1/* A pass for lowering trees to RTL. 2 Copyright (C) 2004, 2005 Free Software Foundation, Inc. 3 4This file is part of GCC. 5 6GCC is free software; you can redistribute it and/or modify 7it under the terms of the GNU General Public License as published by 8the Free Software Foundation; either version 2, or (at your option) 9any later version. 10 11GCC is distributed in the hope that it will be useful, 12but WITHOUT ANY WARRANTY; without even the implied warranty of 13MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14GNU General Public License for more details. 15 16You should have received a copy of the GNU General Public License 17along with GCC; see the file COPYING. If not, write to 18the Free Software Foundation, 51 Franklin Street, Fifth Floor, 19Boston, MA 02110-1301, USA. */ 20 21#include "config.h" 22#include "system.h" 23#include "coretypes.h" 24#include "tm.h" 25#include "tree.h" 26#include "rtl.h" 27#include "tm_p.h" 28#include "basic-block.h" 29#include "function.h" 30#include "expr.h" 31#include "langhooks.h" 32#include "tree-flow.h" 33#include "timevar.h" 34#include "tree-dump.h" 35#include "tree-pass.h" 36#include "except.h" 37#include "flags.h" 38#include "diagnostic.h" 39#include "toplev.h" 40#include "debug.h" 41#include "params.h" 42 43/* Verify that there is exactly single jump instruction since last and attach 44 REG_BR_PROB note specifying probability. 45 ??? We really ought to pass the probability down to RTL expanders and let it 46 re-distribute it when the conditional expands into multiple conditionals. 47 This is however difficult to do. */ 48static void 49add_reg_br_prob_note (rtx last, int probability) 50{ 51 if (profile_status == PROFILE_ABSENT) 52 return; 53 for (last = NEXT_INSN (last); last && NEXT_INSN (last); last = NEXT_INSN (last)) 54 if (JUMP_P (last)) 55 { 56 /* It is common to emit condjump-around-jump sequence when we don't know 57 how to reverse the conditional. Special case this. */ 58 if (!any_condjump_p (last) 59 || !JUMP_P (NEXT_INSN (last)) 60 || !simplejump_p (NEXT_INSN (last)) 61 || !NEXT_INSN (NEXT_INSN (last)) 62 || !BARRIER_P (NEXT_INSN (NEXT_INSN (last))) 63 || !NEXT_INSN (NEXT_INSN (NEXT_INSN (last))) 64 || !LABEL_P (NEXT_INSN (NEXT_INSN (NEXT_INSN (last)))) 65 || NEXT_INSN (NEXT_INSN (NEXT_INSN (NEXT_INSN (last))))) 66 goto failed; 67 gcc_assert (!find_reg_note (last, REG_BR_PROB, 0)); 68 REG_NOTES (last) 69 = gen_rtx_EXPR_LIST (REG_BR_PROB, 70 GEN_INT (REG_BR_PROB_BASE - probability), 71 REG_NOTES (last)); 72 return; 73 } 74 if (!last || !JUMP_P (last) || !any_condjump_p (last)) 75 goto failed; 76 gcc_assert (!find_reg_note (last, REG_BR_PROB, 0)); 77 REG_NOTES (last) 78 = gen_rtx_EXPR_LIST (REG_BR_PROB, 79 GEN_INT (probability), REG_NOTES (last)); 80 return; 81failed: 82 if (dump_file) 83 fprintf (dump_file, "Failed to add probability note\n"); 84} 85 86 87#ifndef LOCAL_ALIGNMENT 88#define LOCAL_ALIGNMENT(TYPE, ALIGNMENT) ALIGNMENT 89#endif 90 91#ifndef STACK_ALIGNMENT_NEEDED 92#define STACK_ALIGNMENT_NEEDED 1 93#endif 94 95 96/* This structure holds data relevant to one variable that will be 97 placed in a stack slot. */ 98struct stack_var 99{ 100 /* The Variable. */ 101 tree decl; 102 103 /* The offset of the variable. During partitioning, this is the 104 offset relative to the partition. After partitioning, this 105 is relative to the stack frame. */ 106 HOST_WIDE_INT offset; 107 108 /* Initially, the size of the variable. Later, the size of the partition, 109 if this variable becomes it's partition's representative. */ 110 HOST_WIDE_INT size; 111 112 /* The *byte* alignment required for this variable. Or as, with the 113 size, the alignment for this partition. */ 114 unsigned int alignb; 115 116 /* The partition representative. */ 117 size_t representative; 118 119 /* The next stack variable in the partition, or EOC. */ 120 size_t next; 121}; 122 123#define EOC ((size_t)-1) 124 125/* We have an array of such objects while deciding allocation. */ 126static struct stack_var *stack_vars; 127static size_t stack_vars_alloc; 128static size_t stack_vars_num; 129 130/* An array of indicies such that stack_vars[stack_vars_sorted[i]].size 131 is non-decreasing. */ 132static size_t *stack_vars_sorted; 133 134/* We have an interference graph between such objects. This graph 135 is lower triangular. */ 136static bool *stack_vars_conflict; 137static size_t stack_vars_conflict_alloc; 138 139/* The phase of the stack frame. This is the known misalignment of 140 virtual_stack_vars_rtx from PREFERRED_STACK_BOUNDARY. That is, 141 (frame_offset+frame_phase) % PREFERRED_STACK_BOUNDARY == 0. */ 142static int frame_phase; 143 144/* Used during expand_used_vars to remember if we saw any decls for 145 which we'd like to enable stack smashing protection. */ 146static bool has_protected_decls; 147 148/* Used during expand_used_vars. Remember if we say a character buffer 149 smaller than our cutoff threshold. Used for -Wstack-protector. */ 150static bool has_short_buffer; 151 152/* Discover the byte alignment to use for DECL. Ignore alignment 153 we can't do with expected alignment of the stack boundary. */ 154 155static unsigned int 156get_decl_align_unit (tree decl) 157{ 158 unsigned int align; 159 160 align = DECL_ALIGN (decl); 161 align = LOCAL_ALIGNMENT (TREE_TYPE (decl), align); 162 if (align > PREFERRED_STACK_BOUNDARY) 163 align = PREFERRED_STACK_BOUNDARY; 164 if (cfun->stack_alignment_needed < align) 165 cfun->stack_alignment_needed = align; 166 167 return align / BITS_PER_UNIT; 168} 169 170/* Allocate SIZE bytes at byte alignment ALIGN from the stack frame. 171 Return the frame offset. */ 172 173static HOST_WIDE_INT 174alloc_stack_frame_space (HOST_WIDE_INT size, HOST_WIDE_INT align) 175{ 176 HOST_WIDE_INT offset, new_frame_offset; 177 178 new_frame_offset = frame_offset; 179 if (FRAME_GROWS_DOWNWARD) 180 { 181 new_frame_offset -= size + frame_phase; 182 new_frame_offset &= -align; 183 new_frame_offset += frame_phase; 184 offset = new_frame_offset; 185 } 186 else 187 { 188 new_frame_offset -= frame_phase; 189 new_frame_offset += align - 1; 190 new_frame_offset &= -align; 191 new_frame_offset += frame_phase; 192 offset = new_frame_offset; 193 new_frame_offset += size; 194 } 195 frame_offset = new_frame_offset; 196 197 if (frame_offset_overflow (frame_offset, cfun->decl)) 198 frame_offset = offset = 0; 199 200 return offset; 201} 202 203/* Accumulate DECL into STACK_VARS. */ 204 205static void 206add_stack_var (tree decl) 207{ 208 if (stack_vars_num >= stack_vars_alloc) 209 { 210 if (stack_vars_alloc) 211 stack_vars_alloc = stack_vars_alloc * 3 / 2; 212 else 213 stack_vars_alloc = 32; 214 stack_vars 215 = XRESIZEVEC (struct stack_var, stack_vars, stack_vars_alloc); 216 } 217 stack_vars[stack_vars_num].decl = decl; 218 stack_vars[stack_vars_num].offset = 0; 219 stack_vars[stack_vars_num].size = tree_low_cst (DECL_SIZE_UNIT (decl), 1); 220 stack_vars[stack_vars_num].alignb = get_decl_align_unit (decl); 221 222 /* All variables are initially in their own partition. */ 223 stack_vars[stack_vars_num].representative = stack_vars_num; 224 stack_vars[stack_vars_num].next = EOC; 225 226 /* Ensure that this decl doesn't get put onto the list twice. */ 227 SET_DECL_RTL (decl, pc_rtx); 228 229 stack_vars_num++; 230} 231 232/* Compute the linear index of a lower-triangular coordinate (I, J). */ 233 234static size_t 235triangular_index (size_t i, size_t j) 236{ 237 if (i < j) 238 { 239 size_t t; 240 t = i, i = j, j = t; 241 } 242 return (i * (i + 1)) / 2 + j; 243} 244 245/* Ensure that STACK_VARS_CONFLICT is large enough for N objects. */ 246 247static void 248resize_stack_vars_conflict (size_t n) 249{ 250 size_t size = triangular_index (n-1, n-1) + 1; 251 252 if (size <= stack_vars_conflict_alloc) 253 return; 254 255 stack_vars_conflict = XRESIZEVEC (bool, stack_vars_conflict, size); 256 memset (stack_vars_conflict + stack_vars_conflict_alloc, 0, 257 (size - stack_vars_conflict_alloc) * sizeof (bool)); 258 stack_vars_conflict_alloc = size; 259} 260 261/* Make the decls associated with luid's X and Y conflict. */ 262 263static void 264add_stack_var_conflict (size_t x, size_t y) 265{ 266 size_t index = triangular_index (x, y); 267 gcc_assert (index < stack_vars_conflict_alloc); 268 stack_vars_conflict[index] = true; 269} 270 271/* Check whether the decls associated with luid's X and Y conflict. */ 272 273static bool 274stack_var_conflict_p (size_t x, size_t y) 275{ 276 size_t index = triangular_index (x, y); 277 gcc_assert (index < stack_vars_conflict_alloc); 278 return stack_vars_conflict[index]; 279} 280 281/* Returns true if TYPE is or contains a union type. */ 282 283static bool 284aggregate_contains_union_type (tree type) 285{ 286 tree field; 287 288 if (TREE_CODE (type) == UNION_TYPE 289 || TREE_CODE (type) == QUAL_UNION_TYPE) 290 return true; 291 if (TREE_CODE (type) == ARRAY_TYPE) 292 return aggregate_contains_union_type (TREE_TYPE (type)); 293 if (TREE_CODE (type) != RECORD_TYPE) 294 return false; 295 296 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) 297 if (TREE_CODE (field) == FIELD_DECL) 298 if (aggregate_contains_union_type (TREE_TYPE (field))) 299 return true; 300 301 return false; 302} 303 304/* A subroutine of expand_used_vars. If two variables X and Y have alias 305 sets that do not conflict, then do add a conflict for these variables 306 in the interference graph. We also need to make sure to add conflicts 307 for union containing structures. Else RTL alias analysis comes along 308 and due to type based aliasing rules decides that for two overlapping 309 union temporaries { short s; int i; } accesses to the same mem through 310 different types may not alias and happily reorders stores across 311 life-time boundaries of the temporaries (See PR25654). 312 We also have to mind MEM_IN_STRUCT_P and MEM_SCALAR_P. */ 313 314static void 315add_alias_set_conflicts (void) 316{ 317 size_t i, j, n = stack_vars_num; 318 319 for (i = 0; i < n; ++i) 320 { 321 tree type_i = TREE_TYPE (stack_vars[i].decl); 322 bool aggr_i = AGGREGATE_TYPE_P (type_i); 323 bool contains_union; 324 325 contains_union = aggregate_contains_union_type (type_i); 326 for (j = 0; j < i; ++j) 327 { 328 tree type_j = TREE_TYPE (stack_vars[j].decl); 329 bool aggr_j = AGGREGATE_TYPE_P (type_j); 330 if (aggr_i != aggr_j 331 /* Either the objects conflict by means of type based 332 aliasing rules, or we need to add a conflict. */ 333 || !objects_must_conflict_p (type_i, type_j) 334 /* In case the types do not conflict ensure that access 335 to elements will conflict. In case of unions we have 336 to be careful as type based aliasing rules may say 337 access to the same memory does not conflict. So play 338 safe and add a conflict in this case. */ 339 || contains_union) 340 add_stack_var_conflict (i, j); 341 } 342 } 343} 344 345/* A subroutine of partition_stack_vars. A comparison function for qsort, 346 sorting an array of indicies by the size of the object. */ 347 348static int 349stack_var_size_cmp (const void *a, const void *b) 350{ 351 HOST_WIDE_INT sa = stack_vars[*(const size_t *)a].size; 352 HOST_WIDE_INT sb = stack_vars[*(const size_t *)b].size; 353 unsigned int uida = DECL_UID (stack_vars[*(const size_t *)a].decl); 354 unsigned int uidb = DECL_UID (stack_vars[*(const size_t *)b].decl); 355 356 if (sa < sb) 357 return -1; 358 if (sa > sb) 359 return 1; 360 /* For stack variables of the same size use the uid of the decl 361 to make the sort stable. */ 362 if (uida < uidb) 363 return -1; 364 if (uida > uidb) 365 return 1; 366 return 0; 367} 368 369/* A subroutine of partition_stack_vars. The UNION portion of a UNION/FIND 370 partitioning algorithm. Partitions A and B are known to be non-conflicting. 371 Merge them into a single partition A. 372 373 At the same time, add OFFSET to all variables in partition B. At the end 374 of the partitioning process we've have a nice block easy to lay out within 375 the stack frame. */ 376 377static void 378union_stack_vars (size_t a, size_t b, HOST_WIDE_INT offset) 379{ 380 size_t i, last; 381 382 /* Update each element of partition B with the given offset, 383 and merge them into partition A. */ 384 for (last = i = b; i != EOC; last = i, i = stack_vars[i].next) 385 { 386 stack_vars[i].offset += offset; 387 stack_vars[i].representative = a; 388 } 389 stack_vars[last].next = stack_vars[a].next; 390 stack_vars[a].next = b; 391 392 /* Update the required alignment of partition A to account for B. */ 393 if (stack_vars[a].alignb < stack_vars[b].alignb) 394 stack_vars[a].alignb = stack_vars[b].alignb; 395 396 /* Update the interference graph and merge the conflicts. */ 397 for (last = stack_vars_num, i = 0; i < last; ++i) 398 if (stack_var_conflict_p (b, i)) 399 add_stack_var_conflict (a, i); 400} 401 402/* A subroutine of expand_used_vars. Binpack the variables into 403 partitions constrained by the interference graph. The overall 404 algorithm used is as follows: 405 406 Sort the objects by size. 407 For each object A { 408 S = size(A) 409 O = 0 410 loop { 411 Look for the largest non-conflicting object B with size <= S. 412 UNION (A, B) 413 offset(B) = O 414 O += size(B) 415 S -= size(B) 416 } 417 } 418*/ 419 420static void 421partition_stack_vars (void) 422{ 423 size_t si, sj, n = stack_vars_num; 424 425 stack_vars_sorted = XNEWVEC (size_t, stack_vars_num); 426 for (si = 0; si < n; ++si) 427 stack_vars_sorted[si] = si; 428 429 if (n == 1) 430 return; 431 432 qsort (stack_vars_sorted, n, sizeof (size_t), stack_var_size_cmp); 433 434 /* Special case: detect when all variables conflict, and thus we can't 435 do anything during the partitioning loop. It isn't uncommon (with 436 C code at least) to declare all variables at the top of the function, 437 and if we're not inlining, then all variables will be in the same scope. 438 Take advantage of very fast libc routines for this scan. */ 439 gcc_assert (sizeof(bool) == sizeof(char)); 440 if (memchr (stack_vars_conflict, false, stack_vars_conflict_alloc) == NULL) 441 return; 442 443 for (si = 0; si < n; ++si) 444 { 445 size_t i = stack_vars_sorted[si]; 446 HOST_WIDE_INT isize = stack_vars[i].size; 447 HOST_WIDE_INT offset = 0; 448 449 for (sj = si; sj-- > 0; ) 450 { 451 size_t j = stack_vars_sorted[sj]; 452 HOST_WIDE_INT jsize = stack_vars[j].size; 453 unsigned int jalign = stack_vars[j].alignb; 454 455 /* Ignore objects that aren't partition representatives. */ 456 if (stack_vars[j].representative != j) 457 continue; 458 459 /* Ignore objects too large for the remaining space. */ 460 if (isize < jsize) 461 continue; 462 463 /* Ignore conflicting objects. */ 464 if (stack_var_conflict_p (i, j)) 465 continue; 466 467 /* Refine the remaining space check to include alignment. */ 468 if (offset & (jalign - 1)) 469 { 470 HOST_WIDE_INT toff = offset; 471 toff += jalign - 1; 472 toff &= -(HOST_WIDE_INT)jalign; 473 if (isize - (toff - offset) < jsize) 474 continue; 475 476 isize -= toff - offset; 477 offset = toff; 478 } 479 480 /* UNION the objects, placing J at OFFSET. */ 481 union_stack_vars (i, j, offset); 482 483 isize -= jsize; 484 if (isize == 0) 485 break; 486 } 487 } 488} 489 490/* A debugging aid for expand_used_vars. Dump the generated partitions. */ 491 492static void 493dump_stack_var_partition (void) 494{ 495 size_t si, i, j, n = stack_vars_num; 496 497 for (si = 0; si < n; ++si) 498 { 499 i = stack_vars_sorted[si]; 500 501 /* Skip variables that aren't partition representatives, for now. */ 502 if (stack_vars[i].representative != i) 503 continue; 504 505 fprintf (dump_file, "Partition %lu: size " HOST_WIDE_INT_PRINT_DEC 506 " align %u\n", (unsigned long) i, stack_vars[i].size, 507 stack_vars[i].alignb); 508 509 for (j = i; j != EOC; j = stack_vars[j].next) 510 { 511 fputc ('\t', dump_file); 512 print_generic_expr (dump_file, stack_vars[j].decl, dump_flags); 513 fprintf (dump_file, ", offset " HOST_WIDE_INT_PRINT_DEC "\n", 514 stack_vars[i].offset); 515 } 516 } 517} 518 519/* Assign rtl to DECL at frame offset OFFSET. */ 520 521static void 522expand_one_stack_var_at (tree decl, HOST_WIDE_INT offset) 523{ 524 HOST_WIDE_INT align; 525 rtx x; 526 527 /* If this fails, we've overflowed the stack frame. Error nicely? */ 528 gcc_assert (offset == trunc_int_for_mode (offset, Pmode)); 529 530 x = plus_constant (virtual_stack_vars_rtx, offset); 531 x = gen_rtx_MEM (DECL_MODE (decl), x); 532 533 /* Set alignment we actually gave this decl. */ 534 offset -= frame_phase; 535 align = offset & -offset; 536 align *= BITS_PER_UNIT; 537 if (align > STACK_BOUNDARY || align == 0) 538 align = STACK_BOUNDARY; 539 DECL_ALIGN (decl) = align; 540 DECL_USER_ALIGN (decl) = 0; 541 542 set_mem_attributes (x, decl, true); 543 SET_DECL_RTL (decl, x); 544} 545 546/* A subroutine of expand_used_vars. Give each partition representative 547 a unique location within the stack frame. Update each partition member 548 with that location. */ 549 550static void 551expand_stack_vars (bool (*pred) (tree)) 552{ 553 size_t si, i, j, n = stack_vars_num; 554 555 for (si = 0; si < n; ++si) 556 { 557 HOST_WIDE_INT offset; 558 559 i = stack_vars_sorted[si]; 560 561 /* Skip variables that aren't partition representatives, for now. */ 562 if (stack_vars[i].representative != i) 563 continue; 564 565 /* Skip variables that have already had rtl assigned. See also 566 add_stack_var where we perpetrate this pc_rtx hack. */ 567 if (DECL_RTL (stack_vars[i].decl) != pc_rtx) 568 continue; 569 570 /* Check the predicate to see whether this variable should be 571 allocated in this pass. */ 572 if (pred && !pred (stack_vars[i].decl)) 573 continue; 574 575 offset = alloc_stack_frame_space (stack_vars[i].size, 576 stack_vars[i].alignb); 577 578 /* Create rtl for each variable based on their location within the 579 partition. */ 580 for (j = i; j != EOC; j = stack_vars[j].next) 581 expand_one_stack_var_at (stack_vars[j].decl, 582 stack_vars[j].offset + offset); 583 } 584} 585 586/* A subroutine of expand_one_var. Called to immediately assign rtl 587 to a variable to be allocated in the stack frame. */ 588 589static void 590expand_one_stack_var (tree var) 591{ 592 HOST_WIDE_INT size, offset, align; 593 594 size = tree_low_cst (DECL_SIZE_UNIT (var), 1); 595 align = get_decl_align_unit (var); 596 offset = alloc_stack_frame_space (size, align); 597 598 expand_one_stack_var_at (var, offset); 599} 600 601/* A subroutine of expand_one_var. Called to assign rtl 602 to a TREE_STATIC VAR_DECL. */ 603 604static void 605expand_one_static_var (tree var) 606{ 607 /* In unit-at-a-time all the static variables are expanded at the end 608 of compilation process. */ 609 if (flag_unit_at_a_time) 610 return; 611 /* If this is an inlined copy of a static local variable, 612 look up the original. */ 613 var = DECL_ORIGIN (var); 614 615 /* If we've already processed this variable because of that, do nothing. */ 616 if (TREE_ASM_WRITTEN (var)) 617 return; 618 619 /* Give the front end a chance to do whatever. In practice, this is 620 resolving duplicate names for IMA in C. */ 621 if (lang_hooks.expand_decl (var)) 622 return; 623 624 /* Otherwise, just emit the variable. */ 625 rest_of_decl_compilation (var, 0, 0); 626} 627 628/* A subroutine of expand_one_var. Called to assign rtl to a VAR_DECL 629 that will reside in a hard register. */ 630 631static void 632expand_one_hard_reg_var (tree var) 633{ 634 rest_of_decl_compilation (var, 0, 0); 635} 636 637/* A subroutine of expand_one_var. Called to assign rtl to a VAR_DECL 638 that will reside in a pseudo register. */ 639 640static void 641expand_one_register_var (tree var) 642{ 643 tree type = TREE_TYPE (var); 644 int unsignedp = TYPE_UNSIGNED (type); 645 enum machine_mode reg_mode 646 = promote_mode (type, DECL_MODE (var), &unsignedp, 0); 647 rtx x = gen_reg_rtx (reg_mode); 648 649 SET_DECL_RTL (var, x); 650 651 /* Note if the object is a user variable. */ 652 if (!DECL_ARTIFICIAL (var)) 653 { 654 mark_user_reg (x); 655 656 /* Trust user variables which have a pointer type to really 657 be pointers. Do not trust compiler generated temporaries 658 as our type system is totally busted as it relates to 659 pointer arithmetic which translates into lots of compiler 660 generated objects with pointer types, but which are not really 661 pointers. */ 662 if (POINTER_TYPE_P (type)) 663 mark_reg_pointer (x, TYPE_ALIGN (TREE_TYPE (TREE_TYPE (var)))); 664 } 665} 666 667/* A subroutine of expand_one_var. Called to assign rtl to a VAR_DECL that 668 has some associated error, e.g. its type is error-mark. We just need 669 to pick something that won't crash the rest of the compiler. */ 670 671static void 672expand_one_error_var (tree var) 673{ 674 enum machine_mode mode = DECL_MODE (var); 675 rtx x; 676 677 if (mode == BLKmode) 678 x = gen_rtx_MEM (BLKmode, const0_rtx); 679 else if (mode == VOIDmode) 680 x = const0_rtx; 681 else 682 x = gen_reg_rtx (mode); 683 684 SET_DECL_RTL (var, x); 685} 686 687/* A subroutine of expand_one_var. VAR is a variable that will be 688 allocated to the local stack frame. Return true if we wish to 689 add VAR to STACK_VARS so that it will be coalesced with other 690 variables. Return false to allocate VAR immediately. 691 692 This function is used to reduce the number of variables considered 693 for coalescing, which reduces the size of the quadratic problem. */ 694 695static bool 696defer_stack_allocation (tree var, bool toplevel) 697{ 698 /* If stack protection is enabled, *all* stack variables must be deferred, 699 so that we can re-order the strings to the top of the frame. */ 700 if (flag_stack_protect) 701 return true; 702 703 /* Variables in the outermost scope automatically conflict with 704 every other variable. The only reason to want to defer them 705 at all is that, after sorting, we can more efficiently pack 706 small variables in the stack frame. Continue to defer at -O2. */ 707 if (toplevel && optimize < 2) 708 return false; 709 710 /* Without optimization, *most* variables are allocated from the 711 stack, which makes the quadratic problem large exactly when we 712 want compilation to proceed as quickly as possible. On the 713 other hand, we don't want the function's stack frame size to 714 get completely out of hand. So we avoid adding scalars and 715 "small" aggregates to the list at all. */ 716 if (optimize == 0 && tree_low_cst (DECL_SIZE_UNIT (var), 1) < 32) 717 return false; 718 719 return true; 720} 721 722/* A subroutine of expand_used_vars. Expand one variable according to 723 its flavor. Variables to be placed on the stack are not actually 724 expanded yet, merely recorded. */ 725 726static void 727expand_one_var (tree var, bool toplevel) 728{ 729 if (TREE_CODE (var) != VAR_DECL) 730 lang_hooks.expand_decl (var); 731 else if (DECL_EXTERNAL (var)) 732 ; 733 else if (DECL_HAS_VALUE_EXPR_P (var)) 734 ; 735 else if (TREE_STATIC (var)) 736 expand_one_static_var (var); 737 else if (DECL_RTL_SET_P (var)) 738 ; 739 else if (TREE_TYPE (var) == error_mark_node) 740 expand_one_error_var (var); 741 else if (DECL_HARD_REGISTER (var)) 742 expand_one_hard_reg_var (var); 743 else if (use_register_for_decl (var)) 744 expand_one_register_var (var); 745 else if (defer_stack_allocation (var, toplevel)) 746 add_stack_var (var); 747 else 748 expand_one_stack_var (var); 749} 750 751/* A subroutine of expand_used_vars. Walk down through the BLOCK tree 752 expanding variables. Those variables that can be put into registers 753 are allocated pseudos; those that can't are put on the stack. 754 755 TOPLEVEL is true if this is the outermost BLOCK. */ 756 757static void 758expand_used_vars_for_block (tree block, bool toplevel) 759{ 760 size_t i, j, old_sv_num, this_sv_num, new_sv_num; 761 tree t; 762 763 old_sv_num = toplevel ? 0 : stack_vars_num; 764 765 /* Expand all variables at this level. */ 766 for (t = BLOCK_VARS (block); t ; t = TREE_CHAIN (t)) 767 if (TREE_USED (t) 768 /* Force local static variables to be output when marked by 769 used attribute. For unit-at-a-time, cgraph code already takes 770 care of this. */ 771 || (!flag_unit_at_a_time && TREE_STATIC (t) 772 && DECL_PRESERVE_P (t))) 773 expand_one_var (t, toplevel); 774 775 this_sv_num = stack_vars_num; 776 777 /* Expand all variables at containing levels. */ 778 for (t = BLOCK_SUBBLOCKS (block); t ; t = BLOCK_CHAIN (t)) 779 expand_used_vars_for_block (t, false); 780 781 /* Since we do not track exact variable lifetimes (which is not even 782 possible for variables whose address escapes), we mirror the block 783 tree in the interference graph. Here we cause all variables at this 784 level, and all sublevels, to conflict. Do make certain that a 785 variable conflicts with itself. */ 786 if (old_sv_num < this_sv_num) 787 { 788 new_sv_num = stack_vars_num; 789 resize_stack_vars_conflict (new_sv_num); 790 791 for (i = old_sv_num; i < new_sv_num; ++i) 792 for (j = i < this_sv_num ? i+1 : this_sv_num; j-- > old_sv_num ;) 793 add_stack_var_conflict (i, j); 794 } 795} 796 797/* A subroutine of expand_used_vars. Walk down through the BLOCK tree 798 and clear TREE_USED on all local variables. */ 799 800static void 801clear_tree_used (tree block) 802{ 803 tree t; 804 805 for (t = BLOCK_VARS (block); t ; t = TREE_CHAIN (t)) 806 /* if (!TREE_STATIC (t) && !DECL_EXTERNAL (t)) */ 807 TREE_USED (t) = 0; 808 809 for (t = BLOCK_SUBBLOCKS (block); t ; t = BLOCK_CHAIN (t)) 810 clear_tree_used (t); 811} 812 813/* Examine TYPE and determine a bit mask of the following features. */ 814 815#define SPCT_HAS_LARGE_CHAR_ARRAY 1 816#define SPCT_HAS_SMALL_CHAR_ARRAY 2 817#define SPCT_HAS_ARRAY 4 818#define SPCT_HAS_AGGREGATE 8 819 820static unsigned int 821stack_protect_classify_type (tree type) 822{ 823 unsigned int ret = 0; 824 tree t; 825 826 switch (TREE_CODE (type)) 827 { 828 case ARRAY_TYPE: 829 t = TYPE_MAIN_VARIANT (TREE_TYPE (type)); 830 if (t == char_type_node 831 || t == signed_char_type_node 832 || t == unsigned_char_type_node) 833 { 834 unsigned HOST_WIDE_INT max = PARAM_VALUE (PARAM_SSP_BUFFER_SIZE); 835 unsigned HOST_WIDE_INT len; 836 837 if (!TYPE_SIZE_UNIT (type) 838 || !host_integerp (TYPE_SIZE_UNIT (type), 1)) 839 len = max; 840 else 841 len = tree_low_cst (TYPE_SIZE_UNIT (type), 1); 842 843 if (len < max) 844 ret = SPCT_HAS_SMALL_CHAR_ARRAY | SPCT_HAS_ARRAY; 845 else 846 ret = SPCT_HAS_LARGE_CHAR_ARRAY | SPCT_HAS_ARRAY; 847 } 848 else 849 ret = SPCT_HAS_ARRAY; 850 break; 851 852 case UNION_TYPE: 853 case QUAL_UNION_TYPE: 854 case RECORD_TYPE: 855 ret = SPCT_HAS_AGGREGATE; 856 for (t = TYPE_FIELDS (type); t ; t = TREE_CHAIN (t)) 857 if (TREE_CODE (t) == FIELD_DECL) 858 ret |= stack_protect_classify_type (TREE_TYPE (t)); 859 break; 860 861 default: 862 break; 863 } 864 865 return ret; 866} 867 868/* Return nonzero if DECL should be segregated into the "vulnerable" upper 869 part of the local stack frame. Remember if we ever return nonzero for 870 any variable in this function. The return value is the phase number in 871 which the variable should be allocated. */ 872 873static int 874stack_protect_decl_phase (tree decl) 875{ 876 unsigned int bits = stack_protect_classify_type (TREE_TYPE (decl)); 877 int ret = 0; 878 879 if (bits & SPCT_HAS_SMALL_CHAR_ARRAY) 880 has_short_buffer = true; 881 882 if (flag_stack_protect == 2) 883 { 884 if ((bits & (SPCT_HAS_SMALL_CHAR_ARRAY | SPCT_HAS_LARGE_CHAR_ARRAY)) 885 && !(bits & SPCT_HAS_AGGREGATE)) 886 ret = 1; 887 else if (bits & SPCT_HAS_ARRAY) 888 ret = 2; 889 } 890 else 891 ret = (bits & SPCT_HAS_LARGE_CHAR_ARRAY) != 0; 892 893 if (ret) 894 has_protected_decls = true; 895 896 return ret; 897} 898 899/* Two helper routines that check for phase 1 and phase 2. These are used 900 as callbacks for expand_stack_vars. */ 901 902static bool 903stack_protect_decl_phase_1 (tree decl) 904{ 905 return stack_protect_decl_phase (decl) == 1; 906} 907 908static bool 909stack_protect_decl_phase_2 (tree decl) 910{ 911 return stack_protect_decl_phase (decl) == 2; 912} 913 914/* Ensure that variables in different stack protection phases conflict 915 so that they are not merged and share the same stack slot. */ 916 917static void 918add_stack_protection_conflicts (void) 919{ 920 size_t i, j, n = stack_vars_num; 921 unsigned char *phase; 922 923 phase = XNEWVEC (unsigned char, n); 924 for (i = 0; i < n; ++i) 925 phase[i] = stack_protect_decl_phase (stack_vars[i].decl); 926 927 for (i = 0; i < n; ++i) 928 { 929 unsigned char ph_i = phase[i]; 930 for (j = 0; j < i; ++j) 931 if (ph_i != phase[j]) 932 add_stack_var_conflict (i, j); 933 } 934 935 XDELETEVEC (phase); 936} 937 938/* Create a decl for the guard at the top of the stack frame. */ 939 940static void 941create_stack_guard (void) 942{ 943 tree guard = build_decl (VAR_DECL, NULL, ptr_type_node); 944 TREE_THIS_VOLATILE (guard) = 1; 945 TREE_USED (guard) = 1; 946 expand_one_stack_var (guard); 947 cfun->stack_protect_guard = guard; 948} 949 950/* Expand all variables used in the function. */ 951 952static void 953expand_used_vars (void) 954{ 955 tree t, outer_block = DECL_INITIAL (current_function_decl); 956 957 /* Compute the phase of the stack frame for this function. */ 958 { 959 int align = PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT; 960 int off = STARTING_FRAME_OFFSET % align; 961 frame_phase = off ? align - off : 0; 962 } 963 964 /* Set TREE_USED on all variables in the unexpanded_var_list. */ 965 for (t = cfun->unexpanded_var_list; t; t = TREE_CHAIN (t)) 966 TREE_USED (TREE_VALUE (t)) = 1; 967 968 /* Clear TREE_USED on all variables associated with a block scope. */ 969 clear_tree_used (outer_block); 970 971 /* Initialize local stack smashing state. */ 972 has_protected_decls = false; 973 has_short_buffer = false; 974 975 /* At this point all variables on the unexpanded_var_list with TREE_USED 976 set are not associated with any block scope. Lay them out. */ 977 for (t = cfun->unexpanded_var_list; t; t = TREE_CHAIN (t)) 978 { 979 tree var = TREE_VALUE (t); 980 bool expand_now = false; 981 982 /* We didn't set a block for static or extern because it's hard 983 to tell the difference between a global variable (re)declared 984 in a local scope, and one that's really declared there to 985 begin with. And it doesn't really matter much, since we're 986 not giving them stack space. Expand them now. */ 987 if (TREE_STATIC (var) || DECL_EXTERNAL (var)) 988 expand_now = true; 989 990 /* Any variable that could have been hoisted into an SSA_NAME 991 will have been propagated anywhere the optimizers chose, 992 i.e. not confined to their original block. Allocate them 993 as if they were defined in the outermost scope. */ 994 else if (is_gimple_reg (var)) 995 expand_now = true; 996 997 /* If the variable is not associated with any block, then it 998 was created by the optimizers, and could be live anywhere 999 in the function. */ 1000 else if (TREE_USED (var)) 1001 expand_now = true; 1002 1003 /* Finally, mark all variables on the list as used. We'll use 1004 this in a moment when we expand those associated with scopes. */ 1005 TREE_USED (var) = 1; 1006 1007 if (expand_now) 1008 expand_one_var (var, true); 1009 } 1010 cfun->unexpanded_var_list = NULL_TREE; 1011 1012 /* At this point, all variables within the block tree with TREE_USED 1013 set are actually used by the optimized function. Lay them out. */ 1014 expand_used_vars_for_block (outer_block, true); 1015 1016 if (stack_vars_num > 0) 1017 { 1018 /* Due to the way alias sets work, no variables with non-conflicting 1019 alias sets may be assigned the same address. Add conflicts to 1020 reflect this. */ 1021 add_alias_set_conflicts (); 1022 1023 /* If stack protection is enabled, we don't share space between 1024 vulnerable data and non-vulnerable data. */ 1025 if (flag_stack_protect) 1026 add_stack_protection_conflicts (); 1027 1028 /* Now that we have collected all stack variables, and have computed a 1029 minimal interference graph, attempt to save some stack space. */ 1030 partition_stack_vars (); 1031 if (dump_file) 1032 dump_stack_var_partition (); 1033 } 1034 1035 /* There are several conditions under which we should create a 1036 stack guard: protect-all, alloca used, protected decls present. */ 1037 if (flag_stack_protect == 2 1038 || (flag_stack_protect 1039 && (current_function_calls_alloca || has_protected_decls))) 1040 create_stack_guard (); 1041 1042 /* Assign rtl to each variable based on these partitions. */ 1043 if (stack_vars_num > 0) 1044 { 1045 /* Reorder decls to be protected by iterating over the variables 1046 array multiple times, and allocating out of each phase in turn. */ 1047 /* ??? We could probably integrate this into the qsort we did 1048 earlier, such that we naturally see these variables first, 1049 and thus naturally allocate things in the right order. */ 1050 if (has_protected_decls) 1051 { 1052 /* Phase 1 contains only character arrays. */ 1053 expand_stack_vars (stack_protect_decl_phase_1); 1054 1055 /* Phase 2 contains other kinds of arrays. */ 1056 if (flag_stack_protect == 2) 1057 expand_stack_vars (stack_protect_decl_phase_2); 1058 } 1059 1060 expand_stack_vars (NULL); 1061 1062 /* Free up stack variable graph data. */ 1063 XDELETEVEC (stack_vars); 1064 XDELETEVEC (stack_vars_sorted); 1065 XDELETEVEC (stack_vars_conflict); 1066 stack_vars = NULL; 1067 stack_vars_alloc = stack_vars_num = 0; 1068 stack_vars_conflict = NULL; 1069 stack_vars_conflict_alloc = 0; 1070 } 1071 1072 /* If the target requires that FRAME_OFFSET be aligned, do it. */ 1073 if (STACK_ALIGNMENT_NEEDED) 1074 { 1075 HOST_WIDE_INT align = PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT; 1076 if (!FRAME_GROWS_DOWNWARD) 1077 frame_offset += align - 1; 1078 frame_offset &= -align; 1079 } 1080} 1081 1082 1083/* If we need to produce a detailed dump, print the tree representation 1084 for STMT to the dump file. SINCE is the last RTX after which the RTL 1085 generated for STMT should have been appended. */ 1086 1087static void 1088maybe_dump_rtl_for_tree_stmt (tree stmt, rtx since) 1089{ 1090 if (dump_file && (dump_flags & TDF_DETAILS)) 1091 { 1092 fprintf (dump_file, "\n;; "); 1093 print_generic_expr (dump_file, stmt, TDF_SLIM); 1094 fprintf (dump_file, "\n"); 1095 1096 print_rtl (dump_file, since ? NEXT_INSN (since) : since); 1097 } 1098} 1099 1100/* A subroutine of expand_gimple_basic_block. Expand one COND_EXPR. 1101 Returns a new basic block if we've terminated the current basic 1102 block and created a new one. */ 1103 1104static basic_block 1105expand_gimple_cond_expr (basic_block bb, tree stmt) 1106{ 1107 basic_block new_bb, dest; 1108 edge new_edge; 1109 edge true_edge; 1110 edge false_edge; 1111 tree pred = COND_EXPR_COND (stmt); 1112 tree then_exp = COND_EXPR_THEN (stmt); 1113 tree else_exp = COND_EXPR_ELSE (stmt); 1114 rtx last2, last; 1115 1116 last2 = last = get_last_insn (); 1117 1118 extract_true_false_edges_from_block (bb, &true_edge, &false_edge); 1119 if (EXPR_LOCUS (stmt)) 1120 { 1121 emit_line_note (*(EXPR_LOCUS (stmt))); 1122 record_block_change (TREE_BLOCK (stmt)); 1123 } 1124 1125 /* These flags have no purpose in RTL land. */ 1126 true_edge->flags &= ~EDGE_TRUE_VALUE; 1127 false_edge->flags &= ~EDGE_FALSE_VALUE; 1128 1129 /* We can either have a pure conditional jump with one fallthru edge or 1130 two-way jump that needs to be decomposed into two basic blocks. */ 1131 if (TREE_CODE (then_exp) == GOTO_EXPR && IS_EMPTY_STMT (else_exp)) 1132 { 1133 jumpif (pred, label_rtx (GOTO_DESTINATION (then_exp))); 1134 add_reg_br_prob_note (last, true_edge->probability); 1135 maybe_dump_rtl_for_tree_stmt (stmt, last); 1136 if (EXPR_LOCUS (then_exp)) 1137 emit_line_note (*(EXPR_LOCUS (then_exp))); 1138 return NULL; 1139 } 1140 if (TREE_CODE (else_exp) == GOTO_EXPR && IS_EMPTY_STMT (then_exp)) 1141 { 1142 jumpifnot (pred, label_rtx (GOTO_DESTINATION (else_exp))); 1143 add_reg_br_prob_note (last, false_edge->probability); 1144 maybe_dump_rtl_for_tree_stmt (stmt, last); 1145 if (EXPR_LOCUS (else_exp)) 1146 emit_line_note (*(EXPR_LOCUS (else_exp))); 1147 return NULL; 1148 } 1149 gcc_assert (TREE_CODE (then_exp) == GOTO_EXPR 1150 && TREE_CODE (else_exp) == GOTO_EXPR); 1151 1152 jumpif (pred, label_rtx (GOTO_DESTINATION (then_exp))); 1153 add_reg_br_prob_note (last, true_edge->probability); 1154 last = get_last_insn (); 1155 expand_expr (else_exp, const0_rtx, VOIDmode, 0); 1156 1157 BB_END (bb) = last; 1158 if (BARRIER_P (BB_END (bb))) 1159 BB_END (bb) = PREV_INSN (BB_END (bb)); 1160 update_bb_for_insn (bb); 1161 1162 new_bb = create_basic_block (NEXT_INSN (last), get_last_insn (), bb); 1163 dest = false_edge->dest; 1164 redirect_edge_succ (false_edge, new_bb); 1165 false_edge->flags |= EDGE_FALLTHRU; 1166 new_bb->count = false_edge->count; 1167 new_bb->frequency = EDGE_FREQUENCY (false_edge); 1168 new_edge = make_edge (new_bb, dest, 0); 1169 new_edge->probability = REG_BR_PROB_BASE; 1170 new_edge->count = new_bb->count; 1171 if (BARRIER_P (BB_END (new_bb))) 1172 BB_END (new_bb) = PREV_INSN (BB_END (new_bb)); 1173 update_bb_for_insn (new_bb); 1174 1175 maybe_dump_rtl_for_tree_stmt (stmt, last2); 1176 1177 if (EXPR_LOCUS (else_exp)) 1178 emit_line_note (*(EXPR_LOCUS (else_exp))); 1179 1180 return new_bb; 1181} 1182 1183/* A subroutine of expand_gimple_basic_block. Expand one CALL_EXPR 1184 that has CALL_EXPR_TAILCALL set. Returns non-null if we actually 1185 generated a tail call (something that might be denied by the ABI 1186 rules governing the call; see calls.c). 1187 1188 Sets CAN_FALLTHRU if we generated a *conditional* tail call, and 1189 can still reach the rest of BB. The case here is __builtin_sqrt, 1190 where the NaN result goes through the external function (with a 1191 tailcall) and the normal result happens via a sqrt instruction. */ 1192 1193static basic_block 1194expand_gimple_tailcall (basic_block bb, tree stmt, bool *can_fallthru) 1195{ 1196 rtx last2, last; 1197 edge e; 1198 edge_iterator ei; 1199 int probability; 1200 gcov_type count; 1201 1202 last2 = last = get_last_insn (); 1203 1204 expand_expr_stmt (stmt); 1205 1206 for (last = NEXT_INSN (last); last; last = NEXT_INSN (last)) 1207 if (CALL_P (last) && SIBLING_CALL_P (last)) 1208 goto found; 1209 1210 maybe_dump_rtl_for_tree_stmt (stmt, last2); 1211 1212 *can_fallthru = true; 1213 return NULL; 1214 1215 found: 1216 /* ??? Wouldn't it be better to just reset any pending stack adjust? 1217 Any instructions emitted here are about to be deleted. */ 1218 do_pending_stack_adjust (); 1219 1220 /* Remove any non-eh, non-abnormal edges that don't go to exit. */ 1221 /* ??? I.e. the fallthrough edge. HOWEVER! If there were to be 1222 EH or abnormal edges, we shouldn't have created a tail call in 1223 the first place. So it seems to me we should just be removing 1224 all edges here, or redirecting the existing fallthru edge to 1225 the exit block. */ 1226 1227 probability = 0; 1228 count = 0; 1229 1230 for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); ) 1231 { 1232 if (!(e->flags & (EDGE_ABNORMAL | EDGE_EH))) 1233 { 1234 if (e->dest != EXIT_BLOCK_PTR) 1235 { 1236 e->dest->count -= e->count; 1237 e->dest->frequency -= EDGE_FREQUENCY (e); 1238 if (e->dest->count < 0) 1239 e->dest->count = 0; 1240 if (e->dest->frequency < 0) 1241 e->dest->frequency = 0; 1242 } 1243 count += e->count; 1244 probability += e->probability; 1245 remove_edge (e); 1246 } 1247 else 1248 ei_next (&ei); 1249 } 1250 1251 /* This is somewhat ugly: the call_expr expander often emits instructions 1252 after the sibcall (to perform the function return). These confuse the 1253 find_many_sub_basic_blocks code, so we need to get rid of these. */ 1254 last = NEXT_INSN (last); 1255 gcc_assert (BARRIER_P (last)); 1256 1257 *can_fallthru = false; 1258 while (NEXT_INSN (last)) 1259 { 1260 /* For instance an sqrt builtin expander expands if with 1261 sibcall in the then and label for `else`. */ 1262 if (LABEL_P (NEXT_INSN (last))) 1263 { 1264 *can_fallthru = true; 1265 break; 1266 } 1267 delete_insn (NEXT_INSN (last)); 1268 } 1269 1270 e = make_edge (bb, EXIT_BLOCK_PTR, EDGE_ABNORMAL | EDGE_SIBCALL); 1271 e->probability += probability; 1272 e->count += count; 1273 BB_END (bb) = last; 1274 update_bb_for_insn (bb); 1275 1276 if (NEXT_INSN (last)) 1277 { 1278 bb = create_basic_block (NEXT_INSN (last), get_last_insn (), bb); 1279 1280 last = BB_END (bb); 1281 if (BARRIER_P (last)) 1282 BB_END (bb) = PREV_INSN (last); 1283 } 1284 1285 maybe_dump_rtl_for_tree_stmt (stmt, last2); 1286 1287 return bb; 1288} 1289 1290/* Expand basic block BB from GIMPLE trees to RTL. */ 1291 1292static basic_block 1293expand_gimple_basic_block (basic_block bb) 1294{ 1295 block_stmt_iterator bsi = bsi_start (bb); 1296 tree stmt = NULL; 1297 rtx note, last; 1298 edge e; 1299 edge_iterator ei; 1300 1301 if (dump_file) 1302 { 1303 fprintf (dump_file, 1304 "\n;; Generating RTL for tree basic block %d\n", 1305 bb->index); 1306 } 1307 1308 init_rtl_bb_info (bb); 1309 bb->flags |= BB_RTL; 1310 1311 if (!bsi_end_p (bsi)) 1312 stmt = bsi_stmt (bsi); 1313 1314 if (stmt && TREE_CODE (stmt) == LABEL_EXPR) 1315 { 1316 last = get_last_insn (); 1317 1318 expand_expr_stmt (stmt); 1319 1320 /* Java emits line number notes in the top of labels. 1321 ??? Make this go away once line number notes are obsoleted. */ 1322 BB_HEAD (bb) = NEXT_INSN (last); 1323 if (NOTE_P (BB_HEAD (bb))) 1324 BB_HEAD (bb) = NEXT_INSN (BB_HEAD (bb)); 1325 bsi_next (&bsi); 1326 note = emit_note_after (NOTE_INSN_BASIC_BLOCK, BB_HEAD (bb)); 1327 1328 maybe_dump_rtl_for_tree_stmt (stmt, last); 1329 } 1330 else 1331 note = BB_HEAD (bb) = emit_note (NOTE_INSN_BASIC_BLOCK); 1332 1333 NOTE_BASIC_BLOCK (note) = bb; 1334 1335 for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); ) 1336 { 1337 /* Clear EDGE_EXECUTABLE. This flag is never used in the backend. */ 1338 e->flags &= ~EDGE_EXECUTABLE; 1339 1340 /* At the moment not all abnormal edges match the RTL representation. 1341 It is safe to remove them here as find_many_sub_basic_blocks will 1342 rediscover them. In the future we should get this fixed properly. */ 1343 if (e->flags & EDGE_ABNORMAL) 1344 remove_edge (e); 1345 else 1346 ei_next (&ei); 1347 } 1348 1349 for (; !bsi_end_p (bsi); bsi_next (&bsi)) 1350 { 1351 tree stmt = bsi_stmt (bsi); 1352 basic_block new_bb; 1353 1354 if (!stmt) 1355 continue; 1356 1357 /* Expand this statement, then evaluate the resulting RTL and 1358 fixup the CFG accordingly. */ 1359 if (TREE_CODE (stmt) == COND_EXPR) 1360 { 1361 new_bb = expand_gimple_cond_expr (bb, stmt); 1362 if (new_bb) 1363 return new_bb; 1364 } 1365 else 1366 { 1367 tree call = get_call_expr_in (stmt); 1368 if (call && CALL_EXPR_TAILCALL (call)) 1369 { 1370 bool can_fallthru; 1371 new_bb = expand_gimple_tailcall (bb, stmt, &can_fallthru); 1372 if (new_bb) 1373 { 1374 if (can_fallthru) 1375 bb = new_bb; 1376 else 1377 return new_bb; 1378 } 1379 } 1380 else 1381 { 1382 last = get_last_insn (); 1383 expand_expr_stmt (stmt); 1384 maybe_dump_rtl_for_tree_stmt (stmt, last); 1385 } 1386 } 1387 } 1388 1389 do_pending_stack_adjust (); 1390 1391 /* Find the block tail. The last insn in the block is the insn 1392 before a barrier and/or table jump insn. */ 1393 last = get_last_insn (); 1394 if (BARRIER_P (last)) 1395 last = PREV_INSN (last); 1396 if (JUMP_TABLE_DATA_P (last)) 1397 last = PREV_INSN (PREV_INSN (last)); 1398 BB_END (bb) = last; 1399 1400 update_bb_for_insn (bb); 1401 1402 return bb; 1403} 1404 1405 1406/* Create a basic block for initialization code. */ 1407 1408static basic_block 1409construct_init_block (void) 1410{ 1411 basic_block init_block, first_block; 1412 edge e = NULL; 1413 int flags; 1414 1415 /* Multiple entry points not supported yet. */ 1416 gcc_assert (EDGE_COUNT (ENTRY_BLOCK_PTR->succs) == 1); 1417 init_rtl_bb_info (ENTRY_BLOCK_PTR); 1418 init_rtl_bb_info (EXIT_BLOCK_PTR); 1419 ENTRY_BLOCK_PTR->flags |= BB_RTL; 1420 EXIT_BLOCK_PTR->flags |= BB_RTL; 1421 1422 e = EDGE_SUCC (ENTRY_BLOCK_PTR, 0); 1423 1424 /* When entry edge points to first basic block, we don't need jump, 1425 otherwise we have to jump into proper target. */ 1426 if (e && e->dest != ENTRY_BLOCK_PTR->next_bb) 1427 { 1428 tree label = tree_block_label (e->dest); 1429 1430 emit_jump (label_rtx (label)); 1431 flags = 0; 1432 } 1433 else 1434 flags = EDGE_FALLTHRU; 1435 1436 init_block = create_basic_block (NEXT_INSN (get_insns ()), 1437 get_last_insn (), 1438 ENTRY_BLOCK_PTR); 1439 init_block->frequency = ENTRY_BLOCK_PTR->frequency; 1440 init_block->count = ENTRY_BLOCK_PTR->count; 1441 if (e) 1442 { 1443 first_block = e->dest; 1444 redirect_edge_succ (e, init_block); 1445 e = make_edge (init_block, first_block, flags); 1446 } 1447 else 1448 e = make_edge (init_block, EXIT_BLOCK_PTR, EDGE_FALLTHRU); 1449 e->probability = REG_BR_PROB_BASE; 1450 e->count = ENTRY_BLOCK_PTR->count; 1451 1452 update_bb_for_insn (init_block); 1453 return init_block; 1454} 1455 1456 1457/* Create a block containing landing pads and similar stuff. */ 1458 1459static void 1460construct_exit_block (void) 1461{ 1462 rtx head = get_last_insn (); 1463 rtx end; 1464 basic_block exit_block; 1465 edge e, e2; 1466 unsigned ix; 1467 edge_iterator ei; 1468 1469 /* Make sure the locus is set to the end of the function, so that 1470 epilogue line numbers and warnings are set properly. */ 1471#ifdef USE_MAPPED_LOCATION 1472 if (cfun->function_end_locus != UNKNOWN_LOCATION) 1473#else 1474 if (cfun->function_end_locus.file) 1475#endif 1476 input_location = cfun->function_end_locus; 1477 1478 /* The following insns belong to the top scope. */ 1479 record_block_change (DECL_INITIAL (current_function_decl)); 1480 1481 /* Generate rtl for function exit. */ 1482 expand_function_end (); 1483 1484 end = get_last_insn (); 1485 if (head == end) 1486 return; 1487 while (NEXT_INSN (head) && NOTE_P (NEXT_INSN (head))) 1488 head = NEXT_INSN (head); 1489 exit_block = create_basic_block (NEXT_INSN (head), end, 1490 EXIT_BLOCK_PTR->prev_bb); 1491 exit_block->frequency = EXIT_BLOCK_PTR->frequency; 1492 exit_block->count = EXIT_BLOCK_PTR->count; 1493 1494 ix = 0; 1495 while (ix < EDGE_COUNT (EXIT_BLOCK_PTR->preds)) 1496 { 1497 e = EDGE_PRED (EXIT_BLOCK_PTR, ix); 1498 if (!(e->flags & EDGE_ABNORMAL)) 1499 redirect_edge_succ (e, exit_block); 1500 else 1501 ix++; 1502 } 1503 1504 e = make_edge (exit_block, EXIT_BLOCK_PTR, EDGE_FALLTHRU); 1505 e->probability = REG_BR_PROB_BASE; 1506 e->count = EXIT_BLOCK_PTR->count; 1507 FOR_EACH_EDGE (e2, ei, EXIT_BLOCK_PTR->preds) 1508 if (e2 != e) 1509 { 1510 e->count -= e2->count; 1511 exit_block->count -= e2->count; 1512 exit_block->frequency -= EDGE_FREQUENCY (e2); 1513 } 1514 if (e->count < 0) 1515 e->count = 0; 1516 if (exit_block->count < 0) 1517 exit_block->count = 0; 1518 if (exit_block->frequency < 0) 1519 exit_block->frequency = 0; 1520 update_bb_for_insn (exit_block); 1521} 1522 1523/* Helper function for discover_nonconstant_array_refs. 1524 Look for ARRAY_REF nodes with non-constant indexes and mark them 1525 addressable. */ 1526 1527static tree 1528discover_nonconstant_array_refs_r (tree * tp, int *walk_subtrees, 1529 void *data ATTRIBUTE_UNUSED) 1530{ 1531 tree t = *tp; 1532 1533 if (IS_TYPE_OR_DECL_P (t)) 1534 *walk_subtrees = 0; 1535 else if (TREE_CODE (t) == ARRAY_REF || TREE_CODE (t) == ARRAY_RANGE_REF) 1536 { 1537 while (((TREE_CODE (t) == ARRAY_REF || TREE_CODE (t) == ARRAY_RANGE_REF) 1538 && is_gimple_min_invariant (TREE_OPERAND (t, 1)) 1539 && (!TREE_OPERAND (t, 2) 1540 || is_gimple_min_invariant (TREE_OPERAND (t, 2)))) 1541 || (TREE_CODE (t) == COMPONENT_REF 1542 && (!TREE_OPERAND (t,2) 1543 || is_gimple_min_invariant (TREE_OPERAND (t, 2)))) 1544 || TREE_CODE (t) == BIT_FIELD_REF 1545 || TREE_CODE (t) == REALPART_EXPR 1546 || TREE_CODE (t) == IMAGPART_EXPR 1547 || TREE_CODE (t) == VIEW_CONVERT_EXPR 1548 || TREE_CODE (t) == NOP_EXPR 1549 || TREE_CODE (t) == CONVERT_EXPR) 1550 t = TREE_OPERAND (t, 0); 1551 1552 if (TREE_CODE (t) == ARRAY_REF || TREE_CODE (t) == ARRAY_RANGE_REF) 1553 { 1554 t = get_base_address (t); 1555 if (t && DECL_P (t)) 1556 TREE_ADDRESSABLE (t) = 1; 1557 } 1558 1559 *walk_subtrees = 0; 1560 } 1561 1562 return NULL_TREE; 1563} 1564 1565/* RTL expansion is not able to compile array references with variable 1566 offsets for arrays stored in single register. Discover such 1567 expressions and mark variables as addressable to avoid this 1568 scenario. */ 1569 1570static void 1571discover_nonconstant_array_refs (void) 1572{ 1573 basic_block bb; 1574 block_stmt_iterator bsi; 1575 1576 FOR_EACH_BB (bb) 1577 { 1578 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) 1579 walk_tree (bsi_stmt_ptr (bsi), discover_nonconstant_array_refs_r, 1580 NULL , NULL); 1581 } 1582} 1583 1584/* Translate the intermediate representation contained in the CFG 1585 from GIMPLE trees to RTL. 1586 1587 We do conversion per basic block and preserve/update the tree CFG. 1588 This implies we have to do some magic as the CFG can simultaneously 1589 consist of basic blocks containing RTL and GIMPLE trees. This can 1590 confuse the CFG hooks, so be careful to not manipulate CFG during 1591 the expansion. */ 1592 1593static unsigned int 1594tree_expand_cfg (void) 1595{ 1596 basic_block bb, init_block; 1597 sbitmap blocks; 1598 edge_iterator ei; 1599 edge e; 1600 1601 /* Some backends want to know that we are expanding to RTL. */ 1602 currently_expanding_to_rtl = 1; 1603 1604 /* Prepare the rtl middle end to start recording block changes. */ 1605 reset_block_changes (); 1606 1607 /* Mark arrays indexed with non-constant indices with TREE_ADDRESSABLE. */ 1608 discover_nonconstant_array_refs (); 1609 1610 /* Expand the variables recorded during gimple lowering. */ 1611 expand_used_vars (); 1612 1613 /* Honor stack protection warnings. */ 1614 if (warn_stack_protect) 1615 { 1616 if (current_function_calls_alloca) 1617 warning (0, "not protecting local variables: variable length buffer"); 1618 if (has_short_buffer && !cfun->stack_protect_guard) 1619 warning (0, "not protecting function: no buffer at least %d bytes long", 1620 (int) PARAM_VALUE (PARAM_SSP_BUFFER_SIZE)); 1621 } 1622 1623 /* Set up parameters and prepare for return, for the function. */ 1624 expand_function_start (current_function_decl); 1625 1626 /* If this function is `main', emit a call to `__main' 1627 to run global initializers, etc. */ 1628 if (DECL_NAME (current_function_decl) 1629 && MAIN_NAME_P (DECL_NAME (current_function_decl)) 1630 && DECL_FILE_SCOPE_P (current_function_decl)) 1631 expand_main_function (); 1632 1633 /* Initialize the stack_protect_guard field. This must happen after the 1634 call to __main (if any) so that the external decl is initialized. */ 1635 if (cfun->stack_protect_guard) 1636 stack_protect_prologue (); 1637 1638 /* Register rtl specific functions for cfg. */ 1639 rtl_register_cfg_hooks (); 1640 1641 init_block = construct_init_block (); 1642 1643 /* Clear EDGE_EXECUTABLE on the entry edge(s). It is cleaned from the 1644 remaining edges in expand_gimple_basic_block. */ 1645 FOR_EACH_EDGE (e, ei, ENTRY_BLOCK_PTR->succs) 1646 e->flags &= ~EDGE_EXECUTABLE; 1647 1648 FOR_BB_BETWEEN (bb, init_block->next_bb, EXIT_BLOCK_PTR, next_bb) 1649 bb = expand_gimple_basic_block (bb); 1650 1651 construct_exit_block (); 1652 1653 /* We're done expanding trees to RTL. */ 1654 currently_expanding_to_rtl = 0; 1655 1656 /* Convert tree EH labels to RTL EH labels, and clean out any unreachable 1657 EH regions. */ 1658 convert_from_eh_region_ranges (); 1659 1660 rebuild_jump_labels (get_insns ()); 1661 find_exception_handler_labels (); 1662 1663 blocks = sbitmap_alloc (last_basic_block); 1664 sbitmap_ones (blocks); 1665 find_many_sub_basic_blocks (blocks); 1666 purge_all_dead_edges (); 1667 sbitmap_free (blocks); 1668 1669 compact_blocks (); 1670#ifdef ENABLE_CHECKING 1671 verify_flow_info(); 1672#endif 1673 1674 /* There's no need to defer outputting this function any more; we 1675 know we want to output it. */ 1676 DECL_DEFER_OUTPUT (current_function_decl) = 0; 1677 1678 /* Now that we're done expanding trees to RTL, we shouldn't have any 1679 more CONCATs anywhere. */ 1680 generating_concat_p = 0; 1681 1682 finalize_block_changes (); 1683 1684 if (dump_file) 1685 { 1686 fprintf (dump_file, 1687 "\n\n;;\n;; Full RTL generated for this function:\n;;\n"); 1688 /* And the pass manager will dump RTL for us. */ 1689 } 1690 1691 /* If we're emitting a nested function, make sure its parent gets 1692 emitted as well. Doing otherwise confuses debug info. */ 1693 { 1694 tree parent; 1695 for (parent = DECL_CONTEXT (current_function_decl); 1696 parent != NULL_TREE; 1697 parent = get_containing_scope (parent)) 1698 if (TREE_CODE (parent) == FUNCTION_DECL) 1699 TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (parent)) = 1; 1700 } 1701 1702 /* We are now committed to emitting code for this function. Do any 1703 preparation, such as emitting abstract debug info for the inline 1704 before it gets mangled by optimization. */ 1705 if (cgraph_function_possibly_inlined_p (current_function_decl)) 1706 (*debug_hooks->outlining_inline_function) (current_function_decl); 1707 1708 TREE_ASM_WRITTEN (current_function_decl) = 1; 1709 1710 /* After expanding, the return labels are no longer needed. */ 1711 return_label = NULL; 1712 naked_return_label = NULL; 1713 return 0; 1714} 1715 1716struct tree_opt_pass pass_expand = 1717{ 1718 "expand", /* name */ 1719 NULL, /* gate */ 1720 tree_expand_cfg, /* execute */ 1721 NULL, /* sub */ 1722 NULL, /* next */ 1723 0, /* static_pass_number */ 1724 TV_EXPAND, /* tv_id */ 1725 /* ??? If TER is enabled, we actually receive GENERIC. */ 1726 PROP_gimple_leh | PROP_cfg, /* properties_required */ 1727 PROP_rtl, /* properties_provided */ 1728 PROP_trees, /* properties_destroyed */ 1729 0, /* todo_flags_start */ 1730 TODO_dump_func, /* todo_flags_finish */ 1731 'r' /* letter */ 1732}; 1733