1 /* Expands front end tree to back end RTL for GCC. 2 Copyright (C) 1987-2018 Free Software Foundation, Inc. 3 4 This file is part of GCC. 5 6 GCC is free software; you can redistribute it and/or modify it under 7 the terms of the GNU General Public License as published by the Free 8 Software Foundation; either version 3, or (at your option) any later 9 version. 10 11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY 12 WARRANTY; without even the implied warranty of MERCHANTABILITY or 13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 14 for more details. 15 16 You should have received a copy of the GNU General Public License 17 along with GCC; see the file COPYING3. If not see 18 <http://www.gnu.org/licenses/>. */ 19 20 /* This file handles the generation of rtl code from tree structure 21 at the level of the function as a whole. 22 It creates the rtl expressions for parameters and auto variables 23 and has full responsibility for allocating stack slots. 24 25 `expand_function_start' is called at the beginning of a function, 26 before the function body is parsed, and `expand_function_end' is 27 called after parsing the body. 28 29 Call `assign_stack_local' to allocate a stack slot for a local variable. 30 This is usually done during the RTL generation for the function body, 31 but it can also be done in the reload pass when a pseudo-register does 32 not get a hard register. */ 33 34 #include "config.h" 35 #include "system.h" 36 #include "coretypes.h" 37 #include "backend.h" 38 #include "target.h" 39 #include "rtl.h" 40 #include "tree.h" 41 #include "gimple-expr.h" 42 #include "cfghooks.h" 43 #include "df.h" 44 #include "memmodel.h" 45 #include "tm_p.h" 46 #include "stringpool.h" 47 #include "expmed.h" 48 #include "optabs.h" 49 #include "regs.h" 50 #include "emit-rtl.h" 51 #include "recog.h" 52 #include "rtl-error.h" 53 #include "alias.h" 54 #include "fold-const.h" 55 #include "stor-layout.h" 56 #include "varasm.h" 57 #include "except.h" 58 #include "dojump.h" 59 #include "explow.h" 60 #include "calls.h" 61 #include "expr.h" 62 #include "optabs-tree.h" 63 #include "output.h" 64 #include "langhooks.h" 65 #include "common/common-target.h" 66 #include "gimplify.h" 67 #include "tree-pass.h" 68 #include "cfgrtl.h" 69 #include "cfganal.h" 70 #include "cfgbuild.h" 71 #include "cfgcleanup.h" 72 #include "cfgexpand.h" 73 #include "shrink-wrap.h" 74 #include "toplev.h" 75 #include "rtl-iter.h" 76 #include "tree-chkp.h" 77 #include "rtl-chkp.h" 78 #include "tree-dfa.h" 79 #include "tree-ssa.h" 80 #include "stringpool.h" 81 #include "attribs.h" 82 #include "gimple.h" 83 #include "options.h" 84 85 /* So we can assign to cfun in this file. */ 86 #undef cfun 87 88 #ifndef STACK_ALIGNMENT_NEEDED 89 #define STACK_ALIGNMENT_NEEDED 1 90 #endif 91 92 #define STACK_BYTES (STACK_BOUNDARY / BITS_PER_UNIT) 93 94 /* Round a value to the lowest integer less than it that is a multiple of 95 the required alignment. Avoid using division in case the value is 96 negative. Assume the alignment is a power of two. */ 97 #define FLOOR_ROUND(VALUE,ALIGN) ((VALUE) & ~((ALIGN) - 1)) 98 99 /* Similar, but round to the next highest integer that meets the 100 alignment. */ 101 #define CEIL_ROUND(VALUE,ALIGN) (((VALUE) + (ALIGN) - 1) & ~((ALIGN)- 1)) 102 103 /* Nonzero once virtual register instantiation has been done. 104 assign_stack_local uses frame_pointer_rtx when this is nonzero. 105 calls.c:emit_library_call_value_1 uses it to set up 106 post-instantiation libcalls. */ 107 int virtuals_instantiated; 108 109 /* Assign unique numbers to labels generated for profiling, debugging, etc. */ 110 static GTY(()) int funcdef_no; 111 112 /* These variables hold pointers to functions to create and destroy 113 target specific, per-function data structures. */ 114 struct machine_function * (*init_machine_status) (void); 115 116 /* The currently compiled function. */ 117 struct function *cfun = 0; 118 119 /* These hashes record the prologue and epilogue insns. */ 120 121 struct insn_cache_hasher : ggc_cache_ptr_hash<rtx_def> 122 { 123 static hashval_t hash (rtx x) { return htab_hash_pointer (x); } 124 static bool equal (rtx a, rtx b) { return a == b; } 125 }; 126 127 static GTY((cache)) 128 hash_table<insn_cache_hasher> *prologue_insn_hash; 129 static GTY((cache)) 130 hash_table<insn_cache_hasher> *epilogue_insn_hash; 131 132 133 hash_table<used_type_hasher> *types_used_by_vars_hash = NULL; 134 vec<tree, va_gc> *types_used_by_cur_var_decl; 135 136 /* Forward declarations. */ 137 138 static struct temp_slot *find_temp_slot_from_address (rtx); 139 static void pad_to_arg_alignment (struct args_size *, int, struct args_size *); 140 static void pad_below (struct args_size *, machine_mode, tree); 141 static void reorder_blocks_1 (rtx_insn *, tree, vec<tree> *); 142 static int all_blocks (tree, tree *); 143 static tree *get_block_vector (tree, int *); 144 extern tree debug_find_var_in_block_tree (tree, tree); 145 /* We always define `record_insns' even if it's not used so that we 146 can always export `prologue_epilogue_contains'. */ 147 static void record_insns (rtx_insn *, rtx, hash_table<insn_cache_hasher> **) 148 ATTRIBUTE_UNUSED; 149 static bool contains (const rtx_insn *, hash_table<insn_cache_hasher> *); 150 static void prepare_function_start (void); 151 static void do_clobber_return_reg (rtx, void *); 152 static void do_use_return_reg (rtx, void *); 153 154 155 /* Stack of nested functions. */ 156 /* Keep track of the cfun stack. */ 157 158 static vec<function *> function_context_stack; 159 160 /* Save the current context for compilation of a nested function. 161 This is called from language-specific code. */ 162 163 void 164 push_function_context (void) 165 { 166 if (cfun == 0) 167 allocate_struct_function (NULL, false); 168 169 function_context_stack.safe_push (cfun); 170 set_cfun (NULL); 171 } 172 173 /* Restore the last saved context, at the end of a nested function. 174 This function is called from language-specific code. */ 175 176 void 177 pop_function_context (void) 178 { 179 struct function *p = function_context_stack.pop (); 180 set_cfun (p); 181 current_function_decl = p->decl; 182 183 /* Reset variables that have known state during rtx generation. */ 184 virtuals_instantiated = 0; 185 generating_concat_p = 1; 186 } 187 188 /* Clear out all parts of the state in F that can safely be discarded 189 after the function has been parsed, but not compiled, to let 190 garbage collection reclaim the memory. */ 191 192 void 193 free_after_parsing (struct function *f) 194 { 195 f->language = 0; 196 } 197 198 /* Clear out all parts of the state in F that can safely be discarded 199 after the function has been compiled, to let garbage collection 200 reclaim the memory. */ 201 202 void 203 free_after_compilation (struct function *f) 204 { 205 prologue_insn_hash = NULL; 206 epilogue_insn_hash = NULL; 207 208 free (crtl->emit.regno_pointer_align); 209 210 memset (crtl, 0, sizeof (struct rtl_data)); 211 f->eh = NULL; 212 f->machine = NULL; 213 f->cfg = NULL; 214 f->curr_properties &= ~PROP_cfg; 215 216 regno_reg_rtx = NULL; 217 } 218 219 /* Return size needed for stack frame based on slots so far allocated. 220 This size counts from zero. It is not rounded to PREFERRED_STACK_BOUNDARY; 221 the caller may have to do that. */ 222 223 poly_int64 224 get_frame_size (void) 225 { 226 if (FRAME_GROWS_DOWNWARD) 227 return -frame_offset; 228 else 229 return frame_offset; 230 } 231 232 /* Issue an error message and return TRUE if frame OFFSET overflows in 233 the signed target pointer arithmetics for function FUNC. Otherwise 234 return FALSE. */ 235 236 bool 237 frame_offset_overflow (poly_int64 offset, tree func) 238 { 239 poly_uint64 size = FRAME_GROWS_DOWNWARD ? -offset : offset; 240 unsigned HOST_WIDE_INT limit 241 = ((HOST_WIDE_INT_1U << (GET_MODE_BITSIZE (Pmode) - 1)) 242 /* Leave room for the fixed part of the frame. */ 243 - 64 * UNITS_PER_WORD); 244 245 if (!coeffs_in_range_p (size, 0U, limit)) 246 { 247 error_at (DECL_SOURCE_LOCATION (func), 248 "total size of local objects too large"); 249 return true; 250 } 251 252 return false; 253 } 254 255 /* Return the minimum spill slot alignment for a register of mode MODE. */ 256 257 unsigned int 258 spill_slot_alignment (machine_mode mode ATTRIBUTE_UNUSED) 259 { 260 return STACK_SLOT_ALIGNMENT (NULL_TREE, mode, GET_MODE_ALIGNMENT (mode)); 261 } 262 263 /* Return stack slot alignment in bits for TYPE and MODE. */ 264 265 static unsigned int 266 get_stack_local_alignment (tree type, machine_mode mode) 267 { 268 unsigned int alignment; 269 270 if (mode == BLKmode) 271 alignment = BIGGEST_ALIGNMENT; 272 else 273 alignment = GET_MODE_ALIGNMENT (mode); 274 275 /* Allow the frond-end to (possibly) increase the alignment of this 276 stack slot. */ 277 if (! type) 278 type = lang_hooks.types.type_for_mode (mode, 0); 279 280 return STACK_SLOT_ALIGNMENT (type, mode, alignment); 281 } 282 283 /* Determine whether it is possible to fit a stack slot of size SIZE and 284 alignment ALIGNMENT into an area in the stack frame that starts at 285 frame offset START and has a length of LENGTH. If so, store the frame 286 offset to be used for the stack slot in *POFFSET and return true; 287 return false otherwise. This function will extend the frame size when 288 given a start/length pair that lies at the end of the frame. */ 289 290 static bool 291 try_fit_stack_local (poly_int64 start, poly_int64 length, 292 poly_int64 size, unsigned int alignment, 293 poly_int64_pod *poffset) 294 { 295 poly_int64 this_frame_offset; 296 int frame_off, frame_alignment, frame_phase; 297 298 /* Calculate how many bytes the start of local variables is off from 299 stack alignment. */ 300 frame_alignment = PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT; 301 frame_off = targetm.starting_frame_offset () % frame_alignment; 302 frame_phase = frame_off ? frame_alignment - frame_off : 0; 303 304 /* Round the frame offset to the specified alignment. */ 305 306 if (FRAME_GROWS_DOWNWARD) 307 this_frame_offset 308 = (aligned_lower_bound (start + length - size - frame_phase, alignment) 309 + frame_phase); 310 else 311 this_frame_offset 312 = aligned_upper_bound (start - frame_phase, alignment) + frame_phase; 313 314 /* See if it fits. If this space is at the edge of the frame, 315 consider extending the frame to make it fit. Our caller relies on 316 this when allocating a new slot. */ 317 if (maybe_lt (this_frame_offset, start)) 318 { 319 if (known_eq (frame_offset, start)) 320 frame_offset = this_frame_offset; 321 else 322 return false; 323 } 324 else if (maybe_gt (this_frame_offset + size, start + length)) 325 { 326 if (known_eq (frame_offset, start + length)) 327 frame_offset = this_frame_offset + size; 328 else 329 return false; 330 } 331 332 *poffset = this_frame_offset; 333 return true; 334 } 335 336 /* Create a new frame_space structure describing free space in the stack 337 frame beginning at START and ending at END, and chain it into the 338 function's frame_space_list. */ 339 340 static void 341 add_frame_space (poly_int64 start, poly_int64 end) 342 { 343 struct frame_space *space = ggc_alloc<frame_space> (); 344 space->next = crtl->frame_space_list; 345 crtl->frame_space_list = space; 346 space->start = start; 347 space->length = end - start; 348 } 349 350 /* Allocate a stack slot of SIZE bytes and return a MEM rtx for it 351 with machine mode MODE. 352 353 ALIGN controls the amount of alignment for the address of the slot: 354 0 means according to MODE, 355 -1 means use BIGGEST_ALIGNMENT and round size to multiple of that, 356 -2 means use BITS_PER_UNIT, 357 positive specifies alignment boundary in bits. 358 359 KIND has ASLK_REDUCE_ALIGN bit set if it is OK to reduce 360 alignment and ASLK_RECORD_PAD bit set if we should remember 361 extra space we allocated for alignment purposes. When we are 362 called from assign_stack_temp_for_type, it is not set so we don't 363 track the same stack slot in two independent lists. 364 365 We do not round to stack_boundary here. */ 366 367 rtx 368 assign_stack_local_1 (machine_mode mode, poly_int64 size, 369 int align, int kind) 370 { 371 rtx x, addr; 372 poly_int64 bigend_correction = 0; 373 poly_int64 slot_offset = 0, old_frame_offset; 374 unsigned int alignment, alignment_in_bits; 375 376 if (align == 0) 377 { 378 alignment = get_stack_local_alignment (NULL, mode); 379 alignment /= BITS_PER_UNIT; 380 } 381 else if (align == -1) 382 { 383 alignment = BIGGEST_ALIGNMENT / BITS_PER_UNIT; 384 size = aligned_upper_bound (size, alignment); 385 } 386 else if (align == -2) 387 alignment = 1; /* BITS_PER_UNIT / BITS_PER_UNIT */ 388 else 389 alignment = align / BITS_PER_UNIT; 390 391 alignment_in_bits = alignment * BITS_PER_UNIT; 392 393 /* Ignore alignment if it exceeds MAX_SUPPORTED_STACK_ALIGNMENT. */ 394 if (alignment_in_bits > MAX_SUPPORTED_STACK_ALIGNMENT) 395 { 396 alignment_in_bits = MAX_SUPPORTED_STACK_ALIGNMENT; 397 alignment = alignment_in_bits / BITS_PER_UNIT; 398 } 399 400 if (SUPPORTS_STACK_ALIGNMENT) 401 { 402 if (crtl->stack_alignment_estimated < alignment_in_bits) 403 { 404 if (!crtl->stack_realign_processed) 405 crtl->stack_alignment_estimated = alignment_in_bits; 406 else 407 { 408 /* If stack is realigned and stack alignment value 409 hasn't been finalized, it is OK not to increase 410 stack_alignment_estimated. The bigger alignment 411 requirement is recorded in stack_alignment_needed 412 below. */ 413 gcc_assert (!crtl->stack_realign_finalized); 414 if (!crtl->stack_realign_needed) 415 { 416 /* It is OK to reduce the alignment as long as the 417 requested size is 0 or the estimated stack 418 alignment >= mode alignment. */ 419 gcc_assert ((kind & ASLK_REDUCE_ALIGN) 420 || known_eq (size, 0) 421 || (crtl->stack_alignment_estimated 422 >= GET_MODE_ALIGNMENT (mode))); 423 alignment_in_bits = crtl->stack_alignment_estimated; 424 alignment = alignment_in_bits / BITS_PER_UNIT; 425 } 426 } 427 } 428 } 429 430 if (crtl->stack_alignment_needed < alignment_in_bits) 431 crtl->stack_alignment_needed = alignment_in_bits; 432 if (crtl->max_used_stack_slot_alignment < alignment_in_bits) 433 crtl->max_used_stack_slot_alignment = alignment_in_bits; 434 435 if (mode != BLKmode || maybe_ne (size, 0)) 436 { 437 if (kind & ASLK_RECORD_PAD) 438 { 439 struct frame_space **psp; 440 441 for (psp = &crtl->frame_space_list; *psp; psp = &(*psp)->next) 442 { 443 struct frame_space *space = *psp; 444 if (!try_fit_stack_local (space->start, space->length, size, 445 alignment, &slot_offset)) 446 continue; 447 *psp = space->next; 448 if (known_gt (slot_offset, space->start)) 449 add_frame_space (space->start, slot_offset); 450 if (known_lt (slot_offset + size, space->start + space->length)) 451 add_frame_space (slot_offset + size, 452 space->start + space->length); 453 goto found_space; 454 } 455 } 456 } 457 else if (!STACK_ALIGNMENT_NEEDED) 458 { 459 slot_offset = frame_offset; 460 goto found_space; 461 } 462 463 old_frame_offset = frame_offset; 464 465 if (FRAME_GROWS_DOWNWARD) 466 { 467 frame_offset -= size; 468 try_fit_stack_local (frame_offset, size, size, alignment, &slot_offset); 469 470 if (kind & ASLK_RECORD_PAD) 471 { 472 if (known_gt (slot_offset, frame_offset)) 473 add_frame_space (frame_offset, slot_offset); 474 if (known_lt (slot_offset + size, old_frame_offset)) 475 add_frame_space (slot_offset + size, old_frame_offset); 476 } 477 } 478 else 479 { 480 frame_offset += size; 481 try_fit_stack_local (old_frame_offset, size, size, alignment, &slot_offset); 482 483 if (kind & ASLK_RECORD_PAD) 484 { 485 if (known_gt (slot_offset, old_frame_offset)) 486 add_frame_space (old_frame_offset, slot_offset); 487 if (known_lt (slot_offset + size, frame_offset)) 488 add_frame_space (slot_offset + size, frame_offset); 489 } 490 } 491 492 found_space: 493 /* On a big-endian machine, if we are allocating more space than we will use, 494 use the least significant bytes of those that are allocated. */ 495 if (mode != BLKmode) 496 { 497 /* The slot size can sometimes be smaller than the mode size; 498 e.g. the rs6000 port allocates slots with a vector mode 499 that have the size of only one element. However, the slot 500 size must always be ordered wrt to the mode size, in the 501 same way as for a subreg. */ 502 gcc_checking_assert (ordered_p (GET_MODE_SIZE (mode), size)); 503 if (BYTES_BIG_ENDIAN && maybe_lt (GET_MODE_SIZE (mode), size)) 504 bigend_correction = size - GET_MODE_SIZE (mode); 505 } 506 507 /* If we have already instantiated virtual registers, return the actual 508 address relative to the frame pointer. */ 509 if (virtuals_instantiated) 510 addr = plus_constant (Pmode, frame_pointer_rtx, 511 trunc_int_for_mode 512 (slot_offset + bigend_correction 513 + targetm.starting_frame_offset (), Pmode)); 514 else 515 addr = plus_constant (Pmode, virtual_stack_vars_rtx, 516 trunc_int_for_mode 517 (slot_offset + bigend_correction, 518 Pmode)); 519 520 x = gen_rtx_MEM (mode, addr); 521 set_mem_align (x, alignment_in_bits); 522 MEM_NOTRAP_P (x) = 1; 523 524 vec_safe_push (stack_slot_list, x); 525 526 if (frame_offset_overflow (frame_offset, current_function_decl)) 527 frame_offset = 0; 528 529 return x; 530 } 531 532 /* Wrap up assign_stack_local_1 with last parameter as false. */ 533 534 rtx 535 assign_stack_local (machine_mode mode, poly_int64 size, int align) 536 { 537 return assign_stack_local_1 (mode, size, align, ASLK_RECORD_PAD); 538 } 539 540 /* In order to evaluate some expressions, such as function calls returning 541 structures in memory, we need to temporarily allocate stack locations. 542 We record each allocated temporary in the following structure. 543 544 Associated with each temporary slot is a nesting level. When we pop up 545 one level, all temporaries associated with the previous level are freed. 546 Normally, all temporaries are freed after the execution of the statement 547 in which they were created. However, if we are inside a ({...}) grouping, 548 the result may be in a temporary and hence must be preserved. If the 549 result could be in a temporary, we preserve it if we can determine which 550 one it is in. If we cannot determine which temporary may contain the 551 result, all temporaries are preserved. A temporary is preserved by 552 pretending it was allocated at the previous nesting level. */ 553 554 struct GTY(()) temp_slot { 555 /* Points to next temporary slot. */ 556 struct temp_slot *next; 557 /* Points to previous temporary slot. */ 558 struct temp_slot *prev; 559 /* The rtx to used to reference the slot. */ 560 rtx slot; 561 /* The size, in units, of the slot. */ 562 poly_int64 size; 563 /* The type of the object in the slot, or zero if it doesn't correspond 564 to a type. We use this to determine whether a slot can be reused. 565 It can be reused if objects of the type of the new slot will always 566 conflict with objects of the type of the old slot. */ 567 tree type; 568 /* The alignment (in bits) of the slot. */ 569 unsigned int align; 570 /* Nonzero if this temporary is currently in use. */ 571 char in_use; 572 /* Nesting level at which this slot is being used. */ 573 int level; 574 /* The offset of the slot from the frame_pointer, including extra space 575 for alignment. This info is for combine_temp_slots. */ 576 poly_int64 base_offset; 577 /* The size of the slot, including extra space for alignment. This 578 info is for combine_temp_slots. */ 579 poly_int64 full_size; 580 }; 581 582 /* Entry for the below hash table. */ 583 struct GTY((for_user)) temp_slot_address_entry { 584 hashval_t hash; 585 rtx address; 586 struct temp_slot *temp_slot; 587 }; 588 589 struct temp_address_hasher : ggc_ptr_hash<temp_slot_address_entry> 590 { 591 static hashval_t hash (temp_slot_address_entry *); 592 static bool equal (temp_slot_address_entry *, temp_slot_address_entry *); 593 }; 594 595 /* A table of addresses that represent a stack slot. The table is a mapping 596 from address RTXen to a temp slot. */ 597 static GTY(()) hash_table<temp_address_hasher> *temp_slot_address_table; 598 static size_t n_temp_slots_in_use; 599 600 /* Removes temporary slot TEMP from LIST. */ 601 602 static void 603 cut_slot_from_list (struct temp_slot *temp, struct temp_slot **list) 604 { 605 if (temp->next) 606 temp->next->prev = temp->prev; 607 if (temp->prev) 608 temp->prev->next = temp->next; 609 else 610 *list = temp->next; 611 612 temp->prev = temp->next = NULL; 613 } 614 615 /* Inserts temporary slot TEMP to LIST. */ 616 617 static void 618 insert_slot_to_list (struct temp_slot *temp, struct temp_slot **list) 619 { 620 temp->next = *list; 621 if (*list) 622 (*list)->prev = temp; 623 temp->prev = NULL; 624 *list = temp; 625 } 626 627 /* Returns the list of used temp slots at LEVEL. */ 628 629 static struct temp_slot ** 630 temp_slots_at_level (int level) 631 { 632 if (level >= (int) vec_safe_length (used_temp_slots)) 633 vec_safe_grow_cleared (used_temp_slots, level + 1); 634 635 return &(*used_temp_slots)[level]; 636 } 637 638 /* Returns the maximal temporary slot level. */ 639 640 static int 641 max_slot_level (void) 642 { 643 if (!used_temp_slots) 644 return -1; 645 646 return used_temp_slots->length () - 1; 647 } 648 649 /* Moves temporary slot TEMP to LEVEL. */ 650 651 static void 652 move_slot_to_level (struct temp_slot *temp, int level) 653 { 654 cut_slot_from_list (temp, temp_slots_at_level (temp->level)); 655 insert_slot_to_list (temp, temp_slots_at_level (level)); 656 temp->level = level; 657 } 658 659 /* Make temporary slot TEMP available. */ 660 661 static void 662 make_slot_available (struct temp_slot *temp) 663 { 664 cut_slot_from_list (temp, temp_slots_at_level (temp->level)); 665 insert_slot_to_list (temp, &avail_temp_slots); 666 temp->in_use = 0; 667 temp->level = -1; 668 n_temp_slots_in_use--; 669 } 670 671 /* Compute the hash value for an address -> temp slot mapping. 672 The value is cached on the mapping entry. */ 673 static hashval_t 674 temp_slot_address_compute_hash (struct temp_slot_address_entry *t) 675 { 676 int do_not_record = 0; 677 return hash_rtx (t->address, GET_MODE (t->address), 678 &do_not_record, NULL, false); 679 } 680 681 /* Return the hash value for an address -> temp slot mapping. */ 682 hashval_t 683 temp_address_hasher::hash (temp_slot_address_entry *t) 684 { 685 return t->hash; 686 } 687 688 /* Compare two address -> temp slot mapping entries. */ 689 bool 690 temp_address_hasher::equal (temp_slot_address_entry *t1, 691 temp_slot_address_entry *t2) 692 { 693 return exp_equiv_p (t1->address, t2->address, 0, true); 694 } 695 696 /* Add ADDRESS as an alias of TEMP_SLOT to the addess -> temp slot mapping. */ 697 static void 698 insert_temp_slot_address (rtx address, struct temp_slot *temp_slot) 699 { 700 struct temp_slot_address_entry *t = ggc_alloc<temp_slot_address_entry> (); 701 t->address = address; 702 t->temp_slot = temp_slot; 703 t->hash = temp_slot_address_compute_hash (t); 704 *temp_slot_address_table->find_slot_with_hash (t, t->hash, INSERT) = t; 705 } 706 707 /* Remove an address -> temp slot mapping entry if the temp slot is 708 not in use anymore. Callback for remove_unused_temp_slot_addresses. */ 709 int 710 remove_unused_temp_slot_addresses_1 (temp_slot_address_entry **slot, void *) 711 { 712 const struct temp_slot_address_entry *t = *slot; 713 if (! t->temp_slot->in_use) 714 temp_slot_address_table->clear_slot (slot); 715 return 1; 716 } 717 718 /* Remove all mappings of addresses to unused temp slots. */ 719 static void 720 remove_unused_temp_slot_addresses (void) 721 { 722 /* Use quicker clearing if there aren't any active temp slots. */ 723 if (n_temp_slots_in_use) 724 temp_slot_address_table->traverse 725 <void *, remove_unused_temp_slot_addresses_1> (NULL); 726 else 727 temp_slot_address_table->empty (); 728 } 729 730 /* Find the temp slot corresponding to the object at address X. */ 731 732 static struct temp_slot * 733 find_temp_slot_from_address (rtx x) 734 { 735 struct temp_slot *p; 736 struct temp_slot_address_entry tmp, *t; 737 738 /* First try the easy way: 739 See if X exists in the address -> temp slot mapping. */ 740 tmp.address = x; 741 tmp.temp_slot = NULL; 742 tmp.hash = temp_slot_address_compute_hash (&tmp); 743 t = temp_slot_address_table->find_with_hash (&tmp, tmp.hash); 744 if (t) 745 return t->temp_slot; 746 747 /* If we have a sum involving a register, see if it points to a temp 748 slot. */ 749 if (GET_CODE (x) == PLUS && REG_P (XEXP (x, 0)) 750 && (p = find_temp_slot_from_address (XEXP (x, 0))) != 0) 751 return p; 752 else if (GET_CODE (x) == PLUS && REG_P (XEXP (x, 1)) 753 && (p = find_temp_slot_from_address (XEXP (x, 1))) != 0) 754 return p; 755 756 /* Last resort: Address is a virtual stack var address. */ 757 poly_int64 offset; 758 if (strip_offset (x, &offset) == virtual_stack_vars_rtx) 759 { 760 int i; 761 for (i = max_slot_level (); i >= 0; i--) 762 for (p = *temp_slots_at_level (i); p; p = p->next) 763 if (known_in_range_p (offset, p->base_offset, p->full_size)) 764 return p; 765 } 766 767 return NULL; 768 } 769 770 /* Allocate a temporary stack slot and record it for possible later 771 reuse. 772 773 MODE is the machine mode to be given to the returned rtx. 774 775 SIZE is the size in units of the space required. We do no rounding here 776 since assign_stack_local will do any required rounding. 777 778 TYPE is the type that will be used for the stack slot. */ 779 780 rtx 781 assign_stack_temp_for_type (machine_mode mode, poly_int64 size, tree type) 782 { 783 unsigned int align; 784 struct temp_slot *p, *best_p = 0, *selected = NULL, **pp; 785 rtx slot; 786 787 gcc_assert (known_size_p (size)); 788 789 align = get_stack_local_alignment (type, mode); 790 791 /* Try to find an available, already-allocated temporary of the proper 792 mode which meets the size and alignment requirements. Choose the 793 smallest one with the closest alignment. 794 795 If assign_stack_temp is called outside of the tree->rtl expansion, 796 we cannot reuse the stack slots (that may still refer to 797 VIRTUAL_STACK_VARS_REGNUM). */ 798 if (!virtuals_instantiated) 799 { 800 for (p = avail_temp_slots; p; p = p->next) 801 { 802 if (p->align >= align 803 && known_ge (p->size, size) 804 && GET_MODE (p->slot) == mode 805 && objects_must_conflict_p (p->type, type) 806 && (best_p == 0 807 || (known_eq (best_p->size, p->size) 808 ? best_p->align > p->align 809 : known_ge (best_p->size, p->size)))) 810 { 811 if (p->align == align && known_eq (p->size, size)) 812 { 813 selected = p; 814 cut_slot_from_list (selected, &avail_temp_slots); 815 best_p = 0; 816 break; 817 } 818 best_p = p; 819 } 820 } 821 } 822 823 /* Make our best, if any, the one to use. */ 824 if (best_p) 825 { 826 selected = best_p; 827 cut_slot_from_list (selected, &avail_temp_slots); 828 829 /* If there are enough aligned bytes left over, make them into a new 830 temp_slot so that the extra bytes don't get wasted. Do this only 831 for BLKmode slots, so that we can be sure of the alignment. */ 832 if (GET_MODE (best_p->slot) == BLKmode) 833 { 834 int alignment = best_p->align / BITS_PER_UNIT; 835 poly_int64 rounded_size = aligned_upper_bound (size, alignment); 836 837 if (known_ge (best_p->size - rounded_size, alignment)) 838 { 839 p = ggc_alloc<temp_slot> (); 840 p->in_use = 0; 841 p->size = best_p->size - rounded_size; 842 p->base_offset = best_p->base_offset + rounded_size; 843 p->full_size = best_p->full_size - rounded_size; 844 p->slot = adjust_address_nv (best_p->slot, BLKmode, rounded_size); 845 p->align = best_p->align; 846 p->type = best_p->type; 847 insert_slot_to_list (p, &avail_temp_slots); 848 849 vec_safe_push (stack_slot_list, p->slot); 850 851 best_p->size = rounded_size; 852 best_p->full_size = rounded_size; 853 } 854 } 855 } 856 857 /* If we still didn't find one, make a new temporary. */ 858 if (selected == 0) 859 { 860 poly_int64 frame_offset_old = frame_offset; 861 862 p = ggc_alloc<temp_slot> (); 863 864 /* We are passing an explicit alignment request to assign_stack_local. 865 One side effect of that is assign_stack_local will not round SIZE 866 to ensure the frame offset remains suitably aligned. 867 868 So for requests which depended on the rounding of SIZE, we go ahead 869 and round it now. We also make sure ALIGNMENT is at least 870 BIGGEST_ALIGNMENT. */ 871 gcc_assert (mode != BLKmode || align == BIGGEST_ALIGNMENT); 872 p->slot = assign_stack_local_1 (mode, 873 (mode == BLKmode 874 ? aligned_upper_bound (size, 875 (int) align 876 / BITS_PER_UNIT) 877 : size), 878 align, 0); 879 880 p->align = align; 881 882 /* The following slot size computation is necessary because we don't 883 know the actual size of the temporary slot until assign_stack_local 884 has performed all the frame alignment and size rounding for the 885 requested temporary. Note that extra space added for alignment 886 can be either above or below this stack slot depending on which 887 way the frame grows. We include the extra space if and only if it 888 is above this slot. */ 889 if (FRAME_GROWS_DOWNWARD) 890 p->size = frame_offset_old - frame_offset; 891 else 892 p->size = size; 893 894 /* Now define the fields used by combine_temp_slots. */ 895 if (FRAME_GROWS_DOWNWARD) 896 { 897 p->base_offset = frame_offset; 898 p->full_size = frame_offset_old - frame_offset; 899 } 900 else 901 { 902 p->base_offset = frame_offset_old; 903 p->full_size = frame_offset - frame_offset_old; 904 } 905 906 selected = p; 907 } 908 909 p = selected; 910 p->in_use = 1; 911 p->type = type; 912 p->level = temp_slot_level; 913 n_temp_slots_in_use++; 914 915 pp = temp_slots_at_level (p->level); 916 insert_slot_to_list (p, pp); 917 insert_temp_slot_address (XEXP (p->slot, 0), p); 918 919 /* Create a new MEM rtx to avoid clobbering MEM flags of old slots. */ 920 slot = gen_rtx_MEM (mode, XEXP (p->slot, 0)); 921 vec_safe_push (stack_slot_list, slot); 922 923 /* If we know the alias set for the memory that will be used, use 924 it. If there's no TYPE, then we don't know anything about the 925 alias set for the memory. */ 926 set_mem_alias_set (slot, type ? get_alias_set (type) : 0); 927 set_mem_align (slot, align); 928 929 /* If a type is specified, set the relevant flags. */ 930 if (type != 0) 931 MEM_VOLATILE_P (slot) = TYPE_VOLATILE (type); 932 MEM_NOTRAP_P (slot) = 1; 933 934 return slot; 935 } 936 937 /* Allocate a temporary stack slot and record it for possible later 938 reuse. First two arguments are same as in preceding function. */ 939 940 rtx 941 assign_stack_temp (machine_mode mode, poly_int64 size) 942 { 943 return assign_stack_temp_for_type (mode, size, NULL_TREE); 944 } 945 946 /* Assign a temporary. 947 If TYPE_OR_DECL is a decl, then we are doing it on behalf of the decl 948 and so that should be used in error messages. In either case, we 949 allocate of the given type. 950 MEMORY_REQUIRED is 1 if the result must be addressable stack memory; 951 it is 0 if a register is OK. 952 DONT_PROMOTE is 1 if we should not promote values in register 953 to wider modes. */ 954 955 rtx 956 assign_temp (tree type_or_decl, int memory_required, 957 int dont_promote ATTRIBUTE_UNUSED) 958 { 959 tree type, decl; 960 machine_mode mode; 961 #ifdef PROMOTE_MODE 962 int unsignedp; 963 #endif 964 965 if (DECL_P (type_or_decl)) 966 decl = type_or_decl, type = TREE_TYPE (decl); 967 else 968 decl = NULL, type = type_or_decl; 969 970 mode = TYPE_MODE (type); 971 #ifdef PROMOTE_MODE 972 unsignedp = TYPE_UNSIGNED (type); 973 #endif 974 975 /* Allocating temporaries of TREE_ADDRESSABLE type must be done in the front 976 end. See also create_tmp_var for the gimplification-time check. */ 977 gcc_assert (!TREE_ADDRESSABLE (type) && COMPLETE_TYPE_P (type)); 978 979 if (mode == BLKmode || memory_required) 980 { 981 HOST_WIDE_INT size = int_size_in_bytes (type); 982 rtx tmp; 983 984 /* Zero sized arrays are GNU C extension. Set size to 1 to avoid 985 problems with allocating the stack space. */ 986 if (size == 0) 987 size = 1; 988 989 /* Unfortunately, we don't yet know how to allocate variable-sized 990 temporaries. However, sometimes we can find a fixed upper limit on 991 the size, so try that instead. */ 992 else if (size == -1) 993 size = max_int_size_in_bytes (type); 994 995 /* The size of the temporary may be too large to fit into an integer. */ 996 /* ??? Not sure this should happen except for user silliness, so limit 997 this to things that aren't compiler-generated temporaries. The 998 rest of the time we'll die in assign_stack_temp_for_type. */ 999 if (decl && size == -1 1000 && TREE_CODE (TYPE_SIZE_UNIT (type)) == INTEGER_CST) 1001 { 1002 error ("size of variable %q+D is too large", decl); 1003 size = 1; 1004 } 1005 1006 tmp = assign_stack_temp_for_type (mode, size, type); 1007 return tmp; 1008 } 1009 1010 #ifdef PROMOTE_MODE 1011 if (! dont_promote) 1012 mode = promote_mode (type, mode, &unsignedp); 1013 #endif 1014 1015 return gen_reg_rtx (mode); 1016 } 1017 1018 /* Combine temporary stack slots which are adjacent on the stack. 1019 1020 This allows for better use of already allocated stack space. This is only 1021 done for BLKmode slots because we can be sure that we won't have alignment 1022 problems in this case. */ 1023 1024 static void 1025 combine_temp_slots (void) 1026 { 1027 struct temp_slot *p, *q, *next, *next_q; 1028 int num_slots; 1029 1030 /* We can't combine slots, because the information about which slot 1031 is in which alias set will be lost. */ 1032 if (flag_strict_aliasing) 1033 return; 1034 1035 /* If there are a lot of temp slots, don't do anything unless 1036 high levels of optimization. */ 1037 if (! flag_expensive_optimizations) 1038 for (p = avail_temp_slots, num_slots = 0; p; p = p->next, num_slots++) 1039 if (num_slots > 100 || (num_slots > 10 && optimize == 0)) 1040 return; 1041 1042 for (p = avail_temp_slots; p; p = next) 1043 { 1044 int delete_p = 0; 1045 1046 next = p->next; 1047 1048 if (GET_MODE (p->slot) != BLKmode) 1049 continue; 1050 1051 for (q = p->next; q; q = next_q) 1052 { 1053 int delete_q = 0; 1054 1055 next_q = q->next; 1056 1057 if (GET_MODE (q->slot) != BLKmode) 1058 continue; 1059 1060 if (known_eq (p->base_offset + p->full_size, q->base_offset)) 1061 { 1062 /* Q comes after P; combine Q into P. */ 1063 p->size += q->size; 1064 p->full_size += q->full_size; 1065 delete_q = 1; 1066 } 1067 else if (known_eq (q->base_offset + q->full_size, p->base_offset)) 1068 { 1069 /* P comes after Q; combine P into Q. */ 1070 q->size += p->size; 1071 q->full_size += p->full_size; 1072 delete_p = 1; 1073 break; 1074 } 1075 if (delete_q) 1076 cut_slot_from_list (q, &avail_temp_slots); 1077 } 1078 1079 /* Either delete P or advance past it. */ 1080 if (delete_p) 1081 cut_slot_from_list (p, &avail_temp_slots); 1082 } 1083 } 1084 1085 /* Indicate that NEW_RTX is an alternate way of referring to the temp 1086 slot that previously was known by OLD_RTX. */ 1087 1088 void 1089 update_temp_slot_address (rtx old_rtx, rtx new_rtx) 1090 { 1091 struct temp_slot *p; 1092 1093 if (rtx_equal_p (old_rtx, new_rtx)) 1094 return; 1095 1096 p = find_temp_slot_from_address (old_rtx); 1097 1098 /* If we didn't find one, see if both OLD_RTX is a PLUS. If so, and 1099 NEW_RTX is a register, see if one operand of the PLUS is a 1100 temporary location. If so, NEW_RTX points into it. Otherwise, 1101 if both OLD_RTX and NEW_RTX are a PLUS and if there is a register 1102 in common between them. If so, try a recursive call on those 1103 values. */ 1104 if (p == 0) 1105 { 1106 if (GET_CODE (old_rtx) != PLUS) 1107 return; 1108 1109 if (REG_P (new_rtx)) 1110 { 1111 update_temp_slot_address (XEXP (old_rtx, 0), new_rtx); 1112 update_temp_slot_address (XEXP (old_rtx, 1), new_rtx); 1113 return; 1114 } 1115 else if (GET_CODE (new_rtx) != PLUS) 1116 return; 1117 1118 if (rtx_equal_p (XEXP (old_rtx, 0), XEXP (new_rtx, 0))) 1119 update_temp_slot_address (XEXP (old_rtx, 1), XEXP (new_rtx, 1)); 1120 else if (rtx_equal_p (XEXP (old_rtx, 1), XEXP (new_rtx, 0))) 1121 update_temp_slot_address (XEXP (old_rtx, 0), XEXP (new_rtx, 1)); 1122 else if (rtx_equal_p (XEXP (old_rtx, 0), XEXP (new_rtx, 1))) 1123 update_temp_slot_address (XEXP (old_rtx, 1), XEXP (new_rtx, 0)); 1124 else if (rtx_equal_p (XEXP (old_rtx, 1), XEXP (new_rtx, 1))) 1125 update_temp_slot_address (XEXP (old_rtx, 0), XEXP (new_rtx, 0)); 1126 1127 return; 1128 } 1129 1130 /* Otherwise add an alias for the temp's address. */ 1131 insert_temp_slot_address (new_rtx, p); 1132 } 1133 1134 /* If X could be a reference to a temporary slot, mark that slot as 1135 belonging to the to one level higher than the current level. If X 1136 matched one of our slots, just mark that one. Otherwise, we can't 1137 easily predict which it is, so upgrade all of them. 1138 1139 This is called when an ({...}) construct occurs and a statement 1140 returns a value in memory. */ 1141 1142 void 1143 preserve_temp_slots (rtx x) 1144 { 1145 struct temp_slot *p = 0, *next; 1146 1147 if (x == 0) 1148 return; 1149 1150 /* If X is a register that is being used as a pointer, see if we have 1151 a temporary slot we know it points to. */ 1152 if (REG_P (x) && REG_POINTER (x)) 1153 p = find_temp_slot_from_address (x); 1154 1155 /* If X is not in memory or is at a constant address, it cannot be in 1156 a temporary slot. */ 1157 if (p == 0 && (!MEM_P (x) || CONSTANT_P (XEXP (x, 0)))) 1158 return; 1159 1160 /* First see if we can find a match. */ 1161 if (p == 0) 1162 p = find_temp_slot_from_address (XEXP (x, 0)); 1163 1164 if (p != 0) 1165 { 1166 if (p->level == temp_slot_level) 1167 move_slot_to_level (p, temp_slot_level - 1); 1168 return; 1169 } 1170 1171 /* Otherwise, preserve all non-kept slots at this level. */ 1172 for (p = *temp_slots_at_level (temp_slot_level); p; p = next) 1173 { 1174 next = p->next; 1175 move_slot_to_level (p, temp_slot_level - 1); 1176 } 1177 } 1178 1179 /* Free all temporaries used so far. This is normally called at the 1180 end of generating code for a statement. */ 1181 1182 void 1183 free_temp_slots (void) 1184 { 1185 struct temp_slot *p, *next; 1186 bool some_available = false; 1187 1188 for (p = *temp_slots_at_level (temp_slot_level); p; p = next) 1189 { 1190 next = p->next; 1191 make_slot_available (p); 1192 some_available = true; 1193 } 1194 1195 if (some_available) 1196 { 1197 remove_unused_temp_slot_addresses (); 1198 combine_temp_slots (); 1199 } 1200 } 1201 1202 /* Push deeper into the nesting level for stack temporaries. */ 1203 1204 void 1205 push_temp_slots (void) 1206 { 1207 temp_slot_level++; 1208 } 1209 1210 /* Pop a temporary nesting level. All slots in use in the current level 1211 are freed. */ 1212 1213 void 1214 pop_temp_slots (void) 1215 { 1216 free_temp_slots (); 1217 temp_slot_level--; 1218 } 1219 1220 /* Initialize temporary slots. */ 1221 1222 void 1223 init_temp_slots (void) 1224 { 1225 /* We have not allocated any temporaries yet. */ 1226 avail_temp_slots = 0; 1227 vec_alloc (used_temp_slots, 0); 1228 temp_slot_level = 0; 1229 n_temp_slots_in_use = 0; 1230 1231 /* Set up the table to map addresses to temp slots. */ 1232 if (! temp_slot_address_table) 1233 temp_slot_address_table = hash_table<temp_address_hasher>::create_ggc (32); 1234 else 1235 temp_slot_address_table->empty (); 1236 } 1237 1238 /* Functions and data structures to keep track of the values hard regs 1239 had at the start of the function. */ 1240 1241 /* Private type used by get_hard_reg_initial_reg, get_hard_reg_initial_val, 1242 and has_hard_reg_initial_val.. */ 1243 struct GTY(()) initial_value_pair { 1244 rtx hard_reg; 1245 rtx pseudo; 1246 }; 1247 /* ??? This could be a VEC but there is currently no way to define an 1248 opaque VEC type. This could be worked around by defining struct 1249 initial_value_pair in function.h. */ 1250 struct GTY(()) initial_value_struct { 1251 int num_entries; 1252 int max_entries; 1253 initial_value_pair * GTY ((length ("%h.num_entries"))) entries; 1254 }; 1255 1256 /* If a pseudo represents an initial hard reg (or expression), return 1257 it, else return NULL_RTX. */ 1258 1259 rtx 1260 get_hard_reg_initial_reg (rtx reg) 1261 { 1262 struct initial_value_struct *ivs = crtl->hard_reg_initial_vals; 1263 int i; 1264 1265 if (ivs == 0) 1266 return NULL_RTX; 1267 1268 for (i = 0; i < ivs->num_entries; i++) 1269 if (rtx_equal_p (ivs->entries[i].pseudo, reg)) 1270 return ivs->entries[i].hard_reg; 1271 1272 return NULL_RTX; 1273 } 1274 1275 /* Make sure that there's a pseudo register of mode MODE that stores the 1276 initial value of hard register REGNO. Return an rtx for such a pseudo. */ 1277 1278 rtx 1279 get_hard_reg_initial_val (machine_mode mode, unsigned int regno) 1280 { 1281 struct initial_value_struct *ivs; 1282 rtx rv; 1283 1284 rv = has_hard_reg_initial_val (mode, regno); 1285 if (rv) 1286 return rv; 1287 1288 ivs = crtl->hard_reg_initial_vals; 1289 if (ivs == 0) 1290 { 1291 ivs = ggc_alloc<initial_value_struct> (); 1292 ivs->num_entries = 0; 1293 ivs->max_entries = 5; 1294 ivs->entries = ggc_vec_alloc<initial_value_pair> (5); 1295 crtl->hard_reg_initial_vals = ivs; 1296 } 1297 1298 if (ivs->num_entries >= ivs->max_entries) 1299 { 1300 ivs->max_entries += 5; 1301 ivs->entries = GGC_RESIZEVEC (initial_value_pair, ivs->entries, 1302 ivs->max_entries); 1303 } 1304 1305 ivs->entries[ivs->num_entries].hard_reg = gen_rtx_REG (mode, regno); 1306 ivs->entries[ivs->num_entries].pseudo = gen_reg_rtx (mode); 1307 1308 return ivs->entries[ivs->num_entries++].pseudo; 1309 } 1310 1311 /* See if get_hard_reg_initial_val has been used to create a pseudo 1312 for the initial value of hard register REGNO in mode MODE. Return 1313 the associated pseudo if so, otherwise return NULL. */ 1314 1315 rtx 1316 has_hard_reg_initial_val (machine_mode mode, unsigned int regno) 1317 { 1318 struct initial_value_struct *ivs; 1319 int i; 1320 1321 ivs = crtl->hard_reg_initial_vals; 1322 if (ivs != 0) 1323 for (i = 0; i < ivs->num_entries; i++) 1324 if (GET_MODE (ivs->entries[i].hard_reg) == mode 1325 && REGNO (ivs->entries[i].hard_reg) == regno) 1326 return ivs->entries[i].pseudo; 1327 1328 return NULL_RTX; 1329 } 1330 1331 unsigned int 1332 emit_initial_value_sets (void) 1333 { 1334 struct initial_value_struct *ivs = crtl->hard_reg_initial_vals; 1335 int i; 1336 rtx_insn *seq; 1337 1338 if (ivs == 0) 1339 return 0; 1340 1341 start_sequence (); 1342 for (i = 0; i < ivs->num_entries; i++) 1343 emit_move_insn (ivs->entries[i].pseudo, ivs->entries[i].hard_reg); 1344 seq = get_insns (); 1345 end_sequence (); 1346 1347 emit_insn_at_entry (seq); 1348 return 0; 1349 } 1350 1351 /* Return the hardreg-pseudoreg initial values pair entry I and 1352 TRUE if I is a valid entry, or FALSE if I is not a valid entry. */ 1353 bool 1354 initial_value_entry (int i, rtx *hreg, rtx *preg) 1355 { 1356 struct initial_value_struct *ivs = crtl->hard_reg_initial_vals; 1357 if (!ivs || i >= ivs->num_entries) 1358 return false; 1359 1360 *hreg = ivs->entries[i].hard_reg; 1361 *preg = ivs->entries[i].pseudo; 1362 return true; 1363 } 1364 1365 /* These routines are responsible for converting virtual register references 1366 to the actual hard register references once RTL generation is complete. 1367 1368 The following four variables are used for communication between the 1369 routines. They contain the offsets of the virtual registers from their 1370 respective hard registers. */ 1371 1372 static poly_int64 in_arg_offset; 1373 static poly_int64 var_offset; 1374 static poly_int64 dynamic_offset; 1375 static poly_int64 out_arg_offset; 1376 static poly_int64 cfa_offset; 1377 1378 /* In most machines, the stack pointer register is equivalent to the bottom 1379 of the stack. */ 1380 1381 #ifndef STACK_POINTER_OFFSET 1382 #define STACK_POINTER_OFFSET 0 1383 #endif 1384 1385 #if defined (REG_PARM_STACK_SPACE) && !defined (INCOMING_REG_PARM_STACK_SPACE) 1386 #define INCOMING_REG_PARM_STACK_SPACE REG_PARM_STACK_SPACE 1387 #endif 1388 1389 /* If not defined, pick an appropriate default for the offset of dynamically 1390 allocated memory depending on the value of ACCUMULATE_OUTGOING_ARGS, 1391 INCOMING_REG_PARM_STACK_SPACE, and OUTGOING_REG_PARM_STACK_SPACE. */ 1392 1393 #ifndef STACK_DYNAMIC_OFFSET 1394 1395 /* The bottom of the stack points to the actual arguments. If 1396 REG_PARM_STACK_SPACE is defined, this includes the space for the register 1397 parameters. However, if OUTGOING_REG_PARM_STACK space is not defined, 1398 stack space for register parameters is not pushed by the caller, but 1399 rather part of the fixed stack areas and hence not included in 1400 `crtl->outgoing_args_size'. Nevertheless, we must allow 1401 for it when allocating stack dynamic objects. */ 1402 1403 #ifdef INCOMING_REG_PARM_STACK_SPACE 1404 #define STACK_DYNAMIC_OFFSET(FNDECL) \ 1405 ((ACCUMULATE_OUTGOING_ARGS \ 1406 ? (crtl->outgoing_args_size \ 1407 + (OUTGOING_REG_PARM_STACK_SPACE ((!(FNDECL) ? NULL_TREE : TREE_TYPE (FNDECL))) ? 0 \ 1408 : INCOMING_REG_PARM_STACK_SPACE (FNDECL))) \ 1409 : 0) + (STACK_POINTER_OFFSET)) 1410 #else 1411 #define STACK_DYNAMIC_OFFSET(FNDECL) \ 1412 ((ACCUMULATE_OUTGOING_ARGS ? crtl->outgoing_args_size : poly_int64 (0)) \ 1413 + (STACK_POINTER_OFFSET)) 1414 #endif 1415 #endif 1416 1417 1418 /* Given a piece of RTX and a pointer to a HOST_WIDE_INT, if the RTX 1419 is a virtual register, return the equivalent hard register and set the 1420 offset indirectly through the pointer. Otherwise, return 0. */ 1421 1422 static rtx 1423 instantiate_new_reg (rtx x, poly_int64_pod *poffset) 1424 { 1425 rtx new_rtx; 1426 poly_int64 offset; 1427 1428 if (x == virtual_incoming_args_rtx) 1429 { 1430 if (stack_realign_drap) 1431 { 1432 /* Replace virtual_incoming_args_rtx with internal arg 1433 pointer if DRAP is used to realign stack. */ 1434 new_rtx = crtl->args.internal_arg_pointer; 1435 offset = 0; 1436 } 1437 else 1438 new_rtx = arg_pointer_rtx, offset = in_arg_offset; 1439 } 1440 else if (x == virtual_stack_vars_rtx) 1441 new_rtx = frame_pointer_rtx, offset = var_offset; 1442 else if (x == virtual_stack_dynamic_rtx) 1443 new_rtx = stack_pointer_rtx, offset = dynamic_offset; 1444 else if (x == virtual_outgoing_args_rtx) 1445 new_rtx = stack_pointer_rtx, offset = out_arg_offset; 1446 else if (x == virtual_cfa_rtx) 1447 { 1448 #ifdef FRAME_POINTER_CFA_OFFSET 1449 new_rtx = frame_pointer_rtx; 1450 #else 1451 new_rtx = arg_pointer_rtx; 1452 #endif 1453 offset = cfa_offset; 1454 } 1455 else if (x == virtual_preferred_stack_boundary_rtx) 1456 { 1457 new_rtx = GEN_INT (crtl->preferred_stack_boundary / BITS_PER_UNIT); 1458 offset = 0; 1459 } 1460 else 1461 return NULL_RTX; 1462 1463 *poffset = offset; 1464 return new_rtx; 1465 } 1466 1467 /* A subroutine of instantiate_virtual_regs. Instantiate any virtual 1468 registers present inside of *LOC. The expression is simplified, 1469 as much as possible, but is not to be considered "valid" in any sense 1470 implied by the target. Return true if any change is made. */ 1471 1472 static bool 1473 instantiate_virtual_regs_in_rtx (rtx *loc) 1474 { 1475 if (!*loc) 1476 return false; 1477 bool changed = false; 1478 subrtx_ptr_iterator::array_type array; 1479 FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) 1480 { 1481 rtx *loc = *iter; 1482 if (rtx x = *loc) 1483 { 1484 rtx new_rtx; 1485 poly_int64 offset; 1486 switch (GET_CODE (x)) 1487 { 1488 case REG: 1489 new_rtx = instantiate_new_reg (x, &offset); 1490 if (new_rtx) 1491 { 1492 *loc = plus_constant (GET_MODE (x), new_rtx, offset); 1493 changed = true; 1494 } 1495 iter.skip_subrtxes (); 1496 break; 1497 1498 case PLUS: 1499 new_rtx = instantiate_new_reg (XEXP (x, 0), &offset); 1500 if (new_rtx) 1501 { 1502 XEXP (x, 0) = new_rtx; 1503 *loc = plus_constant (GET_MODE (x), x, offset, true); 1504 changed = true; 1505 iter.skip_subrtxes (); 1506 break; 1507 } 1508 1509 /* FIXME -- from old code */ 1510 /* If we have (plus (subreg (virtual-reg)) (const_int)), we know 1511 we can commute the PLUS and SUBREG because pointers into the 1512 frame are well-behaved. */ 1513 break; 1514 1515 default: 1516 break; 1517 } 1518 } 1519 } 1520 return changed; 1521 } 1522 1523 /* A subroutine of instantiate_virtual_regs_in_insn. Return true if X 1524 matches the predicate for insn CODE operand OPERAND. */ 1525 1526 static int 1527 safe_insn_predicate (int code, int operand, rtx x) 1528 { 1529 return code < 0 || insn_operand_matches ((enum insn_code) code, operand, x); 1530 } 1531 1532 /* A subroutine of instantiate_virtual_regs. Instantiate any virtual 1533 registers present inside of insn. The result will be a valid insn. */ 1534 1535 static void 1536 instantiate_virtual_regs_in_insn (rtx_insn *insn) 1537 { 1538 poly_int64 offset; 1539 int insn_code, i; 1540 bool any_change = false; 1541 rtx set, new_rtx, x; 1542 rtx_insn *seq; 1543 1544 /* There are some special cases to be handled first. */ 1545 set = single_set (insn); 1546 if (set) 1547 { 1548 /* We're allowed to assign to a virtual register. This is interpreted 1549 to mean that the underlying register gets assigned the inverse 1550 transformation. This is used, for example, in the handling of 1551 non-local gotos. */ 1552 new_rtx = instantiate_new_reg (SET_DEST (set), &offset); 1553 if (new_rtx) 1554 { 1555 start_sequence (); 1556 1557 instantiate_virtual_regs_in_rtx (&SET_SRC (set)); 1558 x = simplify_gen_binary (PLUS, GET_MODE (new_rtx), SET_SRC (set), 1559 gen_int_mode (-offset, GET_MODE (new_rtx))); 1560 x = force_operand (x, new_rtx); 1561 if (x != new_rtx) 1562 emit_move_insn (new_rtx, x); 1563 1564 seq = get_insns (); 1565 end_sequence (); 1566 1567 emit_insn_before (seq, insn); 1568 delete_insn (insn); 1569 return; 1570 } 1571 1572 /* Handle a straight copy from a virtual register by generating a 1573 new add insn. The difference between this and falling through 1574 to the generic case is avoiding a new pseudo and eliminating a 1575 move insn in the initial rtl stream. */ 1576 new_rtx = instantiate_new_reg (SET_SRC (set), &offset); 1577 if (new_rtx 1578 && maybe_ne (offset, 0) 1579 && REG_P (SET_DEST (set)) 1580 && REGNO (SET_DEST (set)) > LAST_VIRTUAL_REGISTER) 1581 { 1582 start_sequence (); 1583 1584 x = expand_simple_binop (GET_MODE (SET_DEST (set)), PLUS, new_rtx, 1585 gen_int_mode (offset, 1586 GET_MODE (SET_DEST (set))), 1587 SET_DEST (set), 1, OPTAB_LIB_WIDEN); 1588 if (x != SET_DEST (set)) 1589 emit_move_insn (SET_DEST (set), x); 1590 1591 seq = get_insns (); 1592 end_sequence (); 1593 1594 emit_insn_before (seq, insn); 1595 delete_insn (insn); 1596 return; 1597 } 1598 1599 extract_insn (insn); 1600 insn_code = INSN_CODE (insn); 1601 1602 /* Handle a plus involving a virtual register by determining if the 1603 operands remain valid if they're modified in place. */ 1604 poly_int64 delta; 1605 if (GET_CODE (SET_SRC (set)) == PLUS 1606 && recog_data.n_operands >= 3 1607 && recog_data.operand_loc[1] == &XEXP (SET_SRC (set), 0) 1608 && recog_data.operand_loc[2] == &XEXP (SET_SRC (set), 1) 1609 && poly_int_rtx_p (recog_data.operand[2], &delta) 1610 && (new_rtx = instantiate_new_reg (recog_data.operand[1], &offset))) 1611 { 1612 offset += delta; 1613 1614 /* If the sum is zero, then replace with a plain move. */ 1615 if (known_eq (offset, 0) 1616 && REG_P (SET_DEST (set)) 1617 && REGNO (SET_DEST (set)) > LAST_VIRTUAL_REGISTER) 1618 { 1619 start_sequence (); 1620 emit_move_insn (SET_DEST (set), new_rtx); 1621 seq = get_insns (); 1622 end_sequence (); 1623 1624 emit_insn_before (seq, insn); 1625 delete_insn (insn); 1626 return; 1627 } 1628 1629 x = gen_int_mode (offset, recog_data.operand_mode[2]); 1630 1631 /* Using validate_change and apply_change_group here leaves 1632 recog_data in an invalid state. Since we know exactly what 1633 we want to check, do those two by hand. */ 1634 if (safe_insn_predicate (insn_code, 1, new_rtx) 1635 && safe_insn_predicate (insn_code, 2, x)) 1636 { 1637 *recog_data.operand_loc[1] = recog_data.operand[1] = new_rtx; 1638 *recog_data.operand_loc[2] = recog_data.operand[2] = x; 1639 any_change = true; 1640 1641 /* Fall through into the regular operand fixup loop in 1642 order to take care of operands other than 1 and 2. */ 1643 } 1644 } 1645 } 1646 else 1647 { 1648 extract_insn (insn); 1649 insn_code = INSN_CODE (insn); 1650 } 1651 1652 /* In the general case, we expect virtual registers to appear only in 1653 operands, and then only as either bare registers or inside memories. */ 1654 for (i = 0; i < recog_data.n_operands; ++i) 1655 { 1656 x = recog_data.operand[i]; 1657 switch (GET_CODE (x)) 1658 { 1659 case MEM: 1660 { 1661 rtx addr = XEXP (x, 0); 1662 1663 if (!instantiate_virtual_regs_in_rtx (&addr)) 1664 continue; 1665 1666 start_sequence (); 1667 x = replace_equiv_address (x, addr, true); 1668 /* It may happen that the address with the virtual reg 1669 was valid (e.g. based on the virtual stack reg, which might 1670 be acceptable to the predicates with all offsets), whereas 1671 the address now isn't anymore, for instance when the address 1672 is still offsetted, but the base reg isn't virtual-stack-reg 1673 anymore. Below we would do a force_reg on the whole operand, 1674 but this insn might actually only accept memory. Hence, 1675 before doing that last resort, try to reload the address into 1676 a register, so this operand stays a MEM. */ 1677 if (!safe_insn_predicate (insn_code, i, x)) 1678 { 1679 addr = force_reg (GET_MODE (addr), addr); 1680 x = replace_equiv_address (x, addr, true); 1681 } 1682 seq = get_insns (); 1683 end_sequence (); 1684 if (seq) 1685 emit_insn_before (seq, insn); 1686 } 1687 break; 1688 1689 case REG: 1690 new_rtx = instantiate_new_reg (x, &offset); 1691 if (new_rtx == NULL) 1692 continue; 1693 if (known_eq (offset, 0)) 1694 x = new_rtx; 1695 else 1696 { 1697 start_sequence (); 1698 1699 /* Careful, special mode predicates may have stuff in 1700 insn_data[insn_code].operand[i].mode that isn't useful 1701 to us for computing a new value. */ 1702 /* ??? Recognize address_operand and/or "p" constraints 1703 to see if (plus new offset) is a valid before we put 1704 this through expand_simple_binop. */ 1705 x = expand_simple_binop (GET_MODE (x), PLUS, new_rtx, 1706 gen_int_mode (offset, GET_MODE (x)), 1707 NULL_RTX, 1, OPTAB_LIB_WIDEN); 1708 seq = get_insns (); 1709 end_sequence (); 1710 emit_insn_before (seq, insn); 1711 } 1712 break; 1713 1714 case SUBREG: 1715 new_rtx = instantiate_new_reg (SUBREG_REG (x), &offset); 1716 if (new_rtx == NULL) 1717 continue; 1718 if (maybe_ne (offset, 0)) 1719 { 1720 start_sequence (); 1721 new_rtx = expand_simple_binop 1722 (GET_MODE (new_rtx), PLUS, new_rtx, 1723 gen_int_mode (offset, GET_MODE (new_rtx)), 1724 NULL_RTX, 1, OPTAB_LIB_WIDEN); 1725 seq = get_insns (); 1726 end_sequence (); 1727 emit_insn_before (seq, insn); 1728 } 1729 x = simplify_gen_subreg (recog_data.operand_mode[i], new_rtx, 1730 GET_MODE (new_rtx), SUBREG_BYTE (x)); 1731 gcc_assert (x); 1732 break; 1733 1734 default: 1735 continue; 1736 } 1737 1738 /* At this point, X contains the new value for the operand. 1739 Validate the new value vs the insn predicate. Note that 1740 asm insns will have insn_code -1 here. */ 1741 if (!safe_insn_predicate (insn_code, i, x)) 1742 { 1743 start_sequence (); 1744 if (REG_P (x)) 1745 { 1746 gcc_assert (REGNO (x) <= LAST_VIRTUAL_REGISTER); 1747 x = copy_to_reg (x); 1748 } 1749 else 1750 x = force_reg (insn_data[insn_code].operand[i].mode, x); 1751 seq = get_insns (); 1752 end_sequence (); 1753 if (seq) 1754 emit_insn_before (seq, insn); 1755 } 1756 1757 *recog_data.operand_loc[i] = recog_data.operand[i] = x; 1758 any_change = true; 1759 } 1760 1761 if (any_change) 1762 { 1763 /* Propagate operand changes into the duplicates. */ 1764 for (i = 0; i < recog_data.n_dups; ++i) 1765 *recog_data.dup_loc[i] 1766 = copy_rtx (recog_data.operand[(unsigned)recog_data.dup_num[i]]); 1767 1768 /* Force re-recognition of the instruction for validation. */ 1769 INSN_CODE (insn) = -1; 1770 } 1771 1772 if (asm_noperands (PATTERN (insn)) >= 0) 1773 { 1774 if (!check_asm_operands (PATTERN (insn))) 1775 { 1776 error_for_asm (insn, "impossible constraint in %<asm%>"); 1777 /* For asm goto, instead of fixing up all the edges 1778 just clear the template and clear input operands 1779 (asm goto doesn't have any output operands). */ 1780 if (JUMP_P (insn)) 1781 { 1782 rtx asm_op = extract_asm_operands (PATTERN (insn)); 1783 ASM_OPERANDS_TEMPLATE (asm_op) = ggc_strdup (""); 1784 ASM_OPERANDS_INPUT_VEC (asm_op) = rtvec_alloc (0); 1785 ASM_OPERANDS_INPUT_CONSTRAINT_VEC (asm_op) = rtvec_alloc (0); 1786 } 1787 else 1788 delete_insn (insn); 1789 } 1790 } 1791 else 1792 { 1793 if (recog_memoized (insn) < 0) 1794 fatal_insn_not_found (insn); 1795 } 1796 } 1797 1798 /* Subroutine of instantiate_decls. Given RTL representing a decl, 1799 do any instantiation required. */ 1800 1801 void 1802 instantiate_decl_rtl (rtx x) 1803 { 1804 rtx addr; 1805 1806 if (x == 0) 1807 return; 1808 1809 /* If this is a CONCAT, recurse for the pieces. */ 1810 if (GET_CODE (x) == CONCAT) 1811 { 1812 instantiate_decl_rtl (XEXP (x, 0)); 1813 instantiate_decl_rtl (XEXP (x, 1)); 1814 return; 1815 } 1816 1817 /* If this is not a MEM, no need to do anything. Similarly if the 1818 address is a constant or a register that is not a virtual register. */ 1819 if (!MEM_P (x)) 1820 return; 1821 1822 addr = XEXP (x, 0); 1823 if (CONSTANT_P (addr) 1824 || (REG_P (addr) 1825 && (REGNO (addr) < FIRST_VIRTUAL_REGISTER 1826 || REGNO (addr) > LAST_VIRTUAL_REGISTER))) 1827 return; 1828 1829 instantiate_virtual_regs_in_rtx (&XEXP (x, 0)); 1830 } 1831 1832 /* Helper for instantiate_decls called via walk_tree: Process all decls 1833 in the given DECL_VALUE_EXPR. */ 1834 1835 static tree 1836 instantiate_expr (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED) 1837 { 1838 tree t = *tp; 1839 if (! EXPR_P (t)) 1840 { 1841 *walk_subtrees = 0; 1842 if (DECL_P (t)) 1843 { 1844 if (DECL_RTL_SET_P (t)) 1845 instantiate_decl_rtl (DECL_RTL (t)); 1846 if (TREE_CODE (t) == PARM_DECL && DECL_NAMELESS (t) 1847 && DECL_INCOMING_RTL (t)) 1848 instantiate_decl_rtl (DECL_INCOMING_RTL (t)); 1849 if ((VAR_P (t) || TREE_CODE (t) == RESULT_DECL) 1850 && DECL_HAS_VALUE_EXPR_P (t)) 1851 { 1852 tree v = DECL_VALUE_EXPR (t); 1853 walk_tree (&v, instantiate_expr, NULL, NULL); 1854 } 1855 } 1856 } 1857 return NULL; 1858 } 1859 1860 /* Subroutine of instantiate_decls: Process all decls in the given 1861 BLOCK node and all its subblocks. */ 1862 1863 static void 1864 instantiate_decls_1 (tree let) 1865 { 1866 tree t; 1867 1868 for (t = BLOCK_VARS (let); t; t = DECL_CHAIN (t)) 1869 { 1870 if (DECL_RTL_SET_P (t)) 1871 instantiate_decl_rtl (DECL_RTL (t)); 1872 if (VAR_P (t) && DECL_HAS_VALUE_EXPR_P (t)) 1873 { 1874 tree v = DECL_VALUE_EXPR (t); 1875 walk_tree (&v, instantiate_expr, NULL, NULL); 1876 } 1877 } 1878 1879 /* Process all subblocks. */ 1880 for (t = BLOCK_SUBBLOCKS (let); t; t = BLOCK_CHAIN (t)) 1881 instantiate_decls_1 (t); 1882 } 1883 1884 /* Scan all decls in FNDECL (both variables and parameters) and instantiate 1885 all virtual registers in their DECL_RTL's. */ 1886 1887 static void 1888 instantiate_decls (tree fndecl) 1889 { 1890 tree decl; 1891 unsigned ix; 1892 1893 /* Process all parameters of the function. */ 1894 for (decl = DECL_ARGUMENTS (fndecl); decl; decl = DECL_CHAIN (decl)) 1895 { 1896 instantiate_decl_rtl (DECL_RTL (decl)); 1897 instantiate_decl_rtl (DECL_INCOMING_RTL (decl)); 1898 if (DECL_HAS_VALUE_EXPR_P (decl)) 1899 { 1900 tree v = DECL_VALUE_EXPR (decl); 1901 walk_tree (&v, instantiate_expr, NULL, NULL); 1902 } 1903 } 1904 1905 if ((decl = DECL_RESULT (fndecl)) 1906 && TREE_CODE (decl) == RESULT_DECL) 1907 { 1908 if (DECL_RTL_SET_P (decl)) 1909 instantiate_decl_rtl (DECL_RTL (decl)); 1910 if (DECL_HAS_VALUE_EXPR_P (decl)) 1911 { 1912 tree v = DECL_VALUE_EXPR (decl); 1913 walk_tree (&v, instantiate_expr, NULL, NULL); 1914 } 1915 } 1916 1917 /* Process the saved static chain if it exists. */ 1918 decl = DECL_STRUCT_FUNCTION (fndecl)->static_chain_decl; 1919 if (decl && DECL_HAS_VALUE_EXPR_P (decl)) 1920 instantiate_decl_rtl (DECL_RTL (DECL_VALUE_EXPR (decl))); 1921 1922 /* Now process all variables defined in the function or its subblocks. */ 1923 if (DECL_INITIAL (fndecl)) 1924 instantiate_decls_1 (DECL_INITIAL (fndecl)); 1925 1926 FOR_EACH_LOCAL_DECL (cfun, ix, decl) 1927 if (DECL_RTL_SET_P (decl)) 1928 instantiate_decl_rtl (DECL_RTL (decl)); 1929 vec_free (cfun->local_decls); 1930 } 1931 1932 /* Pass through the INSNS of function FNDECL and convert virtual register 1933 references to hard register references. */ 1934 1935 static unsigned int 1936 instantiate_virtual_regs (void) 1937 { 1938 rtx_insn *insn; 1939 1940 /* Compute the offsets to use for this function. */ 1941 in_arg_offset = FIRST_PARM_OFFSET (current_function_decl); 1942 var_offset = targetm.starting_frame_offset (); 1943 dynamic_offset = STACK_DYNAMIC_OFFSET (current_function_decl); 1944 out_arg_offset = STACK_POINTER_OFFSET; 1945 #ifdef FRAME_POINTER_CFA_OFFSET 1946 cfa_offset = FRAME_POINTER_CFA_OFFSET (current_function_decl); 1947 #else 1948 cfa_offset = ARG_POINTER_CFA_OFFSET (current_function_decl); 1949 #endif 1950 1951 /* Initialize recognition, indicating that volatile is OK. */ 1952 init_recog (); 1953 1954 /* Scan through all the insns, instantiating every virtual register still 1955 present. */ 1956 for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) 1957 if (INSN_P (insn)) 1958 { 1959 /* These patterns in the instruction stream can never be recognized. 1960 Fortunately, they shouldn't contain virtual registers either. */ 1961 if (GET_CODE (PATTERN (insn)) == USE 1962 || GET_CODE (PATTERN (insn)) == CLOBBER 1963 || GET_CODE (PATTERN (insn)) == ASM_INPUT 1964 || DEBUG_MARKER_INSN_P (insn)) 1965 continue; 1966 else if (DEBUG_BIND_INSN_P (insn)) 1967 instantiate_virtual_regs_in_rtx (INSN_VAR_LOCATION_PTR (insn)); 1968 else 1969 instantiate_virtual_regs_in_insn (insn); 1970 1971 if (insn->deleted ()) 1972 continue; 1973 1974 instantiate_virtual_regs_in_rtx (®_NOTES (insn)); 1975 1976 /* Instantiate any virtual registers in CALL_INSN_FUNCTION_USAGE. */ 1977 if (CALL_P (insn)) 1978 instantiate_virtual_regs_in_rtx (&CALL_INSN_FUNCTION_USAGE (insn)); 1979 } 1980 1981 /* Instantiate the virtual registers in the DECLs for debugging purposes. */ 1982 instantiate_decls (current_function_decl); 1983 1984 targetm.instantiate_decls (); 1985 1986 /* Indicate that, from now on, assign_stack_local should use 1987 frame_pointer_rtx. */ 1988 virtuals_instantiated = 1; 1989 1990 return 0; 1991 } 1992 1993 namespace { 1994 1995 const pass_data pass_data_instantiate_virtual_regs = 1996 { 1997 RTL_PASS, /* type */ 1998 "vregs", /* name */ 1999 OPTGROUP_NONE, /* optinfo_flags */ 2000 TV_NONE, /* tv_id */ 2001 0, /* properties_required */ 2002 0, /* properties_provided */ 2003 0, /* properties_destroyed */ 2004 0, /* todo_flags_start */ 2005 0, /* todo_flags_finish */ 2006 }; 2007 2008 class pass_instantiate_virtual_regs : public rtl_opt_pass 2009 { 2010 public: 2011 pass_instantiate_virtual_regs (gcc::context *ctxt) 2012 : rtl_opt_pass (pass_data_instantiate_virtual_regs, ctxt) 2013 {} 2014 2015 /* opt_pass methods: */ 2016 virtual unsigned int execute (function *) 2017 { 2018 return instantiate_virtual_regs (); 2019 } 2020 2021 }; // class pass_instantiate_virtual_regs 2022 2023 } // anon namespace 2024 2025 rtl_opt_pass * 2026 make_pass_instantiate_virtual_regs (gcc::context *ctxt) 2027 { 2028 return new pass_instantiate_virtual_regs (ctxt); 2029 } 2030 2031 2032 /* Return 1 if EXP is an aggregate type (or a value with aggregate type). 2033 This means a type for which function calls must pass an address to the 2034 function or get an address back from the function. 2035 EXP may be a type node or an expression (whose type is tested). */ 2036 2037 int 2038 aggregate_value_p (const_tree exp, const_tree fntype) 2039 { 2040 const_tree type = (TYPE_P (exp)) ? exp : TREE_TYPE (exp); 2041 int i, regno, nregs; 2042 rtx reg; 2043 2044 if (fntype) 2045 switch (TREE_CODE (fntype)) 2046 { 2047 case CALL_EXPR: 2048 { 2049 tree fndecl = get_callee_fndecl (fntype); 2050 if (fndecl) 2051 fntype = TREE_TYPE (fndecl); 2052 else if (CALL_EXPR_FN (fntype)) 2053 fntype = TREE_TYPE (TREE_TYPE (CALL_EXPR_FN (fntype))); 2054 else 2055 /* For internal functions, assume nothing needs to be 2056 returned in memory. */ 2057 return 0; 2058 } 2059 break; 2060 case FUNCTION_DECL: 2061 fntype = TREE_TYPE (fntype); 2062 break; 2063 case FUNCTION_TYPE: 2064 case METHOD_TYPE: 2065 break; 2066 case IDENTIFIER_NODE: 2067 fntype = NULL_TREE; 2068 break; 2069 default: 2070 /* We don't expect other tree types here. */ 2071 gcc_unreachable (); 2072 } 2073 2074 if (VOID_TYPE_P (type)) 2075 return 0; 2076 2077 /* If a record should be passed the same as its first (and only) member 2078 don't pass it as an aggregate. */ 2079 if (TREE_CODE (type) == RECORD_TYPE && TYPE_TRANSPARENT_AGGR (type)) 2080 return aggregate_value_p (first_field (type), fntype); 2081 2082 /* If the front end has decided that this needs to be passed by 2083 reference, do so. */ 2084 if ((TREE_CODE (exp) == PARM_DECL || TREE_CODE (exp) == RESULT_DECL) 2085 && DECL_BY_REFERENCE (exp)) 2086 return 1; 2087 2088 /* Function types that are TREE_ADDRESSABLE force return in memory. */ 2089 if (fntype && TREE_ADDRESSABLE (fntype)) 2090 return 1; 2091 2092 /* Types that are TREE_ADDRESSABLE must be constructed in memory, 2093 and thus can't be returned in registers. */ 2094 if (TREE_ADDRESSABLE (type)) 2095 return 1; 2096 2097 if (TYPE_EMPTY_P (type)) 2098 return 0; 2099 2100 if (flag_pcc_struct_return && AGGREGATE_TYPE_P (type)) 2101 return 1; 2102 2103 if (targetm.calls.return_in_memory (type, fntype)) 2104 return 1; 2105 2106 /* Make sure we have suitable call-clobbered regs to return 2107 the value in; if not, we must return it in memory. */ 2108 reg = hard_function_value (type, 0, fntype, 0); 2109 2110 /* If we have something other than a REG (e.g. a PARALLEL), then assume 2111 it is OK. */ 2112 if (!REG_P (reg)) 2113 return 0; 2114 2115 regno = REGNO (reg); 2116 nregs = hard_regno_nregs (regno, TYPE_MODE (type)); 2117 for (i = 0; i < nregs; i++) 2118 if (! call_used_regs[regno + i]) 2119 return 1; 2120 2121 return 0; 2122 } 2123 2124 /* Return true if we should assign DECL a pseudo register; false if it 2125 should live on the local stack. */ 2126 2127 bool 2128 use_register_for_decl (const_tree decl) 2129 { 2130 if (TREE_CODE (decl) == SSA_NAME) 2131 { 2132 /* We often try to use the SSA_NAME, instead of its underlying 2133 decl, to get type information and guide decisions, to avoid 2134 differences of behavior between anonymous and named 2135 variables, but in this one case we have to go for the actual 2136 variable if there is one. The main reason is that, at least 2137 at -O0, we want to place user variables on the stack, but we 2138 don't mind using pseudos for anonymous or ignored temps. 2139 Should we take the SSA_NAME, we'd conclude all SSA_NAMEs 2140 should go in pseudos, whereas their corresponding variables 2141 might have to go on the stack. So, disregarding the decl 2142 here would negatively impact debug info at -O0, enable 2143 coalescing between SSA_NAMEs that ought to get different 2144 stack/pseudo assignments, and get the incoming argument 2145 processing thoroughly confused by PARM_DECLs expected to live 2146 in stack slots but assigned to pseudos. */ 2147 if (!SSA_NAME_VAR (decl)) 2148 return TYPE_MODE (TREE_TYPE (decl)) != BLKmode 2149 && !(flag_float_store && FLOAT_TYPE_P (TREE_TYPE (decl))); 2150 2151 decl = SSA_NAME_VAR (decl); 2152 } 2153 2154 /* Honor volatile. */ 2155 if (TREE_SIDE_EFFECTS (decl)) 2156 return false; 2157 2158 /* Honor addressability. */ 2159 if (TREE_ADDRESSABLE (decl)) 2160 return false; 2161 2162 /* RESULT_DECLs are a bit special in that they're assigned without 2163 regard to use_register_for_decl, but we generally only store in 2164 them. If we coalesce their SSA NAMEs, we'd better return a 2165 result that matches the assignment in expand_function_start. */ 2166 if (TREE_CODE (decl) == RESULT_DECL) 2167 { 2168 /* If it's not an aggregate, we're going to use a REG or a 2169 PARALLEL containing a REG. */ 2170 if (!aggregate_value_p (decl, current_function_decl)) 2171 return true; 2172 2173 /* If expand_function_start determines the return value, we'll 2174 use MEM if it's not by reference. */ 2175 if (cfun->returns_pcc_struct 2176 || (targetm.calls.struct_value_rtx 2177 (TREE_TYPE (current_function_decl), 1))) 2178 return DECL_BY_REFERENCE (decl); 2179 2180 /* Otherwise, we're taking an extra all.function_result_decl 2181 argument. It's set up in assign_parms_augmented_arg_list, 2182 under the (negated) conditions above, and then it's used to 2183 set up the RESULT_DECL rtl in assign_params, after looping 2184 over all parameters. Now, if the RESULT_DECL is not by 2185 reference, we'll use a MEM either way. */ 2186 if (!DECL_BY_REFERENCE (decl)) 2187 return false; 2188 2189 /* Otherwise, if RESULT_DECL is DECL_BY_REFERENCE, it will take 2190 the function_result_decl's assignment. Since it's a pointer, 2191 we can short-circuit a number of the tests below, and we must 2192 duplicat e them because we don't have the 2193 function_result_decl to test. */ 2194 if (!targetm.calls.allocate_stack_slots_for_args ()) 2195 return true; 2196 /* We don't set DECL_IGNORED_P for the function_result_decl. */ 2197 if (optimize) 2198 return true; 2199 /* We don't set DECL_REGISTER for the function_result_decl. */ 2200 return false; 2201 } 2202 2203 /* Decl is implicitly addressible by bound stores and loads 2204 if it is an aggregate holding bounds. */ 2205 if (chkp_function_instrumented_p (current_function_decl) 2206 && TREE_TYPE (decl) 2207 && !BOUNDED_P (decl) 2208 && chkp_type_has_pointer (TREE_TYPE (decl))) 2209 return false; 2210 2211 /* Only register-like things go in registers. */ 2212 if (DECL_MODE (decl) == BLKmode) 2213 return false; 2214 2215 /* If -ffloat-store specified, don't put explicit float variables 2216 into registers. */ 2217 /* ??? This should be checked after DECL_ARTIFICIAL, but tree-ssa 2218 propagates values across these stores, and it probably shouldn't. */ 2219 if (flag_float_store && FLOAT_TYPE_P (TREE_TYPE (decl))) 2220 return false; 2221 2222 if (!targetm.calls.allocate_stack_slots_for_args ()) 2223 return true; 2224 2225 /* If we're not interested in tracking debugging information for 2226 this decl, then we can certainly put it in a register. */ 2227 if (DECL_IGNORED_P (decl)) 2228 return true; 2229 2230 if (optimize) 2231 return true; 2232 2233 if (!DECL_REGISTER (decl)) 2234 return false; 2235 2236 /* When not optimizing, disregard register keyword for types that 2237 could have methods, otherwise the methods won't be callable from 2238 the debugger. */ 2239 if (RECORD_OR_UNION_TYPE_P (TREE_TYPE (decl))) 2240 return false; 2241 2242 return true; 2243 } 2244 2245 /* Structures to communicate between the subroutines of assign_parms. 2246 The first holds data persistent across all parameters, the second 2247 is cleared out for each parameter. */ 2248 2249 struct assign_parm_data_all 2250 { 2251 /* When INIT_CUMULATIVE_ARGS gets revamped, allocating CUMULATIVE_ARGS 2252 should become a job of the target or otherwise encapsulated. */ 2253 CUMULATIVE_ARGS args_so_far_v; 2254 cumulative_args_t args_so_far; 2255 struct args_size stack_args_size; 2256 tree function_result_decl; 2257 tree orig_fnargs; 2258 rtx_insn *first_conversion_insn; 2259 rtx_insn *last_conversion_insn; 2260 HOST_WIDE_INT pretend_args_size; 2261 HOST_WIDE_INT extra_pretend_bytes; 2262 int reg_parm_stack_space; 2263 }; 2264 2265 struct assign_parm_data_one 2266 { 2267 tree nominal_type; 2268 tree passed_type; 2269 rtx entry_parm; 2270 rtx stack_parm; 2271 machine_mode nominal_mode; 2272 machine_mode passed_mode; 2273 machine_mode promoted_mode; 2274 struct locate_and_pad_arg_data locate; 2275 int partial; 2276 BOOL_BITFIELD named_arg : 1; 2277 BOOL_BITFIELD passed_pointer : 1; 2278 BOOL_BITFIELD on_stack : 1; 2279 BOOL_BITFIELD loaded_in_reg : 1; 2280 }; 2281 2282 struct bounds_parm_data 2283 { 2284 assign_parm_data_one parm_data; 2285 tree bounds_parm; 2286 tree ptr_parm; 2287 rtx ptr_entry; 2288 int bound_no; 2289 }; 2290 2291 /* A subroutine of assign_parms. Initialize ALL. */ 2292 2293 static void 2294 assign_parms_initialize_all (struct assign_parm_data_all *all) 2295 { 2296 tree fntype ATTRIBUTE_UNUSED; 2297 2298 memset (all, 0, sizeof (*all)); 2299 2300 fntype = TREE_TYPE (current_function_decl); 2301 2302 #ifdef INIT_CUMULATIVE_INCOMING_ARGS 2303 INIT_CUMULATIVE_INCOMING_ARGS (all->args_so_far_v, fntype, NULL_RTX); 2304 #else 2305 INIT_CUMULATIVE_ARGS (all->args_so_far_v, fntype, NULL_RTX, 2306 current_function_decl, -1); 2307 #endif 2308 all->args_so_far = pack_cumulative_args (&all->args_so_far_v); 2309 2310 #ifdef INCOMING_REG_PARM_STACK_SPACE 2311 all->reg_parm_stack_space 2312 = INCOMING_REG_PARM_STACK_SPACE (current_function_decl); 2313 #endif 2314 } 2315 2316 /* If ARGS contains entries with complex types, split the entry into two 2317 entries of the component type. Return a new list of substitutions are 2318 needed, else the old list. */ 2319 2320 static void 2321 split_complex_args (vec<tree> *args) 2322 { 2323 unsigned i; 2324 tree p; 2325 2326 FOR_EACH_VEC_ELT (*args, i, p) 2327 { 2328 tree type = TREE_TYPE (p); 2329 if (TREE_CODE (type) == COMPLEX_TYPE 2330 && targetm.calls.split_complex_arg (type)) 2331 { 2332 tree decl; 2333 tree subtype = TREE_TYPE (type); 2334 bool addressable = TREE_ADDRESSABLE (p); 2335 2336 /* Rewrite the PARM_DECL's type with its component. */ 2337 p = copy_node (p); 2338 TREE_TYPE (p) = subtype; 2339 DECL_ARG_TYPE (p) = TREE_TYPE (DECL_ARG_TYPE (p)); 2340 SET_DECL_MODE (p, VOIDmode); 2341 DECL_SIZE (p) = NULL; 2342 DECL_SIZE_UNIT (p) = NULL; 2343 /* If this arg must go in memory, put it in a pseudo here. 2344 We can't allow it to go in memory as per normal parms, 2345 because the usual place might not have the imag part 2346 adjacent to the real part. */ 2347 DECL_ARTIFICIAL (p) = addressable; 2348 DECL_IGNORED_P (p) = addressable; 2349 TREE_ADDRESSABLE (p) = 0; 2350 layout_decl (p, 0); 2351 (*args)[i] = p; 2352 2353 /* Build a second synthetic decl. */ 2354 decl = build_decl (EXPR_LOCATION (p), 2355 PARM_DECL, NULL_TREE, subtype); 2356 DECL_ARG_TYPE (decl) = DECL_ARG_TYPE (p); 2357 DECL_ARTIFICIAL (decl) = addressable; 2358 DECL_IGNORED_P (decl) = addressable; 2359 layout_decl (decl, 0); 2360 args->safe_insert (++i, decl); 2361 } 2362 } 2363 } 2364 2365 /* A subroutine of assign_parms. Adjust the parameter list to incorporate 2366 the hidden struct return argument, and (abi willing) complex args. 2367 Return the new parameter list. */ 2368 2369 static vec<tree> 2370 assign_parms_augmented_arg_list (struct assign_parm_data_all *all) 2371 { 2372 tree fndecl = current_function_decl; 2373 tree fntype = TREE_TYPE (fndecl); 2374 vec<tree> fnargs = vNULL; 2375 tree arg; 2376 2377 for (arg = DECL_ARGUMENTS (fndecl); arg; arg = DECL_CHAIN (arg)) 2378 fnargs.safe_push (arg); 2379 2380 all->orig_fnargs = DECL_ARGUMENTS (fndecl); 2381 2382 /* If struct value address is treated as the first argument, make it so. */ 2383 if (aggregate_value_p (DECL_RESULT (fndecl), fndecl) 2384 && ! cfun->returns_pcc_struct 2385 && targetm.calls.struct_value_rtx (TREE_TYPE (fndecl), 1) == 0) 2386 { 2387 tree type = build_pointer_type (TREE_TYPE (fntype)); 2388 tree decl; 2389 2390 decl = build_decl (DECL_SOURCE_LOCATION (fndecl), 2391 PARM_DECL, get_identifier (".result_ptr"), type); 2392 DECL_ARG_TYPE (decl) = type; 2393 DECL_ARTIFICIAL (decl) = 1; 2394 DECL_NAMELESS (decl) = 1; 2395 TREE_CONSTANT (decl) = 1; 2396 /* We don't set DECL_IGNORED_P or DECL_REGISTER here. If this 2397 changes, the end of the RESULT_DECL handling block in 2398 use_register_for_decl must be adjusted to match. */ 2399 2400 DECL_CHAIN (decl) = all->orig_fnargs; 2401 all->orig_fnargs = decl; 2402 fnargs.safe_insert (0, decl); 2403 2404 all->function_result_decl = decl; 2405 2406 /* If function is instrumented then bounds of the 2407 passed structure address is the second argument. */ 2408 if (chkp_function_instrumented_p (fndecl)) 2409 { 2410 decl = build_decl (DECL_SOURCE_LOCATION (fndecl), 2411 PARM_DECL, get_identifier (".result_bnd"), 2412 pointer_bounds_type_node); 2413 DECL_ARG_TYPE (decl) = pointer_bounds_type_node; 2414 DECL_ARTIFICIAL (decl) = 1; 2415 DECL_NAMELESS (decl) = 1; 2416 TREE_CONSTANT (decl) = 1; 2417 2418 DECL_CHAIN (decl) = DECL_CHAIN (all->orig_fnargs); 2419 DECL_CHAIN (all->orig_fnargs) = decl; 2420 fnargs.safe_insert (1, decl); 2421 } 2422 } 2423 2424 /* If the target wants to split complex arguments into scalars, do so. */ 2425 if (targetm.calls.split_complex_arg) 2426 split_complex_args (&fnargs); 2427 2428 return fnargs; 2429 } 2430 2431 /* A subroutine of assign_parms. Examine PARM and pull out type and mode 2432 data for the parameter. Incorporate ABI specifics such as pass-by- 2433 reference and type promotion. */ 2434 2435 static void 2436 assign_parm_find_data_types (struct assign_parm_data_all *all, tree parm, 2437 struct assign_parm_data_one *data) 2438 { 2439 tree nominal_type, passed_type; 2440 machine_mode nominal_mode, passed_mode, promoted_mode; 2441 int unsignedp; 2442 2443 memset (data, 0, sizeof (*data)); 2444 2445 /* NAMED_ARG is a misnomer. We really mean 'non-variadic'. */ 2446 if (!cfun->stdarg) 2447 data->named_arg = 1; /* No variadic parms. */ 2448 else if (DECL_CHAIN (parm)) 2449 data->named_arg = 1; /* Not the last non-variadic parm. */ 2450 else if (targetm.calls.strict_argument_naming (all->args_so_far)) 2451 data->named_arg = 1; /* Only variadic ones are unnamed. */ 2452 else 2453 data->named_arg = 0; /* Treat as variadic. */ 2454 2455 nominal_type = TREE_TYPE (parm); 2456 passed_type = DECL_ARG_TYPE (parm); 2457 2458 /* Look out for errors propagating this far. Also, if the parameter's 2459 type is void then its value doesn't matter. */ 2460 if (TREE_TYPE (parm) == error_mark_node 2461 /* This can happen after weird syntax errors 2462 or if an enum type is defined among the parms. */ 2463 || TREE_CODE (parm) != PARM_DECL 2464 || passed_type == NULL 2465 || VOID_TYPE_P (nominal_type)) 2466 { 2467 nominal_type = passed_type = void_type_node; 2468 nominal_mode = passed_mode = promoted_mode = VOIDmode; 2469 goto egress; 2470 } 2471 2472 /* Find mode of arg as it is passed, and mode of arg as it should be 2473 during execution of this function. */ 2474 passed_mode = TYPE_MODE (passed_type); 2475 nominal_mode = TYPE_MODE (nominal_type); 2476 2477 /* If the parm is to be passed as a transparent union or record, use the 2478 type of the first field for the tests below. We have already verified 2479 that the modes are the same. */ 2480 if (RECORD_OR_UNION_TYPE_P (passed_type) 2481 && TYPE_TRANSPARENT_AGGR (passed_type)) 2482 passed_type = TREE_TYPE (first_field (passed_type)); 2483 2484 /* See if this arg was passed by invisible reference. */ 2485 if (pass_by_reference (&all->args_so_far_v, passed_mode, 2486 passed_type, data->named_arg)) 2487 { 2488 passed_type = nominal_type = build_pointer_type (passed_type); 2489 data->passed_pointer = true; 2490 passed_mode = nominal_mode = TYPE_MODE (nominal_type); 2491 } 2492 2493 /* Find mode as it is passed by the ABI. */ 2494 unsignedp = TYPE_UNSIGNED (passed_type); 2495 promoted_mode = promote_function_mode (passed_type, passed_mode, &unsignedp, 2496 TREE_TYPE (current_function_decl), 0); 2497 2498 egress: 2499 data->nominal_type = nominal_type; 2500 data->passed_type = passed_type; 2501 data->nominal_mode = nominal_mode; 2502 data->passed_mode = passed_mode; 2503 data->promoted_mode = promoted_mode; 2504 } 2505 2506 /* A subroutine of assign_parms. Invoke setup_incoming_varargs. */ 2507 2508 static void 2509 assign_parms_setup_varargs (struct assign_parm_data_all *all, 2510 struct assign_parm_data_one *data, bool no_rtl) 2511 { 2512 int varargs_pretend_bytes = 0; 2513 2514 targetm.calls.setup_incoming_varargs (all->args_so_far, 2515 data->promoted_mode, 2516 data->passed_type, 2517 &varargs_pretend_bytes, no_rtl); 2518 2519 /* If the back-end has requested extra stack space, record how much is 2520 needed. Do not change pretend_args_size otherwise since it may be 2521 nonzero from an earlier partial argument. */ 2522 if (varargs_pretend_bytes > 0) 2523 all->pretend_args_size = varargs_pretend_bytes; 2524 } 2525 2526 /* A subroutine of assign_parms. Set DATA->ENTRY_PARM corresponding to 2527 the incoming location of the current parameter. */ 2528 2529 static void 2530 assign_parm_find_entry_rtl (struct assign_parm_data_all *all, 2531 struct assign_parm_data_one *data) 2532 { 2533 HOST_WIDE_INT pretend_bytes = 0; 2534 rtx entry_parm; 2535 bool in_regs; 2536 2537 if (data->promoted_mode == VOIDmode) 2538 { 2539 data->entry_parm = data->stack_parm = const0_rtx; 2540 return; 2541 } 2542 2543 targetm.calls.warn_parameter_passing_abi (all->args_so_far, 2544 data->passed_type); 2545 2546 entry_parm = targetm.calls.function_incoming_arg (all->args_so_far, 2547 data->promoted_mode, 2548 data->passed_type, 2549 data->named_arg); 2550 2551 if (entry_parm == 0) 2552 data->promoted_mode = data->passed_mode; 2553 2554 /* Determine parm's home in the stack, in case it arrives in the stack 2555 or we should pretend it did. Compute the stack position and rtx where 2556 the argument arrives and its size. 2557 2558 There is one complexity here: If this was a parameter that would 2559 have been passed in registers, but wasn't only because it is 2560 __builtin_va_alist, we want locate_and_pad_parm to treat it as if 2561 it came in a register so that REG_PARM_STACK_SPACE isn't skipped. 2562 In this case, we call FUNCTION_ARG with NAMED set to 1 instead of 0 2563 as it was the previous time. */ 2564 in_regs = (entry_parm != 0) || POINTER_BOUNDS_TYPE_P (data->passed_type); 2565 #ifdef STACK_PARMS_IN_REG_PARM_AREA 2566 in_regs = true; 2567 #endif 2568 if (!in_regs && !data->named_arg) 2569 { 2570 if (targetm.calls.pretend_outgoing_varargs_named (all->args_so_far)) 2571 { 2572 rtx tem; 2573 tem = targetm.calls.function_incoming_arg (all->args_so_far, 2574 data->promoted_mode, 2575 data->passed_type, true); 2576 in_regs = tem != NULL; 2577 } 2578 } 2579 2580 /* If this parameter was passed both in registers and in the stack, use 2581 the copy on the stack. */ 2582 if (targetm.calls.must_pass_in_stack (data->promoted_mode, 2583 data->passed_type)) 2584 entry_parm = 0; 2585 2586 if (entry_parm) 2587 { 2588 int partial; 2589 2590 partial = targetm.calls.arg_partial_bytes (all->args_so_far, 2591 data->promoted_mode, 2592 data->passed_type, 2593 data->named_arg); 2594 data->partial = partial; 2595 2596 /* The caller might already have allocated stack space for the 2597 register parameters. */ 2598 if (partial != 0 && all->reg_parm_stack_space == 0) 2599 { 2600 /* Part of this argument is passed in registers and part 2601 is passed on the stack. Ask the prologue code to extend 2602 the stack part so that we can recreate the full value. 2603 2604 PRETEND_BYTES is the size of the registers we need to store. 2605 CURRENT_FUNCTION_PRETEND_ARGS_SIZE is the amount of extra 2606 stack space that the prologue should allocate. 2607 2608 Internally, gcc assumes that the argument pointer is aligned 2609 to STACK_BOUNDARY bits. This is used both for alignment 2610 optimizations (see init_emit) and to locate arguments that are 2611 aligned to more than PARM_BOUNDARY bits. We must preserve this 2612 invariant by rounding CURRENT_FUNCTION_PRETEND_ARGS_SIZE up to 2613 a stack boundary. */ 2614 2615 /* We assume at most one partial arg, and it must be the first 2616 argument on the stack. */ 2617 gcc_assert (!all->extra_pretend_bytes && !all->pretend_args_size); 2618 2619 pretend_bytes = partial; 2620 all->pretend_args_size = CEIL_ROUND (pretend_bytes, STACK_BYTES); 2621 2622 /* We want to align relative to the actual stack pointer, so 2623 don't include this in the stack size until later. */ 2624 all->extra_pretend_bytes = all->pretend_args_size; 2625 } 2626 } 2627 2628 locate_and_pad_parm (data->promoted_mode, data->passed_type, in_regs, 2629 all->reg_parm_stack_space, 2630 entry_parm ? data->partial : 0, current_function_decl, 2631 &all->stack_args_size, &data->locate); 2632 2633 /* Update parm_stack_boundary if this parameter is passed in the 2634 stack. */ 2635 if (!in_regs && crtl->parm_stack_boundary < data->locate.boundary) 2636 crtl->parm_stack_boundary = data->locate.boundary; 2637 2638 /* Adjust offsets to include the pretend args. */ 2639 pretend_bytes = all->extra_pretend_bytes - pretend_bytes; 2640 data->locate.slot_offset.constant += pretend_bytes; 2641 data->locate.offset.constant += pretend_bytes; 2642 2643 data->entry_parm = entry_parm; 2644 } 2645 2646 /* A subroutine of assign_parms. If there is actually space on the stack 2647 for this parm, count it in stack_args_size and return true. */ 2648 2649 static bool 2650 assign_parm_is_stack_parm (struct assign_parm_data_all *all, 2651 struct assign_parm_data_one *data) 2652 { 2653 /* Bounds are never passed on the stack to keep compatibility 2654 with not instrumented code. */ 2655 if (POINTER_BOUNDS_TYPE_P (data->passed_type)) 2656 return false; 2657 /* Trivially true if we've no incoming register. */ 2658 else if (data->entry_parm == NULL) 2659 ; 2660 /* Also true if we're partially in registers and partially not, 2661 since we've arranged to drop the entire argument on the stack. */ 2662 else if (data->partial != 0) 2663 ; 2664 /* Also true if the target says that it's passed in both registers 2665 and on the stack. */ 2666 else if (GET_CODE (data->entry_parm) == PARALLEL 2667 && XEXP (XVECEXP (data->entry_parm, 0, 0), 0) == NULL_RTX) 2668 ; 2669 /* Also true if the target says that there's stack allocated for 2670 all register parameters. */ 2671 else if (all->reg_parm_stack_space > 0) 2672 ; 2673 /* Otherwise, no, this parameter has no ABI defined stack slot. */ 2674 else 2675 return false; 2676 2677 all->stack_args_size.constant += data->locate.size.constant; 2678 if (data->locate.size.var) 2679 ADD_PARM_SIZE (all->stack_args_size, data->locate.size.var); 2680 2681 return true; 2682 } 2683 2684 /* A subroutine of assign_parms. Given that this parameter is allocated 2685 stack space by the ABI, find it. */ 2686 2687 static void 2688 assign_parm_find_stack_rtl (tree parm, struct assign_parm_data_one *data) 2689 { 2690 rtx offset_rtx, stack_parm; 2691 unsigned int align, boundary; 2692 2693 /* If we're passing this arg using a reg, make its stack home the 2694 aligned stack slot. */ 2695 if (data->entry_parm) 2696 offset_rtx = ARGS_SIZE_RTX (data->locate.slot_offset); 2697 else 2698 offset_rtx = ARGS_SIZE_RTX (data->locate.offset); 2699 2700 stack_parm = crtl->args.internal_arg_pointer; 2701 if (offset_rtx != const0_rtx) 2702 stack_parm = gen_rtx_PLUS (Pmode, stack_parm, offset_rtx); 2703 stack_parm = gen_rtx_MEM (data->promoted_mode, stack_parm); 2704 2705 if (!data->passed_pointer) 2706 { 2707 set_mem_attributes (stack_parm, parm, 1); 2708 /* set_mem_attributes could set MEM_SIZE to the passed mode's size, 2709 while promoted mode's size is needed. */ 2710 if (data->promoted_mode != BLKmode 2711 && data->promoted_mode != DECL_MODE (parm)) 2712 { 2713 set_mem_size (stack_parm, GET_MODE_SIZE (data->promoted_mode)); 2714 if (MEM_EXPR (stack_parm) && MEM_OFFSET_KNOWN_P (stack_parm)) 2715 { 2716 poly_int64 offset = subreg_lowpart_offset (DECL_MODE (parm), 2717 data->promoted_mode); 2718 if (maybe_ne (offset, 0)) 2719 set_mem_offset (stack_parm, MEM_OFFSET (stack_parm) - offset); 2720 } 2721 } 2722 } 2723 2724 boundary = data->locate.boundary; 2725 align = BITS_PER_UNIT; 2726 2727 /* If we're padding upward, we know that the alignment of the slot 2728 is TARGET_FUNCTION_ARG_BOUNDARY. If we're using slot_offset, we're 2729 intentionally forcing upward padding. Otherwise we have to come 2730 up with a guess at the alignment based on OFFSET_RTX. */ 2731 poly_int64 offset; 2732 if (data->locate.where_pad != PAD_DOWNWARD || data->entry_parm) 2733 align = boundary; 2734 else if (poly_int_rtx_p (offset_rtx, &offset)) 2735 { 2736 align = least_bit_hwi (boundary); 2737 unsigned int offset_align = known_alignment (offset) * BITS_PER_UNIT; 2738 if (offset_align != 0) 2739 align = MIN (align, offset_align); 2740 } 2741 set_mem_align (stack_parm, align); 2742 2743 if (data->entry_parm) 2744 set_reg_attrs_for_parm (data->entry_parm, stack_parm); 2745 2746 data->stack_parm = stack_parm; 2747 } 2748 2749 /* A subroutine of assign_parms. Adjust DATA->ENTRY_RTL such that it's 2750 always valid and contiguous. */ 2751 2752 static void 2753 assign_parm_adjust_entry_rtl (struct assign_parm_data_one *data) 2754 { 2755 rtx entry_parm = data->entry_parm; 2756 rtx stack_parm = data->stack_parm; 2757 2758 /* If this parm was passed part in regs and part in memory, pretend it 2759 arrived entirely in memory by pushing the register-part onto the stack. 2760 In the special case of a DImode or DFmode that is split, we could put 2761 it together in a pseudoreg directly, but for now that's not worth 2762 bothering with. */ 2763 if (data->partial != 0) 2764 { 2765 /* Handle calls that pass values in multiple non-contiguous 2766 locations. The Irix 6 ABI has examples of this. */ 2767 if (GET_CODE (entry_parm) == PARALLEL) 2768 emit_group_store (validize_mem (copy_rtx (stack_parm)), entry_parm, 2769 data->passed_type, 2770 int_size_in_bytes (data->passed_type)); 2771 else 2772 { 2773 gcc_assert (data->partial % UNITS_PER_WORD == 0); 2774 move_block_from_reg (REGNO (entry_parm), 2775 validize_mem (copy_rtx (stack_parm)), 2776 data->partial / UNITS_PER_WORD); 2777 } 2778 2779 entry_parm = stack_parm; 2780 } 2781 2782 /* If we didn't decide this parm came in a register, by default it came 2783 on the stack. */ 2784 else if (entry_parm == NULL) 2785 entry_parm = stack_parm; 2786 2787 /* When an argument is passed in multiple locations, we can't make use 2788 of this information, but we can save some copying if the whole argument 2789 is passed in a single register. */ 2790 else if (GET_CODE (entry_parm) == PARALLEL 2791 && data->nominal_mode != BLKmode 2792 && data->passed_mode != BLKmode) 2793 { 2794 size_t i, len = XVECLEN (entry_parm, 0); 2795 2796 for (i = 0; i < len; i++) 2797 if (XEXP (XVECEXP (entry_parm, 0, i), 0) != NULL_RTX 2798 && REG_P (XEXP (XVECEXP (entry_parm, 0, i), 0)) 2799 && (GET_MODE (XEXP (XVECEXP (entry_parm, 0, i), 0)) 2800 == data->passed_mode) 2801 && INTVAL (XEXP (XVECEXP (entry_parm, 0, i), 1)) == 0) 2802 { 2803 entry_parm = XEXP (XVECEXP (entry_parm, 0, i), 0); 2804 break; 2805 } 2806 } 2807 2808 data->entry_parm = entry_parm; 2809 } 2810 2811 /* A subroutine of assign_parms. Reconstitute any values which were 2812 passed in multiple registers and would fit in a single register. */ 2813 2814 static void 2815 assign_parm_remove_parallels (struct assign_parm_data_one *data) 2816 { 2817 rtx entry_parm = data->entry_parm; 2818 2819 /* Convert the PARALLEL to a REG of the same mode as the parallel. 2820 This can be done with register operations rather than on the 2821 stack, even if we will store the reconstituted parameter on the 2822 stack later. */ 2823 if (GET_CODE (entry_parm) == PARALLEL && GET_MODE (entry_parm) != BLKmode) 2824 { 2825 rtx parmreg = gen_reg_rtx (GET_MODE (entry_parm)); 2826 emit_group_store (parmreg, entry_parm, data->passed_type, 2827 GET_MODE_SIZE (GET_MODE (entry_parm))); 2828 entry_parm = parmreg; 2829 } 2830 2831 data->entry_parm = entry_parm; 2832 } 2833 2834 /* A subroutine of assign_parms. Adjust DATA->STACK_RTL such that it's 2835 always valid and properly aligned. */ 2836 2837 static void 2838 assign_parm_adjust_stack_rtl (struct assign_parm_data_one *data) 2839 { 2840 rtx stack_parm = data->stack_parm; 2841 2842 /* If we can't trust the parm stack slot to be aligned enough for its 2843 ultimate type, don't use that slot after entry. We'll make another 2844 stack slot, if we need one. */ 2845 if (stack_parm 2846 && ((STRICT_ALIGNMENT 2847 && GET_MODE_ALIGNMENT (data->nominal_mode) > MEM_ALIGN (stack_parm)) 2848 || (data->nominal_type 2849 && TYPE_ALIGN (data->nominal_type) > MEM_ALIGN (stack_parm) 2850 && MEM_ALIGN (stack_parm) < PREFERRED_STACK_BOUNDARY))) 2851 stack_parm = NULL; 2852 2853 /* If parm was passed in memory, and we need to convert it on entry, 2854 don't store it back in that same slot. */ 2855 else if (data->entry_parm == stack_parm 2856 && data->nominal_mode != BLKmode 2857 && data->nominal_mode != data->passed_mode) 2858 stack_parm = NULL; 2859 2860 /* If stack protection is in effect for this function, don't leave any 2861 pointers in their passed stack slots. */ 2862 else if (crtl->stack_protect_guard 2863 && (flag_stack_protect == 2 2864 || data->passed_pointer 2865 || POINTER_TYPE_P (data->nominal_type))) 2866 stack_parm = NULL; 2867 2868 data->stack_parm = stack_parm; 2869 } 2870 2871 /* A subroutine of assign_parms. Return true if the current parameter 2872 should be stored as a BLKmode in the current frame. */ 2873 2874 static bool 2875 assign_parm_setup_block_p (struct assign_parm_data_one *data) 2876 { 2877 if (data->nominal_mode == BLKmode) 2878 return true; 2879 if (GET_MODE (data->entry_parm) == BLKmode) 2880 return true; 2881 2882 #ifdef BLOCK_REG_PADDING 2883 /* Only assign_parm_setup_block knows how to deal with register arguments 2884 that are padded at the least significant end. */ 2885 if (REG_P (data->entry_parm) 2886 && known_lt (GET_MODE_SIZE (data->promoted_mode), UNITS_PER_WORD) 2887 && (BLOCK_REG_PADDING (data->passed_mode, data->passed_type, 1) 2888 == (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD))) 2889 return true; 2890 #endif 2891 2892 return false; 2893 } 2894 2895 /* A subroutine of assign_parms. Arrange for the parameter to be 2896 present and valid in DATA->STACK_RTL. */ 2897 2898 static void 2899 assign_parm_setup_block (struct assign_parm_data_all *all, 2900 tree parm, struct assign_parm_data_one *data) 2901 { 2902 rtx entry_parm = data->entry_parm; 2903 rtx stack_parm = data->stack_parm; 2904 rtx target_reg = NULL_RTX; 2905 bool in_conversion_seq = false; 2906 HOST_WIDE_INT size; 2907 HOST_WIDE_INT size_stored; 2908 2909 if (GET_CODE (entry_parm) == PARALLEL) 2910 entry_parm = emit_group_move_into_temps (entry_parm); 2911 2912 /* If we want the parameter in a pseudo, don't use a stack slot. */ 2913 if (is_gimple_reg (parm) && use_register_for_decl (parm)) 2914 { 2915 tree def = ssa_default_def (cfun, parm); 2916 gcc_assert (def); 2917 machine_mode mode = promote_ssa_mode (def, NULL); 2918 rtx reg = gen_reg_rtx (mode); 2919 if (GET_CODE (reg) != CONCAT) 2920 stack_parm = reg; 2921 else 2922 { 2923 target_reg = reg; 2924 /* Avoid allocating a stack slot, if there isn't one 2925 preallocated by the ABI. It might seem like we should 2926 always prefer a pseudo, but converting between 2927 floating-point and integer modes goes through the stack 2928 on various machines, so it's better to use the reserved 2929 stack slot than to risk wasting it and allocating more 2930 for the conversion. */ 2931 if (stack_parm == NULL_RTX) 2932 { 2933 int save = generating_concat_p; 2934 generating_concat_p = 0; 2935 stack_parm = gen_reg_rtx (mode); 2936 generating_concat_p = save; 2937 } 2938 } 2939 data->stack_parm = NULL; 2940 } 2941 2942 size = int_size_in_bytes (data->passed_type); 2943 size_stored = CEIL_ROUND (size, UNITS_PER_WORD); 2944 if (stack_parm == 0) 2945 { 2946 SET_DECL_ALIGN (parm, MAX (DECL_ALIGN (parm), BITS_PER_WORD)); 2947 stack_parm = assign_stack_local (BLKmode, size_stored, 2948 DECL_ALIGN (parm)); 2949 if (known_eq (GET_MODE_SIZE (GET_MODE (entry_parm)), size)) 2950 PUT_MODE (stack_parm, GET_MODE (entry_parm)); 2951 set_mem_attributes (stack_parm, parm, 1); 2952 } 2953 2954 /* If a BLKmode arrives in registers, copy it to a stack slot. Handle 2955 calls that pass values in multiple non-contiguous locations. */ 2956 if (REG_P (entry_parm) || GET_CODE (entry_parm) == PARALLEL) 2957 { 2958 rtx mem; 2959 2960 /* Note that we will be storing an integral number of words. 2961 So we have to be careful to ensure that we allocate an 2962 integral number of words. We do this above when we call 2963 assign_stack_local if space was not allocated in the argument 2964 list. If it was, this will not work if PARM_BOUNDARY is not 2965 a multiple of BITS_PER_WORD. It isn't clear how to fix this 2966 if it becomes a problem. Exception is when BLKmode arrives 2967 with arguments not conforming to word_mode. */ 2968 2969 if (data->stack_parm == 0) 2970 ; 2971 else if (GET_CODE (entry_parm) == PARALLEL) 2972 ; 2973 else 2974 gcc_assert (!size || !(PARM_BOUNDARY % BITS_PER_WORD)); 2975 2976 mem = validize_mem (copy_rtx (stack_parm)); 2977 2978 /* Handle values in multiple non-contiguous locations. */ 2979 if (GET_CODE (entry_parm) == PARALLEL && !MEM_P (mem)) 2980 emit_group_store (mem, entry_parm, data->passed_type, size); 2981 else if (GET_CODE (entry_parm) == PARALLEL) 2982 { 2983 push_to_sequence2 (all->first_conversion_insn, 2984 all->last_conversion_insn); 2985 emit_group_store (mem, entry_parm, data->passed_type, size); 2986 all->first_conversion_insn = get_insns (); 2987 all->last_conversion_insn = get_last_insn (); 2988 end_sequence (); 2989 in_conversion_seq = true; 2990 } 2991 2992 else if (size == 0) 2993 ; 2994 2995 /* If SIZE is that of a mode no bigger than a word, just use 2996 that mode's store operation. */ 2997 else if (size <= UNITS_PER_WORD) 2998 { 2999 unsigned int bits = size * BITS_PER_UNIT; 3000 machine_mode mode = int_mode_for_size (bits, 0).else_blk (); 3001 3002 if (mode != BLKmode 3003 #ifdef BLOCK_REG_PADDING 3004 && (size == UNITS_PER_WORD 3005 || (BLOCK_REG_PADDING (mode, data->passed_type, 1) 3006 != (BYTES_BIG_ENDIAN ? PAD_UPWARD : PAD_DOWNWARD))) 3007 #endif 3008 ) 3009 { 3010 rtx reg; 3011 3012 /* We are really truncating a word_mode value containing 3013 SIZE bytes into a value of mode MODE. If such an 3014 operation requires no actual instructions, we can refer 3015 to the value directly in mode MODE, otherwise we must 3016 start with the register in word_mode and explicitly 3017 convert it. */ 3018 if (targetm.truly_noop_truncation (size * BITS_PER_UNIT, 3019 BITS_PER_WORD)) 3020 reg = gen_rtx_REG (mode, REGNO (entry_parm)); 3021 else 3022 { 3023 reg = gen_rtx_REG (word_mode, REGNO (entry_parm)); 3024 reg = convert_to_mode (mode, copy_to_reg (reg), 1); 3025 } 3026 emit_move_insn (change_address (mem, mode, 0), reg); 3027 } 3028 3029 #ifdef BLOCK_REG_PADDING 3030 /* Storing the register in memory as a full word, as 3031 move_block_from_reg below would do, and then using the 3032 MEM in a smaller mode, has the effect of shifting right 3033 if BYTES_BIG_ENDIAN. If we're bypassing memory, the 3034 shifting must be explicit. */ 3035 else if (!MEM_P (mem)) 3036 { 3037 rtx x; 3038 3039 /* If the assert below fails, we should have taken the 3040 mode != BLKmode path above, unless we have downward 3041 padding of smaller-than-word arguments on a machine 3042 with little-endian bytes, which would likely require 3043 additional changes to work correctly. */ 3044 gcc_checking_assert (BYTES_BIG_ENDIAN 3045 && (BLOCK_REG_PADDING (mode, 3046 data->passed_type, 1) 3047 == PAD_UPWARD)); 3048 3049 int by = (UNITS_PER_WORD - size) * BITS_PER_UNIT; 3050 3051 x = gen_rtx_REG (word_mode, REGNO (entry_parm)); 3052 x = expand_shift (RSHIFT_EXPR, word_mode, x, by, 3053 NULL_RTX, 1); 3054 x = force_reg (word_mode, x); 3055 x = gen_lowpart_SUBREG (GET_MODE (mem), x); 3056 3057 emit_move_insn (mem, x); 3058 } 3059 #endif 3060 3061 /* Blocks smaller than a word on a BYTES_BIG_ENDIAN 3062 machine must be aligned to the left before storing 3063 to memory. Note that the previous test doesn't 3064 handle all cases (e.g. SIZE == 3). */ 3065 else if (size != UNITS_PER_WORD 3066 #ifdef BLOCK_REG_PADDING 3067 && (BLOCK_REG_PADDING (mode, data->passed_type, 1) 3068 == PAD_DOWNWARD) 3069 #else 3070 && BYTES_BIG_ENDIAN 3071 #endif 3072 ) 3073 { 3074 rtx tem, x; 3075 int by = (UNITS_PER_WORD - size) * BITS_PER_UNIT; 3076 rtx reg = gen_rtx_REG (word_mode, REGNO (entry_parm)); 3077 3078 x = expand_shift (LSHIFT_EXPR, word_mode, reg, by, NULL_RTX, 1); 3079 tem = change_address (mem, word_mode, 0); 3080 emit_move_insn (tem, x); 3081 } 3082 else 3083 move_block_from_reg (REGNO (entry_parm), mem, 3084 size_stored / UNITS_PER_WORD); 3085 } 3086 else if (!MEM_P (mem)) 3087 { 3088 gcc_checking_assert (size > UNITS_PER_WORD); 3089 #ifdef BLOCK_REG_PADDING 3090 gcc_checking_assert (BLOCK_REG_PADDING (GET_MODE (mem), 3091 data->passed_type, 0) 3092 == PAD_UPWARD); 3093 #endif 3094 emit_move_insn (mem, entry_parm); 3095 } 3096 else 3097 move_block_from_reg (REGNO (entry_parm), mem, 3098 size_stored / UNITS_PER_WORD); 3099 } 3100 else if (data->stack_parm == 0 && !TYPE_EMPTY_P (data->passed_type)) 3101 { 3102 push_to_sequence2 (all->first_conversion_insn, all->last_conversion_insn); 3103 emit_block_move (stack_parm, data->entry_parm, GEN_INT (size), 3104 BLOCK_OP_NORMAL); 3105 all->first_conversion_insn = get_insns (); 3106 all->last_conversion_insn = get_last_insn (); 3107 end_sequence (); 3108 in_conversion_seq = true; 3109 } 3110 3111 if (target_reg) 3112 { 3113 if (!in_conversion_seq) 3114 emit_move_insn (target_reg, stack_parm); 3115 else 3116 { 3117 push_to_sequence2 (all->first_conversion_insn, 3118 all->last_conversion_insn); 3119 emit_move_insn (target_reg, stack_parm); 3120 all->first_conversion_insn = get_insns (); 3121 all->last_conversion_insn = get_last_insn (); 3122 end_sequence (); 3123 } 3124 stack_parm = target_reg; 3125 } 3126 3127 data->stack_parm = stack_parm; 3128 set_parm_rtl (parm, stack_parm); 3129 } 3130 3131 /* A subroutine of assign_parms. Allocate a pseudo to hold the current 3132 parameter. Get it there. Perform all ABI specified conversions. */ 3133 3134 static void 3135 assign_parm_setup_reg (struct assign_parm_data_all *all, tree parm, 3136 struct assign_parm_data_one *data) 3137 { 3138 rtx parmreg, validated_mem; 3139 rtx equiv_stack_parm; 3140 machine_mode promoted_nominal_mode; 3141 int unsignedp = TYPE_UNSIGNED (TREE_TYPE (parm)); 3142 bool did_conversion = false; 3143 bool need_conversion, moved; 3144 rtx rtl; 3145 3146 /* Store the parm in a pseudoregister during the function, but we may 3147 need to do it in a wider mode. Using 2 here makes the result 3148 consistent with promote_decl_mode and thus expand_expr_real_1. */ 3149 promoted_nominal_mode 3150 = promote_function_mode (data->nominal_type, data->nominal_mode, &unsignedp, 3151 TREE_TYPE (current_function_decl), 2); 3152 3153 parmreg = gen_reg_rtx (promoted_nominal_mode); 3154 if (!DECL_ARTIFICIAL (parm)) 3155 mark_user_reg (parmreg); 3156 3157 /* If this was an item that we received a pointer to, 3158 set rtl appropriately. */ 3159 if (data->passed_pointer) 3160 { 3161 rtl = gen_rtx_MEM (TYPE_MODE (TREE_TYPE (data->passed_type)), parmreg); 3162 set_mem_attributes (rtl, parm, 1); 3163 } 3164 else 3165 rtl = parmreg; 3166 3167 assign_parm_remove_parallels (data); 3168 3169 /* Copy the value into the register, thus bridging between 3170 assign_parm_find_data_types and expand_expr_real_1. */ 3171 3172 equiv_stack_parm = data->stack_parm; 3173 validated_mem = validize_mem (copy_rtx (data->entry_parm)); 3174 3175 need_conversion = (data->nominal_mode != data->passed_mode 3176 || promoted_nominal_mode != data->promoted_mode); 3177 moved = false; 3178 3179 if (need_conversion 3180 && GET_MODE_CLASS (data->nominal_mode) == MODE_INT 3181 && data->nominal_mode == data->passed_mode 3182 && data->nominal_mode == GET_MODE (data->entry_parm)) 3183 { 3184 /* ENTRY_PARM has been converted to PROMOTED_MODE, its 3185 mode, by the caller. We now have to convert it to 3186 NOMINAL_MODE, if different. However, PARMREG may be in 3187 a different mode than NOMINAL_MODE if it is being stored 3188 promoted. 3189 3190 If ENTRY_PARM is a hard register, it might be in a register 3191 not valid for operating in its mode (e.g., an odd-numbered 3192 register for a DFmode). In that case, moves are the only 3193 thing valid, so we can't do a convert from there. This 3194 occurs when the calling sequence allow such misaligned 3195 usages. 3196 3197 In addition, the conversion may involve a call, which could 3198 clobber parameters which haven't been copied to pseudo 3199 registers yet. 3200 3201 First, we try to emit an insn which performs the necessary 3202 conversion. We verify that this insn does not clobber any 3203 hard registers. */ 3204 3205 enum insn_code icode; 3206 rtx op0, op1; 3207 3208 icode = can_extend_p (promoted_nominal_mode, data->passed_mode, 3209 unsignedp); 3210 3211 op0 = parmreg; 3212 op1 = validated_mem; 3213 if (icode != CODE_FOR_nothing 3214 && insn_operand_matches (icode, 0, op0) 3215 && insn_operand_matches (icode, 1, op1)) 3216 { 3217 enum rtx_code code = unsignedp ? ZERO_EXTEND : SIGN_EXTEND; 3218 rtx_insn *insn, *insns; 3219 rtx t = op1; 3220 HARD_REG_SET hardregs; 3221 3222 start_sequence (); 3223 /* If op1 is a hard register that is likely spilled, first 3224 force it into a pseudo, otherwise combiner might extend 3225 its lifetime too much. */ 3226 if (GET_CODE (t) == SUBREG) 3227 t = SUBREG_REG (t); 3228 if (REG_P (t) 3229 && HARD_REGISTER_P (t) 3230 && ! TEST_HARD_REG_BIT (fixed_reg_set, REGNO (t)) 3231 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (t)))) 3232 { 3233 t = gen_reg_rtx (GET_MODE (op1)); 3234 emit_move_insn (t, op1); 3235 } 3236 else 3237 t = op1; 3238 rtx_insn *pat = gen_extend_insn (op0, t, promoted_nominal_mode, 3239 data->passed_mode, unsignedp); 3240 emit_insn (pat); 3241 insns = get_insns (); 3242 3243 moved = true; 3244 CLEAR_HARD_REG_SET (hardregs); 3245 for (insn = insns; insn && moved; insn = NEXT_INSN (insn)) 3246 { 3247 if (INSN_P (insn)) 3248 note_stores (PATTERN (insn), record_hard_reg_sets, 3249 &hardregs); 3250 if (!hard_reg_set_empty_p (hardregs)) 3251 moved = false; 3252 } 3253 3254 end_sequence (); 3255 3256 if (moved) 3257 { 3258 emit_insn (insns); 3259 if (equiv_stack_parm != NULL_RTX) 3260 equiv_stack_parm = gen_rtx_fmt_e (code, GET_MODE (parmreg), 3261 equiv_stack_parm); 3262 } 3263 } 3264 } 3265 3266 if (moved) 3267 /* Nothing to do. */ 3268 ; 3269 else if (need_conversion) 3270 { 3271 /* We did not have an insn to convert directly, or the sequence 3272 generated appeared unsafe. We must first copy the parm to a 3273 pseudo reg, and save the conversion until after all 3274 parameters have been moved. */ 3275 3276 int save_tree_used; 3277 rtx tempreg = gen_reg_rtx (GET_MODE (data->entry_parm)); 3278 3279 emit_move_insn (tempreg, validated_mem); 3280 3281 push_to_sequence2 (all->first_conversion_insn, all->last_conversion_insn); 3282 tempreg = convert_to_mode (data->nominal_mode, tempreg, unsignedp); 3283 3284 if (partial_subreg_p (tempreg) 3285 && GET_MODE (tempreg) == data->nominal_mode 3286 && REG_P (SUBREG_REG (tempreg)) 3287 && data->nominal_mode == data->passed_mode 3288 && GET_MODE (SUBREG_REG (tempreg)) == GET_MODE (data->entry_parm)) 3289 { 3290 /* The argument is already sign/zero extended, so note it 3291 into the subreg. */ 3292 SUBREG_PROMOTED_VAR_P (tempreg) = 1; 3293 SUBREG_PROMOTED_SET (tempreg, unsignedp); 3294 } 3295 3296 /* TREE_USED gets set erroneously during expand_assignment. */ 3297 save_tree_used = TREE_USED (parm); 3298 SET_DECL_RTL (parm, rtl); 3299 expand_assignment (parm, make_tree (data->nominal_type, tempreg), false); 3300 SET_DECL_RTL (parm, NULL_RTX); 3301 TREE_USED (parm) = save_tree_used; 3302 all->first_conversion_insn = get_insns (); 3303 all->last_conversion_insn = get_last_insn (); 3304 end_sequence (); 3305 3306 did_conversion = true; 3307 } 3308 else 3309 emit_move_insn (parmreg, validated_mem); 3310 3311 /* If we were passed a pointer but the actual value can safely live 3312 in a register, retrieve it and use it directly. */ 3313 if (data->passed_pointer && TYPE_MODE (TREE_TYPE (parm)) != BLKmode) 3314 { 3315 /* We can't use nominal_mode, because it will have been set to 3316 Pmode above. We must use the actual mode of the parm. */ 3317 if (use_register_for_decl (parm)) 3318 { 3319 parmreg = gen_reg_rtx (TYPE_MODE (TREE_TYPE (parm))); 3320 mark_user_reg (parmreg); 3321 } 3322 else 3323 { 3324 int align = STACK_SLOT_ALIGNMENT (TREE_TYPE (parm), 3325 TYPE_MODE (TREE_TYPE (parm)), 3326 TYPE_ALIGN (TREE_TYPE (parm))); 3327 parmreg 3328 = assign_stack_local (TYPE_MODE (TREE_TYPE (parm)), 3329 GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (parm))), 3330 align); 3331 set_mem_attributes (parmreg, parm, 1); 3332 } 3333 3334 /* We need to preserve an address based on VIRTUAL_STACK_VARS_REGNUM for 3335 the debug info in case it is not legitimate. */ 3336 if (GET_MODE (parmreg) != GET_MODE (rtl)) 3337 { 3338 rtx tempreg = gen_reg_rtx (GET_MODE (rtl)); 3339 int unsigned_p = TYPE_UNSIGNED (TREE_TYPE (parm)); 3340 3341 push_to_sequence2 (all->first_conversion_insn, 3342 all->last_conversion_insn); 3343 emit_move_insn (tempreg, rtl); 3344 tempreg = convert_to_mode (GET_MODE (parmreg), tempreg, unsigned_p); 3345 emit_move_insn (MEM_P (parmreg) ? copy_rtx (parmreg) : parmreg, 3346 tempreg); 3347 all->first_conversion_insn = get_insns (); 3348 all->last_conversion_insn = get_last_insn (); 3349 end_sequence (); 3350 3351 did_conversion = true; 3352 } 3353 else 3354 emit_move_insn (MEM_P (parmreg) ? copy_rtx (parmreg) : parmreg, rtl); 3355 3356 rtl = parmreg; 3357 3358 /* STACK_PARM is the pointer, not the parm, and PARMREG is 3359 now the parm. */ 3360 data->stack_parm = NULL; 3361 } 3362 3363 set_parm_rtl (parm, rtl); 3364 3365 /* Mark the register as eliminable if we did no conversion and it was 3366 copied from memory at a fixed offset, and the arg pointer was not 3367 copied to a pseudo-reg. If the arg pointer is a pseudo reg or the 3368 offset formed an invalid address, such memory-equivalences as we 3369 make here would screw up life analysis for it. */ 3370 if (data->nominal_mode == data->passed_mode 3371 && !did_conversion 3372 && data->stack_parm != 0 3373 && MEM_P (data->stack_parm) 3374 && data->locate.offset.var == 0 3375 && reg_mentioned_p (virtual_incoming_args_rtx, 3376 XEXP (data->stack_parm, 0))) 3377 { 3378 rtx_insn *linsn = get_last_insn (); 3379 rtx_insn *sinsn; 3380 rtx set; 3381 3382 /* Mark complex types separately. */ 3383 if (GET_CODE (parmreg) == CONCAT) 3384 { 3385 scalar_mode submode = GET_MODE_INNER (GET_MODE (parmreg)); 3386 int regnor = REGNO (XEXP (parmreg, 0)); 3387 int regnoi = REGNO (XEXP (parmreg, 1)); 3388 rtx stackr = adjust_address_nv (data->stack_parm, submode, 0); 3389 rtx stacki = adjust_address_nv (data->stack_parm, submode, 3390 GET_MODE_SIZE (submode)); 3391 3392 /* Scan backwards for the set of the real and 3393 imaginary parts. */ 3394 for (sinsn = linsn; sinsn != 0; 3395 sinsn = prev_nonnote_insn (sinsn)) 3396 { 3397 set = single_set (sinsn); 3398 if (set == 0) 3399 continue; 3400 3401 if (SET_DEST (set) == regno_reg_rtx [regnoi]) 3402 set_unique_reg_note (sinsn, REG_EQUIV, stacki); 3403 else if (SET_DEST (set) == regno_reg_rtx [regnor]) 3404 set_unique_reg_note (sinsn, REG_EQUIV, stackr); 3405 } 3406 } 3407 else 3408 set_dst_reg_note (linsn, REG_EQUIV, equiv_stack_parm, parmreg); 3409 } 3410 3411 /* For pointer data type, suggest pointer register. */ 3412 if (POINTER_TYPE_P (TREE_TYPE (parm))) 3413 mark_reg_pointer (parmreg, 3414 TYPE_ALIGN (TREE_TYPE (TREE_TYPE (parm)))); 3415 } 3416 3417 /* A subroutine of assign_parms. Allocate stack space to hold the current 3418 parameter. Get it there. Perform all ABI specified conversions. */ 3419 3420 static void 3421 assign_parm_setup_stack (struct assign_parm_data_all *all, tree parm, 3422 struct assign_parm_data_one *data) 3423 { 3424 /* Value must be stored in the stack slot STACK_PARM during function 3425 execution. */ 3426 bool to_conversion = false; 3427 3428 assign_parm_remove_parallels (data); 3429 3430 if (data->promoted_mode != data->nominal_mode) 3431 { 3432 /* Conversion is required. */ 3433 rtx tempreg = gen_reg_rtx (GET_MODE (data->entry_parm)); 3434 3435 emit_move_insn (tempreg, validize_mem (copy_rtx (data->entry_parm))); 3436 3437 push_to_sequence2 (all->first_conversion_insn, all->last_conversion_insn); 3438 to_conversion = true; 3439 3440 data->entry_parm = convert_to_mode (data->nominal_mode, tempreg, 3441 TYPE_UNSIGNED (TREE_TYPE (parm))); 3442 3443 if (data->stack_parm) 3444 { 3445 poly_int64 offset 3446 = subreg_lowpart_offset (data->nominal_mode, 3447 GET_MODE (data->stack_parm)); 3448 /* ??? This may need a big-endian conversion on sparc64. */ 3449 data->stack_parm 3450 = adjust_address (data->stack_parm, data->nominal_mode, 0); 3451 if (maybe_ne (offset, 0) && MEM_OFFSET_KNOWN_P (data->stack_parm)) 3452 set_mem_offset (data->stack_parm, 3453 MEM_OFFSET (data->stack_parm) + offset); 3454 } 3455 } 3456 3457 if (data->entry_parm != data->stack_parm) 3458 { 3459 rtx src, dest; 3460 3461 if (data->stack_parm == 0) 3462 { 3463 int align = STACK_SLOT_ALIGNMENT (data->passed_type, 3464 GET_MODE (data->entry_parm), 3465 TYPE_ALIGN (data->passed_type)); 3466 data->stack_parm 3467 = assign_stack_local (GET_MODE (data->entry_parm), 3468 GET_MODE_SIZE (GET_MODE (data->entry_parm)), 3469 align); 3470 set_mem_attributes (data->stack_parm, parm, 1); 3471 } 3472 3473 dest = validize_mem (copy_rtx (data->stack_parm)); 3474 src = validize_mem (copy_rtx (data->entry_parm)); 3475 3476 if (TYPE_EMPTY_P (data->passed_type)) 3477 /* Empty types don't really need to be copied. */; 3478 else if (MEM_P (src)) 3479 { 3480 /* Use a block move to handle potentially misaligned entry_parm. */ 3481 if (!to_conversion) 3482 push_to_sequence2 (all->first_conversion_insn, 3483 all->last_conversion_insn); 3484 to_conversion = true; 3485 3486 emit_block_move (dest, src, 3487 GEN_INT (int_size_in_bytes (data->passed_type)), 3488 BLOCK_OP_NORMAL); 3489 } 3490 else 3491 { 3492 if (!REG_P (src)) 3493 src = force_reg (GET_MODE (src), src); 3494 emit_move_insn (dest, src); 3495 } 3496 } 3497 3498 if (to_conversion) 3499 { 3500 all->first_conversion_insn = get_insns (); 3501 all->last_conversion_insn = get_last_insn (); 3502 end_sequence (); 3503 } 3504 3505 set_parm_rtl (parm, data->stack_parm); 3506 } 3507 3508 /* A subroutine of assign_parms. If the ABI splits complex arguments, then 3509 undo the frobbing that we did in assign_parms_augmented_arg_list. */ 3510 3511 static void 3512 assign_parms_unsplit_complex (struct assign_parm_data_all *all, 3513 vec<tree> fnargs) 3514 { 3515 tree parm; 3516 tree orig_fnargs = all->orig_fnargs; 3517 unsigned i = 0; 3518 3519 for (parm = orig_fnargs; parm; parm = TREE_CHAIN (parm), ++i) 3520 { 3521 if (TREE_CODE (TREE_TYPE (parm)) == COMPLEX_TYPE 3522 && targetm.calls.split_complex_arg (TREE_TYPE (parm))) 3523 { 3524 rtx tmp, real, imag; 3525 scalar_mode inner = GET_MODE_INNER (DECL_MODE (parm)); 3526 3527 real = DECL_RTL (fnargs[i]); 3528 imag = DECL_RTL (fnargs[i + 1]); 3529 if (inner != GET_MODE (real)) 3530 { 3531 real = gen_lowpart_SUBREG (inner, real); 3532 imag = gen_lowpart_SUBREG (inner, imag); 3533 } 3534 3535 if (TREE_ADDRESSABLE (parm)) 3536 { 3537 rtx rmem, imem; 3538 HOST_WIDE_INT size = int_size_in_bytes (TREE_TYPE (parm)); 3539 int align = STACK_SLOT_ALIGNMENT (TREE_TYPE (parm), 3540 DECL_MODE (parm), 3541 TYPE_ALIGN (TREE_TYPE (parm))); 3542 3543 /* split_complex_arg put the real and imag parts in 3544 pseudos. Move them to memory. */ 3545 tmp = assign_stack_local (DECL_MODE (parm), size, align); 3546 set_mem_attributes (tmp, parm, 1); 3547 rmem = adjust_address_nv (tmp, inner, 0); 3548 imem = adjust_address_nv (tmp, inner, GET_MODE_SIZE (inner)); 3549 push_to_sequence2 (all->first_conversion_insn, 3550 all->last_conversion_insn); 3551 emit_move_insn (rmem, real); 3552 emit_move_insn (imem, imag); 3553 all->first_conversion_insn = get_insns (); 3554 all->last_conversion_insn = get_last_insn (); 3555 end_sequence (); 3556 } 3557 else 3558 tmp = gen_rtx_CONCAT (DECL_MODE (parm), real, imag); 3559 set_parm_rtl (parm, tmp); 3560 3561 real = DECL_INCOMING_RTL (fnargs[i]); 3562 imag = DECL_INCOMING_RTL (fnargs[i + 1]); 3563 if (inner != GET_MODE (real)) 3564 { 3565 real = gen_lowpart_SUBREG (inner, real); 3566 imag = gen_lowpart_SUBREG (inner, imag); 3567 } 3568 tmp = gen_rtx_CONCAT (DECL_MODE (parm), real, imag); 3569 set_decl_incoming_rtl (parm, tmp, false); 3570 i++; 3571 } 3572 } 3573 } 3574 3575 /* Load bounds of PARM from bounds table. */ 3576 static void 3577 assign_parm_load_bounds (struct assign_parm_data_one *data, 3578 tree parm, 3579 rtx entry, 3580 unsigned bound_no) 3581 { 3582 bitmap_iterator bi; 3583 unsigned i, offs = 0; 3584 int bnd_no = -1; 3585 rtx slot = NULL, ptr = NULL; 3586 3587 if (parm) 3588 { 3589 bitmap slots; 3590 bitmap_obstack_initialize (NULL); 3591 slots = BITMAP_ALLOC (NULL); 3592 chkp_find_bound_slots (TREE_TYPE (parm), slots); 3593 EXECUTE_IF_SET_IN_BITMAP (slots, 0, i, bi) 3594 { 3595 if (bound_no) 3596 bound_no--; 3597 else 3598 { 3599 bnd_no = i; 3600 break; 3601 } 3602 } 3603 BITMAP_FREE (slots); 3604 bitmap_obstack_release (NULL); 3605 } 3606 3607 /* We may have bounds not associated with any pointer. */ 3608 if (bnd_no != -1) 3609 offs = bnd_no * POINTER_SIZE / BITS_PER_UNIT; 3610 3611 /* Find associated pointer. */ 3612 if (bnd_no == -1) 3613 { 3614 /* If bounds are not associated with any bounds, 3615 then it is passed in a register or special slot. */ 3616 gcc_assert (data->entry_parm); 3617 ptr = const0_rtx; 3618 } 3619 else if (MEM_P (entry)) 3620 slot = adjust_address (entry, Pmode, offs); 3621 else if (REG_P (entry)) 3622 ptr = gen_rtx_REG (Pmode, REGNO (entry) + bnd_no); 3623 else if (GET_CODE (entry) == PARALLEL) 3624 ptr = chkp_get_value_with_offs (entry, GEN_INT (offs)); 3625 else 3626 gcc_unreachable (); 3627 data->entry_parm = targetm.calls.load_bounds_for_arg (slot, ptr, 3628 data->entry_parm); 3629 } 3630 3631 /* Assign RTL expressions to the function's bounds parameters BNDARGS. */ 3632 3633 static void 3634 assign_bounds (vec<bounds_parm_data> &bndargs, 3635 struct assign_parm_data_all &all, 3636 bool assign_regs, bool assign_special, 3637 bool assign_bt) 3638 { 3639 unsigned i, pass; 3640 bounds_parm_data *pbdata; 3641 3642 if (!bndargs.exists ()) 3643 return; 3644 3645 /* We make few passes to store input bounds. Firstly handle bounds 3646 passed in registers. After that we load bounds passed in special 3647 slots. Finally we load bounds from Bounds Table. */ 3648 for (pass = 0; pass < 3; pass++) 3649 FOR_EACH_VEC_ELT (bndargs, i, pbdata) 3650 { 3651 /* Pass 0 => regs only. */ 3652 if (pass == 0 3653 && (!assign_regs 3654 ||(!pbdata->parm_data.entry_parm 3655 || GET_CODE (pbdata->parm_data.entry_parm) != REG))) 3656 continue; 3657 /* Pass 1 => slots only. */ 3658 else if (pass == 1 3659 && (!assign_special 3660 || (!pbdata->parm_data.entry_parm 3661 || GET_CODE (pbdata->parm_data.entry_parm) == REG))) 3662 continue; 3663 /* Pass 2 => BT only. */ 3664 else if (pass == 2 3665 && (!assign_bt 3666 || pbdata->parm_data.entry_parm)) 3667 continue; 3668 3669 if (!pbdata->parm_data.entry_parm 3670 || GET_CODE (pbdata->parm_data.entry_parm) != REG) 3671 assign_parm_load_bounds (&pbdata->parm_data, pbdata->ptr_parm, 3672 pbdata->ptr_entry, pbdata->bound_no); 3673 3674 set_decl_incoming_rtl (pbdata->bounds_parm, 3675 pbdata->parm_data.entry_parm, false); 3676 3677 if (assign_parm_setup_block_p (&pbdata->parm_data)) 3678 assign_parm_setup_block (&all, pbdata->bounds_parm, 3679 &pbdata->parm_data); 3680 else if (pbdata->parm_data.passed_pointer 3681 || use_register_for_decl (pbdata->bounds_parm)) 3682 assign_parm_setup_reg (&all, pbdata->bounds_parm, 3683 &pbdata->parm_data); 3684 else 3685 assign_parm_setup_stack (&all, pbdata->bounds_parm, 3686 &pbdata->parm_data); 3687 } 3688 } 3689 3690 /* Assign RTL expressions to the function's parameters. This may involve 3691 copying them into registers and using those registers as the DECL_RTL. */ 3692 3693 static void 3694 assign_parms (tree fndecl) 3695 { 3696 struct assign_parm_data_all all; 3697 tree parm; 3698 vec<tree> fnargs; 3699 unsigned i, bound_no = 0; 3700 tree last_arg = NULL; 3701 rtx last_arg_entry = NULL; 3702 vec<bounds_parm_data> bndargs = vNULL; 3703 bounds_parm_data bdata; 3704 3705 crtl->args.internal_arg_pointer 3706 = targetm.calls.internal_arg_pointer (); 3707 3708 assign_parms_initialize_all (&all); 3709 fnargs = assign_parms_augmented_arg_list (&all); 3710 3711 FOR_EACH_VEC_ELT (fnargs, i, parm) 3712 { 3713 struct assign_parm_data_one data; 3714 3715 /* Extract the type of PARM; adjust it according to ABI. */ 3716 assign_parm_find_data_types (&all, parm, &data); 3717 3718 /* Early out for errors and void parameters. */ 3719 if (data.passed_mode == VOIDmode) 3720 { 3721 SET_DECL_RTL (parm, const0_rtx); 3722 DECL_INCOMING_RTL (parm) = DECL_RTL (parm); 3723 continue; 3724 } 3725 3726 /* Estimate stack alignment from parameter alignment. */ 3727 if (SUPPORTS_STACK_ALIGNMENT) 3728 { 3729 unsigned int align 3730 = targetm.calls.function_arg_boundary (data.promoted_mode, 3731 data.passed_type); 3732 align = MINIMUM_ALIGNMENT (data.passed_type, data.promoted_mode, 3733 align); 3734 if (TYPE_ALIGN (data.nominal_type) > align) 3735 align = MINIMUM_ALIGNMENT (data.nominal_type, 3736 TYPE_MODE (data.nominal_type), 3737 TYPE_ALIGN (data.nominal_type)); 3738 if (crtl->stack_alignment_estimated < align) 3739 { 3740 gcc_assert (!crtl->stack_realign_processed); 3741 crtl->stack_alignment_estimated = align; 3742 } 3743 } 3744 3745 /* Find out where the parameter arrives in this function. */ 3746 assign_parm_find_entry_rtl (&all, &data); 3747 3748 /* Find out where stack space for this parameter might be. */ 3749 if (assign_parm_is_stack_parm (&all, &data)) 3750 { 3751 assign_parm_find_stack_rtl (parm, &data); 3752 assign_parm_adjust_entry_rtl (&data); 3753 /* For arguments that occupy no space in the parameter 3754 passing area, have non-zero size and have address taken, 3755 force creation of a stack slot so that they have distinct 3756 address from other parameters. */ 3757 if (TYPE_EMPTY_P (data.passed_type) 3758 && TREE_ADDRESSABLE (parm) 3759 && data.entry_parm == data.stack_parm 3760 && MEM_P (data.entry_parm) 3761 && int_size_in_bytes (data.passed_type)) 3762 data.stack_parm = NULL_RTX; 3763 } 3764 if (!POINTER_BOUNDS_TYPE_P (data.passed_type)) 3765 { 3766 /* Remember where last non bounds arg was passed in case 3767 we have to load associated bounds for it from Bounds 3768 Table. */ 3769 last_arg = parm; 3770 last_arg_entry = data.entry_parm; 3771 bound_no = 0; 3772 } 3773 /* Record permanently how this parm was passed. */ 3774 if (data.passed_pointer) 3775 { 3776 rtx incoming_rtl 3777 = gen_rtx_MEM (TYPE_MODE (TREE_TYPE (data.passed_type)), 3778 data.entry_parm); 3779 set_decl_incoming_rtl (parm, incoming_rtl, true); 3780 } 3781 else 3782 set_decl_incoming_rtl (parm, data.entry_parm, false); 3783 3784 assign_parm_adjust_stack_rtl (&data); 3785 3786 /* Bounds should be loaded in the particular order to 3787 have registers allocated correctly. Collect info about 3788 input bounds and load them later. */ 3789 if (POINTER_BOUNDS_TYPE_P (data.passed_type)) 3790 { 3791 /* Expect bounds in instrumented functions only. */ 3792 gcc_assert (chkp_function_instrumented_p (fndecl)); 3793 3794 bdata.parm_data = data; 3795 bdata.bounds_parm = parm; 3796 bdata.ptr_parm = last_arg; 3797 bdata.ptr_entry = last_arg_entry; 3798 bdata.bound_no = bound_no; 3799 bndargs.safe_push (bdata); 3800 } 3801 else 3802 { 3803 if (assign_parm_setup_block_p (&data)) 3804 assign_parm_setup_block (&all, parm, &data); 3805 else if (data.passed_pointer || use_register_for_decl (parm)) 3806 assign_parm_setup_reg (&all, parm, &data); 3807 else 3808 assign_parm_setup_stack (&all, parm, &data); 3809 } 3810 3811 if (cfun->stdarg && !DECL_CHAIN (parm)) 3812 { 3813 int pretend_bytes = 0; 3814 3815 assign_parms_setup_varargs (&all, &data, false); 3816 3817 if (chkp_function_instrumented_p (fndecl)) 3818 { 3819 /* We expect this is the last parm. Otherwise it is wrong 3820 to assign bounds right now. */ 3821 gcc_assert (i == (fnargs.length () - 1)); 3822 assign_bounds (bndargs, all, true, false, false); 3823 targetm.calls.setup_incoming_vararg_bounds (all.args_so_far, 3824 data.promoted_mode, 3825 data.passed_type, 3826 &pretend_bytes, 3827 false); 3828 assign_bounds (bndargs, all, false, true, true); 3829 bndargs.release (); 3830 } 3831 } 3832 3833 /* Update info on where next arg arrives in registers. */ 3834 targetm.calls.function_arg_advance (all.args_so_far, data.promoted_mode, 3835 data.passed_type, data.named_arg); 3836 3837 if (POINTER_BOUNDS_TYPE_P (data.passed_type)) 3838 bound_no++; 3839 } 3840 3841 assign_bounds (bndargs, all, true, true, true); 3842 bndargs.release (); 3843 3844 if (targetm.calls.split_complex_arg) 3845 assign_parms_unsplit_complex (&all, fnargs); 3846 3847 fnargs.release (); 3848 3849 /* Output all parameter conversion instructions (possibly including calls) 3850 now that all parameters have been copied out of hard registers. */ 3851 emit_insn (all.first_conversion_insn); 3852 3853 /* Estimate reload stack alignment from scalar return mode. */ 3854 if (SUPPORTS_STACK_ALIGNMENT) 3855 { 3856 if (DECL_RESULT (fndecl)) 3857 { 3858 tree type = TREE_TYPE (DECL_RESULT (fndecl)); 3859 machine_mode mode = TYPE_MODE (type); 3860 3861 if (mode != BLKmode 3862 && mode != VOIDmode 3863 && !AGGREGATE_TYPE_P (type)) 3864 { 3865 unsigned int align = GET_MODE_ALIGNMENT (mode); 3866 if (crtl->stack_alignment_estimated < align) 3867 { 3868 gcc_assert (!crtl->stack_realign_processed); 3869 crtl->stack_alignment_estimated = align; 3870 } 3871 } 3872 } 3873 } 3874 3875 /* If we are receiving a struct value address as the first argument, set up 3876 the RTL for the function result. As this might require code to convert 3877 the transmitted address to Pmode, we do this here to ensure that possible 3878 preliminary conversions of the address have been emitted already. */ 3879 if (all.function_result_decl) 3880 { 3881 tree result = DECL_RESULT (current_function_decl); 3882 rtx addr = DECL_RTL (all.function_result_decl); 3883 rtx x; 3884 3885 if (DECL_BY_REFERENCE (result)) 3886 { 3887 SET_DECL_VALUE_EXPR (result, all.function_result_decl); 3888 x = addr; 3889 } 3890 else 3891 { 3892 SET_DECL_VALUE_EXPR (result, 3893 build1 (INDIRECT_REF, TREE_TYPE (result), 3894 all.function_result_decl)); 3895 addr = convert_memory_address (Pmode, addr); 3896 x = gen_rtx_MEM (DECL_MODE (result), addr); 3897 set_mem_attributes (x, result, 1); 3898 } 3899 3900 DECL_HAS_VALUE_EXPR_P (result) = 1; 3901 3902 set_parm_rtl (result, x); 3903 } 3904 3905 /* We have aligned all the args, so add space for the pretend args. */ 3906 crtl->args.pretend_args_size = all.pretend_args_size; 3907 all.stack_args_size.constant += all.extra_pretend_bytes; 3908 crtl->args.size = all.stack_args_size.constant; 3909 3910 /* Adjust function incoming argument size for alignment and 3911 minimum length. */ 3912 3913 crtl->args.size = upper_bound (crtl->args.size, all.reg_parm_stack_space); 3914 crtl->args.size = aligned_upper_bound (crtl->args.size, 3915 PARM_BOUNDARY / BITS_PER_UNIT); 3916 3917 if (ARGS_GROW_DOWNWARD) 3918 { 3919 crtl->args.arg_offset_rtx 3920 = (all.stack_args_size.var == 0 3921 ? gen_int_mode (-all.stack_args_size.constant, Pmode) 3922 : expand_expr (size_diffop (all.stack_args_size.var, 3923 size_int (-all.stack_args_size.constant)), 3924 NULL_RTX, VOIDmode, EXPAND_NORMAL)); 3925 } 3926 else 3927 crtl->args.arg_offset_rtx = ARGS_SIZE_RTX (all.stack_args_size); 3928 3929 /* See how many bytes, if any, of its args a function should try to pop 3930 on return. */ 3931 3932 crtl->args.pops_args = targetm.calls.return_pops_args (fndecl, 3933 TREE_TYPE (fndecl), 3934 crtl->args.size); 3935 3936 /* For stdarg.h function, save info about 3937 regs and stack space used by the named args. */ 3938 3939 crtl->args.info = all.args_so_far_v; 3940 3941 /* Set the rtx used for the function return value. Put this in its 3942 own variable so any optimizers that need this information don't have 3943 to include tree.h. Do this here so it gets done when an inlined 3944 function gets output. */ 3945 3946 crtl->return_rtx 3947 = (DECL_RTL_SET_P (DECL_RESULT (fndecl)) 3948 ? DECL_RTL (DECL_RESULT (fndecl)) : NULL_RTX); 3949 3950 /* If scalar return value was computed in a pseudo-reg, or was a named 3951 return value that got dumped to the stack, copy that to the hard 3952 return register. */ 3953 if (DECL_RTL_SET_P (DECL_RESULT (fndecl))) 3954 { 3955 tree decl_result = DECL_RESULT (fndecl); 3956 rtx decl_rtl = DECL_RTL (decl_result); 3957 3958 if (REG_P (decl_rtl) 3959 ? REGNO (decl_rtl) >= FIRST_PSEUDO_REGISTER 3960 : DECL_REGISTER (decl_result)) 3961 { 3962 rtx real_decl_rtl; 3963 3964 real_decl_rtl = targetm.calls.function_value (TREE_TYPE (decl_result), 3965 fndecl, true); 3966 if (chkp_function_instrumented_p (fndecl)) 3967 crtl->return_bnd 3968 = targetm.calls.chkp_function_value_bounds (TREE_TYPE (decl_result), 3969 fndecl, true); 3970 REG_FUNCTION_VALUE_P (real_decl_rtl) = 1; 3971 /* The delay slot scheduler assumes that crtl->return_rtx 3972 holds the hard register containing the return value, not a 3973 temporary pseudo. */ 3974 crtl->return_rtx = real_decl_rtl; 3975 } 3976 } 3977 } 3978 3979 /* A subroutine of gimplify_parameters, invoked via walk_tree. 3980 For all seen types, gimplify their sizes. */ 3981 3982 static tree 3983 gimplify_parm_type (tree *tp, int *walk_subtrees, void *data) 3984 { 3985 tree t = *tp; 3986 3987 *walk_subtrees = 0; 3988 if (TYPE_P (t)) 3989 { 3990 if (POINTER_TYPE_P (t)) 3991 *walk_subtrees = 1; 3992 else if (TYPE_SIZE (t) && !TREE_CONSTANT (TYPE_SIZE (t)) 3993 && !TYPE_SIZES_GIMPLIFIED (t)) 3994 { 3995 gimplify_type_sizes (t, (gimple_seq *) data); 3996 *walk_subtrees = 1; 3997 } 3998 } 3999 4000 return NULL; 4001 } 4002 4003 /* Gimplify the parameter list for current_function_decl. This involves 4004 evaluating SAVE_EXPRs of variable sized parameters and generating code 4005 to implement callee-copies reference parameters. Returns a sequence of 4006 statements to add to the beginning of the function. */ 4007 4008 gimple_seq 4009 gimplify_parameters (gimple_seq *cleanup) 4010 { 4011 struct assign_parm_data_all all; 4012 tree parm; 4013 gimple_seq stmts = NULL; 4014 vec<tree> fnargs; 4015 unsigned i; 4016 4017 assign_parms_initialize_all (&all); 4018 fnargs = assign_parms_augmented_arg_list (&all); 4019 4020 FOR_EACH_VEC_ELT (fnargs, i, parm) 4021 { 4022 struct assign_parm_data_one data; 4023 4024 /* Extract the type of PARM; adjust it according to ABI. */ 4025 assign_parm_find_data_types (&all, parm, &data); 4026 4027 /* Early out for errors and void parameters. */ 4028 if (data.passed_mode == VOIDmode || DECL_SIZE (parm) == NULL) 4029 continue; 4030 4031 /* Update info on where next arg arrives in registers. */ 4032 targetm.calls.function_arg_advance (all.args_so_far, data.promoted_mode, 4033 data.passed_type, data.named_arg); 4034 4035 /* ??? Once upon a time variable_size stuffed parameter list 4036 SAVE_EXPRs (amongst others) onto a pending sizes list. This 4037 turned out to be less than manageable in the gimple world. 4038 Now we have to hunt them down ourselves. */ 4039 walk_tree_without_duplicates (&data.passed_type, 4040 gimplify_parm_type, &stmts); 4041 4042 if (TREE_CODE (DECL_SIZE_UNIT (parm)) != INTEGER_CST) 4043 { 4044 gimplify_one_sizepos (&DECL_SIZE (parm), &stmts); 4045 gimplify_one_sizepos (&DECL_SIZE_UNIT (parm), &stmts); 4046 } 4047 4048 if (data.passed_pointer) 4049 { 4050 tree type = TREE_TYPE (data.passed_type); 4051 if (reference_callee_copied (&all.args_so_far_v, TYPE_MODE (type), 4052 type, data.named_arg)) 4053 { 4054 tree local, t; 4055 4056 /* For constant-sized objects, this is trivial; for 4057 variable-sized objects, we have to play games. */ 4058 if (TREE_CODE (DECL_SIZE_UNIT (parm)) == INTEGER_CST 4059 && !(flag_stack_check == GENERIC_STACK_CHECK 4060 && compare_tree_int (DECL_SIZE_UNIT (parm), 4061 STACK_CHECK_MAX_VAR_SIZE) > 0)) 4062 { 4063 local = create_tmp_var (type, get_name (parm)); 4064 DECL_IGNORED_P (local) = 0; 4065 /* If PARM was addressable, move that flag over 4066 to the local copy, as its address will be taken, 4067 not the PARMs. Keep the parms address taken 4068 as we'll query that flag during gimplification. */ 4069 if (TREE_ADDRESSABLE (parm)) 4070 TREE_ADDRESSABLE (local) = 1; 4071 else if (TREE_CODE (type) == COMPLEX_TYPE 4072 || TREE_CODE (type) == VECTOR_TYPE) 4073 DECL_GIMPLE_REG_P (local) = 1; 4074 4075 if (!is_gimple_reg (local) 4076 && flag_stack_reuse != SR_NONE) 4077 { 4078 tree clobber = build_constructor (type, NULL); 4079 gimple *clobber_stmt; 4080 TREE_THIS_VOLATILE (clobber) = 1; 4081 clobber_stmt = gimple_build_assign (local, clobber); 4082 gimple_seq_add_stmt (cleanup, clobber_stmt); 4083 } 4084 } 4085 else 4086 { 4087 tree ptr_type, addr; 4088 4089 ptr_type = build_pointer_type (type); 4090 addr = create_tmp_reg (ptr_type, get_name (parm)); 4091 DECL_IGNORED_P (addr) = 0; 4092 local = build_fold_indirect_ref (addr); 4093 4094 t = build_alloca_call_expr (DECL_SIZE_UNIT (parm), 4095 DECL_ALIGN (parm), 4096 max_int_size_in_bytes (type)); 4097 /* The call has been built for a variable-sized object. */ 4098 CALL_ALLOCA_FOR_VAR_P (t) = 1; 4099 t = fold_convert (ptr_type, t); 4100 t = build2 (MODIFY_EXPR, TREE_TYPE (addr), addr, t); 4101 gimplify_and_add (t, &stmts); 4102 } 4103 4104 gimplify_assign (local, parm, &stmts); 4105 4106 SET_DECL_VALUE_EXPR (parm, local); 4107 DECL_HAS_VALUE_EXPR_P (parm) = 1; 4108 } 4109 } 4110 } 4111 4112 fnargs.release (); 4113 4114 return stmts; 4115 } 4116 4117 /* Compute the size and offset from the start of the stacked arguments for a 4118 parm passed in mode PASSED_MODE and with type TYPE. 4119 4120 INITIAL_OFFSET_PTR points to the current offset into the stacked 4121 arguments. 4122 4123 The starting offset and size for this parm are returned in 4124 LOCATE->OFFSET and LOCATE->SIZE, respectively. When IN_REGS is 4125 nonzero, the offset is that of stack slot, which is returned in 4126 LOCATE->SLOT_OFFSET. LOCATE->ALIGNMENT_PAD is the amount of 4127 padding required from the initial offset ptr to the stack slot. 4128 4129 IN_REGS is nonzero if the argument will be passed in registers. It will 4130 never be set if REG_PARM_STACK_SPACE is not defined. 4131 4132 REG_PARM_STACK_SPACE is the number of bytes of stack space reserved 4133 for arguments which are passed in registers. 4134 4135 FNDECL is the function in which the argument was defined. 4136 4137 There are two types of rounding that are done. The first, controlled by 4138 TARGET_FUNCTION_ARG_BOUNDARY, forces the offset from the start of the 4139 argument list to be aligned to the specific boundary (in bits). This 4140 rounding affects the initial and starting offsets, but not the argument 4141 size. 4142 4143 The second, controlled by TARGET_FUNCTION_ARG_PADDING and PARM_BOUNDARY, 4144 optionally rounds the size of the parm to PARM_BOUNDARY. The 4145 initial offset is not affected by this rounding, while the size always 4146 is and the starting offset may be. */ 4147 4148 /* LOCATE->OFFSET will be negative for ARGS_GROW_DOWNWARD case; 4149 INITIAL_OFFSET_PTR is positive because locate_and_pad_parm's 4150 callers pass in the total size of args so far as 4151 INITIAL_OFFSET_PTR. LOCATE->SIZE is always positive. */ 4152 4153 void 4154 locate_and_pad_parm (machine_mode passed_mode, tree type, int in_regs, 4155 int reg_parm_stack_space, int partial, 4156 tree fndecl ATTRIBUTE_UNUSED, 4157 struct args_size *initial_offset_ptr, 4158 struct locate_and_pad_arg_data *locate) 4159 { 4160 tree sizetree; 4161 pad_direction where_pad; 4162 unsigned int boundary, round_boundary; 4163 int part_size_in_regs; 4164 4165 /* If we have found a stack parm before we reach the end of the 4166 area reserved for registers, skip that area. */ 4167 if (! in_regs) 4168 { 4169 if (reg_parm_stack_space > 0) 4170 { 4171 if (initial_offset_ptr->var 4172 || !ordered_p (initial_offset_ptr->constant, 4173 reg_parm_stack_space)) 4174 { 4175 initial_offset_ptr->var 4176 = size_binop (MAX_EXPR, ARGS_SIZE_TREE (*initial_offset_ptr), 4177 ssize_int (reg_parm_stack_space)); 4178 initial_offset_ptr->constant = 0; 4179 } 4180 else 4181 initial_offset_ptr->constant 4182 = ordered_max (initial_offset_ptr->constant, 4183 reg_parm_stack_space); 4184 } 4185 } 4186 4187 part_size_in_regs = (reg_parm_stack_space == 0 ? partial : 0); 4188 4189 sizetree = (type 4190 ? arg_size_in_bytes (type) 4191 : size_int (GET_MODE_SIZE (passed_mode))); 4192 where_pad = targetm.calls.function_arg_padding (passed_mode, type); 4193 boundary = targetm.calls.function_arg_boundary (passed_mode, type); 4194 round_boundary = targetm.calls.function_arg_round_boundary (passed_mode, 4195 type); 4196 locate->where_pad = where_pad; 4197 4198 /* Alignment can't exceed MAX_SUPPORTED_STACK_ALIGNMENT. */ 4199 if (boundary > MAX_SUPPORTED_STACK_ALIGNMENT) 4200 boundary = MAX_SUPPORTED_STACK_ALIGNMENT; 4201 4202 locate->boundary = boundary; 4203 4204 if (SUPPORTS_STACK_ALIGNMENT) 4205 { 4206 /* stack_alignment_estimated can't change after stack has been 4207 realigned. */ 4208 if (crtl->stack_alignment_estimated < boundary) 4209 { 4210 if (!crtl->stack_realign_processed) 4211 crtl->stack_alignment_estimated = boundary; 4212 else 4213 { 4214 /* If stack is realigned and stack alignment value 4215 hasn't been finalized, it is OK not to increase 4216 stack_alignment_estimated. The bigger alignment 4217 requirement is recorded in stack_alignment_needed 4218 below. */ 4219 gcc_assert (!crtl->stack_realign_finalized 4220 && crtl->stack_realign_needed); 4221 } 4222 } 4223 } 4224 4225 /* Remember if the outgoing parameter requires extra alignment on the 4226 calling function side. */ 4227 if (crtl->stack_alignment_needed < boundary) 4228 crtl->stack_alignment_needed = boundary; 4229 if (crtl->preferred_stack_boundary < boundary) 4230 crtl->preferred_stack_boundary = boundary; 4231 4232 if (ARGS_GROW_DOWNWARD) 4233 { 4234 locate->slot_offset.constant = -initial_offset_ptr->constant; 4235 if (initial_offset_ptr->var) 4236 locate->slot_offset.var = size_binop (MINUS_EXPR, ssize_int (0), 4237 initial_offset_ptr->var); 4238 4239 { 4240 tree s2 = sizetree; 4241 if (where_pad != PAD_NONE 4242 && (!tree_fits_uhwi_p (sizetree) 4243 || (tree_to_uhwi (sizetree) * BITS_PER_UNIT) % round_boundary)) 4244 s2 = round_up (s2, round_boundary / BITS_PER_UNIT); 4245 SUB_PARM_SIZE (locate->slot_offset, s2); 4246 } 4247 4248 locate->slot_offset.constant += part_size_in_regs; 4249 4250 if (!in_regs || reg_parm_stack_space > 0) 4251 pad_to_arg_alignment (&locate->slot_offset, boundary, 4252 &locate->alignment_pad); 4253 4254 locate->size.constant = (-initial_offset_ptr->constant 4255 - locate->slot_offset.constant); 4256 if (initial_offset_ptr->var) 4257 locate->size.var = size_binop (MINUS_EXPR, 4258 size_binop (MINUS_EXPR, 4259 ssize_int (0), 4260 initial_offset_ptr->var), 4261 locate->slot_offset.var); 4262 4263 /* Pad_below needs the pre-rounded size to know how much to pad 4264 below. */ 4265 locate->offset = locate->slot_offset; 4266 if (where_pad == PAD_DOWNWARD) 4267 pad_below (&locate->offset, passed_mode, sizetree); 4268 4269 } 4270 else 4271 { 4272 if (!in_regs || reg_parm_stack_space > 0) 4273 pad_to_arg_alignment (initial_offset_ptr, boundary, 4274 &locate->alignment_pad); 4275 locate->slot_offset = *initial_offset_ptr; 4276 4277 #ifdef PUSH_ROUNDING 4278 if (passed_mode != BLKmode) 4279 sizetree = size_int (PUSH_ROUNDING (TREE_INT_CST_LOW (sizetree))); 4280 #endif 4281 4282 /* Pad_below needs the pre-rounded size to know how much to pad below 4283 so this must be done before rounding up. */ 4284 locate->offset = locate->slot_offset; 4285 if (where_pad == PAD_DOWNWARD) 4286 pad_below (&locate->offset, passed_mode, sizetree); 4287 4288 if (where_pad != PAD_NONE 4289 && (!tree_fits_uhwi_p (sizetree) 4290 || (tree_to_uhwi (sizetree) * BITS_PER_UNIT) % round_boundary)) 4291 sizetree = round_up (sizetree, round_boundary / BITS_PER_UNIT); 4292 4293 ADD_PARM_SIZE (locate->size, sizetree); 4294 4295 locate->size.constant -= part_size_in_regs; 4296 } 4297 4298 locate->offset.constant 4299 += targetm.calls.function_arg_offset (passed_mode, type); 4300 } 4301 4302 /* Round the stack offset in *OFFSET_PTR up to a multiple of BOUNDARY. 4303 BOUNDARY is measured in bits, but must be a multiple of a storage unit. */ 4304 4305 static void 4306 pad_to_arg_alignment (struct args_size *offset_ptr, int boundary, 4307 struct args_size *alignment_pad) 4308 { 4309 tree save_var = NULL_TREE; 4310 poly_int64 save_constant = 0; 4311 int boundary_in_bytes = boundary / BITS_PER_UNIT; 4312 poly_int64 sp_offset = STACK_POINTER_OFFSET; 4313 4314 #ifdef SPARC_STACK_BOUNDARY_HACK 4315 /* ??? The SPARC port may claim a STACK_BOUNDARY higher than 4316 the real alignment of %sp. However, when it does this, the 4317 alignment of %sp+STACK_POINTER_OFFSET is STACK_BOUNDARY. */ 4318 if (SPARC_STACK_BOUNDARY_HACK) 4319 sp_offset = 0; 4320 #endif 4321 4322 if (boundary > PARM_BOUNDARY) 4323 { 4324 save_var = offset_ptr->var; 4325 save_constant = offset_ptr->constant; 4326 } 4327 4328 alignment_pad->var = NULL_TREE; 4329 alignment_pad->constant = 0; 4330 4331 if (boundary > BITS_PER_UNIT) 4332 { 4333 int misalign; 4334 if (offset_ptr->var 4335 || !known_misalignment (offset_ptr->constant + sp_offset, 4336 boundary_in_bytes, &misalign)) 4337 { 4338 tree sp_offset_tree = ssize_int (sp_offset); 4339 tree offset = size_binop (PLUS_EXPR, 4340 ARGS_SIZE_TREE (*offset_ptr), 4341 sp_offset_tree); 4342 tree rounded; 4343 if (ARGS_GROW_DOWNWARD) 4344 rounded = round_down (offset, boundary / BITS_PER_UNIT); 4345 else 4346 rounded = round_up (offset, boundary / BITS_PER_UNIT); 4347 4348 offset_ptr->var = size_binop (MINUS_EXPR, rounded, sp_offset_tree); 4349 /* ARGS_SIZE_TREE includes constant term. */ 4350 offset_ptr->constant = 0; 4351 if (boundary > PARM_BOUNDARY) 4352 alignment_pad->var = size_binop (MINUS_EXPR, offset_ptr->var, 4353 save_var); 4354 } 4355 else 4356 { 4357 if (ARGS_GROW_DOWNWARD) 4358 offset_ptr->constant -= misalign; 4359 else 4360 offset_ptr->constant += -misalign & (boundary_in_bytes - 1); 4361 4362 if (boundary > PARM_BOUNDARY) 4363 alignment_pad->constant = offset_ptr->constant - save_constant; 4364 } 4365 } 4366 } 4367 4368 static void 4369 pad_below (struct args_size *offset_ptr, machine_mode passed_mode, tree sizetree) 4370 { 4371 unsigned int align = PARM_BOUNDARY / BITS_PER_UNIT; 4372 int misalign; 4373 if (passed_mode != BLKmode 4374 && known_misalignment (GET_MODE_SIZE (passed_mode), align, &misalign)) 4375 offset_ptr->constant += -misalign & (align - 1); 4376 else 4377 { 4378 if (TREE_CODE (sizetree) != INTEGER_CST 4379 || (TREE_INT_CST_LOW (sizetree) & (align - 1)) != 0) 4380 { 4381 /* Round the size up to multiple of PARM_BOUNDARY bits. */ 4382 tree s2 = round_up (sizetree, align); 4383 /* Add it in. */ 4384 ADD_PARM_SIZE (*offset_ptr, s2); 4385 SUB_PARM_SIZE (*offset_ptr, sizetree); 4386 } 4387 } 4388 } 4389 4390 4391 /* True if register REGNO was alive at a place where `setjmp' was 4392 called and was set more than once or is an argument. Such regs may 4393 be clobbered by `longjmp'. */ 4394 4395 static bool 4396 regno_clobbered_at_setjmp (bitmap setjmp_crosses, int regno) 4397 { 4398 /* There appear to be cases where some local vars never reach the 4399 backend but have bogus regnos. */ 4400 if (regno >= max_reg_num ()) 4401 return false; 4402 4403 return ((REG_N_SETS (regno) > 1 4404 || REGNO_REG_SET_P (df_get_live_out (ENTRY_BLOCK_PTR_FOR_FN (cfun)), 4405 regno)) 4406 && REGNO_REG_SET_P (setjmp_crosses, regno)); 4407 } 4408 4409 /* Walk the tree of blocks describing the binding levels within a 4410 function and warn about variables the might be killed by setjmp or 4411 vfork. This is done after calling flow_analysis before register 4412 allocation since that will clobber the pseudo-regs to hard 4413 regs. */ 4414 4415 static void 4416 setjmp_vars_warning (bitmap setjmp_crosses, tree block) 4417 { 4418 tree decl, sub; 4419 4420 for (decl = BLOCK_VARS (block); decl; decl = DECL_CHAIN (decl)) 4421 { 4422 if (VAR_P (decl) 4423 && DECL_RTL_SET_P (decl) 4424 && REG_P (DECL_RTL (decl)) 4425 && regno_clobbered_at_setjmp (setjmp_crosses, REGNO (DECL_RTL (decl)))) 4426 warning (OPT_Wclobbered, "variable %q+D might be clobbered by" 4427 " %<longjmp%> or %<vfork%>", decl); 4428 } 4429 4430 for (sub = BLOCK_SUBBLOCKS (block); sub; sub = BLOCK_CHAIN (sub)) 4431 setjmp_vars_warning (setjmp_crosses, sub); 4432 } 4433 4434 /* Do the appropriate part of setjmp_vars_warning 4435 but for arguments instead of local variables. */ 4436 4437 static void 4438 setjmp_args_warning (bitmap setjmp_crosses) 4439 { 4440 tree decl; 4441 for (decl = DECL_ARGUMENTS (current_function_decl); 4442 decl; decl = DECL_CHAIN (decl)) 4443 if (DECL_RTL (decl) != 0 4444 && REG_P (DECL_RTL (decl)) 4445 && regno_clobbered_at_setjmp (setjmp_crosses, REGNO (DECL_RTL (decl)))) 4446 warning (OPT_Wclobbered, 4447 "argument %q+D might be clobbered by %<longjmp%> or %<vfork%>", 4448 decl); 4449 } 4450 4451 /* Generate warning messages for variables live across setjmp. */ 4452 4453 void 4454 generate_setjmp_warnings (void) 4455 { 4456 bitmap setjmp_crosses = regstat_get_setjmp_crosses (); 4457 4458 if (n_basic_blocks_for_fn (cfun) == NUM_FIXED_BLOCKS 4459 || bitmap_empty_p (setjmp_crosses)) 4460 return; 4461 4462 setjmp_vars_warning (setjmp_crosses, DECL_INITIAL (current_function_decl)); 4463 setjmp_args_warning (setjmp_crosses); 4464 } 4465 4466 4467 /* Reverse the order of elements in the fragment chain T of blocks, 4468 and return the new head of the chain (old last element). 4469 In addition to that clear BLOCK_SAME_RANGE flags when needed 4470 and adjust BLOCK_SUPERCONTEXT from the super fragment to 4471 its super fragment origin. */ 4472 4473 static tree 4474 block_fragments_nreverse (tree t) 4475 { 4476 tree prev = 0, block, next, prev_super = 0; 4477 tree super = BLOCK_SUPERCONTEXT (t); 4478 if (BLOCK_FRAGMENT_ORIGIN (super)) 4479 super = BLOCK_FRAGMENT_ORIGIN (super); 4480 for (block = t; block; block = next) 4481 { 4482 next = BLOCK_FRAGMENT_CHAIN (block); 4483 BLOCK_FRAGMENT_CHAIN (block) = prev; 4484 if ((prev && !BLOCK_SAME_RANGE (prev)) 4485 || (BLOCK_FRAGMENT_CHAIN (BLOCK_SUPERCONTEXT (block)) 4486 != prev_super)) 4487 BLOCK_SAME_RANGE (block) = 0; 4488 prev_super = BLOCK_SUPERCONTEXT (block); 4489 BLOCK_SUPERCONTEXT (block) = super; 4490 prev = block; 4491 } 4492 t = BLOCK_FRAGMENT_ORIGIN (t); 4493 if (BLOCK_FRAGMENT_CHAIN (BLOCK_SUPERCONTEXT (t)) 4494 != prev_super) 4495 BLOCK_SAME_RANGE (t) = 0; 4496 BLOCK_SUPERCONTEXT (t) = super; 4497 return prev; 4498 } 4499 4500 /* Reverse the order of elements in the chain T of blocks, 4501 and return the new head of the chain (old last element). 4502 Also do the same on subblocks and reverse the order of elements 4503 in BLOCK_FRAGMENT_CHAIN as well. */ 4504 4505 static tree 4506 blocks_nreverse_all (tree t) 4507 { 4508 tree prev = 0, block, next; 4509 for (block = t; block; block = next) 4510 { 4511 next = BLOCK_CHAIN (block); 4512 BLOCK_CHAIN (block) = prev; 4513 if (BLOCK_FRAGMENT_CHAIN (block) 4514 && BLOCK_FRAGMENT_ORIGIN (block) == NULL_TREE) 4515 { 4516 BLOCK_FRAGMENT_CHAIN (block) 4517 = block_fragments_nreverse (BLOCK_FRAGMENT_CHAIN (block)); 4518 if (!BLOCK_SAME_RANGE (BLOCK_FRAGMENT_CHAIN (block))) 4519 BLOCK_SAME_RANGE (block) = 0; 4520 } 4521 BLOCK_SUBBLOCKS (block) = blocks_nreverse_all (BLOCK_SUBBLOCKS (block)); 4522 prev = block; 4523 } 4524 return prev; 4525 } 4526 4527 4528 /* Identify BLOCKs referenced by more than one NOTE_INSN_BLOCK_{BEG,END}, 4529 and create duplicate blocks. */ 4530 /* ??? Need an option to either create block fragments or to create 4531 abstract origin duplicates of a source block. It really depends 4532 on what optimization has been performed. */ 4533 4534 void 4535 reorder_blocks (void) 4536 { 4537 tree block = DECL_INITIAL (current_function_decl); 4538 4539 if (block == NULL_TREE) 4540 return; 4541 4542 auto_vec<tree, 10> block_stack; 4543 4544 /* Reset the TREE_ASM_WRITTEN bit for all blocks. */ 4545 clear_block_marks (block); 4546 4547 /* Prune the old trees away, so that they don't get in the way. */ 4548 BLOCK_SUBBLOCKS (block) = NULL_TREE; 4549 BLOCK_CHAIN (block) = NULL_TREE; 4550 4551 /* Recreate the block tree from the note nesting. */ 4552 reorder_blocks_1 (get_insns (), block, &block_stack); 4553 BLOCK_SUBBLOCKS (block) = blocks_nreverse_all (BLOCK_SUBBLOCKS (block)); 4554 } 4555 4556 /* Helper function for reorder_blocks. Reset TREE_ASM_WRITTEN. */ 4557 4558 void 4559 clear_block_marks (tree block) 4560 { 4561 while (block) 4562 { 4563 TREE_ASM_WRITTEN (block) = 0; 4564 clear_block_marks (BLOCK_SUBBLOCKS (block)); 4565 block = BLOCK_CHAIN (block); 4566 } 4567 } 4568 4569 static void 4570 reorder_blocks_1 (rtx_insn *insns, tree current_block, 4571 vec<tree> *p_block_stack) 4572 { 4573 rtx_insn *insn; 4574 tree prev_beg = NULL_TREE, prev_end = NULL_TREE; 4575 4576 for (insn = insns; insn; insn = NEXT_INSN (insn)) 4577 { 4578 if (NOTE_P (insn)) 4579 { 4580 if (NOTE_KIND (insn) == NOTE_INSN_BLOCK_BEG) 4581 { 4582 tree block = NOTE_BLOCK (insn); 4583 tree origin; 4584 4585 gcc_assert (BLOCK_FRAGMENT_ORIGIN (block) == NULL_TREE); 4586 origin = block; 4587 4588 if (prev_end) 4589 BLOCK_SAME_RANGE (prev_end) = 0; 4590 prev_end = NULL_TREE; 4591 4592 /* If we have seen this block before, that means it now 4593 spans multiple address regions. Create a new fragment. */ 4594 if (TREE_ASM_WRITTEN (block)) 4595 { 4596 tree new_block = copy_node (block); 4597 4598 BLOCK_SAME_RANGE (new_block) = 0; 4599 BLOCK_FRAGMENT_ORIGIN (new_block) = origin; 4600 BLOCK_FRAGMENT_CHAIN (new_block) 4601 = BLOCK_FRAGMENT_CHAIN (origin); 4602 BLOCK_FRAGMENT_CHAIN (origin) = new_block; 4603 4604 NOTE_BLOCK (insn) = new_block; 4605 block = new_block; 4606 } 4607 4608 if (prev_beg == current_block && prev_beg) 4609 BLOCK_SAME_RANGE (block) = 1; 4610 4611 prev_beg = origin; 4612 4613 BLOCK_SUBBLOCKS (block) = 0; 4614 TREE_ASM_WRITTEN (block) = 1; 4615 /* When there's only one block for the entire function, 4616 current_block == block and we mustn't do this, it 4617 will cause infinite recursion. */ 4618 if (block != current_block) 4619 { 4620 tree super; 4621 if (block != origin) 4622 gcc_assert (BLOCK_SUPERCONTEXT (origin) == current_block 4623 || BLOCK_FRAGMENT_ORIGIN (BLOCK_SUPERCONTEXT 4624 (origin)) 4625 == current_block); 4626 if (p_block_stack->is_empty ()) 4627 super = current_block; 4628 else 4629 { 4630 super = p_block_stack->last (); 4631 gcc_assert (super == current_block 4632 || BLOCK_FRAGMENT_ORIGIN (super) 4633 == current_block); 4634 } 4635 BLOCK_SUPERCONTEXT (block) = super; 4636 BLOCK_CHAIN (block) = BLOCK_SUBBLOCKS (current_block); 4637 BLOCK_SUBBLOCKS (current_block) = block; 4638 current_block = origin; 4639 } 4640 p_block_stack->safe_push (block); 4641 } 4642 else if (NOTE_KIND (insn) == NOTE_INSN_BLOCK_END) 4643 { 4644 NOTE_BLOCK (insn) = p_block_stack->pop (); 4645 current_block = BLOCK_SUPERCONTEXT (current_block); 4646 if (BLOCK_FRAGMENT_ORIGIN (current_block)) 4647 current_block = BLOCK_FRAGMENT_ORIGIN (current_block); 4648 prev_beg = NULL_TREE; 4649 prev_end = BLOCK_SAME_RANGE (NOTE_BLOCK (insn)) 4650 ? NOTE_BLOCK (insn) : NULL_TREE; 4651 } 4652 } 4653 else 4654 { 4655 prev_beg = NULL_TREE; 4656 if (prev_end) 4657 BLOCK_SAME_RANGE (prev_end) = 0; 4658 prev_end = NULL_TREE; 4659 } 4660 } 4661 } 4662 4663 /* Reverse the order of elements in the chain T of blocks, 4664 and return the new head of the chain (old last element). */ 4665 4666 tree 4667 blocks_nreverse (tree t) 4668 { 4669 tree prev = 0, block, next; 4670 for (block = t; block; block = next) 4671 { 4672 next = BLOCK_CHAIN (block); 4673 BLOCK_CHAIN (block) = prev; 4674 prev = block; 4675 } 4676 return prev; 4677 } 4678 4679 /* Concatenate two chains of blocks (chained through BLOCK_CHAIN) 4680 by modifying the last node in chain 1 to point to chain 2. */ 4681 4682 tree 4683 block_chainon (tree op1, tree op2) 4684 { 4685 tree t1; 4686 4687 if (!op1) 4688 return op2; 4689 if (!op2) 4690 return op1; 4691 4692 for (t1 = op1; BLOCK_CHAIN (t1); t1 = BLOCK_CHAIN (t1)) 4693 continue; 4694 BLOCK_CHAIN (t1) = op2; 4695 4696 #ifdef ENABLE_TREE_CHECKING 4697 { 4698 tree t2; 4699 for (t2 = op2; t2; t2 = BLOCK_CHAIN (t2)) 4700 gcc_assert (t2 != t1); 4701 } 4702 #endif 4703 4704 return op1; 4705 } 4706 4707 /* Count the subblocks of the list starting with BLOCK. If VECTOR is 4708 non-NULL, list them all into VECTOR, in a depth-first preorder 4709 traversal of the block tree. Also clear TREE_ASM_WRITTEN in all 4710 blocks. */ 4711 4712 static int 4713 all_blocks (tree block, tree *vector) 4714 { 4715 int n_blocks = 0; 4716 4717 while (block) 4718 { 4719 TREE_ASM_WRITTEN (block) = 0; 4720 4721 /* Record this block. */ 4722 if (vector) 4723 vector[n_blocks] = block; 4724 4725 ++n_blocks; 4726 4727 /* Record the subblocks, and their subblocks... */ 4728 n_blocks += all_blocks (BLOCK_SUBBLOCKS (block), 4729 vector ? vector + n_blocks : 0); 4730 block = BLOCK_CHAIN (block); 4731 } 4732 4733 return n_blocks; 4734 } 4735 4736 /* Return a vector containing all the blocks rooted at BLOCK. The 4737 number of elements in the vector is stored in N_BLOCKS_P. The 4738 vector is dynamically allocated; it is the caller's responsibility 4739 to call `free' on the pointer returned. */ 4740 4741 static tree * 4742 get_block_vector (tree block, int *n_blocks_p) 4743 { 4744 tree *block_vector; 4745 4746 *n_blocks_p = all_blocks (block, NULL); 4747 block_vector = XNEWVEC (tree, *n_blocks_p); 4748 all_blocks (block, block_vector); 4749 4750 return block_vector; 4751 } 4752 4753 static GTY(()) int next_block_index = 2; 4754 4755 /* Set BLOCK_NUMBER for all the blocks in FN. */ 4756 4757 void 4758 number_blocks (tree fn) 4759 { 4760 int i; 4761 int n_blocks; 4762 tree *block_vector; 4763 4764 /* For XCOFF debugging output, we start numbering the blocks 4765 from 1 within each function, rather than keeping a running 4766 count. */ 4767 #if defined (XCOFF_DEBUGGING_INFO) 4768 if (write_symbols == XCOFF_DEBUG) 4769 next_block_index = 1; 4770 #endif 4771 4772 block_vector = get_block_vector (DECL_INITIAL (fn), &n_blocks); 4773 4774 /* The top-level BLOCK isn't numbered at all. */ 4775 for (i = 1; i < n_blocks; ++i) 4776 /* We number the blocks from two. */ 4777 BLOCK_NUMBER (block_vector[i]) = next_block_index++; 4778 4779 free (block_vector); 4780 4781 return; 4782 } 4783 4784 /* If VAR is present in a subblock of BLOCK, return the subblock. */ 4785 4786 DEBUG_FUNCTION tree 4787 debug_find_var_in_block_tree (tree var, tree block) 4788 { 4789 tree t; 4790 4791 for (t = BLOCK_VARS (block); t; t = TREE_CHAIN (t)) 4792 if (t == var) 4793 return block; 4794 4795 for (t = BLOCK_SUBBLOCKS (block); t; t = TREE_CHAIN (t)) 4796 { 4797 tree ret = debug_find_var_in_block_tree (var, t); 4798 if (ret) 4799 return ret; 4800 } 4801 4802 return NULL_TREE; 4803 } 4804 4805 /* Keep track of whether we're in a dummy function context. If we are, 4806 we don't want to invoke the set_current_function hook, because we'll 4807 get into trouble if the hook calls target_reinit () recursively or 4808 when the initial initialization is not yet complete. */ 4809 4810 static bool in_dummy_function; 4811 4812 /* Invoke the target hook when setting cfun. Update the optimization options 4813 if the function uses different options than the default. */ 4814 4815 static void 4816 invoke_set_current_function_hook (tree fndecl) 4817 { 4818 if (!in_dummy_function) 4819 { 4820 tree opts = ((fndecl) 4821 ? DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl) 4822 : optimization_default_node); 4823 4824 if (!opts) 4825 opts = optimization_default_node; 4826 4827 /* Change optimization options if needed. */ 4828 if (optimization_current_node != opts) 4829 { 4830 optimization_current_node = opts; 4831 cl_optimization_restore (&global_options, TREE_OPTIMIZATION (opts)); 4832 } 4833 4834 targetm.set_current_function (fndecl); 4835 this_fn_optabs = this_target_optabs; 4836 4837 if (opts != optimization_default_node) 4838 { 4839 init_tree_optimization_optabs (opts); 4840 if (TREE_OPTIMIZATION_OPTABS (opts)) 4841 this_fn_optabs = (struct target_optabs *) 4842 TREE_OPTIMIZATION_OPTABS (opts); 4843 } 4844 } 4845 } 4846 4847 /* cfun should never be set directly; use this function. */ 4848 4849 void 4850 set_cfun (struct function *new_cfun, bool force) 4851 { 4852 if (cfun != new_cfun || force) 4853 { 4854 cfun = new_cfun; 4855 invoke_set_current_function_hook (new_cfun ? new_cfun->decl : NULL_TREE); 4856 redirect_edge_var_map_empty (); 4857 } 4858 } 4859 4860 /* Initialized with NOGC, making this poisonous to the garbage collector. */ 4861 4862 static vec<function *> cfun_stack; 4863 4864 /* Push the current cfun onto the stack, and set cfun to new_cfun. Also set 4865 current_function_decl accordingly. */ 4866 4867 void 4868 push_cfun (struct function *new_cfun) 4869 { 4870 gcc_assert ((!cfun && !current_function_decl) 4871 || (cfun && current_function_decl == cfun->decl)); 4872 cfun_stack.safe_push (cfun); 4873 current_function_decl = new_cfun ? new_cfun->decl : NULL_TREE; 4874 set_cfun (new_cfun); 4875 } 4876 4877 /* Pop cfun from the stack. Also set current_function_decl accordingly. */ 4878 4879 void 4880 pop_cfun (void) 4881 { 4882 struct function *new_cfun = cfun_stack.pop (); 4883 /* When in_dummy_function, we do have a cfun but current_function_decl is 4884 NULL. We also allow pushing NULL cfun and subsequently changing 4885 current_function_decl to something else and have both restored by 4886 pop_cfun. */ 4887 gcc_checking_assert (in_dummy_function 4888 || !cfun 4889 || current_function_decl == cfun->decl); 4890 set_cfun (new_cfun); 4891 current_function_decl = new_cfun ? new_cfun->decl : NULL_TREE; 4892 } 4893 4894 /* Return value of funcdef and increase it. */ 4895 int 4896 get_next_funcdef_no (void) 4897 { 4898 return funcdef_no++; 4899 } 4900 4901 /* Return value of funcdef. */ 4902 int 4903 get_last_funcdef_no (void) 4904 { 4905 return funcdef_no; 4906 } 4907 4908 /* Allocate a function structure for FNDECL and set its contents 4909 to the defaults. Set cfun to the newly-allocated object. 4910 Some of the helper functions invoked during initialization assume 4911 that cfun has already been set. Therefore, assign the new object 4912 directly into cfun and invoke the back end hook explicitly at the 4913 very end, rather than initializing a temporary and calling set_cfun 4914 on it. 4915 4916 ABSTRACT_P is true if this is a function that will never be seen by 4917 the middle-end. Such functions are front-end concepts (like C++ 4918 function templates) that do not correspond directly to functions 4919 placed in object files. */ 4920 4921 void 4922 allocate_struct_function (tree fndecl, bool abstract_p) 4923 { 4924 tree fntype = fndecl ? TREE_TYPE (fndecl) : NULL_TREE; 4925 4926 cfun = ggc_cleared_alloc<function> (); 4927 4928 init_eh_for_function (); 4929 4930 if (init_machine_status) 4931 cfun->machine = (*init_machine_status) (); 4932 4933 #ifdef OVERRIDE_ABI_FORMAT 4934 OVERRIDE_ABI_FORMAT (fndecl); 4935 #endif 4936 4937 if (fndecl != NULL_TREE) 4938 { 4939 DECL_STRUCT_FUNCTION (fndecl) = cfun; 4940 cfun->decl = fndecl; 4941 current_function_funcdef_no = get_next_funcdef_no (); 4942 } 4943 4944 invoke_set_current_function_hook (fndecl); 4945 4946 if (fndecl != NULL_TREE) 4947 { 4948 tree result = DECL_RESULT (fndecl); 4949 4950 if (!abstract_p) 4951 { 4952 /* Now that we have activated any function-specific attributes 4953 that might affect layout, particularly vector modes, relayout 4954 each of the parameters and the result. */ 4955 relayout_decl (result); 4956 for (tree parm = DECL_ARGUMENTS (fndecl); parm; 4957 parm = DECL_CHAIN (parm)) 4958 relayout_decl (parm); 4959 4960 /* Similarly relayout the function decl. */ 4961 targetm.target_option.relayout_function (fndecl); 4962 } 4963 4964 if (!abstract_p && aggregate_value_p (result, fndecl)) 4965 { 4966 #ifdef PCC_STATIC_STRUCT_RETURN 4967 cfun->returns_pcc_struct = 1; 4968 #endif 4969 cfun->returns_struct = 1; 4970 } 4971 4972 cfun->stdarg = stdarg_p (fntype); 4973 4974 /* Assume all registers in stdarg functions need to be saved. */ 4975 cfun->va_list_gpr_size = VA_LIST_MAX_GPR_SIZE; 4976 cfun->va_list_fpr_size = VA_LIST_MAX_FPR_SIZE; 4977 4978 /* ??? This could be set on a per-function basis by the front-end 4979 but is this worth the hassle? */ 4980 cfun->can_throw_non_call_exceptions = flag_non_call_exceptions; 4981 cfun->can_delete_dead_exceptions = flag_delete_dead_exceptions; 4982 4983 if (!profile_flag && !flag_instrument_function_entry_exit) 4984 DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (fndecl) = 1; 4985 } 4986 4987 /* Don't enable begin stmt markers if var-tracking at assignments is 4988 disabled. The markers make little sense without the variable 4989 binding annotations among them. */ 4990 cfun->debug_nonbind_markers = lang_hooks.emits_begin_stmt 4991 && MAY_HAVE_DEBUG_MARKER_STMTS; 4992 } 4993 4994 /* This is like allocate_struct_function, but pushes a new cfun for FNDECL 4995 instead of just setting it. */ 4996 4997 void 4998 push_struct_function (tree fndecl) 4999 { 5000 /* When in_dummy_function we might be in the middle of a pop_cfun and 5001 current_function_decl and cfun may not match. */ 5002 gcc_assert (in_dummy_function 5003 || (!cfun && !current_function_decl) 5004 || (cfun && current_function_decl == cfun->decl)); 5005 cfun_stack.safe_push (cfun); 5006 current_function_decl = fndecl; 5007 allocate_struct_function (fndecl, false); 5008 } 5009 5010 /* Reset crtl and other non-struct-function variables to defaults as 5011 appropriate for emitting rtl at the start of a function. */ 5012 5013 static void 5014 prepare_function_start (void) 5015 { 5016 gcc_assert (!get_last_insn ()); 5017 init_temp_slots (); 5018 init_emit (); 5019 init_varasm_status (); 5020 init_expr (); 5021 default_rtl_profile (); 5022 5023 if (flag_stack_usage_info) 5024 { 5025 cfun->su = ggc_cleared_alloc<stack_usage> (); 5026 cfun->su->static_stack_size = -1; 5027 } 5028 5029 cse_not_expected = ! optimize; 5030 5031 /* Caller save not needed yet. */ 5032 caller_save_needed = 0; 5033 5034 /* We haven't done register allocation yet. */ 5035 reg_renumber = 0; 5036 5037 /* Indicate that we have not instantiated virtual registers yet. */ 5038 virtuals_instantiated = 0; 5039 5040 /* Indicate that we want CONCATs now. */ 5041 generating_concat_p = 1; 5042 5043 /* Indicate we have no need of a frame pointer yet. */ 5044 frame_pointer_needed = 0; 5045 } 5046 5047 void 5048 push_dummy_function (bool with_decl) 5049 { 5050 tree fn_decl, fn_type, fn_result_decl; 5051 5052 gcc_assert (!in_dummy_function); 5053 in_dummy_function = true; 5054 5055 if (with_decl) 5056 { 5057 fn_type = build_function_type_list (void_type_node, NULL_TREE); 5058 fn_decl = build_decl (UNKNOWN_LOCATION, FUNCTION_DECL, NULL_TREE, 5059 fn_type); 5060 fn_result_decl = build_decl (UNKNOWN_LOCATION, RESULT_DECL, 5061 NULL_TREE, void_type_node); 5062 DECL_RESULT (fn_decl) = fn_result_decl; 5063 } 5064 else 5065 fn_decl = NULL_TREE; 5066 5067 push_struct_function (fn_decl); 5068 } 5069 5070 /* Initialize the rtl expansion mechanism so that we can do simple things 5071 like generate sequences. This is used to provide a context during global 5072 initialization of some passes. You must call expand_dummy_function_end 5073 to exit this context. */ 5074 5075 void 5076 init_dummy_function_start (void) 5077 { 5078 push_dummy_function (false); 5079 prepare_function_start (); 5080 } 5081 5082 /* Generate RTL for the start of the function SUBR (a FUNCTION_DECL tree node) 5083 and initialize static variables for generating RTL for the statements 5084 of the function. */ 5085 5086 void 5087 init_function_start (tree subr) 5088 { 5089 /* Initialize backend, if needed. */ 5090 initialize_rtl (); 5091 5092 prepare_function_start (); 5093 decide_function_section (subr); 5094 5095 /* Warn if this value is an aggregate type, 5096 regardless of which calling convention we are using for it. */ 5097 if (AGGREGATE_TYPE_P (TREE_TYPE (DECL_RESULT (subr)))) 5098 warning (OPT_Waggregate_return, "function returns an aggregate"); 5099 } 5100 5101 /* Expand code to verify the stack_protect_guard. This is invoked at 5102 the end of a function to be protected. */ 5103 5104 void 5105 stack_protect_epilogue (void) 5106 { 5107 tree guard_decl = targetm.stack_protect_guard (); 5108 rtx_code_label *label = gen_label_rtx (); 5109 rtx x, y; 5110 rtx_insn *seq; 5111 5112 x = expand_normal (crtl->stack_protect_guard); 5113 if (guard_decl) 5114 y = expand_normal (guard_decl); 5115 else 5116 y = const0_rtx; 5117 5118 /* Allow the target to compare Y with X without leaking either into 5119 a register. */ 5120 if (targetm.have_stack_protect_test () 5121 && ((seq = targetm.gen_stack_protect_test (x, y, label)) != NULL_RTX)) 5122 emit_insn (seq); 5123 else 5124 emit_cmp_and_jump_insns (x, y, EQ, NULL_RTX, ptr_mode, 1, label); 5125 5126 /* The noreturn predictor has been moved to the tree level. The rtl-level 5127 predictors estimate this branch about 20%, which isn't enough to get 5128 things moved out of line. Since this is the only extant case of adding 5129 a noreturn function at the rtl level, it doesn't seem worth doing ought 5130 except adding the prediction by hand. */ 5131 rtx_insn *tmp = get_last_insn (); 5132 if (JUMP_P (tmp)) 5133 predict_insn_def (tmp, PRED_NORETURN, TAKEN); 5134 5135 expand_call (targetm.stack_protect_fail (), NULL_RTX, /*ignore=*/true); 5136 free_temp_slots (); 5137 emit_label (label); 5138 } 5139 5140 /* Start the RTL for a new function, and set variables used for 5141 emitting RTL. 5142 SUBR is the FUNCTION_DECL node. 5143 PARMS_HAVE_CLEANUPS is nonzero if there are cleanups associated with 5144 the function's parameters, which must be run at any return statement. */ 5145 5146 void 5147 expand_function_start (tree subr) 5148 { 5149 /* Make sure volatile mem refs aren't considered 5150 valid operands of arithmetic insns. */ 5151 init_recog_no_volatile (); 5152 5153 crtl->profile 5154 = (profile_flag 5155 && ! DECL_NO_INSTRUMENT_FUNCTION_ENTRY_EXIT (subr)); 5156 5157 crtl->limit_stack 5158 = (stack_limit_rtx != NULL_RTX && ! DECL_NO_LIMIT_STACK (subr)); 5159 5160 /* Make the label for return statements to jump to. Do not special 5161 case machines with special return instructions -- they will be 5162 handled later during jump, ifcvt, or epilogue creation. */ 5163 return_label = gen_label_rtx (); 5164 5165 /* Initialize rtx used to return the value. */ 5166 /* Do this before assign_parms so that we copy the struct value address 5167 before any library calls that assign parms might generate. */ 5168 5169 /* Decide whether to return the value in memory or in a register. */ 5170 tree res = DECL_RESULT (subr); 5171 if (aggregate_value_p (res, subr)) 5172 { 5173 /* Returning something that won't go in a register. */ 5174 rtx value_address = 0; 5175 5176 #ifdef PCC_STATIC_STRUCT_RETURN 5177 if (cfun->returns_pcc_struct) 5178 { 5179 int size = int_size_in_bytes (TREE_TYPE (res)); 5180 value_address = assemble_static_space (size); 5181 } 5182 else 5183 #endif 5184 { 5185 rtx sv = targetm.calls.struct_value_rtx (TREE_TYPE (subr), 2); 5186 /* Expect to be passed the address of a place to store the value. 5187 If it is passed as an argument, assign_parms will take care of 5188 it. */ 5189 if (sv) 5190 { 5191 value_address = gen_reg_rtx (Pmode); 5192 emit_move_insn (value_address, sv); 5193 } 5194 } 5195 if (value_address) 5196 { 5197 rtx x = value_address; 5198 if (!DECL_BY_REFERENCE (res)) 5199 { 5200 x = gen_rtx_MEM (DECL_MODE (res), x); 5201 set_mem_attributes (x, res, 1); 5202 } 5203 set_parm_rtl (res, x); 5204 } 5205 } 5206 else if (DECL_MODE (res) == VOIDmode) 5207 /* If return mode is void, this decl rtl should not be used. */ 5208 set_parm_rtl (res, NULL_RTX); 5209 else 5210 { 5211 /* Compute the return values into a pseudo reg, which we will copy 5212 into the true return register after the cleanups are done. */ 5213 tree return_type = TREE_TYPE (res); 5214 5215 /* If we may coalesce this result, make sure it has the expected mode 5216 in case it was promoted. But we need not bother about BLKmode. */ 5217 machine_mode promoted_mode 5218 = flag_tree_coalesce_vars && is_gimple_reg (res) 5219 ? promote_ssa_mode (ssa_default_def (cfun, res), NULL) 5220 : BLKmode; 5221 5222 if (promoted_mode != BLKmode) 5223 set_parm_rtl (res, gen_reg_rtx (promoted_mode)); 5224 else if (TYPE_MODE (return_type) != BLKmode 5225 && targetm.calls.return_in_msb (return_type)) 5226 /* expand_function_end will insert the appropriate padding in 5227 this case. Use the return value's natural (unpadded) mode 5228 within the function proper. */ 5229 set_parm_rtl (res, gen_reg_rtx (TYPE_MODE (return_type))); 5230 else 5231 { 5232 /* In order to figure out what mode to use for the pseudo, we 5233 figure out what the mode of the eventual return register will 5234 actually be, and use that. */ 5235 rtx hard_reg = hard_function_value (return_type, subr, 0, 1); 5236 5237 /* Structures that are returned in registers are not 5238 aggregate_value_p, so we may see a PARALLEL or a REG. */ 5239 if (REG_P (hard_reg)) 5240 set_parm_rtl (res, gen_reg_rtx (GET_MODE (hard_reg))); 5241 else 5242 { 5243 gcc_assert (GET_CODE (hard_reg) == PARALLEL); 5244 set_parm_rtl (res, gen_group_rtx (hard_reg)); 5245 } 5246 } 5247 5248 /* Set DECL_REGISTER flag so that expand_function_end will copy the 5249 result to the real return register(s). */ 5250 DECL_REGISTER (res) = 1; 5251 5252 if (chkp_function_instrumented_p (current_function_decl)) 5253 { 5254 tree return_type = TREE_TYPE (res); 5255 rtx bounds = targetm.calls.chkp_function_value_bounds (return_type, 5256 subr, 1); 5257 SET_DECL_BOUNDS_RTL (res, bounds); 5258 } 5259 } 5260 5261 /* Initialize rtx for parameters and local variables. 5262 In some cases this requires emitting insns. */ 5263 assign_parms (subr); 5264 5265 /* If function gets a static chain arg, store it. */ 5266 if (cfun->static_chain_decl) 5267 { 5268 tree parm = cfun->static_chain_decl; 5269 rtx local, chain; 5270 rtx_insn *insn; 5271 int unsignedp; 5272 5273 local = gen_reg_rtx (promote_decl_mode (parm, &unsignedp)); 5274 chain = targetm.calls.static_chain (current_function_decl, true); 5275 5276 set_decl_incoming_rtl (parm, chain, false); 5277 set_parm_rtl (parm, local); 5278 mark_reg_pointer (local, TYPE_ALIGN (TREE_TYPE (TREE_TYPE (parm)))); 5279 5280 if (GET_MODE (local) != GET_MODE (chain)) 5281 { 5282 convert_move (local, chain, unsignedp); 5283 insn = get_last_insn (); 5284 } 5285 else 5286 insn = emit_move_insn (local, chain); 5287 5288 /* Mark the register as eliminable, similar to parameters. */ 5289 if (MEM_P (chain) 5290 && reg_mentioned_p (arg_pointer_rtx, XEXP (chain, 0))) 5291 set_dst_reg_note (insn, REG_EQUIV, chain, local); 5292 5293 /* If we aren't optimizing, save the static chain onto the stack. */ 5294 if (!optimize) 5295 { 5296 tree saved_static_chain_decl 5297 = build_decl (DECL_SOURCE_LOCATION (parm), VAR_DECL, 5298 DECL_NAME (parm), TREE_TYPE (parm)); 5299 rtx saved_static_chain_rtx 5300 = assign_stack_local (Pmode, GET_MODE_SIZE (Pmode), 0); 5301 SET_DECL_RTL (saved_static_chain_decl, saved_static_chain_rtx); 5302 emit_move_insn (saved_static_chain_rtx, chain); 5303 SET_DECL_VALUE_EXPR (parm, saved_static_chain_decl); 5304 DECL_HAS_VALUE_EXPR_P (parm) = 1; 5305 } 5306 } 5307 5308 /* The following was moved from init_function_start. 5309 The move was supposed to make sdb output more accurate. */ 5310 /* Indicate the beginning of the function body, 5311 as opposed to parm setup. */ 5312 emit_note (NOTE_INSN_FUNCTION_BEG); 5313 5314 gcc_assert (NOTE_P (get_last_insn ())); 5315 5316 parm_birth_insn = get_last_insn (); 5317 5318 /* If the function receives a non-local goto, then store the 5319 bits we need to restore the frame pointer. */ 5320 if (cfun->nonlocal_goto_save_area) 5321 { 5322 tree t_save; 5323 rtx r_save; 5324 5325 tree var = TREE_OPERAND (cfun->nonlocal_goto_save_area, 0); 5326 gcc_assert (DECL_RTL_SET_P (var)); 5327 5328 t_save = build4 (ARRAY_REF, 5329 TREE_TYPE (TREE_TYPE (cfun->nonlocal_goto_save_area)), 5330 cfun->nonlocal_goto_save_area, 5331 integer_zero_node, NULL_TREE, NULL_TREE); 5332 r_save = expand_expr (t_save, NULL_RTX, VOIDmode, EXPAND_WRITE); 5333 gcc_assert (GET_MODE (r_save) == Pmode); 5334 5335 emit_move_insn (r_save, targetm.builtin_setjmp_frame_value ()); 5336 update_nonlocal_goto_save_area (); 5337 } 5338 5339 if (crtl->profile) 5340 { 5341 #ifdef PROFILE_HOOK 5342 PROFILE_HOOK (current_function_funcdef_no); 5343 #endif 5344 } 5345 5346 /* If we are doing generic stack checking, the probe should go here. */ 5347 if (flag_stack_check == GENERIC_STACK_CHECK) 5348 stack_check_probe_note = emit_note (NOTE_INSN_DELETED); 5349 } 5350 5351 void 5352 pop_dummy_function (void) 5353 { 5354 pop_cfun (); 5355 in_dummy_function = false; 5356 } 5357 5358 /* Undo the effects of init_dummy_function_start. */ 5359 void 5360 expand_dummy_function_end (void) 5361 { 5362 gcc_assert (in_dummy_function); 5363 5364 /* End any sequences that failed to be closed due to syntax errors. */ 5365 while (in_sequence_p ()) 5366 end_sequence (); 5367 5368 /* Outside function body, can't compute type's actual size 5369 until next function's body starts. */ 5370 5371 free_after_parsing (cfun); 5372 free_after_compilation (cfun); 5373 pop_dummy_function (); 5374 } 5375 5376 /* Helper for diddle_return_value. */ 5377 5378 void 5379 diddle_return_value_1 (void (*doit) (rtx, void *), void *arg, rtx outgoing) 5380 { 5381 if (! outgoing) 5382 return; 5383 5384 if (REG_P (outgoing)) 5385 (*doit) (outgoing, arg); 5386 else if (GET_CODE (outgoing) == PARALLEL) 5387 { 5388 int i; 5389 5390 for (i = 0; i < XVECLEN (outgoing, 0); i++) 5391 { 5392 rtx x = XEXP (XVECEXP (outgoing, 0, i), 0); 5393 5394 if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER) 5395 (*doit) (x, arg); 5396 } 5397 } 5398 } 5399 5400 /* Call DOIT for each hard register used as a return value from 5401 the current function. */ 5402 5403 void 5404 diddle_return_value (void (*doit) (rtx, void *), void *arg) 5405 { 5406 diddle_return_value_1 (doit, arg, crtl->return_bnd); 5407 diddle_return_value_1 (doit, arg, crtl->return_rtx); 5408 } 5409 5410 static void 5411 do_clobber_return_reg (rtx reg, void *arg ATTRIBUTE_UNUSED) 5412 { 5413 emit_clobber (reg); 5414 } 5415 5416 void 5417 clobber_return_register (void) 5418 { 5419 diddle_return_value (do_clobber_return_reg, NULL); 5420 5421 /* In case we do use pseudo to return value, clobber it too. */ 5422 if (DECL_RTL_SET_P (DECL_RESULT (current_function_decl))) 5423 { 5424 tree decl_result = DECL_RESULT (current_function_decl); 5425 rtx decl_rtl = DECL_RTL (decl_result); 5426 if (REG_P (decl_rtl) && REGNO (decl_rtl) >= FIRST_PSEUDO_REGISTER) 5427 { 5428 do_clobber_return_reg (decl_rtl, NULL); 5429 } 5430 } 5431 } 5432 5433 static void 5434 do_use_return_reg (rtx reg, void *arg ATTRIBUTE_UNUSED) 5435 { 5436 emit_use (reg); 5437 } 5438 5439 static void 5440 use_return_register (void) 5441 { 5442 diddle_return_value (do_use_return_reg, NULL); 5443 } 5444 5445 /* Set the location of the insn chain starting at INSN to LOC. */ 5446 5447 static void 5448 set_insn_locations (rtx_insn *insn, int loc) 5449 { 5450 while (insn != NULL) 5451 { 5452 if (INSN_P (insn)) 5453 INSN_LOCATION (insn) = loc; 5454 insn = NEXT_INSN (insn); 5455 } 5456 } 5457 5458 /* Generate RTL for the end of the current function. */ 5459 5460 void 5461 expand_function_end (void) 5462 { 5463 /* If arg_pointer_save_area was referenced only from a nested 5464 function, we will not have initialized it yet. Do that now. */ 5465 if (arg_pointer_save_area && ! crtl->arg_pointer_save_area_init) 5466 get_arg_pointer_save_area (); 5467 5468 /* If we are doing generic stack checking and this function makes calls, 5469 do a stack probe at the start of the function to ensure we have enough 5470 space for another stack frame. */ 5471 if (flag_stack_check == GENERIC_STACK_CHECK) 5472 { 5473 rtx_insn *insn, *seq; 5474 5475 for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) 5476 if (CALL_P (insn)) 5477 { 5478 rtx max_frame_size = GEN_INT (STACK_CHECK_MAX_FRAME_SIZE); 5479 start_sequence (); 5480 if (STACK_CHECK_MOVING_SP) 5481 anti_adjust_stack_and_probe (max_frame_size, true); 5482 else 5483 probe_stack_range (STACK_OLD_CHECK_PROTECT, max_frame_size); 5484 seq = get_insns (); 5485 end_sequence (); 5486 set_insn_locations (seq, prologue_location); 5487 emit_insn_before (seq, stack_check_probe_note); 5488 break; 5489 } 5490 } 5491 5492 /* End any sequences that failed to be closed due to syntax errors. */ 5493 while (in_sequence_p ()) 5494 end_sequence (); 5495 5496 clear_pending_stack_adjust (); 5497 do_pending_stack_adjust (); 5498 5499 /* Output a linenumber for the end of the function. 5500 SDB depended on this. */ 5501 set_curr_insn_location (input_location); 5502 5503 /* Before the return label (if any), clobber the return 5504 registers so that they are not propagated live to the rest of 5505 the function. This can only happen with functions that drop 5506 through; if there had been a return statement, there would 5507 have either been a return rtx, or a jump to the return label. 5508 5509 We delay actual code generation after the current_function_value_rtx 5510 is computed. */ 5511 rtx_insn *clobber_after = get_last_insn (); 5512 5513 /* Output the label for the actual return from the function. */ 5514 emit_label (return_label); 5515 5516 if (targetm_common.except_unwind_info (&global_options) == UI_SJLJ) 5517 { 5518 /* Let except.c know where it should emit the call to unregister 5519 the function context for sjlj exceptions. */ 5520 if (flag_exceptions) 5521 sjlj_emit_function_exit_after (get_last_insn ()); 5522 } 5523 else 5524 { 5525 /* We want to ensure that instructions that may trap are not 5526 moved into the epilogue by scheduling, because we don't 5527 always emit unwind information for the epilogue. */ 5528 if (cfun->can_throw_non_call_exceptions) 5529 emit_insn (gen_blockage ()); 5530 } 5531 5532 /* If this is an implementation of throw, do what's necessary to 5533 communicate between __builtin_eh_return and the epilogue. */ 5534 expand_eh_return (); 5535 5536 /* If scalar return value was computed in a pseudo-reg, or was a named 5537 return value that got dumped to the stack, copy that to the hard 5538 return register. */ 5539 if (DECL_RTL_SET_P (DECL_RESULT (current_function_decl))) 5540 { 5541 tree decl_result = DECL_RESULT (current_function_decl); 5542 rtx decl_rtl = DECL_RTL (decl_result); 5543 5544 if (REG_P (decl_rtl) 5545 ? REGNO (decl_rtl) >= FIRST_PSEUDO_REGISTER 5546 : DECL_REGISTER (decl_result)) 5547 { 5548 rtx real_decl_rtl = crtl->return_rtx; 5549 complex_mode cmode; 5550 5551 /* This should be set in assign_parms. */ 5552 gcc_assert (REG_FUNCTION_VALUE_P (real_decl_rtl)); 5553 5554 /* If this is a BLKmode structure being returned in registers, 5555 then use the mode computed in expand_return. Note that if 5556 decl_rtl is memory, then its mode may have been changed, 5557 but that crtl->return_rtx has not. */ 5558 if (GET_MODE (real_decl_rtl) == BLKmode) 5559 PUT_MODE (real_decl_rtl, GET_MODE (decl_rtl)); 5560 5561 /* If a non-BLKmode return value should be padded at the least 5562 significant end of the register, shift it left by the appropriate 5563 amount. BLKmode results are handled using the group load/store 5564 machinery. */ 5565 if (TYPE_MODE (TREE_TYPE (decl_result)) != BLKmode 5566 && REG_P (real_decl_rtl) 5567 && targetm.calls.return_in_msb (TREE_TYPE (decl_result))) 5568 { 5569 emit_move_insn (gen_rtx_REG (GET_MODE (decl_rtl), 5570 REGNO (real_decl_rtl)), 5571 decl_rtl); 5572 shift_return_value (GET_MODE (decl_rtl), true, real_decl_rtl); 5573 } 5574 else if (GET_CODE (real_decl_rtl) == PARALLEL) 5575 { 5576 /* If expand_function_start has created a PARALLEL for decl_rtl, 5577 move the result to the real return registers. Otherwise, do 5578 a group load from decl_rtl for a named return. */ 5579 if (GET_CODE (decl_rtl) == PARALLEL) 5580 emit_group_move (real_decl_rtl, decl_rtl); 5581 else 5582 emit_group_load (real_decl_rtl, decl_rtl, 5583 TREE_TYPE (decl_result), 5584 int_size_in_bytes (TREE_TYPE (decl_result))); 5585 } 5586 /* In the case of complex integer modes smaller than a word, we'll 5587 need to generate some non-trivial bitfield insertions. Do that 5588 on a pseudo and not the hard register. */ 5589 else if (GET_CODE (decl_rtl) == CONCAT 5590 && is_complex_int_mode (GET_MODE (decl_rtl), &cmode) 5591 && GET_MODE_BITSIZE (cmode) <= BITS_PER_WORD) 5592 { 5593 int old_generating_concat_p; 5594 rtx tmp; 5595 5596 old_generating_concat_p = generating_concat_p; 5597 generating_concat_p = 0; 5598 tmp = gen_reg_rtx (GET_MODE (decl_rtl)); 5599 generating_concat_p = old_generating_concat_p; 5600 5601 emit_move_insn (tmp, decl_rtl); 5602 emit_move_insn (real_decl_rtl, tmp); 5603 } 5604 /* If a named return value dumped decl_return to memory, then 5605 we may need to re-do the PROMOTE_MODE signed/unsigned 5606 extension. */ 5607 else if (GET_MODE (real_decl_rtl) != GET_MODE (decl_rtl)) 5608 { 5609 int unsignedp = TYPE_UNSIGNED (TREE_TYPE (decl_result)); 5610 promote_function_mode (TREE_TYPE (decl_result), 5611 GET_MODE (decl_rtl), &unsignedp, 5612 TREE_TYPE (current_function_decl), 1); 5613 5614 convert_move (real_decl_rtl, decl_rtl, unsignedp); 5615 } 5616 else 5617 emit_move_insn (real_decl_rtl, decl_rtl); 5618 } 5619 } 5620 5621 /* If returning a structure, arrange to return the address of the value 5622 in a place where debuggers expect to find it. 5623 5624 If returning a structure PCC style, 5625 the caller also depends on this value. 5626 And cfun->returns_pcc_struct is not necessarily set. */ 5627 if ((cfun->returns_struct || cfun->returns_pcc_struct) 5628 && !targetm.calls.omit_struct_return_reg) 5629 { 5630 rtx value_address = DECL_RTL (DECL_RESULT (current_function_decl)); 5631 tree type = TREE_TYPE (DECL_RESULT (current_function_decl)); 5632 rtx outgoing; 5633 5634 if (DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))) 5635 type = TREE_TYPE (type); 5636 else 5637 value_address = XEXP (value_address, 0); 5638 5639 outgoing = targetm.calls.function_value (build_pointer_type (type), 5640 current_function_decl, true); 5641 5642 /* Mark this as a function return value so integrate will delete the 5643 assignment and USE below when inlining this function. */ 5644 REG_FUNCTION_VALUE_P (outgoing) = 1; 5645 5646 /* The address may be ptr_mode and OUTGOING may be Pmode. */ 5647 scalar_int_mode mode = as_a <scalar_int_mode> (GET_MODE (outgoing)); 5648 value_address = convert_memory_address (mode, value_address); 5649 5650 emit_move_insn (outgoing, value_address); 5651 5652 /* Show return register used to hold result (in this case the address 5653 of the result. */ 5654 crtl->return_rtx = outgoing; 5655 } 5656 5657 /* Emit the actual code to clobber return register. Don't emit 5658 it if clobber_after is a barrier, then the previous basic block 5659 certainly doesn't fall thru into the exit block. */ 5660 if (!BARRIER_P (clobber_after)) 5661 { 5662 start_sequence (); 5663 clobber_return_register (); 5664 rtx_insn *seq = get_insns (); 5665 end_sequence (); 5666 5667 emit_insn_after (seq, clobber_after); 5668 } 5669 5670 /* Output the label for the naked return from the function. */ 5671 if (naked_return_label) 5672 emit_label (naked_return_label); 5673 5674 /* @@@ This is a kludge. We want to ensure that instructions that 5675 may trap are not moved into the epilogue by scheduling, because 5676 we don't always emit unwind information for the epilogue. */ 5677 if (cfun->can_throw_non_call_exceptions 5678 && targetm_common.except_unwind_info (&global_options) != UI_SJLJ) 5679 emit_insn (gen_blockage ()); 5680 5681 /* If stack protection is enabled for this function, check the guard. */ 5682 if (crtl->stack_protect_guard && targetm.stack_protect_runtime_enabled_p ()) 5683 stack_protect_epilogue (); 5684 5685 /* If we had calls to alloca, and this machine needs 5686 an accurate stack pointer to exit the function, 5687 insert some code to save and restore the stack pointer. */ 5688 if (! EXIT_IGNORE_STACK 5689 && cfun->calls_alloca) 5690 { 5691 rtx tem = 0; 5692 5693 start_sequence (); 5694 emit_stack_save (SAVE_FUNCTION, &tem); 5695 rtx_insn *seq = get_insns (); 5696 end_sequence (); 5697 emit_insn_before (seq, parm_birth_insn); 5698 5699 emit_stack_restore (SAVE_FUNCTION, tem); 5700 } 5701 5702 /* ??? This should no longer be necessary since stupid is no longer with 5703 us, but there are some parts of the compiler (eg reload_combine, and 5704 sh mach_dep_reorg) that still try and compute their own lifetime info 5705 instead of using the general framework. */ 5706 use_return_register (); 5707 } 5708 5709 rtx 5710 get_arg_pointer_save_area (void) 5711 { 5712 rtx ret = arg_pointer_save_area; 5713 5714 if (! ret) 5715 { 5716 ret = assign_stack_local (Pmode, GET_MODE_SIZE (Pmode), 0); 5717 arg_pointer_save_area = ret; 5718 } 5719 5720 if (! crtl->arg_pointer_save_area_init) 5721 { 5722 /* Save the arg pointer at the beginning of the function. The 5723 generated stack slot may not be a valid memory address, so we 5724 have to check it and fix it if necessary. */ 5725 start_sequence (); 5726 emit_move_insn (validize_mem (copy_rtx (ret)), 5727 crtl->args.internal_arg_pointer); 5728 rtx_insn *seq = get_insns (); 5729 end_sequence (); 5730 5731 push_topmost_sequence (); 5732 emit_insn_after (seq, entry_of_function ()); 5733 pop_topmost_sequence (); 5734 5735 crtl->arg_pointer_save_area_init = true; 5736 } 5737 5738 return ret; 5739 } 5740 5741 5742 /* If debugging dumps are requested, dump information about how the 5743 target handled -fstack-check=clash for the prologue. 5744 5745 PROBES describes what if any probes were emitted. 5746 5747 RESIDUALS indicates if the prologue had any residual allocation 5748 (i.e. total allocation was not a multiple of PROBE_INTERVAL). */ 5749 5750 void 5751 dump_stack_clash_frame_info (enum stack_clash_probes probes, bool residuals) 5752 { 5753 if (!dump_file) 5754 return; 5755 5756 switch (probes) 5757 { 5758 case NO_PROBE_NO_FRAME: 5759 fprintf (dump_file, 5760 "Stack clash no probe no stack adjustment in prologue.\n"); 5761 break; 5762 case NO_PROBE_SMALL_FRAME: 5763 fprintf (dump_file, 5764 "Stack clash no probe small stack adjustment in prologue.\n"); 5765 break; 5766 case PROBE_INLINE: 5767 fprintf (dump_file, "Stack clash inline probes in prologue.\n"); 5768 break; 5769 case PROBE_LOOP: 5770 fprintf (dump_file, "Stack clash probe loop in prologue.\n"); 5771 break; 5772 } 5773 5774 if (residuals) 5775 fprintf (dump_file, "Stack clash residual allocation in prologue.\n"); 5776 else 5777 fprintf (dump_file, "Stack clash no residual allocation in prologue.\n"); 5778 5779 if (frame_pointer_needed) 5780 fprintf (dump_file, "Stack clash frame pointer needed.\n"); 5781 else 5782 fprintf (dump_file, "Stack clash no frame pointer needed.\n"); 5783 5784 if (TREE_THIS_VOLATILE (cfun->decl)) 5785 fprintf (dump_file, 5786 "Stack clash noreturn prologue, assuming no implicit" 5787 " probes in caller.\n"); 5788 else 5789 fprintf (dump_file, 5790 "Stack clash not noreturn prologue.\n"); 5791 } 5792 5793 /* Add a list of INSNS to the hash HASHP, possibly allocating HASHP 5794 for the first time. */ 5795 5796 static void 5797 record_insns (rtx_insn *insns, rtx end, hash_table<insn_cache_hasher> **hashp) 5798 { 5799 rtx_insn *tmp; 5800 hash_table<insn_cache_hasher> *hash = *hashp; 5801 5802 if (hash == NULL) 5803 *hashp = hash = hash_table<insn_cache_hasher>::create_ggc (17); 5804 5805 for (tmp = insns; tmp != end; tmp = NEXT_INSN (tmp)) 5806 { 5807 rtx *slot = hash->find_slot (tmp, INSERT); 5808 gcc_assert (*slot == NULL); 5809 *slot = tmp; 5810 } 5811 } 5812 5813 /* INSN has been duplicated or replaced by as COPY, perhaps by duplicating a 5814 basic block, splitting or peepholes. If INSN is a prologue or epilogue 5815 insn, then record COPY as well. */ 5816 5817 void 5818 maybe_copy_prologue_epilogue_insn (rtx insn, rtx copy) 5819 { 5820 hash_table<insn_cache_hasher> *hash; 5821 rtx *slot; 5822 5823 hash = epilogue_insn_hash; 5824 if (!hash || !hash->find (insn)) 5825 { 5826 hash = prologue_insn_hash; 5827 if (!hash || !hash->find (insn)) 5828 return; 5829 } 5830 5831 slot = hash->find_slot (copy, INSERT); 5832 gcc_assert (*slot == NULL); 5833 *slot = copy; 5834 } 5835 5836 /* Determine if any INSNs in HASH are, or are part of, INSN. Because 5837 we can be running after reorg, SEQUENCE rtl is possible. */ 5838 5839 static bool 5840 contains (const rtx_insn *insn, hash_table<insn_cache_hasher> *hash) 5841 { 5842 if (hash == NULL) 5843 return false; 5844 5845 if (NONJUMP_INSN_P (insn) && GET_CODE (PATTERN (insn)) == SEQUENCE) 5846 { 5847 rtx_sequence *seq = as_a <rtx_sequence *> (PATTERN (insn)); 5848 int i; 5849 for (i = seq->len () - 1; i >= 0; i--) 5850 if (hash->find (seq->element (i))) 5851 return true; 5852 return false; 5853 } 5854 5855 return hash->find (const_cast<rtx_insn *> (insn)) != NULL; 5856 } 5857 5858 int 5859 prologue_contains (const rtx_insn *insn) 5860 { 5861 return contains (insn, prologue_insn_hash); 5862 } 5863 5864 int 5865 epilogue_contains (const rtx_insn *insn) 5866 { 5867 return contains (insn, epilogue_insn_hash); 5868 } 5869 5870 int 5871 prologue_epilogue_contains (const rtx_insn *insn) 5872 { 5873 if (contains (insn, prologue_insn_hash)) 5874 return 1; 5875 if (contains (insn, epilogue_insn_hash)) 5876 return 1; 5877 return 0; 5878 } 5879 5880 void 5881 record_prologue_seq (rtx_insn *seq) 5882 { 5883 record_insns (seq, NULL, &prologue_insn_hash); 5884 } 5885 5886 void 5887 record_epilogue_seq (rtx_insn *seq) 5888 { 5889 record_insns (seq, NULL, &epilogue_insn_hash); 5890 } 5891 5892 /* Set JUMP_LABEL for a return insn. */ 5893 5894 void 5895 set_return_jump_label (rtx_insn *returnjump) 5896 { 5897 rtx pat = PATTERN (returnjump); 5898 if (GET_CODE (pat) == PARALLEL) 5899 pat = XVECEXP (pat, 0, 0); 5900 if (ANY_RETURN_P (pat)) 5901 JUMP_LABEL (returnjump) = pat; 5902 else 5903 JUMP_LABEL (returnjump) = ret_rtx; 5904 } 5905 5906 /* Return a sequence to be used as the split prologue for the current 5907 function, or NULL. */ 5908 5909 static rtx_insn * 5910 make_split_prologue_seq (void) 5911 { 5912 if (!flag_split_stack 5913 || lookup_attribute ("no_split_stack", DECL_ATTRIBUTES (cfun->decl))) 5914 return NULL; 5915 5916 start_sequence (); 5917 emit_insn (targetm.gen_split_stack_prologue ()); 5918 rtx_insn *seq = get_insns (); 5919 end_sequence (); 5920 5921 record_insns (seq, NULL, &prologue_insn_hash); 5922 set_insn_locations (seq, prologue_location); 5923 5924 return seq; 5925 } 5926 5927 /* Return a sequence to be used as the prologue for the current function, 5928 or NULL. */ 5929 5930 static rtx_insn * 5931 make_prologue_seq (void) 5932 { 5933 if (!targetm.have_prologue ()) 5934 return NULL; 5935 5936 start_sequence (); 5937 rtx_insn *seq = targetm.gen_prologue (); 5938 emit_insn (seq); 5939 5940 /* Insert an explicit USE for the frame pointer 5941 if the profiling is on and the frame pointer is required. */ 5942 if (crtl->profile && frame_pointer_needed) 5943 emit_use (hard_frame_pointer_rtx); 5944 5945 /* Retain a map of the prologue insns. */ 5946 record_insns (seq, NULL, &prologue_insn_hash); 5947 emit_note (NOTE_INSN_PROLOGUE_END); 5948 5949 /* Ensure that instructions are not moved into the prologue when 5950 profiling is on. The call to the profiling routine can be 5951 emitted within the live range of a call-clobbered register. */ 5952 if (!targetm.profile_before_prologue () && crtl->profile) 5953 emit_insn (gen_blockage ()); 5954 5955 seq = get_insns (); 5956 end_sequence (); 5957 set_insn_locations (seq, prologue_location); 5958 5959 return seq; 5960 } 5961 5962 /* Return a sequence to be used as the epilogue for the current function, 5963 or NULL. */ 5964 5965 static rtx_insn * 5966 make_epilogue_seq (void) 5967 { 5968 if (!targetm.have_epilogue ()) 5969 return NULL; 5970 5971 start_sequence (); 5972 emit_note (NOTE_INSN_EPILOGUE_BEG); 5973 rtx_insn *seq = targetm.gen_epilogue (); 5974 if (seq) 5975 emit_jump_insn (seq); 5976 5977 /* Retain a map of the epilogue insns. */ 5978 record_insns (seq, NULL, &epilogue_insn_hash); 5979 set_insn_locations (seq, epilogue_location); 5980 5981 seq = get_insns (); 5982 rtx_insn *returnjump = get_last_insn (); 5983 end_sequence (); 5984 5985 if (JUMP_P (returnjump)) 5986 set_return_jump_label (returnjump); 5987 5988 return seq; 5989 } 5990 5991 5992 /* Generate the prologue and epilogue RTL if the machine supports it. Thread 5993 this into place with notes indicating where the prologue ends and where 5994 the epilogue begins. Update the basic block information when possible. 5995 5996 Notes on epilogue placement: 5997 There are several kinds of edges to the exit block: 5998 * a single fallthru edge from LAST_BB 5999 * possibly, edges from blocks containing sibcalls 6000 * possibly, fake edges from infinite loops 6001 6002 The epilogue is always emitted on the fallthru edge from the last basic 6003 block in the function, LAST_BB, into the exit block. 6004 6005 If LAST_BB is empty except for a label, it is the target of every 6006 other basic block in the function that ends in a return. If a 6007 target has a return or simple_return pattern (possibly with 6008 conditional variants), these basic blocks can be changed so that a 6009 return insn is emitted into them, and their target is adjusted to 6010 the real exit block. 6011 6012 Notes on shrink wrapping: We implement a fairly conservative 6013 version of shrink-wrapping rather than the textbook one. We only 6014 generate a single prologue and a single epilogue. This is 6015 sufficient to catch a number of interesting cases involving early 6016 exits. 6017 6018 First, we identify the blocks that require the prologue to occur before 6019 them. These are the ones that modify a call-saved register, or reference 6020 any of the stack or frame pointer registers. To simplify things, we then 6021 mark everything reachable from these blocks as also requiring a prologue. 6022 This takes care of loops automatically, and avoids the need to examine 6023 whether MEMs reference the frame, since it is sufficient to check for 6024 occurrences of the stack or frame pointer. 6025 6026 We then compute the set of blocks for which the need for a prologue 6027 is anticipatable (borrowing terminology from the shrink-wrapping 6028 description in Muchnick's book). These are the blocks which either 6029 require a prologue themselves, or those that have only successors 6030 where the prologue is anticipatable. The prologue needs to be 6031 inserted on all edges from BB1->BB2 where BB2 is in ANTIC and BB1 6032 is not. For the moment, we ensure that only one such edge exists. 6033 6034 The epilogue is placed as described above, but we make a 6035 distinction between inserting return and simple_return patterns 6036 when modifying other blocks that end in a return. Blocks that end 6037 in a sibcall omit the sibcall_epilogue if the block is not in 6038 ANTIC. */ 6039 6040 void 6041 thread_prologue_and_epilogue_insns (void) 6042 { 6043 df_analyze (); 6044 6045 /* Can't deal with multiple successors of the entry block at the 6046 moment. Function should always have at least one entry 6047 point. */ 6048 gcc_assert (single_succ_p (ENTRY_BLOCK_PTR_FOR_FN (cfun))); 6049 6050 edge entry_edge = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)); 6051 edge orig_entry_edge = entry_edge; 6052 6053 rtx_insn *split_prologue_seq = make_split_prologue_seq (); 6054 rtx_insn *prologue_seq = make_prologue_seq (); 6055 rtx_insn *epilogue_seq = make_epilogue_seq (); 6056 6057 /* Try to perform a kind of shrink-wrapping, making sure the 6058 prologue/epilogue is emitted only around those parts of the 6059 function that require it. */ 6060 try_shrink_wrapping (&entry_edge, prologue_seq); 6061 6062 /* If the target can handle splitting the prologue/epilogue into separate 6063 components, try to shrink-wrap these components separately. */ 6064 try_shrink_wrapping_separate (entry_edge->dest); 6065 6066 /* If that did anything for any component we now need the generate the 6067 "main" prologue again. Because some targets require some of these 6068 to be called in a specific order (i386 requires the split prologue 6069 to be first, for example), we create all three sequences again here. 6070 If this does not work for some target, that target should not enable 6071 separate shrink-wrapping. */ 6072 if (crtl->shrink_wrapped_separate) 6073 { 6074 split_prologue_seq = make_split_prologue_seq (); 6075 prologue_seq = make_prologue_seq (); 6076 epilogue_seq = make_epilogue_seq (); 6077 } 6078 6079 rtl_profile_for_bb (EXIT_BLOCK_PTR_FOR_FN (cfun)); 6080 6081 /* A small fib -- epilogue is not yet completed, but we wish to re-use 6082 this marker for the splits of EH_RETURN patterns, and nothing else 6083 uses the flag in the meantime. */ 6084 epilogue_completed = 1; 6085 6086 /* Find non-fallthru edges that end with EH_RETURN instructions. On 6087 some targets, these get split to a special version of the epilogue 6088 code. In order to be able to properly annotate these with unwind 6089 info, try to split them now. If we get a valid split, drop an 6090 EPILOGUE_BEG note and mark the insns as epilogue insns. */ 6091 edge e; 6092 edge_iterator ei; 6093 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) 6094 { 6095 rtx_insn *prev, *last, *trial; 6096 6097 if (e->flags & EDGE_FALLTHRU) 6098 continue; 6099 last = BB_END (e->src); 6100 if (!eh_returnjump_p (last)) 6101 continue; 6102 6103 prev = PREV_INSN (last); 6104 trial = try_split (PATTERN (last), last, 1); 6105 if (trial == last) 6106 continue; 6107 6108 record_insns (NEXT_INSN (prev), NEXT_INSN (trial), &epilogue_insn_hash); 6109 emit_note_after (NOTE_INSN_EPILOGUE_BEG, prev); 6110 } 6111 6112 edge exit_fallthru_edge = find_fallthru_edge (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds); 6113 6114 if (exit_fallthru_edge) 6115 { 6116 if (epilogue_seq) 6117 { 6118 insert_insn_on_edge (epilogue_seq, exit_fallthru_edge); 6119 commit_edge_insertions (); 6120 6121 /* The epilogue insns we inserted may cause the exit edge to no longer 6122 be fallthru. */ 6123 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) 6124 { 6125 if (((e->flags & EDGE_FALLTHRU) != 0) 6126 && returnjump_p (BB_END (e->src))) 6127 e->flags &= ~EDGE_FALLTHRU; 6128 } 6129 } 6130 else if (next_active_insn (BB_END (exit_fallthru_edge->src))) 6131 { 6132 /* We have a fall-through edge to the exit block, the source is not 6133 at the end of the function, and there will be an assembler epilogue 6134 at the end of the function. 6135 We can't use force_nonfallthru here, because that would try to 6136 use return. Inserting a jump 'by hand' is extremely messy, so 6137 we take advantage of cfg_layout_finalize using 6138 fixup_fallthru_exit_predecessor. */ 6139 cfg_layout_initialize (0); 6140 basic_block cur_bb; 6141 FOR_EACH_BB_FN (cur_bb, cfun) 6142 if (cur_bb->index >= NUM_FIXED_BLOCKS 6143 && cur_bb->next_bb->index >= NUM_FIXED_BLOCKS) 6144 cur_bb->aux = cur_bb->next_bb; 6145 cfg_layout_finalize (); 6146 } 6147 } 6148 6149 /* Insert the prologue. */ 6150 6151 rtl_profile_for_bb (ENTRY_BLOCK_PTR_FOR_FN (cfun)); 6152 6153 if (split_prologue_seq || prologue_seq) 6154 { 6155 rtx_insn *split_prologue_insn = split_prologue_seq; 6156 if (split_prologue_seq) 6157 { 6158 while (split_prologue_insn && !NONDEBUG_INSN_P (split_prologue_insn)) 6159 split_prologue_insn = NEXT_INSN (split_prologue_insn); 6160 insert_insn_on_edge (split_prologue_seq, orig_entry_edge); 6161 } 6162 6163 rtx_insn *prologue_insn = prologue_seq; 6164 if (prologue_seq) 6165 { 6166 while (prologue_insn && !NONDEBUG_INSN_P (prologue_insn)) 6167 prologue_insn = NEXT_INSN (prologue_insn); 6168 insert_insn_on_edge (prologue_seq, entry_edge); 6169 } 6170 6171 commit_edge_insertions (); 6172 6173 /* Look for basic blocks within the prologue insns. */ 6174 if (split_prologue_insn 6175 && BLOCK_FOR_INSN (split_prologue_insn) == NULL) 6176 split_prologue_insn = NULL; 6177 if (prologue_insn 6178 && BLOCK_FOR_INSN (prologue_insn) == NULL) 6179 prologue_insn = NULL; 6180 if (split_prologue_insn || prologue_insn) 6181 { 6182 auto_sbitmap blocks (last_basic_block_for_fn (cfun)); 6183 bitmap_clear (blocks); 6184 if (split_prologue_insn) 6185 bitmap_set_bit (blocks, 6186 BLOCK_FOR_INSN (split_prologue_insn)->index); 6187 if (prologue_insn) 6188 bitmap_set_bit (blocks, BLOCK_FOR_INSN (prologue_insn)->index); 6189 find_many_sub_basic_blocks (blocks); 6190 } 6191 } 6192 6193 default_rtl_profile (); 6194 6195 /* Emit sibling epilogues before any sibling call sites. */ 6196 for (ei = ei_start (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds); 6197 (e = ei_safe_edge (ei)); 6198 ei_next (&ei)) 6199 { 6200 /* Skip those already handled, the ones that run without prologue. */ 6201 if (e->flags & EDGE_IGNORE) 6202 { 6203 e->flags &= ~EDGE_IGNORE; 6204 continue; 6205 } 6206 6207 rtx_insn *insn = BB_END (e->src); 6208 6209 if (!(CALL_P (insn) && SIBLING_CALL_P (insn))) 6210 continue; 6211 6212 if (rtx_insn *ep_seq = targetm.gen_sibcall_epilogue ()) 6213 { 6214 start_sequence (); 6215 emit_note (NOTE_INSN_EPILOGUE_BEG); 6216 emit_insn (ep_seq); 6217 rtx_insn *seq = get_insns (); 6218 end_sequence (); 6219 6220 /* Retain a map of the epilogue insns. Used in life analysis to 6221 avoid getting rid of sibcall epilogue insns. Do this before we 6222 actually emit the sequence. */ 6223 record_insns (seq, NULL, &epilogue_insn_hash); 6224 set_insn_locations (seq, epilogue_location); 6225 6226 emit_insn_before (seq, insn); 6227 } 6228 } 6229 6230 if (epilogue_seq) 6231 { 6232 rtx_insn *insn, *next; 6233 6234 /* Similarly, move any line notes that appear after the epilogue. 6235 There is no need, however, to be quite so anal about the existence 6236 of such a note. Also possibly move 6237 NOTE_INSN_FUNCTION_BEG notes, as those can be relevant for debug 6238 info generation. */ 6239 for (insn = epilogue_seq; insn; insn = next) 6240 { 6241 next = NEXT_INSN (insn); 6242 if (NOTE_P (insn) 6243 && (NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG)) 6244 reorder_insns (insn, insn, PREV_INSN (epilogue_seq)); 6245 } 6246 } 6247 6248 /* Threading the prologue and epilogue changes the artificial refs 6249 in the entry and exit blocks. */ 6250 epilogue_completed = 1; 6251 df_update_entry_exit_and_calls (); 6252 } 6253 6254 /* Reposition the prologue-end and epilogue-begin notes after 6255 instruction scheduling. */ 6256 6257 void 6258 reposition_prologue_and_epilogue_notes (void) 6259 { 6260 if (!targetm.have_prologue () 6261 && !targetm.have_epilogue () 6262 && !targetm.have_sibcall_epilogue ()) 6263 return; 6264 6265 /* Since the hash table is created on demand, the fact that it is 6266 non-null is a signal that it is non-empty. */ 6267 if (prologue_insn_hash != NULL) 6268 { 6269 size_t len = prologue_insn_hash->elements (); 6270 rtx_insn *insn, *last = NULL, *note = NULL; 6271 6272 /* Scan from the beginning until we reach the last prologue insn. */ 6273 /* ??? While we do have the CFG intact, there are two problems: 6274 (1) The prologue can contain loops (typically probing the stack), 6275 which means that the end of the prologue isn't in the first bb. 6276 (2) Sometimes the PROLOGUE_END note gets pushed into the next bb. */ 6277 for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) 6278 { 6279 if (NOTE_P (insn)) 6280 { 6281 if (NOTE_KIND (insn) == NOTE_INSN_PROLOGUE_END) 6282 note = insn; 6283 } 6284 else if (contains (insn, prologue_insn_hash)) 6285 { 6286 last = insn; 6287 if (--len == 0) 6288 break; 6289 } 6290 } 6291 6292 if (last) 6293 { 6294 if (note == NULL) 6295 { 6296 /* Scan forward looking for the PROLOGUE_END note. It should 6297 be right at the beginning of the block, possibly with other 6298 insn notes that got moved there. */ 6299 for (note = NEXT_INSN (last); ; note = NEXT_INSN (note)) 6300 { 6301 if (NOTE_P (note) 6302 && NOTE_KIND (note) == NOTE_INSN_PROLOGUE_END) 6303 break; 6304 } 6305 } 6306 6307 /* Avoid placing note between CODE_LABEL and BASIC_BLOCK note. */ 6308 if (LABEL_P (last)) 6309 last = NEXT_INSN (last); 6310 reorder_insns (note, note, last); 6311 } 6312 } 6313 6314 if (epilogue_insn_hash != NULL) 6315 { 6316 edge_iterator ei; 6317 edge e; 6318 6319 FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) 6320 { 6321 rtx_insn *insn, *first = NULL, *note = NULL; 6322 basic_block bb = e->src; 6323 6324 /* Scan from the beginning until we reach the first epilogue insn. */ 6325 FOR_BB_INSNS (bb, insn) 6326 { 6327 if (NOTE_P (insn)) 6328 { 6329 if (NOTE_KIND (insn) == NOTE_INSN_EPILOGUE_BEG) 6330 { 6331 note = insn; 6332 if (first != NULL) 6333 break; 6334 } 6335 } 6336 else if (first == NULL && contains (insn, epilogue_insn_hash)) 6337 { 6338 first = insn; 6339 if (note != NULL) 6340 break; 6341 } 6342 } 6343 6344 if (note) 6345 { 6346 /* If the function has a single basic block, and no real 6347 epilogue insns (e.g. sibcall with no cleanup), the 6348 epilogue note can get scheduled before the prologue 6349 note. If we have frame related prologue insns, having 6350 them scanned during the epilogue will result in a crash. 6351 In this case re-order the epilogue note to just before 6352 the last insn in the block. */ 6353 if (first == NULL) 6354 first = BB_END (bb); 6355 6356 if (PREV_INSN (first) != note) 6357 reorder_insns (note, note, PREV_INSN (first)); 6358 } 6359 } 6360 } 6361 } 6362 6363 /* Returns the name of function declared by FNDECL. */ 6364 const char * 6365 fndecl_name (tree fndecl) 6366 { 6367 if (fndecl == NULL) 6368 return "(nofn)"; 6369 return lang_hooks.decl_printable_name (fndecl, 1); 6370 } 6371 6372 /* Returns the name of function FN. */ 6373 const char * 6374 function_name (struct function *fn) 6375 { 6376 tree fndecl = (fn == NULL) ? NULL : fn->decl; 6377 return fndecl_name (fndecl); 6378 } 6379 6380 /* Returns the name of the current function. */ 6381 const char * 6382 current_function_name (void) 6383 { 6384 return function_name (cfun); 6385 } 6386 6387 6388 static unsigned int 6389 rest_of_handle_check_leaf_regs (void) 6390 { 6391 #ifdef LEAF_REGISTERS 6392 crtl->uses_only_leaf_regs 6393 = optimize > 0 && only_leaf_regs_used () && leaf_function_p (); 6394 #endif 6395 return 0; 6396 } 6397 6398 /* Insert a TYPE into the used types hash table of CFUN. */ 6399 6400 static void 6401 used_types_insert_helper (tree type, struct function *func) 6402 { 6403 if (type != NULL && func != NULL) 6404 { 6405 if (func->used_types_hash == NULL) 6406 func->used_types_hash = hash_set<tree>::create_ggc (37); 6407 6408 func->used_types_hash->add (type); 6409 } 6410 } 6411 6412 /* Given a type, insert it into the used hash table in cfun. */ 6413 void 6414 used_types_insert (tree t) 6415 { 6416 while (POINTER_TYPE_P (t) || TREE_CODE (t) == ARRAY_TYPE) 6417 if (TYPE_NAME (t)) 6418 break; 6419 else 6420 t = TREE_TYPE (t); 6421 if (TREE_CODE (t) == ERROR_MARK) 6422 return; 6423 if (TYPE_NAME (t) == NULL_TREE 6424 || TYPE_NAME (t) == TYPE_NAME (TYPE_MAIN_VARIANT (t))) 6425 t = TYPE_MAIN_VARIANT (t); 6426 if (debug_info_level > DINFO_LEVEL_NONE) 6427 { 6428 if (cfun) 6429 used_types_insert_helper (t, cfun); 6430 else 6431 { 6432 /* So this might be a type referenced by a global variable. 6433 Record that type so that we can later decide to emit its 6434 debug information. */ 6435 vec_safe_push (types_used_by_cur_var_decl, t); 6436 } 6437 } 6438 } 6439 6440 /* Helper to Hash a struct types_used_by_vars_entry. */ 6441 6442 static hashval_t 6443 hash_types_used_by_vars_entry (const struct types_used_by_vars_entry *entry) 6444 { 6445 gcc_assert (entry && entry->var_decl && entry->type); 6446 6447 return iterative_hash_object (entry->type, 6448 iterative_hash_object (entry->var_decl, 0)); 6449 } 6450 6451 /* Hash function of the types_used_by_vars_entry hash table. */ 6452 6453 hashval_t 6454 used_type_hasher::hash (types_used_by_vars_entry *entry) 6455 { 6456 return hash_types_used_by_vars_entry (entry); 6457 } 6458 6459 /*Equality function of the types_used_by_vars_entry hash table. */ 6460 6461 bool 6462 used_type_hasher::equal (types_used_by_vars_entry *e1, 6463 types_used_by_vars_entry *e2) 6464 { 6465 return (e1->var_decl == e2->var_decl && e1->type == e2->type); 6466 } 6467 6468 /* Inserts an entry into the types_used_by_vars_hash hash table. */ 6469 6470 void 6471 types_used_by_var_decl_insert (tree type, tree var_decl) 6472 { 6473 if (type != NULL && var_decl != NULL) 6474 { 6475 types_used_by_vars_entry **slot; 6476 struct types_used_by_vars_entry e; 6477 e.var_decl = var_decl; 6478 e.type = type; 6479 if (types_used_by_vars_hash == NULL) 6480 types_used_by_vars_hash 6481 = hash_table<used_type_hasher>::create_ggc (37); 6482 6483 slot = types_used_by_vars_hash->find_slot (&e, INSERT); 6484 if (*slot == NULL) 6485 { 6486 struct types_used_by_vars_entry *entry; 6487 entry = ggc_alloc<types_used_by_vars_entry> (); 6488 entry->type = type; 6489 entry->var_decl = var_decl; 6490 *slot = entry; 6491 } 6492 } 6493 } 6494 6495 namespace { 6496 6497 const pass_data pass_data_leaf_regs = 6498 { 6499 RTL_PASS, /* type */ 6500 "*leaf_regs", /* name */ 6501 OPTGROUP_NONE, /* optinfo_flags */ 6502 TV_NONE, /* tv_id */ 6503 0, /* properties_required */ 6504 0, /* properties_provided */ 6505 0, /* properties_destroyed */ 6506 0, /* todo_flags_start */ 6507 0, /* todo_flags_finish */ 6508 }; 6509 6510 class pass_leaf_regs : public rtl_opt_pass 6511 { 6512 public: 6513 pass_leaf_regs (gcc::context *ctxt) 6514 : rtl_opt_pass (pass_data_leaf_regs, ctxt) 6515 {} 6516 6517 /* opt_pass methods: */ 6518 virtual unsigned int execute (function *) 6519 { 6520 return rest_of_handle_check_leaf_regs (); 6521 } 6522 6523 }; // class pass_leaf_regs 6524 6525 } // anon namespace 6526 6527 rtl_opt_pass * 6528 make_pass_leaf_regs (gcc::context *ctxt) 6529 { 6530 return new pass_leaf_regs (ctxt); 6531 } 6532 6533 static unsigned int 6534 rest_of_handle_thread_prologue_and_epilogue (void) 6535 { 6536 /* prepare_shrink_wrap is sensitive to the block structure of the control 6537 flow graph, so clean it up first. */ 6538 if (optimize) 6539 cleanup_cfg (0); 6540 6541 /* On some machines, the prologue and epilogue code, or parts thereof, 6542 can be represented as RTL. Doing so lets us schedule insns between 6543 it and the rest of the code and also allows delayed branch 6544 scheduling to operate in the epilogue. */ 6545 thread_prologue_and_epilogue_insns (); 6546 6547 /* Some non-cold blocks may now be only reachable from cold blocks. 6548 Fix that up. */ 6549 fixup_partitions (); 6550 6551 /* Shrink-wrapping can result in unreachable edges in the epilogue, 6552 see PR57320. */ 6553 cleanup_cfg (optimize ? CLEANUP_EXPENSIVE : 0); 6554 6555 /* The stack usage info is finalized during prologue expansion. */ 6556 if (flag_stack_usage_info) 6557 output_stack_usage (); 6558 6559 return 0; 6560 } 6561 6562 namespace { 6563 6564 const pass_data pass_data_thread_prologue_and_epilogue = 6565 { 6566 RTL_PASS, /* type */ 6567 "pro_and_epilogue", /* name */ 6568 OPTGROUP_NONE, /* optinfo_flags */ 6569 TV_THREAD_PROLOGUE_AND_EPILOGUE, /* tv_id */ 6570 0, /* properties_required */ 6571 0, /* properties_provided */ 6572 0, /* properties_destroyed */ 6573 0, /* todo_flags_start */ 6574 ( TODO_df_verify | TODO_df_finish ), /* todo_flags_finish */ 6575 }; 6576 6577 class pass_thread_prologue_and_epilogue : public rtl_opt_pass 6578 { 6579 public: 6580 pass_thread_prologue_and_epilogue (gcc::context *ctxt) 6581 : rtl_opt_pass (pass_data_thread_prologue_and_epilogue, ctxt) 6582 {} 6583 6584 /* opt_pass methods: */ 6585 virtual unsigned int execute (function *) 6586 { 6587 return rest_of_handle_thread_prologue_and_epilogue (); 6588 } 6589 6590 }; // class pass_thread_prologue_and_epilogue 6591 6592 } // anon namespace 6593 6594 rtl_opt_pass * 6595 make_pass_thread_prologue_and_epilogue (gcc::context *ctxt) 6596 { 6597 return new pass_thread_prologue_and_epilogue (ctxt); 6598 } 6599 6600 6601 /* This mini-pass fixes fall-out from SSA in asm statements that have 6602 in-out constraints. Say you start with 6603 6604 orig = inout; 6605 asm ("": "+mr" (inout)); 6606 use (orig); 6607 6608 which is transformed very early to use explicit output and match operands: 6609 6610 orig = inout; 6611 asm ("": "=mr" (inout) : "0" (inout)); 6612 use (orig); 6613 6614 Or, after SSA and copyprop, 6615 6616 asm ("": "=mr" (inout_2) : "0" (inout_1)); 6617 use (inout_1); 6618 6619 Clearly inout_2 and inout_1 can't be coalesced easily anymore, as 6620 they represent two separate values, so they will get different pseudo 6621 registers during expansion. Then, since the two operands need to match 6622 per the constraints, but use different pseudo registers, reload can 6623 only register a reload for these operands. But reloads can only be 6624 satisfied by hardregs, not by memory, so we need a register for this 6625 reload, just because we are presented with non-matching operands. 6626 So, even though we allow memory for this operand, no memory can be 6627 used for it, just because the two operands don't match. This can 6628 cause reload failures on register-starved targets. 6629 6630 So it's a symptom of reload not being able to use memory for reloads 6631 or, alternatively it's also a symptom of both operands not coming into 6632 reload as matching (in which case the pseudo could go to memory just 6633 fine, as the alternative allows it, and no reload would be necessary). 6634 We fix the latter problem here, by transforming 6635 6636 asm ("": "=mr" (inout_2) : "0" (inout_1)); 6637 6638 back to 6639 6640 inout_2 = inout_1; 6641 asm ("": "=mr" (inout_2) : "0" (inout_2)); */ 6642 6643 static void 6644 match_asm_constraints_1 (rtx_insn *insn, rtx *p_sets, int noutputs) 6645 { 6646 int i; 6647 bool changed = false; 6648 rtx op = SET_SRC (p_sets[0]); 6649 int ninputs = ASM_OPERANDS_INPUT_LENGTH (op); 6650 rtvec inputs = ASM_OPERANDS_INPUT_VEC (op); 6651 bool *output_matched = XALLOCAVEC (bool, noutputs); 6652 6653 memset (output_matched, 0, noutputs * sizeof (bool)); 6654 for (i = 0; i < ninputs; i++) 6655 { 6656 rtx input, output; 6657 rtx_insn *insns; 6658 const char *constraint = ASM_OPERANDS_INPUT_CONSTRAINT (op, i); 6659 char *end; 6660 int match, j; 6661 6662 if (*constraint == '%') 6663 constraint++; 6664 6665 match = strtoul (constraint, &end, 10); 6666 if (end == constraint) 6667 continue; 6668 6669 gcc_assert (match < noutputs); 6670 output = SET_DEST (p_sets[match]); 6671 input = RTVEC_ELT (inputs, i); 6672 /* Only do the transformation for pseudos. */ 6673 if (! REG_P (output) 6674 || rtx_equal_p (output, input) 6675 || !(REG_P (input) || SUBREG_P (input) 6676 || MEM_P (input) || CONSTANT_P (input)) 6677 || !general_operand (input, GET_MODE (output))) 6678 continue; 6679 6680 /* We can't do anything if the output is also used as input, 6681 as we're going to overwrite it. */ 6682 for (j = 0; j < ninputs; j++) 6683 if (reg_overlap_mentioned_p (output, RTVEC_ELT (inputs, j))) 6684 break; 6685 if (j != ninputs) 6686 continue; 6687 6688 /* Avoid changing the same input several times. For 6689 asm ("" : "=mr" (out1), "=mr" (out2) : "0" (in), "1" (in)); 6690 only change in once (to out1), rather than changing it 6691 first to out1 and afterwards to out2. */ 6692 if (i > 0) 6693 { 6694 for (j = 0; j < noutputs; j++) 6695 if (output_matched[j] && input == SET_DEST (p_sets[j])) 6696 break; 6697 if (j != noutputs) 6698 continue; 6699 } 6700 output_matched[match] = true; 6701 6702 start_sequence (); 6703 emit_move_insn (output, copy_rtx (input)); 6704 insns = get_insns (); 6705 end_sequence (); 6706 emit_insn_before (insns, insn); 6707 6708 /* Now replace all mentions of the input with output. We can't 6709 just replace the occurrence in inputs[i], as the register might 6710 also be used in some other input (or even in an address of an 6711 output), which would mean possibly increasing the number of 6712 inputs by one (namely 'output' in addition), which might pose 6713 a too complicated problem for reload to solve. E.g. this situation: 6714 6715 asm ("" : "=r" (output), "=m" (input) : "0" (input)) 6716 6717 Here 'input' is used in two occurrences as input (once for the 6718 input operand, once for the address in the second output operand). 6719 If we would replace only the occurrence of the input operand (to 6720 make the matching) we would be left with this: 6721 6722 output = input 6723 asm ("" : "=r" (output), "=m" (input) : "0" (output)) 6724 6725 Now we suddenly have two different input values (containing the same 6726 value, but different pseudos) where we formerly had only one. 6727 With more complicated asms this might lead to reload failures 6728 which wouldn't have happen without this pass. So, iterate over 6729 all operands and replace all occurrences of the register used. */ 6730 for (j = 0; j < noutputs; j++) 6731 if (!rtx_equal_p (SET_DEST (p_sets[j]), input) 6732 && reg_overlap_mentioned_p (input, SET_DEST (p_sets[j]))) 6733 SET_DEST (p_sets[j]) = replace_rtx (SET_DEST (p_sets[j]), 6734 input, output); 6735 for (j = 0; j < ninputs; j++) 6736 if (reg_overlap_mentioned_p (input, RTVEC_ELT (inputs, j))) 6737 RTVEC_ELT (inputs, j) = replace_rtx (RTVEC_ELT (inputs, j), 6738 input, output); 6739 6740 changed = true; 6741 } 6742 6743 if (changed) 6744 df_insn_rescan (insn); 6745 } 6746 6747 /* Add the decl D to the local_decls list of FUN. */ 6748 6749 void 6750 add_local_decl (struct function *fun, tree d) 6751 { 6752 gcc_assert (VAR_P (d)); 6753 vec_safe_push (fun->local_decls, d); 6754 } 6755 6756 namespace { 6757 6758 const pass_data pass_data_match_asm_constraints = 6759 { 6760 RTL_PASS, /* type */ 6761 "asmcons", /* name */ 6762 OPTGROUP_NONE, /* optinfo_flags */ 6763 TV_NONE, /* tv_id */ 6764 0, /* properties_required */ 6765 0, /* properties_provided */ 6766 0, /* properties_destroyed */ 6767 0, /* todo_flags_start */ 6768 0, /* todo_flags_finish */ 6769 }; 6770 6771 class pass_match_asm_constraints : public rtl_opt_pass 6772 { 6773 public: 6774 pass_match_asm_constraints (gcc::context *ctxt) 6775 : rtl_opt_pass (pass_data_match_asm_constraints, ctxt) 6776 {} 6777 6778 /* opt_pass methods: */ 6779 virtual unsigned int execute (function *); 6780 6781 }; // class pass_match_asm_constraints 6782 6783 unsigned 6784 pass_match_asm_constraints::execute (function *fun) 6785 { 6786 basic_block bb; 6787 rtx_insn *insn; 6788 rtx pat, *p_sets; 6789 int noutputs; 6790 6791 if (!crtl->has_asm_statement) 6792 return 0; 6793 6794 df_set_flags (DF_DEFER_INSN_RESCAN); 6795 FOR_EACH_BB_FN (bb, fun) 6796 { 6797 FOR_BB_INSNS (bb, insn) 6798 { 6799 if (!INSN_P (insn)) 6800 continue; 6801 6802 pat = PATTERN (insn); 6803 if (GET_CODE (pat) == PARALLEL) 6804 p_sets = &XVECEXP (pat, 0, 0), noutputs = XVECLEN (pat, 0); 6805 else if (GET_CODE (pat) == SET) 6806 p_sets = &PATTERN (insn), noutputs = 1; 6807 else 6808 continue; 6809 6810 if (GET_CODE (*p_sets) == SET 6811 && GET_CODE (SET_SRC (*p_sets)) == ASM_OPERANDS) 6812 match_asm_constraints_1 (insn, p_sets, noutputs); 6813 } 6814 } 6815 6816 return TODO_df_finish; 6817 } 6818 6819 } // anon namespace 6820 6821 rtl_opt_pass * 6822 make_pass_match_asm_constraints (gcc::context *ctxt) 6823 { 6824 return new pass_match_asm_constraints (ctxt); 6825 } 6826 6827 6828 #include "gt-function.h" 6829