| 1 | /* IRA hard register and memory cost calculation for allocnos or pseudos. |
| 2 | Copyright (C) 2006-2025 Free Software Foundation, Inc. |
| 3 | Contributed by Vladimir Makarov <vmakarov@redhat.com>. |
| 4 | |
| 5 | This file is part of GCC. |
| 6 | |
| 7 | GCC is free software; you can redistribute it and/or modify it under |
| 8 | the terms of the GNU General Public License as published by the Free |
| 9 | Software Foundation; either version 3, or (at your option) any later |
| 10 | version. |
| 11 | |
| 12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
| 13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
| 14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| 15 | for more details. |
| 16 | |
| 17 | You should have received a copy of the GNU General Public License |
| 18 | along with GCC; see the file COPYING3. If not see |
| 19 | <http://www.gnu.org/licenses/>. */ |
| 20 | |
| 21 | #include "config.h" |
| 22 | #include "system.h" |
| 23 | #include "coretypes.h" |
| 24 | #include "backend.h" |
| 25 | #include "target.h" |
| 26 | #include "rtl.h" |
| 27 | #include "tree.h" |
| 28 | #include "predict.h" |
| 29 | #include "memmodel.h" |
| 30 | #include "tm_p.h" |
| 31 | #include "insn-config.h" |
| 32 | #include "regs.h" |
| 33 | #include "regset.h" |
| 34 | #include "ira.h" |
| 35 | #include "ira-int.h" |
| 36 | #include "addresses.h" |
| 37 | #include "reload.h" |
| 38 | #include "print-rtl.h" |
| 39 | |
| 40 | /* The flags is set up every time when we calculate pseudo register |
| 41 | classes through function ira_set_pseudo_classes. */ |
| 42 | static bool pseudo_classes_defined_p = false; |
| 43 | |
| 44 | /* TRUE if we work with allocnos. Otherwise we work with pseudos. */ |
| 45 | static bool allocno_p; |
| 46 | |
| 47 | /* Number of elements in array `costs'. */ |
| 48 | static int cost_elements_num; |
| 49 | |
| 50 | /* The `costs' struct records the cost of using hard registers of each |
| 51 | class considered for the calculation and of using memory for each |
| 52 | allocno or pseudo. */ |
| 53 | struct costs |
| 54 | { |
| 55 | int mem_cost; |
| 56 | /* Costs for register classes start here. We process only some |
| 57 | allocno classes. */ |
| 58 | int cost[1]; |
| 59 | }; |
| 60 | |
| 61 | #define max_struct_costs_size \ |
| 62 | (this_target_ira_int->x_max_struct_costs_size) |
| 63 | #define init_cost \ |
| 64 | (this_target_ira_int->x_init_cost) |
| 65 | #define temp_costs \ |
| 66 | (this_target_ira_int->x_temp_costs) |
| 67 | #define op_costs \ |
| 68 | (this_target_ira_int->x_op_costs) |
| 69 | #define this_op_costs \ |
| 70 | (this_target_ira_int->x_this_op_costs) |
| 71 | |
| 72 | /* Costs of each class for each allocno or pseudo. */ |
| 73 | static struct costs *costs; |
| 74 | |
| 75 | /* Accumulated costs of each class for each allocno. */ |
| 76 | static struct costs *total_allocno_costs; |
| 77 | |
| 78 | /* It is the current size of struct costs. */ |
| 79 | static size_t struct_costs_size; |
| 80 | |
| 81 | /* Return pointer to structure containing costs of allocno or pseudo |
| 82 | with given NUM in array ARR. */ |
| 83 | #define COSTS(arr, num) \ |
| 84 | ((struct costs *) ((char *) (arr) + (num) * struct_costs_size)) |
| 85 | |
| 86 | /* Return index in COSTS when processing reg with REGNO. */ |
| 87 | #define COST_INDEX(regno) (allocno_p \ |
| 88 | ? ALLOCNO_NUM (ira_curr_regno_allocno_map[regno]) \ |
| 89 | : (int) regno) |
| 90 | |
| 91 | /* Record register class preferences of each allocno or pseudo. Null |
| 92 | value means no preferences. It happens on the 1st iteration of the |
| 93 | cost calculation. */ |
| 94 | static enum reg_class *pref; |
| 95 | |
| 96 | /* Allocated buffers for pref. */ |
| 97 | static enum reg_class *pref_buffer; |
| 98 | |
| 99 | /* Record allocno class of each allocno with the same regno. */ |
| 100 | static enum reg_class *regno_aclass; |
| 101 | |
| 102 | /* Record cost gains for not allocating a register with an invariant |
| 103 | equivalence. */ |
| 104 | static int *regno_equiv_gains; |
| 105 | |
| 106 | /* Execution frequency of the current insn. */ |
| 107 | static int frequency; |
| 108 | |
| 109 | |
| 110 | |
| 111 | /* Info about reg classes whose costs are calculated for a pseudo. */ |
| 112 | struct cost_classes |
| 113 | { |
| 114 | /* Number of the cost classes in the subsequent array. */ |
| 115 | int num; |
| 116 | /* Container of the cost classes. */ |
| 117 | enum reg_class classes[N_REG_CLASSES]; |
| 118 | /* Map reg class -> index of the reg class in the previous array. |
| 119 | -1 if it is not a cost class. */ |
| 120 | int index[N_REG_CLASSES]; |
| 121 | /* Map hard regno index of first class in array CLASSES containing |
| 122 | the hard regno, -1 otherwise. */ |
| 123 | int hard_regno_index[FIRST_PSEUDO_REGISTER]; |
| 124 | }; |
| 125 | |
| 126 | /* Types of pointers to the structure above. */ |
| 127 | typedef struct cost_classes *cost_classes_t; |
| 128 | typedef const struct cost_classes *const_cost_classes_t; |
| 129 | |
| 130 | /* Info about cost classes for each pseudo. */ |
| 131 | static cost_classes_t *regno_cost_classes; |
| 132 | |
| 133 | /* Helper for cost_classes hashing. */ |
| 134 | |
| 135 | struct cost_classes_hasher : pointer_hash <cost_classes> |
| 136 | { |
| 137 | static inline hashval_t hash (const cost_classes *); |
| 138 | static inline bool equal (const cost_classes *, const cost_classes *); |
| 139 | static inline void remove (cost_classes *); |
| 140 | }; |
| 141 | |
| 142 | /* Returns hash value for cost classes info HV. */ |
| 143 | inline hashval_t |
| 144 | cost_classes_hasher::hash (const cost_classes *hv) |
| 145 | { |
| 146 | return iterative_hash (&hv->classes, sizeof (enum reg_class) * hv->num, 0); |
| 147 | } |
| 148 | |
| 149 | /* Compares cost classes info HV1 and HV2. */ |
| 150 | inline bool |
| 151 | cost_classes_hasher::equal (const cost_classes *hv1, const cost_classes *hv2) |
| 152 | { |
| 153 | return (hv1->num == hv2->num |
| 154 | && memcmp (s1: hv1->classes, s2: hv2->classes, |
| 155 | n: sizeof (enum reg_class) * hv1->num) == 0); |
| 156 | } |
| 157 | |
| 158 | /* Delete cost classes info V from the hash table. */ |
| 159 | inline void |
| 160 | cost_classes_hasher::remove (cost_classes *v) |
| 161 | { |
| 162 | ira_free (addr: v); |
| 163 | } |
| 164 | |
| 165 | /* Hash table of unique cost classes. */ |
| 166 | static hash_table<cost_classes_hasher> *cost_classes_htab; |
| 167 | |
| 168 | /* Map allocno class -> cost classes for pseudo of given allocno |
| 169 | class. */ |
| 170 | static cost_classes_t cost_classes_aclass_cache[N_REG_CLASSES]; |
| 171 | |
| 172 | /* Map mode -> cost classes for pseudo of give mode. */ |
| 173 | static cost_classes_t cost_classes_mode_cache[MAX_MACHINE_MODE]; |
| 174 | |
| 175 | /* Cost classes that include all classes in ira_important_classes. */ |
| 176 | static cost_classes all_cost_classes; |
| 177 | |
| 178 | /* Use the array of classes in CLASSES_PTR to fill out the rest of |
| 179 | the structure. */ |
| 180 | static void |
| 181 | complete_cost_classes (cost_classes_t classes_ptr) |
| 182 | { |
| 183 | for (int i = 0; i < N_REG_CLASSES; i++) |
| 184 | classes_ptr->index[i] = -1; |
| 185 | for (int i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
| 186 | classes_ptr->hard_regno_index[i] = -1; |
| 187 | for (int i = 0; i < classes_ptr->num; i++) |
| 188 | { |
| 189 | enum reg_class cl = classes_ptr->classes[i]; |
| 190 | classes_ptr->index[cl] = i; |
| 191 | for (int j = ira_class_hard_regs_num[cl] - 1; j >= 0; j--) |
| 192 | { |
| 193 | unsigned int hard_regno = ira_class_hard_regs[cl][j]; |
| 194 | if (classes_ptr->hard_regno_index[hard_regno] < 0) |
| 195 | classes_ptr->hard_regno_index[hard_regno] = i; |
| 196 | } |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | /* Initialize info about the cost classes for each pseudo. */ |
| 201 | static void |
| 202 | initiate_regno_cost_classes (void) |
| 203 | { |
| 204 | int size = sizeof (cost_classes_t) * max_reg_num (); |
| 205 | |
| 206 | regno_cost_classes = (cost_classes_t *) ira_allocate (size); |
| 207 | memset (s: regno_cost_classes, c: 0, n: size); |
| 208 | memset (s: cost_classes_aclass_cache, c: 0, |
| 209 | n: sizeof (cost_classes_t) * N_REG_CLASSES); |
| 210 | memset (s: cost_classes_mode_cache, c: 0, |
| 211 | n: sizeof (cost_classes_t) * MAX_MACHINE_MODE); |
| 212 | cost_classes_htab = new hash_table<cost_classes_hasher> (200); |
| 213 | all_cost_classes.num = ira_important_classes_num; |
| 214 | for (int i = 0; i < ira_important_classes_num; i++) |
| 215 | all_cost_classes.classes[i] = ira_important_classes[i]; |
| 216 | complete_cost_classes (classes_ptr: &all_cost_classes); |
| 217 | } |
| 218 | |
| 219 | /* Create new cost classes from cost classes FROM and set up members |
| 220 | index and hard_regno_index. Return the new classes. The function |
| 221 | implements some common code of two functions |
| 222 | setup_regno_cost_classes_by_aclass and |
| 223 | setup_regno_cost_classes_by_mode. */ |
| 224 | static cost_classes_t |
| 225 | setup_cost_classes (cost_classes_t from) |
| 226 | { |
| 227 | cost_classes_t classes_ptr; |
| 228 | |
| 229 | classes_ptr = (cost_classes_t) ira_allocate (sizeof (struct cost_classes)); |
| 230 | classes_ptr->num = from->num; |
| 231 | for (int i = 0; i < from->num; i++) |
| 232 | classes_ptr->classes[i] = from->classes[i]; |
| 233 | complete_cost_classes (classes_ptr); |
| 234 | return classes_ptr; |
| 235 | } |
| 236 | |
| 237 | /* Return a version of FULL that only considers registers in REGS that are |
| 238 | valid for mode MODE. Both FULL and the returned class are globally |
| 239 | allocated. */ |
| 240 | static cost_classes_t |
| 241 | restrict_cost_classes (cost_classes_t full, machine_mode mode, |
| 242 | const_hard_reg_set regs) |
| 243 | { |
| 244 | static struct cost_classes narrow; |
| 245 | int map[N_REG_CLASSES]; |
| 246 | narrow.num = 0; |
| 247 | for (int i = 0; i < full->num; i++) |
| 248 | { |
| 249 | /* Assume that we'll drop the class. */ |
| 250 | map[i] = -1; |
| 251 | |
| 252 | /* Ignore classes that are too small for the mode. */ |
| 253 | enum reg_class cl = full->classes[i]; |
| 254 | if (!contains_reg_of_mode[cl][mode]) |
| 255 | continue; |
| 256 | |
| 257 | /* Calculate the set of registers in CL that belong to REGS and |
| 258 | are valid for MODE. */ |
| 259 | HARD_REG_SET valid_for_cl = reg_class_contents[cl] & regs; |
| 260 | valid_for_cl &= ~(ira_prohibited_class_mode_regs[cl][mode] |
| 261 | | ira_no_alloc_regs); |
| 262 | if (hard_reg_set_empty_p (x: valid_for_cl)) |
| 263 | continue; |
| 264 | |
| 265 | /* Don't use this class if the set of valid registers is a subset |
| 266 | of an existing class. For example, suppose we have two classes |
| 267 | GR_REGS and FR_REGS and a union class GR_AND_FR_REGS. Suppose |
| 268 | that the mode changes allowed by FR_REGS are not as general as |
| 269 | the mode changes allowed by GR_REGS. |
| 270 | |
| 271 | In this situation, the mode changes for GR_AND_FR_REGS could |
| 272 | either be seen as the union or the intersection of the mode |
| 273 | changes allowed by the two subclasses. The justification for |
| 274 | the union-based definition would be that, if you want a mode |
| 275 | change that's only allowed by GR_REGS, you can pick a register |
| 276 | from the GR_REGS subclass. The justification for the |
| 277 | intersection-based definition would be that every register |
| 278 | from the class would allow the mode change. |
| 279 | |
| 280 | However, if we have a register that needs to be in GR_REGS, |
| 281 | using GR_AND_FR_REGS with the intersection-based definition |
| 282 | would be too pessimistic, since it would bring in restrictions |
| 283 | that only apply to FR_REGS. Conversely, if we have a register |
| 284 | that needs to be in FR_REGS, using GR_AND_FR_REGS with the |
| 285 | union-based definition would lose the extra restrictions |
| 286 | placed on FR_REGS. GR_AND_FR_REGS is therefore only useful |
| 287 | for cases where GR_REGS and FP_REGS are both valid. */ |
| 288 | int pos; |
| 289 | for (pos = 0; pos < narrow.num; ++pos) |
| 290 | { |
| 291 | enum reg_class cl2 = narrow.classes[pos]; |
| 292 | if (hard_reg_set_subset_p (x: valid_for_cl, reg_class_contents[cl2])) |
| 293 | break; |
| 294 | } |
| 295 | map[i] = pos; |
| 296 | if (pos == narrow.num) |
| 297 | { |
| 298 | /* If several classes are equivalent, prefer to use the one |
| 299 | that was chosen as the allocno class. */ |
| 300 | enum reg_class cl2 = ira_allocno_class_translate[cl]; |
| 301 | if (ira_class_hard_regs_num[cl] == ira_class_hard_regs_num[cl2]) |
| 302 | cl = cl2; |
| 303 | narrow.classes[narrow.num++] = cl; |
| 304 | } |
| 305 | } |
| 306 | if (narrow.num == full->num) |
| 307 | return full; |
| 308 | |
| 309 | cost_classes **slot = cost_classes_htab->find_slot (value: &narrow, insert: INSERT); |
| 310 | if (*slot == NULL) |
| 311 | { |
| 312 | cost_classes_t classes = setup_cost_classes (&narrow); |
| 313 | /* Map equivalent classes to the representative that we chose above. */ |
| 314 | for (int i = 0; i < ira_important_classes_num; i++) |
| 315 | { |
| 316 | enum reg_class cl = ira_important_classes[i]; |
| 317 | int index = full->index[cl]; |
| 318 | if (index >= 0) |
| 319 | classes->index[cl] = map[index]; |
| 320 | } |
| 321 | *slot = classes; |
| 322 | } |
| 323 | return *slot; |
| 324 | } |
| 325 | |
| 326 | /* Setup cost classes for pseudo REGNO whose allocno class is ACLASS. |
| 327 | This function is used when we know an initial approximation of |
| 328 | allocno class of the pseudo already, e.g. on the second iteration |
| 329 | of class cost calculation or after class cost calculation in |
| 330 | register-pressure sensitive insn scheduling or register-pressure |
| 331 | sensitive loop-invariant motion. */ |
| 332 | static void |
| 333 | setup_regno_cost_classes_by_aclass (int regno, enum reg_class aclass) |
| 334 | { |
| 335 | static struct cost_classes classes; |
| 336 | cost_classes_t classes_ptr; |
| 337 | enum reg_class cl; |
| 338 | int i; |
| 339 | cost_classes **slot; |
| 340 | HARD_REG_SET temp, temp2; |
| 341 | bool exclude_p; |
| 342 | |
| 343 | if ((classes_ptr = cost_classes_aclass_cache[aclass]) == NULL) |
| 344 | { |
| 345 | temp = reg_class_contents[aclass] & ~ira_no_alloc_regs; |
| 346 | /* We exclude classes from consideration which are subsets of |
| 347 | ACLASS only if ACLASS is an uniform class. */ |
| 348 | exclude_p = ira_uniform_class_p[aclass]; |
| 349 | classes.num = 0; |
| 350 | for (i = 0; i < ira_important_classes_num; i++) |
| 351 | { |
| 352 | cl = ira_important_classes[i]; |
| 353 | if (exclude_p) |
| 354 | { |
| 355 | /* Exclude non-uniform classes which are subsets of |
| 356 | ACLASS. */ |
| 357 | temp2 = reg_class_contents[cl] & ~ira_no_alloc_regs; |
| 358 | if (hard_reg_set_subset_p (x: temp2, y: temp) && cl != aclass) |
| 359 | continue; |
| 360 | } |
| 361 | classes.classes[classes.num++] = cl; |
| 362 | } |
| 363 | slot = cost_classes_htab->find_slot (value: &classes, insert: INSERT); |
| 364 | if (*slot == NULL) |
| 365 | { |
| 366 | classes_ptr = setup_cost_classes (&classes); |
| 367 | *slot = classes_ptr; |
| 368 | } |
| 369 | classes_ptr = cost_classes_aclass_cache[aclass] = (cost_classes_t) *slot; |
| 370 | } |
| 371 | if (regno_reg_rtx[regno] != NULL_RTX) |
| 372 | { |
| 373 | /* Restrict the classes to those that are valid for REGNO's mode |
| 374 | (which might for example exclude singleton classes if the mode |
| 375 | requires two registers). Also restrict the classes to those that |
| 376 | are valid for subregs of REGNO. */ |
| 377 | const HARD_REG_SET *valid_regs = valid_mode_changes_for_regno (regno); |
| 378 | if (!valid_regs) |
| 379 | valid_regs = ®_class_contents[ALL_REGS]; |
| 380 | classes_ptr = restrict_cost_classes (full: classes_ptr, |
| 381 | PSEUDO_REGNO_MODE (regno), |
| 382 | regs: *valid_regs); |
| 383 | } |
| 384 | regno_cost_classes[regno] = classes_ptr; |
| 385 | } |
| 386 | |
| 387 | /* Setup cost classes for pseudo REGNO with MODE. Usage of MODE can |
| 388 | decrease number of cost classes for the pseudo, if hard registers |
| 389 | of some important classes cannot hold a value of MODE. So the |
| 390 | pseudo cannot get hard register of some important classes and cost |
| 391 | calculation for such important classes is only wasting CPU |
| 392 | time. */ |
| 393 | static void |
| 394 | setup_regno_cost_classes_by_mode (int regno, machine_mode mode) |
| 395 | { |
| 396 | if (const HARD_REG_SET *valid_regs = valid_mode_changes_for_regno (regno)) |
| 397 | regno_cost_classes[regno] = restrict_cost_classes (full: &all_cost_classes, |
| 398 | mode, regs: *valid_regs); |
| 399 | else |
| 400 | { |
| 401 | if (cost_classes_mode_cache[mode] == NULL) |
| 402 | cost_classes_mode_cache[mode] |
| 403 | = restrict_cost_classes (full: &all_cost_classes, mode, |
| 404 | reg_class_contents[ALL_REGS]); |
| 405 | regno_cost_classes[regno] = cost_classes_mode_cache[mode]; |
| 406 | } |
| 407 | } |
| 408 | |
| 409 | /* Finalize info about the cost classes for each pseudo. */ |
| 410 | static void |
| 411 | finish_regno_cost_classes (void) |
| 412 | { |
| 413 | ira_free (addr: regno_cost_classes); |
| 414 | delete cost_classes_htab; |
| 415 | cost_classes_htab = NULL; |
| 416 | } |
| 417 | |
| 418 | |
| 419 | |
| 420 | /* Compute the cost of loading X into (if TO_P is TRUE) or from (if |
| 421 | TO_P is FALSE) a register of class RCLASS in mode MODE. X must not |
| 422 | be a pseudo register. */ |
| 423 | static int |
| 424 | copy_cost (rtx x, machine_mode mode, reg_class_t rclass, bool to_p, |
| 425 | secondary_reload_info *prev_sri) |
| 426 | { |
| 427 | secondary_reload_info sri; |
| 428 | reg_class_t secondary_class = NO_REGS; |
| 429 | |
| 430 | /* If X is a SCRATCH, there is actually nothing to move since we are |
| 431 | assuming optimal allocation. */ |
| 432 | if (GET_CODE (x) == SCRATCH) |
| 433 | return 0; |
| 434 | |
| 435 | /* Get the class we will actually use for a reload. */ |
| 436 | rclass = targetm.preferred_reload_class (x, rclass); |
| 437 | |
| 438 | /* If we need a secondary reload for an intermediate, the cost is |
| 439 | that to load the input into the intermediate register, then to |
| 440 | copy it. */ |
| 441 | sri.prev_sri = prev_sri; |
| 442 | sri.extra_cost = 0; |
| 443 | /* PR 68770: Secondary reload might examine the t_icode field. */ |
| 444 | sri.t_icode = CODE_FOR_nothing; |
| 445 | |
| 446 | secondary_class = targetm.secondary_reload (to_p, x, rclass, mode, &sri); |
| 447 | |
| 448 | if (secondary_class != NO_REGS) |
| 449 | { |
| 450 | ira_init_register_move_cost_if_necessary (mode); |
| 451 | return (ira_register_move_cost[mode][(int) secondary_class][(int) rclass] |
| 452 | + sri.extra_cost |
| 453 | + copy_cost (x, mode, rclass: secondary_class, to_p, prev_sri: &sri)); |
| 454 | } |
| 455 | |
| 456 | /* For memory, use the memory move cost, for (hard) registers, use |
| 457 | the cost to move between the register classes, and use 2 for |
| 458 | everything else (constants). */ |
| 459 | if (MEM_P (x) || rclass == NO_REGS) |
| 460 | return sri.extra_cost |
| 461 | + ira_memory_move_cost[mode][(int) rclass][to_p != 0]; |
| 462 | else if (REG_P (x)) |
| 463 | { |
| 464 | reg_class_t x_class = REGNO_REG_CLASS (REGNO (x)); |
| 465 | |
| 466 | ira_init_register_move_cost_if_necessary (mode); |
| 467 | return (sri.extra_cost |
| 468 | + ira_register_move_cost[mode][(int) x_class][(int) rclass]); |
| 469 | } |
| 470 | else |
| 471 | /* If this is a constant, we may eventually want to call rtx_cost |
| 472 | here. */ |
| 473 | return sri.extra_cost + COSTS_N_INSNS (1); |
| 474 | } |
| 475 | |
| 476 | |
| 477 | |
| 478 | /* Record the cost of using memory or hard registers of various |
| 479 | classes for the operands in INSN. |
| 480 | |
| 481 | N_ALTS is the number of alternatives. |
| 482 | N_OPS is the number of operands. |
| 483 | OPS is an array of the operands. |
| 484 | MODES are the modes of the operands, in case any are VOIDmode. |
| 485 | CONSTRAINTS are the constraints to use for the operands. This array |
| 486 | is modified by this procedure. |
| 487 | |
| 488 | This procedure works alternative by alternative. For each |
| 489 | alternative we assume that we will be able to allocate all allocnos |
| 490 | to their ideal register class and calculate the cost of using that |
| 491 | alternative. Then we compute, for each operand that is a |
| 492 | pseudo-register, the cost of having the allocno allocated to each |
| 493 | register class and using it in that alternative. To this cost is |
| 494 | added the cost of the alternative. |
| 495 | |
| 496 | The cost of each class for this insn is its lowest cost among all |
| 497 | the alternatives. */ |
| 498 | static void |
| 499 | record_reg_classes (int n_alts, int n_ops, rtx *ops, |
| 500 | machine_mode *modes, const char **constraints, |
| 501 | rtx_insn *insn, enum reg_class *pref) |
| 502 | { |
| 503 | int alt; |
| 504 | int i, j, k; |
| 505 | int insn_allows_mem[MAX_RECOG_OPERANDS]; |
| 506 | move_table *move_in_cost, *move_out_cost; |
| 507 | short (*mem_cost)[2]; |
| 508 | const char *p; |
| 509 | |
| 510 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5) |
| 511 | { |
| 512 | fprintf (stream: ira_dump_file, format: " Processing insn %u" , INSN_UID (insn)); |
| 513 | if (INSN_CODE (insn) >= 0 |
| 514 | && (p = get_insn_name (INSN_CODE (insn))) != NULL) |
| 515 | fprintf (stream: ira_dump_file, format: " {%s}" , p); |
| 516 | fprintf (stream: ira_dump_file, format: " (freq=%d)\n" , |
| 517 | REG_FREQ_FROM_BB (BLOCK_FOR_INSN (insn))); |
| 518 | dump_insn_slim (ira_dump_file, insn); |
| 519 | } |
| 520 | |
| 521 | for (i = 0; i < n_ops; i++) |
| 522 | insn_allows_mem[i] = 0; |
| 523 | |
| 524 | /* Process each alternative, each time minimizing an operand's cost |
| 525 | with the cost for each operand in that alternative. */ |
| 526 | alternative_mask preferred = get_preferred_alternatives (insn); |
| 527 | for (alt = 0; alt < n_alts; alt++) |
| 528 | { |
| 529 | enum reg_class classes[MAX_RECOG_OPERANDS]; |
| 530 | int allows_mem[MAX_RECOG_OPERANDS]; |
| 531 | enum reg_class rclass; |
| 532 | int alt_fail = 0; |
| 533 | int alt_cost = 0, op_cost_add; |
| 534 | |
| 535 | if (!TEST_BIT (preferred, alt)) |
| 536 | { |
| 537 | for (i = 0; i < recog_data.n_operands; i++) |
| 538 | constraints[i] = skip_alternative (p: constraints[i]); |
| 539 | |
| 540 | continue; |
| 541 | } |
| 542 | |
| 543 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5) |
| 544 | { |
| 545 | fprintf (stream: ira_dump_file, format: " Alt %d:" , alt); |
| 546 | for (i = 0; i < n_ops; i++) |
| 547 | { |
| 548 | p = constraints[i]; |
| 549 | if (*p == '\0') |
| 550 | continue; |
| 551 | fprintf (stream: ira_dump_file, format: " (%d) " , i); |
| 552 | for (; *p != '\0' && *p != ',' && *p != '#'; p++) |
| 553 | fputc (c: *p, stream: ira_dump_file); |
| 554 | } |
| 555 | fprintf (stream: ira_dump_file, format: "\n" ); |
| 556 | } |
| 557 | |
| 558 | for (i = 0; i < n_ops; i++) |
| 559 | { |
| 560 | unsigned char c; |
| 561 | const char *p = constraints[i]; |
| 562 | rtx op = ops[i]; |
| 563 | machine_mode mode = modes[i]; |
| 564 | int allows_addr = 0; |
| 565 | int win = 0; |
| 566 | |
| 567 | /* Initially show we know nothing about the register class. */ |
| 568 | classes[i] = NO_REGS; |
| 569 | allows_mem[i] = 0; |
| 570 | |
| 571 | /* If this operand has no constraints at all, we can |
| 572 | conclude nothing about it since anything is valid. */ |
| 573 | if (*p == 0) |
| 574 | { |
| 575 | if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER) |
| 576 | memset (this_op_costs[i], c: 0, n: struct_costs_size); |
| 577 | continue; |
| 578 | } |
| 579 | |
| 580 | /* If this alternative is only relevant when this operand |
| 581 | matches a previous operand, we do different things |
| 582 | depending on whether this operand is a allocno-reg or not. |
| 583 | We must process any modifiers for the operand before we |
| 584 | can make this test. */ |
| 585 | while (*p == '%' || *p == '=' || *p == '+' || *p == '&') |
| 586 | p++; |
| 587 | |
| 588 | if (p[0] >= '0' && p[0] <= '0' + i) |
| 589 | { |
| 590 | /* Copy class and whether memory is allowed from the |
| 591 | matching alternative. Then perform any needed cost |
| 592 | computations and/or adjustments. */ |
| 593 | j = p[0] - '0'; |
| 594 | classes[i] = classes[j]; |
| 595 | allows_mem[i] = allows_mem[j]; |
| 596 | if (allows_mem[i]) |
| 597 | insn_allows_mem[i] = 1; |
| 598 | |
| 599 | if (! REG_P (op) || REGNO (op) < FIRST_PSEUDO_REGISTER) |
| 600 | { |
| 601 | /* If this matches the other operand, we have no |
| 602 | added cost and we win. */ |
| 603 | if (rtx_equal_p (ops[j], op)) |
| 604 | win = 1; |
| 605 | /* If we can put the other operand into a register, |
| 606 | add to the cost of this alternative the cost to |
| 607 | copy this operand to the register used for the |
| 608 | other operand. */ |
| 609 | else if (classes[j] != NO_REGS) |
| 610 | { |
| 611 | alt_cost += copy_cost (x: op, mode, rclass: classes[j], to_p: 1, NULL); |
| 612 | win = 1; |
| 613 | } |
| 614 | } |
| 615 | else if (! REG_P (ops[j]) |
| 616 | || REGNO (ops[j]) < FIRST_PSEUDO_REGISTER) |
| 617 | { |
| 618 | /* This op is an allocno but the one it matches is |
| 619 | not. */ |
| 620 | |
| 621 | /* If we can't put the other operand into a |
| 622 | register, this alternative can't be used. */ |
| 623 | |
| 624 | if (classes[j] == NO_REGS) |
| 625 | { |
| 626 | alt_fail = 1; |
| 627 | } |
| 628 | else |
| 629 | /* Otherwise, add to the cost of this alternative the cost |
| 630 | to copy the other operand to the hard register used for |
| 631 | this operand. */ |
| 632 | { |
| 633 | alt_cost += copy_cost (x: ops[j], mode, rclass: classes[j], to_p: 1, NULL); |
| 634 | } |
| 635 | } |
| 636 | else |
| 637 | { |
| 638 | /* The costs of this operand are not the same as the |
| 639 | other operand since move costs are not symmetric. |
| 640 | Moreover, if we cannot tie them, this alternative |
| 641 | needs to do a copy, which is one insn. */ |
| 642 | struct costs *pp = this_op_costs[i]; |
| 643 | int *pp_costs = pp->cost; |
| 644 | cost_classes_t cost_classes_ptr |
| 645 | = regno_cost_classes[REGNO (op)]; |
| 646 | enum reg_class *cost_classes = cost_classes_ptr->classes; |
| 647 | bool in_p = recog_data.operand_type[i] != OP_OUT; |
| 648 | bool out_p = recog_data.operand_type[i] != OP_IN; |
| 649 | enum reg_class op_class = classes[i]; |
| 650 | |
| 651 | ira_init_register_move_cost_if_necessary (mode); |
| 652 | if (! in_p) |
| 653 | { |
| 654 | ira_assert (out_p); |
| 655 | if (op_class == NO_REGS) |
| 656 | { |
| 657 | mem_cost = ira_memory_move_cost[mode]; |
| 658 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 659 | { |
| 660 | rclass = cost_classes[k]; |
| 661 | pp_costs[k] = mem_cost[rclass][0] * frequency; |
| 662 | } |
| 663 | } |
| 664 | else |
| 665 | { |
| 666 | move_out_cost = ira_may_move_out_cost[mode]; |
| 667 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 668 | { |
| 669 | rclass = cost_classes[k]; |
| 670 | pp_costs[k] |
| 671 | = move_out_cost[op_class][rclass] * frequency; |
| 672 | } |
| 673 | } |
| 674 | } |
| 675 | else if (! out_p) |
| 676 | { |
| 677 | ira_assert (in_p); |
| 678 | if (op_class == NO_REGS) |
| 679 | { |
| 680 | mem_cost = ira_memory_move_cost[mode]; |
| 681 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 682 | { |
| 683 | rclass = cost_classes[k]; |
| 684 | pp_costs[k] = mem_cost[rclass][1] * frequency; |
| 685 | } |
| 686 | } |
| 687 | else |
| 688 | { |
| 689 | move_in_cost = ira_may_move_in_cost[mode]; |
| 690 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 691 | { |
| 692 | rclass = cost_classes[k]; |
| 693 | pp_costs[k] |
| 694 | = move_in_cost[rclass][op_class] * frequency; |
| 695 | } |
| 696 | } |
| 697 | } |
| 698 | else |
| 699 | { |
| 700 | if (op_class == NO_REGS) |
| 701 | { |
| 702 | mem_cost = ira_memory_move_cost[mode]; |
| 703 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 704 | { |
| 705 | rclass = cost_classes[k]; |
| 706 | pp_costs[k] = ((mem_cost[rclass][0] |
| 707 | + mem_cost[rclass][1]) |
| 708 | * frequency); |
| 709 | } |
| 710 | } |
| 711 | else |
| 712 | { |
| 713 | move_in_cost = ira_may_move_in_cost[mode]; |
| 714 | move_out_cost = ira_may_move_out_cost[mode]; |
| 715 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 716 | { |
| 717 | rclass = cost_classes[k]; |
| 718 | pp_costs[k] = ((move_in_cost[rclass][op_class] |
| 719 | + move_out_cost[op_class][rclass]) |
| 720 | * frequency); |
| 721 | } |
| 722 | } |
| 723 | } |
| 724 | |
| 725 | /* If the alternative actually allows memory, make |
| 726 | things a bit cheaper since we won't need an extra |
| 727 | insn to load it. */ |
| 728 | pp->mem_cost |
| 729 | = ((out_p ? ira_memory_move_cost[mode][op_class][0] : 0) |
| 730 | + (in_p ? ira_memory_move_cost[mode][op_class][1] : 0) |
| 731 | - allows_mem[i]) * frequency; |
| 732 | |
| 733 | /* If we have assigned a class to this allocno in |
| 734 | our first pass, add a cost to this alternative |
| 735 | corresponding to what we would add if this |
| 736 | allocno were not in the appropriate class. */ |
| 737 | if (pref) |
| 738 | { |
| 739 | enum reg_class pref_class = pref[COST_INDEX (REGNO (op))]; |
| 740 | |
| 741 | if (pref_class == NO_REGS) |
| 742 | alt_cost |
| 743 | += ((out_p |
| 744 | ? ira_memory_move_cost[mode][op_class][0] : 0) |
| 745 | + (in_p |
| 746 | ? ira_memory_move_cost[mode][op_class][1] |
| 747 | : 0)); |
| 748 | else if (ira_reg_class_intersect |
| 749 | [pref_class][op_class] == NO_REGS) |
| 750 | alt_cost |
| 751 | += ira_register_move_cost[mode][pref_class][op_class]; |
| 752 | } |
| 753 | if (REGNO (ops[i]) != REGNO (ops[j]) |
| 754 | && ! find_reg_note (insn, REG_DEAD, op)) |
| 755 | alt_cost += 2; |
| 756 | |
| 757 | p++; |
| 758 | } |
| 759 | } |
| 760 | |
| 761 | /* Scan all the constraint letters. See if the operand |
| 762 | matches any of the constraints. Collect the valid |
| 763 | register classes and see if this operand accepts |
| 764 | memory. */ |
| 765 | while ((c = *p)) |
| 766 | { |
| 767 | switch (c) |
| 768 | { |
| 769 | case '*': |
| 770 | /* Ignore the next letter for this pass. */ |
| 771 | c = *++p; |
| 772 | break; |
| 773 | |
| 774 | case '^': |
| 775 | alt_cost += 2; |
| 776 | break; |
| 777 | |
| 778 | case '?': |
| 779 | alt_cost += 2; |
| 780 | break; |
| 781 | |
| 782 | case 'g': |
| 783 | if (MEM_P (op) |
| 784 | || (CONSTANT_P (op) |
| 785 | && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (op)))) |
| 786 | win = 1; |
| 787 | insn_allows_mem[i] = allows_mem[i] = 1; |
| 788 | classes[i] = ira_reg_class_subunion[classes[i]][GENERAL_REGS]; |
| 789 | break; |
| 790 | |
| 791 | default: |
| 792 | enum constraint_num cn = lookup_constraint (p); |
| 793 | enum reg_class cl; |
| 794 | switch (get_constraint_type (c: cn)) |
| 795 | { |
| 796 | case CT_REGISTER: |
| 797 | cl = reg_class_for_constraint (c: cn); |
| 798 | if (cl != NO_REGS) |
| 799 | classes[i] = ira_reg_class_subunion[classes[i]][cl]; |
| 800 | break; |
| 801 | |
| 802 | case CT_CONST_INT: |
| 803 | if (CONST_INT_P (op) |
| 804 | && insn_const_int_ok_for_constraint (INTVAL (op), cn)) |
| 805 | win = 1; |
| 806 | break; |
| 807 | |
| 808 | case CT_MEMORY: |
| 809 | case CT_RELAXED_MEMORY: |
| 810 | /* Every MEM can be reloaded to fit. */ |
| 811 | insn_allows_mem[i] = allows_mem[i] = 1; |
| 812 | if (MEM_P (op)) |
| 813 | win = 1; |
| 814 | break; |
| 815 | |
| 816 | case CT_SPECIAL_MEMORY: |
| 817 | insn_allows_mem[i] = allows_mem[i] = 1; |
| 818 | if (MEM_P (extract_mem_from_operand (op)) |
| 819 | && constraint_satisfied_p (x: op, c: cn)) |
| 820 | win = 1; |
| 821 | break; |
| 822 | |
| 823 | case CT_ADDRESS: |
| 824 | /* Every address can be reloaded to fit. */ |
| 825 | allows_addr = 1; |
| 826 | if (address_operand (op, GET_MODE (op)) |
| 827 | || constraint_satisfied_p (x: op, c: cn)) |
| 828 | win = 1; |
| 829 | /* We know this operand is an address, so we |
| 830 | want it to be allocated to a hard register |
| 831 | that can be the base of an address, |
| 832 | i.e. BASE_REG_CLASS. */ |
| 833 | classes[i] |
| 834 | = ira_reg_class_subunion[classes[i]] |
| 835 | [base_reg_class (VOIDmode, ADDR_SPACE_GENERIC, |
| 836 | outer_code: ADDRESS, index_code: SCRATCH)]; |
| 837 | break; |
| 838 | |
| 839 | case CT_FIXED_FORM: |
| 840 | if (constraint_satisfied_p (x: op, c: cn)) |
| 841 | win = 1; |
| 842 | break; |
| 843 | } |
| 844 | break; |
| 845 | } |
| 846 | p += CONSTRAINT_LEN (c, p); |
| 847 | if (c == ',') |
| 848 | break; |
| 849 | } |
| 850 | |
| 851 | constraints[i] = p; |
| 852 | |
| 853 | if (alt_fail) |
| 854 | break; |
| 855 | |
| 856 | /* How we account for this operand now depends on whether it |
| 857 | is a pseudo register or not. If it is, we first check if |
| 858 | any register classes are valid. If not, we ignore this |
| 859 | alternative, since we want to assume that all allocnos get |
| 860 | allocated for register preferencing. If some register |
| 861 | class is valid, compute the costs of moving the allocno |
| 862 | into that class. */ |
| 863 | if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER) |
| 864 | { |
| 865 | if (classes[i] == NO_REGS && ! allows_mem[i]) |
| 866 | { |
| 867 | /* We must always fail if the operand is a REG, but |
| 868 | we did not find a suitable class and memory is |
| 869 | not allowed. |
| 870 | |
| 871 | Otherwise we may perform an uninitialized read |
| 872 | from this_op_costs after the `continue' statement |
| 873 | below. */ |
| 874 | alt_fail = 1; |
| 875 | } |
| 876 | else |
| 877 | { |
| 878 | unsigned int regno = REGNO (op); |
| 879 | struct costs *pp = this_op_costs[i]; |
| 880 | int *pp_costs = pp->cost; |
| 881 | cost_classes_t cost_classes_ptr = regno_cost_classes[regno]; |
| 882 | enum reg_class *cost_classes = cost_classes_ptr->classes; |
| 883 | bool in_p = recog_data.operand_type[i] != OP_OUT; |
| 884 | bool out_p = recog_data.operand_type[i] != OP_IN; |
| 885 | enum reg_class op_class = classes[i]; |
| 886 | |
| 887 | ira_init_register_move_cost_if_necessary (mode); |
| 888 | if (! in_p) |
| 889 | { |
| 890 | ira_assert (out_p); |
| 891 | if (op_class == NO_REGS) |
| 892 | { |
| 893 | mem_cost = ira_memory_move_cost[mode]; |
| 894 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 895 | { |
| 896 | rclass = cost_classes[k]; |
| 897 | pp_costs[k] = mem_cost[rclass][0] * frequency; |
| 898 | } |
| 899 | } |
| 900 | else |
| 901 | { |
| 902 | move_out_cost = ira_may_move_out_cost[mode]; |
| 903 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 904 | { |
| 905 | rclass = cost_classes[k]; |
| 906 | pp_costs[k] |
| 907 | = move_out_cost[op_class][rclass] * frequency; |
| 908 | } |
| 909 | } |
| 910 | } |
| 911 | else if (! out_p) |
| 912 | { |
| 913 | ira_assert (in_p); |
| 914 | if (op_class == NO_REGS) |
| 915 | { |
| 916 | mem_cost = ira_memory_move_cost[mode]; |
| 917 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 918 | { |
| 919 | rclass = cost_classes[k]; |
| 920 | pp_costs[k] = mem_cost[rclass][1] * frequency; |
| 921 | } |
| 922 | } |
| 923 | else |
| 924 | { |
| 925 | move_in_cost = ira_may_move_in_cost[mode]; |
| 926 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 927 | { |
| 928 | rclass = cost_classes[k]; |
| 929 | pp_costs[k] |
| 930 | = move_in_cost[rclass][op_class] * frequency; |
| 931 | } |
| 932 | } |
| 933 | } |
| 934 | else |
| 935 | { |
| 936 | if (op_class == NO_REGS) |
| 937 | { |
| 938 | mem_cost = ira_memory_move_cost[mode]; |
| 939 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 940 | { |
| 941 | rclass = cost_classes[k]; |
| 942 | pp_costs[k] = ((mem_cost[rclass][0] |
| 943 | + mem_cost[rclass][1]) |
| 944 | * frequency); |
| 945 | } |
| 946 | } |
| 947 | else |
| 948 | { |
| 949 | move_in_cost = ira_may_move_in_cost[mode]; |
| 950 | move_out_cost = ira_may_move_out_cost[mode]; |
| 951 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 952 | { |
| 953 | rclass = cost_classes[k]; |
| 954 | pp_costs[k] = ((move_in_cost[rclass][op_class] |
| 955 | + move_out_cost[op_class][rclass]) |
| 956 | * frequency); |
| 957 | } |
| 958 | } |
| 959 | } |
| 960 | |
| 961 | if (op_class == NO_REGS) |
| 962 | /* Although we don't need insn to reload from |
| 963 | memory, still accessing memory is usually more |
| 964 | expensive than a register. */ |
| 965 | pp->mem_cost = frequency; |
| 966 | else |
| 967 | /* If the alternative actually allows memory, make |
| 968 | things a bit cheaper since we won't need an |
| 969 | extra insn to load it. */ |
| 970 | pp->mem_cost |
| 971 | = ((out_p ? ira_memory_move_cost[mode][op_class][0] : 0) |
| 972 | + (in_p ? ira_memory_move_cost[mode][op_class][1] : 0) |
| 973 | - allows_mem[i]) * frequency; |
| 974 | /* If we have assigned a class to this allocno in |
| 975 | our first pass, add a cost to this alternative |
| 976 | corresponding to what we would add if this |
| 977 | allocno were not in the appropriate class. */ |
| 978 | if (pref) |
| 979 | { |
| 980 | enum reg_class pref_class = pref[COST_INDEX (REGNO (op))]; |
| 981 | |
| 982 | if (pref_class == NO_REGS) |
| 983 | { |
| 984 | if (op_class != NO_REGS) |
| 985 | alt_cost |
| 986 | += ((out_p |
| 987 | ? ira_memory_move_cost[mode][op_class][0] |
| 988 | : 0) |
| 989 | + (in_p |
| 990 | ? ira_memory_move_cost[mode][op_class][1] |
| 991 | : 0)); |
| 992 | } |
| 993 | else if (op_class == NO_REGS) |
| 994 | alt_cost |
| 995 | += ((out_p |
| 996 | ? ira_memory_move_cost[mode][pref_class][1] |
| 997 | : 0) |
| 998 | + (in_p |
| 999 | ? ira_memory_move_cost[mode][pref_class][0] |
| 1000 | : 0)); |
| 1001 | else if (ira_reg_class_intersect[pref_class][op_class] |
| 1002 | == NO_REGS) |
| 1003 | alt_cost += (ira_register_move_cost |
| 1004 | [mode][pref_class][op_class]); |
| 1005 | } |
| 1006 | } |
| 1007 | } |
| 1008 | |
| 1009 | /* Otherwise, if this alternative wins, either because we |
| 1010 | have already determined that or if we have a hard |
| 1011 | register of the proper class, there is no cost for this |
| 1012 | alternative. */ |
| 1013 | else if (win || (REG_P (op) |
| 1014 | && reg_fits_class_p (op, classes[i], |
| 1015 | 0, GET_MODE (op)))) |
| 1016 | ; |
| 1017 | |
| 1018 | /* If registers are valid, the cost of this alternative |
| 1019 | includes copying the object to and/or from a |
| 1020 | register. */ |
| 1021 | else if (classes[i] != NO_REGS) |
| 1022 | { |
| 1023 | if (recog_data.operand_type[i] != OP_OUT) |
| 1024 | alt_cost += copy_cost (x: op, mode, rclass: classes[i], to_p: 1, NULL); |
| 1025 | |
| 1026 | if (recog_data.operand_type[i] != OP_IN) |
| 1027 | alt_cost += copy_cost (x: op, mode, rclass: classes[i], to_p: 0, NULL); |
| 1028 | } |
| 1029 | /* The only other way this alternative can be used is if |
| 1030 | this is a constant that could be placed into memory. */ |
| 1031 | else if (CONSTANT_P (op) && (allows_addr || allows_mem[i])) |
| 1032 | alt_cost += ira_memory_move_cost[mode][classes[i]][1]; |
| 1033 | else |
| 1034 | alt_fail = 1; |
| 1035 | |
| 1036 | if (alt_fail) |
| 1037 | break; |
| 1038 | } |
| 1039 | |
| 1040 | if (alt_fail) |
| 1041 | { |
| 1042 | /* The loop above might have exited early once the failure |
| 1043 | was seen. Skip over the constraints for the remaining |
| 1044 | operands. */ |
| 1045 | i += 1; |
| 1046 | for (; i < n_ops; ++i) |
| 1047 | constraints[i] = skip_alternative (p: constraints[i]); |
| 1048 | continue; |
| 1049 | } |
| 1050 | |
| 1051 | op_cost_add = alt_cost * frequency; |
| 1052 | /* Finally, update the costs with the information we've |
| 1053 | calculated about this alternative. */ |
| 1054 | for (i = 0; i < n_ops; i++) |
| 1055 | if (REG_P (ops[i]) && REGNO (ops[i]) >= FIRST_PSEUDO_REGISTER) |
| 1056 | { |
| 1057 | int old_cost; |
| 1058 | bool cost_change_p = false; |
| 1059 | struct costs *pp = op_costs[i], *qq = this_op_costs[i]; |
| 1060 | int *pp_costs = pp->cost, *qq_costs = qq->cost; |
| 1061 | int scale = 1 + (recog_data.operand_type[i] == OP_INOUT); |
| 1062 | cost_classes_t cost_classes_ptr |
| 1063 | = regno_cost_classes[REGNO (ops[i])]; |
| 1064 | |
| 1065 | old_cost = pp->mem_cost; |
| 1066 | pp->mem_cost = MIN (old_cost, |
| 1067 | (qq->mem_cost + op_cost_add) * scale); |
| 1068 | |
| 1069 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5 |
| 1070 | && pp->mem_cost < old_cost) |
| 1071 | { |
| 1072 | cost_change_p = true; |
| 1073 | fprintf (stream: ira_dump_file, format: " op %d(r=%u) new costs MEM:%d" , |
| 1074 | i, REGNO(ops[i]), pp->mem_cost); |
| 1075 | } |
| 1076 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 1077 | { |
| 1078 | old_cost = pp_costs[k]; |
| 1079 | pp_costs[k] |
| 1080 | = MIN (old_cost, (qq_costs[k] + op_cost_add) * scale); |
| 1081 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5 |
| 1082 | && pp_costs[k] < old_cost) |
| 1083 | { |
| 1084 | if (!cost_change_p) |
| 1085 | fprintf (stream: ira_dump_file, format: " op %d(r=%u) new costs" , |
| 1086 | i, REGNO(ops[i])); |
| 1087 | cost_change_p = true; |
| 1088 | fprintf (stream: ira_dump_file, format: " %s:%d" , |
| 1089 | reg_class_names[cost_classes_ptr->classes[k]], |
| 1090 | pp_costs[k]); |
| 1091 | } |
| 1092 | } |
| 1093 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5 |
| 1094 | && cost_change_p) |
| 1095 | fprintf (stream: ira_dump_file, format: "\n" ); |
| 1096 | } |
| 1097 | } |
| 1098 | |
| 1099 | if (allocno_p) |
| 1100 | for (i = 0; i < n_ops; i++) |
| 1101 | { |
| 1102 | ira_allocno_t a; |
| 1103 | rtx op = ops[i]; |
| 1104 | |
| 1105 | if (! REG_P (op) || REGNO (op) < FIRST_PSEUDO_REGISTER) |
| 1106 | continue; |
| 1107 | a = ira_curr_regno_allocno_map [REGNO (op)]; |
| 1108 | if (! ALLOCNO_BAD_SPILL_P (a) && insn_allows_mem[i] == 0) |
| 1109 | ALLOCNO_BAD_SPILL_P (a) = true; |
| 1110 | } |
| 1111 | |
| 1112 | } |
| 1113 | |
| 1114 | |
| 1115 | |
| 1116 | /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudo registers. */ |
| 1117 | static inline bool |
| 1118 | ok_for_index_p_nonstrict (rtx reg) |
| 1119 | { |
| 1120 | unsigned regno = REGNO (reg); |
| 1121 | |
| 1122 | return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno); |
| 1123 | } |
| 1124 | |
| 1125 | /* A version of regno_ok_for_base_p for use here, when all |
| 1126 | pseudo-registers should count as OK. Arguments as for |
| 1127 | regno_ok_for_base_p. */ |
| 1128 | static inline bool |
| 1129 | ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as, |
| 1130 | enum rtx_code outer_code, enum rtx_code index_code) |
| 1131 | { |
| 1132 | unsigned regno = REGNO (reg); |
| 1133 | |
| 1134 | if (regno >= FIRST_PSEUDO_REGISTER) |
| 1135 | return true; |
| 1136 | return ok_for_base_p_1 (regno, mode, as, outer_code, index_code); |
| 1137 | } |
| 1138 | |
| 1139 | /* Record the pseudo registers we must reload into hard registers in a |
| 1140 | subexpression of a memory address, X. |
| 1141 | |
| 1142 | If CONTEXT is 0, we are looking at the base part of an address, |
| 1143 | otherwise we are looking at the index part. |
| 1144 | |
| 1145 | MODE and AS are the mode and address space of the memory reference; |
| 1146 | OUTER_CODE and INDEX_CODE give the context that the rtx appears in. |
| 1147 | These four arguments are passed down to base_reg_class. |
| 1148 | |
| 1149 | SCALE is twice the amount to multiply the cost by (it is twice so |
| 1150 | we can represent half-cost adjustments). */ |
| 1151 | static void |
| 1152 | record_address_regs (machine_mode mode, addr_space_t as, rtx x, |
| 1153 | int context, enum rtx_code outer_code, |
| 1154 | enum rtx_code index_code, int scale) |
| 1155 | { |
| 1156 | enum rtx_code code = GET_CODE (x); |
| 1157 | enum reg_class rclass; |
| 1158 | |
| 1159 | if (context == 1) |
| 1160 | rclass = INDEX_REG_CLASS; |
| 1161 | else |
| 1162 | rclass = base_reg_class (mode, as, outer_code, index_code); |
| 1163 | |
| 1164 | switch (code) |
| 1165 | { |
| 1166 | case CONST_INT: |
| 1167 | case CONST: |
| 1168 | case PC: |
| 1169 | case SYMBOL_REF: |
| 1170 | case LABEL_REF: |
| 1171 | return; |
| 1172 | |
| 1173 | case PLUS: |
| 1174 | /* When we have an address that is a sum, we must determine |
| 1175 | whether registers are "base" or "index" regs. If there is a |
| 1176 | sum of two registers, we must choose one to be the "base". |
| 1177 | Luckily, we can use the REG_POINTER to make a good choice |
| 1178 | most of the time. We only need to do this on machines that |
| 1179 | can have two registers in an address and where the base and |
| 1180 | index register classes are different. |
| 1181 | |
| 1182 | ??? This code used to set REGNO_POINTER_FLAG in some cases, |
| 1183 | but that seems bogus since it should only be set when we are |
| 1184 | sure the register is being used as a pointer. */ |
| 1185 | { |
| 1186 | rtx arg0 = XEXP (x, 0); |
| 1187 | rtx arg1 = XEXP (x, 1); |
| 1188 | enum rtx_code code0 = GET_CODE (arg0); |
| 1189 | enum rtx_code code1 = GET_CODE (arg1); |
| 1190 | |
| 1191 | /* Look inside subregs. */ |
| 1192 | if (code0 == SUBREG) |
| 1193 | arg0 = SUBREG_REG (arg0), code0 = GET_CODE (arg0); |
| 1194 | if (code1 == SUBREG) |
| 1195 | arg1 = SUBREG_REG (arg1), code1 = GET_CODE (arg1); |
| 1196 | |
| 1197 | /* If index registers do not appear, or coincide with base registers, |
| 1198 | just record registers in any non-constant operands. We |
| 1199 | assume here, as well as in the tests below, that all |
| 1200 | addresses are in canonical form. */ |
| 1201 | if (MAX_REGS_PER_ADDRESS == 1 |
| 1202 | || INDEX_REG_CLASS == base_reg_class (VOIDmode, as, outer_code: PLUS, index_code: SCRATCH)) |
| 1203 | { |
| 1204 | record_address_regs (mode, as, x: arg0, context, outer_code: PLUS, index_code: code1, scale); |
| 1205 | if (! CONSTANT_P (arg1)) |
| 1206 | record_address_regs (mode, as, x: arg1, context, outer_code: PLUS, index_code: code0, scale); |
| 1207 | } |
| 1208 | |
| 1209 | /* If the second operand is a constant integer, it doesn't |
| 1210 | change what class the first operand must be. */ |
| 1211 | else if (CONST_SCALAR_INT_P (arg1)) |
| 1212 | record_address_regs (mode, as, x: arg0, context, outer_code: PLUS, index_code: code1, scale); |
| 1213 | /* If the second operand is a symbolic constant, the first |
| 1214 | operand must be an index register. */ |
| 1215 | else if (code1 == SYMBOL_REF || code1 == CONST || code1 == LABEL_REF) |
| 1216 | record_address_regs (mode, as, x: arg0, context: 1, outer_code: PLUS, index_code: code1, scale); |
| 1217 | /* If both operands are registers but one is already a hard |
| 1218 | register of index or reg-base class, give the other the |
| 1219 | class that the hard register is not. */ |
| 1220 | else if (code0 == REG && code1 == REG |
| 1221 | && REGNO (arg0) < FIRST_PSEUDO_REGISTER |
| 1222 | && (ok_for_base_p_nonstrict (reg: arg0, mode, as, outer_code: PLUS, index_code: REG) |
| 1223 | || ok_for_index_p_nonstrict (reg: arg0))) |
| 1224 | record_address_regs (mode, as, x: arg1, |
| 1225 | context: ok_for_base_p_nonstrict (reg: arg0, mode, as, |
| 1226 | outer_code: PLUS, index_code: REG) ? 1 : 0, |
| 1227 | outer_code: PLUS, index_code: REG, scale); |
| 1228 | else if (code0 == REG && code1 == REG |
| 1229 | && REGNO (arg1) < FIRST_PSEUDO_REGISTER |
| 1230 | && (ok_for_base_p_nonstrict (reg: arg1, mode, as, outer_code: PLUS, index_code: REG) |
| 1231 | || ok_for_index_p_nonstrict (reg: arg1))) |
| 1232 | record_address_regs (mode, as, x: arg0, |
| 1233 | context: ok_for_base_p_nonstrict (reg: arg1, mode, as, |
| 1234 | outer_code: PLUS, index_code: REG) ? 1 : 0, |
| 1235 | outer_code: PLUS, index_code: REG, scale); |
| 1236 | /* If one operand is known to be a pointer, it must be the |
| 1237 | base with the other operand the index. Likewise if the |
| 1238 | other operand is a MULT. */ |
| 1239 | else if ((code0 == REG && REG_POINTER (arg0)) || code1 == MULT) |
| 1240 | { |
| 1241 | record_address_regs (mode, as, x: arg0, context: 0, outer_code: PLUS, index_code: code1, scale); |
| 1242 | record_address_regs (mode, as, x: arg1, context: 1, outer_code: PLUS, index_code: code0, scale); |
| 1243 | } |
| 1244 | else if ((code1 == REG && REG_POINTER (arg1)) || code0 == MULT) |
| 1245 | { |
| 1246 | record_address_regs (mode, as, x: arg0, context: 1, outer_code: PLUS, index_code: code1, scale); |
| 1247 | record_address_regs (mode, as, x: arg1, context: 0, outer_code: PLUS, index_code: code0, scale); |
| 1248 | } |
| 1249 | /* Otherwise, count equal chances that each might be a base or |
| 1250 | index register. This case should be rare. */ |
| 1251 | else |
| 1252 | { |
| 1253 | record_address_regs (mode, as, x: arg0, context: 0, outer_code: PLUS, index_code: code1, scale: scale / 2); |
| 1254 | record_address_regs (mode, as, x: arg0, context: 1, outer_code: PLUS, index_code: code1, scale: scale / 2); |
| 1255 | record_address_regs (mode, as, x: arg1, context: 0, outer_code: PLUS, index_code: code0, scale: scale / 2); |
| 1256 | record_address_regs (mode, as, x: arg1, context: 1, outer_code: PLUS, index_code: code0, scale: scale / 2); |
| 1257 | } |
| 1258 | } |
| 1259 | break; |
| 1260 | |
| 1261 | /* Double the importance of an allocno that is incremented or |
| 1262 | decremented, since it would take two extra insns if it ends |
| 1263 | up in the wrong place. */ |
| 1264 | case POST_MODIFY: |
| 1265 | case PRE_MODIFY: |
| 1266 | record_address_regs (mode, as, XEXP (x, 0), context: 0, outer_code: code, |
| 1267 | GET_CODE (XEXP (XEXP (x, 1), 1)), scale: 2 * scale); |
| 1268 | if (REG_P (XEXP (XEXP (x, 1), 1))) |
| 1269 | record_address_regs (mode, as, XEXP (XEXP (x, 1), 1), context: 1, outer_code: code, index_code: REG, |
| 1270 | scale: 2 * scale); |
| 1271 | break; |
| 1272 | |
| 1273 | case POST_INC: |
| 1274 | case PRE_INC: |
| 1275 | case POST_DEC: |
| 1276 | case PRE_DEC: |
| 1277 | /* Double the importance of an allocno that is incremented or |
| 1278 | decremented, since it would take two extra insns if it ends |
| 1279 | up in the wrong place. */ |
| 1280 | record_address_regs (mode, as, XEXP (x, 0), context: 0, outer_code: code, index_code: SCRATCH, scale: 2 * scale); |
| 1281 | break; |
| 1282 | |
| 1283 | case REG: |
| 1284 | { |
| 1285 | struct costs *pp; |
| 1286 | int *pp_costs; |
| 1287 | enum reg_class i; |
| 1288 | int k, regno, add_cost; |
| 1289 | cost_classes_t cost_classes_ptr; |
| 1290 | enum reg_class *cost_classes; |
| 1291 | move_table *move_in_cost; |
| 1292 | |
| 1293 | if (REGNO (x) < FIRST_PSEUDO_REGISTER) |
| 1294 | break; |
| 1295 | |
| 1296 | regno = REGNO (x); |
| 1297 | if (allocno_p) |
| 1298 | ALLOCNO_BAD_SPILL_P (ira_curr_regno_allocno_map[regno]) = true; |
| 1299 | pp = COSTS (costs, COST_INDEX (regno)); |
| 1300 | add_cost = (ira_memory_move_cost[Pmode][rclass][1] * scale) / 2; |
| 1301 | if (INT_MAX - add_cost < pp->mem_cost) |
| 1302 | pp->mem_cost = INT_MAX; |
| 1303 | else |
| 1304 | pp->mem_cost += add_cost; |
| 1305 | cost_classes_ptr = regno_cost_classes[regno]; |
| 1306 | cost_classes = cost_classes_ptr->classes; |
| 1307 | pp_costs = pp->cost; |
| 1308 | ira_init_register_move_cost_if_necessary (Pmode); |
| 1309 | move_in_cost = ira_may_move_in_cost[Pmode]; |
| 1310 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 1311 | { |
| 1312 | i = cost_classes[k]; |
| 1313 | add_cost = (move_in_cost[i][rclass] * scale) / 2; |
| 1314 | if (INT_MAX - add_cost < pp_costs[k]) |
| 1315 | pp_costs[k] = INT_MAX; |
| 1316 | else |
| 1317 | pp_costs[k] += add_cost; |
| 1318 | } |
| 1319 | } |
| 1320 | break; |
| 1321 | |
| 1322 | default: |
| 1323 | { |
| 1324 | const char *fmt = GET_RTX_FORMAT (code); |
| 1325 | int i; |
| 1326 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
| 1327 | if (fmt[i] == 'e') |
| 1328 | record_address_regs (mode, as, XEXP (x, i), context, outer_code: code, index_code: SCRATCH, |
| 1329 | scale); |
| 1330 | } |
| 1331 | } |
| 1332 | } |
| 1333 | |
| 1334 | |
| 1335 | |
| 1336 | /* Calculate the costs of insn operands. */ |
| 1337 | static void |
| 1338 | record_operand_costs (rtx_insn *insn, enum reg_class *pref) |
| 1339 | { |
| 1340 | const char *constraints[MAX_RECOG_OPERANDS]; |
| 1341 | machine_mode modes[MAX_RECOG_OPERANDS]; |
| 1342 | rtx set; |
| 1343 | int i; |
| 1344 | |
| 1345 | if ((set = single_set (insn)) != NULL_RTX |
| 1346 | /* In rare cases the single set insn might have less 2 operands |
| 1347 | as the source can be a fixed special reg. */ |
| 1348 | && recog_data.n_operands > 1 |
| 1349 | && recog_data.operand[0] == SET_DEST (set) |
| 1350 | && recog_data.operand[1] == SET_SRC (set)) |
| 1351 | { |
| 1352 | int regno, other_regno; |
| 1353 | rtx dest = SET_DEST (set); |
| 1354 | rtx src = SET_SRC (set); |
| 1355 | |
| 1356 | if (GET_CODE (dest) == SUBREG |
| 1357 | && known_eq (GET_MODE_SIZE (GET_MODE (dest)), |
| 1358 | GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest))))) |
| 1359 | dest = SUBREG_REG (dest); |
| 1360 | if (GET_CODE (src) == SUBREG |
| 1361 | && known_eq (GET_MODE_SIZE (GET_MODE (src)), |
| 1362 | GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))))) |
| 1363 | src = SUBREG_REG (src); |
| 1364 | if (REG_P (src) && REG_P (dest) |
| 1365 | && (((regno = REGNO (src)) >= FIRST_PSEUDO_REGISTER |
| 1366 | && (other_regno = REGNO (dest)) < FIRST_PSEUDO_REGISTER) |
| 1367 | || ((regno = REGNO (dest)) >= FIRST_PSEUDO_REGISTER |
| 1368 | && (other_regno = REGNO (src)) < FIRST_PSEUDO_REGISTER))) |
| 1369 | { |
| 1370 | machine_mode mode = GET_MODE (SET_SRC (set)), cost_mode = mode; |
| 1371 | machine_mode hard_reg_mode = GET_MODE(regno_reg_rtx[other_regno]); |
| 1372 | poly_int64 pmode_size = GET_MODE_SIZE (mode); |
| 1373 | poly_int64 phard_reg_mode_size = GET_MODE_SIZE (mode: hard_reg_mode); |
| 1374 | HOST_WIDE_INT mode_size, hard_reg_mode_size; |
| 1375 | cost_classes_t cost_classes_ptr = regno_cost_classes[regno]; |
| 1376 | enum reg_class *cost_classes = cost_classes_ptr->classes; |
| 1377 | reg_class_t rclass, hard_reg_class, bigger_hard_reg_class; |
| 1378 | int cost_factor = 1, cost, k; |
| 1379 | move_table *move_costs; |
| 1380 | bool dead_p = find_regno_note (insn, REG_DEAD, REGNO (src)); |
| 1381 | |
| 1382 | hard_reg_class = REGNO_REG_CLASS (other_regno); |
| 1383 | bigger_hard_reg_class = ira_pressure_class_translate[hard_reg_class]; |
| 1384 | /* Target code may return any cost for mode which does not fit the |
| 1385 | hard reg class (e.g. DImode for AREG on i386). Check this and use |
| 1386 | a bigger class to get the right cost. */ |
| 1387 | if (bigger_hard_reg_class != NO_REGS |
| 1388 | && ! ira_hard_reg_in_set_p (hard_regno: other_regno, mode, |
| 1389 | reg_class_contents[hard_reg_class])) |
| 1390 | hard_reg_class = bigger_hard_reg_class; |
| 1391 | ira_init_register_move_cost_if_necessary (mode); |
| 1392 | ira_init_register_move_cost_if_necessary (mode: hard_reg_mode); |
| 1393 | /* Use smaller movement cost for natural hard reg mode or its mode as |
| 1394 | operand. */ |
| 1395 | if (pmode_size.is_constant (const_value: &mode_size) |
| 1396 | && phard_reg_mode_size.is_constant (const_value: &hard_reg_mode_size)) |
| 1397 | { |
| 1398 | /* Assume we are moving in the natural modes: */ |
| 1399 | cost_factor = mode_size / hard_reg_mode_size; |
| 1400 | if (mode_size % hard_reg_mode_size != 0) |
| 1401 | cost_factor++; |
| 1402 | if (cost_factor |
| 1403 | * (ira_register_move_cost |
| 1404 | [hard_reg_mode][hard_reg_class][hard_reg_class]) |
| 1405 | < (ira_register_move_cost |
| 1406 | [mode][hard_reg_class][hard_reg_class])) |
| 1407 | cost_mode = hard_reg_mode; |
| 1408 | else |
| 1409 | cost_factor = 1; |
| 1410 | } |
| 1411 | move_costs = ira_register_move_cost[cost_mode]; |
| 1412 | i = regno == (int) REGNO (src) ? 1 : 0; |
| 1413 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 1414 | { |
| 1415 | rclass = cost_classes[k]; |
| 1416 | cost = (i == 0 |
| 1417 | ? move_costs[hard_reg_class][rclass] |
| 1418 | : move_costs[rclass][hard_reg_class]); |
| 1419 | cost *= cost_factor; |
| 1420 | op_costs[i]->cost[k] = cost * frequency; |
| 1421 | /* If this insn is a single set copying operand 1 to |
| 1422 | operand 0 and one operand is an allocno with the |
| 1423 | other a hard reg or an allocno that prefers a hard |
| 1424 | register that is in its own register class then we |
| 1425 | may want to adjust the cost of that register class to |
| 1426 | -1. |
| 1427 | |
| 1428 | Avoid the adjustment if the source does not die to |
| 1429 | avoid stressing of register allocator by preferencing |
| 1430 | two colliding registers into single class. */ |
| 1431 | if (dead_p |
| 1432 | && TEST_HARD_REG_BIT (reg_class_contents[rclass], bit: other_regno) |
| 1433 | && (reg_class_size[(int) rclass] |
| 1434 | == (ira_reg_class_max_nregs |
| 1435 | [(int) rclass][(int) GET_MODE(src)]))) |
| 1436 | { |
| 1437 | if (reg_class_size[rclass] == 1) |
| 1438 | op_costs[i]->cost[k] = -frequency; |
| 1439 | else if (in_hard_reg_set_p (reg_class_contents[rclass], |
| 1440 | GET_MODE(src), regno: other_regno)) |
| 1441 | op_costs[i]->cost[k] = -frequency; |
| 1442 | } |
| 1443 | } |
| 1444 | op_costs[i]->mem_cost |
| 1445 | = ira_memory_move_cost[mode][hard_reg_class][i] * frequency; |
| 1446 | return; |
| 1447 | } |
| 1448 | } |
| 1449 | |
| 1450 | for (i = 0; i < recog_data.n_operands; i++) |
| 1451 | { |
| 1452 | constraints[i] = recog_data.constraints[i]; |
| 1453 | modes[i] = recog_data.operand_mode[i]; |
| 1454 | } |
| 1455 | |
| 1456 | /* If we get here, we are set up to record the costs of all the |
| 1457 | operands for this insn. Start by initializing the costs. Then |
| 1458 | handle any address registers. Finally record the desired classes |
| 1459 | for any allocnos, doing it twice if some pair of operands are |
| 1460 | commutative. */ |
| 1461 | for (i = 0; i < recog_data.n_operands; i++) |
| 1462 | { |
| 1463 | rtx op_mem = extract_mem_from_operand (recog_data.operand[i]); |
| 1464 | memcpy (op_costs[i], init_cost, n: struct_costs_size); |
| 1465 | |
| 1466 | if (GET_CODE (recog_data.operand[i]) == SUBREG) |
| 1467 | recog_data.operand[i] = SUBREG_REG (recog_data.operand[i]); |
| 1468 | |
| 1469 | if (MEM_P (op_mem)) |
| 1470 | record_address_regs (GET_MODE (op_mem), |
| 1471 | MEM_ADDR_SPACE (op_mem), |
| 1472 | XEXP (op_mem, 0), |
| 1473 | context: 0, outer_code: MEM, index_code: SCRATCH, scale: frequency * 2); |
| 1474 | else if (constraints[i][0] == 'p' |
| 1475 | || (insn_extra_address_constraint |
| 1476 | (c: lookup_constraint (p: constraints[i])))) |
| 1477 | record_address_regs (VOIDmode, ADDR_SPACE_GENERIC, |
| 1478 | x: recog_data.operand[i], context: 0, outer_code: ADDRESS, index_code: SCRATCH, |
| 1479 | scale: frequency * 2); |
| 1480 | } |
| 1481 | |
| 1482 | /* Check for commutative in a separate loop so everything will have |
| 1483 | been initialized. We must do this even if one operand is a |
| 1484 | constant--see addsi3 in m68k.md. */ |
| 1485 | for (i = 0; i < (int) recog_data.n_operands - 1; i++) |
| 1486 | if (constraints[i][0] == '%') |
| 1487 | { |
| 1488 | const char *xconstraints[MAX_RECOG_OPERANDS]; |
| 1489 | int j; |
| 1490 | |
| 1491 | /* Handle commutative operands by swapping the |
| 1492 | constraints. We assume the modes are the same. */ |
| 1493 | for (j = 0; j < recog_data.n_operands; j++) |
| 1494 | xconstraints[j] = constraints[j]; |
| 1495 | |
| 1496 | xconstraints[i] = constraints[i+1]; |
| 1497 | xconstraints[i+1] = constraints[i]; |
| 1498 | record_reg_classes (n_alts: recog_data.n_alternatives, n_ops: recog_data.n_operands, |
| 1499 | ops: recog_data.operand, modes, |
| 1500 | constraints: xconstraints, insn, pref); |
| 1501 | } |
| 1502 | record_reg_classes (n_alts: recog_data.n_alternatives, n_ops: recog_data.n_operands, |
| 1503 | ops: recog_data.operand, modes, |
| 1504 | constraints, insn, pref); |
| 1505 | } |
| 1506 | |
| 1507 | |
| 1508 | |
| 1509 | /* Process one insn INSN. Scan it and record each time it would save |
| 1510 | code to put a certain allocnos in a certain class. Return the last |
| 1511 | insn processed, so that the scan can be continued from there. */ |
| 1512 | static rtx_insn * |
| 1513 | scan_one_insn (rtx_insn *insn) |
| 1514 | { |
| 1515 | enum rtx_code pat_code; |
| 1516 | rtx set, note; |
| 1517 | int i, k; |
| 1518 | bool counted_mem; |
| 1519 | |
| 1520 | if (!NONDEBUG_INSN_P (insn)) |
| 1521 | return insn; |
| 1522 | |
| 1523 | pat_code = GET_CODE (PATTERN (insn)); |
| 1524 | if (pat_code == ASM_INPUT) |
| 1525 | return insn; |
| 1526 | |
| 1527 | /* If INSN is a USE/CLOBBER of a pseudo in a mode M then go ahead |
| 1528 | and initialize the register move costs of mode M. |
| 1529 | |
| 1530 | The pseudo may be related to another pseudo via a copy (implicit or |
| 1531 | explicit) and if there are no mode M uses/sets of the original |
| 1532 | pseudo, then we may leave the register move costs uninitialized for |
| 1533 | mode M. */ |
| 1534 | if (pat_code == USE || pat_code == CLOBBER) |
| 1535 | { |
| 1536 | rtx x = XEXP (PATTERN (insn), 0); |
| 1537 | if (GET_CODE (x) == REG |
| 1538 | && REGNO (x) >= FIRST_PSEUDO_REGISTER |
| 1539 | && have_regs_of_mode[GET_MODE (x)]) |
| 1540 | ira_init_register_move_cost_if_necessary (GET_MODE (x)); |
| 1541 | return insn; |
| 1542 | } |
| 1543 | |
| 1544 | counted_mem = false; |
| 1545 | set = single_set (insn); |
| 1546 | extract_insn (insn); |
| 1547 | |
| 1548 | /* If this insn loads a parameter from its stack slot, then it |
| 1549 | represents a savings, rather than a cost, if the parameter is |
| 1550 | stored in memory. Record this fact. |
| 1551 | |
| 1552 | Similarly if we're loading other constants from memory (constant |
| 1553 | pool, TOC references, small data areas, etc) and this is the only |
| 1554 | assignment to the destination pseudo. |
| 1555 | |
| 1556 | Don't do this if SET_SRC (set) isn't a general operand, if it is |
| 1557 | a memory requiring special instructions to load it, decreasing |
| 1558 | mem_cost might result in it being loaded using the specialized |
| 1559 | instruction into a register, then stored into stack and loaded |
| 1560 | again from the stack. See PR52208. |
| 1561 | |
| 1562 | Don't do this if SET_SRC (set) has side effect. See PR56124. */ |
| 1563 | if (set != 0 && REG_P (SET_DEST (set)) && MEM_P (SET_SRC (set)) |
| 1564 | && (note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) != NULL_RTX |
| 1565 | && ((MEM_P (XEXP (note, 0)) |
| 1566 | && !side_effects_p (SET_SRC (set))) |
| 1567 | || (CONSTANT_P (XEXP (note, 0)) |
| 1568 | && targetm.legitimate_constant_p (GET_MODE (SET_DEST (set)), |
| 1569 | XEXP (note, 0)) |
| 1570 | && REG_N_SETS (REGNO (SET_DEST (set))) == 1)) |
| 1571 | && general_operand (SET_SRC (set), GET_MODE (SET_SRC (set))) |
| 1572 | /* LRA does not use equiv with a symbol for PIC code. */ |
| 1573 | && (! ira_use_lra_p || ! pic_offset_table_rtx |
| 1574 | || ! contains_symbol_ref_p (XEXP (note, 0)))) |
| 1575 | { |
| 1576 | enum reg_class cl = GENERAL_REGS; |
| 1577 | rtx reg = SET_DEST (set); |
| 1578 | int num = COST_INDEX (REGNO (reg)); |
| 1579 | /* Costs for NO_REGS are used in cost calculation on the |
| 1580 | 1st pass when the preferred register classes are not |
| 1581 | known yet. In this case we take the best scenario when |
| 1582 | mode can't be put into GENERAL_REGS. */ |
| 1583 | if (!targetm.hard_regno_mode_ok (ira_class_hard_regs[cl][0], |
| 1584 | GET_MODE (reg))) |
| 1585 | cl = NO_REGS; |
| 1586 | |
| 1587 | COSTS (costs, num)->mem_cost |
| 1588 | -= ira_memory_move_cost[GET_MODE (reg)][cl][1] * frequency; |
| 1589 | record_address_regs (GET_MODE (SET_SRC (set)), |
| 1590 | MEM_ADDR_SPACE (SET_SRC (set)), |
| 1591 | XEXP (SET_SRC (set), 0), context: 0, outer_code: MEM, index_code: SCRATCH, |
| 1592 | scale: frequency * 2); |
| 1593 | counted_mem = true; |
| 1594 | } |
| 1595 | |
| 1596 | record_operand_costs (insn, pref); |
| 1597 | |
| 1598 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5) |
| 1599 | { |
| 1600 | const char *p; |
| 1601 | fprintf (stream: ira_dump_file, format: " Final costs after insn %u" , INSN_UID (insn)); |
| 1602 | if (INSN_CODE (insn) >= 0 |
| 1603 | && (p = get_insn_name (INSN_CODE (insn))) != NULL) |
| 1604 | fprintf (stream: ira_dump_file, format: " {%s}" , p); |
| 1605 | fprintf (stream: ira_dump_file, format: " (freq=%d)\n" , |
| 1606 | REG_FREQ_FROM_BB (BLOCK_FOR_INSN (insn))); |
| 1607 | dump_insn_slim (ira_dump_file, insn); |
| 1608 | } |
| 1609 | |
| 1610 | /* Now add the cost for each operand to the total costs for its |
| 1611 | allocno. */ |
| 1612 | for (i = 0; i < recog_data.n_operands; i++) |
| 1613 | { |
| 1614 | rtx op = recog_data.operand[i]; |
| 1615 | |
| 1616 | if (GET_CODE (op) == SUBREG) |
| 1617 | op = SUBREG_REG (op); |
| 1618 | if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER) |
| 1619 | { |
| 1620 | int regno = REGNO (op); |
| 1621 | struct costs *p = COSTS (costs, COST_INDEX (regno)); |
| 1622 | struct costs *q = op_costs[i]; |
| 1623 | int *p_costs = p->cost, *q_costs = q->cost; |
| 1624 | cost_classes_t cost_classes_ptr = regno_cost_classes[regno]; |
| 1625 | int add_cost = 0; |
| 1626 | |
| 1627 | /* If the already accounted for the memory "cost" above, don't |
| 1628 | do so again. */ |
| 1629 | if (!counted_mem) |
| 1630 | { |
| 1631 | add_cost = q->mem_cost; |
| 1632 | if (add_cost > 0 && INT_MAX - add_cost < p->mem_cost) |
| 1633 | p->mem_cost = INT_MAX; |
| 1634 | else |
| 1635 | p->mem_cost += add_cost; |
| 1636 | } |
| 1637 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5) |
| 1638 | { |
| 1639 | fprintf (stream: ira_dump_file, format: " op %d(r=%u) MEM:%d(+%d)" , |
| 1640 | i, REGNO(op), p->mem_cost, add_cost); |
| 1641 | } |
| 1642 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 1643 | { |
| 1644 | add_cost = q_costs[k]; |
| 1645 | if (add_cost > 0 && INT_MAX - add_cost < p_costs[k]) |
| 1646 | p_costs[k] = INT_MAX; |
| 1647 | else |
| 1648 | p_costs[k] += add_cost; |
| 1649 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5) |
| 1650 | { |
| 1651 | fprintf (stream: ira_dump_file, format: " %s:%d(+%d)" , |
| 1652 | reg_class_names[cost_classes_ptr->classes[k]], |
| 1653 | p_costs[k], add_cost); |
| 1654 | } |
| 1655 | } |
| 1656 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5) |
| 1657 | fprintf (stream: ira_dump_file, format: "\n" ); |
| 1658 | } |
| 1659 | } |
| 1660 | return insn; |
| 1661 | } |
| 1662 | |
| 1663 | |
| 1664 | |
| 1665 | /* Print allocnos costs to the dump file. */ |
| 1666 | static void |
| 1667 | print_allocno_costs (void) |
| 1668 | { |
| 1669 | int k; |
| 1670 | ira_allocno_t a; |
| 1671 | ira_allocno_iterator ai; |
| 1672 | |
| 1673 | ira_assert (allocno_p); |
| 1674 | fprintf (stream: ira_dump_file, format: "\n" ); |
| 1675 | FOR_EACH_ALLOCNO (a, ai) |
| 1676 | { |
| 1677 | int i, rclass; |
| 1678 | basic_block bb; |
| 1679 | int regno = ALLOCNO_REGNO (a); |
| 1680 | cost_classes_t cost_classes_ptr = regno_cost_classes[regno]; |
| 1681 | enum reg_class *cost_classes = cost_classes_ptr->classes; |
| 1682 | |
| 1683 | i = ALLOCNO_NUM (a); |
| 1684 | fprintf (stream: ira_dump_file, format: " a%d(r%d," , i, regno); |
| 1685 | if ((bb = ALLOCNO_LOOP_TREE_NODE (a)->bb) != NULL) |
| 1686 | fprintf (stream: ira_dump_file, format: "b%d" , bb->index); |
| 1687 | else |
| 1688 | fprintf (stream: ira_dump_file, format: "l%d" , ALLOCNO_LOOP_TREE_NODE (a)->loop_num); |
| 1689 | fprintf (stream: ira_dump_file, format: ") costs:" ); |
| 1690 | for (k = 0; k < cost_classes_ptr->num; k++) |
| 1691 | { |
| 1692 | rclass = cost_classes[k]; |
| 1693 | fprintf (stream: ira_dump_file, format: " %s:%d" , reg_class_names[rclass], |
| 1694 | COSTS (costs, i)->cost[k]); |
| 1695 | if (flag_ira_region == IRA_REGION_ALL |
| 1696 | || flag_ira_region == IRA_REGION_MIXED) |
| 1697 | fprintf (stream: ira_dump_file, format: ",%d" , |
| 1698 | COSTS (total_allocno_costs, i)->cost[k]); |
| 1699 | } |
| 1700 | fprintf (stream: ira_dump_file, format: " MEM:%i" , COSTS (costs, i)->mem_cost); |
| 1701 | if (flag_ira_region == IRA_REGION_ALL |
| 1702 | || flag_ira_region == IRA_REGION_MIXED) |
| 1703 | fprintf (stream: ira_dump_file, format: ",%d" , |
| 1704 | COSTS (total_allocno_costs, i)->mem_cost); |
| 1705 | fprintf (stream: ira_dump_file, format: "\n" ); |
| 1706 | } |
| 1707 | } |
| 1708 | |
| 1709 | /* Print pseudo costs to the dump file. */ |
| 1710 | static void |
| 1711 | print_pseudo_costs (void) |
| 1712 | { |
| 1713 | int regno, k; |
| 1714 | int rclass; |
| 1715 | cost_classes_t cost_classes_ptr; |
| 1716 | enum reg_class *cost_classes; |
| 1717 | |
| 1718 | ira_assert (! allocno_p); |
| 1719 | fprintf (stream: ira_dump_file, format: "\n" ); |
| 1720 | for (regno = max_reg_num () - 1; regno >= FIRST_PSEUDO_REGISTER; regno--) |
| 1721 | { |
| 1722 | if (REG_N_REFS (regno) <= 0) |
| 1723 | continue; |
| 1724 | cost_classes_ptr = regno_cost_classes[regno]; |
| 1725 | cost_classes = cost_classes_ptr->classes; |
| 1726 | fprintf (stream: ira_dump_file, format: " r%d costs:" , regno); |
| 1727 | for (k = 0; k < cost_classes_ptr->num; k++) |
| 1728 | { |
| 1729 | rclass = cost_classes[k]; |
| 1730 | fprintf (stream: ira_dump_file, format: " %s:%d" , reg_class_names[rclass], |
| 1731 | COSTS (costs, regno)->cost[k]); |
| 1732 | } |
| 1733 | fprintf (stream: ira_dump_file, format: " MEM:%i\n" , COSTS (costs, regno)->mem_cost); |
| 1734 | } |
| 1735 | } |
| 1736 | |
| 1737 | /* Traverse the BB represented by LOOP_TREE_NODE to update the allocno |
| 1738 | costs. */ |
| 1739 | static void |
| 1740 | process_bb_for_costs (basic_block bb) |
| 1741 | { |
| 1742 | rtx_insn *insn; |
| 1743 | |
| 1744 | frequency = REG_FREQ_FROM_BB (bb); |
| 1745 | if (frequency == 0) |
| 1746 | frequency = 1; |
| 1747 | FOR_BB_INSNS (bb, insn) |
| 1748 | insn = scan_one_insn (insn); |
| 1749 | } |
| 1750 | |
| 1751 | /* Traverse the BB represented by LOOP_TREE_NODE to update the allocno |
| 1752 | costs. */ |
| 1753 | static void |
| 1754 | process_bb_node_for_costs (ira_loop_tree_node_t loop_tree_node) |
| 1755 | { |
| 1756 | basic_block bb; |
| 1757 | |
| 1758 | bb = loop_tree_node->bb; |
| 1759 | if (bb != NULL) |
| 1760 | process_bb_for_costs (bb); |
| 1761 | } |
| 1762 | |
| 1763 | /* Return true if all autoinc rtx in X change only a register and memory is |
| 1764 | valid. */ |
| 1765 | static bool |
| 1766 | validate_autoinc_and_mem_addr_p (rtx x) |
| 1767 | { |
| 1768 | enum rtx_code code = GET_CODE (x); |
| 1769 | if (GET_RTX_CLASS (code) == RTX_AUTOINC) |
| 1770 | return REG_P (XEXP (x, 0)); |
| 1771 | const char *fmt = GET_RTX_FORMAT (code); |
| 1772 | for (int i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
| 1773 | if (fmt[i] == 'e') |
| 1774 | { |
| 1775 | if (!validate_autoinc_and_mem_addr_p (XEXP (x, i))) |
| 1776 | return false; |
| 1777 | } |
| 1778 | else if (fmt[i] == 'E') |
| 1779 | { |
| 1780 | for (int j = 0; j < XVECLEN (x, i); j++) |
| 1781 | if (!validate_autoinc_and_mem_addr_p (XVECEXP (x, i, j))) |
| 1782 | return false; |
| 1783 | } |
| 1784 | /* Check memory after checking autoinc to guarantee that autoinc is already |
| 1785 | valid for machine-dependent code checking memory address. */ |
| 1786 | return (!MEM_P (x) |
| 1787 | || memory_address_addr_space_p (GET_MODE (x), XEXP (x, 0), |
| 1788 | MEM_ADDR_SPACE (x))); |
| 1789 | } |
| 1790 | |
| 1791 | /* Check that reg REGNO in INSN can be changed by TO (which is an invariant |
| 1792 | equiv when INVARIANT_P is true). Return true in case the result insn would |
| 1793 | be valid one. */ |
| 1794 | static bool |
| 1795 | equiv_can_be_consumed_p (int regno, rtx to, rtx_insn *insn, bool invariant_p) |
| 1796 | { |
| 1797 | validate_replace_src_group (regno_reg_rtx[regno], to, insn); |
| 1798 | /* We can change register to equivalent memory in autoinc rtl. Some code |
| 1799 | including verify_changes assumes that autoinc contains only a register. |
| 1800 | So check this first. */ |
| 1801 | bool res = validate_autoinc_and_mem_addr_p (x: PATTERN (insn)); |
| 1802 | if (res) |
| 1803 | res = verify_changes (0); |
| 1804 | cancel_changes (0); |
| 1805 | if (!res && invariant_p) |
| 1806 | { |
| 1807 | /* Here we use more expensive code for the invariant because we need to |
| 1808 | simplify the result insn as the invariant can be arithmetic rtx |
| 1809 | inserted into another arithmetic rtx, e.g. into memory address. */ |
| 1810 | rtx pat = PATTERN (insn); |
| 1811 | int code = INSN_CODE (insn); |
| 1812 | PATTERN (insn) = copy_rtx (pat); |
| 1813 | PATTERN (insn) |
| 1814 | = simplify_replace_rtx (PATTERN (insn), regno_reg_rtx[regno], to); |
| 1815 | res = !insn_invalid_p (insn, false); |
| 1816 | PATTERN (insn) = pat; |
| 1817 | INSN_CODE (insn) = code; |
| 1818 | } |
| 1819 | return res; |
| 1820 | } |
| 1821 | |
| 1822 | /* Return true if X contains a pseudo with equivalence. In this case also |
| 1823 | return the pseudo through parameter REG. If the pseudo is a part of subreg, |
| 1824 | return the subreg through parameter SUBREG. */ |
| 1825 | |
| 1826 | static bool |
| 1827 | get_equiv_regno (rtx x, int ®no, rtx &subreg) |
| 1828 | { |
| 1829 | subreg = NULL_RTX; |
| 1830 | if (GET_CODE (x) == SUBREG) |
| 1831 | { |
| 1832 | subreg = x; |
| 1833 | x = SUBREG_REG (x); |
| 1834 | } |
| 1835 | if (REG_P (x) |
| 1836 | && (ira_reg_equiv[REGNO (x)].memory != NULL |
| 1837 | || ira_reg_equiv[REGNO (x)].invariant != NULL |
| 1838 | || ira_reg_equiv[REGNO (x)].constant != NULL)) |
| 1839 | { |
| 1840 | regno = REGNO (x); |
| 1841 | return true; |
| 1842 | } |
| 1843 | RTX_CODE code = GET_CODE (x); |
| 1844 | const char *fmt = GET_RTX_FORMAT (code); |
| 1845 | |
| 1846 | for (int i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
| 1847 | if (fmt[i] == 'e') |
| 1848 | { |
| 1849 | if (get_equiv_regno (XEXP (x, i), regno, subreg)) |
| 1850 | return true; |
| 1851 | } |
| 1852 | else if (fmt[i] == 'E') |
| 1853 | { |
| 1854 | for (int j = 0; j < XVECLEN (x, i); j++) |
| 1855 | if (get_equiv_regno (XVECEXP (x, i, j), regno, subreg)) |
| 1856 | return true; |
| 1857 | } |
| 1858 | return false; |
| 1859 | } |
| 1860 | |
| 1861 | /* A pass through the current function insns. Calculate costs of using |
| 1862 | equivalences for pseudos and store them in regno_equiv_gains. */ |
| 1863 | |
| 1864 | static void |
| 1865 | calculate_equiv_gains (void) |
| 1866 | { |
| 1867 | basic_block bb; |
| 1868 | int regno, freq, cost; |
| 1869 | rtx subreg; |
| 1870 | rtx_insn *insn; |
| 1871 | machine_mode mode; |
| 1872 | enum reg_class rclass; |
| 1873 | bitmap_head equiv_pseudos; |
| 1874 | |
| 1875 | ira_assert (allocno_p); |
| 1876 | bitmap_initialize (head: &equiv_pseudos, obstack: ®_obstack); |
| 1877 | for (regno = max_reg_num () - 1; regno >= FIRST_PSEUDO_REGISTER; regno--) |
| 1878 | if (ira_reg_equiv[regno].init_insns != NULL |
| 1879 | && (ira_reg_equiv[regno].memory != NULL |
| 1880 | || ira_reg_equiv[regno].invariant != NULL |
| 1881 | || (ira_reg_equiv[regno].constant != NULL |
| 1882 | /* Ignore complicated constants which probably will be placed |
| 1883 | in memory: */ |
| 1884 | && GET_CODE (ira_reg_equiv[regno].constant) != CONST_DOUBLE |
| 1885 | && GET_CODE (ira_reg_equiv[regno].constant) != CONST_VECTOR |
| 1886 | && GET_CODE (ira_reg_equiv[regno].constant) != LABEL_REF))) |
| 1887 | { |
| 1888 | rtx_insn_list *x; |
| 1889 | for (x = ira_reg_equiv[regno].init_insns; x != NULL; x = x->next ()) |
| 1890 | { |
| 1891 | insn = x->insn (); |
| 1892 | rtx set = single_set (insn); |
| 1893 | |
| 1894 | if (set == NULL_RTX || SET_DEST (set) != regno_reg_rtx[regno]) |
| 1895 | break; |
| 1896 | bb = BLOCK_FOR_INSN (insn); |
| 1897 | ira_curr_regno_allocno_map |
| 1898 | = ira_bb_nodes[bb->index].parent->regno_allocno_map; |
| 1899 | mode = PSEUDO_REGNO_MODE (regno); |
| 1900 | rclass = pref[COST_INDEX (regno)]; |
| 1901 | ira_init_register_move_cost_if_necessary (mode); |
| 1902 | if (ira_reg_equiv[regno].memory != NULL) |
| 1903 | cost = ira_memory_move_cost[mode][rclass][1]; |
| 1904 | else |
| 1905 | cost = ira_register_move_cost[mode][rclass][rclass]; |
| 1906 | freq = REG_FREQ_FROM_BB (bb); |
| 1907 | regno_equiv_gains[regno] += cost * freq; |
| 1908 | } |
| 1909 | if (x != NULL) |
| 1910 | /* We found complicated equiv or reverse equiv mem=reg. Ignore |
| 1911 | them. */ |
| 1912 | regno_equiv_gains[regno] = 0; |
| 1913 | else |
| 1914 | bitmap_set_bit (&equiv_pseudos, regno); |
| 1915 | } |
| 1916 | |
| 1917 | FOR_EACH_BB_FN (bb, cfun) |
| 1918 | { |
| 1919 | freq = REG_FREQ_FROM_BB (bb); |
| 1920 | ira_curr_regno_allocno_map |
| 1921 | = ira_bb_nodes[bb->index].parent->regno_allocno_map; |
| 1922 | FOR_BB_INSNS (bb, insn) |
| 1923 | { |
| 1924 | if (!NONDEBUG_INSN_P (insn) |
| 1925 | || !get_equiv_regno (x: PATTERN (insn), regno, subreg) |
| 1926 | || !bitmap_bit_p (&equiv_pseudos, regno)) |
| 1927 | continue; |
| 1928 | |
| 1929 | if (ira_reg_equiv[regno].invariant != NULL) |
| 1930 | { |
| 1931 | rtx_insn_list *x = ira_reg_equiv[regno].init_insns; |
| 1932 | for (; x != NULL; x = x->next ()) |
| 1933 | if (insn == x->insn ()) |
| 1934 | break; |
| 1935 | if (x != NULL) |
| 1936 | continue; /* skip equiv init insn for invariant */ |
| 1937 | } |
| 1938 | |
| 1939 | rtx subst = ira_reg_equiv[regno].memory; |
| 1940 | |
| 1941 | if (subst == NULL) |
| 1942 | subst = ira_reg_equiv[regno].constant; |
| 1943 | if (subst == NULL) |
| 1944 | subst = ira_reg_equiv[regno].invariant; |
| 1945 | ira_assert (subst != NULL); |
| 1946 | mode = PSEUDO_REGNO_MODE (regno); |
| 1947 | ira_init_register_move_cost_if_necessary (mode); |
| 1948 | bool consumed_p |
| 1949 | = equiv_can_be_consumed_p (regno, to: subst, insn, |
| 1950 | invariant_p: subst == ira_reg_equiv[regno].invariant); |
| 1951 | |
| 1952 | rclass = pref[COST_INDEX (regno)]; |
| 1953 | if (MEM_P (subst) |
| 1954 | /* If it is a change of constant into double for example, the |
| 1955 | result constant probably will be placed in memory. */ |
| 1956 | || (ira_reg_equiv[regno].invariant == NULL |
| 1957 | && subreg != NULL_RTX |
| 1958 | && !INTEGRAL_MODE_P (GET_MODE (subreg)))) |
| 1959 | cost = ira_memory_move_cost[mode][rclass][1] + (consumed_p ? 0 : 1); |
| 1960 | else if (consumed_p) |
| 1961 | continue; |
| 1962 | else |
| 1963 | cost = ira_register_move_cost[mode][rclass][rclass]; |
| 1964 | regno_equiv_gains[regno] -= cost * freq; |
| 1965 | } |
| 1966 | } |
| 1967 | bitmap_clear (&equiv_pseudos); |
| 1968 | } |
| 1969 | |
| 1970 | /* Find costs of register classes and memory for allocnos or pseudos |
| 1971 | and their best costs. Set up preferred, alternative and allocno |
| 1972 | classes for pseudos. */ |
| 1973 | static void |
| 1974 | find_costs_and_classes (void) |
| 1975 | { |
| 1976 | int i, k, start, max_cost_classes_num; |
| 1977 | int pass; |
| 1978 | basic_block bb; |
| 1979 | enum reg_class *regno_best_class, new_class; |
| 1980 | |
| 1981 | init_recog (); |
| 1982 | regno_best_class |
| 1983 | = (enum reg_class *) ira_allocate (max_reg_num () |
| 1984 | * sizeof (enum reg_class)); |
| 1985 | for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--) |
| 1986 | regno_best_class[i] = NO_REGS; |
| 1987 | if (!resize_reg_info () && allocno_p |
| 1988 | && pseudo_classes_defined_p && flag_expensive_optimizations) |
| 1989 | { |
| 1990 | ira_allocno_t a; |
| 1991 | ira_allocno_iterator ai; |
| 1992 | |
| 1993 | pref = pref_buffer; |
| 1994 | max_cost_classes_num = 1; |
| 1995 | FOR_EACH_ALLOCNO (a, ai) |
| 1996 | { |
| 1997 | pref[ALLOCNO_NUM (a)] = reg_preferred_class (ALLOCNO_REGNO (a)); |
| 1998 | setup_regno_cost_classes_by_aclass |
| 1999 | (ALLOCNO_REGNO (a), aclass: pref[ALLOCNO_NUM (a)]); |
| 2000 | max_cost_classes_num |
| 2001 | = MAX (max_cost_classes_num, |
| 2002 | regno_cost_classes[ALLOCNO_REGNO (a)]->num); |
| 2003 | } |
| 2004 | start = 1; |
| 2005 | } |
| 2006 | else |
| 2007 | { |
| 2008 | pref = NULL; |
| 2009 | max_cost_classes_num = ira_important_classes_num; |
| 2010 | for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--) |
| 2011 | if (regno_reg_rtx[i] != NULL_RTX) |
| 2012 | setup_regno_cost_classes_by_mode (regno: i, PSEUDO_REGNO_MODE (i)); |
| 2013 | else |
| 2014 | setup_regno_cost_classes_by_aclass (regno: i, aclass: ALL_REGS); |
| 2015 | start = 0; |
| 2016 | } |
| 2017 | if (allocno_p) |
| 2018 | /* Clear the flag for the next compiled function. */ |
| 2019 | pseudo_classes_defined_p = false; |
| 2020 | /* Normally we scan the insns once and determine the best class to |
| 2021 | use for each allocno. However, if -fexpensive-optimizations are |
| 2022 | on, we do so twice, the second time using the tentative best |
| 2023 | classes to guide the selection. */ |
| 2024 | for (pass = start; pass <= flag_expensive_optimizations; pass++) |
| 2025 | { |
| 2026 | if ((!allocno_p || internal_flag_ira_verbose > 0) && ira_dump_file) |
| 2027 | fprintf (stream: ira_dump_file, |
| 2028 | format: "\nPass %i for finding pseudo/allocno costs\n\n" , pass); |
| 2029 | |
| 2030 | if (pass != start) |
| 2031 | { |
| 2032 | max_cost_classes_num = 1; |
| 2033 | for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--) |
| 2034 | { |
| 2035 | setup_regno_cost_classes_by_aclass (regno: i, aclass: regno_best_class[i]); |
| 2036 | max_cost_classes_num |
| 2037 | = MAX (max_cost_classes_num, regno_cost_classes[i]->num); |
| 2038 | } |
| 2039 | } |
| 2040 | |
| 2041 | struct_costs_size |
| 2042 | = sizeof (struct costs) + sizeof (int) * (max_cost_classes_num - 1); |
| 2043 | /* Zero out our accumulation of the cost of each class for each |
| 2044 | allocno. */ |
| 2045 | memset (s: costs, c: 0, n: cost_elements_num * struct_costs_size); |
| 2046 | |
| 2047 | if (allocno_p) |
| 2048 | { |
| 2049 | /* Scan the instructions and record each time it would save code |
| 2050 | to put a certain allocno in a certain class. */ |
| 2051 | ira_traverse_loop_tree (true, ira_loop_tree_root, |
| 2052 | process_bb_node_for_costs, NULL); |
| 2053 | |
| 2054 | memcpy (dest: total_allocno_costs, src: costs, |
| 2055 | max_struct_costs_size * ira_allocnos_num); |
| 2056 | } |
| 2057 | else |
| 2058 | { |
| 2059 | basic_block bb; |
| 2060 | |
| 2061 | FOR_EACH_BB_FN (bb, cfun) |
| 2062 | process_bb_for_costs (bb); |
| 2063 | } |
| 2064 | |
| 2065 | if (pass == 0) |
| 2066 | pref = pref_buffer; |
| 2067 | |
| 2068 | if (ira_use_lra_p && allocno_p && pass == 1) |
| 2069 | /* It is a pass through all insns. So do it once and only for RA (not |
| 2070 | for insn scheduler) when we already found preferable pseudo register |
| 2071 | classes on the previous pass. */ |
| 2072 | calculate_equiv_gains (); |
| 2073 | |
| 2074 | /* Now for each allocno look at how desirable each class is and |
| 2075 | find which class is preferred. */ |
| 2076 | for (i = max_reg_num () - 1; i >= FIRST_PSEUDO_REGISTER; i--) |
| 2077 | { |
| 2078 | ira_allocno_t a, parent_a; |
| 2079 | int rclass, a_num, parent_a_num, add_cost; |
| 2080 | ira_loop_tree_node_t parent; |
| 2081 | int best_cost, allocno_cost; |
| 2082 | enum reg_class best, alt_class; |
| 2083 | cost_classes_t cost_classes_ptr = regno_cost_classes[i]; |
| 2084 | enum reg_class *cost_classes; |
| 2085 | int *i_costs = temp_costs->cost; |
| 2086 | int i_mem_cost; |
| 2087 | int equiv_savings = regno_equiv_gains[i]; |
| 2088 | |
| 2089 | if (! allocno_p) |
| 2090 | { |
| 2091 | if (regno_reg_rtx[i] == NULL_RTX) |
| 2092 | continue; |
| 2093 | memcpy (temp_costs, COSTS (costs, i), n: struct_costs_size); |
| 2094 | i_mem_cost = temp_costs->mem_cost; |
| 2095 | cost_classes = cost_classes_ptr->classes; |
| 2096 | } |
| 2097 | else |
| 2098 | { |
| 2099 | if (ira_regno_allocno_map[i] == NULL) |
| 2100 | continue; |
| 2101 | memset (temp_costs, c: 0, n: struct_costs_size); |
| 2102 | i_mem_cost = 0; |
| 2103 | cost_classes = cost_classes_ptr->classes; |
| 2104 | /* Find cost of all allocnos with the same regno. */ |
| 2105 | for (a = ira_regno_allocno_map[i]; |
| 2106 | a != NULL; |
| 2107 | a = ALLOCNO_NEXT_REGNO_ALLOCNO (a)) |
| 2108 | { |
| 2109 | int *a_costs, *p_costs; |
| 2110 | |
| 2111 | a_num = ALLOCNO_NUM (a); |
| 2112 | if ((flag_ira_region == IRA_REGION_ALL |
| 2113 | || flag_ira_region == IRA_REGION_MIXED) |
| 2114 | && (parent = ALLOCNO_LOOP_TREE_NODE (a)->parent) != NULL |
| 2115 | && (parent_a = parent->regno_allocno_map[i]) != NULL |
| 2116 | /* There are no caps yet. */ |
| 2117 | && bitmap_bit_p (ALLOCNO_LOOP_TREE_NODE |
| 2118 | (a)->border_allocnos, |
| 2119 | ALLOCNO_NUM (a))) |
| 2120 | { |
| 2121 | /* Propagate costs to upper levels in the region |
| 2122 | tree. */ |
| 2123 | parent_a_num = ALLOCNO_NUM (parent_a); |
| 2124 | a_costs = COSTS (total_allocno_costs, a_num)->cost; |
| 2125 | p_costs = COSTS (total_allocno_costs, parent_a_num)->cost; |
| 2126 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 2127 | { |
| 2128 | add_cost = a_costs[k]; |
| 2129 | if (add_cost > 0 && INT_MAX - add_cost < p_costs[k]) |
| 2130 | p_costs[k] = INT_MAX; |
| 2131 | else |
| 2132 | p_costs[k] += add_cost; |
| 2133 | } |
| 2134 | add_cost = COSTS (total_allocno_costs, a_num)->mem_cost; |
| 2135 | if (add_cost > 0 |
| 2136 | && (INT_MAX - add_cost |
| 2137 | < COSTS (total_allocno_costs, |
| 2138 | parent_a_num)->mem_cost)) |
| 2139 | COSTS (total_allocno_costs, parent_a_num)->mem_cost |
| 2140 | = INT_MAX; |
| 2141 | else |
| 2142 | COSTS (total_allocno_costs, parent_a_num)->mem_cost |
| 2143 | += add_cost; |
| 2144 | |
| 2145 | if (i >= first_moveable_pseudo && i < last_moveable_pseudo) |
| 2146 | COSTS (total_allocno_costs, parent_a_num)->mem_cost = 0; |
| 2147 | } |
| 2148 | a_costs = COSTS (costs, a_num)->cost; |
| 2149 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 2150 | { |
| 2151 | add_cost = a_costs[k]; |
| 2152 | if (add_cost > 0 && INT_MAX - add_cost < i_costs[k]) |
| 2153 | i_costs[k] = INT_MAX; |
| 2154 | else |
| 2155 | i_costs[k] += add_cost; |
| 2156 | } |
| 2157 | add_cost = COSTS (costs, a_num)->mem_cost; |
| 2158 | if (add_cost > 0 && INT_MAX - add_cost < i_mem_cost) |
| 2159 | i_mem_cost = INT_MAX; |
| 2160 | else |
| 2161 | i_mem_cost += add_cost; |
| 2162 | } |
| 2163 | } |
| 2164 | if (i >= first_moveable_pseudo && i < last_moveable_pseudo) |
| 2165 | i_mem_cost = 0; |
| 2166 | else if (ira_use_lra_p) |
| 2167 | { |
| 2168 | if (equiv_savings > 0) |
| 2169 | { |
| 2170 | i_mem_cost = 0; |
| 2171 | if (ira_dump_file != NULL && internal_flag_ira_verbose > 5) |
| 2172 | fprintf (stream: ira_dump_file, |
| 2173 | format: " Use MEM for r%d as the equiv savings is %d\n" , |
| 2174 | i, equiv_savings); |
| 2175 | } |
| 2176 | } |
| 2177 | else if (equiv_savings < 0) |
| 2178 | i_mem_cost = -equiv_savings; |
| 2179 | else if (equiv_savings > 0) |
| 2180 | { |
| 2181 | i_mem_cost = 0; |
| 2182 | for (k = cost_classes_ptr->num - 1; k >= 0; k--) |
| 2183 | i_costs[k] += equiv_savings; |
| 2184 | } |
| 2185 | |
| 2186 | best_cost = (1 << (HOST_BITS_PER_INT - 2)) - 1; |
| 2187 | best = ALL_REGS; |
| 2188 | alt_class = NO_REGS; |
| 2189 | /* Find best common class for all allocnos with the same |
| 2190 | regno. */ |
| 2191 | for (k = 0; k < cost_classes_ptr->num; k++) |
| 2192 | { |
| 2193 | rclass = cost_classes[k]; |
| 2194 | if (i_costs[k] < best_cost) |
| 2195 | { |
| 2196 | best_cost = i_costs[k]; |
| 2197 | best = (enum reg_class) rclass; |
| 2198 | } |
| 2199 | else if (i_costs[k] == best_cost) |
| 2200 | best = ira_reg_class_subunion[best][rclass]; |
| 2201 | if (pass == flag_expensive_optimizations |
| 2202 | /* We still prefer registers to memory even at this |
| 2203 | stage if their costs are the same. We will make |
| 2204 | a final decision during assigning hard registers |
| 2205 | when we have all info including more accurate |
| 2206 | costs which might be affected by assigning hard |
| 2207 | registers to other pseudos because the pseudos |
| 2208 | involved in moves can be coalesced. */ |
| 2209 | && i_costs[k] <= i_mem_cost |
| 2210 | && (reg_class_size[reg_class_subunion[alt_class][rclass]] |
| 2211 | > reg_class_size[alt_class])) |
| 2212 | alt_class = reg_class_subunion[alt_class][rclass]; |
| 2213 | } |
| 2214 | alt_class = ira_allocno_class_translate[alt_class]; |
| 2215 | if (best_cost > i_mem_cost |
| 2216 | && ! non_spilled_static_chain_regno_p (regno: i)) |
| 2217 | regno_aclass[i] = NO_REGS; |
| 2218 | else if (!optimize && !targetm.class_likely_spilled_p (best)) |
| 2219 | /* Registers in the alternative class are likely to need |
| 2220 | longer or slower sequences than registers in the best class. |
| 2221 | When optimizing we make some effort to use the best class |
| 2222 | over the alternative class where possible, but at -O0 we |
| 2223 | effectively give the alternative class equal weight. |
| 2224 | We then run the risk of using slower alternative registers |
| 2225 | when plenty of registers from the best class are still free. |
| 2226 | This is especially true because live ranges tend to be very |
| 2227 | short in -O0 code and so register pressure tends to be low. |
| 2228 | |
| 2229 | Avoid that by ignoring the alternative class if the best |
| 2230 | class has plenty of registers. |
| 2231 | |
| 2232 | The union class arrays give important classes and only |
| 2233 | part of it are allocno classes. So translate them into |
| 2234 | allocno classes. */ |
| 2235 | regno_aclass[i] = ira_allocno_class_translate[best]; |
| 2236 | else |
| 2237 | { |
| 2238 | /* Make the common class the biggest class of best and |
| 2239 | alt_class. Translate the common class into an |
| 2240 | allocno class too. */ |
| 2241 | regno_aclass[i] = (ira_allocno_class_translate |
| 2242 | [ira_reg_class_superunion[best][alt_class]]); |
| 2243 | ira_assert (regno_aclass[i] != NO_REGS |
| 2244 | && ira_reg_allocno_class_p[regno_aclass[i]]); |
| 2245 | } |
| 2246 | if (pic_offset_table_rtx != NULL |
| 2247 | && i == (int) REGNO (pic_offset_table_rtx)) |
| 2248 | { |
| 2249 | /* For some targets, integer pseudos can be assigned to fp |
| 2250 | regs. As we don't want reload pic offset table pseudo, we |
| 2251 | should avoid using non-integer regs. */ |
| 2252 | regno_aclass[i] |
| 2253 | = ira_reg_class_intersect[regno_aclass[i]][GENERAL_REGS]; |
| 2254 | alt_class = ira_reg_class_intersect[alt_class][GENERAL_REGS]; |
| 2255 | } |
| 2256 | if ((new_class |
| 2257 | = (reg_class) (targetm.ira_change_pseudo_allocno_class |
| 2258 | (i, regno_aclass[i], best))) != regno_aclass[i]) |
| 2259 | { |
| 2260 | regno_aclass[i] = new_class; |
| 2261 | if (hard_reg_set_subset_p (reg_class_contents[new_class], |
| 2262 | reg_class_contents[best])) |
| 2263 | best = new_class; |
| 2264 | if (hard_reg_set_subset_p (reg_class_contents[new_class], |
| 2265 | reg_class_contents[alt_class])) |
| 2266 | alt_class = new_class; |
| 2267 | } |
| 2268 | if (pass == flag_expensive_optimizations) |
| 2269 | { |
| 2270 | if (best_cost > i_mem_cost |
| 2271 | /* Do not assign NO_REGS to static chain pointer |
| 2272 | pseudo when non-local goto is used. */ |
| 2273 | && ! non_spilled_static_chain_regno_p (regno: i)) |
| 2274 | best = alt_class = NO_REGS; |
| 2275 | else if (best == alt_class) |
| 2276 | alt_class = NO_REGS; |
| 2277 | setup_reg_classes (i, best, alt_class, regno_aclass[i]); |
| 2278 | if ((!allocno_p || internal_flag_ira_verbose > 2) |
| 2279 | && ira_dump_file != NULL) |
| 2280 | fprintf (stream: ira_dump_file, |
| 2281 | format: " r%d: preferred %s, alternative %s, allocno %s\n" , |
| 2282 | i, reg_class_names[best], reg_class_names[alt_class], |
| 2283 | reg_class_names[regno_aclass[i]]); |
| 2284 | } |
| 2285 | regno_best_class[i] = best; |
| 2286 | if (! allocno_p) |
| 2287 | { |
| 2288 | pref[i] = (best_cost > i_mem_cost |
| 2289 | && ! non_spilled_static_chain_regno_p (regno: i) |
| 2290 | ? NO_REGS : best); |
| 2291 | continue; |
| 2292 | } |
| 2293 | for (a = ira_regno_allocno_map[i]; |
| 2294 | a != NULL; |
| 2295 | a = ALLOCNO_NEXT_REGNO_ALLOCNO (a)) |
| 2296 | { |
| 2297 | enum reg_class aclass = regno_aclass[i]; |
| 2298 | int a_num = ALLOCNO_NUM (a); |
| 2299 | int *total_a_costs = COSTS (total_allocno_costs, a_num)->cost; |
| 2300 | int *a_costs = COSTS (costs, a_num)->cost; |
| 2301 | |
| 2302 | if (aclass == NO_REGS) |
| 2303 | best = NO_REGS; |
| 2304 | else |
| 2305 | { |
| 2306 | /* Finding best class which is subset of the common |
| 2307 | class. */ |
| 2308 | best_cost = (1 << (HOST_BITS_PER_INT - 2)) - 1; |
| 2309 | allocno_cost = best_cost; |
| 2310 | best = ALL_REGS; |
| 2311 | for (k = 0; k < cost_classes_ptr->num; k++) |
| 2312 | { |
| 2313 | rclass = cost_classes[k]; |
| 2314 | if (! ira_class_subset_p[rclass][aclass]) |
| 2315 | continue; |
| 2316 | if (total_a_costs[k] < best_cost) |
| 2317 | { |
| 2318 | best_cost = total_a_costs[k]; |
| 2319 | allocno_cost = a_costs[k]; |
| 2320 | best = (enum reg_class) rclass; |
| 2321 | } |
| 2322 | else if (total_a_costs[k] == best_cost) |
| 2323 | { |
| 2324 | best = ira_reg_class_subunion[best][rclass]; |
| 2325 | allocno_cost = MAX (allocno_cost, a_costs[k]); |
| 2326 | } |
| 2327 | } |
| 2328 | ALLOCNO_CLASS_COST (a) = allocno_cost; |
| 2329 | } |
| 2330 | if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL |
| 2331 | && (pass == 0 || pref[a_num] != best)) |
| 2332 | { |
| 2333 | fprintf (stream: ira_dump_file, format: " a%d (r%d," , a_num, i); |
| 2334 | if ((bb = ALLOCNO_LOOP_TREE_NODE (a)->bb) != NULL) |
| 2335 | fprintf (stream: ira_dump_file, format: "b%d" , bb->index); |
| 2336 | else |
| 2337 | fprintf (stream: ira_dump_file, format: "l%d" , |
| 2338 | ALLOCNO_LOOP_TREE_NODE (a)->loop_num); |
| 2339 | fprintf (stream: ira_dump_file, format: ") best %s, allocno %s\n" , |
| 2340 | reg_class_names[best], |
| 2341 | reg_class_names[aclass]); |
| 2342 | } |
| 2343 | pref[a_num] = best; |
| 2344 | if (pass == flag_expensive_optimizations && best != aclass |
| 2345 | && ira_class_hard_regs_num[best] > 0 |
| 2346 | && (ira_reg_class_max_nregs[best][ALLOCNO_MODE (a)] |
| 2347 | >= ira_class_hard_regs_num[best])) |
| 2348 | { |
| 2349 | int ind = cost_classes_ptr->index[aclass]; |
| 2350 | |
| 2351 | ira_assert (ind >= 0); |
| 2352 | ira_init_register_move_cost_if_necessary (ALLOCNO_MODE (a)); |
| 2353 | ira_add_allocno_pref (a, ira_class_hard_regs[best][0], |
| 2354 | (a_costs[ind] - ALLOCNO_CLASS_COST (a)) |
| 2355 | / (ira_register_move_cost |
| 2356 | [ALLOCNO_MODE (a)][best][aclass])); |
| 2357 | for (k = 0; k < cost_classes_ptr->num; k++) |
| 2358 | if (ira_class_subset_p[cost_classes[k]][best]) |
| 2359 | a_costs[k] = a_costs[ind]; |
| 2360 | } |
| 2361 | } |
| 2362 | } |
| 2363 | |
| 2364 | if (internal_flag_ira_verbose > 4 && ira_dump_file) |
| 2365 | { |
| 2366 | if (allocno_p) |
| 2367 | print_allocno_costs (); |
| 2368 | else |
| 2369 | print_pseudo_costs (); |
| 2370 | fprintf (stream: ira_dump_file,format: "\n" ); |
| 2371 | } |
| 2372 | } |
| 2373 | ira_free (addr: regno_best_class); |
| 2374 | } |
| 2375 | |
| 2376 | |
| 2377 | |
| 2378 | /* Process moves involving hard regs to modify allocno hard register |
| 2379 | costs. We can do this only after determining allocno class. If a |
| 2380 | hard register forms a register class, then moves with the hard |
| 2381 | register are already taken into account in class costs for the |
| 2382 | allocno. */ |
| 2383 | static void |
| 2384 | process_bb_node_for_hard_reg_moves (ira_loop_tree_node_t loop_tree_node) |
| 2385 | { |
| 2386 | int i, freq, src_regno, dst_regno, hard_regno, a_regno; |
| 2387 | bool to_p; |
| 2388 | ira_allocno_t a, curr_a; |
| 2389 | ira_loop_tree_node_t curr_loop_tree_node; |
| 2390 | enum reg_class rclass; |
| 2391 | basic_block bb; |
| 2392 | rtx_insn *insn; |
| 2393 | rtx set, src, dst; |
| 2394 | |
| 2395 | bb = loop_tree_node->bb; |
| 2396 | if (bb == NULL) |
| 2397 | return; |
| 2398 | freq = REG_FREQ_FROM_BB (bb); |
| 2399 | if (freq == 0) |
| 2400 | freq = 1; |
| 2401 | FOR_BB_INSNS (bb, insn) |
| 2402 | { |
| 2403 | if (!NONDEBUG_INSN_P (insn)) |
| 2404 | continue; |
| 2405 | set = single_set (insn); |
| 2406 | if (set == NULL_RTX) |
| 2407 | continue; |
| 2408 | dst = SET_DEST (set); |
| 2409 | src = SET_SRC (set); |
| 2410 | if (! REG_P (dst) || ! REG_P (src)) |
| 2411 | continue; |
| 2412 | dst_regno = REGNO (dst); |
| 2413 | src_regno = REGNO (src); |
| 2414 | if (dst_regno >= FIRST_PSEUDO_REGISTER |
| 2415 | && src_regno < FIRST_PSEUDO_REGISTER) |
| 2416 | { |
| 2417 | hard_regno = src_regno; |
| 2418 | a = ira_curr_regno_allocno_map[dst_regno]; |
| 2419 | to_p = true; |
| 2420 | } |
| 2421 | else if (src_regno >= FIRST_PSEUDO_REGISTER |
| 2422 | && dst_regno < FIRST_PSEUDO_REGISTER) |
| 2423 | { |
| 2424 | hard_regno = dst_regno; |
| 2425 | a = ira_curr_regno_allocno_map[src_regno]; |
| 2426 | to_p = false; |
| 2427 | } |
| 2428 | else |
| 2429 | continue; |
| 2430 | if (reg_class_size[(int) REGNO_REG_CLASS (hard_regno)] |
| 2431 | == (ira_reg_class_max_nregs |
| 2432 | [REGNO_REG_CLASS (hard_regno)][(int) ALLOCNO_MODE(a)])) |
| 2433 | /* If the class can provide only one hard reg to the allocno, |
| 2434 | we processed the insn record_operand_costs already and we |
| 2435 | actually updated the hard reg cost there. */ |
| 2436 | continue; |
| 2437 | rclass = ALLOCNO_CLASS (a); |
| 2438 | if (! TEST_HARD_REG_BIT (reg_class_contents[rclass], bit: hard_regno)) |
| 2439 | continue; |
| 2440 | i = ira_class_hard_reg_index[rclass][hard_regno]; |
| 2441 | if (i < 0) |
| 2442 | continue; |
| 2443 | a_regno = ALLOCNO_REGNO (a); |
| 2444 | for (curr_loop_tree_node = ALLOCNO_LOOP_TREE_NODE (a); |
| 2445 | curr_loop_tree_node != NULL; |
| 2446 | curr_loop_tree_node = curr_loop_tree_node->parent) |
| 2447 | if ((curr_a = curr_loop_tree_node->regno_allocno_map[a_regno]) != NULL) |
| 2448 | ira_add_allocno_pref (curr_a, hard_regno, freq); |
| 2449 | { |
| 2450 | int cost; |
| 2451 | enum reg_class hard_reg_class; |
| 2452 | machine_mode mode; |
| 2453 | |
| 2454 | mode = ALLOCNO_MODE (a); |
| 2455 | hard_reg_class = REGNO_REG_CLASS (hard_regno); |
| 2456 | ira_init_register_move_cost_if_necessary (mode); |
| 2457 | cost = (to_p ? ira_register_move_cost[mode][hard_reg_class][rclass] |
| 2458 | : ira_register_move_cost[mode][rclass][hard_reg_class]) * freq; |
| 2459 | ira_allocate_and_set_costs (vec: &ALLOCNO_HARD_REG_COSTS (a), aclass: rclass, |
| 2460 | ALLOCNO_CLASS_COST (a)); |
| 2461 | ira_allocate_and_set_costs (vec: &ALLOCNO_CONFLICT_HARD_REG_COSTS (a), |
| 2462 | aclass: rclass, val: 0); |
| 2463 | ALLOCNO_HARD_REG_COSTS (a)[i] -= cost; |
| 2464 | ALLOCNO_CONFLICT_HARD_REG_COSTS (a)[i] -= cost; |
| 2465 | ALLOCNO_CLASS_COST (a) = MIN (ALLOCNO_CLASS_COST (a), |
| 2466 | ALLOCNO_HARD_REG_COSTS (a)[i]); |
| 2467 | } |
| 2468 | } |
| 2469 | } |
| 2470 | |
| 2471 | /* After we find hard register and memory costs for allocnos, define |
| 2472 | its class and modify hard register cost because insns moving |
| 2473 | allocno to/from hard registers. */ |
| 2474 | static void |
| 2475 | setup_allocno_class_and_costs (void) |
| 2476 | { |
| 2477 | int i, j, n, regno, hard_regno, num; |
| 2478 | int *reg_costs; |
| 2479 | enum reg_class aclass, rclass; |
| 2480 | ira_allocno_t a; |
| 2481 | ira_allocno_iterator ai; |
| 2482 | cost_classes_t cost_classes_ptr; |
| 2483 | |
| 2484 | ira_assert (allocno_p); |
| 2485 | FOR_EACH_ALLOCNO (a, ai) |
| 2486 | { |
| 2487 | i = ALLOCNO_NUM (a); |
| 2488 | regno = ALLOCNO_REGNO (a); |
| 2489 | aclass = regno_aclass[regno]; |
| 2490 | cost_classes_ptr = regno_cost_classes[regno]; |
| 2491 | ira_assert (pref[i] == NO_REGS || aclass != NO_REGS); |
| 2492 | ALLOCNO_MEMORY_COST (a) = COSTS (costs, i)->mem_cost; |
| 2493 | ira_set_allocno_class (a, aclass); |
| 2494 | if (aclass == NO_REGS) |
| 2495 | continue; |
| 2496 | if (optimize && ALLOCNO_CLASS (a) != pref[i]) |
| 2497 | { |
| 2498 | n = ira_class_hard_regs_num[aclass]; |
| 2499 | ALLOCNO_HARD_REG_COSTS (a) |
| 2500 | = reg_costs = ira_allocate_cost_vector (aclass); |
| 2501 | for (j = n - 1; j >= 0; j--) |
| 2502 | { |
| 2503 | hard_regno = ira_class_hard_regs[aclass][j]; |
| 2504 | if (TEST_HARD_REG_BIT (reg_class_contents[pref[i]], bit: hard_regno)) |
| 2505 | reg_costs[j] = ALLOCNO_CLASS_COST (a); |
| 2506 | else |
| 2507 | { |
| 2508 | rclass = REGNO_REG_CLASS (hard_regno); |
| 2509 | num = cost_classes_ptr->index[rclass]; |
| 2510 | if (num < 0) |
| 2511 | { |
| 2512 | num = cost_classes_ptr->hard_regno_index[hard_regno]; |
| 2513 | ira_assert (num >= 0); |
| 2514 | } |
| 2515 | reg_costs[j] = COSTS (costs, i)->cost[num]; |
| 2516 | } |
| 2517 | } |
| 2518 | } |
| 2519 | } |
| 2520 | if (optimize) |
| 2521 | ira_traverse_loop_tree (true, ira_loop_tree_root, |
| 2522 | process_bb_node_for_hard_reg_moves, NULL); |
| 2523 | } |
| 2524 | |
| 2525 | |
| 2526 | |
| 2527 | /* Function called once during compiler work. */ |
| 2528 | void |
| 2529 | ira_init_costs_once (void) |
| 2530 | { |
| 2531 | int i; |
| 2532 | |
| 2533 | init_cost = NULL; |
| 2534 | for (i = 0; i < MAX_RECOG_OPERANDS; i++) |
| 2535 | { |
| 2536 | op_costs[i] = NULL; |
| 2537 | this_op_costs[i] = NULL; |
| 2538 | } |
| 2539 | temp_costs = NULL; |
| 2540 | } |
| 2541 | |
| 2542 | /* Free allocated temporary cost vectors. */ |
| 2543 | void |
| 2544 | target_ira_int::free_ira_costs () |
| 2545 | { |
| 2546 | int i; |
| 2547 | |
| 2548 | free (ptr: x_init_cost); |
| 2549 | x_init_cost = NULL; |
| 2550 | for (i = 0; i < MAX_RECOG_OPERANDS; i++) |
| 2551 | { |
| 2552 | free (ptr: x_op_costs[i]); |
| 2553 | free (ptr: x_this_op_costs[i]); |
| 2554 | x_op_costs[i] = x_this_op_costs[i] = NULL; |
| 2555 | } |
| 2556 | free (ptr: x_temp_costs); |
| 2557 | x_temp_costs = NULL; |
| 2558 | } |
| 2559 | |
| 2560 | /* This is called each time register related information is |
| 2561 | changed. */ |
| 2562 | void |
| 2563 | ira_init_costs (void) |
| 2564 | { |
| 2565 | int i; |
| 2566 | |
| 2567 | this_target_ira_int->free_ira_costs (); |
| 2568 | max_struct_costs_size |
| 2569 | = sizeof (struct costs) + sizeof (int) * (ira_important_classes_num - 1); |
| 2570 | /* Don't use ira_allocate because vectors live through several IRA |
| 2571 | calls. */ |
| 2572 | init_cost = (struct costs *) xmalloc (max_struct_costs_size); |
| 2573 | init_cost->mem_cost = 1000000; |
| 2574 | for (i = 0; i < ira_important_classes_num; i++) |
| 2575 | init_cost->cost[i] = 1000000; |
| 2576 | for (i = 0; i < MAX_RECOG_OPERANDS; i++) |
| 2577 | { |
| 2578 | op_costs[i] = (struct costs *) xmalloc (max_struct_costs_size); |
| 2579 | this_op_costs[i] = (struct costs *) xmalloc (max_struct_costs_size); |
| 2580 | } |
| 2581 | temp_costs = (struct costs *) xmalloc (max_struct_costs_size); |
| 2582 | } |
| 2583 | |
| 2584 | |
| 2585 | |
| 2586 | /* Common initialization function for ira_costs and |
| 2587 | ira_set_pseudo_classes. */ |
| 2588 | static void |
| 2589 | init_costs (void) |
| 2590 | { |
| 2591 | init_subregs_of_mode (); |
| 2592 | costs = (struct costs *) ira_allocate (max_struct_costs_size |
| 2593 | * cost_elements_num); |
| 2594 | pref_buffer = (enum reg_class *) ira_allocate (sizeof (enum reg_class) |
| 2595 | * cost_elements_num); |
| 2596 | regno_aclass = (enum reg_class *) ira_allocate (sizeof (enum reg_class) |
| 2597 | * max_reg_num ()); |
| 2598 | regno_equiv_gains = (int *) ira_allocate (sizeof (int) * max_reg_num ()); |
| 2599 | memset (s: regno_equiv_gains, c: 0, n: sizeof (int) * max_reg_num ()); |
| 2600 | } |
| 2601 | |
| 2602 | /* Common finalization function for ira_costs and |
| 2603 | ira_set_pseudo_classes. */ |
| 2604 | static void |
| 2605 | finish_costs (void) |
| 2606 | { |
| 2607 | finish_subregs_of_mode (); |
| 2608 | ira_free (addr: regno_equiv_gains); |
| 2609 | ira_free (addr: regno_aclass); |
| 2610 | ira_free (addr: pref_buffer); |
| 2611 | ira_free (addr: costs); |
| 2612 | } |
| 2613 | |
| 2614 | /* Entry function which defines register class, memory and hard |
| 2615 | register costs for each allocno. */ |
| 2616 | void |
| 2617 | ira_costs (void) |
| 2618 | { |
| 2619 | allocno_p = true; |
| 2620 | cost_elements_num = ira_allocnos_num; |
| 2621 | init_costs (); |
| 2622 | total_allocno_costs = (struct costs *) ira_allocate (max_struct_costs_size |
| 2623 | * ira_allocnos_num); |
| 2624 | initiate_regno_cost_classes (); |
| 2625 | if (!ira_use_lra_p) |
| 2626 | /* Process equivs in reload to update costs through hook |
| 2627 | ira_adjust_equiv_reg_cost. */ |
| 2628 | calculate_elim_costs_all_insns (); |
| 2629 | find_costs_and_classes (); |
| 2630 | setup_allocno_class_and_costs (); |
| 2631 | finish_regno_cost_classes (); |
| 2632 | finish_costs (); |
| 2633 | ira_free (addr: total_allocno_costs); |
| 2634 | } |
| 2635 | |
| 2636 | /* Entry function which defines classes for pseudos. |
| 2637 | Set pseudo_classes_defined_p only if DEFINE_PSEUDO_CLASSES is true. */ |
| 2638 | void |
| 2639 | ira_set_pseudo_classes (bool define_pseudo_classes, FILE *dump_file) |
| 2640 | { |
| 2641 | FILE *saved_file = ira_dump_file; |
| 2642 | allocno_p = false; |
| 2643 | internal_flag_ira_verbose = flag_ira_verbose; |
| 2644 | ira_dump_file = dump_file; |
| 2645 | cost_elements_num = max_reg_num (); |
| 2646 | init_costs (); |
| 2647 | initiate_regno_cost_classes (); |
| 2648 | find_costs_and_classes (); |
| 2649 | finish_regno_cost_classes (); |
| 2650 | if (define_pseudo_classes) |
| 2651 | pseudo_classes_defined_p = true; |
| 2652 | |
| 2653 | finish_costs (); |
| 2654 | ira_dump_file = saved_file; |
| 2655 | } |
| 2656 | |
| 2657 | |
| 2658 | |
| 2659 | /* Change hard register costs for allocnos which lives through |
| 2660 | function calls. This is called only when we found all intersected |
| 2661 | calls during building allocno live ranges. */ |
| 2662 | void |
| 2663 | ira_tune_allocno_costs (void) |
| 2664 | { |
| 2665 | int j, n, regno; |
| 2666 | int cost, min_cost, *reg_costs; |
| 2667 | enum reg_class aclass; |
| 2668 | machine_mode mode; |
| 2669 | ira_allocno_t a; |
| 2670 | ira_allocno_iterator ai; |
| 2671 | ira_allocno_object_iterator oi; |
| 2672 | ira_object_t obj; |
| 2673 | bool skip_p; |
| 2674 | |
| 2675 | FOR_EACH_ALLOCNO (a, ai) |
| 2676 | { |
| 2677 | aclass = ALLOCNO_CLASS (a); |
| 2678 | if (aclass == NO_REGS) |
| 2679 | continue; |
| 2680 | mode = ALLOCNO_MODE (a); |
| 2681 | n = ira_class_hard_regs_num[aclass]; |
| 2682 | min_cost = INT_MAX; |
| 2683 | if (ALLOCNO_CALLS_CROSSED_NUM (a) |
| 2684 | != ALLOCNO_CHEAP_CALLS_CROSSED_NUM (a)) |
| 2685 | { |
| 2686 | ira_allocate_and_set_costs |
| 2687 | (vec: &ALLOCNO_HARD_REG_COSTS (a), aclass, |
| 2688 | ALLOCNO_CLASS_COST (a)); |
| 2689 | reg_costs = ALLOCNO_HARD_REG_COSTS (a); |
| 2690 | for (j = n - 1; j >= 0; j--) |
| 2691 | { |
| 2692 | regno = ira_class_hard_regs[aclass][j]; |
| 2693 | skip_p = false; |
| 2694 | FOR_EACH_ALLOCNO_OBJECT (a, obj, oi) |
| 2695 | { |
| 2696 | if (ira_hard_reg_set_intersection_p (hard_regno: regno, mode, |
| 2697 | OBJECT_CONFLICT_HARD_REGS |
| 2698 | (obj))) |
| 2699 | { |
| 2700 | skip_p = true; |
| 2701 | break; |
| 2702 | } |
| 2703 | } |
| 2704 | if (skip_p) |
| 2705 | continue; |
| 2706 | cost = 0; |
| 2707 | if (ira_need_caller_save_p (a, regno)) |
| 2708 | cost += ira_caller_save_cost (a); |
| 2709 | #ifdef IRA_HARD_REGNO_ADD_COST_MULTIPLIER |
| 2710 | { |
| 2711 | auto rclass = REGNO_REG_CLASS (regno); |
| 2712 | cost += ((ira_memory_move_cost[mode][rclass][0] |
| 2713 | + ira_memory_move_cost[mode][rclass][1]) |
| 2714 | * ALLOCNO_FREQ (a) |
| 2715 | * IRA_HARD_REGNO_ADD_COST_MULTIPLIER (regno) / 2); |
| 2716 | } |
| 2717 | #endif |
| 2718 | if (INT_MAX - cost < reg_costs[j]) |
| 2719 | reg_costs[j] = INT_MAX; |
| 2720 | else |
| 2721 | reg_costs[j] += cost; |
| 2722 | if (min_cost > reg_costs[j]) |
| 2723 | min_cost = reg_costs[j]; |
| 2724 | } |
| 2725 | } |
| 2726 | if (min_cost != INT_MAX) |
| 2727 | ALLOCNO_CLASS_COST (a) = min_cost; |
| 2728 | |
| 2729 | /* Some targets allow pseudos to be allocated to unaligned sequences |
| 2730 | of hard registers. However, selecting an unaligned sequence can |
| 2731 | unnecessarily restrict later allocations. So increase the cost of |
| 2732 | unaligned hard regs to encourage the use of aligned hard regs. */ |
| 2733 | { |
| 2734 | const int nregs = ira_reg_class_max_nregs[aclass][ALLOCNO_MODE (a)]; |
| 2735 | |
| 2736 | if (nregs > 1) |
| 2737 | { |
| 2738 | ira_allocate_and_set_costs |
| 2739 | (vec: &ALLOCNO_HARD_REG_COSTS (a), aclass, ALLOCNO_CLASS_COST (a)); |
| 2740 | reg_costs = ALLOCNO_HARD_REG_COSTS (a); |
| 2741 | for (j = n - 1; j >= 0; j--) |
| 2742 | { |
| 2743 | regno = ira_non_ordered_class_hard_regs[aclass][j]; |
| 2744 | if ((regno % nregs) != 0) |
| 2745 | { |
| 2746 | int index = ira_class_hard_reg_index[aclass][regno]; |
| 2747 | ira_assert (index != -1); |
| 2748 | reg_costs[index] += ALLOCNO_FREQ (a); |
| 2749 | } |
| 2750 | } |
| 2751 | } |
| 2752 | } |
| 2753 | } |
| 2754 | } |
| 2755 | |
| 2756 | /* A hook from the reload pass. Add COST to the estimated gain for eliminating |
| 2757 | REGNO with its equivalence. If COST is zero, record that no such |
| 2758 | elimination is possible. */ |
| 2759 | |
| 2760 | void |
| 2761 | ira_adjust_equiv_reg_cost (unsigned regno, int cost) |
| 2762 | { |
| 2763 | ira_assert (!ira_use_lra_p); |
| 2764 | if (cost == 0) |
| 2765 | regno_equiv_gains[regno] = 0; |
| 2766 | else |
| 2767 | regno_equiv_gains[regno] += cost; |
| 2768 | } |
| 2769 | |
| 2770 | void |
| 2771 | ira_costs_cc_finalize (void) |
| 2772 | { |
| 2773 | this_target_ira_int->free_ira_costs (); |
| 2774 | } |
| 2775 | |