| 1 | /* RTL-based forward propagation pass for GNU compiler. |
|---|---|
| 2 | Copyright (C) 2005-2025 Free Software Foundation, Inc. |
| 3 | Contributed by Paolo Bonzini and Steven Bosscher. |
| 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 | #define INCLUDE_ALGORITHM |
| 22 | #define INCLUDE_FUNCTIONAL |
| 23 | #define INCLUDE_ARRAY |
| 24 | #include "config.h" |
| 25 | #include "system.h" |
| 26 | #include "coretypes.h" |
| 27 | #include "backend.h" |
| 28 | #include "rtl.h" |
| 29 | #include "rtlanal.h" |
| 30 | #include "df.h" |
| 31 | #include "rtl-ssa.h" |
| 32 | |
| 33 | #include "predict.h" |
| 34 | #include "cfgrtl.h" |
| 35 | #include "cfgcleanup.h" |
| 36 | #include "cfgloop.h" |
| 37 | #include "tree-pass.h" |
| 38 | #include "rtl-iter.h" |
| 39 | #include "target.h" |
| 40 | |
| 41 | /* This pass does simple forward propagation and simplification when an |
| 42 | operand of an insn can only come from a single def. This pass uses |
| 43 | RTL SSA, so it is global. However, we only do limited analysis of |
| 44 | available expressions. |
| 45 | |
| 46 | 1) The pass tries to propagate the source of the def into the use, |
| 47 | and checks if the result is independent of the substituted value. |
| 48 | For example, the high word of a (zero_extend:DI (reg:SI M)) is always |
| 49 | zero, independent of the source register. |
| 50 | |
| 51 | In particular, we propagate constants into the use site. Sometimes |
| 52 | RTL expansion did not put the constant in the same insn on purpose, |
| 53 | to satisfy a predicate, and the result will fail to be recognized; |
| 54 | but this happens rarely and in this case we can still create a |
| 55 | REG_EQUAL note. For multi-word operations, this |
| 56 | |
| 57 | (set (subreg:SI (reg:DI 120) 0) (const_int 0)) |
| 58 | (set (subreg:SI (reg:DI 120) 4) (const_int -1)) |
| 59 | (set (subreg:SI (reg:DI 122) 0) |
| 60 | (ior:SI (subreg:SI (reg:DI 119) 0) (subreg:SI (reg:DI 120) 0))) |
| 61 | (set (subreg:SI (reg:DI 122) 4) |
| 62 | (ior:SI (subreg:SI (reg:DI 119) 4) (subreg:SI (reg:DI 120) 4))) |
| 63 | |
| 64 | can be simplified to the much simpler |
| 65 | |
| 66 | (set (subreg:SI (reg:DI 122) 0) (subreg:SI (reg:DI 119))) |
| 67 | (set (subreg:SI (reg:DI 122) 4) (const_int -1)) |
| 68 | |
| 69 | This particular propagation is also effective at putting together |
| 70 | complex addressing modes. We are more aggressive inside MEMs, in |
| 71 | that all definitions are propagated if the use is in a MEM; if the |
| 72 | result is a valid memory address we check address_cost to decide |
| 73 | whether the substitution is worthwhile. |
| 74 | |
| 75 | 2) The pass propagates register copies. This is not as effective as |
| 76 | the copy propagation done by CSE's canon_reg, which works by walking |
| 77 | the instruction chain, it can help the other transformations. |
| 78 | |
| 79 | We should consider removing this optimization, and instead reorder the |
| 80 | RTL passes, because GCSE does this transformation too. With some luck, |
| 81 | the CSE pass at the end of rest_of_handle_gcse could also go away. |
| 82 | |
| 83 | 3) The pass looks for paradoxical subregs that are actually unnecessary. |
| 84 | Things like this: |
| 85 | |
| 86 | (set (reg:QI 120) (subreg:QI (reg:SI 118) 0)) |
| 87 | (set (reg:QI 121) (subreg:QI (reg:SI 119) 0)) |
| 88 | (set (reg:SI 122) (plus:SI (subreg:SI (reg:QI 120) 0) |
| 89 | (subreg:SI (reg:QI 121) 0))) |
| 90 | |
| 91 | are very common on machines that can only do word-sized operations. |
| 92 | For each use of a paradoxical subreg (subreg:WIDER (reg:NARROW N) 0), |
| 93 | if it has a single def and it is (subreg:NARROW (reg:WIDE M) 0), |
| 94 | we can replace the paradoxical subreg with simply (reg:WIDE M). The |
| 95 | above will simplify this to |
| 96 | |
| 97 | (set (reg:QI 120) (subreg:QI (reg:SI 118) 0)) |
| 98 | (set (reg:QI 121) (subreg:QI (reg:SI 119) 0)) |
| 99 | (set (reg:SI 122) (plus:SI (reg:SI 118) (reg:SI 119))) |
| 100 | |
| 101 | where the first two insns are now dead. */ |
| 102 | |
| 103 | using namespace rtl_ssa; |
| 104 | |
| 105 | static int num_changes; |
| 106 | |
| 107 | /* Do not try to replace constant addresses or addresses of local and |
| 108 | argument slots. These MEM expressions are made only once and inserted |
| 109 | in many instructions, as well as being used to control symbol table |
| 110 | output. It is not safe to clobber them. |
| 111 | |
| 112 | There are some uncommon cases where the address is already in a register |
| 113 | for some reason, but we cannot take advantage of that because we have |
| 114 | no easy way to unshare the MEM. In addition, looking up all stack |
| 115 | addresses is costly. */ |
| 116 | |
| 117 | static bool |
| 118 | can_simplify_addr (rtx addr) |
| 119 | { |
| 120 | rtx reg; |
| 121 | |
| 122 | if (CONSTANT_ADDRESS_P (addr)) |
| 123 | return false; |
| 124 | |
| 125 | if (GET_CODE (addr) == PLUS) |
| 126 | reg = XEXP (addr, 0); |
| 127 | else |
| 128 | reg = addr; |
| 129 | |
| 130 | return (!REG_P (reg) |
| 131 | || (REGNO (reg) != FRAME_POINTER_REGNUM |
| 132 | && REGNO (reg) != HARD_FRAME_POINTER_REGNUM |
| 133 | && REGNO (reg) != ARG_POINTER_REGNUM)); |
| 134 | } |
| 135 | |
| 136 | /* MEM is the result of an address simplification, and temporarily |
| 137 | undoing changes OLD_NUM_CHANGES onwards restores the original address. |
| 138 | Return whether it is good to use the new address instead of the |
| 139 | old one. INSN is the containing instruction. */ |
| 140 | |
| 141 | static bool |
| 142 | should_replace_address (int old_num_changes, rtx mem, rtx_insn *insn) |
| 143 | { |
| 144 | int gain; |
| 145 | |
| 146 | /* Prefer the new address if it is less expensive. */ |
| 147 | bool speed = optimize_bb_for_speed_p (BLOCK_FOR_INSN (insn)); |
| 148 | { |
| 149 | undo_recog_changes undo (old_num_changes); |
| 150 | gain = address_cost (XEXP (mem, 0), GET_MODE (mem), |
| 151 | MEM_ADDR_SPACE (mem), speed); |
| 152 | } |
| 153 | gain -= address_cost (XEXP (mem, 0), GET_MODE (mem), |
| 154 | MEM_ADDR_SPACE (mem), speed); |
| 155 | |
| 156 | /* If the addresses have equivalent cost, prefer the new address |
| 157 | if it has the highest `set_src_cost'. That has the potential of |
| 158 | eliminating the most insns without additional costs, and it |
| 159 | is the same that cse.cc used to do. */ |
| 160 | if (gain == 0) |
| 161 | { |
| 162 | gain = set_src_cost (XEXP (mem, 0), VOIDmode, speed_p: speed); |
| 163 | undo_recog_changes undo (old_num_changes); |
| 164 | gain -= set_src_cost (XEXP (mem, 0), VOIDmode, speed_p: speed); |
| 165 | } |
| 166 | |
| 167 | return (gain > 0); |
| 168 | } |
| 169 | |
| 170 | |
| 171 | namespace |
| 172 | { |
| 173 | class fwprop_propagation : public insn_propagation |
| 174 | { |
| 175 | public: |
| 176 | static const uint16_t CHANGED_MEM = FIRST_SPARE_RESULT; |
| 177 | static const uint16_t CONSTANT = FIRST_SPARE_RESULT << 1; |
| 178 | static const uint16_t PROFITABLE = FIRST_SPARE_RESULT << 2; |
| 179 | |
| 180 | fwprop_propagation (insn_info *, set_info *, rtx, rtx); |
| 181 | |
| 182 | bool changed_mem_p () const { return result_flags & CHANGED_MEM; } |
| 183 | bool folded_to_constants_p () const; |
| 184 | bool likely_profitable_p () const; |
| 185 | |
| 186 | bool check_mem (int, rtx) final override; |
| 187 | void note_simplification (int, uint16_t, rtx, rtx) final override; |
| 188 | uint16_t classify_result (rtx, rtx); |
| 189 | |
| 190 | private: |
| 191 | const bool single_use_p; |
| 192 | const bool single_ebb_p; |
| 193 | }; |
| 194 | } |
| 195 | |
| 196 | /* Prepare to replace FROM with TO in USE_INSN. */ |
| 197 | |
| 198 | fwprop_propagation::fwprop_propagation (insn_info *use_insn, |
| 199 | set_info *def, rtx from, rtx to) |
| 200 | : insn_propagation (use_insn->rtl (), from, to), |
| 201 | single_use_p (def->single_nondebug_use ()), |
| 202 | single_ebb_p (use_insn->ebb () == def->ebb ()) |
| 203 | { |
| 204 | should_check_mems = true; |
| 205 | should_note_simplifications = true; |
| 206 | } |
| 207 | |
| 208 | /* MEM is the result of an address simplification, and temporarily |
| 209 | undoing changes OLD_NUM_CHANGES onwards restores the original address. |
| 210 | Return true if the propagation should continue, false if it has failed. */ |
| 211 | |
| 212 | bool |
| 213 | fwprop_propagation::check_mem (int old_num_changes, rtx mem) |
| 214 | { |
| 215 | if (!memory_address_addr_space_p (GET_MODE (mem), XEXP (mem, 0), |
| 216 | MEM_ADDR_SPACE (mem))) |
| 217 | { |
| 218 | failure_reason = "would create an invalid MEM"; |
| 219 | return false; |
| 220 | } |
| 221 | |
| 222 | bool can_simplify = [&]() |
| 223 | { |
| 224 | undo_recog_changes undo (old_num_changes); |
| 225 | return can_simplify_addr (XEXP (mem, 0)); |
| 226 | } (); |
| 227 | if (!can_simplify) |
| 228 | { |
| 229 | failure_reason = "would replace a frame address"; |
| 230 | return false; |
| 231 | } |
| 232 | |
| 233 | /* Copy propagations are always ok. Otherwise check the costs. */ |
| 234 | if (!(REG_P (from) && REG_P (to)) |
| 235 | && !should_replace_address (old_num_changes, mem, insn)) |
| 236 | { |
| 237 | failure_reason = "would increase the cost of a MEM"; |
| 238 | return false; |
| 239 | } |
| 240 | |
| 241 | result_flags |= CHANGED_MEM; |
| 242 | return true; |
| 243 | } |
| 244 | |
| 245 | /* OLDX has been simplified to NEWX. Describe the change in terms of |
| 246 | result_flags. */ |
| 247 | |
| 248 | uint16_t |
| 249 | fwprop_propagation::classify_result (rtx old_rtx, rtx new_rtx) |
| 250 | { |
| 251 | if (CONSTANT_P (new_rtx)) |
| 252 | { |
| 253 | /* If OLD_RTX is a LO_SUM, then it presumably exists for a reason, |
| 254 | and NEW_RTX is likely not a legitimate address. We want it to |
| 255 | disappear if it is invalid. |
| 256 | |
| 257 | ??? Using the mode of the LO_SUM as the mode of the address |
| 258 | seems odd, but it was what the pre-SSA code did. */ |
| 259 | if (GET_CODE (old_rtx) == LO_SUM |
| 260 | && !memory_address_p (GET_MODE (old_rtx), new_rtx)) |
| 261 | return CONSTANT; |
| 262 | return CONSTANT | PROFITABLE; |
| 263 | } |
| 264 | |
| 265 | /* Allow replacements that simplify operations on a vector or complex |
| 266 | value to a component. The most prominent case is |
| 267 | (subreg ([vec_]concat ...)). */ |
| 268 | if (REG_P (new_rtx) |
| 269 | && !HARD_REGISTER_P (new_rtx) |
| 270 | && (VECTOR_MODE_P (GET_MODE (from)) |
| 271 | || COMPLEX_MODE_P (GET_MODE (from))) |
| 272 | && GET_MODE (new_rtx) == GET_MODE_INNER (GET_MODE (from))) |
| 273 | return PROFITABLE; |
| 274 | |
| 275 | /* Allow (subreg (mem)) -> (mem) simplifications with the following |
| 276 | exceptions: |
| 277 | 1) Propagating (mem)s into multiple uses is not profitable. |
| 278 | 2) Propagating (mem)s across EBBs may not be profitable if the source EBB |
| 279 | runs less frequently. |
| 280 | 3) Propagating (mem)s into paradoxical (subreg)s is not profitable. |
| 281 | 4) Creating new (mem/v)s is not correct, since DCE will not remove the old |
| 282 | ones. */ |
| 283 | if (single_use_p |
| 284 | && single_ebb_p |
| 285 | && SUBREG_P (old_rtx) |
| 286 | && !paradoxical_subreg_p (x: old_rtx) |
| 287 | && MEM_P (new_rtx) |
| 288 | && !MEM_VOLATILE_P (new_rtx)) |
| 289 | return PROFITABLE; |
| 290 | |
| 291 | return 0; |
| 292 | } |
| 293 | |
| 294 | /* Record that OLD_RTX has been simplified to NEW_RTX. OLD_NUM_CHANGES |
| 295 | is the number of unrelated changes that had been made before processing |
| 296 | OLD_RTX and its subrtxes. OLD_RESULT_FLAGS is the value that result_flags |
| 297 | had at that point. */ |
| 298 | |
| 299 | void |
| 300 | fwprop_propagation::note_simplification (int old_num_changes, |
| 301 | uint16_t old_result_flags, |
| 302 | rtx old_rtx, rtx new_rtx) |
| 303 | { |
| 304 | result_flags &= ~(CONSTANT | PROFITABLE); |
| 305 | uint16_t new_flags = classify_result (old_rtx, new_rtx); |
| 306 | if (old_num_changes) |
| 307 | new_flags &= old_result_flags; |
| 308 | result_flags |= new_flags; |
| 309 | } |
| 310 | |
| 311 | /* Return true if all substitutions eventually folded to constants. */ |
| 312 | |
| 313 | bool |
| 314 | fwprop_propagation::folded_to_constants_p () const |
| 315 | { |
| 316 | /* If we're propagating a HIGH, require it to be folded with a |
| 317 | partnering LO_SUM. For example, a REG_EQUAL note with a register |
| 318 | replaced by an unfolded HIGH is not useful. */ |
| 319 | if (CONSTANT_P (to) && GET_CODE (to) != HIGH) |
| 320 | return true; |
| 321 | return !(result_flags & UNSIMPLIFIED) && (result_flags & CONSTANT); |
| 322 | } |
| 323 | |
| 324 | |
| 325 | /* Return true if it is worth keeping the result of the propagation, |
| 326 | false if it would increase the complexity of the pattern too much. */ |
| 327 | |
| 328 | bool |
| 329 | fwprop_propagation::likely_profitable_p () const |
| 330 | { |
| 331 | if (changed_mem_p ()) |
| 332 | return true; |
| 333 | |
| 334 | if (!(result_flags & UNSIMPLIFIED) |
| 335 | && (result_flags & PROFITABLE)) |
| 336 | return true; |
| 337 | |
| 338 | if (REG_P (to)) |
| 339 | return true; |
| 340 | |
| 341 | if (GET_CODE (to) == SUBREG |
| 342 | && REG_P (SUBREG_REG (to)) |
| 343 | && !paradoxical_subreg_p (x: to)) |
| 344 | return true; |
| 345 | |
| 346 | if (CONSTANT_P (to)) |
| 347 | return true; |
| 348 | |
| 349 | return false; |
| 350 | } |
| 351 | |
| 352 | /* Check that X has a single def. */ |
| 353 | |
| 354 | static bool |
| 355 | reg_single_def_p (rtx x) |
| 356 | { |
| 357 | return REG_P (x) && crtl->ssa->single_dominating_def (REGNO (x)); |
| 358 | } |
| 359 | |
| 360 | /* Try to substitute (set DEST SRC), which defines DEF, into note NOTE of |
| 361 | USE_INSN. Return the number of substitutions on success, otherwise return |
| 362 | -1 and leave USE_INSN unchanged. |
| 363 | |
| 364 | If REQUIRE_CONSTANT is true, require all substituted occurrences of SRC |
| 365 | to fold to a constant, so that the note does not use any more registers |
| 366 | than it did previously. If REQUIRE_CONSTANT is false, also allow the |
| 367 | substitution if it's something we'd normally allow for the main |
| 368 | instruction pattern. */ |
| 369 | |
| 370 | static int |
| 371 | try_fwprop_subst_note (insn_info *use_insn, set_info *def, |
| 372 | rtx note, rtx dest, rtx src, bool require_constant) |
| 373 | { |
| 374 | rtx_insn *use_rtl = use_insn->rtl (); |
| 375 | insn_info *def_insn = def->insn (); |
| 376 | |
| 377 | insn_change_watermark watermark; |
| 378 | fwprop_propagation prop (use_insn, def, dest, src); |
| 379 | if (!prop.apply_to_rvalue (&XEXP (note, 0))) |
| 380 | { |
| 381 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 382 | fprintf (stream: dump_file, format: "cannot propagate from insn %d into" |
| 383 | " notes of insn %d: %s\n", def_insn->uid (), |
| 384 | use_insn->uid (), prop.failure_reason); |
| 385 | return -1; |
| 386 | } |
| 387 | |
| 388 | if (prop.num_replacements == 0) |
| 389 | return 0; |
| 390 | |
| 391 | if (require_constant) |
| 392 | { |
| 393 | if (!prop.folded_to_constants_p ()) |
| 394 | { |
| 395 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 396 | fprintf (stream: dump_file, format: "cannot propagate from insn %d into" |
| 397 | " notes of insn %d: %s\n", def_insn->uid (), |
| 398 | use_insn->uid (), "wouldn't fold to constants"); |
| 399 | return -1; |
| 400 | } |
| 401 | } |
| 402 | else |
| 403 | { |
| 404 | if (!prop.folded_to_constants_p () && !prop.likely_profitable_p ()) |
| 405 | { |
| 406 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 407 | fprintf (stream: dump_file, format: "cannot propagate from insn %d into" |
| 408 | " notes of insn %d: %s\n", def_insn->uid (), |
| 409 | use_insn->uid (), "would increase complexity of node"); |
| 410 | return -1; |
| 411 | } |
| 412 | } |
| 413 | |
| 414 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 415 | { |
| 416 | fprintf (stream: dump_file, format: "\nin notes of insn %d, replacing:\n ", |
| 417 | INSN_UID (insn: use_rtl)); |
| 418 | { |
| 419 | undo_recog_changes undo (0); |
| 420 | print_inline_rtx (dump_file, note, 2); |
| 421 | } |
| 422 | fprintf (stream: dump_file, format: "\n with:\n "); |
| 423 | print_inline_rtx (dump_file, note, 2); |
| 424 | fprintf (stream: dump_file, format: "\n"); |
| 425 | } |
| 426 | watermark.keep (); |
| 427 | return prop.num_replacements; |
| 428 | } |
| 429 | |
| 430 | /* Try to substitute (set DEST SRC), which defines DEF, into location LOC of |
| 431 | USE_INSN's pattern. Return true on success, otherwise leave USE_INSN |
| 432 | unchanged. */ |
| 433 | |
| 434 | static bool |
| 435 | try_fwprop_subst_pattern (obstack_watermark &attempt, insn_change &use_change, |
| 436 | set_info *def, rtx *loc, rtx dest, rtx src) |
| 437 | { |
| 438 | insn_info *use_insn = use_change.insn (); |
| 439 | rtx_insn *use_rtl = use_insn->rtl (); |
| 440 | insn_info *def_insn = def->insn (); |
| 441 | |
| 442 | insn_change_watermark watermark; |
| 443 | fwprop_propagation prop (use_insn, def, dest, src); |
| 444 | if (!prop.apply_to_pattern (loc)) |
| 445 | { |
| 446 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 447 | fprintf (stream: dump_file, format: "cannot propagate from insn %d into" |
| 448 | " insn %d: %s\n", def_insn->uid (), use_insn->uid (), |
| 449 | prop.failure_reason); |
| 450 | return false; |
| 451 | } |
| 452 | |
| 453 | if (prop.num_replacements == 0) |
| 454 | return false; |
| 455 | |
| 456 | if (!prop.likely_profitable_p () |
| 457 | && (prop.changed_mem_p () |
| 458 | || contains_mem_rtx_p (x: src) |
| 459 | || use_insn->is_asm () |
| 460 | || use_insn->is_debug_insn ())) |
| 461 | { |
| 462 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 463 | fprintf (stream: dump_file, format: "cannot propagate from insn %d into" |
| 464 | " insn %d: %s\n", def_insn->uid (), use_insn->uid (), |
| 465 | "would increase complexity of pattern"); |
| 466 | return false; |
| 467 | } |
| 468 | |
| 469 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 470 | { |
| 471 | fprintf (stream: dump_file, format: "\npropagating insn %d into insn %d, replacing:\n", |
| 472 | def_insn->uid (), use_insn->uid ()); |
| 473 | undo_recog_changes undo (0); |
| 474 | print_rtl_single (dump_file, PATTERN (insn: use_rtl)); |
| 475 | } |
| 476 | |
| 477 | bool ok = recog (watermark&: attempt, change&: use_change); |
| 478 | if (ok |
| 479 | && !prop.changed_mem_p () |
| 480 | && !use_insn->is_asm () |
| 481 | && !use_insn->is_debug_insn ()) |
| 482 | { |
| 483 | bool strict_p = !prop.likely_profitable_p (); |
| 484 | if (!change_is_worthwhile (change&: use_change, strict_p)) |
| 485 | { |
| 486 | if (dump_file) |
| 487 | fprintf (stream: dump_file, format: "change not profitable"); |
| 488 | ok = false; |
| 489 | } |
| 490 | } |
| 491 | |
| 492 | if (!ok) |
| 493 | { |
| 494 | /* The pattern didn't match, but if all uses of SRC folded to |
| 495 | constants, we can add a REG_EQUAL note for the result, if there |
| 496 | isn't one already. */ |
| 497 | if (!prop.folded_to_constants_p ()) |
| 498 | return false; |
| 499 | |
| 500 | /* Test this first to avoid creating an unnecessary copy of SRC. */ |
| 501 | if (find_reg_note (use_rtl, REG_EQUAL, NULL_RTX)) |
| 502 | return false; |
| 503 | |
| 504 | rtx set = set_for_reg_notes (use_rtl); |
| 505 | if (!set || !REG_P (SET_DEST (set))) |
| 506 | return false; |
| 507 | |
| 508 | rtx value = copy_rtx (SET_SRC (set)); |
| 509 | cancel_changes (0); |
| 510 | |
| 511 | /* If there are any paradoxical SUBREGs, drop the REG_EQUAL note, |
| 512 | because the bits in there can be anything and so might not |
| 513 | match the REG_EQUAL note content. See PR70574. */ |
| 514 | if (contains_paradoxical_subreg_p (SET_SRC (set))) |
| 515 | return false; |
| 516 | |
| 517 | if (dump_file && (dump_flags & TDF_DETAILS)) |
| 518 | fprintf (stream: dump_file, format: " Setting REG_EQUAL note\n"); |
| 519 | |
| 520 | return set_unique_reg_note (use_rtl, REG_EQUAL, value); |
| 521 | } |
| 522 | |
| 523 | rtx *note_ptr = ®_NOTES (use_rtl); |
| 524 | while (rtx note = *note_ptr) |
| 525 | { |
| 526 | if ((REG_NOTE_KIND (note) == REG_EQUAL |
| 527 | || REG_NOTE_KIND (note) == REG_EQUIV) |
| 528 | && try_fwprop_subst_note (use_insn, def, note, dest, src, require_constant: false) < 0) |
| 529 | { |
| 530 | *note_ptr = XEXP (note, 1); |
| 531 | free_EXPR_LIST_node (note); |
| 532 | } |
| 533 | else |
| 534 | note_ptr = &XEXP (note, 1); |
| 535 | } |
| 536 | |
| 537 | confirm_change_group (); |
| 538 | crtl->ssa->change_insn (change&: use_change); |
| 539 | num_changes++; |
| 540 | return true; |
| 541 | } |
| 542 | |
| 543 | /* Try to substitute (set DEST SRC), which defines DEF, into USE_INSN's notes, |
| 544 | given that it was not possible to do this for USE_INSN's main pattern. |
| 545 | Return true on success, otherwise leave USE_INSN unchanged. */ |
| 546 | |
| 547 | static bool |
| 548 | try_fwprop_subst_notes (insn_info *use_insn, set_info *def, |
| 549 | rtx dest, rtx src) |
| 550 | { |
| 551 | rtx_insn *use_rtl = use_insn->rtl (); |
| 552 | for (rtx note = REG_NOTES (use_rtl); note; note = XEXP (note, 1)) |
| 553 | if ((REG_NOTE_KIND (note) == REG_EQUAL |
| 554 | || REG_NOTE_KIND (note) == REG_EQUIV) |
| 555 | && try_fwprop_subst_note (use_insn, def, note, dest, src, require_constant: true) > 0) |
| 556 | { |
| 557 | confirm_change_group (); |
| 558 | return true; |
| 559 | } |
| 560 | |
| 561 | return false; |
| 562 | } |
| 563 | |
| 564 | /* Check whether we could validly substitute (set DEST SRC), which defines DEF, |
| 565 | into USE. If so, first try performing the substitution in location LOC |
| 566 | of USE->insn ()'s pattern. If that fails, try instead to substitute |
| 567 | into the notes. |
| 568 | |
| 569 | Return true on success, otherwise leave USE_INSN unchanged. */ |
| 570 | |
| 571 | static bool |
| 572 | try_fwprop_subst (use_info *use, set_info *def, |
| 573 | rtx *loc, rtx dest, rtx src) |
| 574 | { |
| 575 | insn_info *use_insn = use->insn (); |
| 576 | insn_info *def_insn = def->insn (); |
| 577 | |
| 578 | auto attempt = crtl->ssa->new_change_attempt (); |
| 579 | use_array src_uses = remove_note_accesses (watermark&: attempt, accesses: def_insn->uses ()); |
| 580 | |
| 581 | /* ??? Not really a meaningful test: it means we can propagate arithmetic |
| 582 | involving hard registers but not bare references to them. A better |
| 583 | test would be to iterate over src_uses looking for hard registers |
| 584 | that are not fixed. */ |
| 585 | if (REG_P (src) && HARD_REGISTER_P (src)) |
| 586 | return false; |
| 587 | |
| 588 | /* ??? It would be better to make this EBB-based instead. That would |
| 589 | involve checking for equal EBBs rather than equal BBs and trying |
| 590 | to make the uses available at use_insn->ebb ()->first_bb (). */ |
| 591 | if (def_insn->bb () != use_insn->bb ()) |
| 592 | { |
| 593 | src_uses = crtl->ssa->make_uses_available (watermark&: attempt, uses: src_uses, |
| 594 | bb: use_insn->bb (), |
| 595 | will_be_debug_uses: use_insn->is_debug_insn ()); |
| 596 | if (!src_uses.is_valid ()) |
| 597 | return false; |
| 598 | } |
| 599 | |
| 600 | insn_change use_change (use_insn); |
| 601 | use_change.new_uses = merge_access_arrays (watermark&: attempt, accesses1: use_change.new_uses, |
| 602 | accesses2: src_uses); |
| 603 | if (!use_change.new_uses.is_valid ()) |
| 604 | return false; |
| 605 | |
| 606 | /* ??? We could allow movement within the EBB by adding: |
| 607 | |
| 608 | use_change.move_range = use_insn->ebb ()->insn_range (); */ |
| 609 | if (!restrict_movement (change&: use_change)) |
| 610 | return false; |
| 611 | |
| 612 | return (try_fwprop_subst_pattern (attempt, use_change, def, loc, dest, src) |
| 613 | || try_fwprop_subst_notes (use_insn, def, dest, src)); |
| 614 | } |
| 615 | |
| 616 | /* For the given single_set INSN, containing SRC known to be a |
| 617 | ZERO_EXTEND or SIGN_EXTEND of a register, return true if INSN |
| 618 | is redundant due to the register being set by a LOAD_EXTEND_OP |
| 619 | load from memory. */ |
| 620 | |
| 621 | static bool |
| 622 | free_load_extend (rtx src, insn_info *insn) |
| 623 | { |
| 624 | rtx reg = XEXP (src, 0); |
| 625 | if (load_extend_op (GET_MODE (reg)) != GET_CODE (src)) |
| 626 | return false; |
| 627 | |
| 628 | def_info *def = nullptr; |
| 629 | for (use_info *use : insn->uses ()) |
| 630 | if (use->regno () == REGNO (reg)) |
| 631 | { |
| 632 | def = use->def (); |
| 633 | break; |
| 634 | } |
| 635 | |
| 636 | if (!def) |
| 637 | return false; |
| 638 | |
| 639 | insn_info *def_insn = def->insn (); |
| 640 | if (def_insn->is_artificial ()) |
| 641 | return false; |
| 642 | |
| 643 | rtx_insn *def_rtl = def_insn->rtl (); |
| 644 | if (NONJUMP_INSN_P (def_rtl)) |
| 645 | { |
| 646 | rtx patt = PATTERN (insn: def_rtl); |
| 647 | |
| 648 | if (GET_CODE (patt) == SET |
| 649 | && GET_CODE (SET_SRC (patt)) == MEM |
| 650 | && rtx_equal_p (SET_DEST (patt), reg)) |
| 651 | return true; |
| 652 | } |
| 653 | return false; |
| 654 | } |
| 655 | |
| 656 | /* Subroutine of forward_propagate_subreg that handles a use of DEST |
| 657 | in REF. The other parameters are the same. */ |
| 658 | |
| 659 | static bool |
| 660 | forward_propagate_subreg (use_info *use, set_info *def, |
| 661 | rtx dest, rtx src, df_ref ref) |
| 662 | { |
| 663 | scalar_int_mode int_use_mode, src_mode; |
| 664 | |
| 665 | /* Only consider subregs... */ |
| 666 | rtx use_reg = DF_REF_REG (ref); |
| 667 | machine_mode use_mode = GET_MODE (use_reg); |
| 668 | if (GET_CODE (use_reg) != SUBREG |
| 669 | || GET_MODE (SUBREG_REG (use_reg)) != GET_MODE (dest)) |
| 670 | return false; |
| 671 | |
| 672 | /* ??? Replacing throughout the pattern would help for match_dups. */ |
| 673 | rtx *loc = DF_REF_LOC (ref); |
| 674 | if (paradoxical_subreg_p (x: use_reg)) |
| 675 | { |
| 676 | /* If this is a paradoxical SUBREG, we have no idea what value the |
| 677 | extra bits would have. However, if the operand is equivalent to |
| 678 | a SUBREG whose operand is the same as our mode, and all the modes |
| 679 | are within a word, we can just use the inner operand because |
| 680 | these SUBREGs just say how to treat the register. */ |
| 681 | if (GET_CODE (src) == SUBREG |
| 682 | && REG_P (SUBREG_REG (src)) |
| 683 | && REGNO (SUBREG_REG (src)) >= FIRST_PSEUDO_REGISTER |
| 684 | && GET_MODE (SUBREG_REG (src)) == use_mode |
| 685 | && subreg_lowpart_p (src)) |
| 686 | return try_fwprop_subst (use, def, loc, dest: use_reg, SUBREG_REG (src)); |
| 687 | } |
| 688 | |
| 689 | /* If this is a SUBREG of a ZERO_EXTEND or SIGN_EXTEND, and the SUBREG |
| 690 | is the low part of the reg being extended then just use the inner |
| 691 | operand. Don't do this if the ZERO_EXTEND or SIGN_EXTEND insn will |
| 692 | be removed due to it matching a LOAD_EXTEND_OP load from memory, |
| 693 | or due to the operation being a no-op when applied to registers. |
| 694 | For example, if we have: |
| 695 | |
| 696 | A: (set (reg:DI X) (sign_extend:DI (reg:SI Y))) |
| 697 | B: (... (subreg:SI (reg:DI X)) ...) |
| 698 | |
| 699 | and mode_rep_extended says that Y is already sign-extended, |
| 700 | the backend will typically allow A to be combined with the |
| 701 | definition of Y or, failing that, allow A to be deleted after |
| 702 | reload through register tying. Introducing more uses of Y |
| 703 | prevents both optimisations. */ |
| 704 | else if (is_a <scalar_int_mode> (m: use_mode, result: &int_use_mode) |
| 705 | && subreg_lowpart_p (use_reg)) |
| 706 | { |
| 707 | if ((GET_CODE (src) == ZERO_EXTEND |
| 708 | || GET_CODE (src) == SIGN_EXTEND) |
| 709 | && is_a <scalar_int_mode> (GET_MODE (src), result: &src_mode) |
| 710 | && REG_P (XEXP (src, 0)) |
| 711 | && REGNO (XEXP (src, 0)) >= FIRST_PSEUDO_REGISTER |
| 712 | && GET_MODE (XEXP (src, 0)) == use_mode |
| 713 | && !free_load_extend (src, insn: def->insn ()) |
| 714 | && (targetm.mode_rep_extended (int_use_mode, src_mode) |
| 715 | != (int) GET_CODE (src))) |
| 716 | return try_fwprop_subst (use, def, loc, dest: use_reg, XEXP (src, 0)); |
| 717 | } |
| 718 | |
| 719 | return false; |
| 720 | } |
| 721 | |
| 722 | /* Try to substitute (set DEST SRC), which defines DEF, into USE and simplify |
| 723 | the result, handling cases where DEST is used in a subreg and where |
| 724 | applying that subreg to SRC results in a useful simplification. */ |
| 725 | |
| 726 | static bool |
| 727 | forward_propagate_subreg (use_info *use, set_info *def, rtx dest, rtx src) |
| 728 | { |
| 729 | if (!use->includes_subregs () || !REG_P (dest)) |
| 730 | return false; |
| 731 | |
| 732 | if (GET_CODE (src) != SUBREG |
| 733 | && GET_CODE (src) != ZERO_EXTEND |
| 734 | && GET_CODE (src) != SIGN_EXTEND) |
| 735 | return false; |
| 736 | |
| 737 | rtx_insn *use_rtl = use->insn ()->rtl (); |
| 738 | df_ref ref; |
| 739 | |
| 740 | FOR_EACH_INSN_USE (ref, use_rtl) |
| 741 | if (DF_REF_REGNO (ref) == use->regno () |
| 742 | && forward_propagate_subreg (use, def, dest, src, ref)) |
| 743 | return true; |
| 744 | |
| 745 | FOR_EACH_INSN_EQ_USE (ref, use_rtl) |
| 746 | if (DF_REF_REGNO (ref) == use->regno () |
| 747 | && forward_propagate_subreg (use, def, dest, src, ref)) |
| 748 | return true; |
| 749 | |
| 750 | return false; |
| 751 | } |
| 752 | |
| 753 | /* Try to substitute (set DEST SRC), which defines DEF, into USE and |
| 754 | simplify the result. */ |
| 755 | |
| 756 | static bool |
| 757 | forward_propagate_and_simplify (use_info *use, set_info *def, |
| 758 | rtx dest, rtx src) |
| 759 | { |
| 760 | insn_info *use_insn = use->insn (); |
| 761 | rtx_insn *use_rtl = use_insn->rtl (); |
| 762 | insn_info *def_insn = def->insn (); |
| 763 | |
| 764 | /* ??? This check seems unnecessary. We should be able to propagate |
| 765 | into any kind of instruction, regardless of whether it's a single set. |
| 766 | It seems odd to be more permissive with asms than normal instructions. */ |
| 767 | bool need_single_set = (!use_insn->is_asm () && !use_insn->is_debug_insn ()); |
| 768 | rtx use_set = single_set (insn: use_rtl); |
| 769 | if (need_single_set && !use_set) |
| 770 | return false; |
| 771 | |
| 772 | /* Do not propagate into PC etc. |
| 773 | |
| 774 | ??? This too seems unnecessary. The current code should work correctly |
| 775 | without it, including cases where jumps become unconditional. */ |
| 776 | if (use_set && GET_MODE (SET_DEST (use_set)) == VOIDmode) |
| 777 | return false; |
| 778 | |
| 779 | /* In __asm don't replace if src might need more registers than |
| 780 | reg, as that could increase register pressure on the __asm. */ |
| 781 | if (use_insn->is_asm () && def_insn->uses ().size () > 1) |
| 782 | return false; |
| 783 | |
| 784 | /* Check if the def is loading something from the constant pool; in this |
| 785 | case we would undo optimization such as compress_float_constant. |
| 786 | Still, we can set a REG_EQUAL note. */ |
| 787 | if (MEM_P (src) && MEM_READONLY_P (src)) |
| 788 | { |
| 789 | rtx x = avoid_constant_pool_reference (src); |
| 790 | rtx note_set; |
| 791 | if (x != src |
| 792 | && (note_set = set_for_reg_notes (use_rtl)) |
| 793 | && REG_P (SET_DEST (note_set)) |
| 794 | && !contains_paradoxical_subreg_p (SET_SRC (note_set))) |
| 795 | { |
| 796 | rtx note = find_reg_note (use_rtl, REG_EQUAL, NULL_RTX); |
| 797 | rtx old_rtx = note ? XEXP (note, 0) : SET_SRC (note_set); |
| 798 | rtx new_rtx = simplify_replace_rtx (old_rtx, src, x); |
| 799 | if (old_rtx != new_rtx) |
| 800 | set_unique_reg_note (use_rtl, REG_EQUAL, copy_rtx (new_rtx)); |
| 801 | } |
| 802 | return false; |
| 803 | } |
| 804 | |
| 805 | /* ??? Unconditionally propagating into PATTERN would work better |
| 806 | for instructions that have match_dups. */ |
| 807 | rtx *loc = need_single_set ? &use_set : &PATTERN (insn: use_rtl); |
| 808 | return try_fwprop_subst (use, def, loc, dest, src); |
| 809 | } |
| 810 | |
| 811 | /* Given a use USE of an insn, if it has a single reaching |
| 812 | definition, try to forward propagate it into that insn. |
| 813 | Return true if something changed. |
| 814 | |
| 815 | REG_PROP_ONLY is true if we should only propagate register copies. */ |
| 816 | |
| 817 | static bool |
| 818 | forward_propagate_into (use_info *use, bool reg_prop_only = false) |
| 819 | { |
| 820 | if (use->includes_read_writes ()) |
| 821 | return false; |
| 822 | |
| 823 | /* Disregard uninitialized uses. */ |
| 824 | set_info *def = use->def (); |
| 825 | if (!def) |
| 826 | return false; |
| 827 | |
| 828 | /* Only consider single-register definitions. This could be relaxed, |
| 829 | but it should rarely be needed before RA. */ |
| 830 | def = look_through_degenerate_phi (access: def); |
| 831 | if (def->includes_multiregs ()) |
| 832 | return false; |
| 833 | |
| 834 | /* Only consider uses whose definition comes from a real instruction. */ |
| 835 | insn_info *def_insn = def->insn (); |
| 836 | if (def_insn->is_artificial ()) |
| 837 | return false; |
| 838 | |
| 839 | rtx_insn *def_rtl = def_insn->rtl (); |
| 840 | if (!NONJUMP_INSN_P (def_rtl)) |
| 841 | return false; |
| 842 | /* ??? This seems an unnecessary restriction. We can easily tell |
| 843 | which set the definition comes from. */ |
| 844 | if (multiple_sets (def_rtl)) |
| 845 | return false; |
| 846 | rtx def_set = simple_regno_set (PATTERN (insn: def_rtl), def->regno ()); |
| 847 | if (!def_set) |
| 848 | return false; |
| 849 | |
| 850 | rtx dest = SET_DEST (def_set); |
| 851 | rtx src = SET_SRC (def_set); |
| 852 | if (volatile_refs_p (src)) |
| 853 | return false; |
| 854 | |
| 855 | /* Allow propagations into a loop only for reg-to-reg copies, since |
| 856 | replacing one register by another shouldn't increase the cost. |
| 857 | Propagations from inner loop to outer loop should also be ok. */ |
| 858 | struct loop *def_loop = def_insn->bb ()->cfg_bb ()->loop_father; |
| 859 | struct loop *use_loop = use->bb ()->cfg_bb ()->loop_father; |
| 860 | if ((reg_prop_only |
| 861 | || (def_loop != use_loop |
| 862 | && !flow_loop_nested_p (use_loop, def_loop))) |
| 863 | && (!reg_single_def_p (x: dest) || !reg_single_def_p (x: src))) |
| 864 | return false; |
| 865 | |
| 866 | /* Don't substitute into a non-local goto, this confuses CFG. */ |
| 867 | insn_info *use_insn = use->insn (); |
| 868 | rtx_insn *use_rtl = use_insn->rtl (); |
| 869 | if (JUMP_P (use_rtl) |
| 870 | && find_reg_note (use_rtl, REG_NON_LOCAL_GOTO, NULL_RTX)) |
| 871 | return false; |
| 872 | |
| 873 | if (forward_propagate_and_simplify (use, def, dest, src) |
| 874 | || forward_propagate_subreg (use, def, dest, src)) |
| 875 | return true; |
| 876 | |
| 877 | return false; |
| 878 | } |
| 879 | |
| 880 | static void |
| 881 | fwprop_init (void) |
| 882 | { |
| 883 | num_changes = 0; |
| 884 | calculate_dominance_info (CDI_DOMINATORS); |
| 885 | |
| 886 | /* We do not always want to propagate into loops, so we have to find |
| 887 | loops and be careful about them. Avoid CFG modifications so that |
| 888 | we don't have to update dominance information afterwards for |
| 889 | build_single_def_use_links. */ |
| 890 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS); |
| 891 | |
| 892 | df_analyze (); |
| 893 | crtl->ssa = new rtl_ssa::function_info (cfun); |
| 894 | } |
| 895 | |
| 896 | static void |
| 897 | fwprop_done (void) |
| 898 | { |
| 899 | loop_optimizer_finalize (); |
| 900 | |
| 901 | crtl->ssa->perform_pending_updates (); |
| 902 | free_dominance_info (CDI_DOMINATORS); |
| 903 | cleanup_cfg (0); |
| 904 | |
| 905 | delete crtl->ssa; |
| 906 | crtl->ssa = nullptr; |
| 907 | |
| 908 | delete_trivially_dead_insns (get_insns (), max_reg_num ()); |
| 909 | |
| 910 | if (dump_file) |
| 911 | fprintf (stream: dump_file, |
| 912 | format: "\nNumber of successful forward propagations: %d\n\n", |
| 913 | num_changes); |
| 914 | } |
| 915 | |
| 916 | /* Try to optimize INSN, returning true if something changes. |
| 917 | FWPROP_ADDR_P is true if we are running fwprop_addr rather than |
| 918 | the full fwprop. */ |
| 919 | |
| 920 | static bool |
| 921 | fwprop_insn (insn_info *insn, bool fwprop_addr_p) |
| 922 | { |
| 923 | for (use_info *use : insn->uses ()) |
| 924 | { |
| 925 | if (use->is_mem ()) |
| 926 | continue; |
| 927 | /* ??? The choices here follow those in the pre-SSA code. */ |
| 928 | if (!use->includes_address_uses ()) |
| 929 | { |
| 930 | if (forward_propagate_into (use, reg_prop_only: fwprop_addr_p)) |
| 931 | return true; |
| 932 | } |
| 933 | else |
| 934 | { |
| 935 | struct loop *loop = insn->bb ()->cfg_bb ()->loop_father; |
| 936 | /* The outermost loop is not really a loop. */ |
| 937 | if (loop == NULL || loop_outer (loop) == NULL) |
| 938 | { |
| 939 | if (forward_propagate_into (use, reg_prop_only: fwprop_addr_p)) |
| 940 | return true; |
| 941 | } |
| 942 | else if (fwprop_addr_p) |
| 943 | { |
| 944 | if (forward_propagate_into (use, reg_prop_only: false)) |
| 945 | return true; |
| 946 | } |
| 947 | } |
| 948 | } |
| 949 | return false; |
| 950 | } |
| 951 | |
| 952 | /* Main entry point. */ |
| 953 | |
| 954 | static bool |
| 955 | gate_fwprop (void) |
| 956 | { |
| 957 | return optimize > 0 && flag_forward_propagate; |
| 958 | } |
| 959 | |
| 960 | static unsigned int |
| 961 | fwprop (bool fwprop_addr_p) |
| 962 | { |
| 963 | fwprop_init (); |
| 964 | |
| 965 | /* Go through all the instructions (including debug instructions) looking |
| 966 | for uses that we could propagate into. |
| 967 | |
| 968 | Do not forward propagate addresses into loops until after unrolling. |
| 969 | CSE did so because it was able to fix its own mess, but we are not. */ |
| 970 | |
| 971 | insn_info *next; |
| 972 | |
| 973 | /* ??? This code uses a worklist in order to preserve the behavior |
| 974 | of the pre-SSA implementation. It would be better to instead |
| 975 | iterate on each instruction until no more propagations are |
| 976 | possible, then move on to the next. */ |
| 977 | auto_vec<insn_info *> worklist; |
| 978 | for (insn_info *insn = crtl->ssa->first_insn (); insn; insn = next) |
| 979 | { |
| 980 | next = insn->next_any_insn (); |
| 981 | if (insn->can_be_optimized () || insn->is_debug_insn ()) |
| 982 | if (fwprop_insn (insn, fwprop_addr_p)) |
| 983 | worklist.safe_push (obj: insn); |
| 984 | } |
| 985 | for (unsigned int i = 0; i < worklist.length (); ++i) |
| 986 | { |
| 987 | insn_info *insn = worklist[i]; |
| 988 | if (fwprop_insn (insn, fwprop_addr_p)) |
| 989 | worklist.safe_push (obj: insn); |
| 990 | } |
| 991 | |
| 992 | fwprop_done (); |
| 993 | return 0; |
| 994 | } |
| 995 | |
| 996 | namespace { |
| 997 | |
| 998 | const pass_data pass_data_rtl_fwprop = |
| 999 | { |
| 1000 | .type: RTL_PASS, /* type */ |
| 1001 | .name: "fwprop1", /* name */ |
| 1002 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
| 1003 | .tv_id: TV_FWPROP, /* tv_id */ |
| 1004 | .properties_required: 0, /* properties_required */ |
| 1005 | .properties_provided: 0, /* properties_provided */ |
| 1006 | .properties_destroyed: 0, /* properties_destroyed */ |
| 1007 | .todo_flags_start: 0, /* todo_flags_start */ |
| 1008 | TODO_df_finish, /* todo_flags_finish */ |
| 1009 | }; |
| 1010 | |
| 1011 | class pass_rtl_fwprop : public rtl_opt_pass |
| 1012 | { |
| 1013 | public: |
| 1014 | pass_rtl_fwprop (gcc::context *ctxt) |
| 1015 | : rtl_opt_pass (pass_data_rtl_fwprop, ctxt) |
| 1016 | {} |
| 1017 | |
| 1018 | /* opt_pass methods: */ |
| 1019 | bool gate (function *) final override { return gate_fwprop (); } |
| 1020 | unsigned int execute (function *) final override { return fwprop (fwprop_addr_p: false); } |
| 1021 | |
| 1022 | }; // class pass_rtl_fwprop |
| 1023 | |
| 1024 | } // anon namespace |
| 1025 | |
| 1026 | rtl_opt_pass * |
| 1027 | make_pass_rtl_fwprop (gcc::context *ctxt) |
| 1028 | { |
| 1029 | return new pass_rtl_fwprop (ctxt); |
| 1030 | } |
| 1031 | |
| 1032 | namespace { |
| 1033 | |
| 1034 | const pass_data pass_data_rtl_fwprop_addr = |
| 1035 | { |
| 1036 | .type: RTL_PASS, /* type */ |
| 1037 | .name: "fwprop2", /* name */ |
| 1038 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
| 1039 | .tv_id: TV_FWPROP, /* tv_id */ |
| 1040 | .properties_required: 0, /* properties_required */ |
| 1041 | .properties_provided: 0, /* properties_provided */ |
| 1042 | .properties_destroyed: 0, /* properties_destroyed */ |
| 1043 | .todo_flags_start: 0, /* todo_flags_start */ |
| 1044 | TODO_df_finish, /* todo_flags_finish */ |
| 1045 | }; |
| 1046 | |
| 1047 | class pass_rtl_fwprop_addr : public rtl_opt_pass |
| 1048 | { |
| 1049 | public: |
| 1050 | pass_rtl_fwprop_addr (gcc::context *ctxt) |
| 1051 | : rtl_opt_pass (pass_data_rtl_fwprop_addr, ctxt) |
| 1052 | {} |
| 1053 | |
| 1054 | /* opt_pass methods: */ |
| 1055 | bool gate (function *) final override { return gate_fwprop (); } |
| 1056 | unsigned int execute (function *) final override { return fwprop (fwprop_addr_p: true); } |
| 1057 | |
| 1058 | }; // class pass_rtl_fwprop_addr |
| 1059 | |
| 1060 | } // anon namespace |
| 1061 | |
| 1062 | rtl_opt_pass * |
| 1063 | make_pass_rtl_fwprop_addr (gcc::context *ctxt) |
| 1064 | { |
| 1065 | return new pass_rtl_fwprop_addr (ctxt); |
| 1066 | } |
| 1067 |
Definitions
- num_changes
- can_simplify_addr
- should_replace_address
- fwprop_propagation
- changed_mem_p
- fwprop_propagation
- check_mem
- classify_result
- note_simplification
- folded_to_constants_p
- likely_profitable_p
- reg_single_def_p
- try_fwprop_subst_note
- try_fwprop_subst_pattern
- try_fwprop_subst_notes
- try_fwprop_subst
- free_load_extend
- forward_propagate_subreg
- forward_propagate_subreg
- forward_propagate_and_simplify
- forward_propagate_into
- fwprop_init
- fwprop_done
- fwprop_insn
- gate_fwprop
- fwprop
- pass_data_rtl_fwprop
- pass_rtl_fwprop
- pass_rtl_fwprop
- gate
- execute
- make_pass_rtl_fwprop
- pass_data_rtl_fwprop_addr
- pass_rtl_fwprop_addr
- pass_rtl_fwprop_addr
- gate
- execute
Update your C++ knowledge – Modern C++11/14/17 Training
Find out more
