| 1 | /* LRA (local register allocator) driver and LRA utilities. |
| 2 | Copyright (C) 2010-2025 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 | |
| 22 | /* The Local Register Allocator (LRA) is a replacement of former |
| 23 | reload pass. It is focused to simplify code solving the reload |
| 24 | pass tasks, to make the code maintenance easier, and to implement new |
| 25 | perspective optimizations. |
| 26 | |
| 27 | The major LRA design solutions are: |
| 28 | o division small manageable, separated sub-tasks |
| 29 | o reflection of all transformations and decisions in RTL as more |
| 30 | as possible |
| 31 | o insn constraints as a primary source of the info (minimizing |
| 32 | number of target-depended macros/hooks) |
| 33 | |
| 34 | In brief LRA works by iterative insn process with the final goal is |
| 35 | to satisfy all insn and address constraints: |
| 36 | o New reload insns (in brief reloads) and reload pseudos might be |
| 37 | generated; |
| 38 | o Some pseudos might be spilled to assign hard registers to |
| 39 | new reload pseudos; |
| 40 | o Recalculating spilled pseudo values (rematerialization); |
| 41 | o Changing spilled pseudos to stack memory or their equivalences; |
| 42 | o Allocation stack memory changes the address displacement and |
| 43 | new iteration is needed. |
| 44 | |
| 45 | Here is block diagram of LRA passes: |
| 46 | |
| 47 | ------------------------ |
| 48 | --------------- | Undo inheritance for | --------------- |
| 49 | | Memory-memory | | spilled pseudos, | | New (and old) | |
| 50 | | move coalesce |<---| splits for pseudos got |<-- | pseudos | |
| 51 | --------------- | the same hard regs, | | assignment | |
| 52 | Start | | and optional reloads | --------------- |
| 53 | | | ------------------------ ^ |
| 54 | V | ---------------- | |
| 55 | ----------- V | Update virtual | | |
| 56 | | Remove |----> ------------>| register | | |
| 57 | | scratches | ^ | displacements | | |
| 58 | ----------- | ---------------- | |
| 59 | | | | |
| 60 | | V New | |
| 61 | | ------------ pseudos ------------------- |
| 62 | | |Constraints:| or insns | Inheritance/split | |
| 63 | | | RTL |--------->| transformations | |
| 64 | | | transfor- | | in EBB scope | |
| 65 | | substi- | mations | ------------------- |
| 66 | | tutions ------------ |
| 67 | | | No change |
| 68 | ---------------- V |
| 69 | | Spilled pseudo | ------------------- |
| 70 | | to memory |<----| Rematerialization | |
| 71 | | substitution | ------------------- |
| 72 | ---------------- |
| 73 | | No susbtitions |
| 74 | V |
| 75 | ------------------------- |
| 76 | | Hard regs substitution, | |
| 77 | | devirtalization, and |------> Finish |
| 78 | | restoring scratches got | |
| 79 | | memory | |
| 80 | ------------------------- |
| 81 | |
| 82 | To speed up the process: |
| 83 | o We process only insns affected by changes on previous |
| 84 | iterations; |
| 85 | o We don't use DFA-infrastructure because it results in much slower |
| 86 | compiler speed than a special IR described below does; |
| 87 | o We use a special insn representation for quick access to insn |
| 88 | info which is always *synchronized* with the current RTL; |
| 89 | o Insn IR is minimized by memory. It is divided on three parts: |
| 90 | o one specific for each insn in RTL (only operand locations); |
| 91 | o one common for all insns in RTL with the same insn code |
| 92 | (different operand attributes from machine descriptions); |
| 93 | o one oriented for maintenance of live info (list of pseudos). |
| 94 | o Pseudo data: |
| 95 | o all insns where the pseudo is referenced; |
| 96 | o live info (conflicting hard regs, live ranges, # of |
| 97 | references etc); |
| 98 | o data used for assigning (preferred hard regs, costs etc). |
| 99 | |
| 100 | This file contains LRA driver, LRA utility functions and data, and |
| 101 | code for dealing with scratches. */ |
| 102 | |
| 103 | #include "config.h" |
| 104 | #include "system.h" |
| 105 | #include "coretypes.h" |
| 106 | #include "backend.h" |
| 107 | #include "target.h" |
| 108 | #include "rtl.h" |
| 109 | #include "rtl-error.h" |
| 110 | #include "tree.h" |
| 111 | #include "predict.h" |
| 112 | #include "df.h" |
| 113 | #include "memmodel.h" |
| 114 | #include "tm_p.h" |
| 115 | #include "optabs.h" |
| 116 | #include "regs.h" |
| 117 | #include "ira.h" |
| 118 | #include "recog.h" |
| 119 | #include "expr.h" |
| 120 | #include "cfgrtl.h" |
| 121 | #include "cfgbuild.h" |
| 122 | #include "lra.h" |
| 123 | #include "lra-int.h" |
| 124 | #include "print-rtl.h" |
| 125 | #include "function-abi.h" |
| 126 | |
| 127 | /* Dump bitmap SET with TITLE and BB INDEX. */ |
| 128 | void |
| 129 | lra_dump_bitmap_with_title (const char *title, bitmap set, int index) |
| 130 | { |
| 131 | unsigned int i; |
| 132 | int count; |
| 133 | bitmap_iterator bi; |
| 134 | static const int max_nums_on_line = 10; |
| 135 | |
| 136 | if (bitmap_empty_p (map: set)) |
| 137 | return; |
| 138 | fprintf (stream: lra_dump_file, format: " %s %d:" , title, index); |
| 139 | fprintf (stream: lra_dump_file, format: "\n" ); |
| 140 | count = max_nums_on_line + 1; |
| 141 | EXECUTE_IF_SET_IN_BITMAP (set, 0, i, bi) |
| 142 | { |
| 143 | if (count > max_nums_on_line) |
| 144 | { |
| 145 | fprintf (stream: lra_dump_file, format: "\n " ); |
| 146 | count = 0; |
| 147 | } |
| 148 | fprintf (stream: lra_dump_file, format: " %4u" , i); |
| 149 | count++; |
| 150 | } |
| 151 | fprintf (stream: lra_dump_file, format: "\n" ); |
| 152 | } |
| 153 | |
| 154 | /* Hard registers currently not available for allocation. It can |
| 155 | changed after some hard registers become not eliminable. */ |
| 156 | HARD_REG_SET lra_no_alloc_regs; |
| 157 | |
| 158 | static int get_new_reg_value (void); |
| 159 | static void expand_reg_info (void); |
| 160 | static void invalidate_insn_recog_data (int); |
| 161 | static int get_insn_freq (rtx_insn *); |
| 162 | static void invalidate_insn_data_regno_info (lra_insn_recog_data_t, |
| 163 | rtx_insn *, int); |
| 164 | /* Expand all regno related info needed for LRA. */ |
| 165 | static void |
| 166 | expand_reg_data (int old) |
| 167 | { |
| 168 | resize_reg_info (); |
| 169 | expand_reg_info (); |
| 170 | ira_expand_reg_equiv (); |
| 171 | for (int i = (int) max_reg_num () - 1; i >= old; i--) |
| 172 | lra_change_class (regno: i, new_class: ALL_REGS, title: " Set" , nl_p: true); |
| 173 | } |
| 174 | |
| 175 | /* Create and return a new reg of ORIGINAL mode. If ORIGINAL is NULL |
| 176 | or of VOIDmode, use MD_MODE for the new reg. Initialize its |
| 177 | register class to RCLASS. Print message about assigning class |
| 178 | RCLASS containing new register name TITLE unless it is NULL. Use |
| 179 | attributes of ORIGINAL if it is a register. The created register |
| 180 | will have unique held value. */ |
| 181 | rtx |
| 182 | lra_create_new_reg_with_unique_value (machine_mode md_mode, rtx original, |
| 183 | enum reg_class rclass, |
| 184 | HARD_REG_SET *exclude_start_hard_regs, |
| 185 | const char *title) |
| 186 | { |
| 187 | machine_mode mode; |
| 188 | rtx new_reg; |
| 189 | |
| 190 | if (original == NULL_RTX || (mode = GET_MODE (original)) == VOIDmode) |
| 191 | mode = md_mode; |
| 192 | lra_assert (mode != VOIDmode); |
| 193 | new_reg = gen_reg_rtx (mode); |
| 194 | if (original == NULL_RTX || ! REG_P (original)) |
| 195 | { |
| 196 | if (lra_dump_file != NULL) |
| 197 | fprintf (stream: lra_dump_file, format: " Creating newreg=%i" , REGNO (new_reg)); |
| 198 | } |
| 199 | else |
| 200 | { |
| 201 | if (ORIGINAL_REGNO (original) >= FIRST_PSEUDO_REGISTER) |
| 202 | ORIGINAL_REGNO (new_reg) = ORIGINAL_REGNO (original); |
| 203 | REG_USERVAR_P (new_reg) = REG_USERVAR_P (original); |
| 204 | REG_POINTER (new_reg) = REG_POINTER (original); |
| 205 | REG_ATTRS (new_reg) = REG_ATTRS (original); |
| 206 | if (lra_dump_file != NULL) |
| 207 | fprintf (stream: lra_dump_file, format: " Creating newreg=%i from oldreg=%i" , |
| 208 | REGNO (new_reg), REGNO (original)); |
| 209 | } |
| 210 | if (lra_dump_file != NULL) |
| 211 | { |
| 212 | if (title != NULL) |
| 213 | fprintf (stream: lra_dump_file, format: ", assigning class %s to%s%s r%d" , |
| 214 | reg_class_names[rclass], *title == '\0' ? "" : " " , |
| 215 | title, REGNO (new_reg)); |
| 216 | fprintf (stream: lra_dump_file, format: "\n" ); |
| 217 | } |
| 218 | expand_reg_data (old: max_reg_num ()); |
| 219 | setup_reg_classes (REGNO (new_reg), rclass, NO_REGS, rclass); |
| 220 | if (exclude_start_hard_regs != NULL) |
| 221 | lra_reg_info[REGNO (new_reg)].exclude_start_hard_regs |
| 222 | = *exclude_start_hard_regs; |
| 223 | return new_reg; |
| 224 | } |
| 225 | |
| 226 | /* Analogous to the previous function but also inherits value of |
| 227 | ORIGINAL. */ |
| 228 | rtx |
| 229 | lra_create_new_reg (machine_mode md_mode, rtx original, enum reg_class rclass, |
| 230 | HARD_REG_SET *exclude_start_hard_regs, const char *title) |
| 231 | { |
| 232 | rtx new_reg; |
| 233 | |
| 234 | new_reg |
| 235 | = lra_create_new_reg_with_unique_value (md_mode, original, rclass, |
| 236 | exclude_start_hard_regs, title); |
| 237 | if (original != NULL_RTX && REG_P (original)) |
| 238 | lra_assign_reg_val (REGNO (original), REGNO (new_reg)); |
| 239 | return new_reg; |
| 240 | } |
| 241 | |
| 242 | /* Set up for REGNO unique hold value. */ |
| 243 | void |
| 244 | lra_set_regno_unique_value (int regno) |
| 245 | { |
| 246 | lra_reg_info[regno].val = get_new_reg_value (); |
| 247 | } |
| 248 | |
| 249 | /* Invalidate INSN related info used by LRA. The info should never be |
| 250 | used after that. */ |
| 251 | void |
| 252 | lra_invalidate_insn_data (rtx_insn *insn) |
| 253 | { |
| 254 | lra_invalidate_insn_regno_info (insn); |
| 255 | invalidate_insn_recog_data (INSN_UID (insn)); |
| 256 | } |
| 257 | |
| 258 | /* Mark INSN deleted and invalidate the insn related info used by |
| 259 | LRA. */ |
| 260 | void |
| 261 | lra_set_insn_deleted (rtx_insn *insn) |
| 262 | { |
| 263 | lra_invalidate_insn_data (insn); |
| 264 | SET_INSN_DELETED (insn); |
| 265 | } |
| 266 | |
| 267 | /* Delete an unneeded INSN and any previous insns who sole purpose is |
| 268 | loading data that is dead in INSN. */ |
| 269 | void |
| 270 | lra_delete_dead_insn (rtx_insn *insn) |
| 271 | { |
| 272 | rtx_insn *prev = prev_real_insn (insn); |
| 273 | rtx prev_dest; |
| 274 | |
| 275 | /* If the previous insn sets a register that dies in our insn, |
| 276 | delete it too. */ |
| 277 | if (prev && GET_CODE (PATTERN (prev)) == SET |
| 278 | && (prev_dest = SET_DEST (PATTERN (prev)), REG_P (prev_dest)) |
| 279 | && reg_mentioned_p (prev_dest, PATTERN (insn)) |
| 280 | && find_regno_note (insn, REG_DEAD, REGNO (prev_dest)) |
| 281 | && ! side_effects_p (SET_SRC (PATTERN (prev)))) |
| 282 | lra_delete_dead_insn (insn: prev); |
| 283 | |
| 284 | lra_set_insn_deleted (insn); |
| 285 | } |
| 286 | |
| 287 | /* Emit insn x = y + z. Return NULL if we failed to do it. |
| 288 | Otherwise, return the insn. We don't use gen_add3_insn as it might |
| 289 | clobber CC. */ |
| 290 | static rtx_insn * |
| 291 | emit_add3_insn (rtx x, rtx y, rtx z) |
| 292 | { |
| 293 | rtx_insn *last; |
| 294 | |
| 295 | last = get_last_insn (); |
| 296 | |
| 297 | if (have_addptr3_insn (x, y, z)) |
| 298 | { |
| 299 | rtx_insn *insn = gen_addptr3_insn (x, y, z); |
| 300 | |
| 301 | /* If the target provides an "addptr" pattern it hopefully does |
| 302 | for a reason. So falling back to the normal add would be |
| 303 | a bug. */ |
| 304 | lra_assert (insn != NULL_RTX); |
| 305 | emit_insn (insn); |
| 306 | return insn; |
| 307 | } |
| 308 | |
| 309 | rtx_insn *insn = emit_insn (gen_rtx_SET (x, gen_rtx_PLUS (GET_MODE (y), |
| 310 | y, z))); |
| 311 | if (recog_memoized (insn) < 0) |
| 312 | { |
| 313 | delete_insns_since (last); |
| 314 | insn = NULL; |
| 315 | } |
| 316 | return insn; |
| 317 | } |
| 318 | |
| 319 | /* Emit insn x = x + y. Return the insn. We use gen_add2_insn as the |
| 320 | last resort. */ |
| 321 | static rtx_insn * |
| 322 | emit_add2_insn (rtx x, rtx y) |
| 323 | { |
| 324 | rtx_insn *insn = emit_add3_insn (x, y: x, z: y); |
| 325 | if (insn == NULL_RTX) |
| 326 | { |
| 327 | insn = gen_add2_insn (x, y); |
| 328 | if (insn != NULL_RTX) |
| 329 | emit_insn (insn); |
| 330 | } |
| 331 | return insn; |
| 332 | } |
| 333 | |
| 334 | /* Target checks operands through operand predicates to recognize an |
| 335 | insn. We should have a special precaution to generate add insns |
| 336 | which are frequent results of elimination. |
| 337 | |
| 338 | Emit insns for x = y + z. X can be used to store intermediate |
| 339 | values and should be not in Y and Z when we use X to store an |
| 340 | intermediate value. Y + Z should form [base] [+ index[ * scale]] [ |
| 341 | + disp] where base and index are registers, disp and scale are |
| 342 | constants. Y should contain base if it is present, Z should |
| 343 | contain disp if any. index[*scale] can be part of Y or Z. */ |
| 344 | void |
| 345 | lra_emit_add (rtx x, rtx y, rtx z) |
| 346 | { |
| 347 | int old; |
| 348 | rtx_insn *last; |
| 349 | rtx a1, a2, base, index, disp, scale, index_scale; |
| 350 | bool ok_p; |
| 351 | |
| 352 | rtx_insn *add3_insn = emit_add3_insn (x, y, z); |
| 353 | old = max_reg_num (); |
| 354 | if (add3_insn != NULL) |
| 355 | ; |
| 356 | else |
| 357 | { |
| 358 | disp = a2 = NULL_RTX; |
| 359 | if (GET_CODE (y) == PLUS) |
| 360 | { |
| 361 | a1 = XEXP (y, 0); |
| 362 | a2 = XEXP (y, 1); |
| 363 | disp = z; |
| 364 | } |
| 365 | else |
| 366 | { |
| 367 | a1 = y; |
| 368 | if (CONSTANT_P (z)) |
| 369 | disp = z; |
| 370 | else |
| 371 | a2 = z; |
| 372 | } |
| 373 | index_scale = scale = NULL_RTX; |
| 374 | if (GET_CODE (a1) == MULT) |
| 375 | { |
| 376 | index_scale = a1; |
| 377 | index = XEXP (a1, 0); |
| 378 | scale = XEXP (a1, 1); |
| 379 | base = a2; |
| 380 | } |
| 381 | else if (a2 != NULL_RTX && GET_CODE (a2) == MULT) |
| 382 | { |
| 383 | index_scale = a2; |
| 384 | index = XEXP (a2, 0); |
| 385 | scale = XEXP (a2, 1); |
| 386 | base = a1; |
| 387 | } |
| 388 | else |
| 389 | { |
| 390 | base = a1; |
| 391 | index = a2; |
| 392 | } |
| 393 | if ((base != NULL_RTX && ! (REG_P (base) || GET_CODE (base) == SUBREG)) |
| 394 | || (index != NULL_RTX |
| 395 | && ! (REG_P (index) || GET_CODE (index) == SUBREG)) |
| 396 | || (disp != NULL_RTX && ! CONSTANT_P (disp)) |
| 397 | || (scale != NULL_RTX && ! CONSTANT_P (scale))) |
| 398 | { |
| 399 | /* Probably we have no 3 op add. Last chance is to use 2-op |
| 400 | add insn. To succeed, don't move Z to X as an address |
| 401 | segment always comes in Y. Otherwise, we might fail when |
| 402 | adding the address segment to register. */ |
| 403 | lra_assert (x != y && x != z); |
| 404 | emit_move_insn (x, y); |
| 405 | rtx_insn *insn = emit_add2_insn (x, y: z); |
| 406 | lra_assert (insn != NULL_RTX); |
| 407 | } |
| 408 | else |
| 409 | { |
| 410 | if (index_scale == NULL_RTX) |
| 411 | index_scale = index; |
| 412 | if (disp == NULL_RTX) |
| 413 | { |
| 414 | /* Generate x = index_scale; x = x + base. */ |
| 415 | lra_assert (index_scale != NULL_RTX && base != NULL_RTX); |
| 416 | emit_move_insn (x, index_scale); |
| 417 | rtx_insn *insn = emit_add2_insn (x, y: base); |
| 418 | lra_assert (insn != NULL_RTX); |
| 419 | } |
| 420 | else if (scale == NULL_RTX) |
| 421 | { |
| 422 | /* Try x = base + disp. */ |
| 423 | lra_assert (base != NULL_RTX); |
| 424 | last = get_last_insn (); |
| 425 | rtx_insn *move_insn = |
| 426 | emit_move_insn (x, gen_rtx_PLUS (GET_MODE (base), base, disp)); |
| 427 | if (recog_memoized (insn: move_insn) < 0) |
| 428 | { |
| 429 | delete_insns_since (last); |
| 430 | /* Generate x = disp; x = x + base. */ |
| 431 | emit_move_insn (x, disp); |
| 432 | rtx_insn *add2_insn = emit_add2_insn (x, y: base); |
| 433 | lra_assert (add2_insn != NULL_RTX); |
| 434 | } |
| 435 | /* Generate x = x + index. */ |
| 436 | if (index != NULL_RTX) |
| 437 | { |
| 438 | rtx_insn *insn = emit_add2_insn (x, y: index); |
| 439 | lra_assert (insn != NULL_RTX); |
| 440 | } |
| 441 | } |
| 442 | else |
| 443 | { |
| 444 | /* Try x = index_scale; x = x + disp; x = x + base. */ |
| 445 | last = get_last_insn (); |
| 446 | rtx_insn *move_insn = emit_move_insn (x, index_scale); |
| 447 | ok_p = false; |
| 448 | if (recog_memoized (insn: move_insn) >= 0) |
| 449 | { |
| 450 | rtx_insn *insn = emit_add2_insn (x, y: disp); |
| 451 | if (insn != NULL_RTX) |
| 452 | { |
| 453 | if (base == NULL_RTX) |
| 454 | ok_p = true; |
| 455 | else |
| 456 | { |
| 457 | insn = emit_add2_insn (x, y: base); |
| 458 | if (insn != NULL_RTX) |
| 459 | ok_p = true; |
| 460 | } |
| 461 | } |
| 462 | } |
| 463 | if (! ok_p) |
| 464 | { |
| 465 | rtx_insn *insn; |
| 466 | |
| 467 | delete_insns_since (last); |
| 468 | /* Generate x = disp; x = x + base; x = x + index_scale. */ |
| 469 | emit_move_insn (x, disp); |
| 470 | if (base != NULL_RTX) |
| 471 | { |
| 472 | insn = emit_add2_insn (x, y: base); |
| 473 | lra_assert (insn != NULL_RTX); |
| 474 | } |
| 475 | insn = emit_add2_insn (x, y: index_scale); |
| 476 | lra_assert (insn != NULL_RTX); |
| 477 | } |
| 478 | } |
| 479 | } |
| 480 | } |
| 481 | /* Functions emit_... can create pseudos -- so expand the pseudo |
| 482 | data. */ |
| 483 | if (old != max_reg_num ()) |
| 484 | expand_reg_data (old); |
| 485 | } |
| 486 | |
| 487 | /* The number of emitted reload insns so far. */ |
| 488 | int lra_curr_reload_num; |
| 489 | |
| 490 | static void remove_insn_scratches (rtx_insn *insn); |
| 491 | |
| 492 | /* Emit x := y, processing special case when y = u + v or y = u + v * |
| 493 | scale + w through emit_add (Y can be an address which is base + |
| 494 | index reg * scale + displacement in general case). X may be used |
| 495 | as intermediate result therefore it should be not in Y. */ |
| 496 | void |
| 497 | lra_emit_move (rtx x, rtx y) |
| 498 | { |
| 499 | int old; |
| 500 | rtx_insn *insn; |
| 501 | |
| 502 | if (GET_CODE (y) != PLUS) |
| 503 | { |
| 504 | if (rtx_equal_p (x, y)) |
| 505 | return; |
| 506 | old = max_reg_num (); |
| 507 | |
| 508 | insn = (GET_CODE (x) != STRICT_LOW_PART |
| 509 | ? emit_move_insn (x, y) : emit_insn (gen_rtx_SET (x, y))); |
| 510 | /* The move pattern may require scratch registers, so convert them |
| 511 | into real registers now. */ |
| 512 | if (insn != NULL_RTX) |
| 513 | remove_insn_scratches (insn); |
| 514 | if (REG_P (x)) |
| 515 | lra_reg_info[ORIGINAL_REGNO (x)].last_reload = ++lra_curr_reload_num; |
| 516 | /* Function emit_move can create pseudos -- so expand the pseudo |
| 517 | data. */ |
| 518 | if (old != max_reg_num ()) |
| 519 | expand_reg_data (old); |
| 520 | return; |
| 521 | } |
| 522 | lra_emit_add (x, XEXP (y, 0), XEXP (y, 1)); |
| 523 | } |
| 524 | |
| 525 | /* Update insn operands which are duplication of operands whose |
| 526 | numbers are in array of NOPS (with end marker -1). The insn is |
| 527 | represented by its LRA internal representation ID. */ |
| 528 | void |
| 529 | lra_update_dups (lra_insn_recog_data_t id, signed char *nops) |
| 530 | { |
| 531 | int i, j, nop; |
| 532 | struct lra_static_insn_data *static_id = id->insn_static_data; |
| 533 | |
| 534 | for (i = 0; i < static_id->n_dups; i++) |
| 535 | for (j = 0; (nop = nops[j]) >= 0; j++) |
| 536 | if (static_id->dup_num[i] == nop) |
| 537 | *id->dup_loc[i] = *id->operand_loc[nop]; |
| 538 | } |
| 539 | |
| 540 | /* Report asm insn error and modify the asm insn. */ |
| 541 | void |
| 542 | lra_asm_insn_error (rtx_insn *insn) |
| 543 | { |
| 544 | lra_asm_error_p = true; |
| 545 | error_for_asm (insn, |
| 546 | "%<asm%> operand has impossible constraints" |
| 547 | " or there are not enough registers" ); |
| 548 | /* Avoid further trouble with this insn. */ |
| 549 | if (JUMP_P (insn)) |
| 550 | { |
| 551 | ira_nullify_asm_goto (insn); |
| 552 | lra_invalidate_insn_data (insn); |
| 553 | } |
| 554 | else |
| 555 | { |
| 556 | PATTERN (insn) = gen_rtx_USE (VOIDmode, const0_rtx); |
| 557 | lra_set_insn_deleted (insn); |
| 558 | } |
| 559 | } |
| 560 | |
| 561 | |
| 562 | |
| 563 | /* This page contains code dealing with info about registers in the |
| 564 | insns. */ |
| 565 | |
| 566 | /* Pools for insn reg info. */ |
| 567 | object_allocator<lra_insn_reg> lra_insn_reg_pool ("insn regs" ); |
| 568 | |
| 569 | /* Create LRA insn related info about a reference to REGNO in INSN |
| 570 | with TYPE (in/out/inout), biggest reference mode MODE, flag that it |
| 571 | is reference through subreg (SUBREG_P), and reference to the next |
| 572 | insn reg info (NEXT). If REGNO can be early clobbered, |
| 573 | alternatives in which it can be early clobbered are given by |
| 574 | EARLY_CLOBBER_ALTS. */ |
| 575 | static struct lra_insn_reg * |
| 576 | new_insn_reg (rtx_insn *insn, int regno, enum op_type type, |
| 577 | machine_mode mode, bool subreg_p, |
| 578 | alternative_mask early_clobber_alts, |
| 579 | struct lra_insn_reg *next) |
| 580 | { |
| 581 | lra_insn_reg *ir = lra_insn_reg_pool.allocate (); |
| 582 | ir->type = type; |
| 583 | ir->biggest_mode = mode; |
| 584 | if (NONDEBUG_INSN_P (insn)) |
| 585 | lra_update_biggest_mode (regno, mode); |
| 586 | ir->subreg_p = subreg_p; |
| 587 | ir->early_clobber_alts = early_clobber_alts; |
| 588 | ir->regno = regno; |
| 589 | ir->next = next; |
| 590 | return ir; |
| 591 | } |
| 592 | |
| 593 | /* Free insn reg info list IR. */ |
| 594 | static void |
| 595 | free_insn_regs (struct lra_insn_reg *ir) |
| 596 | { |
| 597 | struct lra_insn_reg *next_ir; |
| 598 | |
| 599 | for (; ir != NULL; ir = next_ir) |
| 600 | { |
| 601 | next_ir = ir->next; |
| 602 | lra_insn_reg_pool.remove (object: ir); |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | /* Finish pool for insn reg info. */ |
| 607 | static void |
| 608 | finish_insn_regs (void) |
| 609 | { |
| 610 | lra_insn_reg_pool.release (); |
| 611 | } |
| 612 | |
| 613 | |
| 614 | |
| 615 | /* This page contains code dealing LRA insn info (or in other words |
| 616 | LRA internal insn representation). */ |
| 617 | |
| 618 | /* Map INSN_CODE -> the static insn data. This info is valid during |
| 619 | all translation unit. */ |
| 620 | struct lra_static_insn_data *insn_code_data[NUM_INSN_CODES]; |
| 621 | |
| 622 | /* Debug insns are represented as a special insn with one input |
| 623 | operand which is RTL expression in var_location. */ |
| 624 | |
| 625 | /* The following data are used as static insn operand data for all |
| 626 | debug insns. If structure lra_operand_data is changed, the |
| 627 | initializer should be changed too. */ |
| 628 | static struct lra_operand_data debug_operand_data = |
| 629 | { |
| 630 | NULL, /* alternative */ |
| 631 | .early_clobber_alts: 0, /* early_clobber_alts */ |
| 632 | .mode: E_VOIDmode, /* We are not interesting in the operand mode. */ |
| 633 | .type: OP_IN, |
| 634 | .strict_low: 0, .is_operator: 0, .is_address: 0 |
| 635 | }; |
| 636 | |
| 637 | /* The following data are used as static insn data for all debug |
| 638 | bind insns. If structure lra_static_insn_data is changed, the |
| 639 | initializer should be changed too. */ |
| 640 | static struct lra_static_insn_data debug_bind_static_data = |
| 641 | { |
| 642 | .operand: &debug_operand_data, |
| 643 | .dup_num: 0, /* Duplication operands #. */ |
| 644 | .commutative: -1, /* Commutative operand #. */ |
| 645 | .n_operands: 1, /* Operands #. There is only one operand which is debug RTL |
| 646 | expression. */ |
| 647 | .n_dups: 0, /* Duplications #. */ |
| 648 | .n_alternatives: 0, /* Alternatives #. We are not interesting in alternatives |
| 649 | because we does not proceed debug_insns for reloads. */ |
| 650 | NULL, /* Hard registers referenced in machine description. */ |
| 651 | NULL /* Descriptions of operands in alternatives. */ |
| 652 | }; |
| 653 | |
| 654 | /* The following data are used as static insn data for all debug |
| 655 | marker insns. If structure lra_static_insn_data is changed, the |
| 656 | initializer should be changed too. */ |
| 657 | static struct lra_static_insn_data debug_marker_static_data = |
| 658 | { |
| 659 | .operand: &debug_operand_data, |
| 660 | .dup_num: 0, /* Duplication operands #. */ |
| 661 | .commutative: -1, /* Commutative operand #. */ |
| 662 | .n_operands: 0, /* Operands #. There isn't any operand. */ |
| 663 | .n_dups: 0, /* Duplications #. */ |
| 664 | .n_alternatives: 0, /* Alternatives #. We are not interesting in alternatives |
| 665 | because we does not proceed debug_insns for reloads. */ |
| 666 | NULL, /* Hard registers referenced in machine description. */ |
| 667 | NULL /* Descriptions of operands in alternatives. */ |
| 668 | }; |
| 669 | |
| 670 | /* Called once per compiler work to initialize some LRA data related |
| 671 | to insns. */ |
| 672 | static void |
| 673 | init_insn_code_data_once (void) |
| 674 | { |
| 675 | memset (s: insn_code_data, c: 0, n: sizeof (insn_code_data)); |
| 676 | } |
| 677 | |
| 678 | /* Called once per compiler work to finalize some LRA data related to |
| 679 | insns. */ |
| 680 | static void |
| 681 | finish_insn_code_data_once (void) |
| 682 | { |
| 683 | for (unsigned int i = 0; i < NUM_INSN_CODES; i++) |
| 684 | { |
| 685 | if (insn_code_data[i] != NULL) |
| 686 | { |
| 687 | free (ptr: insn_code_data[i]); |
| 688 | insn_code_data[i] = NULL; |
| 689 | } |
| 690 | } |
| 691 | } |
| 692 | |
| 693 | /* Return static insn data, allocate and setup if necessary. Although |
| 694 | dup_num is static data (it depends only on icode), to set it up we |
| 695 | need to extract insn first. So recog_data should be valid for |
| 696 | normal insn (ICODE >= 0) before the call. */ |
| 697 | static struct lra_static_insn_data * |
| 698 | get_static_insn_data (int icode, int nop, int ndup, int nalt) |
| 699 | { |
| 700 | struct lra_static_insn_data *data; |
| 701 | size_t n_bytes; |
| 702 | |
| 703 | lra_assert (icode < (int) NUM_INSN_CODES); |
| 704 | if (icode >= 0 && (data = insn_code_data[icode]) != NULL) |
| 705 | return data; |
| 706 | lra_assert (nop >= 0 && ndup >= 0 && nalt >= 0); |
| 707 | n_bytes = sizeof (struct lra_static_insn_data) |
| 708 | + sizeof (struct lra_operand_data) * nop |
| 709 | + sizeof (int) * ndup; |
| 710 | data = XNEWVAR (struct lra_static_insn_data, n_bytes); |
| 711 | data->operand_alternative = NULL; |
| 712 | data->n_operands = nop; |
| 713 | data->n_dups = ndup; |
| 714 | data->n_alternatives = nalt; |
| 715 | data->operand = ((struct lra_operand_data *) |
| 716 | ((char *) data + sizeof (struct lra_static_insn_data))); |
| 717 | data->dup_num = ((int *) ((char *) data->operand |
| 718 | + sizeof (struct lra_operand_data) * nop)); |
| 719 | if (icode >= 0) |
| 720 | { |
| 721 | int i; |
| 722 | |
| 723 | insn_code_data[icode] = data; |
| 724 | for (i = 0; i < nop; i++) |
| 725 | { |
| 726 | data->operand[i].constraint |
| 727 | = insn_data[icode].operand[i].constraint; |
| 728 | data->operand[i].mode = insn_data[icode].operand[i].mode; |
| 729 | data->operand[i].strict_low = insn_data[icode].operand[i].strict_low; |
| 730 | data->operand[i].is_operator |
| 731 | = insn_data[icode].operand[i].is_operator; |
| 732 | data->operand[i].type |
| 733 | = (data->operand[i].constraint[0] == '=' ? OP_OUT |
| 734 | : data->operand[i].constraint[0] == '+' ? OP_INOUT |
| 735 | : OP_IN); |
| 736 | data->operand[i].is_address = false; |
| 737 | } |
| 738 | for (i = 0; i < ndup; i++) |
| 739 | data->dup_num[i] = recog_data.dup_num[i]; |
| 740 | } |
| 741 | return data; |
| 742 | } |
| 743 | |
| 744 | /* The current length of the following array. */ |
| 745 | int lra_insn_recog_data_len; |
| 746 | |
| 747 | /* Map INSN_UID -> the insn recog data (NULL if unknown). */ |
| 748 | lra_insn_recog_data_t *lra_insn_recog_data; |
| 749 | |
| 750 | /* Alloc pool we allocate entries for lra_insn_recog_data from. */ |
| 751 | static object_allocator<class lra_insn_recog_data> |
| 752 | lra_insn_recog_data_pool ("insn recog data pool" ); |
| 753 | |
| 754 | /* Initialize LRA data about insns. */ |
| 755 | static void |
| 756 | init_insn_recog_data (void) |
| 757 | { |
| 758 | lra_insn_recog_data_len = 0; |
| 759 | lra_insn_recog_data = NULL; |
| 760 | } |
| 761 | |
| 762 | /* Expand, if necessary, LRA data about insns. */ |
| 763 | static void |
| 764 | check_and_expand_insn_recog_data (int index) |
| 765 | { |
| 766 | int i, old; |
| 767 | |
| 768 | if (lra_insn_recog_data_len > index) |
| 769 | return; |
| 770 | old = lra_insn_recog_data_len; |
| 771 | lra_insn_recog_data_len = index * 3U / 2; |
| 772 | if (lra_insn_recog_data_len <= index) |
| 773 | lra_insn_recog_data_len = index + 1; |
| 774 | lra_insn_recog_data = XRESIZEVEC (lra_insn_recog_data_t, |
| 775 | lra_insn_recog_data, |
| 776 | lra_insn_recog_data_len); |
| 777 | for (i = old; i < lra_insn_recog_data_len; i++) |
| 778 | lra_insn_recog_data[i] = NULL; |
| 779 | } |
| 780 | |
| 781 | /* Finish LRA DATA about insn. */ |
| 782 | static void |
| 783 | free_insn_recog_data (lra_insn_recog_data_t data) |
| 784 | { |
| 785 | if (data->operand_loc != NULL) |
| 786 | free (ptr: data->operand_loc); |
| 787 | if (data->dup_loc != NULL) |
| 788 | free (ptr: data->dup_loc); |
| 789 | if (data->arg_hard_regs != NULL) |
| 790 | free (ptr: data->arg_hard_regs); |
| 791 | if (data->icode < 0 && NONDEBUG_INSN_P (data->insn)) |
| 792 | { |
| 793 | if (data->insn_static_data->operand_alternative != NULL) |
| 794 | free (ptr: const_cast <operand_alternative *> |
| 795 | (data->insn_static_data->operand_alternative)); |
| 796 | free_insn_regs (ir: data->insn_static_data->hard_regs); |
| 797 | free (ptr: data->insn_static_data); |
| 798 | } |
| 799 | free_insn_regs (ir: data->regs); |
| 800 | data->regs = NULL; |
| 801 | lra_insn_recog_data_pool.remove (object: data); |
| 802 | } |
| 803 | |
| 804 | /* Pools for copies. */ |
| 805 | static object_allocator<lra_copy> lra_copy_pool ("lra copies" ); |
| 806 | |
| 807 | /* Finish LRA data about all insns. */ |
| 808 | static void |
| 809 | finish_insn_recog_data (void) |
| 810 | { |
| 811 | int i; |
| 812 | lra_insn_recog_data_t data; |
| 813 | |
| 814 | for (i = 0; i < lra_insn_recog_data_len; i++) |
| 815 | if ((data = lra_insn_recog_data[i]) != NULL) |
| 816 | free_insn_recog_data (data); |
| 817 | finish_insn_regs (); |
| 818 | lra_copy_pool.release (); |
| 819 | lra_insn_reg_pool.release (); |
| 820 | lra_insn_recog_data_pool.release (); |
| 821 | free (ptr: lra_insn_recog_data); |
| 822 | } |
| 823 | |
| 824 | /* Setup info about operands in alternatives of LRA DATA of insn. */ |
| 825 | static void |
| 826 | setup_operand_alternative (lra_insn_recog_data_t data, |
| 827 | const operand_alternative *op_alt) |
| 828 | { |
| 829 | int i, j, nop, nalt; |
| 830 | int icode = data->icode; |
| 831 | struct lra_static_insn_data *static_data = data->insn_static_data; |
| 832 | |
| 833 | static_data->commutative = -1; |
| 834 | nop = static_data->n_operands; |
| 835 | nalt = static_data->n_alternatives; |
| 836 | static_data->operand_alternative = op_alt; |
| 837 | for (i = 0; i < nop; i++) |
| 838 | { |
| 839 | static_data->operand[i].early_clobber_alts = 0; |
| 840 | static_data->operand[i].is_address = false; |
| 841 | if (static_data->operand[i].constraint[0] == '%') |
| 842 | { |
| 843 | /* We currently only support one commutative pair of operands. */ |
| 844 | if (static_data->commutative < 0) |
| 845 | static_data->commutative = i; |
| 846 | else |
| 847 | lra_assert (icode < 0); /* Asm */ |
| 848 | /* The last operand should not be marked commutative. */ |
| 849 | lra_assert (i != nop - 1); |
| 850 | } |
| 851 | } |
| 852 | for (j = 0; j < nalt; j++) |
| 853 | for (i = 0; i < nop; i++, op_alt++) |
| 854 | { |
| 855 | if (op_alt->earlyclobber) |
| 856 | static_data->operand[i].early_clobber_alts |= (alternative_mask) 1 << j; |
| 857 | static_data->operand[i].is_address |= op_alt->is_address; |
| 858 | } |
| 859 | } |
| 860 | |
| 861 | /* Recursively process X and collect info about registers, which are |
| 862 | not the insn operands, in X with TYPE (in/out/inout) and flag that |
| 863 | it is early clobbered in the insn (EARLY_CLOBBER) and add the info |
| 864 | to LIST. X is a part of insn given by DATA. Return the result |
| 865 | list. */ |
| 866 | static struct lra_insn_reg * |
| 867 | collect_non_operand_hard_regs (rtx_insn *insn, rtx *x, |
| 868 | lra_insn_recog_data_t data, |
| 869 | struct lra_insn_reg *list, |
| 870 | enum op_type type, bool early_clobber) |
| 871 | { |
| 872 | int i, j, regno, last; |
| 873 | bool subreg_p; |
| 874 | machine_mode mode; |
| 875 | struct lra_insn_reg *curr; |
| 876 | rtx op = *x; |
| 877 | enum rtx_code code = GET_CODE (op); |
| 878 | const char *fmt = GET_RTX_FORMAT (code); |
| 879 | |
| 880 | for (i = 0; i < data->insn_static_data->n_operands; i++) |
| 881 | if (! data->insn_static_data->operand[i].is_operator |
| 882 | && x == data->operand_loc[i]) |
| 883 | /* It is an operand loc. Stop here. */ |
| 884 | return list; |
| 885 | for (i = 0; i < data->insn_static_data->n_dups; i++) |
| 886 | if (x == data->dup_loc[i]) |
| 887 | /* It is a dup loc. Stop here. */ |
| 888 | return list; |
| 889 | mode = GET_MODE (op); |
| 890 | subreg_p = false; |
| 891 | if (code == SUBREG) |
| 892 | { |
| 893 | mode = wider_subreg_mode (x: op); |
| 894 | if (read_modify_subreg_p (op)) |
| 895 | subreg_p = true; |
| 896 | op = SUBREG_REG (op); |
| 897 | code = GET_CODE (op); |
| 898 | } |
| 899 | if (REG_P (op)) |
| 900 | { |
| 901 | if ((regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER) |
| 902 | return list; |
| 903 | /* Process all regs even unallocatable ones as we need info |
| 904 | about all regs for rematerialization pass. */ |
| 905 | for (last = end_hard_regno (mode, regno); regno < last; regno++) |
| 906 | { |
| 907 | for (curr = list; curr != NULL; curr = curr->next) |
| 908 | if (curr->regno == regno && curr->subreg_p == subreg_p |
| 909 | && curr->biggest_mode == mode) |
| 910 | { |
| 911 | if (curr->type != type) |
| 912 | curr->type = OP_INOUT; |
| 913 | if (early_clobber) |
| 914 | curr->early_clobber_alts = ALL_ALTERNATIVES; |
| 915 | break; |
| 916 | } |
| 917 | if (curr == NULL) |
| 918 | { |
| 919 | /* This is a new hard regno or the info cannot be |
| 920 | integrated into the found structure. */ |
| 921 | #ifdef STACK_REGS |
| 922 | early_clobber |
| 923 | = (early_clobber |
| 924 | /* This clobber is to inform popping floating |
| 925 | point stack only. */ |
| 926 | && ! (FIRST_STACK_REG <= regno |
| 927 | && regno <= LAST_STACK_REG)); |
| 928 | #endif |
| 929 | list = new_insn_reg (insn: data->insn, regno, type, mode, subreg_p, |
| 930 | early_clobber_alts: early_clobber ? ALL_ALTERNATIVES : 0, next: list); |
| 931 | } |
| 932 | } |
| 933 | return list; |
| 934 | } |
| 935 | switch (code) |
| 936 | { |
| 937 | case SET: |
| 938 | list = collect_non_operand_hard_regs (insn, x: &SET_DEST (op), data, |
| 939 | list, type: OP_OUT, early_clobber: false); |
| 940 | list = collect_non_operand_hard_regs (insn, x: &SET_SRC (op), data, |
| 941 | list, type: OP_IN, early_clobber: false); |
| 942 | break; |
| 943 | case CLOBBER: |
| 944 | /* We treat clobber of non-operand hard registers as early clobber. */ |
| 945 | list = collect_non_operand_hard_regs (insn, x: &XEXP (op, 0), data, |
| 946 | list, type: OP_OUT, early_clobber: true); |
| 947 | break; |
| 948 | case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC: |
| 949 | list = collect_non_operand_hard_regs (insn, x: &XEXP (op, 0), data, |
| 950 | list, type: OP_INOUT, early_clobber: false); |
| 951 | break; |
| 952 | case PRE_MODIFY: case POST_MODIFY: |
| 953 | list = collect_non_operand_hard_regs (insn, x: &XEXP (op, 0), data, |
| 954 | list, type: OP_INOUT, early_clobber: false); |
| 955 | list = collect_non_operand_hard_regs (insn, x: &XEXP (op, 1), data, |
| 956 | list, type: OP_IN, early_clobber: false); |
| 957 | break; |
| 958 | default: |
| 959 | fmt = GET_RTX_FORMAT (code); |
| 960 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
| 961 | { |
| 962 | if (fmt[i] == 'e') |
| 963 | list = collect_non_operand_hard_regs (insn, x: &XEXP (op, i), data, |
| 964 | list, type: OP_IN, early_clobber: false); |
| 965 | else if (fmt[i] == 'E') |
| 966 | for (j = XVECLEN (op, i) - 1; j >= 0; j--) |
| 967 | list = collect_non_operand_hard_regs (insn, x: &XVECEXP (op, i, j), |
| 968 | data, list, type: OP_IN, early_clobber: false); |
| 969 | } |
| 970 | } |
| 971 | return list; |
| 972 | } |
| 973 | |
| 974 | /* Set up and return info about INSN. Set up the info if it is not set up |
| 975 | yet. */ |
| 976 | lra_insn_recog_data_t |
| 977 | lra_set_insn_recog_data (rtx_insn *insn) |
| 978 | { |
| 979 | lra_insn_recog_data_t data; |
| 980 | int i, n, icode; |
| 981 | rtx **locs; |
| 982 | unsigned int uid = INSN_UID (insn); |
| 983 | struct lra_static_insn_data *insn_static_data; |
| 984 | |
| 985 | check_and_expand_insn_recog_data (index: uid); |
| 986 | if (DEBUG_INSN_P (insn)) |
| 987 | icode = -1; |
| 988 | else |
| 989 | { |
| 990 | icode = INSN_CODE (insn); |
| 991 | if (icode < 0) |
| 992 | /* It might be a new simple insn which is not recognized yet. */ |
| 993 | INSN_CODE (insn) = icode = recog_memoized (insn); |
| 994 | } |
| 995 | data = lra_insn_recog_data_pool.allocate (); |
| 996 | lra_insn_recog_data[uid] = data; |
| 997 | data->insn = insn; |
| 998 | data->used_insn_alternative = LRA_UNKNOWN_ALT; |
| 999 | data->asm_reloads_num = 0; |
| 1000 | data->icode = icode; |
| 1001 | data->regs = NULL; |
| 1002 | if (DEBUG_INSN_P (insn)) |
| 1003 | { |
| 1004 | data->dup_loc = NULL; |
| 1005 | data->arg_hard_regs = NULL; |
| 1006 | data->preferred_alternatives = ALL_ALTERNATIVES; |
| 1007 | if (DEBUG_BIND_INSN_P (insn)) |
| 1008 | { |
| 1009 | data->insn_static_data = &debug_bind_static_data; |
| 1010 | data->operand_loc = XNEWVEC (rtx *, 1); |
| 1011 | data->operand_loc[0] = &INSN_VAR_LOCATION_LOC (insn); |
| 1012 | } |
| 1013 | else if (DEBUG_MARKER_INSN_P (insn)) |
| 1014 | { |
| 1015 | data->insn_static_data = &debug_marker_static_data; |
| 1016 | data->operand_loc = NULL; |
| 1017 | } |
| 1018 | return data; |
| 1019 | } |
| 1020 | if (icode < 0) |
| 1021 | { |
| 1022 | int nop, nalt; |
| 1023 | machine_mode operand_mode[MAX_RECOG_OPERANDS]; |
| 1024 | const char *constraints[MAX_RECOG_OPERANDS]; |
| 1025 | |
| 1026 | nop = asm_noperands (PATTERN (insn)); |
| 1027 | data->operand_loc = data->dup_loc = NULL; |
| 1028 | nalt = 1; |
| 1029 | if (nop < 0) |
| 1030 | { |
| 1031 | /* It is a special insn like USE or CLOBBER. We should |
| 1032 | recognize any regular insn otherwise LRA can do nothing |
| 1033 | with this insn. */ |
| 1034 | gcc_assert (GET_CODE (PATTERN (insn)) == USE |
| 1035 | || GET_CODE (PATTERN (insn)) == CLOBBER |
| 1036 | || GET_CODE (PATTERN (insn)) == ASM_INPUT); |
| 1037 | data->insn_static_data = insn_static_data |
| 1038 | = get_static_insn_data (icode: -1, nop: 0, ndup: 0, nalt); |
| 1039 | } |
| 1040 | else |
| 1041 | { |
| 1042 | /* expand_asm_operands makes sure there aren't too many |
| 1043 | operands. */ |
| 1044 | lra_assert (nop <= MAX_RECOG_OPERANDS); |
| 1045 | if (nop != 0) |
| 1046 | data->operand_loc = XNEWVEC (rtx *, nop); |
| 1047 | /* Now get the operand values and constraints out of the |
| 1048 | insn. */ |
| 1049 | decode_asm_operands (PATTERN (insn), NULL, |
| 1050 | data->operand_loc, |
| 1051 | constraints, operand_mode, NULL); |
| 1052 | if (nop > 0) |
| 1053 | for (const char *p =constraints[0]; *p; p++) |
| 1054 | nalt += *p == ','; |
| 1055 | data->insn_static_data = insn_static_data |
| 1056 | = get_static_insn_data (icode: -1, nop, ndup: 0, nalt); |
| 1057 | for (i = 0; i < nop; i++) |
| 1058 | { |
| 1059 | insn_static_data->operand[i].mode = operand_mode[i]; |
| 1060 | insn_static_data->operand[i].constraint = constraints[i]; |
| 1061 | insn_static_data->operand[i].strict_low = false; |
| 1062 | insn_static_data->operand[i].is_operator = false; |
| 1063 | insn_static_data->operand[i].is_address = false; |
| 1064 | } |
| 1065 | } |
| 1066 | for (i = 0; i < insn_static_data->n_operands; i++) |
| 1067 | insn_static_data->operand[i].type |
| 1068 | = (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT |
| 1069 | : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT |
| 1070 | : OP_IN); |
| 1071 | data->preferred_alternatives = ALL_ALTERNATIVES; |
| 1072 | if (nop > 0) |
| 1073 | { |
| 1074 | operand_alternative *op_alt = XCNEWVEC (operand_alternative, |
| 1075 | nalt * nop); |
| 1076 | preprocess_constraints (nop, nalt, constraints, op_alt, |
| 1077 | data->operand_loc); |
| 1078 | setup_operand_alternative (data, op_alt); |
| 1079 | } |
| 1080 | } |
| 1081 | else |
| 1082 | { |
| 1083 | insn_extract (insn); |
| 1084 | data->insn_static_data = insn_static_data |
| 1085 | = get_static_insn_data (icode, nop: insn_data[icode].n_operands, |
| 1086 | ndup: insn_data[icode].n_dups, |
| 1087 | nalt: insn_data[icode].n_alternatives); |
| 1088 | n = insn_static_data->n_operands; |
| 1089 | if (n == 0) |
| 1090 | locs = NULL; |
| 1091 | else |
| 1092 | { |
| 1093 | locs = XNEWVEC (rtx *, n); |
| 1094 | memcpy (dest: locs, src: recog_data.operand_loc, n: n * sizeof (rtx *)); |
| 1095 | } |
| 1096 | data->operand_loc = locs; |
| 1097 | n = insn_static_data->n_dups; |
| 1098 | if (n == 0) |
| 1099 | locs = NULL; |
| 1100 | else |
| 1101 | { |
| 1102 | locs = XNEWVEC (rtx *, n); |
| 1103 | memcpy (dest: locs, src: recog_data.dup_loc, n: n * sizeof (rtx *)); |
| 1104 | } |
| 1105 | data->dup_loc = locs; |
| 1106 | data->preferred_alternatives = get_preferred_alternatives (insn); |
| 1107 | const operand_alternative *op_alt = preprocess_insn_constraints (icode); |
| 1108 | if (!insn_static_data->operand_alternative) |
| 1109 | setup_operand_alternative (data, op_alt); |
| 1110 | else if (op_alt != insn_static_data->operand_alternative) |
| 1111 | insn_static_data->operand_alternative = op_alt; |
| 1112 | } |
| 1113 | if (GET_CODE (PATTERN (insn)) == CLOBBER || GET_CODE (PATTERN (insn)) == USE) |
| 1114 | insn_static_data->hard_regs = NULL; |
| 1115 | else |
| 1116 | insn_static_data->hard_regs |
| 1117 | = collect_non_operand_hard_regs (insn, x: &PATTERN (insn), data, |
| 1118 | NULL, type: OP_IN, early_clobber: false); |
| 1119 | data->arg_hard_regs = NULL; |
| 1120 | if (CALL_P (insn)) |
| 1121 | { |
| 1122 | bool use_p; |
| 1123 | rtx link; |
| 1124 | int n_hard_regs, regno, arg_hard_regs[FIRST_PSEUDO_REGISTER]; |
| 1125 | |
| 1126 | n_hard_regs = 0; |
| 1127 | /* Finding implicit hard register usage. We believe it will be |
| 1128 | not changed whatever transformations are used. Call insns |
| 1129 | are such example. */ |
| 1130 | for (link = CALL_INSN_FUNCTION_USAGE (insn); |
| 1131 | link != NULL_RTX; |
| 1132 | link = XEXP (link, 1)) |
| 1133 | if (((use_p = GET_CODE (XEXP (link, 0)) == USE) |
| 1134 | || GET_CODE (XEXP (link, 0)) == CLOBBER) |
| 1135 | && REG_P (XEXP (XEXP (link, 0), 0))) |
| 1136 | { |
| 1137 | regno = REGNO (XEXP (XEXP (link, 0), 0)); |
| 1138 | lra_assert (regno < FIRST_PSEUDO_REGISTER); |
| 1139 | /* It is an argument register. */ |
| 1140 | for (i = REG_NREGS (XEXP (XEXP (link, 0), 0)) - 1; i >= 0; i--) |
| 1141 | arg_hard_regs[n_hard_regs++] |
| 1142 | = regno + i + (use_p ? 0 : FIRST_PSEUDO_REGISTER); |
| 1143 | } |
| 1144 | |
| 1145 | if (n_hard_regs != 0) |
| 1146 | { |
| 1147 | arg_hard_regs[n_hard_regs++] = -1; |
| 1148 | data->arg_hard_regs = XNEWVEC (int, n_hard_regs); |
| 1149 | memcpy (dest: data->arg_hard_regs, src: arg_hard_regs, |
| 1150 | n: sizeof (int) * n_hard_regs); |
| 1151 | } |
| 1152 | } |
| 1153 | /* Some output operand can be recognized only from the context not |
| 1154 | from the constraints which are empty in this case. Call insn may |
| 1155 | contain a hard register in set destination with empty constraint |
| 1156 | and extract_insn treats them as an input. */ |
| 1157 | for (i = 0; i < insn_static_data->n_operands; i++) |
| 1158 | { |
| 1159 | int j; |
| 1160 | rtx pat, set; |
| 1161 | struct lra_operand_data *operand = &insn_static_data->operand[i]; |
| 1162 | |
| 1163 | /* ??? Should we treat 'X' the same way. It looks to me that |
| 1164 | 'X' means anything and empty constraint means we do not |
| 1165 | care. */ |
| 1166 | if (operand->type != OP_IN || *operand->constraint != '\0' |
| 1167 | || operand->is_operator) |
| 1168 | continue; |
| 1169 | pat = PATTERN (insn); |
| 1170 | if (GET_CODE (pat) == SET) |
| 1171 | { |
| 1172 | if (data->operand_loc[i] != &SET_DEST (pat)) |
| 1173 | continue; |
| 1174 | } |
| 1175 | else if (GET_CODE (pat) == PARALLEL) |
| 1176 | { |
| 1177 | for (j = XVECLEN (pat, 0) - 1; j >= 0; j--) |
| 1178 | { |
| 1179 | set = XVECEXP (PATTERN (insn), 0, j); |
| 1180 | if (GET_CODE (set) == SET |
| 1181 | && &SET_DEST (set) == data->operand_loc[i]) |
| 1182 | break; |
| 1183 | } |
| 1184 | if (j < 0) |
| 1185 | continue; |
| 1186 | } |
| 1187 | else |
| 1188 | continue; |
| 1189 | operand->type = OP_OUT; |
| 1190 | } |
| 1191 | return data; |
| 1192 | } |
| 1193 | |
| 1194 | /* Return info about insn give by UID. The info should be already set |
| 1195 | up. */ |
| 1196 | static lra_insn_recog_data_t |
| 1197 | get_insn_recog_data_by_uid (int uid) |
| 1198 | { |
| 1199 | lra_insn_recog_data_t data; |
| 1200 | |
| 1201 | data = lra_insn_recog_data[uid]; |
| 1202 | lra_assert (data != NULL); |
| 1203 | return data; |
| 1204 | } |
| 1205 | |
| 1206 | /* Invalidate all info about insn given by its UID. */ |
| 1207 | static void |
| 1208 | invalidate_insn_recog_data (int uid) |
| 1209 | { |
| 1210 | lra_insn_recog_data_t data; |
| 1211 | |
| 1212 | data = lra_insn_recog_data[uid]; |
| 1213 | lra_assert (data != NULL); |
| 1214 | free_insn_recog_data (data); |
| 1215 | lra_insn_recog_data[uid] = NULL; |
| 1216 | } |
| 1217 | |
| 1218 | /* Update all the insn info about INSN. It is usually called when |
| 1219 | something in the insn was changed. Return the updated info. */ |
| 1220 | lra_insn_recog_data_t |
| 1221 | lra_update_insn_recog_data (rtx_insn *insn) |
| 1222 | { |
| 1223 | lra_insn_recog_data_t data; |
| 1224 | int n; |
| 1225 | unsigned int uid = INSN_UID (insn); |
| 1226 | struct lra_static_insn_data *insn_static_data; |
| 1227 | poly_int64 sp_offset = 0; |
| 1228 | |
| 1229 | check_and_expand_insn_recog_data (index: uid); |
| 1230 | if ((data = lra_insn_recog_data[uid]) != NULL |
| 1231 | && data->icode != INSN_CODE (insn)) |
| 1232 | { |
| 1233 | sp_offset = data->sp_offset; |
| 1234 | invalidate_insn_data_regno_info (data, insn, get_insn_freq (insn)); |
| 1235 | invalidate_insn_recog_data (uid); |
| 1236 | data = NULL; |
| 1237 | } |
| 1238 | if (data == NULL) |
| 1239 | { |
| 1240 | data = lra_get_insn_recog_data (insn); |
| 1241 | /* Initiate or restore SP offset. */ |
| 1242 | data->sp_offset = sp_offset; |
| 1243 | return data; |
| 1244 | } |
| 1245 | insn_static_data = data->insn_static_data; |
| 1246 | data->used_insn_alternative = LRA_UNKNOWN_ALT; |
| 1247 | if (DEBUG_INSN_P (insn)) |
| 1248 | return data; |
| 1249 | if (data->icode < 0) |
| 1250 | { |
| 1251 | int nop; |
| 1252 | machine_mode operand_mode[MAX_RECOG_OPERANDS]; |
| 1253 | const char *constraints[MAX_RECOG_OPERANDS]; |
| 1254 | |
| 1255 | nop = asm_noperands (PATTERN (insn)); |
| 1256 | if (nop >= 0) |
| 1257 | { |
| 1258 | lra_assert (nop == data->insn_static_data->n_operands); |
| 1259 | /* Now get the operand values and constraints out of the |
| 1260 | insn. */ |
| 1261 | decode_asm_operands (PATTERN (insn), NULL, |
| 1262 | data->operand_loc, |
| 1263 | constraints, operand_mode, NULL); |
| 1264 | |
| 1265 | if (flag_checking) |
| 1266 | for (int i = 0; i < nop; i++) |
| 1267 | lra_assert |
| 1268 | (insn_static_data->operand[i].mode == operand_mode[i] |
| 1269 | && insn_static_data->operand[i].constraint == constraints[i] |
| 1270 | && ! insn_static_data->operand[i].is_operator); |
| 1271 | } |
| 1272 | |
| 1273 | if (flag_checking) |
| 1274 | for (int i = 0; i < insn_static_data->n_operands; i++) |
| 1275 | lra_assert |
| 1276 | (insn_static_data->operand[i].type |
| 1277 | == (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT |
| 1278 | : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT |
| 1279 | : OP_IN)); |
| 1280 | } |
| 1281 | else |
| 1282 | { |
| 1283 | insn_extract (insn); |
| 1284 | n = insn_static_data->n_operands; |
| 1285 | if (n != 0) |
| 1286 | memcpy (dest: data->operand_loc, src: recog_data.operand_loc, n: n * sizeof (rtx *)); |
| 1287 | n = insn_static_data->n_dups; |
| 1288 | if (n != 0) |
| 1289 | memcpy (dest: data->dup_loc, src: recog_data.dup_loc, n: n * sizeof (rtx *)); |
| 1290 | lra_assert (check_bool_attrs (insn)); |
| 1291 | } |
| 1292 | return data; |
| 1293 | } |
| 1294 | |
| 1295 | /* Set up that INSN is using alternative ALT now. */ |
| 1296 | void |
| 1297 | lra_set_used_insn_alternative (rtx_insn *insn, int alt) |
| 1298 | { |
| 1299 | lra_insn_recog_data_t data; |
| 1300 | |
| 1301 | data = lra_get_insn_recog_data (insn); |
| 1302 | data->used_insn_alternative = alt; |
| 1303 | } |
| 1304 | |
| 1305 | /* Set up that insn with UID is using alternative ALT now. The insn |
| 1306 | info should be already set up. */ |
| 1307 | void |
| 1308 | lra_set_used_insn_alternative_by_uid (int uid, int alt) |
| 1309 | { |
| 1310 | lra_insn_recog_data_t data; |
| 1311 | |
| 1312 | check_and_expand_insn_recog_data (index: uid); |
| 1313 | data = lra_insn_recog_data[uid]; |
| 1314 | lra_assert (data != NULL); |
| 1315 | data->used_insn_alternative = alt; |
| 1316 | } |
| 1317 | |
| 1318 | |
| 1319 | |
| 1320 | /* This page contains code dealing with common register info and |
| 1321 | pseudo copies. */ |
| 1322 | |
| 1323 | /* The size of the following array. */ |
| 1324 | static int reg_info_size; |
| 1325 | /* Common info about each register. */ |
| 1326 | class lra_reg *lra_reg_info; |
| 1327 | |
| 1328 | HARD_REG_SET hard_regs_spilled_into; |
| 1329 | |
| 1330 | /* Last register value. */ |
| 1331 | static int last_reg_value; |
| 1332 | |
| 1333 | /* Return new register value. */ |
| 1334 | static int |
| 1335 | get_new_reg_value (void) |
| 1336 | { |
| 1337 | return ++last_reg_value; |
| 1338 | } |
| 1339 | |
| 1340 | /* Vec referring to pseudo copies. */ |
| 1341 | static vec<lra_copy_t> copy_vec; |
| 1342 | |
| 1343 | /* Initialize I-th element of lra_reg_info. */ |
| 1344 | static inline void |
| 1345 | initialize_lra_reg_info_element (int i) |
| 1346 | { |
| 1347 | bitmap_initialize (head: &lra_reg_info[i].insn_bitmap, obstack: ®_obstack); |
| 1348 | #ifdef STACK_REGS |
| 1349 | lra_reg_info[i].no_stack_p = false; |
| 1350 | #endif |
| 1351 | CLEAR_HARD_REG_SET (set&: lra_reg_info[i].conflict_hard_regs); |
| 1352 | CLEAR_HARD_REG_SET (set&: lra_reg_info[i].exclude_start_hard_regs); |
| 1353 | lra_reg_info[i].preferred_hard_regno1 = -1; |
| 1354 | lra_reg_info[i].preferred_hard_regno2 = -1; |
| 1355 | lra_reg_info[i].preferred_hard_regno_profit1 = 0; |
| 1356 | lra_reg_info[i].preferred_hard_regno_profit2 = 0; |
| 1357 | lra_reg_info[i].biggest_mode = VOIDmode; |
| 1358 | lra_reg_info[i].live_ranges = NULL; |
| 1359 | lra_reg_info[i].nrefs = lra_reg_info[i].freq = 0; |
| 1360 | lra_reg_info[i].last_reload = 0; |
| 1361 | lra_reg_info[i].restore_rtx = NULL_RTX; |
| 1362 | lra_reg_info[i].val = get_new_reg_value (); |
| 1363 | lra_reg_info[i].offset = 0; |
| 1364 | lra_reg_info[i].copies = NULL; |
| 1365 | } |
| 1366 | |
| 1367 | /* Initialize common reg info and copies. */ |
| 1368 | static void |
| 1369 | init_reg_info (void) |
| 1370 | { |
| 1371 | int i; |
| 1372 | |
| 1373 | last_reg_value = 0; |
| 1374 | reg_info_size = max_reg_num () * 3 / 2 + 1; |
| 1375 | lra_reg_info = XNEWVEC (class lra_reg, reg_info_size); |
| 1376 | for (i = 0; i < reg_info_size; i++) |
| 1377 | initialize_lra_reg_info_element (i); |
| 1378 | copy_vec.truncate (size: 0); |
| 1379 | CLEAR_HARD_REG_SET (set&: hard_regs_spilled_into); |
| 1380 | } |
| 1381 | |
| 1382 | |
| 1383 | /* Finish common reg info and copies. */ |
| 1384 | static void |
| 1385 | finish_reg_info (void) |
| 1386 | { |
| 1387 | int i; |
| 1388 | |
| 1389 | for (i = 0; i < reg_info_size; i++) |
| 1390 | bitmap_clear (&lra_reg_info[i].insn_bitmap); |
| 1391 | free (ptr: lra_reg_info); |
| 1392 | reg_info_size = 0; |
| 1393 | } |
| 1394 | |
| 1395 | /* Expand common reg info if it is necessary. */ |
| 1396 | static void |
| 1397 | expand_reg_info (void) |
| 1398 | { |
| 1399 | int i, old = reg_info_size; |
| 1400 | |
| 1401 | if (reg_info_size > max_reg_num ()) |
| 1402 | return; |
| 1403 | reg_info_size = max_reg_num () * 3 / 2 + 1; |
| 1404 | lra_reg_info = XRESIZEVEC (class lra_reg, lra_reg_info, reg_info_size); |
| 1405 | for (i = old; i < reg_info_size; i++) |
| 1406 | initialize_lra_reg_info_element (i); |
| 1407 | } |
| 1408 | |
| 1409 | /* Free all copies. */ |
| 1410 | void |
| 1411 | lra_free_copies (void) |
| 1412 | { |
| 1413 | lra_copy_t cp; |
| 1414 | |
| 1415 | while (copy_vec.length () != 0) |
| 1416 | { |
| 1417 | cp = copy_vec.pop (); |
| 1418 | lra_reg_info[cp->regno1].copies = lra_reg_info[cp->regno2].copies = NULL; |
| 1419 | lra_copy_pool.remove (object: cp); |
| 1420 | } |
| 1421 | } |
| 1422 | |
| 1423 | /* Create copy of two pseudos REGNO1 and REGNO2. The copy execution |
| 1424 | frequency is FREQ. */ |
| 1425 | void |
| 1426 | lra_create_copy (int regno1, int regno2, int freq) |
| 1427 | { |
| 1428 | bool regno1_dest_p; |
| 1429 | lra_copy_t cp; |
| 1430 | |
| 1431 | lra_assert (regno1 != regno2); |
| 1432 | regno1_dest_p = true; |
| 1433 | if (regno1 > regno2) |
| 1434 | { |
| 1435 | std::swap (a&: regno1, b&: regno2); |
| 1436 | regno1_dest_p = false; |
| 1437 | } |
| 1438 | cp = lra_copy_pool.allocate (); |
| 1439 | copy_vec.safe_push (obj: cp); |
| 1440 | cp->regno1_dest_p = regno1_dest_p; |
| 1441 | cp->freq = freq; |
| 1442 | cp->regno1 = regno1; |
| 1443 | cp->regno2 = regno2; |
| 1444 | cp->regno1_next = lra_reg_info[regno1].copies; |
| 1445 | lra_reg_info[regno1].copies = cp; |
| 1446 | cp->regno2_next = lra_reg_info[regno2].copies; |
| 1447 | lra_reg_info[regno2].copies = cp; |
| 1448 | if (lra_dump_file != NULL) |
| 1449 | fprintf (stream: lra_dump_file, format: " Creating copy r%d%sr%d@%d\n" , |
| 1450 | regno1, regno1_dest_p ? "<-" : "->" , regno2, freq); |
| 1451 | } |
| 1452 | |
| 1453 | /* Return N-th (0, 1, ...) copy. If there is no copy, return |
| 1454 | NULL. */ |
| 1455 | lra_copy_t |
| 1456 | lra_get_copy (int n) |
| 1457 | { |
| 1458 | if (n >= (int) copy_vec.length ()) |
| 1459 | return NULL; |
| 1460 | return copy_vec[n]; |
| 1461 | } |
| 1462 | |
| 1463 | |
| 1464 | |
| 1465 | /* This page contains code dealing with info about registers in |
| 1466 | insns. */ |
| 1467 | |
| 1468 | /* Process X of INSN recursively and add info (operand type is given |
| 1469 | by TYPE) about registers in X to the insn DATA. If X can be early |
| 1470 | clobbered, alternatives in which it can be early clobbered are given |
| 1471 | by EARLY_CLOBBER_ALTS. */ |
| 1472 | static void |
| 1473 | add_regs_to_insn_regno_info (lra_insn_recog_data_t data, rtx x, |
| 1474 | rtx_insn *insn, enum op_type type, |
| 1475 | alternative_mask early_clobber_alts) |
| 1476 | { |
| 1477 | int i, j, regno; |
| 1478 | bool subreg_p; |
| 1479 | machine_mode mode; |
| 1480 | const char *fmt; |
| 1481 | enum rtx_code code; |
| 1482 | struct lra_insn_reg *curr; |
| 1483 | |
| 1484 | code = GET_CODE (x); |
| 1485 | mode = GET_MODE (x); |
| 1486 | subreg_p = false; |
| 1487 | if (GET_CODE (x) == SUBREG) |
| 1488 | { |
| 1489 | mode = wider_subreg_mode (x); |
| 1490 | if (read_modify_subreg_p (x)) |
| 1491 | subreg_p = true; |
| 1492 | x = SUBREG_REG (x); |
| 1493 | code = GET_CODE (x); |
| 1494 | } |
| 1495 | if (REG_P (x)) |
| 1496 | { |
| 1497 | regno = REGNO (x); |
| 1498 | /* Process all regs even unallocatable ones as we need info about |
| 1499 | all regs for rematerialization pass. */ |
| 1500 | expand_reg_info (); |
| 1501 | if (bitmap_set_bit (&lra_reg_info[regno].insn_bitmap, INSN_UID (insn))) |
| 1502 | { |
| 1503 | data->regs = new_insn_reg (insn: data->insn, regno, type, mode, subreg_p, |
| 1504 | early_clobber_alts, next: data->regs); |
| 1505 | return; |
| 1506 | } |
| 1507 | else |
| 1508 | { |
| 1509 | for (curr = data->regs; curr != NULL; curr = curr->next) |
| 1510 | if (curr->regno == regno) |
| 1511 | { |
| 1512 | if (curr->subreg_p != subreg_p || curr->biggest_mode != mode) |
| 1513 | /* The info cannot be integrated into the found |
| 1514 | structure. */ |
| 1515 | data->regs = new_insn_reg (insn: data->insn, regno, type, mode, |
| 1516 | subreg_p, early_clobber_alts, |
| 1517 | next: data->regs); |
| 1518 | else |
| 1519 | { |
| 1520 | if (curr->type != type) |
| 1521 | curr->type = OP_INOUT; |
| 1522 | curr->early_clobber_alts |= early_clobber_alts; |
| 1523 | } |
| 1524 | return; |
| 1525 | } |
| 1526 | gcc_unreachable (); |
| 1527 | } |
| 1528 | } |
| 1529 | |
| 1530 | switch (code) |
| 1531 | { |
| 1532 | case SET: |
| 1533 | add_regs_to_insn_regno_info (data, SET_DEST (x), insn, type: OP_OUT, early_clobber_alts: 0); |
| 1534 | add_regs_to_insn_regno_info (data, SET_SRC (x), insn, type: OP_IN, early_clobber_alts: 0); |
| 1535 | break; |
| 1536 | case CLOBBER: |
| 1537 | /* We treat clobber of non-operand hard registers as early |
| 1538 | clobber. */ |
| 1539 | add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, type: OP_OUT, |
| 1540 | ALL_ALTERNATIVES); |
| 1541 | break; |
| 1542 | case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC: |
| 1543 | add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, type: OP_INOUT, early_clobber_alts: 0); |
| 1544 | break; |
| 1545 | case PRE_MODIFY: case POST_MODIFY: |
| 1546 | add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, type: OP_INOUT, early_clobber_alts: 0); |
| 1547 | add_regs_to_insn_regno_info (data, XEXP (x, 1), insn, type: OP_IN, early_clobber_alts: 0); |
| 1548 | break; |
| 1549 | default: |
| 1550 | if ((code != PARALLEL && code != EXPR_LIST) || type != OP_OUT) |
| 1551 | /* Some targets place small structures in registers for return |
| 1552 | values of functions, and those registers are wrapped in |
| 1553 | PARALLEL that we may see as the destination of a SET. Here |
| 1554 | is an example: |
| 1555 | |
| 1556 | (call_insn 13 12 14 2 (set (parallel:BLK [ |
| 1557 | (expr_list:REG_DEP_TRUE (reg:DI 0 ax) |
| 1558 | (const_int 0 [0])) |
| 1559 | (expr_list:REG_DEP_TRUE (reg:DI 1 dx) |
| 1560 | (const_int 8 [0x8])) |
| 1561 | ]) |
| 1562 | (call (mem:QI (symbol_ref:DI (... */ |
| 1563 | type = OP_IN; |
| 1564 | fmt = GET_RTX_FORMAT (code); |
| 1565 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
| 1566 | { |
| 1567 | if (fmt[i] == 'e') |
| 1568 | add_regs_to_insn_regno_info (data, XEXP (x, i), insn, type, early_clobber_alts: 0); |
| 1569 | else if (fmt[i] == 'E') |
| 1570 | { |
| 1571 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) |
| 1572 | add_regs_to_insn_regno_info (data, XVECEXP (x, i, j), insn, |
| 1573 | type, early_clobber_alts: 0); |
| 1574 | } |
| 1575 | } |
| 1576 | } |
| 1577 | } |
| 1578 | |
| 1579 | /* Return execution frequency of INSN. */ |
| 1580 | static int |
| 1581 | get_insn_freq (rtx_insn *insn) |
| 1582 | { |
| 1583 | basic_block bb = BLOCK_FOR_INSN (insn); |
| 1584 | |
| 1585 | gcc_checking_assert (bb != NULL); |
| 1586 | return REG_FREQ_FROM_BB (bb); |
| 1587 | } |
| 1588 | |
| 1589 | /* Invalidate all reg info of INSN with DATA and execution frequency |
| 1590 | FREQ. Update common info about the invalidated registers. */ |
| 1591 | static void |
| 1592 | invalidate_insn_data_regno_info (lra_insn_recog_data_t data, rtx_insn *insn, |
| 1593 | int freq) |
| 1594 | { |
| 1595 | int uid; |
| 1596 | bool debug_p; |
| 1597 | unsigned int i; |
| 1598 | struct lra_insn_reg *ir, *next_ir; |
| 1599 | |
| 1600 | uid = INSN_UID (insn); |
| 1601 | debug_p = DEBUG_INSN_P (insn); |
| 1602 | for (ir = data->regs; ir != NULL; ir = next_ir) |
| 1603 | { |
| 1604 | i = ir->regno; |
| 1605 | next_ir = ir->next; |
| 1606 | lra_insn_reg_pool.remove (object: ir); |
| 1607 | bitmap_clear_bit (&lra_reg_info[i].insn_bitmap, uid); |
| 1608 | if (i >= FIRST_PSEUDO_REGISTER && ! debug_p) |
| 1609 | { |
| 1610 | lra_reg_info[i].nrefs--; |
| 1611 | lra_reg_info[i].freq -= freq; |
| 1612 | lra_assert (lra_reg_info[i].nrefs >= 0 && lra_reg_info[i].freq >= 0); |
| 1613 | } |
| 1614 | } |
| 1615 | data->regs = NULL; |
| 1616 | } |
| 1617 | |
| 1618 | /* Invalidate all reg info of INSN. Update common info about the |
| 1619 | invalidated registers. */ |
| 1620 | void |
| 1621 | lra_invalidate_insn_regno_info (rtx_insn *insn) |
| 1622 | { |
| 1623 | invalidate_insn_data_regno_info (data: lra_get_insn_recog_data (insn), insn, |
| 1624 | freq: get_insn_freq (insn)); |
| 1625 | } |
| 1626 | |
| 1627 | /* Update common reg info from reg info of insn given by its DATA and |
| 1628 | execution frequency FREQ. */ |
| 1629 | static void |
| 1630 | setup_insn_reg_info (lra_insn_recog_data_t data, int freq) |
| 1631 | { |
| 1632 | unsigned int i; |
| 1633 | struct lra_insn_reg *ir; |
| 1634 | |
| 1635 | for (ir = data->regs; ir != NULL; ir = ir->next) |
| 1636 | if ((i = ir->regno) >= FIRST_PSEUDO_REGISTER) |
| 1637 | { |
| 1638 | lra_reg_info[i].nrefs++; |
| 1639 | lra_reg_info[i].freq += freq; |
| 1640 | } |
| 1641 | } |
| 1642 | |
| 1643 | /* Set up insn reg info of INSN. Update common reg info from reg info |
| 1644 | of INSN. */ |
| 1645 | void |
| 1646 | lra_update_insn_regno_info (rtx_insn *insn) |
| 1647 | { |
| 1648 | int i, freq; |
| 1649 | lra_insn_recog_data_t data; |
| 1650 | struct lra_static_insn_data *static_data; |
| 1651 | enum rtx_code code; |
| 1652 | rtx link; |
| 1653 | |
| 1654 | if (! INSN_P (insn)) |
| 1655 | return; |
| 1656 | data = lra_get_insn_recog_data (insn); |
| 1657 | static_data = data->insn_static_data; |
| 1658 | freq = NONDEBUG_INSN_P (insn) ? get_insn_freq (insn) : 0; |
| 1659 | invalidate_insn_data_regno_info (data, insn, freq); |
| 1660 | for (i = static_data->n_operands - 1; i >= 0; i--) |
| 1661 | add_regs_to_insn_regno_info (data, x: *data->operand_loc[i], insn, |
| 1662 | type: static_data->operand[i].type, |
| 1663 | early_clobber_alts: static_data->operand[i].early_clobber_alts); |
| 1664 | if ((code = GET_CODE (PATTERN (insn))) == CLOBBER || code == USE) |
| 1665 | add_regs_to_insn_regno_info (data, XEXP (PATTERN (insn), 0), insn, |
| 1666 | type: code == USE ? OP_IN : OP_OUT, early_clobber_alts: 0); |
| 1667 | if (CALL_P (insn)) |
| 1668 | /* On some targets call insns can refer to pseudos in memory in |
| 1669 | CALL_INSN_FUNCTION_USAGE list. Process them in order to |
| 1670 | consider their occurrences in calls for different |
| 1671 | transformations (e.g. inheritance) with given pseudos. */ |
| 1672 | for (link = CALL_INSN_FUNCTION_USAGE (insn); |
| 1673 | link != NULL_RTX; |
| 1674 | link = XEXP (link, 1)) |
| 1675 | { |
| 1676 | code = GET_CODE (XEXP (link, 0)); |
| 1677 | if ((code == USE || code == CLOBBER) |
| 1678 | && MEM_P (XEXP (XEXP (link, 0), 0))) |
| 1679 | add_regs_to_insn_regno_info (data, XEXP (XEXP (link, 0), 0), insn, |
| 1680 | type: code == USE ? OP_IN : OP_OUT, early_clobber_alts: 0); |
| 1681 | } |
| 1682 | if (NONDEBUG_INSN_P (insn)) |
| 1683 | setup_insn_reg_info (data, freq); |
| 1684 | } |
| 1685 | |
| 1686 | /* Return reg info of insn given by it UID. */ |
| 1687 | struct lra_insn_reg * |
| 1688 | lra_get_insn_regs (int uid) |
| 1689 | { |
| 1690 | lra_insn_recog_data_t data; |
| 1691 | |
| 1692 | data = get_insn_recog_data_by_uid (uid); |
| 1693 | return data->regs; |
| 1694 | } |
| 1695 | |
| 1696 | |
| 1697 | |
| 1698 | /* Recursive hash function for RTL X. */ |
| 1699 | hashval_t |
| 1700 | lra_rtx_hash (rtx x) |
| 1701 | { |
| 1702 | int i, j; |
| 1703 | enum rtx_code code; |
| 1704 | const char *fmt; |
| 1705 | hashval_t val = 0; |
| 1706 | |
| 1707 | if (x == 0) |
| 1708 | return val; |
| 1709 | |
| 1710 | code = GET_CODE (x); |
| 1711 | val += (int) code + 4095; |
| 1712 | |
| 1713 | /* Some RTL can be compared nonrecursively. */ |
| 1714 | switch (code) |
| 1715 | { |
| 1716 | case REG: |
| 1717 | return val + REGNO (x); |
| 1718 | |
| 1719 | case LABEL_REF: |
| 1720 | return iterative_hash_object (XEXP (x, 0), val); |
| 1721 | |
| 1722 | case SYMBOL_REF: |
| 1723 | return iterative_hash_object (XSTR (x, 0), val); |
| 1724 | |
| 1725 | case SCRATCH: |
| 1726 | case CONST_DOUBLE: |
| 1727 | case CONST_VECTOR: |
| 1728 | return val; |
| 1729 | |
| 1730 | case CONST_INT: |
| 1731 | return val + UINTVAL (x); |
| 1732 | |
| 1733 | case SUBREG: |
| 1734 | val += lra_rtx_hash (SUBREG_REG (x)); |
| 1735 | for (int i = 0; i < NUM_POLY_INT_COEFFS; ++i) |
| 1736 | val += SUBREG_BYTE (x).coeffs[i]; |
| 1737 | return val; |
| 1738 | |
| 1739 | default: |
| 1740 | break; |
| 1741 | } |
| 1742 | |
| 1743 | /* Hash the elements. */ |
| 1744 | fmt = GET_RTX_FORMAT (code); |
| 1745 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
| 1746 | { |
| 1747 | switch (fmt[i]) |
| 1748 | { |
| 1749 | case 'w': |
| 1750 | val += XWINT (x, i); |
| 1751 | break; |
| 1752 | |
| 1753 | case 'n': |
| 1754 | case 'i': |
| 1755 | val += XINT (x, i); |
| 1756 | break; |
| 1757 | |
| 1758 | case 'L': |
| 1759 | val += XLOC (x, i); |
| 1760 | break; |
| 1761 | |
| 1762 | case 'V': |
| 1763 | case 'E': |
| 1764 | val += XVECLEN (x, i); |
| 1765 | |
| 1766 | for (j = 0; j < XVECLEN (x, i); j++) |
| 1767 | val += lra_rtx_hash (XVECEXP (x, i, j)); |
| 1768 | break; |
| 1769 | |
| 1770 | case 'e': |
| 1771 | val += lra_rtx_hash (XEXP (x, i)); |
| 1772 | break; |
| 1773 | |
| 1774 | case 'S': |
| 1775 | case 's': |
| 1776 | val += htab_hash_string (XSTR (x, i)); |
| 1777 | break; |
| 1778 | |
| 1779 | case 'u': |
| 1780 | case '0': |
| 1781 | case 't': |
| 1782 | break; |
| 1783 | |
| 1784 | /* It is believed that rtx's at this level will never |
| 1785 | contain anything but integers and other rtx's, except for |
| 1786 | within LABEL_REFs and SYMBOL_REFs. */ |
| 1787 | default: |
| 1788 | abort (); |
| 1789 | } |
| 1790 | } |
| 1791 | return val; |
| 1792 | } |
| 1793 | |
| 1794 | |
| 1795 | |
| 1796 | /* This page contains code dealing with stack of the insns which |
| 1797 | should be processed by the next constraint pass. */ |
| 1798 | |
| 1799 | /* Bitmap used to put an insn on the stack only in one exemplar. */ |
| 1800 | static sbitmap lra_constraint_insn_stack_bitmap; |
| 1801 | |
| 1802 | /* The stack itself. */ |
| 1803 | vec<rtx_insn *> lra_constraint_insn_stack; |
| 1804 | |
| 1805 | /* Put INSN on the stack. If ALWAYS_UPDATE is true, always update the reg |
| 1806 | info for INSN, otherwise only update it if INSN is not already on the |
| 1807 | stack. */ |
| 1808 | static inline void |
| 1809 | lra_push_insn_1 (rtx_insn *insn, bool always_update) |
| 1810 | { |
| 1811 | unsigned int uid = INSN_UID (insn); |
| 1812 | if (always_update) |
| 1813 | lra_update_insn_regno_info (insn); |
| 1814 | if (uid >= SBITMAP_SIZE (lra_constraint_insn_stack_bitmap)) |
| 1815 | lra_constraint_insn_stack_bitmap = |
| 1816 | sbitmap_resize (lra_constraint_insn_stack_bitmap, 3 * uid / 2, 0); |
| 1817 | if (bitmap_bit_p (map: lra_constraint_insn_stack_bitmap, bitno: uid)) |
| 1818 | return; |
| 1819 | bitmap_set_bit (map: lra_constraint_insn_stack_bitmap, bitno: uid); |
| 1820 | if (! always_update) |
| 1821 | lra_update_insn_regno_info (insn); |
| 1822 | lra_constraint_insn_stack.safe_push (obj: insn); |
| 1823 | } |
| 1824 | |
| 1825 | /* Put INSN on the stack. */ |
| 1826 | void |
| 1827 | lra_push_insn (rtx_insn *insn) |
| 1828 | { |
| 1829 | lra_push_insn_1 (insn, always_update: false); |
| 1830 | } |
| 1831 | |
| 1832 | /* Put INSN on the stack and update its reg info. */ |
| 1833 | void |
| 1834 | lra_push_insn_and_update_insn_regno_info (rtx_insn *insn) |
| 1835 | { |
| 1836 | lra_push_insn_1 (insn, always_update: true); |
| 1837 | } |
| 1838 | |
| 1839 | /* Put insn with UID on the stack. */ |
| 1840 | void |
| 1841 | lra_push_insn_by_uid (unsigned int uid) |
| 1842 | { |
| 1843 | lra_push_insn (insn: lra_insn_recog_data[uid]->insn); |
| 1844 | } |
| 1845 | |
| 1846 | /* Take the last-inserted insns off the stack and return it. */ |
| 1847 | rtx_insn * |
| 1848 | lra_pop_insn (void) |
| 1849 | { |
| 1850 | rtx_insn *insn = lra_constraint_insn_stack.pop (); |
| 1851 | bitmap_clear_bit (map: lra_constraint_insn_stack_bitmap, bitno: INSN_UID (insn)); |
| 1852 | return insn; |
| 1853 | } |
| 1854 | |
| 1855 | /* Return the current size of the insn stack. */ |
| 1856 | unsigned int |
| 1857 | lra_insn_stack_length (void) |
| 1858 | { |
| 1859 | return lra_constraint_insn_stack.length (); |
| 1860 | } |
| 1861 | |
| 1862 | /* Push insns FROM to TO (excluding it) going in reverse order. */ |
| 1863 | static void |
| 1864 | push_insns (rtx_insn *from, rtx_insn *to) |
| 1865 | { |
| 1866 | rtx_insn *insn; |
| 1867 | |
| 1868 | if (from == NULL_RTX) |
| 1869 | return; |
| 1870 | for (insn = from; insn != to; insn = PREV_INSN (insn)) |
| 1871 | if (INSN_P (insn)) |
| 1872 | lra_push_insn (insn); |
| 1873 | } |
| 1874 | |
| 1875 | /* Set up and return sp offset for insns in range [FROM, LAST]. The offset is |
| 1876 | taken from the BB insn before FROM after simulating its effects, |
| 1877 | or zero if there is no such insn. */ |
| 1878 | static poly_int64 |
| 1879 | setup_sp_offset (rtx_insn *from, rtx_insn *last) |
| 1880 | { |
| 1881 | rtx_insn *before = prev_nonnote_nondebug_insn_bb (from); |
| 1882 | poly_int64 offset = 0; |
| 1883 | |
| 1884 | if (before && INSN_P (before)) |
| 1885 | offset = lra_update_sp_offset (PATTERN (insn: before), |
| 1886 | lra_get_insn_recog_data (insn: before)->sp_offset); |
| 1887 | |
| 1888 | for (rtx_insn *insn = from; insn != NEXT_INSN (insn: last); insn = NEXT_INSN (insn)) |
| 1889 | { |
| 1890 | lra_get_insn_recog_data (insn)->sp_offset = offset; |
| 1891 | offset = lra_update_sp_offset (PATTERN (insn), offset); |
| 1892 | } |
| 1893 | return offset; |
| 1894 | } |
| 1895 | |
| 1896 | /* Dump all func insns in a slim form. */ |
| 1897 | void |
| 1898 | lra_dump_insns (FILE *f) |
| 1899 | { |
| 1900 | dump_rtl_slim (f, get_insns (), NULL, -1, 0); |
| 1901 | } |
| 1902 | |
| 1903 | /* Dump all func insns in a slim form with TITLE when the dump file is open and |
| 1904 | lra_verbose >=7. */ |
| 1905 | void |
| 1906 | lra_dump_insns_if_possible (const char *title) |
| 1907 | { |
| 1908 | if (lra_dump_file == NULL || lra_verbose < 7) |
| 1909 | return; |
| 1910 | fprintf (stream: lra_dump_file, format: "%s:" , title); |
| 1911 | lra_dump_insns (f: lra_dump_file); |
| 1912 | } |
| 1913 | |
| 1914 | /* Emit insns BEFORE before INSN and insns AFTER after INSN. Put the |
| 1915 | insns onto the stack. Print about emitting the insns with |
| 1916 | TITLE. */ |
| 1917 | void |
| 1918 | lra_process_new_insns (rtx_insn *insn, rtx_insn *before, rtx_insn *after, |
| 1919 | const char *title) |
| 1920 | { |
| 1921 | if (before == NULL_RTX && after == NULL_RTX) |
| 1922 | return; |
| 1923 | if (lra_dump_file != NULL) |
| 1924 | { |
| 1925 | dump_insn_slim (lra_dump_file, insn); |
| 1926 | if (before != NULL_RTX) |
| 1927 | { |
| 1928 | fprintf (stream: lra_dump_file,format: " %s before:\n" , title); |
| 1929 | dump_rtl_slim (lra_dump_file, before, NULL, -1, 0); |
| 1930 | } |
| 1931 | } |
| 1932 | if (before != NULL_RTX) |
| 1933 | { |
| 1934 | if (cfun->can_throw_non_call_exceptions) |
| 1935 | copy_reg_eh_region_note_forward (insn, before, NULL); |
| 1936 | emit_insn_before (before, insn); |
| 1937 | poly_int64 old_sp_offset = lra_get_insn_recog_data (insn)->sp_offset; |
| 1938 | poly_int64 new_sp_offset = setup_sp_offset (from: before, last: PREV_INSN (insn)); |
| 1939 | if (maybe_ne (a: old_sp_offset, b: new_sp_offset)) |
| 1940 | { |
| 1941 | if (lra_dump_file != NULL) |
| 1942 | { |
| 1943 | fprintf (stream: lra_dump_file, format: " Changing sp offset from " ); |
| 1944 | print_dec (value: old_sp_offset, file: lra_dump_file); |
| 1945 | fprintf (stream: lra_dump_file, format: " to " ); |
| 1946 | print_dec (value: new_sp_offset, file: lra_dump_file); |
| 1947 | fprintf (stream: lra_dump_file, format: " for insn" ); |
| 1948 | dump_rtl_slim (lra_dump_file, insn, NULL, -1, 0); |
| 1949 | } |
| 1950 | lra_get_insn_recog_data (insn)->sp_offset = new_sp_offset; |
| 1951 | eliminate_regs_in_insn (insn, false, false, |
| 1952 | old_sp_offset - new_sp_offset); |
| 1953 | lra_push_insn (insn); |
| 1954 | } |
| 1955 | push_insns (from: PREV_INSN (insn), to: PREV_INSN (insn: before)); |
| 1956 | } |
| 1957 | if (after != NULL_RTX) |
| 1958 | { |
| 1959 | if (cfun->can_throw_non_call_exceptions) |
| 1960 | copy_reg_eh_region_note_forward (insn, after, NULL); |
| 1961 | if (! JUMP_P (insn)) |
| 1962 | { |
| 1963 | rtx_insn *last; |
| 1964 | |
| 1965 | if (lra_dump_file != NULL) |
| 1966 | { |
| 1967 | fprintf (stream: lra_dump_file, format: " %s after:\n" , title); |
| 1968 | dump_rtl_slim (lra_dump_file, after, NULL, -1, 0); |
| 1969 | } |
| 1970 | for (last = after; |
| 1971 | NEXT_INSN (insn: last) != NULL_RTX; |
| 1972 | last = NEXT_INSN (insn: last)) |
| 1973 | ; |
| 1974 | emit_insn_after (after, insn); |
| 1975 | push_insns (from: last, to: insn); |
| 1976 | setup_sp_offset (from: after, last); |
| 1977 | } |
| 1978 | else |
| 1979 | { |
| 1980 | /* Put output reload insns on successor BBs: */ |
| 1981 | edge_iterator ei; |
| 1982 | edge e; |
| 1983 | |
| 1984 | FOR_EACH_EDGE (e, ei, BLOCK_FOR_INSN (insn)->succs) |
| 1985 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
| 1986 | { |
| 1987 | /* We already made the edge no-critical in ira.cc::ira */ |
| 1988 | lra_assert (!EDGE_CRITICAL_P (e)); |
| 1989 | rtx_insn *curr, *tmp = BB_HEAD (e->dest); |
| 1990 | if (LABEL_P (tmp)) |
| 1991 | tmp = NEXT_INSN (insn: tmp); |
| 1992 | if (NOTE_INSN_BASIC_BLOCK_P (tmp)) |
| 1993 | tmp = NEXT_INSN (insn: tmp); |
| 1994 | /* Do not put reload insns if it is the last BB |
| 1995 | without actual insns. */ |
| 1996 | if (tmp == NULL) |
| 1997 | continue; |
| 1998 | start_sequence (); |
| 1999 | for (curr = after; curr != NULL_RTX; curr = NEXT_INSN (insn: curr)) |
| 2000 | emit_insn (copy_insn (PATTERN (insn: curr))); |
| 2001 | rtx_insn *copy = get_insns (), *last = get_last_insn (); |
| 2002 | end_sequence (); |
| 2003 | if (lra_dump_file != NULL) |
| 2004 | { |
| 2005 | fprintf (stream: lra_dump_file, format: " %s after in bb%d:\n" , title, |
| 2006 | e->dest->index); |
| 2007 | dump_rtl_slim (lra_dump_file, copy, NULL, -1, 0); |
| 2008 | } |
| 2009 | /* Use the right emit func for setting up BB_END/BB_HEAD: */ |
| 2010 | if (BB_END (e->dest) == PREV_INSN (insn: tmp)) |
| 2011 | emit_insn_after_noloc (copy, PREV_INSN (insn: tmp), e->dest); |
| 2012 | else |
| 2013 | emit_insn_before_noloc (copy, tmp, e->dest); |
| 2014 | push_insns (from: last, to: PREV_INSN (insn: copy)); |
| 2015 | setup_sp_offset (from: copy, last); |
| 2016 | /* We can ignore BB live info here as it and reg notes |
| 2017 | will be updated before the next assignment |
| 2018 | sub-pass. */ |
| 2019 | } |
| 2020 | } |
| 2021 | } |
| 2022 | if (lra_dump_file != NULL) |
| 2023 | fprintf (stream: lra_dump_file, format: "\n" ); |
| 2024 | if (cfun->can_throw_non_call_exceptions) |
| 2025 | { |
| 2026 | rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX); |
| 2027 | if (note && !insn_could_throw_p (insn)) |
| 2028 | remove_note (insn, note); |
| 2029 | } |
| 2030 | } |
| 2031 | |
| 2032 | |
| 2033 | /* Replace all references to register OLD_REGNO in *LOC with pseudo |
| 2034 | register NEW_REG. Try to simplify subreg of constant if SUBREG_P. |
| 2035 | DEBUG_P is if LOC is within a DEBUG_INSN. Return true if any |
| 2036 | change was made. */ |
| 2037 | bool |
| 2038 | lra_substitute_pseudo (rtx *loc, int old_regno, rtx new_reg, bool subreg_p, |
| 2039 | bool debug_p) |
| 2040 | { |
| 2041 | rtx x = *loc; |
| 2042 | bool result = false; |
| 2043 | enum rtx_code code; |
| 2044 | const char *fmt; |
| 2045 | int i, j; |
| 2046 | |
| 2047 | if (x == NULL_RTX) |
| 2048 | return false; |
| 2049 | |
| 2050 | code = GET_CODE (x); |
| 2051 | if (code == SUBREG && subreg_p) |
| 2052 | { |
| 2053 | rtx subst, inner = SUBREG_REG (x); |
| 2054 | /* Transform subreg of constant while we still have inner mode |
| 2055 | of the subreg. The subreg internal should not be an insn |
| 2056 | operand. */ |
| 2057 | if (REG_P (inner) && (int) REGNO (inner) == old_regno |
| 2058 | && CONSTANT_P (new_reg) |
| 2059 | && (subst = simplify_subreg (GET_MODE (x), op: new_reg, GET_MODE (inner), |
| 2060 | SUBREG_BYTE (x))) != NULL_RTX) |
| 2061 | { |
| 2062 | *loc = subst; |
| 2063 | return true; |
| 2064 | } |
| 2065 | |
| 2066 | } |
| 2067 | else if (code == REG && (int) REGNO (x) == old_regno) |
| 2068 | { |
| 2069 | machine_mode mode = GET_MODE (x); |
| 2070 | machine_mode inner_mode = GET_MODE (new_reg); |
| 2071 | |
| 2072 | if (mode != inner_mode |
| 2073 | && ! (CONST_SCALAR_INT_P (new_reg) && SCALAR_INT_MODE_P (mode))) |
| 2074 | { |
| 2075 | poly_uint64 offset = 0; |
| 2076 | if (partial_subreg_p (outermode: mode, innermode: inner_mode) |
| 2077 | && SCALAR_INT_MODE_P (inner_mode)) |
| 2078 | offset = subreg_lowpart_offset (outermode: mode, innermode: inner_mode); |
| 2079 | if (debug_p) |
| 2080 | new_reg = gen_rtx_raw_SUBREG (mode, new_reg, offset); |
| 2081 | else |
| 2082 | new_reg = gen_rtx_SUBREG (mode, new_reg, offset); |
| 2083 | } |
| 2084 | *loc = new_reg; |
| 2085 | return true; |
| 2086 | } |
| 2087 | |
| 2088 | /* Scan all the operand sub-expressions. */ |
| 2089 | fmt = GET_RTX_FORMAT (code); |
| 2090 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
| 2091 | { |
| 2092 | if (fmt[i] == 'e') |
| 2093 | { |
| 2094 | if (debug_p |
| 2095 | && i == 0 |
| 2096 | && (code == SUBREG |
| 2097 | || code == ZERO_EXTEND |
| 2098 | || code == SIGN_EXTEND |
| 2099 | || code == FLOAT |
| 2100 | || code == UNSIGNED_FLOAT)) |
| 2101 | { |
| 2102 | rtx y = XEXP (x, 0); |
| 2103 | if (lra_substitute_pseudo (loc: &y, old_regno, |
| 2104 | new_reg, subreg_p, debug_p)) |
| 2105 | { |
| 2106 | result = true; |
| 2107 | if (CONST_SCALAR_INT_P (y)) |
| 2108 | { |
| 2109 | if (code == SUBREG) |
| 2110 | y = simplify_subreg (GET_MODE (x), op: y, |
| 2111 | GET_MODE (SUBREG_REG (x)), |
| 2112 | SUBREG_BYTE (x)); |
| 2113 | else |
| 2114 | y = simplify_unary_operation (code, GET_MODE (x), op: y, |
| 2115 | GET_MODE (XEXP (x, 0))); |
| 2116 | if (y) |
| 2117 | *loc = y; |
| 2118 | else |
| 2119 | *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx); |
| 2120 | } |
| 2121 | else |
| 2122 | XEXP (x, 0) = y; |
| 2123 | } |
| 2124 | } |
| 2125 | else if (lra_substitute_pseudo (loc: &XEXP (x, i), old_regno, |
| 2126 | new_reg, subreg_p, debug_p)) |
| 2127 | result = true; |
| 2128 | } |
| 2129 | else if (fmt[i] == 'E') |
| 2130 | { |
| 2131 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) |
| 2132 | if (lra_substitute_pseudo (loc: &XVECEXP (x, i, j), old_regno, |
| 2133 | new_reg, subreg_p, debug_p)) |
| 2134 | result = true; |
| 2135 | } |
| 2136 | } |
| 2137 | return result; |
| 2138 | } |
| 2139 | |
| 2140 | /* Call lra_substitute_pseudo within an insn. Try to simplify subreg |
| 2141 | of constant if SUBREG_P. This won't update the insn ptr, just the |
| 2142 | contents of the insn. */ |
| 2143 | bool |
| 2144 | lra_substitute_pseudo_within_insn (rtx_insn *insn, int old_regno, |
| 2145 | rtx new_reg, bool subreg_p) |
| 2146 | { |
| 2147 | rtx loc = insn; |
| 2148 | return lra_substitute_pseudo (loc: &loc, old_regno, new_reg, subreg_p, |
| 2149 | DEBUG_INSN_P (insn)); |
| 2150 | } |
| 2151 | |
| 2152 | |
| 2153 | |
| 2154 | /* Return new register of the same mode as ORIGINAL of class ALL_REGS. |
| 2155 | Used in ira_remove_scratches. */ |
| 2156 | static rtx |
| 2157 | get_scratch_reg (rtx original) |
| 2158 | { |
| 2159 | return lra_create_new_reg (GET_MODE (original), original, rclass: ALL_REGS, |
| 2160 | NULL, NULL); |
| 2161 | } |
| 2162 | |
| 2163 | /* Remove all insn scratches in INSN. */ |
| 2164 | static void |
| 2165 | remove_insn_scratches (rtx_insn *insn) |
| 2166 | { |
| 2167 | if (ira_remove_insn_scratches (insn, all_p: true, dump_file: lra_dump_file, get_reg: get_scratch_reg)) |
| 2168 | df_insn_rescan (insn); |
| 2169 | } |
| 2170 | |
| 2171 | /* Remove all insn scratches in the current function. */ |
| 2172 | static void |
| 2173 | remove_scratches (void) |
| 2174 | { |
| 2175 | basic_block bb; |
| 2176 | rtx_insn *insn; |
| 2177 | |
| 2178 | FOR_EACH_BB_FN (bb, cfun) |
| 2179 | FOR_BB_INSNS (bb, insn) |
| 2180 | if (INSN_P (insn)) |
| 2181 | remove_insn_scratches (insn); |
| 2182 | } |
| 2183 | |
| 2184 | /* Function checks RTL for correctness. If FINAL_P is true, it is |
| 2185 | done at the end of LRA and the check is more rigorous. */ |
| 2186 | static void |
| 2187 | check_rtl (bool final_p) |
| 2188 | { |
| 2189 | basic_block bb; |
| 2190 | rtx_insn *insn; |
| 2191 | |
| 2192 | lra_assert (! final_p || reload_completed); |
| 2193 | FOR_EACH_BB_FN (bb, cfun) |
| 2194 | FOR_BB_INSNS (bb, insn) |
| 2195 | if (NONDEBUG_INSN_P (insn) |
| 2196 | && GET_CODE (PATTERN (insn)) != USE |
| 2197 | && GET_CODE (PATTERN (insn)) != CLOBBER |
| 2198 | && GET_CODE (PATTERN (insn)) != ASM_INPUT) |
| 2199 | { |
| 2200 | if (final_p) |
| 2201 | { |
| 2202 | extract_constrain_insn (insn); |
| 2203 | continue; |
| 2204 | } |
| 2205 | /* LRA code is based on assumption that all addresses can be |
| 2206 | correctly decomposed. LRA can generate reloads for |
| 2207 | decomposable addresses. The decomposition code checks the |
| 2208 | correctness of the addresses. So we don't need to check |
| 2209 | the addresses here. Don't call insn_invalid_p here, it can |
| 2210 | change the code at this stage. */ |
| 2211 | if (recog_memoized (insn) < 0 && asm_noperands (PATTERN (insn)) < 0) |
| 2212 | fatal_insn_not_found (insn); |
| 2213 | } |
| 2214 | } |
| 2215 | |
| 2216 | /* Determine if the current function has an exception receiver block |
| 2217 | that reaches the exit block via non-exceptional edges */ |
| 2218 | static bool |
| 2219 | has_nonexceptional_receiver (void) |
| 2220 | { |
| 2221 | edge e; |
| 2222 | edge_iterator ei; |
| 2223 | basic_block *tos, *worklist, bb; |
| 2224 | |
| 2225 | /* If we're not optimizing, then just err on the safe side. */ |
| 2226 | if (!optimize) |
| 2227 | return true; |
| 2228 | |
| 2229 | /* First determine which blocks can reach exit via normal paths. */ |
| 2230 | tos = worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) + 1); |
| 2231 | |
| 2232 | FOR_EACH_BB_FN (bb, cfun) |
| 2233 | bb->flags &= ~BB_REACHABLE; |
| 2234 | |
| 2235 | /* Place the exit block on our worklist. */ |
| 2236 | EXIT_BLOCK_PTR_FOR_FN (cfun)->flags |= BB_REACHABLE; |
| 2237 | *tos++ = EXIT_BLOCK_PTR_FOR_FN (cfun); |
| 2238 | |
| 2239 | /* Iterate: find everything reachable from what we've already seen. */ |
| 2240 | while (tos != worklist) |
| 2241 | { |
| 2242 | bb = *--tos; |
| 2243 | |
| 2244 | FOR_EACH_EDGE (e, ei, bb->preds) |
| 2245 | if (e->flags & EDGE_ABNORMAL) |
| 2246 | { |
| 2247 | free (ptr: worklist); |
| 2248 | return true; |
| 2249 | } |
| 2250 | else |
| 2251 | { |
| 2252 | basic_block src = e->src; |
| 2253 | |
| 2254 | if (!(src->flags & BB_REACHABLE)) |
| 2255 | { |
| 2256 | src->flags |= BB_REACHABLE; |
| 2257 | *tos++ = src; |
| 2258 | } |
| 2259 | } |
| 2260 | } |
| 2261 | free (ptr: worklist); |
| 2262 | /* No exceptional block reached exit unexceptionally. */ |
| 2263 | return false; |
| 2264 | } |
| 2265 | |
| 2266 | /* Remove all REG_DEAD and REG_UNUSED notes and regenerate REG_INC. |
| 2267 | We change pseudos by hard registers without notification of DF and |
| 2268 | that can make the notes obsolete. DF-infrastructure does not deal |
| 2269 | with REG_INC notes -- so we should regenerate them here. */ |
| 2270 | static void |
| 2271 | update_inc_notes (void) |
| 2272 | { |
| 2273 | rtx *pnote; |
| 2274 | basic_block bb; |
| 2275 | rtx_insn *insn; |
| 2276 | |
| 2277 | FOR_EACH_BB_FN (bb, cfun) |
| 2278 | FOR_BB_INSNS (bb, insn) |
| 2279 | if (NONDEBUG_INSN_P (insn)) |
| 2280 | { |
| 2281 | pnote = ®_NOTES (insn); |
| 2282 | while (*pnote != 0) |
| 2283 | { |
| 2284 | if (REG_NOTE_KIND (*pnote) == REG_DEAD |
| 2285 | || REG_NOTE_KIND (*pnote) == REG_UNUSED |
| 2286 | || REG_NOTE_KIND (*pnote) == REG_INC) |
| 2287 | *pnote = XEXP (*pnote, 1); |
| 2288 | else |
| 2289 | pnote = &XEXP (*pnote, 1); |
| 2290 | } |
| 2291 | |
| 2292 | if (AUTO_INC_DEC) |
| 2293 | add_auto_inc_notes (insn, PATTERN (insn)); |
| 2294 | } |
| 2295 | } |
| 2296 | |
| 2297 | /* Set to true while in LRA. */ |
| 2298 | bool lra_in_progress = false; |
| 2299 | |
| 2300 | /* Start of pseudo regnos before the LRA. */ |
| 2301 | int lra_new_regno_start; |
| 2302 | |
| 2303 | /* Start of reload pseudo regnos before the new spill pass. */ |
| 2304 | int lra_constraint_new_regno_start; |
| 2305 | |
| 2306 | /* Avoid spilling pseudos with regno more than the following value if |
| 2307 | it is possible. */ |
| 2308 | int lra_bad_spill_regno_start; |
| 2309 | |
| 2310 | /* A pseudo of Pmode. */ |
| 2311 | rtx lra_pmode_pseudo; |
| 2312 | |
| 2313 | /* Inheritance pseudo regnos before the new spill pass. */ |
| 2314 | bitmap_head lra_inheritance_pseudos; |
| 2315 | |
| 2316 | /* Split regnos before the new spill pass. */ |
| 2317 | bitmap_head lra_split_regs; |
| 2318 | |
| 2319 | /* Reload pseudo regnos before the new assignment pass which still can |
| 2320 | be spilled after the assignment pass as memory is also accepted in |
| 2321 | insns for the reload pseudos. */ |
| 2322 | bitmap_head lra_optional_reload_pseudos; |
| 2323 | |
| 2324 | /* Pseudo regnos used for subreg reloads before the new assignment |
| 2325 | pass. Such pseudos still can be spilled after the assignment |
| 2326 | pass. */ |
| 2327 | bitmap_head lra_subreg_reload_pseudos; |
| 2328 | |
| 2329 | /* File used for output of LRA debug information. */ |
| 2330 | FILE *lra_dump_file; |
| 2331 | |
| 2332 | /* How verbose should be the debug information. */ |
| 2333 | int lra_verbose; |
| 2334 | |
| 2335 | /* True if we split hard reg after the last constraint sub-pass. */ |
| 2336 | bool lra_hard_reg_split_p; |
| 2337 | |
| 2338 | /* True if we found an asm error. */ |
| 2339 | bool lra_asm_error_p; |
| 2340 | |
| 2341 | /* True if we should try spill into registers of different classes |
| 2342 | instead of memory. */ |
| 2343 | bool lra_reg_spill_p; |
| 2344 | |
| 2345 | /* Set up value LRA_REG_SPILL_P. */ |
| 2346 | static void |
| 2347 | setup_reg_spill_flag (void) |
| 2348 | { |
| 2349 | int cl, mode; |
| 2350 | |
| 2351 | if (targetm.spill_class != NULL) |
| 2352 | for (cl = 0; cl < (int) LIM_REG_CLASSES; cl++) |
| 2353 | for (mode = 0; mode < MAX_MACHINE_MODE; mode++) |
| 2354 | if (targetm.spill_class ((enum reg_class) cl, |
| 2355 | (machine_mode) mode) != NO_REGS) |
| 2356 | { |
| 2357 | lra_reg_spill_p = true; |
| 2358 | return; |
| 2359 | } |
| 2360 | lra_reg_spill_p = false; |
| 2361 | } |
| 2362 | |
| 2363 | /* True if the current function is too big to use regular algorithms |
| 2364 | in LRA. In other words, we should use simpler and faster algorithms |
| 2365 | in LRA. It also means we should not worry about generation code |
| 2366 | for caller saves. The value is set up in IRA. */ |
| 2367 | bool lra_simple_p; |
| 2368 | |
| 2369 | /* Major LRA entry function. F is a file should be used to dump LRA |
| 2370 | debug info with given verbosity. */ |
| 2371 | void |
| 2372 | lra (FILE *f, int verbose) |
| 2373 | { |
| 2374 | int i; |
| 2375 | bool live_p, inserted_p; |
| 2376 | |
| 2377 | lra_dump_file = f; |
| 2378 | lra_verbose = verbose; |
| 2379 | lra_asm_error_p = false; |
| 2380 | lra_pmode_pseudo = gen_reg_rtx (Pmode); |
| 2381 | |
| 2382 | timevar_push (tv: TV_LRA); |
| 2383 | |
| 2384 | /* Make sure that the last insn is a note. Some subsequent passes |
| 2385 | need it. */ |
| 2386 | emit_note (NOTE_INSN_DELETED); |
| 2387 | |
| 2388 | lra_no_alloc_regs = ira_no_alloc_regs; |
| 2389 | |
| 2390 | init_reg_info (); |
| 2391 | expand_reg_info (); |
| 2392 | |
| 2393 | init_insn_recog_data (); |
| 2394 | |
| 2395 | /* Some quick check on RTL generated by previous passes. */ |
| 2396 | if (flag_checking) |
| 2397 | check_rtl (final_p: false); |
| 2398 | |
| 2399 | lra_in_progress = true; |
| 2400 | |
| 2401 | lra_live_range_iter = lra_coalesce_iter = lra_constraint_iter = 0; |
| 2402 | lra_assignment_iter = lra_assignment_iter_after_spill = 0; |
| 2403 | lra_inheritance_iter = lra_undo_inheritance_iter = 0; |
| 2404 | lra_rematerialization_iter = 0; |
| 2405 | |
| 2406 | setup_reg_spill_flag (); |
| 2407 | |
| 2408 | /* Function remove_scratches can creates new pseudos for clobbers -- |
| 2409 | so set up lra_constraint_new_regno_start before its call to |
| 2410 | permit changing reg classes for pseudos created by this |
| 2411 | simplification. */ |
| 2412 | lra_constraint_new_regno_start = lra_new_regno_start = max_reg_num (); |
| 2413 | lra_bad_spill_regno_start = INT_MAX; |
| 2414 | remove_scratches (); |
| 2415 | |
| 2416 | /* A function that has a non-local label that can reach the exit |
| 2417 | block via non-exceptional paths must save all call-saved |
| 2418 | registers. */ |
| 2419 | if (cfun->has_nonlocal_label && has_nonexceptional_receiver ()) |
| 2420 | crtl->saves_all_registers = 1; |
| 2421 | |
| 2422 | if (crtl->saves_all_registers) |
| 2423 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
| 2424 | if (!crtl->abi->clobbers_full_reg_p (regno: i) |
| 2425 | && !fixed_regs[i] |
| 2426 | && !LOCAL_REGNO (i)) |
| 2427 | df_set_regs_ever_live (i, true); |
| 2428 | |
| 2429 | /* We don't DF from now and avoid its using because it is to |
| 2430 | expensive when a lot of RTL changes are made. */ |
| 2431 | df_set_flags (DF_NO_INSN_RESCAN); |
| 2432 | lra_constraint_insn_stack.create (nelems: get_max_uid ()); |
| 2433 | lra_constraint_insn_stack_bitmap = sbitmap_alloc (get_max_uid ()); |
| 2434 | bitmap_clear (lra_constraint_insn_stack_bitmap); |
| 2435 | lra_live_ranges_init (); |
| 2436 | lra_constraints_init (); |
| 2437 | lra_curr_reload_num = 0; |
| 2438 | push_insns (from: get_last_insn (), NULL); |
| 2439 | /* It is needed for the 1st coalescing. */ |
| 2440 | bitmap_initialize (head: &lra_inheritance_pseudos, obstack: ®_obstack); |
| 2441 | bitmap_initialize (head: &lra_split_regs, obstack: ®_obstack); |
| 2442 | bitmap_initialize (head: &lra_optional_reload_pseudos, obstack: ®_obstack); |
| 2443 | bitmap_initialize (head: &lra_subreg_reload_pseudos, obstack: ®_obstack); |
| 2444 | live_p = false; |
| 2445 | if (maybe_ne (a: get_frame_size (), b: 0) && crtl->stack_alignment_needed) |
| 2446 | /* If we have a stack frame, we must align it now. The stack size |
| 2447 | may be a part of the offset computation for register |
| 2448 | elimination. */ |
| 2449 | assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed); |
| 2450 | lra_init_equiv (); |
| 2451 | for (;;) |
| 2452 | { |
| 2453 | for (;;) |
| 2454 | { |
| 2455 | bool reloads_p = lra_constraints (lra_constraint_iter == 0); |
| 2456 | /* Constraint transformations may result in that eliminable |
| 2457 | hard regs become uneliminable and pseudos which use them |
| 2458 | should be spilled. It is better to do it before pseudo |
| 2459 | assignments. |
| 2460 | |
| 2461 | For example, rs6000 can make |
| 2462 | RS6000_PIC_OFFSET_TABLE_REGNUM uneliminable if we started |
| 2463 | to use a constant pool. */ |
| 2464 | lra_eliminate (false, false); |
| 2465 | /* We should try to assign hard registers to scratches even |
| 2466 | if there were no RTL transformations in lra_constraints. |
| 2467 | Also we should check IRA assignments on the first |
| 2468 | iteration as they can be wrong because of early clobbers |
| 2469 | operands which are ignored in IRA. */ |
| 2470 | if (! reloads_p && lra_constraint_iter > 1) |
| 2471 | { |
| 2472 | /* Stack is not empty here only when there are changes |
| 2473 | during the elimination sub-pass. */ |
| 2474 | if (bitmap_empty_p (lra_constraint_insn_stack_bitmap)) |
| 2475 | break; |
| 2476 | else |
| 2477 | /* If there are no reloads but changing due |
| 2478 | elimination, restart the constraint sub-pass |
| 2479 | first. */ |
| 2480 | continue; |
| 2481 | } |
| 2482 | /* Do inheritance only for regular algorithms. */ |
| 2483 | if (! lra_simple_p) |
| 2484 | lra_inheritance (); |
| 2485 | if (live_p) |
| 2486 | lra_clear_live_ranges (); |
| 2487 | bool fails_p; |
| 2488 | lra_hard_reg_split_p = false; |
| 2489 | int split_fails_num = 0; |
| 2490 | do |
| 2491 | { |
| 2492 | /* We need live ranges for lra_assign -- so build them. |
| 2493 | But don't remove dead insns or change global live |
| 2494 | info as we can undo inheritance transformations after |
| 2495 | inheritance pseudo assigning. */ |
| 2496 | lra_create_live_ranges (true, !lra_simple_p); |
| 2497 | live_p = true; |
| 2498 | /* If we don't spill non-reload and non-inheritance |
| 2499 | pseudos, there is no sense to run memory-memory move |
| 2500 | coalescing. If inheritance pseudos were spilled, the |
| 2501 | memory-memory moves involving them will be removed by |
| 2502 | pass undoing inheritance. */ |
| 2503 | if (lra_simple_p || lra_hard_reg_split_p) |
| 2504 | lra_assign (fails_p); |
| 2505 | else |
| 2506 | { |
| 2507 | bool spill_p = !lra_assign (fails_p); |
| 2508 | |
| 2509 | if (lra_undo_inheritance ()) |
| 2510 | live_p = false; |
| 2511 | if (spill_p && ! fails_p) |
| 2512 | { |
| 2513 | if (! live_p) |
| 2514 | { |
| 2515 | lra_create_live_ranges (true, true); |
| 2516 | live_p = true; |
| 2517 | } |
| 2518 | if (lra_coalesce ()) |
| 2519 | live_p = false; |
| 2520 | } |
| 2521 | if (! live_p) |
| 2522 | lra_clear_live_ranges (); |
| 2523 | } |
| 2524 | if (fails_p) |
| 2525 | { |
| 2526 | /* It is a very rare case. It is the last hope to |
| 2527 | split a hard regno live range for a reload |
| 2528 | pseudo. */ |
| 2529 | if (live_p) |
| 2530 | lra_clear_live_ranges (); |
| 2531 | live_p = false; |
| 2532 | /* See a comment for LRA_MAX_FAILED_SPLITS definition. */ |
| 2533 | bool last_failed_split_p |
| 2534 | = split_fails_num > LRA_MAX_FAILED_SPLITS; |
| 2535 | if (! lra_split_hard_reg_for (fail_p: last_failed_split_p)) |
| 2536 | { |
| 2537 | if (last_failed_split_p) |
| 2538 | break; |
| 2539 | split_fails_num++; |
| 2540 | } |
| 2541 | lra_hard_reg_split_p = true; |
| 2542 | } |
| 2543 | } |
| 2544 | while (fails_p && !lra_asm_error_p); |
| 2545 | if (! live_p) { |
| 2546 | /* We need the correct reg notes for work of constraint sub-pass. */ |
| 2547 | lra_create_live_ranges (true, true); |
| 2548 | live_p = true; |
| 2549 | } |
| 2550 | } |
| 2551 | /* Don't clear optional reloads bitmap until all constraints are |
| 2552 | satisfied as we need to differ them from regular reloads. */ |
| 2553 | bitmap_clear (&lra_optional_reload_pseudos); |
| 2554 | bitmap_clear (&lra_subreg_reload_pseudos); |
| 2555 | bitmap_clear (&lra_inheritance_pseudos); |
| 2556 | bitmap_clear (&lra_split_regs); |
| 2557 | if (! live_p) |
| 2558 | { |
| 2559 | /* We need full live info for spilling pseudos into |
| 2560 | registers instead of memory. */ |
| 2561 | lra_create_live_ranges (lra_reg_spill_p, true); |
| 2562 | live_p = true; |
| 2563 | } |
| 2564 | /* We should check necessity for spilling here as the above live |
| 2565 | range pass can remove spilled pseudos. */ |
| 2566 | if (! lra_need_for_spills_p ()) |
| 2567 | break; |
| 2568 | /* Now we know what pseudos should be spilled. Try to |
| 2569 | rematerialize them first. */ |
| 2570 | if (lra_remat ()) |
| 2571 | { |
| 2572 | /* We need full live info -- see the comment above. We also might |
| 2573 | need live info if we have a pseudo assigned to hard frame pointer |
| 2574 | reg and will need FP for usual purposes. */ |
| 2575 | lra_create_live_ranges (lra_reg_spill_p || lra_fp_pseudo_p (), |
| 2576 | true); |
| 2577 | live_p = true; |
| 2578 | if (! lra_need_for_spills_p ()) |
| 2579 | { |
| 2580 | if (lra_need_for_scratch_reg_p ()) |
| 2581 | continue; |
| 2582 | break; |
| 2583 | } |
| 2584 | } |
| 2585 | lra_spill (); |
| 2586 | /* Assignment of stack slots changes elimination offsets for |
| 2587 | some eliminations. So update the offsets here. */ |
| 2588 | lra_eliminate (false, false); |
| 2589 | lra_constraint_new_regno_start = max_reg_num (); |
| 2590 | if (lra_bad_spill_regno_start == INT_MAX |
| 2591 | && lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES |
| 2592 | && lra_rematerialization_iter > LRA_MAX_REMATERIALIZATION_PASSES) |
| 2593 | /* After switching off inheritance and rematerialization |
| 2594 | passes, avoid spilling reload pseudos will be created to |
| 2595 | prevent LRA cycling in some complicated cases. */ |
| 2596 | lra_bad_spill_regno_start = lra_constraint_new_regno_start; |
| 2597 | lra_assignment_iter_after_spill = 0; |
| 2598 | } |
| 2599 | ira_restore_scratches (dump_file: lra_dump_file); |
| 2600 | lra_eliminate (true, false); |
| 2601 | lra_final_code_change (); |
| 2602 | lra_in_progress = false; |
| 2603 | if (live_p) |
| 2604 | lra_clear_live_ranges (); |
| 2605 | lra_live_ranges_finish (); |
| 2606 | lra_constraints_finish (); |
| 2607 | finish_reg_info (); |
| 2608 | sbitmap_free (map: lra_constraint_insn_stack_bitmap); |
| 2609 | lra_constraint_insn_stack.release (); |
| 2610 | finish_insn_recog_data (); |
| 2611 | regstat_free_n_sets_and_refs (); |
| 2612 | regstat_free_ri (); |
| 2613 | reload_completed = 1; |
| 2614 | update_inc_notes (); |
| 2615 | |
| 2616 | inserted_p = fixup_abnormal_edges (); |
| 2617 | |
| 2618 | /* Split basic blocks if we've possibly turned single trapping insn |
| 2619 | into multiple ones or otherwise the backend requested to do so. */ |
| 2620 | if (cfun->can_throw_non_call_exceptions |
| 2621 | || cfun->split_basic_blocks_after_reload) |
| 2622 | { |
| 2623 | auto_sbitmap blocks (last_basic_block_for_fn (cfun)); |
| 2624 | bitmap_ones (blocks); |
| 2625 | find_many_sub_basic_blocks (blocks); |
| 2626 | } |
| 2627 | |
| 2628 | if (inserted_p) |
| 2629 | commit_edge_insertions (); |
| 2630 | |
| 2631 | /* Subsequent passes expect that rtl is unshared, so unshare everything |
| 2632 | here. */ |
| 2633 | unshare_all_rtl_again (get_insns ()); |
| 2634 | |
| 2635 | if (flag_checking) |
| 2636 | check_rtl (final_p: true); |
| 2637 | |
| 2638 | timevar_pop (tv: TV_LRA); |
| 2639 | } |
| 2640 | |
| 2641 | /* Called once per compiler to initialize LRA data once. */ |
| 2642 | void |
| 2643 | lra_init_once (void) |
| 2644 | { |
| 2645 | init_insn_code_data_once (); |
| 2646 | } |
| 2647 | |
| 2648 | /* Called once per compiler to finish LRA data which are initialize |
| 2649 | once. */ |
| 2650 | void |
| 2651 | lra_finish_once (void) |
| 2652 | { |
| 2653 | finish_insn_code_data_once (); |
| 2654 | } |
| 2655 | |