495 | precision types, instead, add candidate for the corresponding scev in |
| 3496 | unsigned type with the same precision. See PR93674 for more info. */ |
| 3497 | if ((TREE_CODE (basetype) != INTEGER_TYPE && !POINTER_TYPE_P (basetype)) |
| 3498 | || !type_has_mode_precision_p (t: basetype)) |
| 3499 | { |
| 3500 | basetype = lang_hooks.types.type_for_mode (TYPE_MODE (basetype), |
| 3501 | TYPE_UNSIGNED (basetype)); |
| 3502 | add_candidate (data, fold_convert (basetype, iv->base), |
| 3503 | fold_convert (basetype, iv->step), important: false, NULL); |
| 3504 | return; |
| 3505 | } |
| 3506 | |
| 3507 | add_candidate (data, base: iv->base, step: iv->step, important: false, use); |
| 3508 | |
| 3509 | /* Record common candidate for use in case it can be shared by others. */ |
| 3510 | record_common_cand (data, base: iv->base, step: iv->step, use); |
| 3511 | |
| 3512 | /* Record common candidate with initial value zero. */ |
| 3513 | basetype = TREE_TYPE (iv->base); |
| 3514 | if (POINTER_TYPE_P (basetype)) |
| 3515 | basetype = sizetype; |
| 3516 | record_common_cand (data, base: build_int_cst (basetype, 0), step: iv->step, use); |
| 3517 | |
| 3518 | /* Compare the cost of an address with an unscaled index with the cost of |
| 3519 | an address with a scaled index and add candidate if useful. */ |
| 3520 | poly_int64 step; |
| 3521 | if (use != NULL |
| 3522 | && poly_int_tree_p (t: iv->step, value: &step) |
| 3523 | && address_p (type: use->type)) |
| 3524 | { |
| 3525 | poly_int64 new_step; |
| 3526 | unsigned int fact = preferred_mem_scale_factor |
| 3527 | (base: use->iv->base, |
| 3528 | TYPE_MODE (use->mem_type), |
| 3529 | speed: optimize_loop_for_speed_p (data->current_loop)); |
| 3530 | |
| 3531 | if (fact != 1 |
| 3532 | && multiple_p (a: step, b: fact, multiple: &new_step)) |
| 3533 | add_candidate (data, size_int (0), |
| 3534 | step: wide_int_to_tree (sizetype, cst: new_step), |
| 3535 | important: true, NULL); |
| 3536 | } |
| 3537 | |
| 3538 | /* Record common candidate with constant offset stripped in base. |
| 3539 | Like the use itself, we also add candidate directly for it. */ |
| 3540 | base = strip_offset (expr: iv->base, offset: &offset); |
| 3541 | if (maybe_ne (a: offset, b: 0U) || base != iv->base) |
| 3542 | { |
| 3543 | record_common_cand (data, base, step: iv->step, use); |
| 3544 | add_candidate (data, base, step: iv->step, important: false, use); |
| 3545 | } |
| 3546 | |
| 3547 | /* Record common candidate with base_object removed in base. */ |
| 3548 | base = iv->base; |
| 3549 | STRIP_NOPS (base); |
| 3550 | if (iv->base_object != NULL && TREE_CODE (base) == POINTER_PLUS_EXPR) |
| 3551 | { |
| 3552 | tree step = iv->step; |
| 3553 | |
| 3554 | STRIP_NOPS (step); |
| 3555 | base = TREE_OPERAND (base, 1); |
| 3556 | step = fold_convert (sizetype, step); |
| 3557 | record_common_cand (data, base, step, use); |
| 3558 | /* Also record common candidate with offset stripped. */ |
| 3559 | tree alt_base, alt_offset; |
| 3560 | split_constant_offset (base, &alt_base, &alt_offset); |
| 3561 | if (!integer_zerop (alt_offset)) |
| 3562 | record_common_cand (data, base: alt_base, step, use); |
| 3563 | } |
| 3564 | |
| 3565 | /* At last, add auto-incremental candidates. Make such variables |
| 3566 | important since other iv uses with same base object may be based |
| 3567 | on it. */ |
| 3568 | if (use != NULL && address_p (type: use->type)) |
| 3569 | add_autoinc_candidates (data, base: iv->base, step: iv->step, important: true, use); |
| 3570 | } |
| 3571 | |
| 3572 | /* Adds candidates based on the uses. */ |
| 3573 | |
| 3574 | static void |
| 3575 | add_iv_candidate_for_groups (struct ivopts_data *data) |
| 3576 | { |
| 3577 | unsigned i; |
| 3578 | |
| 3579 | /* Only add candidate for the first use in group. */ |
| 3580 | for (i = 0; i < data->vgroups.length (); i++) |
| 3581 | { |
| 3582 | struct iv_group *group = data->vgroups[i]; |
| 3583 | |
| 3584 | gcc_assert (group->vuses[0] != NULL); |
| 3585 | add_iv_candidate_for_use (data, use: group->vuses[0]); |
| 3586 | } |
| 3587 | add_iv_candidate_derived_from_uses (data); |
| 3588 | } |
| 3589 | |
| 3590 | /* Record important candidates and add them to related_cands bitmaps. */ |
| 3591 | |
| 3592 | static void |
| 3593 | record_important_candidates (struct ivopts_data *data) |
| 3594 | { |
| 3595 | unsigned i; |
| 3596 | struct iv_group *group; |
| 3597 | |
| 3598 | for (i = 0; i < data->vcands.length (); i++) |
| 3599 | { |
| 3600 | struct iv_cand *cand = data->vcands[i]; |
| 3601 | |
| 3602 | if (cand->important) |
| 3603 | bitmap_set_bit (data->important_candidates, i); |
| 3604 | } |
| 3605 | |
| 3606 | data->consider_all_candidates = (data->vcands.length () |
| 3607 | <= CONSIDER_ALL_CANDIDATES_BOUND); |
| 3608 | |
| 3609 | /* Add important candidates to groups' related_cands bitmaps. */ |
| 3610 | for (i = 0; i < data->vgroups.length (); i++) |
| 3611 | { |
| 3612 | group = data->vgroups[i]; |
| 3613 | bitmap_ior_into (group->related_cands, data->important_candidates); |
| 3614 | } |
| 3615 | } |
| 3616 | |
| 3617 | /* Allocates the data structure mapping the (use, candidate) pairs to costs. |
| 3618 | If consider_all_candidates is true, we use a two-dimensional array, otherwise |
| 3619 | we allocate a simple list to every use. */ |
| 3620 | |
| 3621 | static void |
| 3622 | alloc_use_cost_map (struct ivopts_data *data) |
| 3623 | { |
| 3624 | unsigned i, size, s; |
| 3625 | |
| 3626 | for (i = 0; i < data->vgroups.length (); i++) |
| 3627 | { |
| 3628 | struct iv_group *group = data->vgroups[i]; |
| 3629 | |
| 3630 | if (data->consider_all_candidates) |
| 3631 | size = data->vcands.length (); |
| 3632 | else |
| 3633 | { |
| 3634 | s = bitmap_count_bits (group->related_cands); |
| 3635 | |
| 3636 | /* Round up to the power of two, so that moduling by it is fast. */ |
| 3637 | size = s ? (1 << ceil_log2 (x: s)) : 1; |
| 3638 | } |
| 3639 | |
| 3640 | group->n_map_members = size; |
| 3641 | group->cost_map = XCNEWVEC (class cost_pair, size); |
| 3642 | } |
| 3643 | } |
| 3644 | |
| 3645 | /* Sets cost of (GROUP, CAND) pair to COST and record that it depends |
| 3646 | on invariants INV_VARS and that the value used in expressing it is |
| 3647 | VALUE, and in case of iv elimination the comparison operator is COMP. */ |
| 3648 | |
| 3649 | static void |
| 3650 | set_group_iv_cost (struct ivopts_data *data, |
| 3651 | struct iv_group *group, struct iv_cand *cand, |
| 3652 | comp_cost cost, bitmap inv_vars, tree value, |
| 3653 | enum tree_code comp, bitmap inv_exprs) |
| 3654 | { |
| 3655 | unsigned i, s; |
| 3656 | |
| 3657 | if (cost.infinite_cost_p ()) |
| 3658 | { |
| 3659 | BITMAP_FREE (inv_vars); |
| 3660 | BITMAP_FREE (inv_exprs); |
| 3661 | return; |
| 3662 | } |
| 3663 | |
| 3664 | if (data->consider_all_candidates) |
| 3665 | { |
| 3666 | group->cost_map[cand->id].cand = cand; |
| 3667 | group->cost_map[cand->id].cost = cost; |
| 3668 | group->cost_map[cand->id].inv_vars = inv_vars; |
| 3669 | group->cost_map[cand->id].inv_exprs = inv_exprs; |
| 3670 | group->cost_map[cand->id].value = value; |
| 3671 | group->cost_map[cand->id].comp = comp; |
| 3672 | return; |
| 3673 | } |
| 3674 | |
| 3675 | /* n_map_members is a power of two, so this computes modulo. */ |
| 3676 | s = cand->id & (group->n_map_members - 1); |
| 3677 | for (i = s; i < group->n_map_members; i++) |
| 3678 | if (!group->cost_map[i].cand) |
| 3679 | goto found; |
| 3680 | for (i = 0; i < s; i++) |
| 3681 | if (!group->cost_map[i].cand) |
| 3682 | goto found; |
| 3683 | |
| 3684 | gcc_unreachable (); |
| 3685 | |
| 3686 | found: |
| 3687 | group->cost_map[i].cand = cand; |
| 3688 | group->cost_map[i].cost = cost; |
| 3689 | group->cost_map[i].inv_vars = inv_vars; |
| 3690 | group->cost_map[i].inv_exprs = inv_exprs; |
| 3691 | group->cost_map[i].value = value; |
| 3692 | group->cost_map[i].comp = comp; |
| 3693 | } |
| 3694 | |
| 3695 | /* Gets cost of (GROUP, CAND) pair. */ |
| 3696 | |
| 3697 | static class cost_pair * |
| 3698 | get_group_iv_cost (struct ivopts_data *data, struct iv_group *group, |
| 3699 | struct iv_cand *cand) |
| 3700 | { |
| 3701 | unsigned i, s; |
| 3702 | class cost_pair *ret; |
| 3703 | |
| 3704 | if (!cand) |
| 3705 | return NULL; |
| 3706 | |
| 3707 | if (data->consider_all_candidates) |
| 3708 | { |
| 3709 | ret = group->cost_map + cand->id; |
| 3710 | if (!ret->cand) |
| 3711 | return NULL; |
| 3712 | |
| 3713 | return ret; |
| 3714 | } |
| 3715 | |
| 3716 | /* n_map_members is a power of two, so this computes modulo. */ |
| 3717 | s = cand->id & (group->n_map_members - 1); |
| 3718 | for (i = s; i < group->n_map_members; i++) |
| 3719 | if (group->cost_map[i].cand == cand) |
| 3720 | return group->cost_map + i; |
| 3721 | else if (group->cost_map[i].cand == NULL) |
| 3722 | return NULL; |
| 3723 | for (i = 0; i < s; i++) |
| 3724 | if (group->cost_map[i].cand == cand) |
| 3725 | return group->cost_map + i; |
| 3726 | else if (group->cost_map[i].cand == NULL) |
| 3727 | return NULL; |
| 3728 | |
| 3729 | return NULL; |
| 3730 | } |
| 3731 | |
| 3732 | /* Produce DECL_RTL for object obj so it looks like it is stored in memory. */ |
| 3733 | static rtx |
| 3734 | produce_memory_decl_rtl (tree obj, int *regno) |
| 3735 | { |
| 3736 | addr_space_t as = TYPE_ADDR_SPACE (TREE_TYPE (obj)); |
| 3737 | machine_mode address_mode = targetm.addr_space.address_mode (as); |
| 3738 | rtx x; |
| 3739 | |
| 3740 | gcc_assert (obj); |
| 3741 | if (TREE_STATIC (obj) || DECL_EXTERNAL (obj)) |
| 3742 | { |
| 3743 | const char *name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (obj)); |
| 3744 | x = gen_rtx_SYMBOL_REF (address_mode, name); |
| 3745 | SET_SYMBOL_REF_DECL (x, obj); |
| 3746 | x = gen_rtx_MEM (DECL_MODE (obj), x); |
| 3747 | set_mem_addr_space (x, as); |
| 3748 | targetm.encode_section_info (obj, x, true); |
| 3749 | } |
| 3750 | else |
| 3751 | { |
| 3752 | x = gen_raw_REG (address_mode, (*regno)++); |
| 3753 | x = gen_rtx_MEM (DECL_MODE (obj), x); |
| 3754 | set_mem_addr_space (x, as); |
| 3755 | } |
| 3756 | |
| 3757 | return x; |
| 3758 | } |
| 3759 | |
| 3760 | /* Prepares decl_rtl for variables referred in *EXPR_P. Callback for |
| 3761 | walk_tree. DATA contains the actual fake register number. */ |
| 3762 | |
| 3763 | static tree |
| 3764 | prepare_decl_rtl (tree *expr_p, int *ws, void *data) |
| 3765 | { |
| 3766 | tree obj = NULL_TREE; |
| 3767 | rtx x = NULL_RTX; |
| 3768 | int *regno = (int *) data; |
| 3769 | |
| 3770 | switch (TREE_CODE (*expr_p)) |
| 3771 | { |
| 3772 | case ADDR_EXPR: |
| 3773 | for (expr_p = &TREE_OPERAND (*expr_p, 0); |
| 3774 | handled_component_p (t: *expr_p); |
| 3775 | expr_p = &TREE_OPERAND (*expr_p, 0)) |
| 3776 | continue; |
| 3777 | obj = *expr_p; |
| 3778 | if (DECL_P (obj) && HAS_RTL_P (obj) && !DECL_RTL_SET_P (obj)) |
| 3779 | x = produce_memory_decl_rtl (obj, regno); |
| 3780 | break; |
| 3781 | |
| 3782 | case SSA_NAME: |
| 3783 | *ws = 0; |
| 3784 | obj = SSA_NAME_VAR (*expr_p); |
| 3785 | /* Defer handling of anonymous SSA_NAMEs to the expander. */ |
| 3786 | if (!obj) |
| 3787 | return NULL_TREE; |
| 3788 | if (!DECL_RTL_SET_P (obj)) |
| 3789 | x = gen_raw_REG (DECL_MODE (obj), (*regno)++); |
| 3790 | break; |
| 3791 | |
| 3792 | case VAR_DECL: |
| 3793 | case PARM_DECL: |
| 3794 | case RESULT_DECL: |
| 3795 | *ws = 0; |
| 3796 | obj = *expr_p; |
| 3797 | |
| 3798 | if (DECL_RTL_SET_P (obj)) |
| 3799 | break; |
| 3800 | |
| 3801 | if (DECL_MODE (obj) == BLKmode) |
| 3802 | x = produce_memory_decl_rtl (obj, regno); |
| 3803 | else |
| 3804 | x = gen_raw_REG (DECL_MODE (obj), (*regno)++); |
| 3805 | |
| 3806 | break; |
| 3807 | |
| 3808 | default: |
| 3809 | break; |
| 3810 | } |
| 3811 | |
| 3812 | if (x) |
| 3813 | { |
| 3814 | decl_rtl_to_reset.safe_push (obj); |
| 3815 | SET_DECL_RTL (obj, x); |
| 3816 | } |
| 3817 | |
| 3818 | return NULL_TREE; |
| 3819 | } |
| 3820 | |
| 3821 | /* Predict whether the given loop will be transformed in the RTL |
| 3822 | doloop_optimize pass. Attempt to duplicate some doloop_optimize checks. |
| 3823 | This is only for target independent checks, see targetm.predict_doloop_p |
| 3824 | for the target dependent ones. |
| 3825 | |
| 3826 | Note that according to some initial investigation, some checks like costly |
| 3827 | niter check and invalid stmt scanning don't have much gains among general |
| 3828 | cases, so keep this as simple as possible first. |
| 3829 | |
| 3830 | Some RTL specific checks seems unable to be checked in gimple, if any new |
| 3831 | checks or easy checks _are_ missing here, please add them. */ |
| 3832 | |
| 3833 | static bool |
| 3834 | generic_predict_doloop_p (struct ivopts_data *data) |
| 3835 | { |
| 3836 | class loop *loop = data->current_loop; |
| 3837 | |
| 3838 | /* Call target hook for target dependent checks. */ |
| 3839 | if (!targetm.predict_doloop_p (loop)) |
| 3840 | { |
| 3841 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 3842 | fprintf (stream: dump_file, format: "Predict doloop failure due to" |
| 3843 | " target specific checks.\n" ); |
| 3844 | return false; |
| 3845 | } |
| 3846 | |
| 3847 | /* Similar to doloop_optimize, check iteration description to know it's |
| 3848 | suitable or not. Keep it as simple as possible, feel free to extend it |
| 3849 | if you find any multiple exits cases matter. */ |
| 3850 | edge exit = single_dom_exit (loop); |
| 3851 | class tree_niter_desc *niter_desc; |
| 3852 | if (!exit || !(niter_desc = niter_for_exit (data, exit))) |
| 3853 | { |
| 3854 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 3855 | fprintf (stream: dump_file, format: "Predict doloop failure due to" |
| 3856 | " unexpected niters.\n" ); |
| 3857 | return false; |
| 3858 | } |
| 3859 | |
| 3860 | /* Similar to doloop_optimize, check whether iteration count too small |
| 3861 | and not profitable. */ |
| 3862 | HOST_WIDE_INT est_niter = get_estimated_loop_iterations_int (loop); |
| 3863 | if (est_niter == -1) |
| 3864 | est_niter = get_likely_max_loop_iterations_int (loop); |
| 3865 | if (est_niter >= 0 && est_niter < 3) |
| 3866 | { |
| 3867 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 3868 | fprintf (stream: dump_file, |
| 3869 | format: "Predict doloop failure due to" |
| 3870 | " too few iterations (%u).\n" , |
| 3871 | (unsigned int) est_niter); |
| 3872 | return false; |
| 3873 | } |
| 3874 | |
| 3875 | return true; |
| 3876 | } |
| 3877 | |
| 3878 | /* Determines cost of the computation of EXPR. */ |
| 3879 | |
| 3880 | static unsigned |
| 3881 | computation_cost (tree expr, bool speed) |
| 3882 | { |
| 3883 | rtx_insn *seq; |
| 3884 | rtx rslt; |
| 3885 | tree type = TREE_TYPE (expr); |
| 3886 | unsigned cost; |
| 3887 | /* Avoid using hard regs in ways which may be unsupported. */ |
| 3888 | int regno = LAST_VIRTUAL_REGISTER + 1; |
| 3889 | struct cgraph_node *node = cgraph_node::get (decl: current_function_decl); |
| 3890 | enum node_frequency real_frequency = node->frequency; |
| 3891 | |
| 3892 | node->frequency = NODE_FREQUENCY_NORMAL; |
| 3893 | crtl->maybe_hot_insn_p = speed; |
| 3894 | walk_tree (&expr, prepare_decl_rtl, ®no, NULL); |
| 3895 | start_sequence (); |
| 3896 | rslt = expand_expr (exp: expr, NULL_RTX, TYPE_MODE (type), modifier: EXPAND_NORMAL); |
| 3897 | seq = end_sequence (); |
| 3898 | default_rtl_profile (); |
| 3899 | node->frequency = real_frequency; |
| 3900 | |
| 3901 | cost = seq_cost (seq, speed); |
| 3902 | if (MEM_P (rslt)) |
| 3903 | cost += address_cost (XEXP (rslt, 0), TYPE_MODE (type), |
| 3904 | TYPE_ADDR_SPACE (type), speed); |
| 3905 | else if (!REG_P (rslt)) |
| 3906 | cost += set_src_cost (x: rslt, TYPE_MODE (type), speed_p: speed); |
| 3907 | |
| 3908 | return cost; |
| 3909 | } |
| 3910 | |
| 3911 | /* Returns variable containing the value of candidate CAND at statement AT. */ |
| 3912 | |
| 3913 | static tree |
| 3914 | var_at_stmt (class loop *loop, struct iv_cand *cand, gimple *stmt) |
| 3915 | { |
| 3916 | if (stmt_after_increment (loop, cand, stmt)) |
| 3917 | return cand->var_after; |
| 3918 | else |
| 3919 | return cand->var_before; |
| 3920 | } |
| 3921 | |
| 3922 | /* If A is (TYPE) BA and B is (TYPE) BB, and the types of BA and BB have the |
| 3923 | same precision that is at least as wide as the precision of TYPE, stores |
| 3924 | BA to A and BB to B, and returns the type of BA. Otherwise, returns the |
| 3925 | type of A and B. */ |
| 3926 | |
| 3927 | static tree |
| 3928 | determine_common_wider_type (tree *a, tree *b) |
| 3929 | { |
| 3930 | tree wider_type = NULL; |
| 3931 | tree suba, subb; |
| 3932 | tree atype = TREE_TYPE (*a); |
| 3933 | |
| 3934 | if (CONVERT_EXPR_P (*a)) |
| 3935 | { |
| 3936 | suba = TREE_OPERAND (*a, 0); |
| 3937 | wider_type = TREE_TYPE (suba); |
| 3938 | if (TYPE_PRECISION (wider_type) < TYPE_PRECISION (atype)) |
| 3939 | return atype; |
| 3940 | } |
| 3941 | else |
| 3942 | return atype; |
| 3943 | |
| 3944 | if (CONVERT_EXPR_P (*b)) |
| 3945 | { |
| 3946 | subb = TREE_OPERAND (*b, 0); |
| 3947 | if (TYPE_PRECISION (wider_type) != TYPE_PRECISION (TREE_TYPE (subb))) |
| 3948 | return atype; |
| 3949 | } |
| 3950 | else |
| 3951 | return atype; |
| 3952 | |
| 3953 | *a = suba; |
| 3954 | *b = subb; |
| 3955 | return wider_type; |
| 3956 | } |
| 3957 | |
| 3958 | /* Determines the expression by that USE is expressed from induction variable |
| 3959 | CAND at statement AT in DATA's current loop. The expression is stored in |
| 3960 | two parts in a decomposed form. The invariant part is stored in AFF_INV; |
| 3961 | while variant part in AFF_VAR. Store ratio of CAND.step over USE.step in |
| 3962 | PRAT if it's non-null. Returns false if USE cannot be expressed using |
| 3963 | CAND. */ |
| 3964 | |
| 3965 | static bool |
| 3966 | get_computation_aff_1 (struct ivopts_data *data, gimple *at, struct iv_use *use, |
| 3967 | struct iv_cand *cand, class aff_tree *aff_inv, |
| 3968 | class aff_tree *aff_var, widest_int *prat = NULL) |
| 3969 | { |
| 3970 | tree ubase = use->iv->base, ustep = use->iv->step; |
| 3971 | tree cbase = cand->iv->base, cstep = cand->iv->step; |
| 3972 | tree common_type, uutype, var, cstep_common; |
| 3973 | tree utype = TREE_TYPE (ubase), ctype = TREE_TYPE (cbase); |
| 3974 | aff_tree aff_cbase; |
| 3975 | widest_int rat; |
| 3976 | |
| 3977 | /* We must have a precision to express the values of use. */ |
| 3978 | if (TYPE_PRECISION (utype) > TYPE_PRECISION (ctype)) |
| 3979 | return false; |
| 3980 | |
| 3981 | var = var_at_stmt (loop: data->current_loop, cand, stmt: at); |
| 3982 | uutype = unsigned_type_for (utype); |
| 3983 | |
| 3984 | /* If the conversion is not noop, perform it. */ |
| 3985 | if (TYPE_PRECISION (utype) < TYPE_PRECISION (ctype)) |
| 3986 | { |
| 3987 | if (cand->orig_iv != NULL && CONVERT_EXPR_P (cbase) |
| 3988 | && (CONVERT_EXPR_P (cstep) || poly_int_tree_p (t: cstep))) |
| 3989 | { |
| 3990 | tree inner_base, inner_step, inner_type; |
| 3991 | inner_base = TREE_OPERAND (cbase, 0); |
| 3992 | if (CONVERT_EXPR_P (cstep)) |
| 3993 | inner_step = TREE_OPERAND (cstep, 0); |
| 3994 | else |
| 3995 | inner_step = cstep; |
| 3996 | |
| 3997 | inner_type = TREE_TYPE (inner_base); |
| 3998 | /* If candidate is added from a biv whose type is smaller than |
| 3999 | ctype, we know both candidate and the biv won't overflow. |
| 4000 | In this case, it's safe to skip the convertion in candidate. |
| 4001 | As an example, (unsigned short)((unsigned long)A) equals to |
| 4002 | (unsigned short)A, if A has a type no larger than short. */ |
| 4003 | if (TYPE_PRECISION (inner_type) <= TYPE_PRECISION (uutype)) |
| 4004 | { |
| 4005 | cbase = inner_base; |
| 4006 | cstep = inner_step; |
| 4007 | } |
| 4008 | } |
| 4009 | cbase = fold_convert (uutype, cbase); |
| 4010 | cstep = fold_convert (uutype, cstep); |
| 4011 | var = fold_convert (uutype, var); |
| 4012 | } |
| 4013 | |
| 4014 | /* Ratio is 1 when computing the value of biv cand by itself. |
| 4015 | We can't rely on constant_multiple_of in this case because the |
| 4016 | use is created after the original biv is selected. The call |
| 4017 | could fail because of inconsistent fold behavior. See PR68021 |
| 4018 | for more information. */ |
| 4019 | if (cand->pos == IP_ORIGINAL && cand->incremented_at == use->stmt) |
| 4020 | { |
| 4021 | gcc_assert (is_gimple_assign (use->stmt)); |
| 4022 | gcc_assert (use->iv->ssa_name == cand->var_after); |
| 4023 | gcc_assert (gimple_assign_lhs (use->stmt) == cand->var_after); |
| 4024 | rat = 1; |
| 4025 | } |
| 4026 | else if (!constant_multiple_of (top: ustep, bot: cstep, mul: &rat, data)) |
| 4027 | return false; |
| 4028 | |
| 4029 | if (prat) |
| 4030 | *prat = rat; |
| 4031 | |
| 4032 | /* In case both UBASE and CBASE are shortened to UUTYPE from some common |
| 4033 | type, we achieve better folding by computing their difference in this |
| 4034 | wider type, and cast the result to UUTYPE. We do not need to worry about |
| 4035 | overflows, as all the arithmetics will in the end be performed in UUTYPE |
| 4036 | anyway. */ |
| 4037 | common_type = determine_common_wider_type (a: &ubase, b: &cbase); |
| 4038 | |
| 4039 | /* use = ubase - ratio * cbase + ratio * var. */ |
| 4040 | tree_to_aff_combination (ubase, common_type, aff_inv); |
| 4041 | tree_to_aff_combination (cbase, common_type, &aff_cbase); |
| 4042 | tree_to_aff_combination (var, uutype, aff_var); |
| 4043 | |
| 4044 | /* We need to shift the value if we are after the increment. */ |
| 4045 | if (stmt_after_increment (loop: data->current_loop, cand, stmt: at)) |
| 4046 | { |
| 4047 | aff_tree cstep_aff; |
| 4048 | |
| 4049 | if (common_type != uutype) |
| 4050 | cstep_common = fold_convert (common_type, cstep); |
| 4051 | else |
| 4052 | cstep_common = cstep; |
| 4053 | |
| 4054 | tree_to_aff_combination (cstep_common, common_type, &cstep_aff); |
| 4055 | aff_combination_add (&aff_cbase, &cstep_aff); |
| 4056 | } |
| 4057 | |
| 4058 | aff_combination_scale (&aff_cbase, -rat); |
| 4059 | aff_combination_add (aff_inv, &aff_cbase); |
| 4060 | if (common_type != uutype) |
| 4061 | aff_combination_convert (aff_inv, uutype); |
| 4062 | |
| 4063 | aff_combination_scale (aff_var, rat); |
| 4064 | return true; |
| 4065 | } |
| 4066 | |
| 4067 | /* Determines the expression by that USE is expressed from induction variable |
| 4068 | CAND at statement AT in DATA's current loop. The expression is stored in a |
| 4069 | decomposed form into AFF. Returns false if USE cannot be expressed using |
| 4070 | CAND. */ |
| 4071 | |
| 4072 | static bool |
| 4073 | get_computation_aff (struct ivopts_data *data, gimple *at, struct iv_use *use, |
| 4074 | struct iv_cand *cand, class aff_tree *aff) |
| 4075 | { |
| 4076 | aff_tree aff_var; |
| 4077 | |
| 4078 | if (!get_computation_aff_1 (data, at, use, cand, aff_inv: aff, aff_var: &aff_var)) |
| 4079 | return false; |
| 4080 | |
| 4081 | aff_combination_add (aff, &aff_var); |
| 4082 | return true; |
| 4083 | } |
| 4084 | |
| 4085 | /* Return the type of USE. */ |
| 4086 | |
| 4087 | static tree |
| 4088 | get_use_type (struct iv_use *use) |
| 4089 | { |
| 4090 | tree base_type = TREE_TYPE (use->iv->base); |
| 4091 | tree type; |
| 4092 | |
| 4093 | if (use->type == USE_REF_ADDRESS) |
| 4094 | { |
| 4095 | /* The base_type may be a void pointer. Create a pointer type based on |
| 4096 | the mem_ref instead. */ |
| 4097 | type = build_pointer_type (TREE_TYPE (*use->op_p)); |
| 4098 | gcc_assert (TYPE_ADDR_SPACE (TREE_TYPE (type)) |
| 4099 | == TYPE_ADDR_SPACE (TREE_TYPE (base_type))); |
| 4100 | } |
| 4101 | else |
| 4102 | type = base_type; |
| 4103 | |
| 4104 | return type; |
| 4105 | } |
| 4106 | |
| 4107 | /* Determines the expression by that USE is expressed from induction variable |
| 4108 | CAND at statement AT in DATA's current loop. The computation is |
| 4109 | unshared. */ |
| 4110 | |
| 4111 | static tree |
| 4112 | get_computation_at (struct ivopts_data *data, gimple *at, |
| 4113 | struct iv_use *use, struct iv_cand *cand) |
| 4114 | { |
| 4115 | aff_tree aff; |
| 4116 | tree type = get_use_type (use); |
| 4117 | |
| 4118 | if (!get_computation_aff (data, at, use, cand, aff: &aff)) |
| 4119 | return NULL_TREE; |
| 4120 | unshare_aff_combination (&aff); |
| 4121 | return fold_convert (type, aff_combination_to_tree (&aff)); |
| 4122 | } |
| 4123 | |
| 4124 | /* Like get_computation_at, but try harder, even if the computation |
| 4125 | is more expensive. Intended for debug stmts. */ |
| 4126 | |
| 4127 | static tree |
| 4128 | get_debug_computation_at (struct ivopts_data *data, gimple *at, |
| 4129 | struct iv_use *use, struct iv_cand *cand) |
| 4130 | { |
| 4131 | if (tree ret = get_computation_at (data, at, use, cand)) |
| 4132 | return ret; |
| 4133 | |
| 4134 | tree ubase = use->iv->base, ustep = use->iv->step; |
| 4135 | tree cbase = cand->iv->base, cstep = cand->iv->step; |
| 4136 | tree var; |
| 4137 | tree utype = TREE_TYPE (ubase), ctype = TREE_TYPE (cbase); |
| 4138 | widest_int rat; |
| 4139 | |
| 4140 | /* We must have a precision to express the values of use. */ |
| 4141 | if (TYPE_PRECISION (utype) >= TYPE_PRECISION (ctype)) |
| 4142 | return NULL_TREE; |
| 4143 | |
| 4144 | /* Try to handle the case that get_computation_at doesn't, |
| 4145 | try to express |
| 4146 | use = ubase + (var - cbase) / ratio. */ |
| 4147 | if (!constant_multiple_of (top: cstep, fold_convert (TREE_TYPE (cstep), ustep), |
| 4148 | mul: &rat, data)) |
| 4149 | return NULL_TREE; |
| 4150 | |
| 4151 | bool neg_p = false; |
| 4152 | if (wi::neg_p (x: rat)) |
| 4153 | { |
| 4154 | if (TYPE_UNSIGNED (ctype)) |
| 4155 | return NULL_TREE; |
| 4156 | neg_p = true; |
| 4157 | rat = wi::neg (x: rat); |
| 4158 | } |
| 4159 | |
| 4160 | /* If both IVs can wrap around and CAND doesn't have a power of two step, |
| 4161 | it is unsafe. Consider uint16_t CAND with step 9, when wrapping around, |
| 4162 | the values will be ... 0xfff0, 0xfff9, 2, 11 ... and when use is say |
| 4163 | uint8_t with step 3, those values divided by 3 cast to uint8_t will be |
| 4164 | ... 0x50, 0x53, 0, 3 ... rather than expected 0x50, 0x53, 0x56, 0x59. */ |
| 4165 | if (!use->iv->no_overflow |
| 4166 | && !cand->iv->no_overflow |
| 4167 | && !integer_pow2p (cstep)) |
| 4168 | return NULL_TREE; |
| 4169 | |
| 4170 | int bits = wi::exact_log2 (rat); |
| 4171 | if (bits == -1) |
| 4172 | bits = wi::floor_log2 (rat) + 1; |
| 4173 | if (!cand->iv->no_overflow |
| 4174 | && TYPE_PRECISION (utype) + bits > TYPE_PRECISION (ctype)) |
| 4175 | return NULL_TREE; |
| 4176 | |
| 4177 | var = var_at_stmt (loop: data->current_loop, cand, stmt: at); |
| 4178 | |
| 4179 | if (POINTER_TYPE_P (ctype)) |
| 4180 | { |
| 4181 | ctype = unsigned_type_for (ctype); |
| 4182 | cbase = fold_convert (ctype, cbase); |
| 4183 | cstep = fold_convert (ctype, cstep); |
| 4184 | var = fold_convert (ctype, var); |
| 4185 | } |
| 4186 | |
| 4187 | if (stmt_after_increment (loop: data->current_loop, cand, stmt: at)) |
| 4188 | var = fold_build2 (MINUS_EXPR, TREE_TYPE (var), var, |
| 4189 | unshare_expr (cstep)); |
| 4190 | |
| 4191 | var = fold_build2 (MINUS_EXPR, TREE_TYPE (var), var, cbase); |
| 4192 | var = fold_build2 (EXACT_DIV_EXPR, TREE_TYPE (var), var, |
| 4193 | wide_int_to_tree (TREE_TYPE (var), rat)); |
| 4194 | if (POINTER_TYPE_P (utype)) |
| 4195 | { |
| 4196 | var = fold_convert (sizetype, var); |
| 4197 | if (neg_p) |
| 4198 | var = fold_build1 (NEGATE_EXPR, sizetype, var); |
| 4199 | var = fold_build2 (POINTER_PLUS_EXPR, utype, ubase, var); |
| 4200 | } |
| 4201 | else |
| 4202 | { |
| 4203 | var = fold_convert (utype, var); |
| 4204 | var = fold_build2 (neg_p ? MINUS_EXPR : PLUS_EXPR, utype, |
| 4205 | ubase, var); |
| 4206 | } |
| 4207 | return var; |
| 4208 | } |
| 4209 | |
| 4210 | /* Adjust the cost COST for being in loop setup rather than loop body. |
| 4211 | If we're optimizing for space, the loop setup overhead is constant; |
| 4212 | if we're optimizing for speed, amortize it over the per-iteration cost. |
| 4213 | If ROUND_UP_P is true, the result is round up rather than to zero when |
| 4214 | optimizing for speed. */ |
| 4215 | static int64_t |
| 4216 | adjust_setup_cost (struct ivopts_data *data, int64_t cost, |
| 4217 | bool round_up_p = false) |
| 4218 | { |
| 4219 | if (cost == INFTY) |
| 4220 | return cost; |
| 4221 | else if (optimize_loop_for_speed_p (data->current_loop)) |
| 4222 | { |
| 4223 | uint64_t niters = avg_loop_niter (loop: data->current_loop); |
| 4224 | if (niters > (uint64_t) cost) |
| 4225 | return (round_up_p && cost != 0) ? 1 : 0; |
| 4226 | return (cost + (round_up_p ? niters - 1 : 0)) / niters; |
| 4227 | } |
| 4228 | else |
| 4229 | return cost; |
| 4230 | } |
| 4231 | |
| 4232 | /* Calculate the SPEED or size cost of shiftadd EXPR in MODE. MULT is the |
| 4233 | EXPR operand holding the shift. COST0 and COST1 are the costs for |
| 4234 | calculating the operands of EXPR. Returns true if successful, and returns |
| 4235 | the cost in COST. */ |
| 4236 | |
| 4237 | static bool |
| 4238 | get_shiftadd_cost (tree expr, scalar_int_mode mode, comp_cost cost0, |
| 4239 | comp_cost cost1, tree mult, bool speed, comp_cost *cost) |
| 4240 | { |
| 4241 | comp_cost res; |
| 4242 | tree op1 = TREE_OPERAND (expr, 1); |
| 4243 | tree cst = TREE_OPERAND (mult, 1); |
| 4244 | tree multop = TREE_OPERAND (mult, 0); |
| 4245 | int m = exact_log2 (x: int_cst_value (cst)); |
| 4246 | int maxm = MIN (BITS_PER_WORD, GET_MODE_BITSIZE (mode)); |
| 4247 | int as_cost, sa_cost; |
| 4248 | bool mult_in_op1; |
| 4249 | |
| 4250 | if (!(m >= 0 && m < maxm)) |
| 4251 | return false; |
| 4252 | |
| 4253 | STRIP_NOPS (op1); |
| 4254 | mult_in_op1 = operand_equal_p (op1, mult, flags: 0); |
| 4255 | |
| 4256 | as_cost = add_cost (speed, mode) + shift_cost (speed, mode, bits: m); |
| 4257 | |
| 4258 | /* If the target has a cheap shift-and-add or shift-and-sub instruction, |
| 4259 | use that in preference to a shift insn followed by an add insn. */ |
| 4260 | sa_cost = (TREE_CODE (expr) != MINUS_EXPR |
| 4261 | ? shiftadd_cost (speed, mode, bits: m) |
| 4262 | : (mult_in_op1 |
| 4263 | ? shiftsub1_cost (speed, mode, bits: m) |
| 4264 | : shiftsub0_cost (speed, mode, bits: m))); |
| 4265 | |
| 4266 | res = comp_cost (MIN (as_cost, sa_cost), 0); |
| 4267 | res += (mult_in_op1 ? cost0 : cost1); |
| 4268 | |
| 4269 | STRIP_NOPS (multop); |
| 4270 | if (!is_gimple_val (multop)) |
| 4271 | res += force_expr_to_var_cost (multop, speed); |
| 4272 | |
| 4273 | *cost = res; |
| 4274 | return true; |
| 4275 | } |
| 4276 | |
| 4277 | /* Estimates cost of forcing expression EXPR into a variable. */ |
| 4278 | |
| 4279 | static comp_cost |
| 4280 | force_expr_to_var_cost (tree expr, bool speed) |
| 4281 | { |
| 4282 | static bool costs_initialized = false; |
| 4283 | static unsigned integer_cost [2]; |
| 4284 | static unsigned symbol_cost [2]; |
| 4285 | static unsigned address_cost [2]; |
| 4286 | tree op0, op1; |
| 4287 | comp_cost cost0, cost1, cost; |
| 4288 | machine_mode mode; |
| 4289 | scalar_int_mode int_mode; |
| 4290 | |
| 4291 | if (!costs_initialized) |
| 4292 | { |
| 4293 | tree type = build_pointer_type (integer_type_node); |
| 4294 | tree var, addr; |
| 4295 | rtx x; |
| 4296 | int i; |
| 4297 | |
| 4298 | var = create_tmp_var_raw (integer_type_node, "test_var" ); |
| 4299 | TREE_STATIC (var) = 1; |
| 4300 | x = produce_memory_decl_rtl (obj: var, NULL); |
| 4301 | SET_DECL_RTL (var, x); |
| 4302 | |
| 4303 | addr = build1 (ADDR_EXPR, type, var); |
| 4304 | |
| 4305 | |
| 4306 | for (i = 0; i < 2; i++) |
| 4307 | { |
| 4308 | integer_cost[i] = computation_cost (expr: build_int_cst (integer_type_node, |
| 4309 | 2000), speed: i); |
| 4310 | |
| 4311 | symbol_cost[i] = computation_cost (expr: addr, speed: i) + 1; |
| 4312 | |
| 4313 | address_cost[i] |
| 4314 | = computation_cost (fold_build_pointer_plus_hwi (addr, 2000), speed: i) + 1; |
| 4315 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 4316 | { |
| 4317 | fprintf (stream: dump_file, format: "force_expr_to_var_cost %s costs:\n" , i ? "speed" : "size" ); |
| 4318 | fprintf (stream: dump_file, format: " integer %d\n" , (int) integer_cost[i]); |
| 4319 | fprintf (stream: dump_file, format: " symbol %d\n" , (int) symbol_cost[i]); |
| 4320 | fprintf (stream: dump_file, format: " address %d\n" , (int) address_cost[i]); |
| 4321 | fprintf (stream: dump_file, format: " other %d\n" , (int) target_spill_cost[i]); |
| 4322 | fprintf (stream: dump_file, format: "\n" ); |
| 4323 | } |
| 4324 | } |
| 4325 | |
| 4326 | costs_initialized = true; |
| 4327 | } |
| 4328 | |
| 4329 | STRIP_NOPS (expr); |
| 4330 | |
| 4331 | if (SSA_VAR_P (expr)) |
| 4332 | return no_cost; |
| 4333 | |
| 4334 | if (is_gimple_min_invariant (expr)) |
| 4335 | { |
| 4336 | if (poly_int_tree_p (t: expr)) |
| 4337 | return comp_cost (integer_cost [speed], 0); |
| 4338 | |
| 4339 | if (TREE_CODE (expr) == ADDR_EXPR) |
| 4340 | { |
| 4341 | tree obj = TREE_OPERAND (expr, 0); |
| 4342 | |
| 4343 | if (VAR_P (obj) |
| 4344 | || TREE_CODE (obj) == PARM_DECL |
| 4345 | || TREE_CODE (obj) == RESULT_DECL) |
| 4346 | return comp_cost (symbol_cost [speed], 0); |
| 4347 | } |
| 4348 | |
| 4349 | return comp_cost (address_cost [speed], 0); |
| 4350 | } |
| 4351 | |
| 4352 | switch (TREE_CODE (expr)) |
| 4353 | { |
| 4354 | case POINTER_PLUS_EXPR: |
| 4355 | case PLUS_EXPR: |
| 4356 | case MINUS_EXPR: |
| 4357 | case MULT_EXPR: |
| 4358 | case EXACT_DIV_EXPR: |
| 4359 | case TRUNC_DIV_EXPR: |
| 4360 | case BIT_AND_EXPR: |
| 4361 | case BIT_IOR_EXPR: |
| 4362 | case LSHIFT_EXPR: |
| 4363 | case RSHIFT_EXPR: |
| 4364 | op0 = TREE_OPERAND (expr, 0); |
| 4365 | op1 = TREE_OPERAND (expr, 1); |
| 4366 | STRIP_NOPS (op0); |
| 4367 | STRIP_NOPS (op1); |
| 4368 | break; |
| 4369 | |
| 4370 | CASE_CONVERT: |
| 4371 | case NEGATE_EXPR: |
| 4372 | case BIT_NOT_EXPR: |
| 4373 | op0 = TREE_OPERAND (expr, 0); |
| 4374 | STRIP_NOPS (op0); |
| 4375 | op1 = NULL_TREE; |
| 4376 | break; |
| 4377 | /* See add_iv_candidate_for_doloop, for doloop may_be_zero case, we |
| 4378 | introduce COND_EXPR for IV base, need to support better cost estimation |
| 4379 | for this COND_EXPR and tcc_comparison. */ |
| 4380 | case COND_EXPR: |
| 4381 | op0 = TREE_OPERAND (expr, 1); |
| 4382 | STRIP_NOPS (op0); |
| 4383 | op1 = TREE_OPERAND (expr, 2); |
| 4384 | STRIP_NOPS (op1); |
| 4385 | break; |
| 4386 | case LT_EXPR: |
| 4387 | case LE_EXPR: |
| 4388 | case GT_EXPR: |
| 4389 | case GE_EXPR: |
| 4390 | case EQ_EXPR: |
| 4391 | case NE_EXPR: |
| 4392 | case UNORDERED_EXPR: |
| 4393 | case ORDERED_EXPR: |
| 4394 | case UNLT_EXPR: |
| 4395 | case UNLE_EXPR: |
| 4396 | case UNGT_EXPR: |
| 4397 | case UNGE_EXPR: |
| 4398 | case UNEQ_EXPR: |
| 4399 | case LTGT_EXPR: |
| 4400 | case MAX_EXPR: |
| 4401 | case MIN_EXPR: |
| 4402 | op0 = TREE_OPERAND (expr, 0); |
| 4403 | STRIP_NOPS (op0); |
| 4404 | op1 = TREE_OPERAND (expr, 1); |
| 4405 | STRIP_NOPS (op1); |
| 4406 | break; |
| 4407 | |
| 4408 | default: |
| 4409 | /* Just an arbitrary value, FIXME. */ |
| 4410 | return comp_cost (target_spill_cost[speed], 0); |
| 4411 | } |
| 4412 | |
| 4413 | if (op0 == NULL_TREE |
| 4414 | || TREE_CODE (op0) == SSA_NAME || CONSTANT_CLASS_P (op0)) |
| 4415 | cost0 = no_cost; |
| 4416 | else |
| 4417 | cost0 = force_expr_to_var_cost (expr: op0, speed); |
| 4418 | |
| 4419 | if (op1 == NULL_TREE |
| 4420 | || TREE_CODE (op1) == SSA_NAME || CONSTANT_CLASS_P (op1)) |
| 4421 | cost1 = no_cost; |
| 4422 | else |
| 4423 | cost1 = force_expr_to_var_cost (expr: op1, speed); |
| 4424 | |
| 4425 | mode = TYPE_MODE (TREE_TYPE (expr)); |
| 4426 | switch (TREE_CODE (expr)) |
| 4427 | { |
| 4428 | case POINTER_PLUS_EXPR: |
| 4429 | case PLUS_EXPR: |
| 4430 | case MINUS_EXPR: |
| 4431 | case NEGATE_EXPR: |
| 4432 | cost = comp_cost (add_cost (speed, mode), 0); |
| 4433 | if (TREE_CODE (expr) != NEGATE_EXPR) |
| 4434 | { |
| 4435 | tree mult = NULL_TREE; |
| 4436 | comp_cost sa_cost; |
| 4437 | if (TREE_CODE (op1) == MULT_EXPR) |
| 4438 | mult = op1; |
| 4439 | else if (TREE_CODE (op0) == MULT_EXPR) |
| 4440 | mult = op0; |
| 4441 | |
| 4442 | if (mult != NULL_TREE |
| 4443 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
| 4444 | && cst_and_fits_in_hwi (TREE_OPERAND (mult, 1)) |
| 4445 | && get_shiftadd_cost (expr, mode: int_mode, cost0, cost1, mult, |
| 4446 | speed, cost: &sa_cost)) |
| 4447 | return sa_cost; |
| 4448 | } |
| 4449 | break; |
| 4450 | |
| 4451 | CASE_CONVERT: |
| 4452 | { |
| 4453 | tree inner_mode, outer_mode; |
| 4454 | outer_mode = TREE_TYPE (expr); |
| 4455 | inner_mode = TREE_TYPE (op0); |
| 4456 | cost = comp_cost (convert_cost (TYPE_MODE (outer_mode), |
| 4457 | TYPE_MODE (inner_mode), speed), 0); |
| 4458 | } |
| 4459 | break; |
| 4460 | |
| 4461 | case MULT_EXPR: |
| 4462 | if (cst_and_fits_in_hwi (op0)) |
| 4463 | cost = comp_cost (mult_by_coeff_cost (int_cst_value (op0), |
| 4464 | mode, speed), 0); |
| 4465 | else if (cst_and_fits_in_hwi (op1)) |
| 4466 | cost = comp_cost (mult_by_coeff_cost (int_cst_value (op1), |
| 4467 | mode, speed), 0); |
| 4468 | else |
| 4469 | return comp_cost (target_spill_cost [speed], 0); |
| 4470 | break; |
| 4471 | |
| 4472 | case EXACT_DIV_EXPR: |
| 4473 | case TRUNC_DIV_EXPR: |
| 4474 | /* Division by power of two is usually cheap, so we allow it. Forbid |
| 4475 | anything else. */ |
| 4476 | if (integer_pow2p (TREE_OPERAND (expr, 1))) |
| 4477 | cost = comp_cost (add_cost (speed, mode), 0); |
| 4478 | else |
| 4479 | cost = comp_cost (target_spill_cost[speed], 0); |
| 4480 | break; |
| 4481 | |
| 4482 | case BIT_AND_EXPR: |
| 4483 | case BIT_IOR_EXPR: |
| 4484 | case BIT_NOT_EXPR: |
| 4485 | case LSHIFT_EXPR: |
| 4486 | case RSHIFT_EXPR: |
| 4487 | cost = comp_cost (add_cost (speed, mode), 0); |
| 4488 | break; |
| 4489 | case COND_EXPR: |
| 4490 | op0 = TREE_OPERAND (expr, 0); |
| 4491 | STRIP_NOPS (op0); |
| 4492 | if (op0 == NULL_TREE || TREE_CODE (op0) == SSA_NAME |
| 4493 | || CONSTANT_CLASS_P (op0)) |
| 4494 | cost = no_cost; |
| 4495 | else |
| 4496 | cost = force_expr_to_var_cost (expr: op0, speed); |
| 4497 | break; |
| 4498 | case LT_EXPR: |
| 4499 | case LE_EXPR: |
| 4500 | case GT_EXPR: |
| 4501 | case GE_EXPR: |
| 4502 | case EQ_EXPR: |
| 4503 | case NE_EXPR: |
| 4504 | case UNORDERED_EXPR: |
| 4505 | case ORDERED_EXPR: |
| 4506 | case UNLT_EXPR: |
| 4507 | case UNLE_EXPR: |
| 4508 | case UNGT_EXPR: |
| 4509 | case UNGE_EXPR: |
| 4510 | case UNEQ_EXPR: |
| 4511 | case LTGT_EXPR: |
| 4512 | case MAX_EXPR: |
| 4513 | case MIN_EXPR: |
| 4514 | /* Simply use add cost for now, FIXME if there is some more accurate cost |
| 4515 | evaluation way. */ |
| 4516 | cost = comp_cost (add_cost (speed, mode), 0); |
| 4517 | break; |
| 4518 | |
| 4519 | default: |
| 4520 | gcc_unreachable (); |
| 4521 | } |
| 4522 | |
| 4523 | cost += cost0; |
| 4524 | cost += cost1; |
| 4525 | return cost; |
| 4526 | } |
| 4527 | |
| 4528 | /* Estimates cost of forcing EXPR into a variable. INV_VARS is a set of the |
| 4529 | invariants the computation depends on. */ |
| 4530 | |
| 4531 | static comp_cost |
| 4532 | force_var_cost (struct ivopts_data *data, tree expr, bitmap *inv_vars) |
| 4533 | { |
| 4534 | if (!expr) |
| 4535 | return no_cost; |
| 4536 | |
| 4537 | find_inv_vars (data, expr_p: &expr, inv_vars); |
| 4538 | return force_expr_to_var_cost (expr, speed: data->speed); |
| 4539 | } |
| 4540 | |
| 4541 | /* Returns cost of auto-modifying address expression in shape base + offset. |
| 4542 | AINC_STEP is step size of the address IV. AINC_OFFSET is offset of the |
| 4543 | address expression. The address expression has ADDR_MODE in addr space |
| 4544 | AS. The memory access has MEM_MODE. SPEED means we are optimizing for |
| 4545 | speed or size. */ |
| 4546 | |
| 4547 | enum ainc_type |
| 4548 | { |
| 4549 | AINC_PRE_INC, /* Pre increment. */ |
| 4550 | AINC_PRE_DEC, /* Pre decrement. */ |
| 4551 | AINC_POST_INC, /* Post increment. */ |
| 4552 | AINC_POST_DEC, /* Post decrement. */ |
| 4553 | AINC_NONE /* Also the number of auto increment types. */ |
| 4554 | }; |
| 4555 | |
| 4556 | struct ainc_cost_data |
| 4557 | { |
| 4558 | int64_t costs[AINC_NONE]; |
| 4559 | }; |
| 4560 | |
| 4561 | static comp_cost |
| 4562 | get_address_cost_ainc (poly_int64 ainc_step, poly_int64 ainc_offset, |
| 4563 | machine_mode addr_mode, machine_mode mem_mode, |
| 4564 | addr_space_t as, bool speed) |
| 4565 | { |
| 4566 | if (!USE_LOAD_PRE_DECREMENT (mem_mode) |
| 4567 | && !USE_STORE_PRE_DECREMENT (mem_mode) |
| 4568 | && !USE_LOAD_POST_DECREMENT (mem_mode) |
| 4569 | && !USE_STORE_POST_DECREMENT (mem_mode) |
| 4570 | && !USE_LOAD_PRE_INCREMENT (mem_mode) |
| 4571 | && !USE_STORE_PRE_INCREMENT (mem_mode) |
| 4572 | && !USE_LOAD_POST_INCREMENT (mem_mode) |
| 4573 | && !USE_STORE_POST_INCREMENT (mem_mode)) |
| 4574 | return infinite_cost; |
| 4575 | |
| 4576 | static vec<ainc_cost_data *> ainc_cost_data_list; |
| 4577 | unsigned idx = (unsigned) as * MAX_MACHINE_MODE + (unsigned) mem_mode; |
| 4578 | if (idx >= ainc_cost_data_list.length ()) |
| 4579 | { |
| 4580 | unsigned nsize = ((unsigned) as + 1) *MAX_MACHINE_MODE; |
| 4581 | |
| 4582 | gcc_assert (nsize > idx); |
| 4583 | ainc_cost_data_list.safe_grow_cleared (len: nsize, exact: true); |
| 4584 | } |
| 4585 | |
| 4586 | ainc_cost_data *data = ainc_cost_data_list[idx]; |
| 4587 | if (data == NULL) |
| 4588 | { |
| 4589 | rtx reg = gen_raw_REG (addr_mode, LAST_VIRTUAL_REGISTER + 1); |
| 4590 | |
| 4591 | data = (ainc_cost_data *) xcalloc (1, sizeof (*data)); |
| 4592 | data->costs[AINC_PRE_DEC] = INFTY; |
| 4593 | data->costs[AINC_POST_DEC] = INFTY; |
| 4594 | data->costs[AINC_PRE_INC] = INFTY; |
| 4595 | data->costs[AINC_POST_INC] = INFTY; |
| 4596 | if (USE_LOAD_PRE_DECREMENT (mem_mode) |
| 4597 | || USE_STORE_PRE_DECREMENT (mem_mode)) |
| 4598 | { |
| 4599 | rtx addr = gen_rtx_PRE_DEC (addr_mode, reg); |
| 4600 | |
| 4601 | if (memory_address_addr_space_p (mem_mode, addr, as)) |
| 4602 | data->costs[AINC_PRE_DEC] |
| 4603 | = address_cost (addr, mem_mode, as, speed); |
| 4604 | } |
| 4605 | if (USE_LOAD_POST_DECREMENT (mem_mode) |
| 4606 | || USE_STORE_POST_DECREMENT (mem_mode)) |
| 4607 | { |
| 4608 | rtx addr = gen_rtx_POST_DEC (addr_mode, reg); |
| 4609 | |
| 4610 | if (memory_address_addr_space_p (mem_mode, addr, as)) |
| 4611 | data->costs[AINC_POST_DEC] |
| 4612 | = address_cost (addr, mem_mode, as, speed); |
| 4613 | } |
| 4614 | if (USE_LOAD_PRE_INCREMENT (mem_mode) |
| 4615 | || USE_STORE_PRE_INCREMENT (mem_mode)) |
| 4616 | { |
| 4617 | rtx addr = gen_rtx_PRE_INC (addr_mode, reg); |
| 4618 | |
| 4619 | if (memory_address_addr_space_p (mem_mode, addr, as)) |
| 4620 | data->costs[AINC_PRE_INC] |
| 4621 | = address_cost (addr, mem_mode, as, speed); |
| 4622 | } |
| 4623 | if (USE_LOAD_POST_INCREMENT (mem_mode) |
| 4624 | || USE_STORE_POST_INCREMENT (mem_mode)) |
| 4625 | { |
| 4626 | rtx addr = gen_rtx_POST_INC (addr_mode, reg); |
| 4627 | |
| 4628 | if (memory_address_addr_space_p (mem_mode, addr, as)) |
| 4629 | data->costs[AINC_POST_INC] |
| 4630 | = address_cost (addr, mem_mode, as, speed); |
| 4631 | } |
| 4632 | ainc_cost_data_list[idx] = data; |
| 4633 | } |
| 4634 | |
| 4635 | poly_int64 msize = GET_MODE_SIZE (mode: mem_mode); |
| 4636 | if (known_eq (ainc_offset, 0) && known_eq (msize, ainc_step)) |
| 4637 | return comp_cost (data->costs[AINC_POST_INC], 0); |
| 4638 | if (known_eq (ainc_offset, 0) && known_eq (msize, -ainc_step)) |
| 4639 | return comp_cost (data->costs[AINC_POST_DEC], 0); |
| 4640 | if (known_eq (ainc_offset, msize) && known_eq (msize, ainc_step)) |
| 4641 | return comp_cost (data->costs[AINC_PRE_INC], 0); |
| 4642 | if (known_eq (ainc_offset, -msize) && known_eq (msize, -ainc_step)) |
| 4643 | return comp_cost (data->costs[AINC_PRE_DEC], 0); |
| 4644 | |
| 4645 | return infinite_cost; |
| 4646 | } |
| 4647 | |
| 4648 | /* Return cost of computing USE's address expression by using CAND. |
| 4649 | AFF_INV and AFF_VAR represent invariant and variant parts of the |
| 4650 | address expression, respectively. If AFF_INV is simple, store |
| 4651 | the loop invariant variables which are depended by it in INV_VARS; |
| 4652 | if AFF_INV is complicated, handle it as a new invariant expression |
| 4653 | and record it in INV_EXPR. RATIO indicates multiple times between |
| 4654 | steps of USE and CAND. If CAN_AUTOINC is nonNULL, store boolean |
| 4655 | value to it indicating if this is an auto-increment address. */ |
| 4656 | |
| 4657 | static comp_cost |
| 4658 | get_address_cost (struct ivopts_data *data, struct iv_use *use, |
| 4659 | struct iv_cand *cand, aff_tree *aff_inv, |
| 4660 | aff_tree *aff_var, HOST_WIDE_INT ratio, |
| 4661 | bitmap *inv_vars, iv_inv_expr_ent **inv_expr, |
| 4662 | bool *can_autoinc, bool speed) |
| 4663 | { |
| 4664 | rtx addr; |
| 4665 | bool simple_inv = true; |
| 4666 | tree comp_inv = NULL_TREE, type = aff_var->type; |
| 4667 | comp_cost var_cost = no_cost, cost = no_cost; |
| 4668 | struct mem_address parts = {NULL_TREE, integer_one_node, |
| 4669 | NULL_TREE, NULL_TREE, NULL_TREE}; |
| 4670 | machine_mode addr_mode = TYPE_MODE (type); |
| 4671 | machine_mode mem_mode = TYPE_MODE (use->mem_type); |
| 4672 | addr_space_t as = TYPE_ADDR_SPACE (TREE_TYPE (use->iv->base)); |
| 4673 | /* Only true if ratio != 1. */ |
| 4674 | bool ok_with_ratio_p = false; |
| 4675 | bool ok_without_ratio_p = false; |
| 4676 | code_helper code = ERROR_MARK; |
| 4677 | |
| 4678 | if (use->type == USE_PTR_ADDRESS) |
| 4679 | { |
| 4680 | gcall *call = as_a<gcall *> (p: use->stmt); |
| 4681 | gcc_assert (gimple_call_internal_p (call)); |
| 4682 | code = gimple_call_internal_fn (gs: call); |
| 4683 | } |
| 4684 | |
| 4685 | if (!aff_combination_const_p (aff: aff_inv)) |
| 4686 | { |
| 4687 | parts.index = integer_one_node; |
| 4688 | /* Addressing mode "base + index". */ |
| 4689 | ok_without_ratio_p = valid_mem_ref_p (mem_mode, as, &parts, code); |
| 4690 | if (ratio != 1) |
| 4691 | { |
| 4692 | parts.step = wide_int_to_tree (type, cst: ratio); |
| 4693 | /* Addressing mode "base + index << scale". */ |
| 4694 | ok_with_ratio_p = valid_mem_ref_p (mem_mode, as, &parts, code); |
| 4695 | if (!ok_with_ratio_p) |
| 4696 | parts.step = NULL_TREE; |
| 4697 | } |
| 4698 | if (ok_with_ratio_p || ok_without_ratio_p) |
| 4699 | { |
| 4700 | if (maybe_ne (a: aff_inv->offset, b: 0)) |
| 4701 | { |
| 4702 | parts.offset = wide_int_to_tree (sizetype, cst: aff_inv->offset); |
| 4703 | /* Addressing mode "base + index [<< scale] + offset". */ |
| 4704 | if (!valid_mem_ref_p (mem_mode, as, &parts, code)) |
| 4705 | parts.offset = NULL_TREE; |
| 4706 | else |
| 4707 | aff_inv->offset = 0; |
| 4708 | } |
| 4709 | |
| 4710 | move_fixed_address_to_symbol (&parts, aff_inv); |
| 4711 | /* Base is fixed address and is moved to symbol part. */ |
| 4712 | if (parts.symbol != NULL_TREE && aff_combination_zero_p (aff: aff_inv)) |
| 4713 | parts.base = NULL_TREE; |
| 4714 | |
| 4715 | /* Addressing mode "symbol + base + index [<< scale] [+ offset]". */ |
| 4716 | if (parts.symbol != NULL_TREE |
| 4717 | && !valid_mem_ref_p (mem_mode, as, &parts, code)) |
| 4718 | { |
| 4719 | aff_combination_add_elt (aff_inv, parts.symbol, 1); |
| 4720 | parts.symbol = NULL_TREE; |
| 4721 | /* Reset SIMPLE_INV since symbol address needs to be computed |
| 4722 | outside of address expression in this case. */ |
| 4723 | simple_inv = false; |
| 4724 | /* Symbol part is moved back to base part, it can't be NULL. */ |
| 4725 | parts.base = integer_one_node; |
| 4726 | } |
| 4727 | } |
| 4728 | else |
| 4729 | parts.index = NULL_TREE; |
| 4730 | } |
| 4731 | else |
| 4732 | { |
| 4733 | poly_int64 ainc_step; |
| 4734 | if (can_autoinc |
| 4735 | && ratio == 1 |
| 4736 | && ptrdiff_tree_p (cand->iv->step, &ainc_step)) |
| 4737 | { |
| 4738 | poly_int64 ainc_offset = (aff_inv->offset).force_shwi (); |
| 4739 | |
| 4740 | if (stmt_after_increment (loop: data->current_loop, cand, stmt: use->stmt)) |
| 4741 | ainc_offset += ainc_step; |
| 4742 | cost = get_address_cost_ainc (ainc_step, ainc_offset, |
| 4743 | addr_mode, mem_mode, as, speed); |
| 4744 | if (!cost.infinite_cost_p ()) |
| 4745 | { |
| 4746 | *can_autoinc = true; |
| 4747 | return cost; |
| 4748 | } |
| 4749 | cost = no_cost; |
| 4750 | } |
| 4751 | if (!aff_combination_zero_p (aff: aff_inv)) |
| 4752 | { |
| 4753 | parts.offset = wide_int_to_tree (sizetype, cst: aff_inv->offset); |
| 4754 | /* Addressing mode "base + offset". */ |
| 4755 | if (!valid_mem_ref_p (mem_mode, as, &parts, code)) |
| 4756 | parts.offset = NULL_TREE; |
| 4757 | else |
| 4758 | aff_inv->offset = 0; |
| 4759 | } |
| 4760 | } |
| 4761 | |
| 4762 | if (simple_inv) |
| 4763 | simple_inv = (aff_inv == NULL |
| 4764 | || aff_combination_const_p (aff: aff_inv) |
| 4765 | || aff_combination_singleton_var_p (aff: aff_inv)); |
| 4766 | if (!aff_combination_zero_p (aff: aff_inv)) |
| 4767 | comp_inv = aff_combination_to_tree (aff_inv); |
| 4768 | if (comp_inv != NULL_TREE) |
| 4769 | cost = force_var_cost (data, expr: comp_inv, inv_vars); |
| 4770 | if (ratio != 1 && parts.step == NULL_TREE) |
| 4771 | var_cost += mult_by_coeff_cost (ratio, addr_mode, speed); |
| 4772 | if (comp_inv != NULL_TREE && parts.index == NULL_TREE) |
| 4773 | var_cost += add_cost (speed, mode: addr_mode); |
| 4774 | |
| 4775 | if (comp_inv && inv_expr && !simple_inv) |
| 4776 | { |
| 4777 | *inv_expr = get_loop_invariant_expr (data, inv_expr: comp_inv); |
| 4778 | /* Clear depends on. */ |
| 4779 | if (*inv_expr != NULL && inv_vars && *inv_vars) |
| 4780 | bitmap_clear (*inv_vars); |
| 4781 | |
| 4782 | /* Cost of small invariant expression adjusted against loop niters |
| 4783 | is usually zero, which makes it difficult to be differentiated |
| 4784 | from candidate based on loop invariant variables. Secondly, the |
| 4785 | generated invariant expression may not be hoisted out of loop by |
| 4786 | following pass. We penalize the cost by rounding up in order to |
| 4787 | neutralize such effects. */ |
| 4788 | cost.cost = adjust_setup_cost (data, cost: cost.cost, round_up_p: true); |
| 4789 | cost.scratch = cost.cost; |
| 4790 | } |
| 4791 | |
| 4792 | cost += var_cost; |
| 4793 | addr = addr_for_mem_ref (&parts, as, false); |
| 4794 | gcc_assert (memory_address_addr_space_p (mem_mode, addr, as)); |
| 4795 | cost += address_cost (addr, mem_mode, as, speed); |
| 4796 | |
| 4797 | if (parts.symbol != NULL_TREE) |
| 4798 | cost.complexity += 1; |
| 4799 | /* Don't increase the complexity of adding a scaled index if it's |
| 4800 | the only kind of index that the target allows. */ |
| 4801 | if (parts.step != NULL_TREE && ok_without_ratio_p) |
| 4802 | cost.complexity += 1; |
| 4803 | if (parts.base != NULL_TREE && parts.index != NULL_TREE) |
| 4804 | cost.complexity += 1; |
| 4805 | if (parts.offset != NULL_TREE && !integer_zerop (parts.offset)) |
| 4806 | cost.complexity += 1; |
| 4807 | |
| 4808 | return cost; |
| 4809 | } |
| 4810 | |
| 4811 | /* Scale (multiply) the computed COST (except scratch part that should be |
| 4812 | hoisted out a loop) by header->frequency / AT->frequency, which makes |
| 4813 | expected cost more accurate. */ |
| 4814 | |
| 4815 | static comp_cost |
| 4816 | get_scaled_computation_cost_at (ivopts_data *data, gimple *at, comp_cost cost) |
| 4817 | { |
| 4818 | if (data->speed |
| 4819 | && data->current_loop->header->count.to_frequency (cfun) > 0) |
| 4820 | { |
| 4821 | basic_block bb = gimple_bb (g: at); |
| 4822 | gcc_assert (cost.scratch <= cost.cost); |
| 4823 | int scale_factor = (int)(intptr_t) bb->aux; |
| 4824 | if (scale_factor == 1) |
| 4825 | return cost; |
| 4826 | |
| 4827 | int64_t scaled_cost |
| 4828 | = cost.scratch + (cost.cost - cost.scratch) * scale_factor; |
| 4829 | |
| 4830 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 4831 | fprintf (stream: dump_file, format: "Scaling cost based on bb prob by %2.2f: " |
| 4832 | "%" PRId64 " (scratch: %" PRId64 ") -> %" PRId64 "\n" , |
| 4833 | 1.0f * scale_factor, cost.cost, cost.scratch, scaled_cost); |
| 4834 | |
| 4835 | cost.cost = scaled_cost; |
| 4836 | } |
| 4837 | |
| 4838 | return cost; |
| 4839 | } |
| 4840 | |
| 4841 | /* Determines the cost of the computation by that USE is expressed |
| 4842 | from induction variable CAND. If ADDRESS_P is true, we just need |
| 4843 | to create an address from it, otherwise we want to get it into |
| 4844 | register. A set of invariants we depend on is stored in INV_VARS. |
| 4845 | If CAN_AUTOINC is nonnull, use it to record whether autoinc |
| 4846 | addressing is likely. If INV_EXPR is nonnull, record invariant |
| 4847 | expr entry in it. */ |
| 4848 | |
| 4849 | static comp_cost |
| 4850 | get_computation_cost (struct ivopts_data *data, struct iv_use *use, |
| 4851 | struct iv_cand *cand, bool address_p, bitmap *inv_vars, |
| 4852 | bool *can_autoinc, iv_inv_expr_ent **inv_expr) |
| 4853 | { |
| 4854 | gimple *at = use->stmt; |
| 4855 | tree ubase = use->iv->base, cbase = cand->iv->base; |
| 4856 | tree utype = TREE_TYPE (ubase), ctype = TREE_TYPE (cbase); |
| 4857 | tree comp_inv = NULL_TREE; |
| 4858 | HOST_WIDE_INT ratio, aratio; |
| 4859 | comp_cost cost; |
| 4860 | widest_int rat; |
| 4861 | aff_tree aff_inv, aff_var; |
| 4862 | bool speed = optimize_bb_for_speed_p (gimple_bb (g: at)); |
| 4863 | |
| 4864 | if (inv_vars) |
| 4865 | *inv_vars = NULL; |
| 4866 | if (can_autoinc) |
| 4867 | *can_autoinc = false; |
| 4868 | if (inv_expr) |
| 4869 | *inv_expr = NULL; |
| 4870 | |
| 4871 | /* Check if we have enough precision to express the values of use. */ |
| 4872 | if (TYPE_PRECISION (utype) > TYPE_PRECISION (ctype)) |
| 4873 | return infinite_cost; |
| 4874 | |
| 4875 | if (address_p |
| 4876 | || (use->iv->base_object |
| 4877 | && cand->iv->base_object |
| 4878 | && POINTER_TYPE_P (TREE_TYPE (use->iv->base_object)) |
| 4879 | && POINTER_TYPE_P (TREE_TYPE (cand->iv->base_object)))) |
| 4880 | { |
| 4881 | /* Do not try to express address of an object with computation based |
| 4882 | on address of a different object. This may cause problems in rtl |
| 4883 | level alias analysis (that does not expect this to be happening, |
| 4884 | as this is illegal in C), and would be unlikely to be useful |
| 4885 | anyway. */ |
| 4886 | if (use->iv->base_object |
| 4887 | && cand->iv->base_object |
| 4888 | && !operand_equal_p (use->iv->base_object, cand->iv->base_object, flags: 0)) |
| 4889 | return infinite_cost; |
| 4890 | } |
| 4891 | |
| 4892 | if (!get_computation_aff_1 (data, at, use, cand, aff_inv: &aff_inv, aff_var: &aff_var, prat: &rat) |
| 4893 | || !wi::fits_shwi_p (x: rat)) |
| 4894 | return infinite_cost; |
| 4895 | |
| 4896 | ratio = rat.to_shwi (); |
| 4897 | if (address_p) |
| 4898 | { |
| 4899 | cost = get_address_cost (data, use, cand, aff_inv: &aff_inv, aff_var: &aff_var, ratio, |
| 4900 | inv_vars, inv_expr, can_autoinc, speed); |
| 4901 | cost = get_scaled_computation_cost_at (data, at, cost); |
| 4902 | /* For doloop IV cand, add on the extra cost. */ |
| 4903 | cost += cand->doloop_p ? targetm.doloop_cost_for_address : 0; |
| 4904 | return cost; |
| 4905 | } |
| 4906 | |
| 4907 | bool simple_inv = (aff_combination_const_p (aff: &aff_inv) |
| 4908 | || aff_combination_singleton_var_p (aff: &aff_inv)); |
| 4909 | tree signed_type = signed_type_for (aff_combination_type (aff: &aff_inv)); |
| 4910 | aff_combination_convert (&aff_inv, signed_type); |
| 4911 | if (!aff_combination_zero_p (aff: &aff_inv)) |
| 4912 | comp_inv = aff_combination_to_tree (&aff_inv); |
| 4913 | |
| 4914 | cost = force_var_cost (data, expr: comp_inv, inv_vars); |
| 4915 | if (comp_inv && inv_expr && !simple_inv) |
| 4916 | { |
| 4917 | *inv_expr = get_loop_invariant_expr (data, inv_expr: comp_inv); |
| 4918 | /* Clear depends on. */ |
| 4919 | if (*inv_expr != NULL && inv_vars && *inv_vars) |
| 4920 | bitmap_clear (*inv_vars); |
| 4921 | |
| 4922 | cost.cost = adjust_setup_cost (data, cost: cost.cost); |
| 4923 | /* Record setup cost in scratch field. */ |
| 4924 | cost.scratch = cost.cost; |
| 4925 | } |
| 4926 | /* Cost of constant integer can be covered when adding invariant part to |
| 4927 | variant part. */ |
| 4928 | else if (comp_inv && CONSTANT_CLASS_P (comp_inv)) |
| 4929 | cost = no_cost; |
| 4930 | |
| 4931 | /* Need type narrowing to represent use with cand. */ |
| 4932 | if (TYPE_PRECISION (utype) < TYPE_PRECISION (ctype)) |
| 4933 | { |
| 4934 | machine_mode outer_mode = TYPE_MODE (utype); |
| 4935 | machine_mode inner_mode = TYPE_MODE (ctype); |
| 4936 | cost += comp_cost (convert_cost (to_mode: outer_mode, from_mode: inner_mode, speed), 0); |
| 4937 | } |
| 4938 | |
| 4939 | /* Turn a + i * (-c) into a - i * c. */ |
| 4940 | if (ratio < 0 && comp_inv && !integer_zerop (comp_inv)) |
| 4941 | aratio = -ratio; |
| 4942 | else |
| 4943 | aratio = ratio; |
| 4944 | |
| 4945 | if (ratio != 1) |
| 4946 | cost += mult_by_coeff_cost (aratio, TYPE_MODE (utype), speed); |
| 4947 | |
| 4948 | /* TODO: We may also need to check if we can compute a + i * 4 in one |
| 4949 | instruction. */ |
| 4950 | /* Need to add up the invariant and variant parts. */ |
| 4951 | if (comp_inv && !integer_zerop (comp_inv)) |
| 4952 | cost += add_cost (speed, TYPE_MODE (utype)); |
| 4953 | |
| 4954 | cost = get_scaled_computation_cost_at (data, at, cost); |
| 4955 | |
| 4956 | /* For doloop IV cand, add on the extra cost. */ |
| 4957 | if (cand->doloop_p && use->type == USE_NONLINEAR_EXPR) |
| 4958 | cost += targetm.doloop_cost_for_generic; |
| 4959 | |
| 4960 | return cost; |
| 4961 | } |
| 4962 | |
| 4963 | /* Determines cost of computing the use in GROUP with CAND in a generic |
| 4964 | expression. */ |
| 4965 | |
| 4966 | static bool |
| 4967 | determine_group_iv_cost_generic (struct ivopts_data *data, |
| 4968 | struct iv_group *group, struct iv_cand *cand) |
| 4969 | { |
| 4970 | comp_cost cost; |
| 4971 | iv_inv_expr_ent *inv_expr = NULL; |
| 4972 | bitmap inv_vars = NULL, inv_exprs = NULL; |
| 4973 | struct iv_use *use = group->vuses[0]; |
| 4974 | |
| 4975 | /* The simple case first -- if we need to express value of the preserved |
| 4976 | original biv, the cost is 0. This also prevents us from counting the |
| 4977 | cost of increment twice -- once at this use and once in the cost of |
| 4978 | the candidate. */ |
| 4979 | if (cand->pos == IP_ORIGINAL && cand->incremented_at == use->stmt) |
| 4980 | cost = no_cost; |
| 4981 | /* If the IV candidate involves undefined SSA values and is not the |
| 4982 | same IV as on the USE avoid using that candidate here. */ |
| 4983 | else if (cand->involves_undefs |
| 4984 | && (!use->iv || !operand_equal_p (cand->iv->base, use->iv->base, flags: 0))) |
| 4985 | return false; |
| 4986 | else |
| 4987 | cost = get_computation_cost (data, use, cand, address_p: false, |
| 4988 | inv_vars: &inv_vars, NULL, inv_expr: &inv_expr); |
| 4989 | |
| 4990 | if (inv_expr) |
| 4991 | { |
| 4992 | inv_exprs = BITMAP_ALLOC (NULL); |
| 4993 | bitmap_set_bit (inv_exprs, inv_expr->id); |
| 4994 | } |
| 4995 | set_group_iv_cost (data, group, cand, cost, inv_vars, |
| 4996 | NULL_TREE, comp: ERROR_MARK, inv_exprs); |
| 4997 | return !cost.infinite_cost_p (); |
| 4998 | } |
| 4999 | |
| 5000 | /* Determines cost of computing uses in GROUP with CAND in addresses. */ |
| 5001 | |
| 5002 | static bool |
| 5003 | determine_group_iv_cost_address (struct ivopts_data *data, |
| 5004 | struct iv_group *group, struct iv_cand *cand) |
| 5005 | { |
| 5006 | unsigned i; |
| 5007 | bitmap inv_vars = NULL, inv_exprs = NULL; |
| 5008 | bool can_autoinc; |
| 5009 | iv_inv_expr_ent *inv_expr = NULL; |
| 5010 | struct iv_use *use = group->vuses[0]; |
| 5011 | comp_cost sum_cost = no_cost, cost; |
| 5012 | |
| 5013 | cost = get_computation_cost (data, use, cand, address_p: true, |
| 5014 | inv_vars: &inv_vars, can_autoinc: &can_autoinc, inv_expr: &inv_expr); |
| 5015 | |
| 5016 | if (inv_expr) |
| 5017 | { |
| 5018 | inv_exprs = BITMAP_ALLOC (NULL); |
| 5019 | bitmap_set_bit (inv_exprs, inv_expr->id); |
| 5020 | } |
| 5021 | sum_cost = cost; |
| 5022 | if (!sum_cost.infinite_cost_p () && cand->ainc_use == use) |
| 5023 | { |
| 5024 | if (can_autoinc) |
| 5025 | sum_cost -= cand->cost_step; |
| 5026 | /* If we generated the candidate solely for exploiting autoincrement |
| 5027 | opportunities, and it turns out it can't be used, set the cost to |
| 5028 | infinity to make sure we ignore it. */ |
| 5029 | else if (cand->pos == IP_AFTER_USE || cand->pos == IP_BEFORE_USE) |
| 5030 | sum_cost = infinite_cost; |
| 5031 | } |
| 5032 | |
| 5033 | /* Compute and add costs for rest uses of this group. */ |
| 5034 | for (i = 1; i < group->vuses.length () && !sum_cost.infinite_cost_p (); i++) |
| 5035 | { |
| 5036 | struct iv_use *next = group->vuses[i]; |
| 5037 | |
| 5038 | /* TODO: We could skip computing cost for sub iv_use when it has the |
| 5039 | same cost as the first iv_use, but the cost really depends on the |
| 5040 | offset and where the iv_use is. */ |
| 5041 | cost = get_computation_cost (data, use: next, cand, address_p: true, |
| 5042 | NULL, can_autoinc: &can_autoinc, inv_expr: &inv_expr); |
| 5043 | if (inv_expr) |
| 5044 | { |
| 5045 | if (!inv_exprs) |
| 5046 | inv_exprs = BITMAP_ALLOC (NULL); |
| 5047 | |
| 5048 | /* Uses in a group can share setup code, |
| 5049 | so only add setup cost once. */ |
| 5050 | if (bitmap_bit_p (inv_exprs, inv_expr->id)) |
| 5051 | cost -= cost.scratch; |
| 5052 | else |
| 5053 | bitmap_set_bit (inv_exprs, inv_expr->id); |
| 5054 | } |
| 5055 | sum_cost += cost; |
| 5056 | } |
| 5057 | set_group_iv_cost (data, group, cand, cost: sum_cost, inv_vars, |
| 5058 | NULL_TREE, comp: ERROR_MARK, inv_exprs); |
| 5059 | |
| 5060 | return !sum_cost.infinite_cost_p (); |
| 5061 | } |
| 5062 | |
| 5063 | /* Computes value of candidate CAND at position AT in iteration DESC->NITER, |
| 5064 | and stores it to VAL. */ |
| 5065 | |
| 5066 | static void |
| 5067 | cand_value_at (class loop *loop, struct iv_cand *cand, gimple *at, |
| 5068 | class tree_niter_desc *desc, aff_tree *val) |
| 5069 | { |
| 5070 | aff_tree step, delta, nit; |
| 5071 | struct iv *iv = cand->iv; |
| 5072 | tree type = TREE_TYPE (iv->base); |
| 5073 | tree niter = desc->niter; |
| 5074 | bool after_adjust = stmt_after_increment (loop, cand, stmt: at); |
| 5075 | tree steptype; |
| 5076 | |
| 5077 | if (POINTER_TYPE_P (type)) |
| 5078 | steptype = sizetype; |
| 5079 | else |
| 5080 | steptype = unsigned_type_for (type); |
| 5081 | |
| 5082 | /* If AFTER_ADJUST is required, the code below generates the equivalent |
| 5083 | of BASE + NITER * STEP + STEP, when ideally we'd prefer the expression |
| 5084 | BASE + (NITER + 1) * STEP, especially when NITER is often of the form |
| 5085 | SSA_NAME - 1. Unfortunately, guaranteeing that adding 1 to NITER |
| 5086 | doesn't overflow is tricky, so we peek inside the TREE_NITER_DESC |
| 5087 | class for common idioms that we know are safe. */ |
| 5088 | if (after_adjust |
| 5089 | && desc->control.no_overflow |
| 5090 | && integer_onep (desc->control.step) |
| 5091 | && (desc->cmp == LT_EXPR |
| 5092 | || desc->cmp == NE_EXPR) |
| 5093 | && TREE_CODE (desc->bound) == SSA_NAME) |
| 5094 | { |
| 5095 | if (integer_onep (desc->control.base)) |
| 5096 | { |
| 5097 | niter = desc->bound; |
| 5098 | after_adjust = false; |
| 5099 | } |
| 5100 | else if (TREE_CODE (niter) == MINUS_EXPR |
| 5101 | && integer_onep (TREE_OPERAND (niter, 1))) |
| 5102 | { |
| 5103 | niter = TREE_OPERAND (niter, 0); |
| 5104 | after_adjust = false; |
| 5105 | } |
| 5106 | } |
| 5107 | |
| 5108 | tree_to_aff_combination (iv->step, TREE_TYPE (iv->step), &step); |
| 5109 | aff_combination_convert (&step, steptype); |
| 5110 | tree_to_aff_combination (niter, TREE_TYPE (niter), &nit); |
| 5111 | aff_combination_convert (&nit, steptype); |
| 5112 | aff_combination_mult (&nit, &step, &delta); |
| 5113 | if (after_adjust) |
| 5114 | aff_combination_add (&delta, &step); |
| 5115 | |
| 5116 | tree_to_aff_combination (iv->base, type, val); |
| 5117 | if (!POINTER_TYPE_P (type)) |
| 5118 | aff_combination_convert (val, steptype); |
| 5119 | aff_combination_add (val, &delta); |
| 5120 | } |
| 5121 | |
| 5122 | /* Returns period of induction variable iv. */ |
| 5123 | |
| 5124 | static tree |
| 5125 | iv_period (struct iv *iv) |
| 5126 | { |
| 5127 | tree step = iv->step, period, type; |
| 5128 | tree pow2div; |
| 5129 | |
| 5130 | gcc_assert (step && TREE_CODE (step) == INTEGER_CST); |
| 5131 | |
| 5132 | type = unsigned_type_for (TREE_TYPE (step)); |
| 5133 | /* Period of the iv is lcm (step, type_range)/step -1, |
| 5134 | i.e., N*type_range/step - 1. Since type range is power |
| 5135 | of two, N == (step >> num_of_ending_zeros_binary (step), |
| 5136 | so the final result is |
| 5137 | |
| 5138 | (type_range >> num_of_ending_zeros_binary (step)) - 1 |
| 5139 | |
| 5140 | */ |
| 5141 | pow2div = num_ending_zeros (step); |
| 5142 | |
| 5143 | period = build_low_bits_mask (type, |
| 5144 | (TYPE_PRECISION (type) |
| 5145 | - tree_to_uhwi (pow2div))); |
| 5146 | |
| 5147 | return period; |
| 5148 | } |
| 5149 | |
| 5150 | /* Returns the comparison operator used when eliminating the iv USE. */ |
| 5151 | |
| 5152 | static enum tree_code |
| 5153 | iv_elimination_compare (struct ivopts_data *data, struct iv_use *use) |
| 5154 | { |
| 5155 | class loop *loop = data->current_loop; |
| 5156 | basic_block ex_bb; |
| 5157 | edge exit; |
| 5158 | |
| 5159 | ex_bb = gimple_bb (g: use->stmt); |
| 5160 | exit = EDGE_SUCC (ex_bb, 0); |
| 5161 | if (flow_bb_inside_loop_p (loop, exit->dest)) |
| 5162 | exit = EDGE_SUCC (ex_bb, 1); |
| 5163 | |
| 5164 | return (exit->flags & EDGE_TRUE_VALUE ? EQ_EXPR : NE_EXPR); |
| 5165 | } |
| 5166 | |
| 5167 | /* Returns true if we can prove that BASE - OFFSET does not overflow. For now, |
| 5168 | we only detect the situation that BASE = SOMETHING + OFFSET, where the |
| 5169 | calculation is performed in non-wrapping type. |
| 5170 | |
| 5171 | TODO: More generally, we could test for the situation that |
| 5172 | BASE = SOMETHING + OFFSET' and OFFSET is between OFFSET' and zero. |
| 5173 | This would require knowing the sign of OFFSET. */ |
| 5174 | |
| 5175 | static bool |
| 5176 | difference_cannot_overflow_p (struct ivopts_data *data, tree base, tree offset) |
| 5177 | { |
| 5178 | enum tree_code code; |
| 5179 | tree e1, e2; |
| 5180 | aff_tree aff_e1, aff_e2, aff_offset; |
| 5181 | |
| 5182 | if (!nowrap_type_p (TREE_TYPE (base))) |
| 5183 | return false; |
| 5184 | |
| 5185 | base = expand_simple_operations (base); |
| 5186 | |
| 5187 | if (TREE_CODE (base) == SSA_NAME) |
| 5188 | { |
| 5189 | gimple *stmt = SSA_NAME_DEF_STMT (base); |
| 5190 | |
| 5191 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
| 5192 | return false; |
| 5193 | |
| 5194 | code = gimple_assign_rhs_code (gs: stmt); |
| 5195 | if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS) |
| 5196 | return false; |
| 5197 | |
| 5198 | e1 = gimple_assign_rhs1 (gs: stmt); |
| 5199 | e2 = gimple_assign_rhs2 (gs: stmt); |
| 5200 | } |
| 5201 | else |
| 5202 | { |
| 5203 | code = TREE_CODE (base); |
| 5204 | if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS) |
| 5205 | return false; |
| 5206 | e1 = TREE_OPERAND (base, 0); |
| 5207 | e2 = TREE_OPERAND (base, 1); |
| 5208 | } |
| 5209 | |
| 5210 | /* Use affine expansion as deeper inspection to prove the equality. */ |
| 5211 | tree_to_aff_combination_expand (e2, TREE_TYPE (e2), |
| 5212 | &aff_e2, &data->name_expansion_cache); |
| 5213 | tree_to_aff_combination_expand (offset, TREE_TYPE (offset), |
| 5214 | &aff_offset, &data->name_expansion_cache); |
| 5215 | aff_combination_scale (&aff_offset, -1); |
| 5216 | switch (code) |
| 5217 | { |
| 5218 | case PLUS_EXPR: |
| 5219 | aff_combination_add (&aff_e2, &aff_offset); |
| 5220 | if (aff_combination_zero_p (aff: &aff_e2)) |
| 5221 | return true; |
| 5222 | |
| 5223 | tree_to_aff_combination_expand (e1, TREE_TYPE (e1), |
| 5224 | &aff_e1, &data->name_expansion_cache); |
| 5225 | aff_combination_add (&aff_e1, &aff_offset); |
| 5226 | return aff_combination_zero_p (aff: &aff_e1); |
| 5227 | |
| 5228 | case POINTER_PLUS_EXPR: |
| 5229 | aff_combination_add (&aff_e2, &aff_offset); |
| 5230 | return aff_combination_zero_p (aff: &aff_e2); |
| 5231 | |
| 5232 | default: |
| 5233 | return false; |
| 5234 | } |
| 5235 | } |
| 5236 | |
| 5237 | /* Tries to replace loop exit by one formulated in terms of a LT_EXPR |
| 5238 | comparison with CAND. NITER describes the number of iterations of |
| 5239 | the loops. If successful, the comparison in COMP_P is altered accordingly. |
| 5240 | |
| 5241 | We aim to handle the following situation: |
| 5242 | |
| 5243 | sometype *base, *p; |
| 5244 | int a, b, i; |
| 5245 | |
| 5246 | i = a; |
| 5247 | p = p_0 = base + a; |
| 5248 | |
| 5249 | do |
| 5250 | { |
| 5251 | bla (*p); |
| 5252 | p++; |
| 5253 | i++; |
| 5254 | } |
| 5255 | while (i < b); |
| 5256 | |
| 5257 | Here, the number of iterations of the loop is (a + 1 > b) ? 0 : b - a - 1. |
| 5258 | We aim to optimize this to |
| 5259 | |
| 5260 | p = p_0 = base + a; |
| 5261 | do |
| 5262 | { |
| 5263 | bla (*p); |
| 5264 | p++; |
| 5265 | } |
| 5266 | while (p < p_0 - a + b); |
| 5267 | |
| 5268 | This preserves the correctness, since the pointer arithmetics does not |
| 5269 | overflow. More precisely: |
| 5270 | |
| 5271 | 1) if a + 1 <= b, then p_0 - a + b is the final value of p, hence there is no |
| 5272 | overflow in computing it or the values of p. |
| 5273 | 2) if a + 1 > b, then we need to verify that the expression p_0 - a does not |
| 5274 | overflow. To prove this, we use the fact that p_0 = base + a. */ |
| 5275 | |
| 5276 | static bool |
| 5277 | iv_elimination_compare_lt (struct ivopts_data *data, |
| 5278 | struct iv_cand *cand, enum tree_code *comp_p, |
| 5279 | class tree_niter_desc *niter) |
| 5280 | { |
| 5281 | tree cand_type, a, b, mbz, nit_type = TREE_TYPE (niter->niter), offset; |
| 5282 | class aff_tree nit, tmpa, tmpb; |
| 5283 | enum tree_code comp; |
| 5284 | HOST_WIDE_INT step; |
| 5285 | |
| 5286 | /* We need to know that the candidate induction variable does not overflow. |
| 5287 | While more complex analysis may be used to prove this, for now just |
| 5288 | check that the variable appears in the original program and that it |
| 5289 | is computed in a type that guarantees no overflows. */ |
| 5290 | cand_type = TREE_TYPE (cand->iv->base); |
| 5291 | if (cand->pos != IP_ORIGINAL || !nowrap_type_p (cand_type)) |
| 5292 | return false; |
| 5293 | |
| 5294 | /* Make sure that the loop iterates till the loop bound is hit, as otherwise |
| 5295 | the calculation of the BOUND could overflow, making the comparison |
| 5296 | invalid. */ |
| 5297 | if (!data->loop_single_exit_p) |
| 5298 | return false; |
| 5299 | |
| 5300 | /* We need to be able to decide whether candidate is increasing or decreasing |
| 5301 | in order to choose the right comparison operator. */ |
| 5302 | if (!cst_and_fits_in_hwi (cand->iv->step)) |
| 5303 | return false; |
| 5304 | step = int_cst_value (cand->iv->step); |
| 5305 | |
| 5306 | /* Check that the number of iterations matches the expected pattern: |
| 5307 | a + 1 > b ? 0 : b - a - 1. */ |
| 5308 | mbz = niter->may_be_zero; |
| 5309 | if (TREE_CODE (mbz) == GT_EXPR) |
| 5310 | { |
| 5311 | /* Handle a + 1 > b. */ |
| 5312 | tree op0 = TREE_OPERAND (mbz, 0); |
| 5313 | if (TREE_CODE (op0) == PLUS_EXPR && integer_onep (TREE_OPERAND (op0, 1))) |
| 5314 | { |
| 5315 | a = TREE_OPERAND (op0, 0); |
| 5316 | b = TREE_OPERAND (mbz, 1); |
| 5317 | } |
| 5318 | else |
| 5319 | return false; |
| 5320 | } |
| 5321 | else if (TREE_CODE (mbz) == LT_EXPR) |
| 5322 | { |
| 5323 | tree op1 = TREE_OPERAND (mbz, 1); |
| 5324 | |
| 5325 | /* Handle b < a + 1. */ |
| 5326 | if (TREE_CODE (op1) == PLUS_EXPR && integer_onep (TREE_OPERAND (op1, 1))) |
| 5327 | { |
| 5328 | a = TREE_OPERAND (op1, 0); |
| 5329 | b = TREE_OPERAND (mbz, 0); |
| 5330 | } |
| 5331 | else |
| 5332 | return false; |
| 5333 | } |
| 5334 | else |
| 5335 | return false; |
| 5336 | |
| 5337 | /* Expected number of iterations is B - A - 1. Check that it matches |
| 5338 | the actual number, i.e., that B - A - NITER = 1. */ |
| 5339 | tree_to_aff_combination (niter->niter, nit_type, &nit); |
| 5340 | tree_to_aff_combination (fold_convert (nit_type, a), nit_type, &tmpa); |
| 5341 | tree_to_aff_combination (fold_convert (nit_type, b), nit_type, &tmpb); |
| 5342 | aff_combination_scale (&nit, -1); |
| 5343 | aff_combination_scale (&tmpa, -1); |
| 5344 | aff_combination_add (&tmpb, &tmpa); |
| 5345 | aff_combination_add (&tmpb, &nit); |
| 5346 | if (tmpb.n != 0 || maybe_ne (a: tmpb.offset, b: 1)) |
| 5347 | return false; |
| 5348 | |
| 5349 | /* Finally, check that CAND->IV->BASE - CAND->IV->STEP * A does not |
| 5350 | overflow. */ |
| 5351 | offset = fold_build2 (MULT_EXPR, TREE_TYPE (cand->iv->step), |
| 5352 | cand->iv->step, |
| 5353 | fold_convert (TREE_TYPE (cand->iv->step), a)); |
| 5354 | if (!difference_cannot_overflow_p (data, base: cand->iv->base, offset)) |
| 5355 | return false; |
| 5356 | |
| 5357 | /* Determine the new comparison operator. */ |
| 5358 | comp = step < 0 ? GT_EXPR : LT_EXPR; |
| 5359 | if (*comp_p == NE_EXPR) |
| 5360 | *comp_p = comp; |
| 5361 | else if (*comp_p == EQ_EXPR) |
| 5362 | *comp_p = invert_tree_comparison (comp, false); |
| 5363 | else |
| 5364 | gcc_unreachable (); |
| 5365 | |
| 5366 | return true; |
| 5367 | } |
| 5368 | |
| 5369 | /* Check whether it is possible to express the condition in USE by comparison |
| 5370 | of candidate CAND. If so, store the value compared with to BOUND, and the |
| 5371 | comparison operator to COMP. */ |
| 5372 | |
| 5373 | static bool |
| 5374 | may_eliminate_iv (struct ivopts_data *data, |
| 5375 | struct iv_use *use, struct iv_cand *cand, tree *bound, |
| 5376 | enum tree_code *comp) |
| 5377 | { |
| 5378 | basic_block ex_bb; |
| 5379 | edge exit; |
| 5380 | tree period; |
| 5381 | class loop *loop = data->current_loop; |
| 5382 | aff_tree bnd; |
| 5383 | class tree_niter_desc *desc = NULL; |
| 5384 | |
| 5385 | if (TREE_CODE (cand->iv->step) != INTEGER_CST) |
| 5386 | return false; |
| 5387 | |
| 5388 | /* For now works only for exits that dominate the loop latch. |
| 5389 | TODO: extend to other conditions inside loop body. */ |
| 5390 | ex_bb = gimple_bb (g: use->stmt); |
| 5391 | if (use->stmt != last_nondebug_stmt (ex_bb) |
| 5392 | || gimple_code (g: use->stmt) != GIMPLE_COND |
| 5393 | || !dominated_by_p (CDI_DOMINATORS, loop->latch, ex_bb)) |
| 5394 | return false; |
| 5395 | |
| 5396 | exit = EDGE_SUCC (ex_bb, 0); |
| 5397 | if (flow_bb_inside_loop_p (loop, exit->dest)) |
| 5398 | exit = EDGE_SUCC (ex_bb, 1); |
| 5399 | if (flow_bb_inside_loop_p (loop, exit->dest)) |
| 5400 | return false; |
| 5401 | |
| 5402 | desc = niter_for_exit (data, exit); |
| 5403 | if (!desc) |
| 5404 | return false; |
| 5405 | |
| 5406 | /* Determine whether we can use the variable to test the exit condition. |
| 5407 | This is the case iff the period of the induction variable is greater |
| 5408 | than the number of iterations for which the exit condition is true. */ |
| 5409 | period = iv_period (iv: cand->iv); |
| 5410 | |
| 5411 | /* If the number of iterations is constant, compare against it directly. */ |
| 5412 | if (TREE_CODE (desc->niter) == INTEGER_CST) |
| 5413 | { |
| 5414 | /* See cand_value_at. */ |
| 5415 | if (stmt_after_increment (loop, cand, stmt: use->stmt)) |
| 5416 | { |
| 5417 | if (!tree_int_cst_lt (t1: desc->niter, t2: period)) |
| 5418 | return false; |
| 5419 | } |
| 5420 | else |
| 5421 | { |
| 5422 | if (tree_int_cst_lt (t1: period, t2: desc->niter)) |
| 5423 | return false; |
| 5424 | } |
| 5425 | } |
| 5426 | |
| 5427 | /* If not, and if this is the only possible exit of the loop, see whether |
| 5428 | we can get a conservative estimate on the number of iterations of the |
| 5429 | entire loop and compare against that instead. */ |
| 5430 | else |
| 5431 | { |
| 5432 | widest_int period_value, max_niter; |
| 5433 | |
| 5434 | max_niter = desc->max; |
| 5435 | if (stmt_after_increment (loop, cand, stmt: use->stmt)) |
| 5436 | max_niter += 1; |
| 5437 | period_value = wi::to_widest (t: period); |
| 5438 | if (wi::gtu_p (x: max_niter, y: period_value)) |
| 5439 | { |
| 5440 | /* See if we can take advantage of inferred loop bound |
| 5441 | information. */ |
| 5442 | if (data->loop_single_exit_p) |
| 5443 | { |
| 5444 | if (!max_loop_iterations (loop, &max_niter)) |
| 5445 | return false; |
| 5446 | /* The loop bound is already adjusted by adding 1. */ |
| 5447 | if (wi::gtu_p (x: max_niter, y: period_value)) |
| 5448 | return false; |
| 5449 | } |
| 5450 | else |
| 5451 | return false; |
| 5452 | } |
| 5453 | } |
| 5454 | |
| 5455 | /* For doloop IV cand, the bound would be zero. It's safe whether |
| 5456 | may_be_zero set or not. */ |
| 5457 | if (cand->doloop_p) |
| 5458 | { |
| 5459 | *bound = build_int_cst (TREE_TYPE (cand->iv->base), 0); |
| 5460 | *comp = iv_elimination_compare (data, use); |
| 5461 | return true; |
| 5462 | } |
| 5463 | |
| 5464 | cand_value_at (loop, cand, at: use->stmt, desc, val: &bnd); |
| 5465 | |
| 5466 | *bound = fold_convert (TREE_TYPE (cand->iv->base), |
| 5467 | aff_combination_to_tree (&bnd)); |
| 5468 | *comp = iv_elimination_compare (data, use); |
| 5469 | |
| 5470 | /* It is unlikely that computing the number of iterations using division |
| 5471 | would be more profitable than keeping the original induction variable. */ |
| 5472 | bool cond_overflow_p; |
| 5473 | if (expression_expensive_p (*bound, &cond_overflow_p)) |
| 5474 | return false; |
| 5475 | |
| 5476 | /* Sometimes, it is possible to handle the situation that the number of |
| 5477 | iterations may be zero unless additional assumptions by using < |
| 5478 | instead of != in the exit condition. |
| 5479 | |
| 5480 | TODO: we could also calculate the value MAY_BE_ZERO ? 0 : NITER and |
| 5481 | base the exit condition on it. However, that is often too |
| 5482 | expensive. */ |
| 5483 | if (!integer_zerop (desc->may_be_zero)) |
| 5484 | return iv_elimination_compare_lt (data, cand, comp_p: comp, niter: desc); |
| 5485 | |
| 5486 | return true; |
| 5487 | } |
| 5488 | |
| 5489 | /* Calculates the cost of BOUND, if it is a PARM_DECL. A PARM_DECL must |
| 5490 | be copied, if it is used in the loop body and DATA->body_includes_call. */ |
| 5491 | |
| 5492 | static int |
| 5493 | parm_decl_cost (struct ivopts_data *data, tree bound) |
| 5494 | { |
| 5495 | tree sbound = bound; |
| 5496 | STRIP_NOPS (sbound); |
| 5497 | |
| 5498 | if (TREE_CODE (sbound) == SSA_NAME |
| 5499 | && SSA_NAME_IS_DEFAULT_DEF (sbound) |
| 5500 | && TREE_CODE (SSA_NAME_VAR (sbound)) == PARM_DECL |
| 5501 | && data->body_includes_call) |
| 5502 | return COSTS_N_INSNS (1); |
| 5503 | |
| 5504 | return 0; |
| 5505 | } |
| 5506 | |
| 5507 | /* Determines cost of computing the use in GROUP with CAND in a condition. */ |
| 5508 | |
| 5509 | static bool |
| 5510 | determine_group_iv_cost_cond (struct ivopts_data *data, |
| 5511 | struct iv_group *group, struct iv_cand *cand) |
| 5512 | { |
| 5513 | tree bound = NULL_TREE; |
| 5514 | struct iv *cmp_iv; |
| 5515 | bitmap inv_exprs = NULL; |
| 5516 | bitmap inv_vars_elim = NULL, inv_vars_express = NULL, inv_vars; |
| 5517 | comp_cost elim_cost = infinite_cost, express_cost, cost, bound_cost; |
| 5518 | enum comp_iv_rewrite rewrite_type; |
| 5519 | iv_inv_expr_ent *inv_expr_elim = NULL, *inv_expr_express = NULL, *inv_expr; |
| 5520 | tree *control_var, *bound_cst; |
| 5521 | enum tree_code comp = ERROR_MARK; |
| 5522 | struct iv_use *use = group->vuses[0]; |
| 5523 | |
| 5524 | /* Extract condition operands. */ |
| 5525 | rewrite_type = extract_cond_operands (data, stmt: use->stmt, control_var: &control_var, |
| 5526 | bound: &bound_cst, NULL, iv_bound: &cmp_iv); |
| 5527 | gcc_assert (rewrite_type != COMP_IV_NA); |
| 5528 | |
| 5529 | /* Try iv elimination. */ |
| 5530 | if (rewrite_type == COMP_IV_ELIM |
| 5531 | && may_eliminate_iv (data, use, cand, bound: &bound, comp: &comp)) |
| 5532 | { |
| 5533 | elim_cost = force_var_cost (data, expr: bound, inv_vars: &inv_vars_elim); |
| 5534 | if (elim_cost.cost == 0) |
| 5535 | elim_cost.cost = parm_decl_cost (data, bound); |
| 5536 | else if (TREE_CODE (bound) == INTEGER_CST) |
| 5537 | elim_cost.cost = 0; |
| 5538 | /* If we replace a loop condition 'i < n' with 'p < base + n', |
| 5539 | inv_vars_elim will have 'base' and 'n' set, which implies that both |
| 5540 | 'base' and 'n' will be live during the loop. More likely, |
| 5541 | 'base + n' will be loop invariant, resulting in only one live value |
| 5542 | during the loop. So in that case we clear inv_vars_elim and set |
| 5543 | inv_expr_elim instead. */ |
| 5544 | if (inv_vars_elim && bitmap_count_bits (inv_vars_elim) > 1) |
| 5545 | { |
| 5546 | inv_expr_elim = get_loop_invariant_expr (data, inv_expr: bound); |
| 5547 | bitmap_clear (inv_vars_elim); |
| 5548 | } |
| 5549 | /* The bound is a loop invariant, so it will be only computed |
| 5550 | once. */ |
| 5551 | elim_cost.cost = adjust_setup_cost (data, cost: elim_cost.cost); |
| 5552 | } |
| 5553 | |
| 5554 | /* When the condition is a comparison of the candidate IV against |
| 5555 | zero, prefer this IV. |
| 5556 | |
| 5557 | TODO: The constant that we're subtracting from the cost should |
| 5558 | be target-dependent. This information should be added to the |
| 5559 | target costs for each backend. */ |
| 5560 | if (!elim_cost.infinite_cost_p () /* Do not try to decrease infinite! */ |
| 5561 | && integer_zerop (*bound_cst) |
| 5562 | && (operand_equal_p (*control_var, cand->var_after, flags: 0) |
| 5563 | || operand_equal_p (*control_var, cand->var_before, flags: 0))) |
| 5564 | elim_cost -= 1; |
| 5565 | |
| 5566 | express_cost = get_computation_cost (data, use, cand, address_p: false, |
| 5567 | inv_vars: &inv_vars_express, NULL, |
| 5568 | inv_expr: &inv_expr_express); |
| 5569 | if (cmp_iv != NULL) |
| 5570 | find_inv_vars (data, expr_p: &cmp_iv->base, inv_vars: &inv_vars_express); |
| 5571 | |
| 5572 | /* Count the cost of the original bound as well. */ |
| 5573 | bound_cost = force_var_cost (data, expr: *bound_cst, NULL); |
| 5574 | if (bound_cost.cost == 0) |
| 5575 | bound_cost.cost = parm_decl_cost (data, bound: *bound_cst); |
| 5576 | else if (TREE_CODE (*bound_cst) == INTEGER_CST) |
| 5577 | bound_cost.cost = 0; |
| 5578 | express_cost += bound_cost; |
| 5579 | |
| 5580 | /* Choose the better approach, preferring the eliminated IV. */ |
| 5581 | if (elim_cost <= express_cost) |
| 5582 | { |
| 5583 | cost = elim_cost; |
| 5584 | inv_vars = inv_vars_elim; |
| 5585 | inv_vars_elim = NULL; |
| 5586 | inv_expr = inv_expr_elim; |
| 5587 | /* For doloop candidate/use pair, adjust to zero cost. */ |
| 5588 | if (group->doloop_p && cand->doloop_p && elim_cost.cost > no_cost.cost) |
| 5589 | cost = no_cost; |
| 5590 | } |
| 5591 | else |
| 5592 | { |
| 5593 | cost = express_cost; |
| 5594 | inv_vars = inv_vars_express; |
| 5595 | inv_vars_express = NULL; |
| 5596 | bound = NULL_TREE; |
| 5597 | comp = ERROR_MARK; |
| 5598 | inv_expr = inv_expr_express; |
| 5599 | } |
| 5600 | |
| 5601 | if (inv_expr) |
| 5602 | { |
| 5603 | inv_exprs = BITMAP_ALLOC (NULL); |
| 5604 | bitmap_set_bit (inv_exprs, inv_expr->id); |
| 5605 | } |
| 5606 | set_group_iv_cost (data, group, cand, cost, |
| 5607 | inv_vars, value: bound, comp, inv_exprs); |
| 5608 | |
| 5609 | if (inv_vars_elim) |
| 5610 | BITMAP_FREE (inv_vars_elim); |
| 5611 | if (inv_vars_express) |
| 5612 | BITMAP_FREE (inv_vars_express); |
| 5613 | |
| 5614 | return !cost.infinite_cost_p (); |
| 5615 | } |
| 5616 | |
| 5617 | /* Determines cost of computing uses in GROUP with CAND. Returns false |
| 5618 | if USE cannot be represented with CAND. */ |
| 5619 | |
| 5620 | static bool |
| 5621 | determine_group_iv_cost (struct ivopts_data *data, |
| 5622 | struct iv_group *group, struct iv_cand *cand) |
| 5623 | { |
| 5624 | switch (group->type) |
| 5625 | { |
| 5626 | case USE_NONLINEAR_EXPR: |
| 5627 | return determine_group_iv_cost_generic (data, group, cand); |
| 5628 | |
| 5629 | case USE_REF_ADDRESS: |
| 5630 | case USE_PTR_ADDRESS: |
| 5631 | return determine_group_iv_cost_address (data, group, cand); |
| 5632 | |
| 5633 | case USE_COMPARE: |
| 5634 | return determine_group_iv_cost_cond (data, group, cand); |
| 5635 | |
| 5636 | default: |
| 5637 | gcc_unreachable (); |
| 5638 | } |
| 5639 | } |
| 5640 | |
| 5641 | /* Return true if get_computation_cost indicates that autoincrement is |
| 5642 | a possibility for the pair of USE and CAND, false otherwise. */ |
| 5643 | |
| 5644 | static bool |
| 5645 | autoinc_possible_for_pair (struct ivopts_data *data, struct iv_use *use, |
| 5646 | struct iv_cand *cand) |
| 5647 | { |
| 5648 | if (!address_p (type: use->type)) |
| 5649 | return false; |
| 5650 | |
| 5651 | bool can_autoinc = false; |
| 5652 | get_computation_cost (data, use, cand, address_p: true, NULL, can_autoinc: &can_autoinc, NULL); |
| 5653 | return can_autoinc; |
| 5654 | } |
| 5655 | |
| 5656 | /* Examine IP_ORIGINAL candidates to see if they are incremented next to a |
| 5657 | use that allows autoincrement, and set their AINC_USE if possible. */ |
| 5658 | |
| 5659 | static void |
| 5660 | set_autoinc_for_original_candidates (struct ivopts_data *data) |
| 5661 | { |
| 5662 | unsigned i, j; |
| 5663 | |
| 5664 | for (i = 0; i < data->vcands.length (); i++) |
| 5665 | { |
| 5666 | struct iv_cand *cand = data->vcands[i]; |
| 5667 | struct iv_use *closest_before = NULL; |
| 5668 | struct iv_use *closest_after = NULL; |
| 5669 | if (cand->pos != IP_ORIGINAL) |
| 5670 | continue; |
| 5671 | |
| 5672 | for (j = 0; j < data->vgroups.length (); j++) |
| 5673 | { |
| 5674 | struct iv_group *group = data->vgroups[j]; |
| 5675 | struct iv_use *use = group->vuses[0]; |
| 5676 | unsigned uid = gimple_uid (g: use->stmt); |
| 5677 | |
| 5678 | if (gimple_bb (g: use->stmt) != gimple_bb (g: cand->incremented_at)) |
| 5679 | continue; |
| 5680 | |
| 5681 | if (uid < gimple_uid (g: cand->incremented_at) |
| 5682 | && (closest_before == NULL |
| 5683 | || uid > gimple_uid (g: closest_before->stmt))) |
| 5684 | closest_before = use; |
| 5685 | |
| 5686 | if (uid > gimple_uid (g: cand->incremented_at) |
| 5687 | && (closest_after == NULL |
| 5688 | || uid < gimple_uid (g: closest_after->stmt))) |
| 5689 | closest_after = use; |
| 5690 | } |
| 5691 | |
| 5692 | if (closest_before != NULL |
| 5693 | && autoinc_possible_for_pair (data, use: closest_before, cand)) |
| 5694 | cand->ainc_use = closest_before; |
| 5695 | else if (closest_after != NULL |
| 5696 | && autoinc_possible_for_pair (data, use: closest_after, cand)) |
| 5697 | cand->ainc_use = closest_after; |
| 5698 | } |
| 5699 | } |
| 5700 | |
| 5701 | /* Relate compare use with all candidates. */ |
| 5702 | |
| 5703 | static void |
| 5704 | relate_compare_use_with_all_cands (struct ivopts_data *data) |
| 5705 | { |
| 5706 | unsigned i, count = data->vcands.length (); |
| 5707 | for (i = 0; i < data->vgroups.length (); i++) |
| 5708 | { |
| 5709 | struct iv_group *group = data->vgroups[i]; |
| 5710 | |
| 5711 | if (group->type == USE_COMPARE) |
| 5712 | bitmap_set_range (group->related_cands, 0, count); |
| 5713 | } |
| 5714 | } |
| 5715 | |
| 5716 | /* If PREFERRED_MODE is suitable and profitable, use the preferred |
| 5717 | PREFERRED_MODE to compute doloop iv base from niter: base = niter + 1. */ |
| 5718 | |
| 5719 | static tree |
| 5720 | compute_doloop_base_on_mode (machine_mode preferred_mode, tree niter, |
| 5721 | const widest_int &iterations_max) |
| 5722 | { |
| 5723 | tree ntype = TREE_TYPE (niter); |
| 5724 | tree pref_type = lang_hooks.types.type_for_mode (preferred_mode, 1); |
| 5725 | if (!pref_type) |
| 5726 | return fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), |
| 5727 | build_int_cst (ntype, 1)); |
| 5728 | |
| 5729 | gcc_assert (TREE_CODE (pref_type) == INTEGER_TYPE); |
| 5730 | |
| 5731 | int prec = TYPE_PRECISION (ntype); |
| 5732 | int pref_prec = TYPE_PRECISION (pref_type); |
| 5733 | |
| 5734 | tree base; |
| 5735 | |
| 5736 | /* Check if the PREFERRED_MODED is able to present niter. */ |
| 5737 | if (pref_prec > prec |
| 5738 | || wi::ltu_p (x: iterations_max, |
| 5739 | y: widest_int::from (x: wi::max_value (pref_prec, UNSIGNED), |
| 5740 | sgn: UNSIGNED))) |
| 5741 | { |
| 5742 | /* No wrap, it is safe to use preferred type after niter + 1. */ |
| 5743 | if (wi::ltu_p (x: iterations_max, |
| 5744 | y: widest_int::from (x: wi::max_value (prec, UNSIGNED), |
| 5745 | sgn: UNSIGNED))) |
| 5746 | { |
| 5747 | /* This could help to optimize "-1 +1" pair when niter looks |
| 5748 | like "n-1": n is in original mode. "base = (n - 1) + 1" |
| 5749 | in PREFERRED_MODED: it could be base = (PREFERRED_TYPE)n. */ |
| 5750 | base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), |
| 5751 | build_int_cst (ntype, 1)); |
| 5752 | base = fold_convert (pref_type, base); |
| 5753 | } |
| 5754 | |
| 5755 | /* To avoid wrap, convert niter to preferred type before plus 1. */ |
| 5756 | else |
| 5757 | { |
| 5758 | niter = fold_convert (pref_type, niter); |
| 5759 | base = fold_build2 (PLUS_EXPR, pref_type, unshare_expr (niter), |
| 5760 | build_int_cst (pref_type, 1)); |
| 5761 | } |
| 5762 | } |
| 5763 | else |
| 5764 | base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), |
| 5765 | build_int_cst (ntype, 1)); |
| 5766 | return base; |
| 5767 | } |
| 5768 | |
| 5769 | /* Add one doloop dedicated IV candidate: |
| 5770 | - Base is (may_be_zero ? 1 : (niter + 1)). |
| 5771 | - Step is -1. */ |
| 5772 | |
| 5773 | static void |
| 5774 | add_iv_candidate_for_doloop (struct ivopts_data *data) |
| 5775 | { |
| 5776 | tree_niter_desc *niter_desc = niter_for_single_dom_exit (data); |
| 5777 | gcc_assert (niter_desc && niter_desc->assumptions); |
| 5778 | |
| 5779 | tree niter = niter_desc->niter; |
| 5780 | tree ntype = TREE_TYPE (niter); |
| 5781 | gcc_assert (TREE_CODE (ntype) == INTEGER_TYPE); |
| 5782 | |
| 5783 | tree may_be_zero = niter_desc->may_be_zero; |
| 5784 | if (may_be_zero && integer_zerop (may_be_zero)) |
| 5785 | may_be_zero = NULL_TREE; |
| 5786 | if (may_be_zero) |
| 5787 | { |
| 5788 | if (COMPARISON_CLASS_P (may_be_zero)) |
| 5789 | { |
| 5790 | niter = fold_build3 (COND_EXPR, ntype, may_be_zero, |
| 5791 | build_int_cst (ntype, 0), |
| 5792 | rewrite_to_non_trapping_overflow (niter)); |
| 5793 | } |
| 5794 | /* Don't try to obtain the iteration count expression when may_be_zero is |
| 5795 | integer_nonzerop (actually iteration count is one) or else. */ |
| 5796 | else |
| 5797 | return; |
| 5798 | } |
| 5799 | |
| 5800 | machine_mode mode = TYPE_MODE (ntype); |
| 5801 | machine_mode pref_mode = targetm.preferred_doloop_mode (mode); |
| 5802 | |
| 5803 | tree base; |
| 5804 | if (mode != pref_mode) |
| 5805 | { |
| 5806 | base = compute_doloop_base_on_mode (preferred_mode: pref_mode, niter, iterations_max: niter_desc->max); |
| 5807 | ntype = TREE_TYPE (base); |
| 5808 | } |
| 5809 | else |
| 5810 | base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter), |
| 5811 | build_int_cst (ntype, 1)); |
| 5812 | |
| 5813 | |
| 5814 | add_candidate (data, base, step: build_int_cst (ntype, -1), important: true, NULL, NULL, doloop: true); |
| 5815 | } |
| 5816 | |
| 5817 | /* Finds the candidates for the induction variables. */ |
| 5818 | |
| 5819 | static void |
| 5820 | find_iv_candidates (struct ivopts_data *data) |
| 5821 | { |
| 5822 | /* Add commonly used ivs. */ |
| 5823 | add_standard_iv_candidates (data); |
| 5824 | |
| 5825 | /* Add doloop dedicated ivs. */ |
| 5826 | if (data->doloop_use_p) |
| 5827 | add_iv_candidate_for_doloop (data); |
| 5828 | |
| 5829 | /* Add old induction variables. */ |
| 5830 | add_iv_candidate_for_bivs (data); |
| 5831 | |
| 5832 | /* Add induction variables derived from uses. */ |
| 5833 | add_iv_candidate_for_groups (data); |
| 5834 | |
| 5835 | set_autoinc_for_original_candidates (data); |
| 5836 | |
| 5837 | /* Record the important candidates. */ |
| 5838 | record_important_candidates (data); |
| 5839 | |
| 5840 | /* Relate compare iv_use with all candidates. */ |
| 5841 | if (!data->consider_all_candidates) |
| 5842 | relate_compare_use_with_all_cands (data); |
| 5843 | |
| 5844 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 5845 | { |
| 5846 | unsigned i; |
| 5847 | |
| 5848 | fprintf (stream: dump_file, format: "\n<Important Candidates>:\t" ); |
| 5849 | for (i = 0; i < data->vcands.length (); i++) |
| 5850 | if (data->vcands[i]->important) |
| 5851 | fprintf (stream: dump_file, format: " %d," , data->vcands[i]->id); |
| 5852 | fprintf (stream: dump_file, format: "\n" ); |
| 5853 | |
| 5854 | fprintf (stream: dump_file, format: "\n<Group, Cand> Related:\n" ); |
| 5855 | for (i = 0; i < data->vgroups.length (); i++) |
| 5856 | { |
| 5857 | struct iv_group *group = data->vgroups[i]; |
| 5858 | |
| 5859 | if (group->related_cands) |
| 5860 | { |
| 5861 | fprintf (stream: dump_file, format: " Group %d:\t" , group->id); |
| 5862 | dump_bitmap (file: dump_file, map: group->related_cands); |
| 5863 | } |
| 5864 | } |
| 5865 | fprintf (stream: dump_file, format: "\n" ); |
| 5866 | } |
| 5867 | } |
| 5868 | |
| 5869 | /* Determines costs of computing use of iv with an iv candidate. */ |
| 5870 | |
| 5871 | static void |
| 5872 | determine_group_iv_costs (struct ivopts_data *data) |
| 5873 | { |
| 5874 | unsigned i, j; |
| 5875 | struct iv_cand *cand; |
| 5876 | struct iv_group *group; |
| 5877 | bitmap to_clear = BITMAP_ALLOC (NULL); |
| 5878 | |
| 5879 | alloc_use_cost_map (data); |
| 5880 | |
| 5881 | for (i = 0; i < data->vgroups.length (); i++) |
| 5882 | { |
| 5883 | group = data->vgroups[i]; |
| 5884 | |
| 5885 | if (data->consider_all_candidates) |
| 5886 | { |
| 5887 | for (j = 0; j < data->vcands.length (); j++) |
| 5888 | { |
| 5889 | cand = data->vcands[j]; |
| 5890 | determine_group_iv_cost (data, group, cand); |
| 5891 | } |
| 5892 | } |
| 5893 | else |
| 5894 | { |
| 5895 | bitmap_iterator bi; |
| 5896 | |
| 5897 | EXECUTE_IF_SET_IN_BITMAP (group->related_cands, 0, j, bi) |
| 5898 | { |
| 5899 | cand = data->vcands[j]; |
| 5900 | if (!determine_group_iv_cost (data, group, cand)) |
| 5901 | bitmap_set_bit (to_clear, j); |
| 5902 | } |
| 5903 | |
| 5904 | /* Remove the candidates for that the cost is infinite from |
| 5905 | the list of related candidates. */ |
| 5906 | bitmap_and_compl_into (group->related_cands, to_clear); |
| 5907 | bitmap_clear (to_clear); |
| 5908 | } |
| 5909 | } |
| 5910 | |
| 5911 | BITMAP_FREE (to_clear); |
| 5912 | |
| 5913 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 5914 | { |
| 5915 | bitmap_iterator bi; |
| 5916 | |
| 5917 | /* Dump invariant variables. */ |
| 5918 | fprintf (stream: dump_file, format: "\n<Invariant Vars>:\n" ); |
| 5919 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
| 5920 | { |
| 5921 | struct version_info *info = ver_info (data, ver: i); |
| 5922 | if (info->inv_id) |
| 5923 | { |
| 5924 | fprintf (stream: dump_file, format: "Inv %d:\t" , info->inv_id); |
| 5925 | print_generic_expr (dump_file, info->name, TDF_SLIM); |
| 5926 | fprintf (stream: dump_file, format: "%s\n" , |
| 5927 | info->has_nonlin_use ? "" : "\t(eliminable)" ); |
| 5928 | } |
| 5929 | } |
| 5930 | |
| 5931 | /* Dump invariant expressions. */ |
| 5932 | fprintf (stream: dump_file, format: "\n<Invariant Expressions>:\n" ); |
| 5933 | auto_vec <iv_inv_expr_ent *> list (data->inv_expr_tab->elements ()); |
| 5934 | |
| 5935 | for (hash_table<iv_inv_expr_hasher>::iterator it |
| 5936 | = data->inv_expr_tab->begin (); it != data->inv_expr_tab->end (); |
| 5937 | ++it) |
| 5938 | list.safe_push (obj: *it); |
| 5939 | |
| 5940 | list.qsort (sort_iv_inv_expr_ent); |
| 5941 | |
| 5942 | for (i = 0; i < list.length (); ++i) |
| 5943 | { |
| 5944 | fprintf (stream: dump_file, format: "inv_expr %d: \t" , list[i]->id); |
| 5945 | print_generic_expr (dump_file, list[i]->expr, TDF_SLIM); |
| 5946 | fprintf (stream: dump_file, format: "\n" ); |
| 5947 | } |
| 5948 | |
| 5949 | fprintf (stream: dump_file, format: "\n<Group-candidate Costs>:\n" ); |
| 5950 | |
| 5951 | for (i = 0; i < data->vgroups.length (); i++) |
| 5952 | { |
| 5953 | group = data->vgroups[i]; |
| 5954 | |
| 5955 | fprintf (stream: dump_file, format: "Group %d:\n" , i); |
| 5956 | fprintf (stream: dump_file, format: " cand\tcost\tcompl.\tinv.expr.\tinv.vars\n" ); |
| 5957 | for (j = 0; j < group->n_map_members; j++) |
| 5958 | { |
| 5959 | if (!group->cost_map[j].cand |
| 5960 | || group->cost_map[j].cost.infinite_cost_p ()) |
| 5961 | continue; |
| 5962 | |
| 5963 | fprintf (stream: dump_file, format: " %d\t%" PRId64 "\t%d\t" , |
| 5964 | group->cost_map[j].cand->id, |
| 5965 | group->cost_map[j].cost.cost, |
| 5966 | group->cost_map[j].cost.complexity); |
| 5967 | if (!group->cost_map[j].inv_exprs |
| 5968 | || bitmap_empty_p (map: group->cost_map[j].inv_exprs)) |
| 5969 | fprintf (stream: dump_file, format: "NIL;\t" ); |
| 5970 | else |
| 5971 | bitmap_print (dump_file, |
| 5972 | group->cost_map[j].inv_exprs, "" , ";\t" ); |
| 5973 | if (!group->cost_map[j].inv_vars |
| 5974 | || bitmap_empty_p (map: group->cost_map[j].inv_vars)) |
| 5975 | fprintf (stream: dump_file, format: "NIL;\n" ); |
| 5976 | else |
| 5977 | bitmap_print (dump_file, |
| 5978 | group->cost_map[j].inv_vars, "" , "\n" ); |
| 5979 | } |
| 5980 | |
| 5981 | fprintf (stream: dump_file, format: "\n" ); |
| 5982 | } |
| 5983 | fprintf (stream: dump_file, format: "\n" ); |
| 5984 | } |
| 5985 | } |
| 5986 | |
| 5987 | /* Determines cost of the candidate CAND. */ |
| 5988 | |
| 5989 | static void |
| 5990 | determine_iv_cost (struct ivopts_data *data, struct iv_cand *cand) |
| 5991 | { |
| 5992 | comp_cost cost_base; |
| 5993 | int64_t cost, cost_step; |
| 5994 | tree base; |
| 5995 | |
| 5996 | gcc_assert (cand->iv != NULL); |
| 5997 | |
| 5998 | /* There are two costs associated with the candidate -- its increment |
| 5999 | and its initialization. The second is almost negligible for any loop |
| 6000 | that rolls enough, so we take it just very little into account. */ |
| 6001 | |
| 6002 | base = cand->iv->base; |
| 6003 | cost_base = force_var_cost (data, expr: base, NULL); |
| 6004 | /* It will be exceptional that the iv register happens to be initialized with |
| 6005 | the proper value at no cost. In general, there will at least be a regcopy |
| 6006 | or a const set. */ |
| 6007 | if (cost_base.cost == 0) |
| 6008 | cost_base.cost = COSTS_N_INSNS (1); |
| 6009 | /* Doloop decrement should be considered as zero cost. */ |
| 6010 | if (cand->doloop_p) |
| 6011 | cost_step = 0; |
| 6012 | else |
| 6013 | cost_step = add_cost (speed: data->speed, TYPE_MODE (TREE_TYPE (base))); |
| 6014 | cost = cost_step + adjust_setup_cost (data, cost: cost_base.cost); |
| 6015 | |
| 6016 | /* Prefer the original ivs unless we may gain something by replacing it. |
| 6017 | The reason is to make debugging simpler; so this is not relevant for |
| 6018 | artificial ivs created by other optimization passes. */ |
| 6019 | if ((cand->pos != IP_ORIGINAL |
| 6020 | || !SSA_NAME_VAR (cand->var_before) |
| 6021 | || DECL_ARTIFICIAL (SSA_NAME_VAR (cand->var_before))) |
| 6022 | /* Prefer doloop as well. */ |
| 6023 | && !cand->doloop_p) |
| 6024 | cost++; |
| 6025 | |
| 6026 | /* Prefer not to insert statements into latch unless there are some |
| 6027 | already (so that we do not create unnecessary jumps). */ |
| 6028 | if (cand->pos == IP_END |
| 6029 | && empty_block_p (ip_end_pos (data->current_loop))) |
| 6030 | cost++; |
| 6031 | |
| 6032 | cand->cost = cost; |
| 6033 | cand->cost_step = cost_step; |
| 6034 | } |
| 6035 | |
| 6036 | /* Determines costs of computation of the candidates. */ |
| 6037 | |
| 6038 | static void |
| 6039 | determine_iv_costs (struct ivopts_data *data) |
| 6040 | { |
| 6041 | unsigned i; |
| 6042 | |
| 6043 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 6044 | { |
| 6045 | fprintf (stream: dump_file, format: "<Candidate Costs>:\n" ); |
| 6046 | fprintf (stream: dump_file, format: " cand\tcost\n" ); |
| 6047 | } |
| 6048 | |
| 6049 | for (i = 0; i < data->vcands.length (); i++) |
| 6050 | { |
| 6051 | struct iv_cand *cand = data->vcands[i]; |
| 6052 | |
| 6053 | determine_iv_cost (data, cand); |
| 6054 | |
| 6055 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 6056 | fprintf (stream: dump_file, format: " %d\t%d\n" , i, cand->cost); |
| 6057 | } |
| 6058 | |
| 6059 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 6060 | fprintf (stream: dump_file, format: "\n" ); |
| 6061 | } |
| 6062 | |
| 6063 | /* Estimate register pressure for loop having N_INVS invariants and N_CANDS |
| 6064 | induction variables. Note N_INVS includes both invariant variables and |
| 6065 | invariant expressions. */ |
| 6066 | |
| 6067 | static unsigned |
| 6068 | ivopts_estimate_reg_pressure (struct ivopts_data *data, unsigned n_invs, |
| 6069 | unsigned n_cands) |
| 6070 | { |
| 6071 | unsigned cost; |
| 6072 | unsigned n_old = data->regs_used, n_new = n_invs + n_cands; |
| 6073 | unsigned regs_needed = n_new + n_old, available_regs = target_avail_regs; |
| 6074 | bool speed = data->speed; |
| 6075 | |
| 6076 | /* If there is a call in the loop body, the call-clobbered registers |
| 6077 | are not available for loop invariants. */ |
| 6078 | if (data->body_includes_call) |
| 6079 | available_regs = available_regs - target_clobbered_regs; |
| 6080 | |
| 6081 | /* If we have enough registers. */ |
| 6082 | if (regs_needed + target_res_regs < available_regs) |
| 6083 | cost = n_new; |
| 6084 | /* If close to running out of registers, try to preserve them. */ |
| 6085 | else if (regs_needed <= available_regs) |
| 6086 | cost = target_reg_cost [speed] * regs_needed; |
| 6087 | /* If we run out of available registers but the number of candidates |
| 6088 | does not, we penalize extra registers using target_spill_cost. */ |
| 6089 | else if (n_cands <= available_regs) |
| 6090 | cost = target_reg_cost [speed] * available_regs |
| 6091 | + target_spill_cost [speed] * (regs_needed - available_regs); |
| 6092 | /* If the number of candidates runs out available registers, we penalize |
| 6093 | extra candidate registers using target_spill_cost * 2. Because it is |
| 6094 | more expensive to spill induction variable than invariant. */ |
| 6095 | else |
| 6096 | cost = target_reg_cost [speed] * available_regs |
| 6097 | + target_spill_cost [speed] * (n_cands - available_regs) * 2 |
| 6098 | + target_spill_cost [speed] * (regs_needed - n_cands); |
| 6099 | |
| 6100 | /* Finally, add the number of candidates, so that we prefer eliminating |
| 6101 | induction variables if possible. */ |
| 6102 | return cost + n_cands; |
| 6103 | } |
| 6104 | |
| 6105 | /* For each size of the induction variable set determine the penalty. */ |
| 6106 | |
| 6107 | static void |
| 6108 | determine_set_costs (struct ivopts_data *data) |
| 6109 | { |
| 6110 | unsigned j, n; |
| 6111 | gphi *phi; |
| 6112 | gphi_iterator psi; |
| 6113 | tree op; |
| 6114 | class loop *loop = data->current_loop; |
| 6115 | bitmap_iterator bi; |
| 6116 | |
| 6117 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 6118 | { |
| 6119 | fprintf (stream: dump_file, format: "<Global Costs>:\n" ); |
| 6120 | fprintf (stream: dump_file, format: " target_avail_regs %d\n" , target_avail_regs); |
| 6121 | fprintf (stream: dump_file, format: " target_clobbered_regs %d\n" , target_clobbered_regs); |
| 6122 | fprintf (stream: dump_file, format: " target_reg_cost %d\n" , target_reg_cost[data->speed]); |
| 6123 | fprintf (stream: dump_file, format: " target_spill_cost %d\n" , target_spill_cost[data->speed]); |
| 6124 | } |
| 6125 | |
| 6126 | n = 0; |
| 6127 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (i: psi); gsi_next (i: &psi)) |
| 6128 | { |
| 6129 | phi = psi.phi (); |
| 6130 | op = PHI_RESULT (phi); |
| 6131 | |
| 6132 | if (virtual_operand_p (op)) |
| 6133 | continue; |
| 6134 | |
| 6135 | if (get_iv (data, var: op)) |
| 6136 | continue; |
| 6137 | |
| 6138 | if (!POINTER_TYPE_P (TREE_TYPE (op)) |
| 6139 | && !INTEGRAL_TYPE_P (TREE_TYPE (op))) |
| 6140 | continue; |
| 6141 | |
| 6142 | n++; |
| 6143 | } |
| 6144 | |
| 6145 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, j, bi) |
| 6146 | { |
| 6147 | struct version_info *info = ver_info (data, ver: j); |
| 6148 | |
| 6149 | if (info->inv_id && info->has_nonlin_use) |
| 6150 | n++; |
| 6151 | } |
| 6152 | |
| 6153 | data->regs_used = n; |
| 6154 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 6155 | fprintf (stream: dump_file, format: " regs_used %d\n" , n); |
| 6156 | |
| 6157 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 6158 | { |
| 6159 | fprintf (stream: dump_file, format: " cost for size:\n" ); |
| 6160 | fprintf (stream: dump_file, format: " ivs\tcost\n" ); |
| 6161 | for (j = 0; j <= 2 * target_avail_regs; j++) |
| 6162 | fprintf (stream: dump_file, format: " %d\t%d\n" , j, |
| 6163 | ivopts_estimate_reg_pressure (data, n_invs: 0, n_cands: j)); |
| 6164 | fprintf (stream: dump_file, format: "\n" ); |
| 6165 | } |
| 6166 | } |
| 6167 | |
| 6168 | /* Returns true if A is a cheaper cost pair than B. */ |
| 6169 | |
| 6170 | static bool |
| 6171 | cheaper_cost_pair (class cost_pair *a, class cost_pair *b) |
| 6172 | { |
| 6173 | if (!a) |
| 6174 | return false; |
| 6175 | |
| 6176 | if (!b) |
| 6177 | return true; |
| 6178 | |
| 6179 | if (a->cost < b->cost) |
| 6180 | return true; |
| 6181 | |
| 6182 | if (b->cost < a->cost) |
| 6183 | return false; |
| 6184 | |
| 6185 | /* In case the costs are the same, prefer the cheaper candidate. */ |
| 6186 | if (a->cand->cost < b->cand->cost) |
| 6187 | return true; |
| 6188 | |
| 6189 | return false; |
| 6190 | } |
| 6191 | |
| 6192 | /* Compare if A is a more expensive cost pair than B. Return 1, 0 and -1 |
| 6193 | for more expensive, equal and cheaper respectively. */ |
| 6194 | |
| 6195 | static int |
| 6196 | compare_cost_pair (class cost_pair *a, class cost_pair *b) |
| 6197 | { |
| 6198 | if (cheaper_cost_pair (a, b)) |
| 6199 | return -1; |
| 6200 | if (cheaper_cost_pair (a: b, b: a)) |
| 6201 | return 1; |
| 6202 | |
| 6203 | return 0; |
| 6204 | } |
| 6205 | |
| 6206 | /* Returns candidate by that USE is expressed in IVS. */ |
| 6207 | |
| 6208 | static class cost_pair * |
| 6209 | iv_ca_cand_for_group (class iv_ca *ivs, struct iv_group *group) |
| 6210 | { |
| 6211 | return ivs->cand_for_group[group->id]; |
| 6212 | } |
| 6213 | |
| 6214 | /* Computes the cost field of IVS structure. */ |
| 6215 | |
| 6216 | static void |
| 6217 | iv_ca_recount_cost (struct ivopts_data *data, class iv_ca *ivs) |
| 6218 | { |
| 6219 | comp_cost cost = ivs->cand_use_cost; |
| 6220 | |
| 6221 | cost += ivs->cand_cost; |
| 6222 | cost += ivopts_estimate_reg_pressure (data, n_invs: ivs->n_invs, n_cands: ivs->n_cands); |
| 6223 | ivs->cost = cost; |
| 6224 | } |
| 6225 | |
| 6226 | /* Remove use of invariants in set INVS by decreasing counter in N_INV_USES |
| 6227 | and IVS. */ |
| 6228 | |
| 6229 | static void |
| 6230 | iv_ca_set_remove_invs (class iv_ca *ivs, bitmap invs, unsigned *n_inv_uses) |
| 6231 | { |
| 6232 | bitmap_iterator bi; |
| 6233 | unsigned iid; |
| 6234 | |
| 6235 | if (!invs) |
| 6236 | return; |
| 6237 | |
| 6238 | gcc_assert (n_inv_uses != NULL); |
| 6239 | EXECUTE_IF_SET_IN_BITMAP (invs, 0, iid, bi) |
| 6240 | { |
| 6241 | n_inv_uses[iid]--; |
| 6242 | if (n_inv_uses[iid] == 0) |
| 6243 | ivs->n_invs--; |
| 6244 | } |
| 6245 | } |
| 6246 | |
| 6247 | /* Set USE not to be expressed by any candidate in IVS. */ |
| 6248 | |
| 6249 | static void |
| 6250 | iv_ca_set_no_cp (struct ivopts_data *data, class iv_ca *ivs, |
| 6251 | struct iv_group *group) |
| 6252 | { |
| 6253 | unsigned gid = group->id, cid; |
| 6254 | class cost_pair *cp; |
| 6255 | |
| 6256 | cp = ivs->cand_for_group[gid]; |
| 6257 | if (!cp) |
| 6258 | return; |
| 6259 | cid = cp->cand->id; |
| 6260 | |
| 6261 | ivs->bad_groups++; |
| 6262 | ivs->cand_for_group[gid] = NULL; |
| 6263 | ivs->n_cand_uses[cid]--; |
| 6264 | |
| 6265 | if (ivs->n_cand_uses[cid] == 0) |
| 6266 | { |
| 6267 | bitmap_clear_bit (ivs->cands, cid); |
| 6268 | if (!cp->cand->doloop_p || !targetm.have_count_reg_decr_p) |
| 6269 | ivs->n_cands--; |
| 6270 | ivs->cand_cost -= cp->cand->cost; |
| 6271 | iv_ca_set_remove_invs (ivs, invs: cp->cand->inv_vars, n_inv_uses: ivs->n_inv_var_uses); |
| 6272 | iv_ca_set_remove_invs (ivs, invs: cp->cand->inv_exprs, n_inv_uses: ivs->n_inv_expr_uses); |
| 6273 | } |
| 6274 | |
| 6275 | ivs->cand_use_cost -= cp->cost; |
| 6276 | iv_ca_set_remove_invs (ivs, invs: cp->inv_vars, n_inv_uses: ivs->n_inv_var_uses); |
| 6277 | iv_ca_set_remove_invs (ivs, invs: cp->inv_exprs, n_inv_uses: ivs->n_inv_expr_uses); |
| 6278 | iv_ca_recount_cost (data, ivs); |
| 6279 | } |
| 6280 | |
| 6281 | /* Add use of invariants in set INVS by increasing counter in N_INV_USES and |
| 6282 | IVS. */ |
| 6283 | |
| 6284 | static void |
| 6285 | iv_ca_set_add_invs (class iv_ca *ivs, bitmap invs, unsigned *n_inv_uses) |
| 6286 | { |
| 6287 | bitmap_iterator bi; |
| 6288 | unsigned iid; |
| 6289 | |
| 6290 | if (!invs) |
| 6291 | return; |
| 6292 | |
| 6293 | gcc_assert (n_inv_uses != NULL); |
| 6294 | EXECUTE_IF_SET_IN_BITMAP (invs, 0, iid, bi) |
| 6295 | { |
| 6296 | n_inv_uses[iid]++; |
| 6297 | if (n_inv_uses[iid] == 1) |
| 6298 | ivs->n_invs++; |
| 6299 | } |
| 6300 | } |
| 6301 | |
| 6302 | /* Set cost pair for GROUP in set IVS to CP. */ |
| 6303 | |
| 6304 | static void |
| 6305 | iv_ca_set_cp (struct ivopts_data *data, class iv_ca *ivs, |
| 6306 | struct iv_group *group, class cost_pair *cp) |
| 6307 | { |
| 6308 | unsigned gid = group->id, cid; |
| 6309 | |
| 6310 | if (ivs->cand_for_group[gid] == cp) |
| 6311 | return; |
| 6312 | |
| 6313 | if (ivs->cand_for_group[gid]) |
| 6314 | iv_ca_set_no_cp (data, ivs, group); |
| 6315 | |
| 6316 | if (cp) |
| 6317 | { |
| 6318 | cid = cp->cand->id; |
| 6319 | |
| 6320 | ivs->bad_groups--; |
| 6321 | ivs->cand_for_group[gid] = cp; |
| 6322 | ivs->n_cand_uses[cid]++; |
| 6323 | if (ivs->n_cand_uses[cid] == 1) |
| 6324 | { |
| 6325 | bitmap_set_bit (ivs->cands, cid); |
| 6326 | if (!cp->cand->doloop_p || !targetm.have_count_reg_decr_p) |
| 6327 | ivs->n_cands++; |
| 6328 | ivs->cand_cost += cp->cand->cost; |
| 6329 | iv_ca_set_add_invs (ivs, invs: cp->cand->inv_vars, n_inv_uses: ivs->n_inv_var_uses); |
| 6330 | iv_ca_set_add_invs (ivs, invs: cp->cand->inv_exprs, n_inv_uses: ivs->n_inv_expr_uses); |
| 6331 | } |
| 6332 | |
| 6333 | ivs->cand_use_cost += cp->cost; |
| 6334 | iv_ca_set_add_invs (ivs, invs: cp->inv_vars, n_inv_uses: ivs->n_inv_var_uses); |
| 6335 | iv_ca_set_add_invs (ivs, invs: cp->inv_exprs, n_inv_uses: ivs->n_inv_expr_uses); |
| 6336 | iv_ca_recount_cost (data, ivs); |
| 6337 | } |
| 6338 | } |
| 6339 | |
| 6340 | /* Extend set IVS by expressing USE by some of the candidates in it |
| 6341 | if possible. Consider all important candidates if candidates in |
| 6342 | set IVS don't give any result. */ |
| 6343 | |
| 6344 | static void |
| 6345 | iv_ca_add_group (struct ivopts_data *data, class iv_ca *ivs, |
| 6346 | struct iv_group *group) |
| 6347 | { |
| 6348 | class cost_pair *best_cp = NULL, *cp; |
| 6349 | bitmap_iterator bi; |
| 6350 | unsigned i; |
| 6351 | struct iv_cand *cand; |
| 6352 | |
| 6353 | gcc_assert (ivs->upto >= group->id); |
| 6354 | ivs->upto++; |
| 6355 | ivs->bad_groups++; |
| 6356 | |
| 6357 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, i, bi) |
| 6358 | { |
| 6359 | cand = data->vcands[i]; |
| 6360 | cp = get_group_iv_cost (data, group, cand); |
| 6361 | if (cheaper_cost_pair (a: cp, b: best_cp)) |
| 6362 | best_cp = cp; |
| 6363 | } |
| 6364 | |
| 6365 | if (best_cp == NULL) |
| 6366 | { |
| 6367 | EXECUTE_IF_SET_IN_BITMAP (data->important_candidates, 0, i, bi) |
| 6368 | { |
| 6369 | cand = data->vcands[i]; |
| 6370 | cp = get_group_iv_cost (data, group, cand); |
| 6371 | if (cheaper_cost_pair (a: cp, b: best_cp)) |
| 6372 | best_cp = cp; |
| 6373 | } |
| 6374 | } |
| 6375 | |
| 6376 | iv_ca_set_cp (data, ivs, group, cp: best_cp); |
| 6377 | } |
| 6378 | |
| 6379 | /* Get cost for assignment IVS. */ |
| 6380 | |
| 6381 | static comp_cost |
| 6382 | iv_ca_cost (class iv_ca *ivs) |
| 6383 | { |
| 6384 | /* This was a conditional expression but it triggered a bug in |
| 6385 | Sun C 5.5. */ |
| 6386 | if (ivs->bad_groups) |
| 6387 | return infinite_cost; |
| 6388 | else |
| 6389 | return ivs->cost; |
| 6390 | } |
| 6391 | |
| 6392 | /* Compare if applying NEW_CP to GROUP for IVS introduces more invariants |
| 6393 | than OLD_CP. Return 1, 0 and -1 for more, equal and fewer invariants |
| 6394 | respectively. */ |
| 6395 | |
| 6396 | static int |
| 6397 | iv_ca_compare_deps (struct ivopts_data *data, class iv_ca *ivs, |
| 6398 | struct iv_group *group, class cost_pair *old_cp, |
| 6399 | class cost_pair *new_cp) |
| 6400 | { |
| 6401 | gcc_assert (old_cp && new_cp && old_cp != new_cp); |
| 6402 | unsigned old_n_invs = ivs->n_invs; |
| 6403 | iv_ca_set_cp (data, ivs, group, cp: new_cp); |
| 6404 | unsigned new_n_invs = ivs->n_invs; |
| 6405 | iv_ca_set_cp (data, ivs, group, cp: old_cp); |
| 6406 | |
| 6407 | return new_n_invs > old_n_invs ? 1 : (new_n_invs < old_n_invs ? -1 : 0); |
| 6408 | } |
| 6409 | |
| 6410 | /* Creates change of expressing GROUP by NEW_CP instead of OLD_CP and chains |
| 6411 | it before NEXT. */ |
| 6412 | |
| 6413 | static struct iv_ca_delta * |
| 6414 | iv_ca_delta_add (struct iv_group *group, class cost_pair *old_cp, |
| 6415 | class cost_pair *new_cp, struct iv_ca_delta *next) |
| 6416 | { |
| 6417 | struct iv_ca_delta *change = XNEW (struct iv_ca_delta); |
| 6418 | |
| 6419 | change->group = group; |
| 6420 | change->old_cp = old_cp; |
| 6421 | change->new_cp = new_cp; |
| 6422 | change->next = next; |
| 6423 | |
| 6424 | return change; |
| 6425 | } |
| 6426 | |
| 6427 | /* Joins two lists of changes L1 and L2. Destructive -- old lists |
| 6428 | are rewritten. */ |
| 6429 | |
| 6430 | static struct iv_ca_delta * |
| 6431 | iv_ca_delta_join (struct iv_ca_delta *l1, struct iv_ca_delta *l2) |
| 6432 | { |
| 6433 | struct iv_ca_delta *last; |
| 6434 | |
| 6435 | if (!l2) |
| 6436 | return l1; |
| 6437 | |
| 6438 | if (!l1) |
| 6439 | return l2; |
| 6440 | |
| 6441 | for (last = l1; last->next; last = last->next) |
| 6442 | continue; |
| 6443 | last->next = l2; |
| 6444 | |
| 6445 | return l1; |
| 6446 | } |
| 6447 | |
| 6448 | /* Reverse the list of changes DELTA, forming the inverse to it. */ |
| 6449 | |
| 6450 | static struct iv_ca_delta * |
| 6451 | iv_ca_delta_reverse (struct iv_ca_delta *delta) |
| 6452 | { |
| 6453 | struct iv_ca_delta *act, *next, *prev = NULL; |
| 6454 | |
| 6455 | for (act = delta; act; act = next) |
| 6456 | { |
| 6457 | next = act->next; |
| 6458 | act->next = prev; |
| 6459 | prev = act; |
| 6460 | |
| 6461 | std::swap (a&: act->old_cp, b&: act->new_cp); |
| 6462 | } |
| 6463 | |
| 6464 | return prev; |
| 6465 | } |
| 6466 | |
| 6467 | /* Commit changes in DELTA to IVS. If FORWARD is false, the changes are |
| 6468 | reverted instead. */ |
| 6469 | |
| 6470 | static void |
| 6471 | iv_ca_delta_commit (struct ivopts_data *data, class iv_ca *ivs, |
| 6472 | struct iv_ca_delta *delta, bool forward) |
| 6473 | { |
| 6474 | class cost_pair *from, *to; |
| 6475 | struct iv_ca_delta *act; |
| 6476 | |
| 6477 | if (!forward) |
| 6478 | delta = iv_ca_delta_reverse (delta); |
| 6479 | |
| 6480 | for (act = delta; act; act = act->next) |
| 6481 | { |
| 6482 | from = act->old_cp; |
| 6483 | to = act->new_cp; |
| 6484 | gcc_assert (iv_ca_cand_for_group (ivs, act->group) == from); |
| 6485 | iv_ca_set_cp (data, ivs, group: act->group, cp: to); |
| 6486 | } |
| 6487 | |
| 6488 | if (!forward) |
| 6489 | iv_ca_delta_reverse (delta); |
| 6490 | } |
| 6491 | |
| 6492 | /* Returns true if CAND is used in IVS. */ |
| 6493 | |
| 6494 | static bool |
| 6495 | iv_ca_cand_used_p (class iv_ca *ivs, struct iv_cand *cand) |
| 6496 | { |
| 6497 | return ivs->n_cand_uses[cand->id] > 0; |
| 6498 | } |
| 6499 | |
| 6500 | /* Returns number of induction variable candidates in the set IVS. */ |
| 6501 | |
| 6502 | static unsigned |
| 6503 | iv_ca_n_cands (class iv_ca *ivs) |
| 6504 | { |
| 6505 | return ivs->n_cands; |
| 6506 | } |
| 6507 | |
| 6508 | /* Free the list of changes DELTA. */ |
| 6509 | |
| 6510 | static void |
| 6511 | iv_ca_delta_free (struct iv_ca_delta **delta) |
| 6512 | { |
| 6513 | struct iv_ca_delta *act, *next; |
| 6514 | |
| 6515 | for (act = *delta; act; act = next) |
| 6516 | { |
| 6517 | next = act->next; |
| 6518 | free (ptr: act); |
| 6519 | } |
| 6520 | |
| 6521 | *delta = NULL; |
| 6522 | } |
| 6523 | |
| 6524 | /* Allocates new iv candidates assignment. */ |
| 6525 | |
| 6526 | static class iv_ca * |
| 6527 | iv_ca_new (struct ivopts_data *data) |
| 6528 | { |
| 6529 | class iv_ca *nw = XNEW (class iv_ca); |
| 6530 | |
| 6531 | nw->upto = 0; |
| 6532 | nw->bad_groups = 0; |
| 6533 | nw->cand_for_group = XCNEWVEC (class cost_pair *, |
| 6534 | data->vgroups.length ()); |
| 6535 | nw->n_cand_uses = XCNEWVEC (unsigned, data->vcands.length ()); |
| 6536 | nw->cands = BITMAP_ALLOC (NULL); |
| 6537 | nw->n_cands = 0; |
| 6538 | nw->n_invs = 0; |
| 6539 | nw->cand_use_cost = no_cost; |
| 6540 | nw->cand_cost = 0; |
| 6541 | nw->n_inv_var_uses = XCNEWVEC (unsigned, data->max_inv_var_id + 1); |
| 6542 | nw->n_inv_expr_uses = XCNEWVEC (unsigned, data->max_inv_expr_id + 1); |
| 6543 | nw->cost = no_cost; |
| 6544 | |
| 6545 | return nw; |
| 6546 | } |
| 6547 | |
| 6548 | /* Free memory occupied by the set IVS. */ |
| 6549 | |
| 6550 | static void |
| 6551 | iv_ca_free (class iv_ca **ivs) |
| 6552 | { |
| 6553 | free (ptr: (*ivs)->cand_for_group); |
| 6554 | free (ptr: (*ivs)->n_cand_uses); |
| 6555 | BITMAP_FREE ((*ivs)->cands); |
| 6556 | free (ptr: (*ivs)->n_inv_var_uses); |
| 6557 | free (ptr: (*ivs)->n_inv_expr_uses); |
| 6558 | free (ptr: *ivs); |
| 6559 | *ivs = NULL; |
| 6560 | } |
| 6561 | |
| 6562 | /* Dumps IVS to FILE. */ |
| 6563 | |
| 6564 | static void |
| 6565 | iv_ca_dump (struct ivopts_data *data, FILE *file, class iv_ca *ivs) |
| 6566 | { |
| 6567 | unsigned i; |
| 6568 | comp_cost cost = iv_ca_cost (ivs); |
| 6569 | |
| 6570 | fprintf (stream: file, format: " cost: %" PRId64 " (complexity %d)\n" , cost.cost, |
| 6571 | cost.complexity); |
| 6572 | fprintf (stream: file, format: " reg_cost: %d\n" , |
| 6573 | ivopts_estimate_reg_pressure (data, n_invs: ivs->n_invs, n_cands: ivs->n_cands)); |
| 6574 | fprintf (stream: file, format: " cand_cost: %" PRId64 "\n cand_group_cost: " |
| 6575 | "%" PRId64 " (complexity %d)\n" , ivs->cand_cost, |
| 6576 | ivs->cand_use_cost.cost, ivs->cand_use_cost.complexity); |
| 6577 | bitmap_print (file, ivs->cands, " candidates: " ,"\n" ); |
| 6578 | |
| 6579 | for (i = 0; i < ivs->upto; i++) |
| 6580 | { |
| 6581 | struct iv_group *group = data->vgroups[i]; |
| 6582 | class cost_pair *cp = iv_ca_cand_for_group (ivs, group); |
| 6583 | if (cp) |
| 6584 | fprintf (stream: file, format: " group:%d --> iv_cand:%d, cost=(" |
| 6585 | "%" PRId64 ",%d)\n" , group->id, cp->cand->id, |
| 6586 | cp->cost.cost, cp->cost.complexity); |
| 6587 | else |
| 6588 | fprintf (stream: file, format: " group:%d --> ??\n" , group->id); |
| 6589 | } |
| 6590 | |
| 6591 | const char *pref = "" ; |
| 6592 | fprintf (stream: file, format: " invariant variables: " ); |
| 6593 | for (i = 1; i <= data->max_inv_var_id; i++) |
| 6594 | if (ivs->n_inv_var_uses[i]) |
| 6595 | { |
| 6596 | fprintf (stream: file, format: "%s%d" , pref, i); |
| 6597 | pref = ", " ; |
| 6598 | } |
| 6599 | |
| 6600 | pref = "" ; |
| 6601 | fprintf (stream: file, format: "\n invariant expressions: " ); |
| 6602 | for (i = 1; i <= data->max_inv_expr_id; i++) |
| 6603 | if (ivs->n_inv_expr_uses[i]) |
| 6604 | { |
| 6605 | fprintf (stream: file, format: "%s%d" , pref, i); |
| 6606 | pref = ", " ; |
| 6607 | } |
| 6608 | |
| 6609 | fprintf (stream: file, format: "\n\n" ); |
| 6610 | } |
| 6611 | |
| 6612 | /* Try changing candidate in IVS to CAND for each use. Return cost of the |
| 6613 | new set, and store differences in DELTA. Number of induction variables |
| 6614 | in the new set is stored to N_IVS. MIN_NCAND is a flag. When it is true |
| 6615 | the function will try to find a solution with mimimal iv candidates. */ |
| 6616 | |
| 6617 | static comp_cost |
| 6618 | iv_ca_extend (struct ivopts_data *data, class iv_ca *ivs, |
| 6619 | struct iv_cand *cand, struct iv_ca_delta **delta, |
| 6620 | unsigned *n_ivs, bool min_ncand) |
| 6621 | { |
| 6622 | unsigned i; |
| 6623 | comp_cost cost; |
| 6624 | struct iv_group *group; |
| 6625 | class cost_pair *old_cp, *new_cp; |
| 6626 | |
| 6627 | *delta = NULL; |
| 6628 | for (i = 0; i < ivs->upto; i++) |
| 6629 | { |
| 6630 | group = data->vgroups[i]; |
| 6631 | old_cp = iv_ca_cand_for_group (ivs, group); |
| 6632 | |
| 6633 | if (old_cp |
| 6634 | && old_cp->cand == cand) |
| 6635 | continue; |
| 6636 | |
| 6637 | new_cp = get_group_iv_cost (data, group, cand); |
| 6638 | if (!new_cp) |
| 6639 | continue; |
| 6640 | |
| 6641 | if (!min_ncand) |
| 6642 | { |
| 6643 | int cmp_invs = iv_ca_compare_deps (data, ivs, group, old_cp, new_cp); |
| 6644 | /* Skip if new_cp depends on more invariants. */ |
| 6645 | if (cmp_invs > 0) |
| 6646 | continue; |
| 6647 | |
| 6648 | int cmp_cost = compare_cost_pair (a: new_cp, b: old_cp); |
| 6649 | /* Skip if new_cp is not cheaper. */ |
| 6650 | if (cmp_cost > 0 || (cmp_cost == 0 && cmp_invs == 0)) |
| 6651 | continue; |
| 6652 | } |
| 6653 | |
| 6654 | *delta = iv_ca_delta_add (group, old_cp, new_cp, next: *delta); |
| 6655 | } |
| 6656 | |
| 6657 | iv_ca_delta_commit (data, ivs, delta: *delta, forward: true); |
| 6658 | cost = iv_ca_cost (ivs); |
| 6659 | if (n_ivs) |
| 6660 | *n_ivs = iv_ca_n_cands (ivs); |
| 6661 | iv_ca_delta_commit (data, ivs, delta: *delta, forward: false); |
| 6662 | |
| 6663 | return cost; |
| 6664 | } |
| 6665 | |
| 6666 | /* Try narrowing set IVS by removing CAND. Return the cost of |
| 6667 | the new set and store the differences in DELTA. START is |
| 6668 | the candidate with which we start narrowing. */ |
| 6669 | |
| 6670 | static comp_cost |
| 6671 | iv_ca_narrow (struct ivopts_data *data, class iv_ca *ivs, |
| 6672 | struct iv_cand *cand, struct iv_cand *start, |
| 6673 | struct iv_ca_delta **delta) |
| 6674 | { |
| 6675 | unsigned i, ci; |
| 6676 | struct iv_group *group; |
| 6677 | class cost_pair *old_cp, *new_cp, *cp; |
| 6678 | bitmap_iterator bi; |
| 6679 | struct iv_cand *cnd; |
| 6680 | comp_cost cost, best_cost, acost; |
| 6681 | |
| 6682 | *delta = NULL; |
| 6683 | for (i = 0; i < data->vgroups.length (); i++) |
| 6684 | { |
| 6685 | group = data->vgroups[i]; |
| 6686 | |
| 6687 | old_cp = iv_ca_cand_for_group (ivs, group); |
| 6688 | if (old_cp->cand != cand) |
| 6689 | continue; |
| 6690 | |
| 6691 | best_cost = iv_ca_cost (ivs); |
| 6692 | /* Start narrowing with START. */ |
| 6693 | new_cp = get_group_iv_cost (data, group, cand: start); |
| 6694 | |
| 6695 | if (data->consider_all_candidates) |
| 6696 | { |
| 6697 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, ci, bi) |
| 6698 | { |
| 6699 | if (ci == cand->id || (start && ci == start->id)) |
| 6700 | continue; |
| 6701 | |
| 6702 | cnd = data->vcands[ci]; |
| 6703 | |
| 6704 | cp = get_group_iv_cost (data, group, cand: cnd); |
| 6705 | if (!cp) |
| 6706 | continue; |
| 6707 | |
| 6708 | iv_ca_set_cp (data, ivs, group, cp); |
| 6709 | acost = iv_ca_cost (ivs); |
| 6710 | |
| 6711 | if (acost < best_cost) |
| 6712 | { |
| 6713 | best_cost = acost; |
| 6714 | new_cp = cp; |
| 6715 | } |
| 6716 | } |
| 6717 | } |
| 6718 | else |
| 6719 | { |
| 6720 | EXECUTE_IF_AND_IN_BITMAP (group->related_cands, ivs->cands, 0, ci, bi) |
| 6721 | { |
| 6722 | if (ci == cand->id || (start && ci == start->id)) |
| 6723 | continue; |
| 6724 | |
| 6725 | cnd = data->vcands[ci]; |
| 6726 | |
| 6727 | cp = get_group_iv_cost (data, group, cand: cnd); |
| 6728 | if (!cp) |
| 6729 | continue; |
| 6730 | |
| 6731 | iv_ca_set_cp (data, ivs, group, cp); |
| 6732 | acost = iv_ca_cost (ivs); |
| 6733 | |
| 6734 | if (acost < best_cost) |
| 6735 | { |
| 6736 | best_cost = acost; |
| 6737 | new_cp = cp; |
| 6738 | } |
| 6739 | } |
| 6740 | } |
| 6741 | /* Restore to old cp for use. */ |
| 6742 | iv_ca_set_cp (data, ivs, group, cp: old_cp); |
| 6743 | |
| 6744 | if (!new_cp) |
| 6745 | { |
| 6746 | iv_ca_delta_free (delta); |
| 6747 | return infinite_cost; |
| 6748 | } |
| 6749 | |
| 6750 | *delta = iv_ca_delta_add (group, old_cp, new_cp, next: *delta); |
| 6751 | } |
| 6752 | |
| 6753 | iv_ca_delta_commit (data, ivs, delta: *delta, forward: true); |
| 6754 | cost = iv_ca_cost (ivs); |
| 6755 | iv_ca_delta_commit (data, ivs, delta: *delta, forward: false); |
| 6756 | |
| 6757 | return cost; |
| 6758 | } |
| 6759 | |
| 6760 | /* Try optimizing the set of candidates IVS by removing candidates different |
| 6761 | from to EXCEPT_CAND from it. Return cost of the new set, and store |
| 6762 | differences in DELTA. */ |
| 6763 | |
| 6764 | static comp_cost |
| 6765 | iv_ca_prune (struct ivopts_data *data, class iv_ca *ivs, |
| 6766 | struct iv_cand *except_cand, struct iv_ca_delta **delta) |
| 6767 | { |
| 6768 | bitmap_iterator bi; |
| 6769 | struct iv_ca_delta *act_delta, *best_delta; |
| 6770 | unsigned i; |
| 6771 | comp_cost best_cost, acost; |
| 6772 | struct iv_cand *cand; |
| 6773 | |
| 6774 | best_delta = NULL; |
| 6775 | best_cost = iv_ca_cost (ivs); |
| 6776 | |
| 6777 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, i, bi) |
| 6778 | { |
| 6779 | cand = data->vcands[i]; |
| 6780 | |
| 6781 | if (cand == except_cand) |
| 6782 | continue; |
| 6783 | |
| 6784 | acost = iv_ca_narrow (data, ivs, cand, start: except_cand, delta: &act_delta); |
| 6785 | |
| 6786 | if (acost < best_cost) |
| 6787 | { |
| 6788 | best_cost = acost; |
| 6789 | iv_ca_delta_free (delta: &best_delta); |
| 6790 | best_delta = act_delta; |
| 6791 | } |
| 6792 | else |
| 6793 | iv_ca_delta_free (delta: &act_delta); |
| 6794 | } |
| 6795 | |
| 6796 | if (!best_delta) |
| 6797 | { |
| 6798 | *delta = NULL; |
| 6799 | return best_cost; |
| 6800 | } |
| 6801 | |
| 6802 | /* Recurse to possibly remove other unnecessary ivs. */ |
| 6803 | iv_ca_delta_commit (data, ivs, delta: best_delta, forward: true); |
| 6804 | best_cost = iv_ca_prune (data, ivs, except_cand, delta); |
| 6805 | iv_ca_delta_commit (data, ivs, delta: best_delta, forward: false); |
| 6806 | *delta = iv_ca_delta_join (l1: best_delta, l2: *delta); |
| 6807 | return best_cost; |
| 6808 | } |
| 6809 | |
| 6810 | /* Check if CAND_IDX is a candidate other than OLD_CAND and has |
| 6811 | cheaper local cost for GROUP than BEST_CP. Return pointer to |
| 6812 | the corresponding cost_pair, otherwise just return BEST_CP. */ |
| 6813 | |
| 6814 | static class cost_pair* |
| 6815 | cheaper_cost_with_cand (struct ivopts_data *data, struct iv_group *group, |
| 6816 | unsigned int cand_idx, struct iv_cand *old_cand, |
| 6817 | class cost_pair *best_cp) |
| 6818 | { |
| 6819 | struct iv_cand *cand; |
| 6820 | class cost_pair *cp; |
| 6821 | |
| 6822 | gcc_assert (old_cand != NULL && best_cp != NULL); |
| 6823 | if (cand_idx == old_cand->id) |
| 6824 | return best_cp; |
| 6825 | |
| 6826 | cand = data->vcands[cand_idx]; |
| 6827 | cp = get_group_iv_cost (data, group, cand); |
| 6828 | if (cp != NULL && cheaper_cost_pair (a: cp, b: best_cp)) |
| 6829 | return cp; |
| 6830 | |
| 6831 | return best_cp; |
| 6832 | } |
| 6833 | |
| 6834 | /* Try breaking local optimal fixed-point for IVS by replacing candidates |
| 6835 | which are used by more than one iv uses. For each of those candidates, |
| 6836 | this function tries to represent iv uses under that candidate using |
| 6837 | other ones with lower local cost, then tries to prune the new set. |
| 6838 | If the new set has lower cost, It returns the new cost after recording |
| 6839 | candidate replacement in list DELTA. */ |
| 6840 | |
| 6841 | static comp_cost |
| 6842 | iv_ca_replace (struct ivopts_data *data, class iv_ca *ivs, |
| 6843 | struct iv_ca_delta **delta) |
| 6844 | { |
| 6845 | bitmap_iterator bi, bj; |
| 6846 | unsigned int i, j, k; |
| 6847 | struct iv_cand *cand; |
| 6848 | comp_cost orig_cost, acost; |
| 6849 | struct iv_ca_delta *act_delta, *tmp_delta; |
| 6850 | class cost_pair *old_cp, *best_cp = NULL; |
| 6851 | |
| 6852 | *delta = NULL; |
| 6853 | orig_cost = iv_ca_cost (ivs); |
| 6854 | |
| 6855 | EXECUTE_IF_SET_IN_BITMAP (ivs->cands, 0, i, bi) |
| 6856 | { |
| 6857 | if (ivs->n_cand_uses[i] == 1 |
| 6858 | || ivs->n_cand_uses[i] > ALWAYS_PRUNE_CAND_SET_BOUND) |
| 6859 | continue; |
| 6860 | |
| 6861 | cand = data->vcands[i]; |
| 6862 | |
| 6863 | act_delta = NULL; |
| 6864 | /* Represent uses under current candidate using other ones with |
| 6865 | lower local cost. */ |
| 6866 | for (j = 0; j < ivs->upto; j++) |
| 6867 | { |
| 6868 | struct iv_group *group = data->vgroups[j]; |
| 6869 | old_cp = iv_ca_cand_for_group (ivs, group); |
| 6870 | |
| 6871 | if (old_cp->cand != cand) |
| 6872 | continue; |
| 6873 | |
| 6874 | best_cp = old_cp; |
| 6875 | if (data->consider_all_candidates) |
| 6876 | for (k = 0; k < data->vcands.length (); k++) |
| 6877 | best_cp = cheaper_cost_with_cand (data, group, cand_idx: k, |
| 6878 | old_cand: old_cp->cand, best_cp); |
| 6879 | else |
| 6880 | EXECUTE_IF_SET_IN_BITMAP (group->related_cands, 0, k, bj) |
| 6881 | best_cp = cheaper_cost_with_cand (data, group, cand_idx: k, |
| 6882 | old_cand: old_cp->cand, best_cp); |
| 6883 | |
| 6884 | if (best_cp == old_cp) |
| 6885 | continue; |
| 6886 | |
| 6887 | act_delta = iv_ca_delta_add (group, old_cp, new_cp: best_cp, next: act_delta); |
| 6888 | } |
| 6889 | /* No need for further prune. */ |
| 6890 | if (!act_delta) |
| 6891 | continue; |
| 6892 | |
| 6893 | /* Prune the new candidate set. */ |
| 6894 | iv_ca_delta_commit (data, ivs, delta: act_delta, forward: true); |
| 6895 | acost = iv_ca_prune (data, ivs, NULL, delta: &tmp_delta); |
| 6896 | iv_ca_delta_commit (data, ivs, delta: act_delta, forward: false); |
| 6897 | act_delta = iv_ca_delta_join (l1: act_delta, l2: tmp_delta); |
| 6898 | |
| 6899 | if (acost < orig_cost) |
| 6900 | { |
| 6901 | *delta = act_delta; |
| 6902 | return acost; |
| 6903 | } |
| 6904 | else |
| 6905 | iv_ca_delta_free (delta: &act_delta); |
| 6906 | } |
| 6907 | |
| 6908 | return orig_cost; |
| 6909 | } |
| 6910 | |
| 6911 | /* Tries to extend the sets IVS in the best possible way in order to |
| 6912 | express the GROUP. If ORIGINALP is true, prefer candidates from |
| 6913 | the original set of IVs, otherwise favor important candidates not |
| 6914 | based on any memory object. */ |
| 6915 | |
| 6916 | static bool |
| 6917 | try_add_cand_for (struct ivopts_data *data, class iv_ca *ivs, |
| 6918 | struct iv_group *group, bool originalp) |
| 6919 | { |
| 6920 | comp_cost best_cost, act_cost; |
| 6921 | unsigned i; |
| 6922 | bitmap_iterator bi; |
| 6923 | struct iv_cand *cand; |
| 6924 | struct iv_ca_delta *best_delta = NULL, *act_delta; |
| 6925 | class cost_pair *cp; |
| 6926 | |
| 6927 | iv_ca_add_group (data, ivs, group); |
| 6928 | best_cost = iv_ca_cost (ivs); |
| 6929 | cp = iv_ca_cand_for_group (ivs, group); |
| 6930 | if (cp) |
| 6931 | { |
| 6932 | best_delta = iv_ca_delta_add (group, NULL, new_cp: cp, NULL); |
| 6933 | iv_ca_set_no_cp (data, ivs, group); |
| 6934 | } |
| 6935 | |
| 6936 | /* If ORIGINALP is true, try to find the original IV for the use. Otherwise |
| 6937 | first try important candidates not based on any memory object. Only if |
| 6938 | this fails, try the specific ones. Rationale -- in loops with many |
| 6939 | variables the best choice often is to use just one generic biv. If we |
| 6940 | added here many ivs specific to the uses, the optimization algorithm later |
| 6941 | would be likely to get stuck in a local minimum, thus causing us to create |
| 6942 | too many ivs. The approach from few ivs to more seems more likely to be |
| 6943 | successful -- starting from few ivs, replacing an expensive use by a |
| 6944 | specific iv should always be a win. */ |
| 6945 | EXECUTE_IF_SET_IN_BITMAP (group->related_cands, 0, i, bi) |
| 6946 | { |
| 6947 | cand = data->vcands[i]; |
| 6948 | |
| 6949 | if (originalp && cand->pos !=IP_ORIGINAL) |
| 6950 | continue; |
| 6951 | |
| 6952 | if (!originalp && cand->iv->base_object != NULL_TREE) |
| 6953 | continue; |
| 6954 | |
| 6955 | if (iv_ca_cand_used_p (ivs, cand)) |
| 6956 | continue; |
| 6957 | |
| 6958 | cp = get_group_iv_cost (data, group, cand); |
| 6959 | if (!cp) |
| 6960 | continue; |
| 6961 | |
| 6962 | iv_ca_set_cp (data, ivs, group, cp); |
| 6963 | act_cost = iv_ca_extend (data, ivs, cand, delta: &act_delta, NULL, |
| 6964 | min_ncand: true); |
| 6965 | iv_ca_set_no_cp (data, ivs, group); |
| 6966 | act_delta = iv_ca_delta_add (group, NULL, new_cp: cp, next: act_delta); |
| 6967 | |
| 6968 | if (act_cost < best_cost) |
| 6969 | { |
| 6970 | best_cost = act_cost; |
| 6971 | |
| 6972 | iv_ca_delta_free (delta: &best_delta); |
| 6973 | best_delta = act_delta; |
| 6974 | } |
| 6975 | else |
| 6976 | iv_ca_delta_free (delta: &act_delta); |
| 6977 | } |
| 6978 | |
| 6979 | if (best_cost.infinite_cost_p ()) |
| 6980 | { |
| 6981 | for (i = 0; i < group->n_map_members; i++) |
| 6982 | { |
| 6983 | cp = group->cost_map + i; |
| 6984 | cand = cp->cand; |
| 6985 | if (!cand) |
| 6986 | continue; |
| 6987 | |
| 6988 | /* Already tried this. */ |
| 6989 | if (cand->important) |
| 6990 | { |
| 6991 | if (originalp && cand->pos == IP_ORIGINAL) |
| 6992 | continue; |
| 6993 | if (!originalp && cand->iv->base_object == NULL_TREE) |
| 6994 | continue; |
| 6995 | } |
| 6996 | |
| 6997 | if (iv_ca_cand_used_p (ivs, cand)) |
| 6998 | continue; |
| 6999 | |
| 7000 | act_delta = NULL; |
| 7001 | iv_ca_set_cp (data, ivs, group, cp); |
| 7002 | act_cost = iv_ca_extend (data, ivs, cand, delta: &act_delta, NULL, min_ncand: true); |
| 7003 | iv_ca_set_no_cp (data, ivs, group); |
| 7004 | act_delta = iv_ca_delta_add (group, |
| 7005 | old_cp: iv_ca_cand_for_group (ivs, group), |
| 7006 | new_cp: cp, next: act_delta); |
| 7007 | |
| 7008 | if (act_cost < best_cost) |
| 7009 | { |
| 7010 | best_cost = act_cost; |
| 7011 | |
| 7012 | if (best_delta) |
| 7013 | iv_ca_delta_free (delta: &best_delta); |
| 7014 | best_delta = act_delta; |
| 7015 | } |
| 7016 | else |
| 7017 | iv_ca_delta_free (delta: &act_delta); |
| 7018 | } |
| 7019 | } |
| 7020 | |
| 7021 | iv_ca_delta_commit (data, ivs, delta: best_delta, forward: true); |
| 7022 | iv_ca_delta_free (delta: &best_delta); |
| 7023 | |
| 7024 | return !best_cost.infinite_cost_p (); |
| 7025 | } |
| 7026 | |
| 7027 | /* Finds an initial assignment of candidates to uses. */ |
| 7028 | |
| 7029 | static class iv_ca * |
| 7030 | get_initial_solution (struct ivopts_data *data, bool originalp) |
| 7031 | { |
| 7032 | unsigned i; |
| 7033 | class iv_ca *ivs = iv_ca_new (data); |
| 7034 | |
| 7035 | for (i = 0; i < data->vgroups.length (); i++) |
| 7036 | if (!try_add_cand_for (data, ivs, group: data->vgroups[i], originalp)) |
| 7037 | { |
| 7038 | iv_ca_free (ivs: &ivs); |
| 7039 | return NULL; |
| 7040 | } |
| 7041 | |
| 7042 | return ivs; |
| 7043 | } |
| 7044 | |
| 7045 | /* Tries to improve set of induction variables IVS. TRY_REPLACE_P |
| 7046 | points to a bool variable, this function tries to break local |
| 7047 | optimal fixed-point by replacing candidates in IVS if it's true. */ |
| 7048 | |
| 7049 | static bool |
| 7050 | try_improve_iv_set (struct ivopts_data *data, |
| 7051 | class iv_ca *ivs, bool *try_replace_p) |
| 7052 | { |
| 7053 | unsigned i, n_ivs; |
| 7054 | comp_cost acost, best_cost = iv_ca_cost (ivs); |
| 7055 | struct iv_ca_delta *best_delta = NULL, *act_delta, *tmp_delta; |
| 7056 | struct iv_cand *cand; |
| 7057 | |
| 7058 | /* Try extending the set of induction variables by one. */ |
| 7059 | for (i = 0; i < data->vcands.length (); i++) |
| 7060 | { |
| 7061 | cand = data->vcands[i]; |
| 7062 | |
| 7063 | if (iv_ca_cand_used_p (ivs, cand)) |
| 7064 | continue; |
| 7065 | |
| 7066 | acost = iv_ca_extend (data, ivs, cand, delta: &act_delta, n_ivs: &n_ivs, min_ncand: false); |
| 7067 | if (!act_delta) |
| 7068 | continue; |
| 7069 | |
| 7070 | /* If we successfully added the candidate and the set is small enough, |
| 7071 | try optimizing it by removing other candidates. */ |
| 7072 | if (n_ivs <= ALWAYS_PRUNE_CAND_SET_BOUND) |
| 7073 | { |
| 7074 | iv_ca_delta_commit (data, ivs, delta: act_delta, forward: true); |
| 7075 | acost = iv_ca_prune (data, ivs, except_cand: cand, delta: &tmp_delta); |
| 7076 | iv_ca_delta_commit (data, ivs, delta: act_delta, forward: false); |
| 7077 | act_delta = iv_ca_delta_join (l1: act_delta, l2: tmp_delta); |
| 7078 | } |
| 7079 | |
| 7080 | if (acost < best_cost) |
| 7081 | { |
| 7082 | best_cost = acost; |
| 7083 | iv_ca_delta_free (delta: &best_delta); |
| 7084 | best_delta = act_delta; |
| 7085 | } |
| 7086 | else |
| 7087 | iv_ca_delta_free (delta: &act_delta); |
| 7088 | } |
| 7089 | |
| 7090 | if (!best_delta) |
| 7091 | { |
| 7092 | /* Try removing the candidates from the set instead. */ |
| 7093 | best_cost = iv_ca_prune (data, ivs, NULL, delta: &best_delta); |
| 7094 | |
| 7095 | if (!best_delta && *try_replace_p) |
| 7096 | { |
| 7097 | *try_replace_p = false; |
| 7098 | /* So far candidate selecting algorithm tends to choose fewer IVs |
| 7099 | so that it can handle cases in which loops have many variables |
| 7100 | but the best choice is often to use only one general biv. One |
| 7101 | weakness is it can't handle opposite cases, in which different |
| 7102 | candidates should be chosen with respect to each use. To solve |
| 7103 | the problem, we replace candidates in a manner described by the |
| 7104 | comments of iv_ca_replace, thus give general algorithm a chance |
| 7105 | to break local optimal fixed-point in these cases. */ |
| 7106 | best_cost = iv_ca_replace (data, ivs, delta: &best_delta); |
| 7107 | } |
| 7108 | |
| 7109 | if (!best_delta) |
| 7110 | return false; |
| 7111 | } |
| 7112 | |
| 7113 | iv_ca_delta_commit (data, ivs, delta: best_delta, forward: true); |
| 7114 | iv_ca_delta_free (delta: &best_delta); |
| 7115 | return best_cost == iv_ca_cost (ivs); |
| 7116 | } |
| 7117 | |
| 7118 | /* Attempts to find the optimal set of induction variables. We do simple |
| 7119 | greedy heuristic -- we try to replace at most one candidate in the selected |
| 7120 | solution and remove the unused ivs while this improves the cost. */ |
| 7121 | |
| 7122 | static class iv_ca * |
| 7123 | find_optimal_iv_set_1 (struct ivopts_data *data, bool originalp) |
| 7124 | { |
| 7125 | class iv_ca *set; |
| 7126 | bool try_replace_p = true; |
| 7127 | |
| 7128 | /* Get the initial solution. */ |
| 7129 | set = get_initial_solution (data, originalp); |
| 7130 | if (!set) |
| 7131 | { |
| 7132 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 7133 | fprintf (stream: dump_file, format: "Unable to substitute for ivs, failed.\n" ); |
| 7134 | return NULL; |
| 7135 | } |
| 7136 | |
| 7137 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 7138 | { |
| 7139 | fprintf (stream: dump_file, format: "Initial set of candidates:\n" ); |
| 7140 | iv_ca_dump (data, file: dump_file, ivs: set); |
| 7141 | } |
| 7142 | |
| 7143 | while (try_improve_iv_set (data, ivs: set, try_replace_p: &try_replace_p)) |
| 7144 | { |
| 7145 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 7146 | { |
| 7147 | fprintf (stream: dump_file, format: "Improved to:\n" ); |
| 7148 | iv_ca_dump (data, file: dump_file, ivs: set); |
| 7149 | } |
| 7150 | } |
| 7151 | |
| 7152 | /* If the set has infinite_cost, it can't be optimal. */ |
| 7153 | if (iv_ca_cost (ivs: set).infinite_cost_p ()) |
| 7154 | { |
| 7155 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 7156 | fprintf (stream: dump_file, |
| 7157 | format: "Overflow to infinite cost in try_improve_iv_set.\n" ); |
| 7158 | iv_ca_free (ivs: &set); |
| 7159 | } |
| 7160 | return set; |
| 7161 | } |
| 7162 | |
| 7163 | static class iv_ca * |
| 7164 | find_optimal_iv_set (struct ivopts_data *data) |
| 7165 | { |
| 7166 | unsigned i; |
| 7167 | comp_cost cost, origcost; |
| 7168 | class iv_ca *set, *origset; |
| 7169 | |
| 7170 | /* Determine the cost based on a strategy that starts with original IVs, |
| 7171 | and try again using a strategy that prefers candidates not based |
| 7172 | on any IVs. */ |
| 7173 | origset = find_optimal_iv_set_1 (data, originalp: true); |
| 7174 | set = find_optimal_iv_set_1 (data, originalp: false); |
| 7175 | |
| 7176 | if (!origset && !set) |
| 7177 | return NULL; |
| 7178 | |
| 7179 | origcost = origset ? iv_ca_cost (ivs: origset) : infinite_cost; |
| 7180 | cost = set ? iv_ca_cost (ivs: set) : infinite_cost; |
| 7181 | |
| 7182 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 7183 | { |
| 7184 | fprintf (stream: dump_file, format: "Original cost %" PRId64 " (complexity %d)\n\n" , |
| 7185 | origcost.cost, origcost.complexity); |
| 7186 | fprintf (stream: dump_file, format: "Final cost %" PRId64 " (complexity %d)\n\n" , |
| 7187 | cost.cost, cost.complexity); |
| 7188 | } |
| 7189 | |
| 7190 | /* Choose the one with the best cost. */ |
| 7191 | if (origcost <= cost) |
| 7192 | { |
| 7193 | if (set) |
| 7194 | iv_ca_free (ivs: &set); |
| 7195 | set = origset; |
| 7196 | } |
| 7197 | else if (origset) |
| 7198 | iv_ca_free (ivs: &origset); |
| 7199 | |
| 7200 | for (i = 0; i < data->vgroups.length (); i++) |
| 7201 | { |
| 7202 | struct iv_group *group = data->vgroups[i]; |
| 7203 | group->selected = iv_ca_cand_for_group (ivs: set, group)->cand; |
| 7204 | } |
| 7205 | |
| 7206 | return set; |
| 7207 | } |
| 7208 | |
| 7209 | /* Creates a new induction variable corresponding to CAND. */ |
| 7210 | |
| 7211 | static void |
| 7212 | create_new_iv (struct ivopts_data *data, struct iv_cand *cand) |
| 7213 | { |
| 7214 | gimple_stmt_iterator incr_pos; |
| 7215 | tree base; |
| 7216 | struct iv_use *use; |
| 7217 | struct iv_group *group; |
| 7218 | bool after = false; |
| 7219 | |
| 7220 | gcc_assert (cand->iv != NULL); |
| 7221 | |
| 7222 | switch (cand->pos) |
| 7223 | { |
| 7224 | case IP_NORMAL: |
| 7225 | incr_pos = gsi_last_bb (bb: ip_normal_pos (data->current_loop)); |
| 7226 | break; |
| 7227 | |
| 7228 | case IP_END: |
| 7229 | incr_pos = gsi_last_bb (bb: ip_end_pos (data->current_loop)); |
| 7230 | after = true; |
| 7231 | gcc_assert (gsi_end_p (incr_pos) || !stmt_ends_bb_p (*incr_pos)); |
| 7232 | break; |
| 7233 | |
| 7234 | case IP_AFTER_USE: |
| 7235 | after = true; |
| 7236 | /* fall through */ |
| 7237 | case IP_BEFORE_USE: |
| 7238 | incr_pos = gsi_for_stmt (cand->incremented_at); |
| 7239 | break; |
| 7240 | |
| 7241 | case IP_ORIGINAL: |
| 7242 | /* Mark that the iv is preserved. */ |
| 7243 | name_info (data, name: cand->var_before)->preserve_biv = true; |
| 7244 | name_info (data, name: cand->var_after)->preserve_biv = true; |
| 7245 | |
| 7246 | /* Rewrite the increment so that it uses var_before directly. */ |
| 7247 | use = find_interesting_uses_op (data, op: cand->var_after); |
| 7248 | group = data->vgroups[use->group_id]; |
| 7249 | group->selected = cand; |
| 7250 | return; |
| 7251 | } |
| 7252 | |
| 7253 | gimple_add_tmp_var (cand->var_before); |
| 7254 | |
| 7255 | base = unshare_expr (cand->iv->base); |
| 7256 | |
| 7257 | /* The step computation could invoke UB when the loop does not iterate. |
| 7258 | Avoid inserting it on the preheader in its native form but rewrite |
| 7259 | it to a well-defined form. This also helps masking SCEV issues |
| 7260 | which freely re-associates the IV computations when building up |
| 7261 | CHRECs without much regard for signed overflow invoking UB. */ |
| 7262 | gimple_seq stmts = NULL; |
| 7263 | tree step = force_gimple_operand (unshare_expr (cand->iv->step), &stmts, |
| 7264 | true, NULL_TREE); |
| 7265 | if (stmts) |
| 7266 | { |
| 7267 | for (auto gsi = gsi_start (seq&: stmts); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
| 7268 | if (gimple_needing_rewrite_undefined (gsi_stmt (i: gsi))) |
| 7269 | rewrite_to_defined_unconditional (&gsi); |
| 7270 | gsi_insert_seq_on_edge_immediate |
| 7271 | (loop_preheader_edge (data->current_loop), stmts); |
| 7272 | } |
| 7273 | |
| 7274 | create_iv (base, PLUS_EXPR, step, |
| 7275 | cand->var_before, data->current_loop, |
| 7276 | &incr_pos, after, &cand->var_before, &cand->var_after); |
| 7277 | } |
| 7278 | |
| 7279 | /* Creates new induction variables described in SET. */ |
| 7280 | |
| 7281 | static void |
| 7282 | create_new_ivs (struct ivopts_data *data, class iv_ca *set) |
| 7283 | { |
| 7284 | unsigned i; |
| 7285 | struct iv_cand *cand; |
| 7286 | bitmap_iterator bi; |
| 7287 | |
| 7288 | EXECUTE_IF_SET_IN_BITMAP (set->cands, 0, i, bi) |
| 7289 | { |
| 7290 | cand = data->vcands[i]; |
| 7291 | create_new_iv (data, cand); |
| 7292 | } |
| 7293 | |
| 7294 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 7295 | { |
| 7296 | fprintf (stream: dump_file, format: "Selected IV set for loop %d" , |
| 7297 | data->current_loop->num); |
| 7298 | if (data->loop_loc != UNKNOWN_LOCATION) |
| 7299 | fprintf (stream: dump_file, format: " at %s:%d" , LOCATION_FILE (data->loop_loc), |
| 7300 | LOCATION_LINE (data->loop_loc)); |
| 7301 | fprintf (stream: dump_file, format: ", " HOST_WIDE_INT_PRINT_UNSIGNED " avg niters" , |
| 7302 | avg_loop_niter (loop: data->current_loop)); |
| 7303 | fprintf (stream: dump_file, format: ", %lu IVs:\n" , bitmap_count_bits (set->cands)); |
| 7304 | EXECUTE_IF_SET_IN_BITMAP (set->cands, 0, i, bi) |
| 7305 | { |
| 7306 | cand = data->vcands[i]; |
| 7307 | dump_cand (file: dump_file, cand); |
| 7308 | } |
| 7309 | fprintf (stream: dump_file, format: "\n" ); |
| 7310 | } |
| 7311 | } |
| 7312 | |
| 7313 | /* Rewrites USE (definition of iv used in a nonlinear expression) |
| 7314 | using candidate CAND. */ |
| 7315 | |
| 7316 | static void |
| 7317 | rewrite_use_nonlinear_expr (struct ivopts_data *data, |
| 7318 | struct iv_use *use, struct iv_cand *cand) |
| 7319 | { |
| 7320 | gassign *ass; |
| 7321 | gimple_stmt_iterator bsi; |
| 7322 | tree comp, type = get_use_type (use), tgt; |
| 7323 | |
| 7324 | /* An important special case -- if we are asked to express value of |
| 7325 | the original iv by itself, just exit; there is no need to |
| 7326 | introduce a new computation (that might also need casting the |
| 7327 | variable to unsigned and back). */ |
| 7328 | if (cand->pos == IP_ORIGINAL |
| 7329 | && cand->incremented_at == use->stmt) |
| 7330 | { |
| 7331 | tree op = NULL_TREE; |
| 7332 | enum tree_code stmt_code; |
| 7333 | |
| 7334 | gcc_assert (is_gimple_assign (use->stmt)); |
| 7335 | gcc_assert (gimple_assign_lhs (use->stmt) == cand->var_after); |
| 7336 | |
| 7337 | /* Check whether we may leave the computation unchanged. |
| 7338 | This is the case only if it does not rely on other |
| 7339 | computations in the loop -- otherwise, the computation |
| 7340 | we rely upon may be removed in remove_unused_ivs, |
| 7341 | thus leading to ICE. */ |
| 7342 | stmt_code = gimple_assign_rhs_code (gs: use->stmt); |
| 7343 | if (stmt_code == PLUS_EXPR |
| 7344 | || stmt_code == MINUS_EXPR |
| 7345 | || stmt_code == POINTER_PLUS_EXPR) |
| 7346 | { |
| 7347 | if (gimple_assign_rhs1 (gs: use->stmt) == cand->var_before) |
| 7348 | op = gimple_assign_rhs2 (gs: use->stmt); |
| 7349 | else if (gimple_assign_rhs2 (gs: use->stmt) == cand->var_before) |
| 7350 | op = gimple_assign_rhs1 (gs: use->stmt); |
| 7351 | } |
| 7352 | |
| 7353 | if (op != NULL_TREE) |
| 7354 | { |
| 7355 | if (expr_invariant_in_loop_p (loop: data->current_loop, expr: op)) |
| 7356 | return; |
| 7357 | if (TREE_CODE (op) == SSA_NAME) |
| 7358 | { |
| 7359 | struct iv *iv = get_iv (data, var: op); |
| 7360 | if (iv != NULL && integer_zerop (iv->step)) |
| 7361 | return; |
| 7362 | } |
| 7363 | } |
| 7364 | } |
| 7365 | |
| 7366 | switch (gimple_code (g: use->stmt)) |
| 7367 | { |
| 7368 | case GIMPLE_PHI: |
| 7369 | tgt = PHI_RESULT (use->stmt); |
| 7370 | |
| 7371 | /* If we should keep the biv, do not replace it. */ |
| 7372 | if (name_info (data, name: tgt)->preserve_biv) |
| 7373 | return; |
| 7374 | |
| 7375 | bsi = gsi_after_labels (bb: gimple_bb (g: use->stmt)); |
| 7376 | break; |
| 7377 | |
| 7378 | case GIMPLE_ASSIGN: |
| 7379 | tgt = gimple_assign_lhs (gs: use->stmt); |
| 7380 | bsi = gsi_for_stmt (use->stmt); |
| 7381 | break; |
| 7382 | |
| 7383 | default: |
| 7384 | gcc_unreachable (); |
| 7385 | } |
| 7386 | |
| 7387 | aff_tree aff_inv, aff_var; |
| 7388 | if (!get_computation_aff_1 (data, at: use->stmt, use, cand, aff_inv: &aff_inv, aff_var: &aff_var)) |
| 7389 | gcc_unreachable (); |
| 7390 | |
| 7391 | unshare_aff_combination (&aff_inv); |
| 7392 | unshare_aff_combination (&aff_var); |
| 7393 | /* Prefer CSE opportunity than loop invariant by adding offset at last |
| 7394 | so that iv_uses have different offsets can be CSEed. */ |
| 7395 | poly_widest_int offset = aff_inv.offset; |
| 7396 | aff_inv.offset = 0; |
| 7397 | |
| 7398 | gimple_seq stmt_list = NULL, seq = NULL; |
| 7399 | tree comp_op1 = aff_combination_to_tree (&aff_inv); |
| 7400 | tree comp_op2 = aff_combination_to_tree (&aff_var); |
| 7401 | gcc_assert (comp_op1 && comp_op2); |
| 7402 | |
| 7403 | comp_op1 = force_gimple_operand (comp_op1, &seq, true, NULL); |
| 7404 | gimple_seq_add_seq (&stmt_list, seq); |
| 7405 | comp_op2 = force_gimple_operand (comp_op2, &seq, true, NULL); |
| 7406 | gimple_seq_add_seq (&stmt_list, seq); |
| 7407 | |
| 7408 | if (POINTER_TYPE_P (TREE_TYPE (comp_op2))) |
| 7409 | std::swap (a&: comp_op1, b&: comp_op2); |
| 7410 | |
| 7411 | if (POINTER_TYPE_P (TREE_TYPE (comp_op1))) |
| 7412 | { |
| 7413 | comp = fold_build_pointer_plus (comp_op1, |
| 7414 | fold_convert (sizetype, comp_op2)); |
| 7415 | comp = fold_build_pointer_plus (comp, |
| 7416 | wide_int_to_tree (sizetype, offset)); |
| 7417 | } |
| 7418 | else |
| 7419 | { |
| 7420 | comp = fold_build2 (PLUS_EXPR, TREE_TYPE (comp_op1), comp_op1, |
| 7421 | fold_convert (TREE_TYPE (comp_op1), comp_op2)); |
| 7422 | comp = fold_build2 (PLUS_EXPR, TREE_TYPE (comp_op1), comp, |
| 7423 | wide_int_to_tree (TREE_TYPE (comp_op1), offset)); |
| 7424 | } |
| 7425 | |
| 7426 | comp = fold_convert (type, comp); |
| 7427 | comp = force_gimple_operand (comp, &seq, false, NULL); |
| 7428 | gimple_seq_add_seq (&stmt_list, seq); |
| 7429 | if (gimple_code (g: use->stmt) != GIMPLE_PHI |
| 7430 | /* We can't allow re-allocating the stmt as it might be pointed |
| 7431 | to still. */ |
| 7432 | && (get_gimple_rhs_num_ops (TREE_CODE (comp)) |
| 7433 | >= gimple_num_ops (gs: gsi_stmt (i: bsi)))) |
| 7434 | { |
| 7435 | comp = force_gimple_operand (comp, &seq, true, NULL); |
| 7436 | gimple_seq_add_seq (&stmt_list, seq); |
| 7437 | if (POINTER_TYPE_P (TREE_TYPE (tgt))) |
| 7438 | { |
| 7439 | duplicate_ssa_name_ptr_info (comp, SSA_NAME_PTR_INFO (tgt)); |
| 7440 | /* As this isn't a plain copy we have to reset alignment |
| 7441 | information. */ |
| 7442 | if (SSA_NAME_PTR_INFO (comp)) |
| 7443 | mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (comp)); |
| 7444 | } |
| 7445 | } |
| 7446 | |
| 7447 | gsi_insert_seq_before (&bsi, stmt_list, GSI_SAME_STMT); |
| 7448 | if (gimple_code (g: use->stmt) == GIMPLE_PHI) |
| 7449 | { |
| 7450 | ass = gimple_build_assign (tgt, comp); |
| 7451 | gsi_insert_before (&bsi, ass, GSI_SAME_STMT); |
| 7452 | |
| 7453 | bsi = gsi_for_stmt (use->stmt); |
| 7454 | remove_phi_node (&bsi, false); |
| 7455 | } |
| 7456 | else |
| 7457 | { |
| 7458 | gimple_assign_set_rhs_from_tree (&bsi, comp); |
| 7459 | use->stmt = gsi_stmt (i: bsi); |
| 7460 | } |
| 7461 | } |
| 7462 | |
| 7463 | /* Performs a peephole optimization to reorder the iv update statement with |
| 7464 | a mem ref to enable instruction combining in later phases. The mem ref uses |
| 7465 | the iv value before the update, so the reordering transformation requires |
| 7466 | adjustment of the offset. CAND is the selected IV_CAND. |
| 7467 | |
| 7468 | Example: |
| 7469 | |
| 7470 | t = MEM_REF (base, iv1, 8, 16); // base, index, stride, offset |
| 7471 | iv2 = iv1 + 1; |
| 7472 | |
| 7473 | if (t < val) (1) |
| 7474 | goto L; |
| 7475 | goto Head; |
| 7476 | |
| 7477 | |
| 7478 | directly propagating t over to (1) will introduce overlapping live range |
| 7479 | thus increase register pressure. This peephole transform it into: |
| 7480 | |
| 7481 | |
| 7482 | iv2 = iv1 + 1; |
| 7483 | t = MEM_REF (base, iv2, 8, 8); |
| 7484 | if (t < val) |
| 7485 | goto L; |
| 7486 | goto Head; |
| 7487 | */ |
| 7488 | |
| 7489 | static void |
| 7490 | adjust_iv_update_pos (struct iv_cand *cand, struct iv_use *use) |
| 7491 | { |
| 7492 | tree var_after; |
| 7493 | gimple *iv_update, *stmt; |
| 7494 | basic_block bb; |
| 7495 | gimple_stmt_iterator gsi, gsi_iv; |
| 7496 | |
| 7497 | if (cand->pos != IP_NORMAL) |
| 7498 | return; |
| 7499 | |
| 7500 | var_after = cand->var_after; |
| 7501 | iv_update = SSA_NAME_DEF_STMT (var_after); |
| 7502 | |
| 7503 | bb = gimple_bb (g: iv_update); |
| 7504 | gsi = gsi_last_nondebug_bb (bb); |
| 7505 | stmt = gsi_stmt (i: gsi); |
| 7506 | |
| 7507 | /* Only handle conditional statement for now. */ |
| 7508 | if (gimple_code (g: stmt) != GIMPLE_COND) |
| 7509 | return; |
| 7510 | |
| 7511 | gsi_prev_nondebug (i: &gsi); |
| 7512 | stmt = gsi_stmt (i: gsi); |
| 7513 | if (stmt != iv_update) |
| 7514 | return; |
| 7515 | |
| 7516 | gsi_prev_nondebug (i: &gsi); |
| 7517 | if (gsi_end_p (i: gsi)) |
| 7518 | return; |
| 7519 | |
| 7520 | stmt = gsi_stmt (i: gsi); |
| 7521 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
| 7522 | return; |
| 7523 | |
| 7524 | if (stmt != use->stmt) |
| 7525 | return; |
| 7526 | |
| 7527 | if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) |
| 7528 | return; |
| 7529 | |
| 7530 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 7531 | { |
| 7532 | fprintf (stream: dump_file, format: "Reordering \n" ); |
| 7533 | print_gimple_stmt (dump_file, iv_update, 0); |
| 7534 | print_gimple_stmt (dump_file, use->stmt, 0); |
| 7535 | fprintf (stream: dump_file, format: "\n" ); |
| 7536 | } |
| 7537 | |
| 7538 | gsi = gsi_for_stmt (use->stmt); |
| 7539 | gsi_iv = gsi_for_stmt (iv_update); |
| 7540 | gsi_move_before (&gsi_iv, &gsi); |
| 7541 | |
| 7542 | cand->pos = IP_BEFORE_USE; |
| 7543 | cand->incremented_at = use->stmt; |
| 7544 | } |
| 7545 | |
| 7546 | /* Return the alias pointer type that should be used for a MEM_REF |
| 7547 | associated with USE, which has type USE_PTR_ADDRESS. */ |
| 7548 | |
| 7549 | static tree |
| 7550 | get_alias_ptr_type_for_ptr_address (iv_use *use) |
| 7551 | { |
| 7552 | gcall *call = as_a <gcall *> (p: use->stmt); |
| 7553 | switch (gimple_call_internal_fn (gs: call)) |
| 7554 | { |
| 7555 | case IFN_MASK_LOAD: |
| 7556 | case IFN_MASK_STORE: |
| 7557 | case IFN_MASK_LOAD_LANES: |
| 7558 | case IFN_MASK_STORE_LANES: |
| 7559 | case IFN_MASK_LEN_LOAD_LANES: |
| 7560 | case IFN_MASK_LEN_STORE_LANES: |
| 7561 | case IFN_LEN_LOAD: |
| 7562 | case IFN_LEN_STORE: |
| 7563 | case IFN_MASK_LEN_LOAD: |
| 7564 | case IFN_MASK_LEN_STORE: |
| 7565 | /* The second argument contains the correct alias type. */ |
| 7566 | gcc_assert (use->op_p == gimple_call_arg_ptr (call, 0)); |
| 7567 | return TREE_TYPE (gimple_call_arg (call, 1)); |
| 7568 | |
| 7569 | default: |
| 7570 | gcc_unreachable (); |
| 7571 | } |
| 7572 | } |
| 7573 | |
| 7574 | |
| 7575 | /* Rewrites USE (address that is an iv) using candidate CAND. */ |
| 7576 | |
| 7577 | static void |
| 7578 | rewrite_use_address (struct ivopts_data *data, |
| 7579 | struct iv_use *use, struct iv_cand *cand) |
| 7580 | { |
| 7581 | aff_tree aff; |
| 7582 | bool ok; |
| 7583 | |
| 7584 | adjust_iv_update_pos (cand, use); |
| 7585 | ok = get_computation_aff (data, at: use->stmt, use, cand, aff: &aff); |
| 7586 | gcc_assert (ok); |
| 7587 | unshare_aff_combination (&aff); |
| 7588 | |
| 7589 | /* To avoid undefined overflow problems, all IV candidates use unsigned |
| 7590 | integer types. The drawback is that this makes it impossible for |
| 7591 | create_mem_ref to distinguish an IV that is based on a memory object |
| 7592 | from one that represents simply an offset. |
| 7593 | |
| 7594 | To work around this problem, we pass a hint to create_mem_ref that |
| 7595 | indicates which variable (if any) in aff is an IV based on a memory |
| 7596 | object. Note that we only consider the candidate. If this is not |
| 7597 | based on an object, the base of the reference is in some subexpression |
| 7598 | of the use -- but these will use pointer types, so they are recognized |
| 7599 | by the create_mem_ref heuristics anyway. */ |
| 7600 | tree iv = var_at_stmt (loop: data->current_loop, cand, stmt: use->stmt); |
| 7601 | tree base_hint = (cand->iv->base_object) ? iv : NULL_TREE; |
| 7602 | gimple_stmt_iterator bsi = gsi_for_stmt (use->stmt); |
| 7603 | tree type = use->mem_type; |
| 7604 | tree alias_ptr_type; |
| 7605 | if (use->type == USE_PTR_ADDRESS) |
| 7606 | alias_ptr_type = get_alias_ptr_type_for_ptr_address (use); |
| 7607 | else |
| 7608 | { |
| 7609 | gcc_assert (type == TREE_TYPE (*use->op_p)); |
| 7610 | unsigned int align = get_object_alignment (*use->op_p); |
| 7611 | if (align != TYPE_ALIGN (type)) |
| 7612 | type = build_aligned_type (type, align); |
| 7613 | alias_ptr_type = reference_alias_ptr_type (*use->op_p); |
| 7614 | } |
| 7615 | tree ref = create_mem_ref (&bsi, type, &aff, alias_ptr_type, |
| 7616 | iv, base_hint, data->speed); |
| 7617 | |
| 7618 | if (use->type == USE_PTR_ADDRESS) |
| 7619 | { |
| 7620 | ref = fold_build1 (ADDR_EXPR, build_pointer_type (use->mem_type), ref); |
| 7621 | ref = fold_convert (get_use_type (use), ref); |
| 7622 | ref = force_gimple_operand_gsi (&bsi, ref, true, NULL_TREE, |
| 7623 | true, GSI_SAME_STMT); |
| 7624 | } |
| 7625 | else |
| 7626 | { |
| 7627 | /* When we end up confused enough and have no suitable base but |
| 7628 | stuffed everything to index2 use a LEA for the address and |
| 7629 | create a plain MEM_REF to avoid basing a memory reference |
| 7630 | on address zero which create_mem_ref_raw does as fallback. */ |
| 7631 | if (TREE_CODE (ref) == TARGET_MEM_REF |
| 7632 | && TMR_INDEX2 (ref) != NULL_TREE |
| 7633 | && integer_zerop (TREE_OPERAND (ref, 0))) |
| 7634 | { |
| 7635 | ref = fold_build1 (ADDR_EXPR, TREE_TYPE (TREE_OPERAND (ref, 0)), ref); |
| 7636 | ref = force_gimple_operand_gsi (&bsi, ref, true, NULL_TREE, |
| 7637 | true, GSI_SAME_STMT); |
| 7638 | ref = build2 (MEM_REF, type, ref, build_zero_cst (alias_ptr_type)); |
| 7639 | } |
| 7640 | copy_ref_info (ref, *use->op_p); |
| 7641 | } |
| 7642 | |
| 7643 | *use->op_p = ref; |
| 7644 | } |
| 7645 | |
| 7646 | /* Rewrites USE (the condition such that one of the arguments is an iv) using |
| 7647 | candidate CAND. */ |
| 7648 | |
| 7649 | static void |
| 7650 | rewrite_use_compare (struct ivopts_data *data, |
| 7651 | struct iv_use *use, struct iv_cand *cand) |
| 7652 | { |
| 7653 | tree comp, op, bound; |
| 7654 | gimple_stmt_iterator bsi = gsi_for_stmt (use->stmt); |
| 7655 | enum tree_code compare; |
| 7656 | struct iv_group *group = data->vgroups[use->group_id]; |
| 7657 | class cost_pair *cp = get_group_iv_cost (data, group, cand); |
| 7658 | |
| 7659 | bound = cp->value; |
| 7660 | if (bound) |
| 7661 | { |
| 7662 | tree var = var_at_stmt (loop: data->current_loop, cand, stmt: use->stmt); |
| 7663 | tree var_type = TREE_TYPE (var); |
| 7664 | gimple_seq stmts; |
| 7665 | |
| 7666 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 7667 | { |
| 7668 | fprintf (stream: dump_file, format: "Replacing exit test: " ); |
| 7669 | print_gimple_stmt (dump_file, use->stmt, 0, TDF_SLIM); |
| 7670 | } |
| 7671 | compare = cp->comp; |
| 7672 | bound = unshare_expr (fold_convert (var_type, bound)); |
| 7673 | op = force_gimple_operand (bound, &stmts, true, NULL_TREE); |
| 7674 | if (stmts) |
| 7675 | gsi_insert_seq_on_edge_immediate ( |
| 7676 | loop_preheader_edge (data->current_loop), |
| 7677 | stmts); |
| 7678 | |
| 7679 | gcond *cond_stmt = as_a <gcond *> (p: use->stmt); |
| 7680 | gimple_cond_set_lhs (gs: cond_stmt, lhs: var); |
| 7681 | gimple_cond_set_code (gs: cond_stmt, code: compare); |
| 7682 | gimple_cond_set_rhs (gs: cond_stmt, rhs: op); |
| 7683 | return; |
| 7684 | } |
| 7685 | |
| 7686 | /* The induction variable elimination failed; just express the original |
| 7687 | giv. */ |
| 7688 | comp = get_computation_at (data, at: use->stmt, use, cand); |
| 7689 | gcc_assert (comp != NULL_TREE); |
| 7690 | gcc_assert (use->op_p != NULL); |
| 7691 | *use->op_p = force_gimple_operand_gsi (&bsi, comp, true, |
| 7692 | SSA_NAME_VAR (*use->op_p), |
| 7693 | true, GSI_SAME_STMT); |
| 7694 | } |
| 7695 | |
| 7696 | /* Rewrite the groups using the selected induction variables. */ |
| 7697 | |
| 7698 | static void |
| 7699 | rewrite_groups (struct ivopts_data *data) |
| 7700 | { |
| 7701 | unsigned i, j; |
| 7702 | |
| 7703 | for (i = 0; i < data->vgroups.length (); i++) |
| 7704 | { |
| 7705 | struct iv_group *group = data->vgroups[i]; |
| 7706 | struct iv_cand *cand = group->selected; |
| 7707 | |
| 7708 | gcc_assert (cand); |
| 7709 | |
| 7710 | if (group->type == USE_NONLINEAR_EXPR) |
| 7711 | { |
| 7712 | for (j = 0; j < group->vuses.length (); j++) |
| 7713 | { |
| 7714 | rewrite_use_nonlinear_expr (data, use: group->vuses[j], cand); |
| 7715 | update_stmt (s: group->vuses[j]->stmt); |
| 7716 | } |
| 7717 | } |
| 7718 | else if (address_p (type: group->type)) |
| 7719 | { |
| 7720 | for (j = 0; j < group->vuses.length (); j++) |
| 7721 | { |
| 7722 | rewrite_use_address (data, use: group->vuses[j], cand); |
| 7723 | update_stmt (s: group->vuses[j]->stmt); |
| 7724 | } |
| 7725 | } |
| 7726 | else |
| 7727 | { |
| 7728 | gcc_assert (group->type == USE_COMPARE); |
| 7729 | |
| 7730 | for (j = 0; j < group->vuses.length (); j++) |
| 7731 | { |
| 7732 | rewrite_use_compare (data, use: group->vuses[j], cand); |
| 7733 | update_stmt (s: group->vuses[j]->stmt); |
| 7734 | } |
| 7735 | } |
| 7736 | } |
| 7737 | } |
| 7738 | |
| 7739 | /* Removes the ivs that are not used after rewriting. */ |
| 7740 | |
| 7741 | static void |
| 7742 | remove_unused_ivs (struct ivopts_data *data, bitmap toremove) |
| 7743 | { |
| 7744 | unsigned j; |
| 7745 | bitmap_iterator bi; |
| 7746 | |
| 7747 | /* Figure out an order in which to release SSA DEFs so that we don't |
| 7748 | release something that we'd have to propagate into a debug stmt |
| 7749 | afterwards. */ |
| 7750 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, j, bi) |
| 7751 | { |
| 7752 | struct version_info *info; |
| 7753 | |
| 7754 | info = ver_info (data, ver: j); |
| 7755 | if (info->iv |
| 7756 | && !integer_zerop (info->iv->step) |
| 7757 | && !info->inv_id |
| 7758 | && !info->iv->nonlin_use |
| 7759 | && !info->preserve_biv) |
| 7760 | { |
| 7761 | bitmap_set_bit (toremove, SSA_NAME_VERSION (info->iv->ssa_name)); |
| 7762 | |
| 7763 | tree def = info->iv->ssa_name; |
| 7764 | |
| 7765 | if (MAY_HAVE_DEBUG_BIND_STMTS && SSA_NAME_DEF_STMT (def)) |
| 7766 | { |
| 7767 | imm_use_iterator imm_iter; |
| 7768 | use_operand_p use_p; |
| 7769 | gimple *stmt; |
| 7770 | int count = 0; |
| 7771 | |
| 7772 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, def) |
| 7773 | { |
| 7774 | if (!gimple_debug_bind_p (s: stmt)) |
| 7775 | continue; |
| 7776 | |
| 7777 | /* We just want to determine whether to do nothing |
| 7778 | (count == 0), to substitute the computed |
| 7779 | expression into a single use of the SSA DEF by |
| 7780 | itself (count == 1), or to use a debug temp |
| 7781 | because the SSA DEF is used multiple times or as |
| 7782 | part of a larger expression (count > 1). */ |
| 7783 | count++; |
| 7784 | if (gimple_debug_bind_get_value (dbg: stmt) != def) |
| 7785 | count++; |
| 7786 | |
| 7787 | if (count > 1) |
| 7788 | break; |
| 7789 | } |
| 7790 | |
| 7791 | if (!count) |
| 7792 | continue; |
| 7793 | |
| 7794 | struct iv_use dummy_use; |
| 7795 | struct iv_cand *best_cand = NULL, *cand; |
| 7796 | unsigned i, best_pref = 0, cand_pref; |
| 7797 | tree comp = NULL_TREE; |
| 7798 | |
| 7799 | memset (s: &dummy_use, c: 0, n: sizeof (dummy_use)); |
| 7800 | dummy_use.iv = info->iv; |
| 7801 | for (i = 0; i < data->vgroups.length () && i < 64; i++) |
| 7802 | { |
| 7803 | cand = data->vgroups[i]->selected; |
| 7804 | if (cand == best_cand) |
| 7805 | continue; |
| 7806 | cand_pref = operand_equal_p (cand->iv->step, |
| 7807 | info->iv->step, flags: 0) |
| 7808 | ? 4 : 0; |
| 7809 | cand_pref |
| 7810 | += TYPE_MODE (TREE_TYPE (cand->iv->base)) |
| 7811 | == TYPE_MODE (TREE_TYPE (info->iv->base)) |
| 7812 | ? 2 : 0; |
| 7813 | cand_pref |
| 7814 | += TREE_CODE (cand->iv->base) == INTEGER_CST |
| 7815 | ? 1 : 0; |
| 7816 | if (best_cand == NULL || best_pref < cand_pref) |
| 7817 | { |
| 7818 | tree this_comp |
| 7819 | = get_debug_computation_at (data, |
| 7820 | SSA_NAME_DEF_STMT (def), |
| 7821 | use: &dummy_use, cand); |
| 7822 | if (this_comp) |
| 7823 | { |
| 7824 | best_cand = cand; |
| 7825 | best_pref = cand_pref; |
| 7826 | comp = this_comp; |
| 7827 | } |
| 7828 | } |
| 7829 | } |
| 7830 | |
| 7831 | if (!best_cand) |
| 7832 | continue; |
| 7833 | |
| 7834 | comp = unshare_expr (comp); |
| 7835 | if (count > 1) |
| 7836 | { |
| 7837 | tree vexpr = build_debug_expr_decl (TREE_TYPE (comp)); |
| 7838 | /* FIXME: Is setting the mode really necessary? */ |
| 7839 | if (SSA_NAME_VAR (def)) |
| 7840 | SET_DECL_MODE (vexpr, DECL_MODE (SSA_NAME_VAR (def))); |
| 7841 | else |
| 7842 | SET_DECL_MODE (vexpr, TYPE_MODE (TREE_TYPE (vexpr))); |
| 7843 | gdebug *def_temp |
| 7844 | = gimple_build_debug_bind (vexpr, comp, NULL); |
| 7845 | gimple_stmt_iterator gsi; |
| 7846 | |
| 7847 | if (gimple_code (SSA_NAME_DEF_STMT (def)) == GIMPLE_PHI) |
| 7848 | gsi = gsi_after_labels (bb: gimple_bb |
| 7849 | (SSA_NAME_DEF_STMT (def))); |
| 7850 | else |
| 7851 | gsi = gsi_for_stmt (SSA_NAME_DEF_STMT (def)); |
| 7852 | |
| 7853 | gsi_insert_before (&gsi, def_temp, GSI_SAME_STMT); |
| 7854 | comp = vexpr; |
| 7855 | } |
| 7856 | |
| 7857 | FOR_EACH_IMM_USE_STMT (stmt, imm_iter, def) |
| 7858 | { |
| 7859 | if (!gimple_debug_bind_p (s: stmt)) |
| 7860 | continue; |
| 7861 | |
| 7862 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) |
| 7863 | SET_USE (use_p, comp); |
| 7864 | |
| 7865 | update_stmt (s: stmt); |
| 7866 | } |
| 7867 | } |
| 7868 | } |
| 7869 | } |
| 7870 | } |
| 7871 | |
| 7872 | /* Frees memory occupied by class tree_niter_desc in *VALUE. Callback |
| 7873 | for hash_map::traverse. */ |
| 7874 | |
| 7875 | bool |
| 7876 | free_tree_niter_desc (edge const &, tree_niter_desc *const &value, void *) |
| 7877 | { |
| 7878 | if (value) |
| 7879 | { |
| 7880 | value->~tree_niter_desc (); |
| 7881 | free (ptr: value); |
| 7882 | } |
| 7883 | return true; |
| 7884 | } |
| 7885 | |
| 7886 | /* Frees data allocated by the optimization of a single loop. */ |
| 7887 | |
| 7888 | static void |
| 7889 | free_loop_data (struct ivopts_data *data) |
| 7890 | { |
| 7891 | unsigned i, j; |
| 7892 | bitmap_iterator bi; |
| 7893 | tree obj; |
| 7894 | |
| 7895 | if (data->niters) |
| 7896 | { |
| 7897 | data->niters->traverse<void *, free_tree_niter_desc> (NULL); |
| 7898 | delete data->niters; |
| 7899 | data->niters = NULL; |
| 7900 | } |
| 7901 | |
| 7902 | EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, i, bi) |
| 7903 | { |
| 7904 | struct version_info *info; |
| 7905 | |
| 7906 | info = ver_info (data, ver: i); |
| 7907 | info->iv = NULL; |
| 7908 | info->has_nonlin_use = false; |
| 7909 | info->preserve_biv = false; |
| 7910 | info->inv_id = 0; |
| 7911 | } |
| 7912 | bitmap_clear (data->relevant); |
| 7913 | bitmap_clear (data->important_candidates); |
| 7914 | |
| 7915 | for (i = 0; i < data->vgroups.length (); i++) |
| 7916 | { |
| 7917 | struct iv_group *group = data->vgroups[i]; |
| 7918 | |
| 7919 | for (j = 0; j < group->vuses.length (); j++) |
| 7920 | free (ptr: group->vuses[j]); |
| 7921 | group->vuses.release (); |
| 7922 | |
| 7923 | BITMAP_FREE (group->related_cands); |
| 7924 | for (j = 0; j < group->n_map_members; j++) |
| 7925 | { |
| 7926 | if (group->cost_map[j].inv_vars) |
| 7927 | BITMAP_FREE (group->cost_map[j].inv_vars); |
| 7928 | if (group->cost_map[j].inv_exprs) |
| 7929 | BITMAP_FREE (group->cost_map[j].inv_exprs); |
| 7930 | } |
| 7931 | |
| 7932 | free (ptr: group->cost_map); |
| 7933 | free (ptr: group); |
| 7934 | } |
| 7935 | data->vgroups.truncate (size: 0); |
| 7936 | |
| 7937 | for (i = 0; i < data->vcands.length (); i++) |
| 7938 | { |
| 7939 | struct iv_cand *cand = data->vcands[i]; |
| 7940 | |
| 7941 | if (cand->inv_vars) |
| 7942 | BITMAP_FREE (cand->inv_vars); |
| 7943 | if (cand->inv_exprs) |
| 7944 | BITMAP_FREE (cand->inv_exprs); |
| 7945 | free (ptr: cand); |
| 7946 | } |
| 7947 | data->vcands.truncate (size: 0); |
| 7948 | |
| 7949 | if (data->version_info_size < num_ssa_names) |
| 7950 | { |
| 7951 | data->version_info_size = 2 * num_ssa_names; |
| 7952 | free (ptr: data->version_info); |
| 7953 | data->version_info = XCNEWVEC (struct version_info, data->version_info_size); |
| 7954 | } |
| 7955 | |
| 7956 | data->max_inv_var_id = 0; |
| 7957 | data->max_inv_expr_id = 0; |
| 7958 | |
| 7959 | FOR_EACH_VEC_ELT (decl_rtl_to_reset, i, obj) |
| 7960 | SET_DECL_RTL (obj, NULL_RTX); |
| 7961 | |
| 7962 | decl_rtl_to_reset.truncate (size: 0); |
| 7963 | |
| 7964 | data->inv_expr_tab->empty (); |
| 7965 | |
| 7966 | data->iv_common_cand_tab->empty (); |
| 7967 | data->iv_common_cands.truncate (size: 0); |
| 7968 | } |
| 7969 | |
| 7970 | /* Finalizes data structures used by the iv optimization pass. LOOPS is the |
| 7971 | loop tree. */ |
| 7972 | |
| 7973 | static void |
| 7974 | tree_ssa_iv_optimize_finalize (struct ivopts_data *data) |
| 7975 | { |
| 7976 | free_loop_data (data); |
| 7977 | free (ptr: data->version_info); |
| 7978 | BITMAP_FREE (data->relevant); |
| 7979 | BITMAP_FREE (data->important_candidates); |
| 7980 | |
| 7981 | decl_rtl_to_reset.release (); |
| 7982 | data->vgroups.release (); |
| 7983 | data->vcands.release (); |
| 7984 | delete data->inv_expr_tab; |
| 7985 | data->inv_expr_tab = NULL; |
| 7986 | free_affine_expand_cache (&data->name_expansion_cache); |
| 7987 | if (data->base_object_map) |
| 7988 | delete data->base_object_map; |
| 7989 | delete data->iv_common_cand_tab; |
| 7990 | data->iv_common_cand_tab = NULL; |
| 7991 | data->iv_common_cands.release (); |
| 7992 | obstack_free (&data->iv_obstack, NULL); |
| 7993 | } |
| 7994 | |
| 7995 | /* Returns true if the loop body BODY includes any function calls. */ |
| 7996 | |
| 7997 | static bool |
| 7998 | loop_body_includes_call (basic_block *body, unsigned num_nodes) |
| 7999 | { |
| 8000 | gimple_stmt_iterator gsi; |
| 8001 | unsigned i; |
| 8002 | |
| 8003 | for (i = 0; i < num_nodes; i++) |
| 8004 | for (gsi = gsi_start_bb (bb: body[i]); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
| 8005 | { |
| 8006 | gimple *stmt = gsi_stmt (i: gsi); |
| 8007 | if (is_gimple_call (gs: stmt) |
| 8008 | && !gimple_call_internal_p (gs: stmt) |
| 8009 | && !is_inexpensive_builtin (gimple_call_fndecl (gs: stmt))) |
| 8010 | return true; |
| 8011 | } |
| 8012 | return false; |
| 8013 | } |
| 8014 | |
| 8015 | /* Determine cost scaling factor for basic blocks in loop. */ |
| 8016 | #define COST_SCALING_FACTOR_BOUND (20) |
| 8017 | |
| 8018 | static void |
| 8019 | determine_scaling_factor (struct ivopts_data *data, basic_block *body) |
| 8020 | { |
| 8021 | int lfreq = data->current_loop->header->count.to_frequency (cfun); |
| 8022 | if (!data->speed || lfreq <= 0) |
| 8023 | return; |
| 8024 | |
| 8025 | int max_freq = lfreq; |
| 8026 | for (unsigned i = 0; i < data->current_loop->num_nodes; i++) |
| 8027 | { |
| 8028 | body[i]->aux = (void *)(intptr_t) 1; |
| 8029 | if (max_freq < body[i]->count.to_frequency (cfun)) |
| 8030 | max_freq = body[i]->count.to_frequency (cfun); |
| 8031 | } |
| 8032 | if (max_freq > lfreq) |
| 8033 | { |
| 8034 | int divisor, factor; |
| 8035 | /* Check if scaling factor itself needs to be scaled by the bound. This |
| 8036 | is to avoid overflow when scaling cost according to profile info. */ |
| 8037 | if (max_freq / lfreq > COST_SCALING_FACTOR_BOUND) |
| 8038 | { |
| 8039 | divisor = max_freq; |
| 8040 | factor = COST_SCALING_FACTOR_BOUND; |
| 8041 | } |
| 8042 | else |
| 8043 | { |
| 8044 | divisor = lfreq; |
| 8045 | factor = 1; |
| 8046 | } |
| 8047 | for (unsigned i = 0; i < data->current_loop->num_nodes; i++) |
| 8048 | { |
| 8049 | int bfreq = body[i]->count.to_frequency (cfun); |
| 8050 | if (bfreq <= lfreq) |
| 8051 | continue; |
| 8052 | |
| 8053 | body[i]->aux = (void*)(intptr_t) (factor * bfreq / divisor); |
| 8054 | } |
| 8055 | } |
| 8056 | } |
| 8057 | |
| 8058 | /* Find doloop comparison use and set its doloop_p on if found. */ |
| 8059 | |
| 8060 | static bool |
| 8061 | find_doloop_use (struct ivopts_data *data) |
| 8062 | { |
| 8063 | struct loop *loop = data->current_loop; |
| 8064 | |
| 8065 | for (unsigned i = 0; i < data->vgroups.length (); i++) |
| 8066 | { |
| 8067 | struct iv_group *group = data->vgroups[i]; |
| 8068 | if (group->type == USE_COMPARE) |
| 8069 | { |
| 8070 | gcc_assert (group->vuses.length () == 1); |
| 8071 | struct iv_use *use = group->vuses[0]; |
| 8072 | gimple *stmt = use->stmt; |
| 8073 | if (gimple_code (g: stmt) == GIMPLE_COND) |
| 8074 | { |
| 8075 | basic_block bb = gimple_bb (g: stmt); |
| 8076 | edge true_edge, false_edge; |
| 8077 | extract_true_false_edges_from_block (bb, &true_edge, &false_edge); |
| 8078 | /* This comparison is used for loop latch. Require latch is empty |
| 8079 | for now. */ |
| 8080 | if ((loop->latch == true_edge->dest |
| 8081 | || loop->latch == false_edge->dest) |
| 8082 | && empty_block_p (loop->latch)) |
| 8083 | { |
| 8084 | group->doloop_p = true; |
| 8085 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 8086 | { |
| 8087 | fprintf (stream: dump_file, format: "Doloop cmp iv use: " ); |
| 8088 | print_gimple_stmt (dump_file, stmt, TDF_DETAILS); |
| 8089 | } |
| 8090 | return true; |
| 8091 | } |
| 8092 | } |
| 8093 | } |
| 8094 | } |
| 8095 | |
| 8096 | return false; |
| 8097 | } |
| 8098 | |
| 8099 | /* For the targets which support doloop, to predict whether later RTL doloop |
| 8100 | transformation will perform on this loop, further detect the doloop use and |
| 8101 | mark the flag doloop_use_p if predicted. */ |
| 8102 | |
| 8103 | void |
| 8104 | analyze_and_mark_doloop_use (struct ivopts_data *data) |
| 8105 | { |
| 8106 | data->doloop_use_p = false; |
| 8107 | |
| 8108 | if (!flag_branch_on_count_reg) |
| 8109 | return; |
| 8110 | |
| 8111 | if (data->current_loop->unroll == USHRT_MAX) |
| 8112 | return; |
| 8113 | |
| 8114 | if (!generic_predict_doloop_p (data)) |
| 8115 | return; |
| 8116 | |
| 8117 | if (find_doloop_use (data)) |
| 8118 | { |
| 8119 | data->doloop_use_p = true; |
| 8120 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 8121 | { |
| 8122 | struct loop *loop = data->current_loop; |
| 8123 | fprintf (stream: dump_file, |
| 8124 | format: "Predict loop %d can perform" |
| 8125 | " doloop optimization later.\n" , |
| 8126 | loop->num); |
| 8127 | flow_loop_dump (loop, dump_file, NULL, 1); |
| 8128 | } |
| 8129 | } |
| 8130 | } |
| 8131 | |
| 8132 | /* Optimizes the LOOP. Returns true if anything changed. */ |
| 8133 | |
| 8134 | static bool |
| 8135 | tree_ssa_iv_optimize_loop (struct ivopts_data *data, class loop *loop, |
| 8136 | bitmap toremove) |
| 8137 | { |
| 8138 | bool changed = false; |
| 8139 | class iv_ca *iv_ca; |
| 8140 | edge exit = single_dom_exit (loop); |
| 8141 | basic_block *body; |
| 8142 | |
| 8143 | gcc_assert (!data->niters); |
| 8144 | data->current_loop = loop; |
| 8145 | data->loop_loc = find_loop_location (loop).get_location_t (); |
| 8146 | data->speed = optimize_loop_for_speed_p (loop); |
| 8147 | |
| 8148 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 8149 | { |
| 8150 | fprintf (stream: dump_file, format: "Processing loop %d" , loop->num); |
| 8151 | if (data->loop_loc != UNKNOWN_LOCATION) |
| 8152 | fprintf (stream: dump_file, format: " at %s:%d" , LOCATION_FILE (data->loop_loc), |
| 8153 | LOCATION_LINE (data->loop_loc)); |
| 8154 | fprintf (stream: dump_file, format: "\n" ); |
| 8155 | |
| 8156 | if (exit) |
| 8157 | { |
| 8158 | fprintf (stream: dump_file, format: " single exit %d -> %d, exit condition " , |
| 8159 | exit->src->index, exit->dest->index); |
| 8160 | print_gimple_stmt (dump_file, *gsi_last_bb (bb: exit->src), |
| 8161 | 0, TDF_SLIM); |
| 8162 | fprintf (stream: dump_file, format: "\n" ); |
| 8163 | } |
| 8164 | |
| 8165 | fprintf (stream: dump_file, format: "\n" ); |
| 8166 | } |
| 8167 | |
| 8168 | body = get_loop_body (loop); |
| 8169 | data->body_includes_call = loop_body_includes_call (body, num_nodes: loop->num_nodes); |
| 8170 | renumber_gimple_stmt_uids_in_blocks (body, loop->num_nodes); |
| 8171 | |
| 8172 | data->loop_single_exit_p |
| 8173 | = exit != NULL && loop_only_exit_p (loop, body, exit); |
| 8174 | |
| 8175 | /* For each ssa name determines whether it behaves as an induction variable |
| 8176 | in some loop. */ |
| 8177 | if (!find_induction_variables (data, body)) |
| 8178 | goto finish; |
| 8179 | |
| 8180 | /* Finds interesting uses (item 1). */ |
| 8181 | find_interesting_uses (data, body); |
| 8182 | if (data->vgroups.length () > MAX_CONSIDERED_GROUPS) |
| 8183 | goto finish; |
| 8184 | |
| 8185 | /* Determine cost scaling factor for basic blocks in loop. */ |
| 8186 | determine_scaling_factor (data, body); |
| 8187 | |
| 8188 | /* Analyze doloop possibility and mark the doloop use if predicted. */ |
| 8189 | analyze_and_mark_doloop_use (data); |
| 8190 | |
| 8191 | /* Finds candidates for the induction variables (item 2). */ |
| 8192 | find_iv_candidates (data); |
| 8193 | |
| 8194 | /* Calculates the costs (item 3, part 1). */ |
| 8195 | determine_iv_costs (data); |
| 8196 | determine_group_iv_costs (data); |
| 8197 | determine_set_costs (data); |
| 8198 | |
| 8199 | /* Find the optimal set of induction variables (item 3, part 2). */ |
| 8200 | iv_ca = find_optimal_iv_set (data); |
| 8201 | /* Cleanup basic block aux field. */ |
| 8202 | for (unsigned i = 0; i < data->current_loop->num_nodes; i++) |
| 8203 | body[i]->aux = NULL; |
| 8204 | if (!iv_ca) |
| 8205 | goto finish; |
| 8206 | changed = true; |
| 8207 | |
| 8208 | /* Create the new induction variables (item 4, part 1). */ |
| 8209 | create_new_ivs (data, set: iv_ca); |
| 8210 | iv_ca_free (ivs: &iv_ca); |
| 8211 | |
| 8212 | /* Rewrite the uses (item 4, part 2). */ |
| 8213 | rewrite_groups (data); |
| 8214 | |
| 8215 | /* Remove the ivs that are unused after rewriting. */ |
| 8216 | remove_unused_ivs (data, toremove); |
| 8217 | |
| 8218 | finish: |
| 8219 | free (ptr: body); |
| 8220 | free_loop_data (data); |
| 8221 | |
| 8222 | return changed; |
| 8223 | } |
| 8224 | |
| 8225 | /* Main entry point. Optimizes induction variables in loops. */ |
| 8226 | |
| 8227 | void |
| 8228 | tree_ssa_iv_optimize (void) |
| 8229 | { |
| 8230 | struct ivopts_data data; |
| 8231 | auto_bitmap toremove; |
| 8232 | |
| 8233 | tree_ssa_iv_optimize_init (data: &data); |
| 8234 | mark_ssa_maybe_undefs (); |
| 8235 | |
| 8236 | /* Optimize the loops starting with the innermost ones. */ |
| 8237 | for (auto loop : loops_list (cfun, LI_FROM_INNERMOST)) |
| 8238 | { |
| 8239 | if (!dbg_cnt (index: ivopts_loop)) |
| 8240 | continue; |
| 8241 | |
| 8242 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 8243 | flow_loop_dump (loop, dump_file, NULL, 1); |
| 8244 | |
| 8245 | tree_ssa_iv_optimize_loop (data: &data, loop, toremove); |
| 8246 | } |
| 8247 | |
| 8248 | /* Remove eliminated IV defs. */ |
| 8249 | release_defs_bitset (toremove); |
| 8250 | |
| 8251 | /* We have changed the structure of induction variables; it might happen |
| 8252 | that definitions in the scev database refer to some of them that were |
| 8253 | eliminated. */ |
| 8254 | scev_reset_htab (); |
| 8255 | /* Likewise niter and control-IV information. */ |
| 8256 | free_numbers_of_iterations_estimates (cfun); |
| 8257 | |
| 8258 | tree_ssa_iv_optimize_finalize (data: &data); |
| 8259 | } |
| 8260 | |
| 8261 | #include "gt-tree-ssa-loop-ivopts.h" |
| 8262 | |