1 /* Convert function calls to rtl insns, for GNU C compiler. 2 Copyright (C) 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 4 2011, 2012 Free Software Foundation, Inc. 5 6 This file is part of GCC. 7 8 GCC is free software; you can redistribute it and/or modify it under 9 the terms of the GNU General Public License as published by the Free 10 Software Foundation; either version 3, or (at your option) any later 11 version. 12 13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY 14 WARRANTY; without even the implied warranty of MERCHANTABILITY or 15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 16 for more details. 17 18 You should have received a copy of the GNU General Public License 19 along with GCC; see the file COPYING3. If not see 20 <http://www.gnu.org/licenses/>. */ 21 22 #include "config.h" 23 #include "system.h" 24 #include "coretypes.h" 25 #include "tm.h" 26 #include "rtl.h" 27 #include "tree.h" 28 #include "gimple.h" 29 #include "flags.h" 30 #include "expr.h" 31 #include "optabs.h" 32 #include "libfuncs.h" 33 #include "function.h" 34 #include "regs.h" 35 #include "diagnostic-core.h" 36 #include "output.h" 37 #include "tm_p.h" 38 #include "timevar.h" 39 #include "sbitmap.h" 40 #include "langhooks.h" 41 #include "target.h" 42 #include "cgraph.h" 43 #include "except.h" 44 #include "dbgcnt.h" 45 #include "tree-flow.h" 46 47 /* Like PREFERRED_STACK_BOUNDARY but in units of bytes, not bits. */ 48 #define STACK_BYTES (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT) 49 50 /* Data structure and subroutines used within expand_call. */ 51 52 struct arg_data 53 { 54 /* Tree node for this argument. */ 55 tree tree_value; 56 /* Mode for value; TYPE_MODE unless promoted. */ 57 enum machine_mode mode; 58 /* Current RTL value for argument, or 0 if it isn't precomputed. */ 59 rtx value; 60 /* Initially-compute RTL value for argument; only for const functions. */ 61 rtx initial_value; 62 /* Register to pass this argument in, 0 if passed on stack, or an 63 PARALLEL if the arg is to be copied into multiple non-contiguous 64 registers. */ 65 rtx reg; 66 /* Register to pass this argument in when generating tail call sequence. 67 This is not the same register as for normal calls on machines with 68 register windows. */ 69 rtx tail_call_reg; 70 /* If REG is a PARALLEL, this is a copy of VALUE pulled into the correct 71 form for emit_group_move. */ 72 rtx parallel_value; 73 /* If REG was promoted from the actual mode of the argument expression, 74 indicates whether the promotion is sign- or zero-extended. */ 75 int unsignedp; 76 /* Number of bytes to put in registers. 0 means put the whole arg 77 in registers. Also 0 if not passed in registers. */ 78 int partial; 79 /* Nonzero if argument must be passed on stack. 80 Note that some arguments may be passed on the stack 81 even though pass_on_stack is zero, just because FUNCTION_ARG says so. 82 pass_on_stack identifies arguments that *cannot* go in registers. */ 83 int pass_on_stack; 84 /* Some fields packaged up for locate_and_pad_parm. */ 85 struct locate_and_pad_arg_data locate; 86 /* Location on the stack at which parameter should be stored. The store 87 has already been done if STACK == VALUE. */ 88 rtx stack; 89 /* Location on the stack of the start of this argument slot. This can 90 differ from STACK if this arg pads downward. This location is known 91 to be aligned to TARGET_FUNCTION_ARG_BOUNDARY. */ 92 rtx stack_slot; 93 /* Place that this stack area has been saved, if needed. */ 94 rtx save_area; 95 /* If an argument's alignment does not permit direct copying into registers, 96 copy in smaller-sized pieces into pseudos. These are stored in a 97 block pointed to by this field. The next field says how many 98 word-sized pseudos we made. */ 99 rtx *aligned_regs; 100 int n_aligned_regs; 101 }; 102 103 /* A vector of one char per byte of stack space. A byte if nonzero if 104 the corresponding stack location has been used. 105 This vector is used to prevent a function call within an argument from 106 clobbering any stack already set up. */ 107 static char *stack_usage_map; 108 109 /* Size of STACK_USAGE_MAP. */ 110 static int highest_outgoing_arg_in_use; 111 112 /* A bitmap of virtual-incoming stack space. Bit is set if the corresponding 113 stack location's tail call argument has been already stored into the stack. 114 This bitmap is used to prevent sibling call optimization if function tries 115 to use parent's incoming argument slots when they have been already 116 overwritten with tail call arguments. */ 117 static sbitmap stored_args_map; 118 119 /* stack_arg_under_construction is nonzero when an argument may be 120 initialized with a constructor call (including a C function that 121 returns a BLKmode struct) and expand_call must take special action 122 to make sure the object being constructed does not overlap the 123 argument list for the constructor call. */ 124 static int stack_arg_under_construction; 125 126 static void emit_call_1 (rtx, tree, tree, tree, HOST_WIDE_INT, HOST_WIDE_INT, 127 HOST_WIDE_INT, rtx, rtx, int, rtx, int, 128 cumulative_args_t); 129 static void precompute_register_parameters (int, struct arg_data *, int *); 130 static int store_one_arg (struct arg_data *, rtx, int, int, int); 131 static void store_unaligned_arguments_into_pseudos (struct arg_data *, int); 132 static int finalize_must_preallocate (int, int, struct arg_data *, 133 struct args_size *); 134 static void precompute_arguments (int, struct arg_data *); 135 static int compute_argument_block_size (int, struct args_size *, tree, tree, int); 136 static void initialize_argument_information (int, struct arg_data *, 137 struct args_size *, int, 138 tree, tree, 139 tree, tree, cumulative_args_t, int, 140 rtx *, int *, int *, int *, 141 bool *, bool); 142 static void compute_argument_addresses (struct arg_data *, rtx, int); 143 static rtx rtx_for_function_call (tree, tree); 144 static void load_register_parameters (struct arg_data *, int, rtx *, int, 145 int, int *); 146 static rtx emit_library_call_value_1 (int, rtx, rtx, enum libcall_type, 147 enum machine_mode, int, va_list); 148 static int special_function_p (const_tree, int); 149 static int check_sibcall_argument_overlap_1 (rtx); 150 static int check_sibcall_argument_overlap (rtx, struct arg_data *, int); 151 152 static int combine_pending_stack_adjustment_and_call (int, struct args_size *, 153 unsigned int); 154 static tree split_complex_types (tree); 155 156 #ifdef REG_PARM_STACK_SPACE 157 static rtx save_fixed_argument_area (int, rtx, int *, int *); 158 static void restore_fixed_argument_area (rtx, rtx, int, int); 159 #endif 160 161 /* Force FUNEXP into a form suitable for the address of a CALL, 162 and return that as an rtx. Also load the static chain register 163 if FNDECL is a nested function. 164 165 CALL_FUSAGE points to a variable holding the prospective 166 CALL_INSN_FUNCTION_USAGE information. */ 167 168 rtx 169 prepare_call_address (tree fndecl, rtx funexp, rtx static_chain_value, 170 rtx *call_fusage, int reg_parm_seen, int sibcallp) 171 { 172 /* Make a valid memory address and copy constants through pseudo-regs, 173 but not for a constant address if -fno-function-cse. */ 174 if (GET_CODE (funexp) != SYMBOL_REF) 175 /* If we are using registers for parameters, force the 176 function address into a register now. */ 177 funexp = ((reg_parm_seen 178 && targetm.small_register_classes_for_mode_p (FUNCTION_MODE)) 179 ? force_not_mem (memory_address (FUNCTION_MODE, funexp)) 180 : memory_address (FUNCTION_MODE, funexp)); 181 else if (! sibcallp) 182 { 183 #ifndef NO_FUNCTION_CSE 184 if (optimize && ! flag_no_function_cse) 185 funexp = force_reg (Pmode, funexp); 186 #endif 187 } 188 189 if (static_chain_value != 0) 190 { 191 rtx chain; 192 193 gcc_assert (fndecl); 194 chain = targetm.calls.static_chain (fndecl, false); 195 static_chain_value = convert_memory_address (Pmode, static_chain_value); 196 197 emit_move_insn (chain, static_chain_value); 198 if (REG_P (chain)) 199 use_reg (call_fusage, chain); 200 } 201 202 return funexp; 203 } 204 205 /* Generate instructions to call function FUNEXP, 206 and optionally pop the results. 207 The CALL_INSN is the first insn generated. 208 209 FNDECL is the declaration node of the function. This is given to the 210 hook TARGET_RETURN_POPS_ARGS to determine whether this function pops 211 its own args. 212 213 FUNTYPE is the data type of the function. This is given to the hook 214 TARGET_RETURN_POPS_ARGS to determine whether this function pops its 215 own args. We used to allow an identifier for library functions, but 216 that doesn't work when the return type is an aggregate type and the 217 calling convention says that the pointer to this aggregate is to be 218 popped by the callee. 219 220 STACK_SIZE is the number of bytes of arguments on the stack, 221 ROUNDED_STACK_SIZE is that number rounded up to 222 PREFERRED_STACK_BOUNDARY; zero if the size is variable. This is 223 both to put into the call insn and to generate explicit popping 224 code if necessary. 225 226 STRUCT_VALUE_SIZE is the number of bytes wanted in a structure value. 227 It is zero if this call doesn't want a structure value. 228 229 NEXT_ARG_REG is the rtx that results from executing 230 targetm.calls.function_arg (&args_so_far, VOIDmode, void_type_node, true) 231 just after all the args have had their registers assigned. 232 This could be whatever you like, but normally it is the first 233 arg-register beyond those used for args in this call, 234 or 0 if all the arg-registers are used in this call. 235 It is passed on to `gen_call' so you can put this info in the call insn. 236 237 VALREG is a hard register in which a value is returned, 238 or 0 if the call does not return a value. 239 240 OLD_INHIBIT_DEFER_POP is the value that `inhibit_defer_pop' had before 241 the args to this call were processed. 242 We restore `inhibit_defer_pop' to that value. 243 244 CALL_FUSAGE is either empty or an EXPR_LIST of USE expressions that 245 denote registers used by the called function. */ 246 247 static void 248 emit_call_1 (rtx funexp, tree fntree ATTRIBUTE_UNUSED, tree fndecl ATTRIBUTE_UNUSED, 249 tree funtype ATTRIBUTE_UNUSED, 250 HOST_WIDE_INT stack_size ATTRIBUTE_UNUSED, 251 HOST_WIDE_INT rounded_stack_size, 252 HOST_WIDE_INT struct_value_size ATTRIBUTE_UNUSED, 253 rtx next_arg_reg ATTRIBUTE_UNUSED, rtx valreg, 254 int old_inhibit_defer_pop, rtx call_fusage, int ecf_flags, 255 cumulative_args_t args_so_far ATTRIBUTE_UNUSED) 256 { 257 rtx rounded_stack_size_rtx = GEN_INT (rounded_stack_size); 258 rtx call_insn, call, funmem; 259 int already_popped = 0; 260 HOST_WIDE_INT n_popped 261 = targetm.calls.return_pops_args (fndecl, funtype, stack_size); 262 263 #ifdef CALL_POPS_ARGS 264 n_popped += CALL_POPS_ARGS (*get_cumulative_args (args_so_far)); 265 #endif 266 267 /* Ensure address is valid. SYMBOL_REF is already valid, so no need, 268 and we don't want to load it into a register as an optimization, 269 because prepare_call_address already did it if it should be done. */ 270 if (GET_CODE (funexp) != SYMBOL_REF) 271 funexp = memory_address (FUNCTION_MODE, funexp); 272 273 funmem = gen_rtx_MEM (FUNCTION_MODE, funexp); 274 if (fndecl && TREE_CODE (fndecl) == FUNCTION_DECL) 275 { 276 tree t = fndecl; 277 278 /* Although a built-in FUNCTION_DECL and its non-__builtin 279 counterpart compare equal and get a shared mem_attrs, they 280 produce different dump output in compare-debug compilations, 281 if an entry gets garbage collected in one compilation, then 282 adds a different (but equivalent) entry, while the other 283 doesn't run the garbage collector at the same spot and then 284 shares the mem_attr with the equivalent entry. */ 285 if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL) 286 { 287 tree t2 = builtin_decl_explicit (DECL_FUNCTION_CODE (t)); 288 if (t2) 289 t = t2; 290 } 291 292 set_mem_expr (funmem, t); 293 } 294 else if (fntree) 295 set_mem_expr (funmem, build_simple_mem_ref (CALL_EXPR_FN (fntree))); 296 297 #if defined (HAVE_sibcall_pop) && defined (HAVE_sibcall_value_pop) 298 if ((ecf_flags & ECF_SIBCALL) 299 && HAVE_sibcall_pop && HAVE_sibcall_value_pop 300 && (n_popped > 0 || stack_size == 0)) 301 { 302 rtx n_pop = GEN_INT (n_popped); 303 rtx pat; 304 305 /* If this subroutine pops its own args, record that in the call insn 306 if possible, for the sake of frame pointer elimination. */ 307 308 if (valreg) 309 pat = GEN_SIBCALL_VALUE_POP (valreg, funmem, rounded_stack_size_rtx, 310 next_arg_reg, n_pop); 311 else 312 pat = GEN_SIBCALL_POP (funmem, rounded_stack_size_rtx, next_arg_reg, 313 n_pop); 314 315 emit_call_insn (pat); 316 already_popped = 1; 317 } 318 else 319 #endif 320 321 #if defined (HAVE_call_pop) && defined (HAVE_call_value_pop) 322 /* If the target has "call" or "call_value" insns, then prefer them 323 if no arguments are actually popped. If the target does not have 324 "call" or "call_value" insns, then we must use the popping versions 325 even if the call has no arguments to pop. */ 326 #if defined (HAVE_call) && defined (HAVE_call_value) 327 if (HAVE_call && HAVE_call_value && HAVE_call_pop && HAVE_call_value_pop 328 && n_popped > 0) 329 #else 330 if (HAVE_call_pop && HAVE_call_value_pop) 331 #endif 332 { 333 rtx n_pop = GEN_INT (n_popped); 334 rtx pat; 335 336 /* If this subroutine pops its own args, record that in the call insn 337 if possible, for the sake of frame pointer elimination. */ 338 339 if (valreg) 340 pat = GEN_CALL_VALUE_POP (valreg, funmem, rounded_stack_size_rtx, 341 next_arg_reg, n_pop); 342 else 343 pat = GEN_CALL_POP (funmem, rounded_stack_size_rtx, next_arg_reg, 344 n_pop); 345 346 emit_call_insn (pat); 347 already_popped = 1; 348 } 349 else 350 #endif 351 352 #if defined (HAVE_sibcall) && defined (HAVE_sibcall_value) 353 if ((ecf_flags & ECF_SIBCALL) 354 && HAVE_sibcall && HAVE_sibcall_value) 355 { 356 if (valreg) 357 emit_call_insn (GEN_SIBCALL_VALUE (valreg, funmem, 358 rounded_stack_size_rtx, 359 next_arg_reg, NULL_RTX)); 360 else 361 emit_call_insn (GEN_SIBCALL (funmem, rounded_stack_size_rtx, 362 next_arg_reg, 363 GEN_INT (struct_value_size))); 364 } 365 else 366 #endif 367 368 #if defined (HAVE_call) && defined (HAVE_call_value) 369 if (HAVE_call && HAVE_call_value) 370 { 371 if (valreg) 372 emit_call_insn (GEN_CALL_VALUE (valreg, funmem, rounded_stack_size_rtx, 373 next_arg_reg, NULL_RTX)); 374 else 375 emit_call_insn (GEN_CALL (funmem, rounded_stack_size_rtx, next_arg_reg, 376 GEN_INT (struct_value_size))); 377 } 378 else 379 #endif 380 gcc_unreachable (); 381 382 /* Find the call we just emitted. */ 383 call_insn = last_call_insn (); 384 385 /* Some target create a fresh MEM instead of reusing the one provided 386 above. Set its MEM_EXPR. */ 387 call = PATTERN (call_insn); 388 if (GET_CODE (call) == PARALLEL) 389 call = XVECEXP (call, 0, 0); 390 if (GET_CODE (call) == SET) 391 call = SET_SRC (call); 392 if (GET_CODE (call) == CALL 393 && MEM_P (XEXP (call, 0)) 394 && MEM_EXPR (XEXP (call, 0)) == NULL_TREE 395 && MEM_EXPR (funmem) != NULL_TREE) 396 set_mem_expr (XEXP (call, 0), MEM_EXPR (funmem)); 397 398 /* Put the register usage information there. */ 399 add_function_usage_to (call_insn, call_fusage); 400 401 /* If this is a const call, then set the insn's unchanging bit. */ 402 if (ecf_flags & ECF_CONST) 403 RTL_CONST_CALL_P (call_insn) = 1; 404 405 /* If this is a pure call, then set the insn's unchanging bit. */ 406 if (ecf_flags & ECF_PURE) 407 RTL_PURE_CALL_P (call_insn) = 1; 408 409 /* If this is a const call, then set the insn's unchanging bit. */ 410 if (ecf_flags & ECF_LOOPING_CONST_OR_PURE) 411 RTL_LOOPING_CONST_OR_PURE_CALL_P (call_insn) = 1; 412 413 /* Create a nothrow REG_EH_REGION note, if needed. */ 414 make_reg_eh_region_note (call_insn, ecf_flags, 0); 415 416 if (ecf_flags & ECF_NORETURN) 417 add_reg_note (call_insn, REG_NORETURN, const0_rtx); 418 419 if (ecf_flags & ECF_RETURNS_TWICE) 420 { 421 add_reg_note (call_insn, REG_SETJMP, const0_rtx); 422 cfun->calls_setjmp = 1; 423 } 424 425 SIBLING_CALL_P (call_insn) = ((ecf_flags & ECF_SIBCALL) != 0); 426 427 /* Restore this now, so that we do defer pops for this call's args 428 if the context of the call as a whole permits. */ 429 inhibit_defer_pop = old_inhibit_defer_pop; 430 431 if (n_popped > 0) 432 { 433 if (!already_popped) 434 CALL_INSN_FUNCTION_USAGE (call_insn) 435 = gen_rtx_EXPR_LIST (VOIDmode, 436 gen_rtx_CLOBBER (VOIDmode, stack_pointer_rtx), 437 CALL_INSN_FUNCTION_USAGE (call_insn)); 438 rounded_stack_size -= n_popped; 439 rounded_stack_size_rtx = GEN_INT (rounded_stack_size); 440 stack_pointer_delta -= n_popped; 441 442 add_reg_note (call_insn, REG_ARGS_SIZE, GEN_INT (stack_pointer_delta)); 443 444 /* If popup is needed, stack realign must use DRAP */ 445 if (SUPPORTS_STACK_ALIGNMENT) 446 crtl->need_drap = true; 447 } 448 /* For noreturn calls when not accumulating outgoing args force 449 REG_ARGS_SIZE note to prevent crossjumping of calls with different 450 args sizes. */ 451 else if (!ACCUMULATE_OUTGOING_ARGS && (ecf_flags & ECF_NORETURN) != 0) 452 add_reg_note (call_insn, REG_ARGS_SIZE, GEN_INT (stack_pointer_delta)); 453 454 if (!ACCUMULATE_OUTGOING_ARGS) 455 { 456 /* If returning from the subroutine does not automatically pop the args, 457 we need an instruction to pop them sooner or later. 458 Perhaps do it now; perhaps just record how much space to pop later. 459 460 If returning from the subroutine does pop the args, indicate that the 461 stack pointer will be changed. */ 462 463 if (rounded_stack_size != 0) 464 { 465 if (ecf_flags & ECF_NORETURN) 466 /* Just pretend we did the pop. */ 467 stack_pointer_delta -= rounded_stack_size; 468 else if (flag_defer_pop && inhibit_defer_pop == 0 469 && ! (ecf_flags & (ECF_CONST | ECF_PURE))) 470 pending_stack_adjust += rounded_stack_size; 471 else 472 adjust_stack (rounded_stack_size_rtx); 473 } 474 } 475 /* When we accumulate outgoing args, we must avoid any stack manipulations. 476 Restore the stack pointer to its original value now. Usually 477 ACCUMULATE_OUTGOING_ARGS targets don't get here, but there are exceptions. 478 On i386 ACCUMULATE_OUTGOING_ARGS can be enabled on demand, and 479 popping variants of functions exist as well. 480 481 ??? We may optimize similar to defer_pop above, but it is 482 probably not worthwhile. 483 484 ??? It will be worthwhile to enable combine_stack_adjustments even for 485 such machines. */ 486 else if (n_popped) 487 anti_adjust_stack (GEN_INT (n_popped)); 488 } 489 490 /* Determine if the function identified by NAME and FNDECL is one with 491 special properties we wish to know about. 492 493 For example, if the function might return more than one time (setjmp), then 494 set RETURNS_TWICE to a nonzero value. 495 496 Similarly set NORETURN if the function is in the longjmp family. 497 498 Set MAY_BE_ALLOCA for any memory allocation function that might allocate 499 space from the stack such as alloca. */ 500 501 static int 502 special_function_p (const_tree fndecl, int flags) 503 { 504 if (fndecl && DECL_NAME (fndecl) 505 && IDENTIFIER_LENGTH (DECL_NAME (fndecl)) <= 17 506 /* Exclude functions not at the file scope, or not `extern', 507 since they are not the magic functions we would otherwise 508 think they are. 509 FIXME: this should be handled with attributes, not with this 510 hacky imitation of DECL_ASSEMBLER_NAME. It's (also) wrong 511 because you can declare fork() inside a function if you 512 wish. */ 513 && (DECL_CONTEXT (fndecl) == NULL_TREE 514 || TREE_CODE (DECL_CONTEXT (fndecl)) == TRANSLATION_UNIT_DECL) 515 && TREE_PUBLIC (fndecl)) 516 { 517 const char *name = IDENTIFIER_POINTER (DECL_NAME (fndecl)); 518 const char *tname = name; 519 520 /* We assume that alloca will always be called by name. It 521 makes no sense to pass it as a pointer-to-function to 522 anything that does not understand its behavior. */ 523 if (((IDENTIFIER_LENGTH (DECL_NAME (fndecl)) == 6 524 && name[0] == 'a' 525 && ! strcmp (name, "alloca")) 526 || (IDENTIFIER_LENGTH (DECL_NAME (fndecl)) == 16 527 && name[0] == '_' 528 && ! strcmp (name, "__builtin_alloca")))) 529 flags |= ECF_MAY_BE_ALLOCA; 530 531 /* Disregard prefix _, __, __x or __builtin_. */ 532 if (name[0] == '_') 533 { 534 if (name[1] == '_' 535 && name[2] == 'b' 536 && !strncmp (name + 3, "uiltin_", 7)) 537 tname += 10; 538 else if (name[1] == '_' && name[2] == 'x') 539 tname += 3; 540 else if (name[1] == '_') 541 tname += 2; 542 else 543 tname += 1; 544 } 545 546 if (tname[0] == 's') 547 { 548 if ((tname[1] == 'e' 549 && (! strcmp (tname, "setjmp") 550 || ! strcmp (tname, "setjmp_syscall"))) 551 || (tname[1] == 'i' 552 && ! strcmp (tname, "sigsetjmp")) 553 || (tname[1] == 'a' 554 && ! strcmp (tname, "savectx"))) 555 flags |= ECF_RETURNS_TWICE; 556 557 if (tname[1] == 'i' 558 && ! strcmp (tname, "siglongjmp")) 559 flags |= ECF_NORETURN; 560 } 561 else if ((tname[0] == 'q' && tname[1] == 's' 562 && ! strcmp (tname, "qsetjmp")) 563 || (tname[0] == 'v' && tname[1] == 'f' 564 && ! strcmp (tname, "vfork")) 565 || (tname[0] == 'g' && tname[1] == 'e' 566 && !strcmp (tname, "getcontext"))) 567 flags |= ECF_RETURNS_TWICE; 568 569 else if (tname[0] == 'l' && tname[1] == 'o' 570 && ! strcmp (tname, "longjmp")) 571 flags |= ECF_NORETURN; 572 } 573 574 return flags; 575 } 576 577 /* Return nonzero when FNDECL represents a call to setjmp. */ 578 579 int 580 setjmp_call_p (const_tree fndecl) 581 { 582 if (DECL_IS_RETURNS_TWICE (fndecl)) 583 return ECF_RETURNS_TWICE; 584 return special_function_p (fndecl, 0) & ECF_RETURNS_TWICE; 585 } 586 587 588 /* Return true if STMT is an alloca call. */ 589 590 bool 591 gimple_alloca_call_p (const_gimple stmt) 592 { 593 tree fndecl; 594 595 if (!is_gimple_call (stmt)) 596 return false; 597 598 fndecl = gimple_call_fndecl (stmt); 599 if (fndecl && (special_function_p (fndecl, 0) & ECF_MAY_BE_ALLOCA)) 600 return true; 601 602 return false; 603 } 604 605 /* Return true when exp contains alloca call. */ 606 607 bool 608 alloca_call_p (const_tree exp) 609 { 610 if (TREE_CODE (exp) == CALL_EXPR 611 && TREE_CODE (CALL_EXPR_FN (exp)) == ADDR_EXPR 612 && (TREE_CODE (TREE_OPERAND (CALL_EXPR_FN (exp), 0)) == FUNCTION_DECL) 613 && (special_function_p (TREE_OPERAND (CALL_EXPR_FN (exp), 0), 0) 614 & ECF_MAY_BE_ALLOCA)) 615 return true; 616 return false; 617 } 618 619 /* Return TRUE if FNDECL is either a TM builtin or a TM cloned 620 function. Return FALSE otherwise. */ 621 622 static bool 623 is_tm_builtin (const_tree fndecl) 624 { 625 if (fndecl == NULL) 626 return false; 627 628 if (decl_is_tm_clone (fndecl)) 629 return true; 630 631 if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) 632 { 633 switch (DECL_FUNCTION_CODE (fndecl)) 634 { 635 case BUILT_IN_TM_COMMIT: 636 case BUILT_IN_TM_COMMIT_EH: 637 case BUILT_IN_TM_ABORT: 638 case BUILT_IN_TM_IRREVOCABLE: 639 case BUILT_IN_TM_GETTMCLONE_IRR: 640 case BUILT_IN_TM_MEMCPY: 641 case BUILT_IN_TM_MEMMOVE: 642 case BUILT_IN_TM_MEMSET: 643 CASE_BUILT_IN_TM_STORE (1): 644 CASE_BUILT_IN_TM_STORE (2): 645 CASE_BUILT_IN_TM_STORE (4): 646 CASE_BUILT_IN_TM_STORE (8): 647 CASE_BUILT_IN_TM_STORE (FLOAT): 648 CASE_BUILT_IN_TM_STORE (DOUBLE): 649 CASE_BUILT_IN_TM_STORE (LDOUBLE): 650 CASE_BUILT_IN_TM_STORE (M64): 651 CASE_BUILT_IN_TM_STORE (M128): 652 CASE_BUILT_IN_TM_STORE (M256): 653 CASE_BUILT_IN_TM_LOAD (1): 654 CASE_BUILT_IN_TM_LOAD (2): 655 CASE_BUILT_IN_TM_LOAD (4): 656 CASE_BUILT_IN_TM_LOAD (8): 657 CASE_BUILT_IN_TM_LOAD (FLOAT): 658 CASE_BUILT_IN_TM_LOAD (DOUBLE): 659 CASE_BUILT_IN_TM_LOAD (LDOUBLE): 660 CASE_BUILT_IN_TM_LOAD (M64): 661 CASE_BUILT_IN_TM_LOAD (M128): 662 CASE_BUILT_IN_TM_LOAD (M256): 663 case BUILT_IN_TM_LOG: 664 case BUILT_IN_TM_LOG_1: 665 case BUILT_IN_TM_LOG_2: 666 case BUILT_IN_TM_LOG_4: 667 case BUILT_IN_TM_LOG_8: 668 case BUILT_IN_TM_LOG_FLOAT: 669 case BUILT_IN_TM_LOG_DOUBLE: 670 case BUILT_IN_TM_LOG_LDOUBLE: 671 case BUILT_IN_TM_LOG_M64: 672 case BUILT_IN_TM_LOG_M128: 673 case BUILT_IN_TM_LOG_M256: 674 return true; 675 default: 676 break; 677 } 678 } 679 return false; 680 } 681 682 /* Detect flags (function attributes) from the function decl or type node. */ 683 684 int 685 flags_from_decl_or_type (const_tree exp) 686 { 687 int flags = 0; 688 689 if (DECL_P (exp)) 690 { 691 /* The function exp may have the `malloc' attribute. */ 692 if (DECL_IS_MALLOC (exp)) 693 flags |= ECF_MALLOC; 694 695 /* The function exp may have the `returns_twice' attribute. */ 696 if (DECL_IS_RETURNS_TWICE (exp)) 697 flags |= ECF_RETURNS_TWICE; 698 699 /* Process the pure and const attributes. */ 700 if (TREE_READONLY (exp)) 701 flags |= ECF_CONST; 702 if (DECL_PURE_P (exp)) 703 flags |= ECF_PURE; 704 if (DECL_LOOPING_CONST_OR_PURE_P (exp)) 705 flags |= ECF_LOOPING_CONST_OR_PURE; 706 707 if (DECL_IS_NOVOPS (exp)) 708 flags |= ECF_NOVOPS; 709 if (lookup_attribute ("leaf", DECL_ATTRIBUTES (exp))) 710 flags |= ECF_LEAF; 711 712 if (TREE_NOTHROW (exp)) 713 flags |= ECF_NOTHROW; 714 715 if (flag_tm) 716 { 717 if (is_tm_builtin (exp)) 718 flags |= ECF_TM_BUILTIN; 719 else if ((flags & (ECF_CONST|ECF_NOVOPS)) != 0 720 || lookup_attribute ("transaction_pure", 721 TYPE_ATTRIBUTES (TREE_TYPE (exp)))) 722 flags |= ECF_TM_PURE; 723 } 724 725 flags = special_function_p (exp, flags); 726 } 727 else if (TYPE_P (exp)) 728 { 729 if (TYPE_READONLY (exp)) 730 flags |= ECF_CONST; 731 732 if (flag_tm 733 && ((flags & ECF_CONST) != 0 734 || lookup_attribute ("transaction_pure", TYPE_ATTRIBUTES (exp)))) 735 flags |= ECF_TM_PURE; 736 } 737 738 if (TREE_THIS_VOLATILE (exp)) 739 { 740 flags |= ECF_NORETURN; 741 if (flags & (ECF_CONST|ECF_PURE)) 742 flags |= ECF_LOOPING_CONST_OR_PURE; 743 } 744 745 return flags; 746 } 747 748 /* Detect flags from a CALL_EXPR. */ 749 750 int 751 call_expr_flags (const_tree t) 752 { 753 int flags; 754 tree decl = get_callee_fndecl (t); 755 756 if (decl) 757 flags = flags_from_decl_or_type (decl); 758 else 759 { 760 t = TREE_TYPE (CALL_EXPR_FN (t)); 761 if (t && TREE_CODE (t) == POINTER_TYPE) 762 flags = flags_from_decl_or_type (TREE_TYPE (t)); 763 else 764 flags = 0; 765 } 766 767 return flags; 768 } 769 770 /* Precompute all register parameters as described by ARGS, storing values 771 into fields within the ARGS array. 772 773 NUM_ACTUALS indicates the total number elements in the ARGS array. 774 775 Set REG_PARM_SEEN if we encounter a register parameter. */ 776 777 static void 778 precompute_register_parameters (int num_actuals, struct arg_data *args, 779 int *reg_parm_seen) 780 { 781 int i; 782 783 *reg_parm_seen = 0; 784 785 for (i = 0; i < num_actuals; i++) 786 if (args[i].reg != 0 && ! args[i].pass_on_stack) 787 { 788 *reg_parm_seen = 1; 789 790 if (args[i].value == 0) 791 { 792 push_temp_slots (); 793 args[i].value = expand_normal (args[i].tree_value); 794 preserve_temp_slots (args[i].value); 795 pop_temp_slots (); 796 } 797 798 /* If we are to promote the function arg to a wider mode, 799 do it now. */ 800 801 if (args[i].mode != TYPE_MODE (TREE_TYPE (args[i].tree_value))) 802 args[i].value 803 = convert_modes (args[i].mode, 804 TYPE_MODE (TREE_TYPE (args[i].tree_value)), 805 args[i].value, args[i].unsignedp); 806 807 /* If the value is a non-legitimate constant, force it into a 808 pseudo now. TLS symbols sometimes need a call to resolve. */ 809 if (CONSTANT_P (args[i].value) 810 && !targetm.legitimate_constant_p (args[i].mode, args[i].value)) 811 args[i].value = force_reg (args[i].mode, args[i].value); 812 813 /* If we're going to have to load the value by parts, pull the 814 parts into pseudos. The part extraction process can involve 815 non-trivial computation. */ 816 if (GET_CODE (args[i].reg) == PARALLEL) 817 { 818 tree type = TREE_TYPE (args[i].tree_value); 819 args[i].parallel_value 820 = emit_group_load_into_temps (args[i].reg, args[i].value, 821 type, int_size_in_bytes (type)); 822 } 823 824 /* If the value is expensive, and we are inside an appropriately 825 short loop, put the value into a pseudo and then put the pseudo 826 into the hard reg. 827 828 For small register classes, also do this if this call uses 829 register parameters. This is to avoid reload conflicts while 830 loading the parameters registers. */ 831 832 else if ((! (REG_P (args[i].value) 833 || (GET_CODE (args[i].value) == SUBREG 834 && REG_P (SUBREG_REG (args[i].value))))) 835 && args[i].mode != BLKmode 836 && set_src_cost (args[i].value, optimize_insn_for_speed_p ()) 837 > COSTS_N_INSNS (1) 838 && ((*reg_parm_seen 839 && targetm.small_register_classes_for_mode_p (args[i].mode)) 840 || optimize)) 841 args[i].value = copy_to_mode_reg (args[i].mode, args[i].value); 842 } 843 } 844 845 #ifdef REG_PARM_STACK_SPACE 846 847 /* The argument list is the property of the called routine and it 848 may clobber it. If the fixed area has been used for previous 849 parameters, we must save and restore it. */ 850 851 static rtx 852 save_fixed_argument_area (int reg_parm_stack_space, rtx argblock, int *low_to_save, int *high_to_save) 853 { 854 int low; 855 int high; 856 857 /* Compute the boundary of the area that needs to be saved, if any. */ 858 high = reg_parm_stack_space; 859 #ifdef ARGS_GROW_DOWNWARD 860 high += 1; 861 #endif 862 if (high > highest_outgoing_arg_in_use) 863 high = highest_outgoing_arg_in_use; 864 865 for (low = 0; low < high; low++) 866 if (stack_usage_map[low] != 0) 867 { 868 int num_to_save; 869 enum machine_mode save_mode; 870 int delta; 871 rtx stack_area; 872 rtx save_area; 873 874 while (stack_usage_map[--high] == 0) 875 ; 876 877 *low_to_save = low; 878 *high_to_save = high; 879 880 num_to_save = high - low + 1; 881 save_mode = mode_for_size (num_to_save * BITS_PER_UNIT, MODE_INT, 1); 882 883 /* If we don't have the required alignment, must do this 884 in BLKmode. */ 885 if ((low & (MIN (GET_MODE_SIZE (save_mode), 886 BIGGEST_ALIGNMENT / UNITS_PER_WORD) - 1))) 887 save_mode = BLKmode; 888 889 #ifdef ARGS_GROW_DOWNWARD 890 delta = -high; 891 #else 892 delta = low; 893 #endif 894 stack_area = gen_rtx_MEM (save_mode, 895 memory_address (save_mode, 896 plus_constant (argblock, 897 delta))); 898 899 set_mem_align (stack_area, PARM_BOUNDARY); 900 if (save_mode == BLKmode) 901 { 902 save_area = assign_stack_temp (BLKmode, num_to_save, 0); 903 emit_block_move (validize_mem (save_area), stack_area, 904 GEN_INT (num_to_save), BLOCK_OP_CALL_PARM); 905 } 906 else 907 { 908 save_area = gen_reg_rtx (save_mode); 909 emit_move_insn (save_area, stack_area); 910 } 911 912 return save_area; 913 } 914 915 return NULL_RTX; 916 } 917 918 static void 919 restore_fixed_argument_area (rtx save_area, rtx argblock, int high_to_save, int low_to_save) 920 { 921 enum machine_mode save_mode = GET_MODE (save_area); 922 int delta; 923 rtx stack_area; 924 925 #ifdef ARGS_GROW_DOWNWARD 926 delta = -high_to_save; 927 #else 928 delta = low_to_save; 929 #endif 930 stack_area = gen_rtx_MEM (save_mode, 931 memory_address (save_mode, 932 plus_constant (argblock, delta))); 933 set_mem_align (stack_area, PARM_BOUNDARY); 934 935 if (save_mode != BLKmode) 936 emit_move_insn (stack_area, save_area); 937 else 938 emit_block_move (stack_area, validize_mem (save_area), 939 GEN_INT (high_to_save - low_to_save + 1), 940 BLOCK_OP_CALL_PARM); 941 } 942 #endif /* REG_PARM_STACK_SPACE */ 943 944 /* If any elements in ARGS refer to parameters that are to be passed in 945 registers, but not in memory, and whose alignment does not permit a 946 direct copy into registers. Copy the values into a group of pseudos 947 which we will later copy into the appropriate hard registers. 948 949 Pseudos for each unaligned argument will be stored into the array 950 args[argnum].aligned_regs. The caller is responsible for deallocating 951 the aligned_regs array if it is nonzero. */ 952 953 static void 954 store_unaligned_arguments_into_pseudos (struct arg_data *args, int num_actuals) 955 { 956 int i, j; 957 958 for (i = 0; i < num_actuals; i++) 959 if (args[i].reg != 0 && ! args[i].pass_on_stack 960 && args[i].mode == BLKmode 961 && MEM_P (args[i].value) 962 && (MEM_ALIGN (args[i].value) 963 < (unsigned int) MIN (BIGGEST_ALIGNMENT, BITS_PER_WORD))) 964 { 965 int bytes = int_size_in_bytes (TREE_TYPE (args[i].tree_value)); 966 int endian_correction = 0; 967 968 if (args[i].partial) 969 { 970 gcc_assert (args[i].partial % UNITS_PER_WORD == 0); 971 args[i].n_aligned_regs = args[i].partial / UNITS_PER_WORD; 972 } 973 else 974 { 975 args[i].n_aligned_regs 976 = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD; 977 } 978 979 args[i].aligned_regs = XNEWVEC (rtx, args[i].n_aligned_regs); 980 981 /* Structures smaller than a word are normally aligned to the 982 least significant byte. On a BYTES_BIG_ENDIAN machine, 983 this means we must skip the empty high order bytes when 984 calculating the bit offset. */ 985 if (bytes < UNITS_PER_WORD 986 #ifdef BLOCK_REG_PADDING 987 && (BLOCK_REG_PADDING (args[i].mode, 988 TREE_TYPE (args[i].tree_value), 1) 989 == downward) 990 #else 991 && BYTES_BIG_ENDIAN 992 #endif 993 ) 994 endian_correction = BITS_PER_WORD - bytes * BITS_PER_UNIT; 995 996 for (j = 0; j < args[i].n_aligned_regs; j++) 997 { 998 rtx reg = gen_reg_rtx (word_mode); 999 rtx word = operand_subword_force (args[i].value, j, BLKmode); 1000 int bitsize = MIN (bytes * BITS_PER_UNIT, BITS_PER_WORD); 1001 1002 args[i].aligned_regs[j] = reg; 1003 word = extract_bit_field (word, bitsize, 0, 1, false, NULL_RTX, 1004 word_mode, word_mode); 1005 1006 /* There is no need to restrict this code to loading items 1007 in TYPE_ALIGN sized hunks. The bitfield instructions can 1008 load up entire word sized registers efficiently. 1009 1010 ??? This may not be needed anymore. 1011 We use to emit a clobber here but that doesn't let later 1012 passes optimize the instructions we emit. By storing 0 into 1013 the register later passes know the first AND to zero out the 1014 bitfield being set in the register is unnecessary. The store 1015 of 0 will be deleted as will at least the first AND. */ 1016 1017 emit_move_insn (reg, const0_rtx); 1018 1019 bytes -= bitsize / BITS_PER_UNIT; 1020 store_bit_field (reg, bitsize, endian_correction, 0, 0, 1021 word_mode, word); 1022 } 1023 } 1024 } 1025 1026 /* Fill in ARGS_SIZE and ARGS array based on the parameters found in 1027 CALL_EXPR EXP. 1028 1029 NUM_ACTUALS is the total number of parameters. 1030 1031 N_NAMED_ARGS is the total number of named arguments. 1032 1033 STRUCT_VALUE_ADDR_VALUE is the implicit argument for a struct return 1034 value, or null. 1035 1036 FNDECL is the tree code for the target of this call (if known) 1037 1038 ARGS_SO_FAR holds state needed by the target to know where to place 1039 the next argument. 1040 1041 REG_PARM_STACK_SPACE is the number of bytes of stack space reserved 1042 for arguments which are passed in registers. 1043 1044 OLD_STACK_LEVEL is a pointer to an rtx which olds the old stack level 1045 and may be modified by this routine. 1046 1047 OLD_PENDING_ADJ, MUST_PREALLOCATE and FLAGS are pointers to integer 1048 flags which may may be modified by this routine. 1049 1050 MAY_TAILCALL is cleared if we encounter an invisible pass-by-reference 1051 that requires allocation of stack space. 1052 1053 CALL_FROM_THUNK_P is true if this call is the jump from a thunk to 1054 the thunked-to function. */ 1055 1056 static void 1057 initialize_argument_information (int num_actuals ATTRIBUTE_UNUSED, 1058 struct arg_data *args, 1059 struct args_size *args_size, 1060 int n_named_args ATTRIBUTE_UNUSED, 1061 tree exp, tree struct_value_addr_value, 1062 tree fndecl, tree fntype, 1063 cumulative_args_t args_so_far, 1064 int reg_parm_stack_space, 1065 rtx *old_stack_level, int *old_pending_adj, 1066 int *must_preallocate, int *ecf_flags, 1067 bool *may_tailcall, bool call_from_thunk_p) 1068 { 1069 CUMULATIVE_ARGS *args_so_far_pnt = get_cumulative_args (args_so_far); 1070 location_t loc = EXPR_LOCATION (exp); 1071 /* 1 if scanning parms front to back, -1 if scanning back to front. */ 1072 int inc; 1073 1074 /* Count arg position in order args appear. */ 1075 int argpos; 1076 1077 int i; 1078 1079 args_size->constant = 0; 1080 args_size->var = 0; 1081 1082 /* In this loop, we consider args in the order they are written. 1083 We fill up ARGS from the front or from the back if necessary 1084 so that in any case the first arg to be pushed ends up at the front. */ 1085 1086 if (PUSH_ARGS_REVERSED) 1087 { 1088 i = num_actuals - 1, inc = -1; 1089 /* In this case, must reverse order of args 1090 so that we compute and push the last arg first. */ 1091 } 1092 else 1093 { 1094 i = 0, inc = 1; 1095 } 1096 1097 /* First fill in the actual arguments in the ARGS array, splitting 1098 complex arguments if necessary. */ 1099 { 1100 int j = i; 1101 call_expr_arg_iterator iter; 1102 tree arg; 1103 1104 if (struct_value_addr_value) 1105 { 1106 args[j].tree_value = struct_value_addr_value; 1107 j += inc; 1108 } 1109 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp) 1110 { 1111 tree argtype = TREE_TYPE (arg); 1112 if (targetm.calls.split_complex_arg 1113 && argtype 1114 && TREE_CODE (argtype) == COMPLEX_TYPE 1115 && targetm.calls.split_complex_arg (argtype)) 1116 { 1117 tree subtype = TREE_TYPE (argtype); 1118 args[j].tree_value = build1 (REALPART_EXPR, subtype, arg); 1119 j += inc; 1120 args[j].tree_value = build1 (IMAGPART_EXPR, subtype, arg); 1121 } 1122 else 1123 args[j].tree_value = arg; 1124 j += inc; 1125 } 1126 } 1127 1128 /* I counts args in order (to be) pushed; ARGPOS counts in order written. */ 1129 for (argpos = 0; argpos < num_actuals; i += inc, argpos++) 1130 { 1131 tree type = TREE_TYPE (args[i].tree_value); 1132 int unsignedp; 1133 enum machine_mode mode; 1134 1135 /* Replace erroneous argument with constant zero. */ 1136 if (type == error_mark_node || !COMPLETE_TYPE_P (type)) 1137 args[i].tree_value = integer_zero_node, type = integer_type_node; 1138 1139 /* If TYPE is a transparent union or record, pass things the way 1140 we would pass the first field of the union or record. We have 1141 already verified that the modes are the same. */ 1142 if ((TREE_CODE (type) == UNION_TYPE || TREE_CODE (type) == RECORD_TYPE) 1143 && TYPE_TRANSPARENT_AGGR (type)) 1144 type = TREE_TYPE (first_field (type)); 1145 1146 /* Decide where to pass this arg. 1147 1148 args[i].reg is nonzero if all or part is passed in registers. 1149 1150 args[i].partial is nonzero if part but not all is passed in registers, 1151 and the exact value says how many bytes are passed in registers. 1152 1153 args[i].pass_on_stack is nonzero if the argument must at least be 1154 computed on the stack. It may then be loaded back into registers 1155 if args[i].reg is nonzero. 1156 1157 These decisions are driven by the FUNCTION_... macros and must agree 1158 with those made by function.c. */ 1159 1160 /* See if this argument should be passed by invisible reference. */ 1161 if (pass_by_reference (args_so_far_pnt, TYPE_MODE (type), 1162 type, argpos < n_named_args)) 1163 { 1164 bool callee_copies; 1165 tree base = NULL_TREE; 1166 1167 callee_copies 1168 = reference_callee_copied (args_so_far_pnt, TYPE_MODE (type), 1169 type, argpos < n_named_args); 1170 1171 /* If we're compiling a thunk, pass through invisible references 1172 instead of making a copy. */ 1173 if (call_from_thunk_p 1174 || (callee_copies 1175 && !TREE_ADDRESSABLE (type) 1176 && (base = get_base_address (args[i].tree_value)) 1177 && TREE_CODE (base) != SSA_NAME 1178 && (!DECL_P (base) || MEM_P (DECL_RTL (base))))) 1179 { 1180 mark_addressable (args[i].tree_value); 1181 1182 /* We can't use sibcalls if a callee-copied argument is 1183 stored in the current function's frame. */ 1184 if (!call_from_thunk_p && DECL_P (base) && !TREE_STATIC (base)) 1185 *may_tailcall = false; 1186 1187 args[i].tree_value = build_fold_addr_expr_loc (loc, 1188 args[i].tree_value); 1189 type = TREE_TYPE (args[i].tree_value); 1190 1191 if (*ecf_flags & ECF_CONST) 1192 *ecf_flags &= ~(ECF_CONST | ECF_LOOPING_CONST_OR_PURE); 1193 } 1194 else 1195 { 1196 /* We make a copy of the object and pass the address to the 1197 function being called. */ 1198 rtx copy; 1199 1200 if (!COMPLETE_TYPE_P (type) 1201 || TREE_CODE (TYPE_SIZE_UNIT (type)) != INTEGER_CST 1202 || (flag_stack_check == GENERIC_STACK_CHECK 1203 && compare_tree_int (TYPE_SIZE_UNIT (type), 1204 STACK_CHECK_MAX_VAR_SIZE) > 0)) 1205 { 1206 /* This is a variable-sized object. Make space on the stack 1207 for it. */ 1208 rtx size_rtx = expr_size (args[i].tree_value); 1209 1210 if (*old_stack_level == 0) 1211 { 1212 emit_stack_save (SAVE_BLOCK, old_stack_level); 1213 *old_pending_adj = pending_stack_adjust; 1214 pending_stack_adjust = 0; 1215 } 1216 1217 /* We can pass TRUE as the 4th argument because we just 1218 saved the stack pointer and will restore it right after 1219 the call. */ 1220 copy = allocate_dynamic_stack_space (size_rtx, 1221 TYPE_ALIGN (type), 1222 TYPE_ALIGN (type), 1223 true); 1224 copy = gen_rtx_MEM (BLKmode, copy); 1225 set_mem_attributes (copy, type, 1); 1226 } 1227 else 1228 copy = assign_temp (type, 0, 1, 0); 1229 1230 store_expr (args[i].tree_value, copy, 0, false); 1231 1232 /* Just change the const function to pure and then let 1233 the next test clear the pure based on 1234 callee_copies. */ 1235 if (*ecf_flags & ECF_CONST) 1236 { 1237 *ecf_flags &= ~ECF_CONST; 1238 *ecf_flags |= ECF_PURE; 1239 } 1240 1241 if (!callee_copies && *ecf_flags & ECF_PURE) 1242 *ecf_flags &= ~(ECF_PURE | ECF_LOOPING_CONST_OR_PURE); 1243 1244 args[i].tree_value 1245 = build_fold_addr_expr_loc (loc, make_tree (type, copy)); 1246 type = TREE_TYPE (args[i].tree_value); 1247 *may_tailcall = false; 1248 } 1249 } 1250 1251 unsignedp = TYPE_UNSIGNED (type); 1252 mode = promote_function_mode (type, TYPE_MODE (type), &unsignedp, 1253 fndecl ? TREE_TYPE (fndecl) : fntype, 0); 1254 1255 args[i].unsignedp = unsignedp; 1256 args[i].mode = mode; 1257 1258 args[i].reg = targetm.calls.function_arg (args_so_far, mode, type, 1259 argpos < n_named_args); 1260 1261 /* If this is a sibling call and the machine has register windows, the 1262 register window has to be unwinded before calling the routine, so 1263 arguments have to go into the incoming registers. */ 1264 if (targetm.calls.function_incoming_arg != targetm.calls.function_arg) 1265 args[i].tail_call_reg 1266 = targetm.calls.function_incoming_arg (args_so_far, mode, type, 1267 argpos < n_named_args); 1268 else 1269 args[i].tail_call_reg = args[i].reg; 1270 1271 if (args[i].reg) 1272 args[i].partial 1273 = targetm.calls.arg_partial_bytes (args_so_far, mode, type, 1274 argpos < n_named_args); 1275 1276 args[i].pass_on_stack = targetm.calls.must_pass_in_stack (mode, type); 1277 1278 /* If FUNCTION_ARG returned a (parallel [(expr_list (nil) ...) ...]), 1279 it means that we are to pass this arg in the register(s) designated 1280 by the PARALLEL, but also to pass it in the stack. */ 1281 if (args[i].reg && GET_CODE (args[i].reg) == PARALLEL 1282 && XEXP (XVECEXP (args[i].reg, 0, 0), 0) == 0) 1283 args[i].pass_on_stack = 1; 1284 1285 /* If this is an addressable type, we must preallocate the stack 1286 since we must evaluate the object into its final location. 1287 1288 If this is to be passed in both registers and the stack, it is simpler 1289 to preallocate. */ 1290 if (TREE_ADDRESSABLE (type) 1291 || (args[i].pass_on_stack && args[i].reg != 0)) 1292 *must_preallocate = 1; 1293 1294 /* Compute the stack-size of this argument. */ 1295 if (args[i].reg == 0 || args[i].partial != 0 1296 || reg_parm_stack_space > 0 1297 || args[i].pass_on_stack) 1298 locate_and_pad_parm (mode, type, 1299 #ifdef STACK_PARMS_IN_REG_PARM_AREA 1300 1, 1301 #else 1302 args[i].reg != 0, 1303 #endif 1304 args[i].pass_on_stack ? 0 : args[i].partial, 1305 fndecl, args_size, &args[i].locate); 1306 #ifdef BLOCK_REG_PADDING 1307 else 1308 /* The argument is passed entirely in registers. See at which 1309 end it should be padded. */ 1310 args[i].locate.where_pad = 1311 BLOCK_REG_PADDING (mode, type, 1312 int_size_in_bytes (type) <= UNITS_PER_WORD); 1313 #endif 1314 1315 /* Update ARGS_SIZE, the total stack space for args so far. */ 1316 1317 args_size->constant += args[i].locate.size.constant; 1318 if (args[i].locate.size.var) 1319 ADD_PARM_SIZE (*args_size, args[i].locate.size.var); 1320 1321 /* Increment ARGS_SO_FAR, which has info about which arg-registers 1322 have been used, etc. */ 1323 1324 targetm.calls.function_arg_advance (args_so_far, TYPE_MODE (type), 1325 type, argpos < n_named_args); 1326 } 1327 } 1328 1329 /* Update ARGS_SIZE to contain the total size for the argument block. 1330 Return the original constant component of the argument block's size. 1331 1332 REG_PARM_STACK_SPACE holds the number of bytes of stack space reserved 1333 for arguments passed in registers. */ 1334 1335 static int 1336 compute_argument_block_size (int reg_parm_stack_space, 1337 struct args_size *args_size, 1338 tree fndecl ATTRIBUTE_UNUSED, 1339 tree fntype ATTRIBUTE_UNUSED, 1340 int preferred_stack_boundary ATTRIBUTE_UNUSED) 1341 { 1342 int unadjusted_args_size = args_size->constant; 1343 1344 /* For accumulate outgoing args mode we don't need to align, since the frame 1345 will be already aligned. Align to STACK_BOUNDARY in order to prevent 1346 backends from generating misaligned frame sizes. */ 1347 if (ACCUMULATE_OUTGOING_ARGS && preferred_stack_boundary > STACK_BOUNDARY) 1348 preferred_stack_boundary = STACK_BOUNDARY; 1349 1350 /* Compute the actual size of the argument block required. The variable 1351 and constant sizes must be combined, the size may have to be rounded, 1352 and there may be a minimum required size. */ 1353 1354 if (args_size->var) 1355 { 1356 args_size->var = ARGS_SIZE_TREE (*args_size); 1357 args_size->constant = 0; 1358 1359 preferred_stack_boundary /= BITS_PER_UNIT; 1360 if (preferred_stack_boundary > 1) 1361 { 1362 /* We don't handle this case yet. To handle it correctly we have 1363 to add the delta, round and subtract the delta. 1364 Currently no machine description requires this support. */ 1365 gcc_assert (!(stack_pointer_delta & (preferred_stack_boundary - 1))); 1366 args_size->var = round_up (args_size->var, preferred_stack_boundary); 1367 } 1368 1369 if (reg_parm_stack_space > 0) 1370 { 1371 args_size->var 1372 = size_binop (MAX_EXPR, args_size->var, 1373 ssize_int (reg_parm_stack_space)); 1374 1375 /* The area corresponding to register parameters is not to count in 1376 the size of the block we need. So make the adjustment. */ 1377 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) 1378 args_size->var 1379 = size_binop (MINUS_EXPR, args_size->var, 1380 ssize_int (reg_parm_stack_space)); 1381 } 1382 } 1383 else 1384 { 1385 preferred_stack_boundary /= BITS_PER_UNIT; 1386 if (preferred_stack_boundary < 1) 1387 preferred_stack_boundary = 1; 1388 args_size->constant = (((args_size->constant 1389 + stack_pointer_delta 1390 + preferred_stack_boundary - 1) 1391 / preferred_stack_boundary 1392 * preferred_stack_boundary) 1393 - stack_pointer_delta); 1394 1395 args_size->constant = MAX (args_size->constant, 1396 reg_parm_stack_space); 1397 1398 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) 1399 args_size->constant -= reg_parm_stack_space; 1400 } 1401 return unadjusted_args_size; 1402 } 1403 1404 /* Precompute parameters as needed for a function call. 1405 1406 FLAGS is mask of ECF_* constants. 1407 1408 NUM_ACTUALS is the number of arguments. 1409 1410 ARGS is an array containing information for each argument; this 1411 routine fills in the INITIAL_VALUE and VALUE fields for each 1412 precomputed argument. */ 1413 1414 static void 1415 precompute_arguments (int num_actuals, struct arg_data *args) 1416 { 1417 int i; 1418 1419 /* If this is a libcall, then precompute all arguments so that we do not 1420 get extraneous instructions emitted as part of the libcall sequence. */ 1421 1422 /* If we preallocated the stack space, and some arguments must be passed 1423 on the stack, then we must precompute any parameter which contains a 1424 function call which will store arguments on the stack. 1425 Otherwise, evaluating the parameter may clobber previous parameters 1426 which have already been stored into the stack. (we have code to avoid 1427 such case by saving the outgoing stack arguments, but it results in 1428 worse code) */ 1429 if (!ACCUMULATE_OUTGOING_ARGS) 1430 return; 1431 1432 for (i = 0; i < num_actuals; i++) 1433 { 1434 tree type; 1435 enum machine_mode mode; 1436 1437 if (TREE_CODE (args[i].tree_value) != CALL_EXPR) 1438 continue; 1439 1440 /* If this is an addressable type, we cannot pre-evaluate it. */ 1441 type = TREE_TYPE (args[i].tree_value); 1442 gcc_assert (!TREE_ADDRESSABLE (type)); 1443 1444 args[i].initial_value = args[i].value 1445 = expand_normal (args[i].tree_value); 1446 1447 mode = TYPE_MODE (type); 1448 if (mode != args[i].mode) 1449 { 1450 int unsignedp = args[i].unsignedp; 1451 args[i].value 1452 = convert_modes (args[i].mode, mode, 1453 args[i].value, args[i].unsignedp); 1454 1455 /* CSE will replace this only if it contains args[i].value 1456 pseudo, so convert it down to the declared mode using 1457 a SUBREG. */ 1458 if (REG_P (args[i].value) 1459 && GET_MODE_CLASS (args[i].mode) == MODE_INT 1460 && promote_mode (type, mode, &unsignedp) != args[i].mode) 1461 { 1462 args[i].initial_value 1463 = gen_lowpart_SUBREG (mode, args[i].value); 1464 SUBREG_PROMOTED_VAR_P (args[i].initial_value) = 1; 1465 SUBREG_PROMOTED_UNSIGNED_SET (args[i].initial_value, 1466 args[i].unsignedp); 1467 } 1468 } 1469 } 1470 } 1471 1472 /* Given the current state of MUST_PREALLOCATE and information about 1473 arguments to a function call in NUM_ACTUALS, ARGS and ARGS_SIZE, 1474 compute and return the final value for MUST_PREALLOCATE. */ 1475 1476 static int 1477 finalize_must_preallocate (int must_preallocate, int num_actuals, 1478 struct arg_data *args, struct args_size *args_size) 1479 { 1480 /* See if we have or want to preallocate stack space. 1481 1482 If we would have to push a partially-in-regs parm 1483 before other stack parms, preallocate stack space instead. 1484 1485 If the size of some parm is not a multiple of the required stack 1486 alignment, we must preallocate. 1487 1488 If the total size of arguments that would otherwise create a copy in 1489 a temporary (such as a CALL) is more than half the total argument list 1490 size, preallocation is faster. 1491 1492 Another reason to preallocate is if we have a machine (like the m88k) 1493 where stack alignment is required to be maintained between every 1494 pair of insns, not just when the call is made. However, we assume here 1495 that such machines either do not have push insns (and hence preallocation 1496 would occur anyway) or the problem is taken care of with 1497 PUSH_ROUNDING. */ 1498 1499 if (! must_preallocate) 1500 { 1501 int partial_seen = 0; 1502 int copy_to_evaluate_size = 0; 1503 int i; 1504 1505 for (i = 0; i < num_actuals && ! must_preallocate; i++) 1506 { 1507 if (args[i].partial > 0 && ! args[i].pass_on_stack) 1508 partial_seen = 1; 1509 else if (partial_seen && args[i].reg == 0) 1510 must_preallocate = 1; 1511 1512 if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode 1513 && (TREE_CODE (args[i].tree_value) == CALL_EXPR 1514 || TREE_CODE (args[i].tree_value) == TARGET_EXPR 1515 || TREE_CODE (args[i].tree_value) == COND_EXPR 1516 || TREE_ADDRESSABLE (TREE_TYPE (args[i].tree_value)))) 1517 copy_to_evaluate_size 1518 += int_size_in_bytes (TREE_TYPE (args[i].tree_value)); 1519 } 1520 1521 if (copy_to_evaluate_size * 2 >= args_size->constant 1522 && args_size->constant > 0) 1523 must_preallocate = 1; 1524 } 1525 return must_preallocate; 1526 } 1527 1528 /* If we preallocated stack space, compute the address of each argument 1529 and store it into the ARGS array. 1530 1531 We need not ensure it is a valid memory address here; it will be 1532 validized when it is used. 1533 1534 ARGBLOCK is an rtx for the address of the outgoing arguments. */ 1535 1536 static void 1537 compute_argument_addresses (struct arg_data *args, rtx argblock, int num_actuals) 1538 { 1539 if (argblock) 1540 { 1541 rtx arg_reg = argblock; 1542 int i, arg_offset = 0; 1543 1544 if (GET_CODE (argblock) == PLUS) 1545 arg_reg = XEXP (argblock, 0), arg_offset = INTVAL (XEXP (argblock, 1)); 1546 1547 for (i = 0; i < num_actuals; i++) 1548 { 1549 rtx offset = ARGS_SIZE_RTX (args[i].locate.offset); 1550 rtx slot_offset = ARGS_SIZE_RTX (args[i].locate.slot_offset); 1551 rtx addr; 1552 unsigned int align, boundary; 1553 unsigned int units_on_stack = 0; 1554 enum machine_mode partial_mode = VOIDmode; 1555 1556 /* Skip this parm if it will not be passed on the stack. */ 1557 if (! args[i].pass_on_stack 1558 && args[i].reg != 0 1559 && args[i].partial == 0) 1560 continue; 1561 1562 if (CONST_INT_P (offset)) 1563 addr = plus_constant (arg_reg, INTVAL (offset)); 1564 else 1565 addr = gen_rtx_PLUS (Pmode, arg_reg, offset); 1566 1567 addr = plus_constant (addr, arg_offset); 1568 1569 if (args[i].partial != 0) 1570 { 1571 /* Only part of the parameter is being passed on the stack. 1572 Generate a simple memory reference of the correct size. */ 1573 units_on_stack = args[i].locate.size.constant; 1574 partial_mode = mode_for_size (units_on_stack * BITS_PER_UNIT, 1575 MODE_INT, 1); 1576 args[i].stack = gen_rtx_MEM (partial_mode, addr); 1577 set_mem_size (args[i].stack, units_on_stack); 1578 } 1579 else 1580 { 1581 args[i].stack = gen_rtx_MEM (args[i].mode, addr); 1582 set_mem_attributes (args[i].stack, 1583 TREE_TYPE (args[i].tree_value), 1); 1584 } 1585 align = BITS_PER_UNIT; 1586 boundary = args[i].locate.boundary; 1587 if (args[i].locate.where_pad != downward) 1588 align = boundary; 1589 else if (CONST_INT_P (offset)) 1590 { 1591 align = INTVAL (offset) * BITS_PER_UNIT | boundary; 1592 align = align & -align; 1593 } 1594 set_mem_align (args[i].stack, align); 1595 1596 if (CONST_INT_P (slot_offset)) 1597 addr = plus_constant (arg_reg, INTVAL (slot_offset)); 1598 else 1599 addr = gen_rtx_PLUS (Pmode, arg_reg, slot_offset); 1600 1601 addr = plus_constant (addr, arg_offset); 1602 1603 if (args[i].partial != 0) 1604 { 1605 /* Only part of the parameter is being passed on the stack. 1606 Generate a simple memory reference of the correct size. 1607 */ 1608 args[i].stack_slot = gen_rtx_MEM (partial_mode, addr); 1609 set_mem_size (args[i].stack_slot, units_on_stack); 1610 } 1611 else 1612 { 1613 args[i].stack_slot = gen_rtx_MEM (args[i].mode, addr); 1614 set_mem_attributes (args[i].stack_slot, 1615 TREE_TYPE (args[i].tree_value), 1); 1616 } 1617 set_mem_align (args[i].stack_slot, args[i].locate.boundary); 1618 1619 /* Function incoming arguments may overlap with sibling call 1620 outgoing arguments and we cannot allow reordering of reads 1621 from function arguments with stores to outgoing arguments 1622 of sibling calls. */ 1623 set_mem_alias_set (args[i].stack, 0); 1624 set_mem_alias_set (args[i].stack_slot, 0); 1625 } 1626 } 1627 } 1628 1629 /* Given a FNDECL and EXP, return an rtx suitable for use as a target address 1630 in a call instruction. 1631 1632 FNDECL is the tree node for the target function. For an indirect call 1633 FNDECL will be NULL_TREE. 1634 1635 ADDR is the operand 0 of CALL_EXPR for this call. */ 1636 1637 static rtx 1638 rtx_for_function_call (tree fndecl, tree addr) 1639 { 1640 rtx funexp; 1641 1642 /* Get the function to call, in the form of RTL. */ 1643 if (fndecl) 1644 { 1645 /* If this is the first use of the function, see if we need to 1646 make an external definition for it. */ 1647 if (!TREE_USED (fndecl) && fndecl != current_function_decl) 1648 { 1649 assemble_external (fndecl); 1650 TREE_USED (fndecl) = 1; 1651 } 1652 1653 /* Get a SYMBOL_REF rtx for the function address. */ 1654 funexp = XEXP (DECL_RTL (fndecl), 0); 1655 } 1656 else 1657 /* Generate an rtx (probably a pseudo-register) for the address. */ 1658 { 1659 push_temp_slots (); 1660 funexp = expand_normal (addr); 1661 pop_temp_slots (); /* FUNEXP can't be BLKmode. */ 1662 } 1663 return funexp; 1664 } 1665 1666 /* Internal state for internal_arg_pointer_based_exp and its helpers. */ 1667 static struct 1668 { 1669 /* Last insn that has been scanned by internal_arg_pointer_based_exp_scan, 1670 or NULL_RTX if none has been scanned yet. */ 1671 rtx scan_start; 1672 /* Vector indexed by REGNO - FIRST_PSEUDO_REGISTER, recording if a pseudo is 1673 based on crtl->args.internal_arg_pointer. The element is NULL_RTX if the 1674 pseudo isn't based on it, a CONST_INT offset if the pseudo is based on it 1675 with fixed offset, or PC if this is with variable or unknown offset. */ 1676 VEC(rtx, heap) *cache; 1677 } internal_arg_pointer_exp_state; 1678 1679 static rtx internal_arg_pointer_based_exp (rtx, bool); 1680 1681 /* Helper function for internal_arg_pointer_based_exp. Scan insns in 1682 the tail call sequence, starting with first insn that hasn't been 1683 scanned yet, and note for each pseudo on the LHS whether it is based 1684 on crtl->args.internal_arg_pointer or not, and what offset from that 1685 that pointer it has. */ 1686 1687 static void 1688 internal_arg_pointer_based_exp_scan (void) 1689 { 1690 rtx insn, scan_start = internal_arg_pointer_exp_state.scan_start; 1691 1692 if (scan_start == NULL_RTX) 1693 insn = get_insns (); 1694 else 1695 insn = NEXT_INSN (scan_start); 1696 1697 while (insn) 1698 { 1699 rtx set = single_set (insn); 1700 if (set && REG_P (SET_DEST (set)) && !HARD_REGISTER_P (SET_DEST (set))) 1701 { 1702 rtx val = NULL_RTX; 1703 unsigned int idx = REGNO (SET_DEST (set)) - FIRST_PSEUDO_REGISTER; 1704 /* Punt on pseudos set multiple times. */ 1705 if (idx < VEC_length (rtx, internal_arg_pointer_exp_state.cache) 1706 && (VEC_index (rtx, internal_arg_pointer_exp_state.cache, idx) 1707 != NULL_RTX)) 1708 val = pc_rtx; 1709 else 1710 val = internal_arg_pointer_based_exp (SET_SRC (set), false); 1711 if (val != NULL_RTX) 1712 { 1713 if (idx 1714 >= VEC_length (rtx, internal_arg_pointer_exp_state.cache)) 1715 VEC_safe_grow_cleared (rtx, heap, 1716 internal_arg_pointer_exp_state.cache, 1717 idx + 1); 1718 VEC_replace (rtx, internal_arg_pointer_exp_state.cache, 1719 idx, val); 1720 } 1721 } 1722 if (NEXT_INSN (insn) == NULL_RTX) 1723 scan_start = insn; 1724 insn = NEXT_INSN (insn); 1725 } 1726 1727 internal_arg_pointer_exp_state.scan_start = scan_start; 1728 } 1729 1730 /* Helper function for internal_arg_pointer_based_exp, called through 1731 for_each_rtx. Return 1 if *LOC is a register based on 1732 crtl->args.internal_arg_pointer. Return -1 if *LOC is not based on it 1733 and the subexpressions need not be examined. Otherwise return 0. */ 1734 1735 static int 1736 internal_arg_pointer_based_exp_1 (rtx *loc, void *data ATTRIBUTE_UNUSED) 1737 { 1738 if (REG_P (*loc) && internal_arg_pointer_based_exp (*loc, false) != NULL_RTX) 1739 return 1; 1740 if (MEM_P (*loc)) 1741 return -1; 1742 return 0; 1743 } 1744 1745 /* Compute whether RTL is based on crtl->args.internal_arg_pointer. Return 1746 NULL_RTX if RTL isn't based on it, a CONST_INT offset if RTL is based on 1747 it with fixed offset, or PC if this is with variable or unknown offset. 1748 TOPLEVEL is true if the function is invoked at the topmost level. */ 1749 1750 static rtx 1751 internal_arg_pointer_based_exp (rtx rtl, bool toplevel) 1752 { 1753 if (CONSTANT_P (rtl)) 1754 return NULL_RTX; 1755 1756 if (rtl == crtl->args.internal_arg_pointer) 1757 return const0_rtx; 1758 1759 if (REG_P (rtl) && HARD_REGISTER_P (rtl)) 1760 return NULL_RTX; 1761 1762 if (GET_CODE (rtl) == PLUS && CONST_INT_P (XEXP (rtl, 1))) 1763 { 1764 rtx val = internal_arg_pointer_based_exp (XEXP (rtl, 0), toplevel); 1765 if (val == NULL_RTX || val == pc_rtx) 1766 return val; 1767 return plus_constant (val, INTVAL (XEXP (rtl, 1))); 1768 } 1769 1770 /* When called at the topmost level, scan pseudo assignments in between the 1771 last scanned instruction in the tail call sequence and the latest insn 1772 in that sequence. */ 1773 if (toplevel) 1774 internal_arg_pointer_based_exp_scan (); 1775 1776 if (REG_P (rtl)) 1777 { 1778 unsigned int idx = REGNO (rtl) - FIRST_PSEUDO_REGISTER; 1779 if (idx < VEC_length (rtx, internal_arg_pointer_exp_state.cache)) 1780 return VEC_index (rtx, internal_arg_pointer_exp_state.cache, idx); 1781 1782 return NULL_RTX; 1783 } 1784 1785 if (for_each_rtx (&rtl, internal_arg_pointer_based_exp_1, NULL)) 1786 return pc_rtx; 1787 1788 return NULL_RTX; 1789 } 1790 1791 /* Return true if and only if SIZE storage units (usually bytes) 1792 starting from address ADDR overlap with already clobbered argument 1793 area. This function is used to determine if we should give up a 1794 sibcall. */ 1795 1796 static bool 1797 mem_overlaps_already_clobbered_arg_p (rtx addr, unsigned HOST_WIDE_INT size) 1798 { 1799 HOST_WIDE_INT i; 1800 rtx val; 1801 1802 if (sbitmap_empty_p (stored_args_map)) 1803 return false; 1804 val = internal_arg_pointer_based_exp (addr, true); 1805 if (val == NULL_RTX) 1806 return false; 1807 else if (val == pc_rtx) 1808 return true; 1809 else 1810 i = INTVAL (val); 1811 #ifdef STACK_GROWS_DOWNWARD 1812 i -= crtl->args.pretend_args_size; 1813 #else 1814 i += crtl->args.pretend_args_size; 1815 #endif 1816 1817 #ifdef ARGS_GROW_DOWNWARD 1818 i = -i - size; 1819 #endif 1820 if (size > 0) 1821 { 1822 unsigned HOST_WIDE_INT k; 1823 1824 for (k = 0; k < size; k++) 1825 if (i + k < stored_args_map->n_bits 1826 && TEST_BIT (stored_args_map, i + k)) 1827 return true; 1828 } 1829 1830 return false; 1831 } 1832 1833 /* Do the register loads required for any wholly-register parms or any 1834 parms which are passed both on the stack and in a register. Their 1835 expressions were already evaluated. 1836 1837 Mark all register-parms as living through the call, putting these USE 1838 insns in the CALL_INSN_FUNCTION_USAGE field. 1839 1840 When IS_SIBCALL, perform the check_sibcall_argument_overlap 1841 checking, setting *SIBCALL_FAILURE if appropriate. */ 1842 1843 static void 1844 load_register_parameters (struct arg_data *args, int num_actuals, 1845 rtx *call_fusage, int flags, int is_sibcall, 1846 int *sibcall_failure) 1847 { 1848 int i, j; 1849 1850 for (i = 0; i < num_actuals; i++) 1851 { 1852 rtx reg = ((flags & ECF_SIBCALL) 1853 ? args[i].tail_call_reg : args[i].reg); 1854 if (reg) 1855 { 1856 int partial = args[i].partial; 1857 int nregs; 1858 int size = 0; 1859 rtx before_arg = get_last_insn (); 1860 /* Set non-negative if we must move a word at a time, even if 1861 just one word (e.g, partial == 4 && mode == DFmode). Set 1862 to -1 if we just use a normal move insn. This value can be 1863 zero if the argument is a zero size structure. */ 1864 nregs = -1; 1865 if (GET_CODE (reg) == PARALLEL) 1866 ; 1867 else if (partial) 1868 { 1869 gcc_assert (partial % UNITS_PER_WORD == 0); 1870 nregs = partial / UNITS_PER_WORD; 1871 } 1872 else if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode) 1873 { 1874 size = int_size_in_bytes (TREE_TYPE (args[i].tree_value)); 1875 nregs = (size + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD; 1876 } 1877 else 1878 size = GET_MODE_SIZE (args[i].mode); 1879 1880 /* Handle calls that pass values in multiple non-contiguous 1881 locations. The Irix 6 ABI has examples of this. */ 1882 1883 if (GET_CODE (reg) == PARALLEL) 1884 emit_group_move (reg, args[i].parallel_value); 1885 1886 /* If simple case, just do move. If normal partial, store_one_arg 1887 has already loaded the register for us. In all other cases, 1888 load the register(s) from memory. */ 1889 1890 else if (nregs == -1) 1891 { 1892 emit_move_insn (reg, args[i].value); 1893 #ifdef BLOCK_REG_PADDING 1894 /* Handle case where we have a value that needs shifting 1895 up to the msb. eg. a QImode value and we're padding 1896 upward on a BYTES_BIG_ENDIAN machine. */ 1897 if (size < UNITS_PER_WORD 1898 && (args[i].locate.where_pad 1899 == (BYTES_BIG_ENDIAN ? upward : downward))) 1900 { 1901 rtx x; 1902 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT; 1903 1904 /* Assigning REG here rather than a temp makes CALL_FUSAGE 1905 report the whole reg as used. Strictly speaking, the 1906 call only uses SIZE bytes at the msb end, but it doesn't 1907 seem worth generating rtl to say that. */ 1908 reg = gen_rtx_REG (word_mode, REGNO (reg)); 1909 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1); 1910 if (x != reg) 1911 emit_move_insn (reg, x); 1912 } 1913 #endif 1914 } 1915 1916 /* If we have pre-computed the values to put in the registers in 1917 the case of non-aligned structures, copy them in now. */ 1918 1919 else if (args[i].n_aligned_regs != 0) 1920 for (j = 0; j < args[i].n_aligned_regs; j++) 1921 emit_move_insn (gen_rtx_REG (word_mode, REGNO (reg) + j), 1922 args[i].aligned_regs[j]); 1923 1924 else if (partial == 0 || args[i].pass_on_stack) 1925 { 1926 rtx mem = validize_mem (args[i].value); 1927 1928 /* Check for overlap with already clobbered argument area, 1929 providing that this has non-zero size. */ 1930 if (is_sibcall 1931 && (size == 0 1932 || mem_overlaps_already_clobbered_arg_p 1933 (XEXP (args[i].value, 0), size))) 1934 *sibcall_failure = 1; 1935 1936 /* Handle a BLKmode that needs shifting. */ 1937 if (nregs == 1 && size < UNITS_PER_WORD 1938 #ifdef BLOCK_REG_PADDING 1939 && args[i].locate.where_pad == downward 1940 #else 1941 && BYTES_BIG_ENDIAN 1942 #endif 1943 ) 1944 { 1945 rtx tem = operand_subword_force (mem, 0, args[i].mode); 1946 rtx ri = gen_rtx_REG (word_mode, REGNO (reg)); 1947 rtx x = gen_reg_rtx (word_mode); 1948 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT; 1949 enum tree_code dir = BYTES_BIG_ENDIAN ? RSHIFT_EXPR 1950 : LSHIFT_EXPR; 1951 1952 emit_move_insn (x, tem); 1953 x = expand_shift (dir, word_mode, x, shift, ri, 1); 1954 if (x != ri) 1955 emit_move_insn (ri, x); 1956 } 1957 else 1958 move_block_to_reg (REGNO (reg), mem, nregs, args[i].mode); 1959 } 1960 1961 /* When a parameter is a block, and perhaps in other cases, it is 1962 possible that it did a load from an argument slot that was 1963 already clobbered. */ 1964 if (is_sibcall 1965 && check_sibcall_argument_overlap (before_arg, &args[i], 0)) 1966 *sibcall_failure = 1; 1967 1968 /* Handle calls that pass values in multiple non-contiguous 1969 locations. The Irix 6 ABI has examples of this. */ 1970 if (GET_CODE (reg) == PARALLEL) 1971 use_group_regs (call_fusage, reg); 1972 else if (nregs == -1) 1973 use_reg_mode (call_fusage, reg, 1974 TYPE_MODE (TREE_TYPE (args[i].tree_value))); 1975 else if (nregs > 0) 1976 use_regs (call_fusage, REGNO (reg), nregs); 1977 } 1978 } 1979 } 1980 1981 /* We need to pop PENDING_STACK_ADJUST bytes. But, if the arguments 1982 wouldn't fill up an even multiple of PREFERRED_UNIT_STACK_BOUNDARY 1983 bytes, then we would need to push some additional bytes to pad the 1984 arguments. So, we compute an adjust to the stack pointer for an 1985 amount that will leave the stack under-aligned by UNADJUSTED_ARGS_SIZE 1986 bytes. Then, when the arguments are pushed the stack will be perfectly 1987 aligned. ARGS_SIZE->CONSTANT is set to the number of bytes that should 1988 be popped after the call. Returns the adjustment. */ 1989 1990 static int 1991 combine_pending_stack_adjustment_and_call (int unadjusted_args_size, 1992 struct args_size *args_size, 1993 unsigned int preferred_unit_stack_boundary) 1994 { 1995 /* The number of bytes to pop so that the stack will be 1996 under-aligned by UNADJUSTED_ARGS_SIZE bytes. */ 1997 HOST_WIDE_INT adjustment; 1998 /* The alignment of the stack after the arguments are pushed, if we 1999 just pushed the arguments without adjust the stack here. */ 2000 unsigned HOST_WIDE_INT unadjusted_alignment; 2001 2002 unadjusted_alignment 2003 = ((stack_pointer_delta + unadjusted_args_size) 2004 % preferred_unit_stack_boundary); 2005 2006 /* We want to get rid of as many of the PENDING_STACK_ADJUST bytes 2007 as possible -- leaving just enough left to cancel out the 2008 UNADJUSTED_ALIGNMENT. In other words, we want to ensure that the 2009 PENDING_STACK_ADJUST is non-negative, and congruent to 2010 -UNADJUSTED_ALIGNMENT modulo the PREFERRED_UNIT_STACK_BOUNDARY. */ 2011 2012 /* Begin by trying to pop all the bytes. */ 2013 unadjusted_alignment 2014 = (unadjusted_alignment 2015 - (pending_stack_adjust % preferred_unit_stack_boundary)); 2016 adjustment = pending_stack_adjust; 2017 /* Push enough additional bytes that the stack will be aligned 2018 after the arguments are pushed. */ 2019 if (preferred_unit_stack_boundary > 1) 2020 { 2021 if (unadjusted_alignment > 0) 2022 adjustment -= preferred_unit_stack_boundary - unadjusted_alignment; 2023 else 2024 adjustment += unadjusted_alignment; 2025 } 2026 2027 /* Now, sets ARGS_SIZE->CONSTANT so that we pop the right number of 2028 bytes after the call. The right number is the entire 2029 PENDING_STACK_ADJUST less our ADJUSTMENT plus the amount required 2030 by the arguments in the first place. */ 2031 args_size->constant 2032 = pending_stack_adjust - adjustment + unadjusted_args_size; 2033 2034 return adjustment; 2035 } 2036 2037 /* Scan X expression if it does not dereference any argument slots 2038 we already clobbered by tail call arguments (as noted in stored_args_map 2039 bitmap). 2040 Return nonzero if X expression dereferences such argument slots, 2041 zero otherwise. */ 2042 2043 static int 2044 check_sibcall_argument_overlap_1 (rtx x) 2045 { 2046 RTX_CODE code; 2047 int i, j; 2048 const char *fmt; 2049 2050 if (x == NULL_RTX) 2051 return 0; 2052 2053 code = GET_CODE (x); 2054 2055 /* We need not check the operands of the CALL expression itself. */ 2056 if (code == CALL) 2057 return 0; 2058 2059 if (code == MEM) 2060 return mem_overlaps_already_clobbered_arg_p (XEXP (x, 0), 2061 GET_MODE_SIZE (GET_MODE (x))); 2062 2063 /* Scan all subexpressions. */ 2064 fmt = GET_RTX_FORMAT (code); 2065 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++) 2066 { 2067 if (*fmt == 'e') 2068 { 2069 if (check_sibcall_argument_overlap_1 (XEXP (x, i))) 2070 return 1; 2071 } 2072 else if (*fmt == 'E') 2073 { 2074 for (j = 0; j < XVECLEN (x, i); j++) 2075 if (check_sibcall_argument_overlap_1 (XVECEXP (x, i, j))) 2076 return 1; 2077 } 2078 } 2079 return 0; 2080 } 2081 2082 /* Scan sequence after INSN if it does not dereference any argument slots 2083 we already clobbered by tail call arguments (as noted in stored_args_map 2084 bitmap). If MARK_STORED_ARGS_MAP, add stack slots for ARG to 2085 stored_args_map bitmap afterwards (when ARG is a register MARK_STORED_ARGS_MAP 2086 should be 0). Return nonzero if sequence after INSN dereferences such argument 2087 slots, zero otherwise. */ 2088 2089 static int 2090 check_sibcall_argument_overlap (rtx insn, struct arg_data *arg, int mark_stored_args_map) 2091 { 2092 int low, high; 2093 2094 if (insn == NULL_RTX) 2095 insn = get_insns (); 2096 else 2097 insn = NEXT_INSN (insn); 2098 2099 for (; insn; insn = NEXT_INSN (insn)) 2100 if (INSN_P (insn) 2101 && check_sibcall_argument_overlap_1 (PATTERN (insn))) 2102 break; 2103 2104 if (mark_stored_args_map) 2105 { 2106 #ifdef ARGS_GROW_DOWNWARD 2107 low = -arg->locate.slot_offset.constant - arg->locate.size.constant; 2108 #else 2109 low = arg->locate.slot_offset.constant; 2110 #endif 2111 2112 for (high = low + arg->locate.size.constant; low < high; low++) 2113 SET_BIT (stored_args_map, low); 2114 } 2115 return insn != NULL_RTX; 2116 } 2117 2118 /* Given that a function returns a value of mode MODE at the most 2119 significant end of hard register VALUE, shift VALUE left or right 2120 as specified by LEFT_P. Return true if some action was needed. */ 2121 2122 bool 2123 shift_return_value (enum machine_mode mode, bool left_p, rtx value) 2124 { 2125 HOST_WIDE_INT shift; 2126 2127 gcc_assert (REG_P (value) && HARD_REGISTER_P (value)); 2128 shift = GET_MODE_BITSIZE (GET_MODE (value)) - GET_MODE_BITSIZE (mode); 2129 if (shift == 0) 2130 return false; 2131 2132 /* Use ashr rather than lshr for right shifts. This is for the benefit 2133 of the MIPS port, which requires SImode values to be sign-extended 2134 when stored in 64-bit registers. */ 2135 if (!force_expand_binop (GET_MODE (value), left_p ? ashl_optab : ashr_optab, 2136 value, GEN_INT (shift), value, 1, OPTAB_WIDEN)) 2137 gcc_unreachable (); 2138 return true; 2139 } 2140 2141 /* If X is a likely-spilled register value, copy it to a pseudo 2142 register and return that register. Return X otherwise. */ 2143 2144 static rtx 2145 avoid_likely_spilled_reg (rtx x) 2146 { 2147 rtx new_rtx; 2148 2149 if (REG_P (x) 2150 && HARD_REGISTER_P (x) 2151 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (x)))) 2152 { 2153 /* Make sure that we generate a REG rather than a CONCAT. 2154 Moves into CONCATs can need nontrivial instructions, 2155 and the whole point of this function is to avoid 2156 using the hard register directly in such a situation. */ 2157 generating_concat_p = 0; 2158 new_rtx = gen_reg_rtx (GET_MODE (x)); 2159 generating_concat_p = 1; 2160 emit_move_insn (new_rtx, x); 2161 return new_rtx; 2162 } 2163 return x; 2164 } 2165 2166 /* Generate all the code for a CALL_EXPR exp 2167 and return an rtx for its value. 2168 Store the value in TARGET (specified as an rtx) if convenient. 2169 If the value is stored in TARGET then TARGET is returned. 2170 If IGNORE is nonzero, then we ignore the value of the function call. */ 2171 2172 rtx 2173 expand_call (tree exp, rtx target, int ignore) 2174 { 2175 /* Nonzero if we are currently expanding a call. */ 2176 static int currently_expanding_call = 0; 2177 2178 /* RTX for the function to be called. */ 2179 rtx funexp; 2180 /* Sequence of insns to perform a normal "call". */ 2181 rtx normal_call_insns = NULL_RTX; 2182 /* Sequence of insns to perform a tail "call". */ 2183 rtx tail_call_insns = NULL_RTX; 2184 /* Data type of the function. */ 2185 tree funtype; 2186 tree type_arg_types; 2187 tree rettype; 2188 /* Declaration of the function being called, 2189 or 0 if the function is computed (not known by name). */ 2190 tree fndecl = 0; 2191 /* The type of the function being called. */ 2192 tree fntype; 2193 bool try_tail_call = CALL_EXPR_TAILCALL (exp); 2194 int pass; 2195 2196 /* Register in which non-BLKmode value will be returned, 2197 or 0 if no value or if value is BLKmode. */ 2198 rtx valreg; 2199 /* Address where we should return a BLKmode value; 2200 0 if value not BLKmode. */ 2201 rtx structure_value_addr = 0; 2202 /* Nonzero if that address is being passed by treating it as 2203 an extra, implicit first parameter. Otherwise, 2204 it is passed by being copied directly into struct_value_rtx. */ 2205 int structure_value_addr_parm = 0; 2206 /* Holds the value of implicit argument for the struct value. */ 2207 tree structure_value_addr_value = NULL_TREE; 2208 /* Size of aggregate value wanted, or zero if none wanted 2209 or if we are using the non-reentrant PCC calling convention 2210 or expecting the value in registers. */ 2211 HOST_WIDE_INT struct_value_size = 0; 2212 /* Nonzero if called function returns an aggregate in memory PCC style, 2213 by returning the address of where to find it. */ 2214 int pcc_struct_value = 0; 2215 rtx struct_value = 0; 2216 2217 /* Number of actual parameters in this call, including struct value addr. */ 2218 int num_actuals; 2219 /* Number of named args. Args after this are anonymous ones 2220 and they must all go on the stack. */ 2221 int n_named_args; 2222 /* Number of complex actual arguments that need to be split. */ 2223 int num_complex_actuals = 0; 2224 2225 /* Vector of information about each argument. 2226 Arguments are numbered in the order they will be pushed, 2227 not the order they are written. */ 2228 struct arg_data *args; 2229 2230 /* Total size in bytes of all the stack-parms scanned so far. */ 2231 struct args_size args_size; 2232 struct args_size adjusted_args_size; 2233 /* Size of arguments before any adjustments (such as rounding). */ 2234 int unadjusted_args_size; 2235 /* Data on reg parms scanned so far. */ 2236 CUMULATIVE_ARGS args_so_far_v; 2237 cumulative_args_t args_so_far; 2238 /* Nonzero if a reg parm has been scanned. */ 2239 int reg_parm_seen; 2240 /* Nonzero if this is an indirect function call. */ 2241 2242 /* Nonzero if we must avoid push-insns in the args for this call. 2243 If stack space is allocated for register parameters, but not by the 2244 caller, then it is preallocated in the fixed part of the stack frame. 2245 So the entire argument block must then be preallocated (i.e., we 2246 ignore PUSH_ROUNDING in that case). */ 2247 2248 int must_preallocate = !PUSH_ARGS; 2249 2250 /* Size of the stack reserved for parameter registers. */ 2251 int reg_parm_stack_space = 0; 2252 2253 /* Address of space preallocated for stack parms 2254 (on machines that lack push insns), or 0 if space not preallocated. */ 2255 rtx argblock = 0; 2256 2257 /* Mask of ECF_ flags. */ 2258 int flags = 0; 2259 #ifdef REG_PARM_STACK_SPACE 2260 /* Define the boundary of the register parm stack space that needs to be 2261 saved, if any. */ 2262 int low_to_save, high_to_save; 2263 rtx save_area = 0; /* Place that it is saved */ 2264 #endif 2265 2266 int initial_highest_arg_in_use = highest_outgoing_arg_in_use; 2267 char *initial_stack_usage_map = stack_usage_map; 2268 char *stack_usage_map_buf = NULL; 2269 2270 int old_stack_allocated; 2271 2272 /* State variables to track stack modifications. */ 2273 rtx old_stack_level = 0; 2274 int old_stack_arg_under_construction = 0; 2275 int old_pending_adj = 0; 2276 int old_inhibit_defer_pop = inhibit_defer_pop; 2277 2278 /* Some stack pointer alterations we make are performed via 2279 allocate_dynamic_stack_space. This modifies the stack_pointer_delta, 2280 which we then also need to save/restore along the way. */ 2281 int old_stack_pointer_delta = 0; 2282 2283 rtx call_fusage; 2284 tree addr = CALL_EXPR_FN (exp); 2285 int i; 2286 /* The alignment of the stack, in bits. */ 2287 unsigned HOST_WIDE_INT preferred_stack_boundary; 2288 /* The alignment of the stack, in bytes. */ 2289 unsigned HOST_WIDE_INT preferred_unit_stack_boundary; 2290 /* The static chain value to use for this call. */ 2291 rtx static_chain_value; 2292 /* See if this is "nothrow" function call. */ 2293 if (TREE_NOTHROW (exp)) 2294 flags |= ECF_NOTHROW; 2295 2296 /* See if we can find a DECL-node for the actual function, and get the 2297 function attributes (flags) from the function decl or type node. */ 2298 fndecl = get_callee_fndecl (exp); 2299 if (fndecl) 2300 { 2301 fntype = TREE_TYPE (fndecl); 2302 flags |= flags_from_decl_or_type (fndecl); 2303 } 2304 else 2305 { 2306 fntype = TREE_TYPE (TREE_TYPE (addr)); 2307 flags |= flags_from_decl_or_type (fntype); 2308 } 2309 rettype = TREE_TYPE (exp); 2310 2311 struct_value = targetm.calls.struct_value_rtx (fntype, 0); 2312 2313 /* Warn if this value is an aggregate type, 2314 regardless of which calling convention we are using for it. */ 2315 if (AGGREGATE_TYPE_P (rettype)) 2316 warning (OPT_Waggregate_return, "function call has aggregate value"); 2317 2318 /* If the result of a non looping pure or const function call is 2319 ignored (or void), and none of its arguments are volatile, we can 2320 avoid expanding the call and just evaluate the arguments for 2321 side-effects. */ 2322 if ((flags & (ECF_CONST | ECF_PURE)) 2323 && (!(flags & ECF_LOOPING_CONST_OR_PURE)) 2324 && (ignore || target == const0_rtx 2325 || TYPE_MODE (rettype) == VOIDmode)) 2326 { 2327 bool volatilep = false; 2328 tree arg; 2329 call_expr_arg_iterator iter; 2330 2331 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp) 2332 if (TREE_THIS_VOLATILE (arg)) 2333 { 2334 volatilep = true; 2335 break; 2336 } 2337 2338 if (! volatilep) 2339 { 2340 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp) 2341 expand_expr (arg, const0_rtx, VOIDmode, EXPAND_NORMAL); 2342 return const0_rtx; 2343 } 2344 } 2345 2346 #ifdef REG_PARM_STACK_SPACE 2347 reg_parm_stack_space = REG_PARM_STACK_SPACE (!fndecl ? fntype : fndecl); 2348 #endif 2349 2350 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))) 2351 && reg_parm_stack_space > 0 && PUSH_ARGS) 2352 must_preallocate = 1; 2353 2354 /* Set up a place to return a structure. */ 2355 2356 /* Cater to broken compilers. */ 2357 if (aggregate_value_p (exp, fntype)) 2358 { 2359 /* This call returns a big structure. */ 2360 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE); 2361 2362 #ifdef PCC_STATIC_STRUCT_RETURN 2363 { 2364 pcc_struct_value = 1; 2365 } 2366 #else /* not PCC_STATIC_STRUCT_RETURN */ 2367 { 2368 struct_value_size = int_size_in_bytes (rettype); 2369 2370 if (target && MEM_P (target) && CALL_EXPR_RETURN_SLOT_OPT (exp)) 2371 structure_value_addr = XEXP (target, 0); 2372 else 2373 { 2374 /* For variable-sized objects, we must be called with a target 2375 specified. If we were to allocate space on the stack here, 2376 we would have no way of knowing when to free it. */ 2377 rtx d = assign_temp (rettype, 0, 1, 1); 2378 2379 mark_temp_addr_taken (d); 2380 structure_value_addr = XEXP (d, 0); 2381 target = 0; 2382 } 2383 } 2384 #endif /* not PCC_STATIC_STRUCT_RETURN */ 2385 } 2386 2387 /* Figure out the amount to which the stack should be aligned. */ 2388 preferred_stack_boundary = PREFERRED_STACK_BOUNDARY; 2389 if (fndecl) 2390 { 2391 struct cgraph_rtl_info *i = cgraph_rtl_info (fndecl); 2392 /* Without automatic stack alignment, we can't increase preferred 2393 stack boundary. With automatic stack alignment, it is 2394 unnecessary since unless we can guarantee that all callers will 2395 align the outgoing stack properly, callee has to align its 2396 stack anyway. */ 2397 if (i 2398 && i->preferred_incoming_stack_boundary 2399 && i->preferred_incoming_stack_boundary < preferred_stack_boundary) 2400 preferred_stack_boundary = i->preferred_incoming_stack_boundary; 2401 } 2402 2403 /* Operand 0 is a pointer-to-function; get the type of the function. */ 2404 funtype = TREE_TYPE (addr); 2405 gcc_assert (POINTER_TYPE_P (funtype)); 2406 funtype = TREE_TYPE (funtype); 2407 2408 /* Count whether there are actual complex arguments that need to be split 2409 into their real and imaginary parts. Munge the type_arg_types 2410 appropriately here as well. */ 2411 if (targetm.calls.split_complex_arg) 2412 { 2413 call_expr_arg_iterator iter; 2414 tree arg; 2415 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp) 2416 { 2417 tree type = TREE_TYPE (arg); 2418 if (type && TREE_CODE (type) == COMPLEX_TYPE 2419 && targetm.calls.split_complex_arg (type)) 2420 num_complex_actuals++; 2421 } 2422 type_arg_types = split_complex_types (TYPE_ARG_TYPES (funtype)); 2423 } 2424 else 2425 type_arg_types = TYPE_ARG_TYPES (funtype); 2426 2427 if (flags & ECF_MAY_BE_ALLOCA) 2428 cfun->calls_alloca = 1; 2429 2430 /* If struct_value_rtx is 0, it means pass the address 2431 as if it were an extra parameter. Put the argument expression 2432 in structure_value_addr_value. */ 2433 if (structure_value_addr && struct_value == 0) 2434 { 2435 /* If structure_value_addr is a REG other than 2436 virtual_outgoing_args_rtx, we can use always use it. If it 2437 is not a REG, we must always copy it into a register. 2438 If it is virtual_outgoing_args_rtx, we must copy it to another 2439 register in some cases. */ 2440 rtx temp = (!REG_P (structure_value_addr) 2441 || (ACCUMULATE_OUTGOING_ARGS 2442 && stack_arg_under_construction 2443 && structure_value_addr == virtual_outgoing_args_rtx) 2444 ? copy_addr_to_reg (convert_memory_address 2445 (Pmode, structure_value_addr)) 2446 : structure_value_addr); 2447 2448 structure_value_addr_value = 2449 make_tree (build_pointer_type (TREE_TYPE (funtype)), temp); 2450 structure_value_addr_parm = 1; 2451 } 2452 2453 /* Count the arguments and set NUM_ACTUALS. */ 2454 num_actuals = 2455 call_expr_nargs (exp) + num_complex_actuals + structure_value_addr_parm; 2456 2457 /* Compute number of named args. 2458 First, do a raw count of the args for INIT_CUMULATIVE_ARGS. */ 2459 2460 if (type_arg_types != 0) 2461 n_named_args 2462 = (list_length (type_arg_types) 2463 /* Count the struct value address, if it is passed as a parm. */ 2464 + structure_value_addr_parm); 2465 else 2466 /* If we know nothing, treat all args as named. */ 2467 n_named_args = num_actuals; 2468 2469 /* Start updating where the next arg would go. 2470 2471 On some machines (such as the PA) indirect calls have a different 2472 calling convention than normal calls. The fourth argument in 2473 INIT_CUMULATIVE_ARGS tells the backend if this is an indirect call 2474 or not. */ 2475 INIT_CUMULATIVE_ARGS (args_so_far_v, funtype, NULL_RTX, fndecl, n_named_args); 2476 args_so_far = pack_cumulative_args (&args_so_far_v); 2477 2478 /* Now possibly adjust the number of named args. 2479 Normally, don't include the last named arg if anonymous args follow. 2480 We do include the last named arg if 2481 targetm.calls.strict_argument_naming() returns nonzero. 2482 (If no anonymous args follow, the result of list_length is actually 2483 one too large. This is harmless.) 2484 2485 If targetm.calls.pretend_outgoing_varargs_named() returns 2486 nonzero, and targetm.calls.strict_argument_naming() returns zero, 2487 this machine will be able to place unnamed args that were passed 2488 in registers into the stack. So treat all args as named. This 2489 allows the insns emitting for a specific argument list to be 2490 independent of the function declaration. 2491 2492 If targetm.calls.pretend_outgoing_varargs_named() returns zero, 2493 we do not have any reliable way to pass unnamed args in 2494 registers, so we must force them into memory. */ 2495 2496 if (type_arg_types != 0 2497 && targetm.calls.strict_argument_naming (args_so_far)) 2498 ; 2499 else if (type_arg_types != 0 2500 && ! targetm.calls.pretend_outgoing_varargs_named (args_so_far)) 2501 /* Don't include the last named arg. */ 2502 --n_named_args; 2503 else 2504 /* Treat all args as named. */ 2505 n_named_args = num_actuals; 2506 2507 /* Make a vector to hold all the information about each arg. */ 2508 args = XALLOCAVEC (struct arg_data, num_actuals); 2509 memset (args, 0, num_actuals * sizeof (struct arg_data)); 2510 2511 /* Build up entries in the ARGS array, compute the size of the 2512 arguments into ARGS_SIZE, etc. */ 2513 initialize_argument_information (num_actuals, args, &args_size, 2514 n_named_args, exp, 2515 structure_value_addr_value, fndecl, fntype, 2516 args_so_far, reg_parm_stack_space, 2517 &old_stack_level, &old_pending_adj, 2518 &must_preallocate, &flags, 2519 &try_tail_call, CALL_FROM_THUNK_P (exp)); 2520 2521 if (args_size.var) 2522 must_preallocate = 1; 2523 2524 /* Now make final decision about preallocating stack space. */ 2525 must_preallocate = finalize_must_preallocate (must_preallocate, 2526 num_actuals, args, 2527 &args_size); 2528 2529 /* If the structure value address will reference the stack pointer, we 2530 must stabilize it. We don't need to do this if we know that we are 2531 not going to adjust the stack pointer in processing this call. */ 2532 2533 if (structure_value_addr 2534 && (reg_mentioned_p (virtual_stack_dynamic_rtx, structure_value_addr) 2535 || reg_mentioned_p (virtual_outgoing_args_rtx, 2536 structure_value_addr)) 2537 && (args_size.var 2538 || (!ACCUMULATE_OUTGOING_ARGS && args_size.constant))) 2539 structure_value_addr = copy_to_reg (structure_value_addr); 2540 2541 /* Tail calls can make things harder to debug, and we've traditionally 2542 pushed these optimizations into -O2. Don't try if we're already 2543 expanding a call, as that means we're an argument. Don't try if 2544 there's cleanups, as we know there's code to follow the call. */ 2545 2546 if (currently_expanding_call++ != 0 2547 || !flag_optimize_sibling_calls 2548 || args_size.var 2549 || dbg_cnt (tail_call) == false) 2550 try_tail_call = 0; 2551 2552 /* Rest of purposes for tail call optimizations to fail. */ 2553 if ( 2554 #ifdef HAVE_sibcall_epilogue 2555 !HAVE_sibcall_epilogue 2556 #else 2557 1 2558 #endif 2559 || !try_tail_call 2560 /* Doing sibling call optimization needs some work, since 2561 structure_value_addr can be allocated on the stack. 2562 It does not seem worth the effort since few optimizable 2563 sibling calls will return a structure. */ 2564 || structure_value_addr != NULL_RTX 2565 #ifdef REG_PARM_STACK_SPACE 2566 /* If outgoing reg parm stack space changes, we can not do sibcall. */ 2567 || (OUTGOING_REG_PARM_STACK_SPACE (funtype) 2568 != OUTGOING_REG_PARM_STACK_SPACE (TREE_TYPE (current_function_decl))) 2569 || (reg_parm_stack_space != REG_PARM_STACK_SPACE (fndecl)) 2570 #endif 2571 /* Check whether the target is able to optimize the call 2572 into a sibcall. */ 2573 || !targetm.function_ok_for_sibcall (fndecl, exp) 2574 /* Functions that do not return exactly once may not be sibcall 2575 optimized. */ 2576 || (flags & (ECF_RETURNS_TWICE | ECF_NORETURN)) 2577 || TYPE_VOLATILE (TREE_TYPE (TREE_TYPE (addr))) 2578 /* If the called function is nested in the current one, it might access 2579 some of the caller's arguments, but could clobber them beforehand if 2580 the argument areas are shared. */ 2581 || (fndecl && decl_function_context (fndecl) == current_function_decl) 2582 /* If this function requires more stack slots than the current 2583 function, we cannot change it into a sibling call. 2584 crtl->args.pretend_args_size is not part of the 2585 stack allocated by our caller. */ 2586 || args_size.constant > (crtl->args.size 2587 - crtl->args.pretend_args_size) 2588 /* If the callee pops its own arguments, then it must pop exactly 2589 the same number of arguments as the current function. */ 2590 || (targetm.calls.return_pops_args (fndecl, funtype, args_size.constant) 2591 != targetm.calls.return_pops_args (current_function_decl, 2592 TREE_TYPE (current_function_decl), 2593 crtl->args.size)) 2594 || !lang_hooks.decls.ok_for_sibcall (fndecl)) 2595 try_tail_call = 0; 2596 2597 /* Check if caller and callee disagree in promotion of function 2598 return value. */ 2599 if (try_tail_call) 2600 { 2601 enum machine_mode caller_mode, caller_promoted_mode; 2602 enum machine_mode callee_mode, callee_promoted_mode; 2603 int caller_unsignedp, callee_unsignedp; 2604 tree caller_res = DECL_RESULT (current_function_decl); 2605 2606 caller_unsignedp = TYPE_UNSIGNED (TREE_TYPE (caller_res)); 2607 caller_mode = DECL_MODE (caller_res); 2608 callee_unsignedp = TYPE_UNSIGNED (TREE_TYPE (funtype)); 2609 callee_mode = TYPE_MODE (TREE_TYPE (funtype)); 2610 caller_promoted_mode 2611 = promote_function_mode (TREE_TYPE (caller_res), caller_mode, 2612 &caller_unsignedp, 2613 TREE_TYPE (current_function_decl), 1); 2614 callee_promoted_mode 2615 = promote_function_mode (TREE_TYPE (funtype), callee_mode, 2616 &callee_unsignedp, 2617 funtype, 1); 2618 if (caller_mode != VOIDmode 2619 && (caller_promoted_mode != callee_promoted_mode 2620 || ((caller_mode != caller_promoted_mode 2621 || callee_mode != callee_promoted_mode) 2622 && (caller_unsignedp != callee_unsignedp 2623 || GET_MODE_BITSIZE (caller_mode) 2624 < GET_MODE_BITSIZE (callee_mode))))) 2625 try_tail_call = 0; 2626 } 2627 2628 /* Ensure current function's preferred stack boundary is at least 2629 what we need. Stack alignment may also increase preferred stack 2630 boundary. */ 2631 if (crtl->preferred_stack_boundary < preferred_stack_boundary) 2632 crtl->preferred_stack_boundary = preferred_stack_boundary; 2633 else 2634 preferred_stack_boundary = crtl->preferred_stack_boundary; 2635 2636 preferred_unit_stack_boundary = preferred_stack_boundary / BITS_PER_UNIT; 2637 2638 /* We want to make two insn chains; one for a sibling call, the other 2639 for a normal call. We will select one of the two chains after 2640 initial RTL generation is complete. */ 2641 for (pass = try_tail_call ? 0 : 1; pass < 2; pass++) 2642 { 2643 int sibcall_failure = 0; 2644 /* We want to emit any pending stack adjustments before the tail 2645 recursion "call". That way we know any adjustment after the tail 2646 recursion call can be ignored if we indeed use the tail 2647 call expansion. */ 2648 int save_pending_stack_adjust = 0; 2649 int save_stack_pointer_delta = 0; 2650 rtx insns; 2651 rtx before_call, next_arg_reg, after_args; 2652 2653 if (pass == 0) 2654 { 2655 /* State variables we need to save and restore between 2656 iterations. */ 2657 save_pending_stack_adjust = pending_stack_adjust; 2658 save_stack_pointer_delta = stack_pointer_delta; 2659 } 2660 if (pass) 2661 flags &= ~ECF_SIBCALL; 2662 else 2663 flags |= ECF_SIBCALL; 2664 2665 /* Other state variables that we must reinitialize each time 2666 through the loop (that are not initialized by the loop itself). */ 2667 argblock = 0; 2668 call_fusage = 0; 2669 2670 /* Start a new sequence for the normal call case. 2671 2672 From this point on, if the sibling call fails, we want to set 2673 sibcall_failure instead of continuing the loop. */ 2674 start_sequence (); 2675 2676 /* Don't let pending stack adjusts add up to too much. 2677 Also, do all pending adjustments now if there is any chance 2678 this might be a call to alloca or if we are expanding a sibling 2679 call sequence. 2680 Also do the adjustments before a throwing call, otherwise 2681 exception handling can fail; PR 19225. */ 2682 if (pending_stack_adjust >= 32 2683 || (pending_stack_adjust > 0 2684 && (flags & ECF_MAY_BE_ALLOCA)) 2685 || (pending_stack_adjust > 0 2686 && flag_exceptions && !(flags & ECF_NOTHROW)) 2687 || pass == 0) 2688 do_pending_stack_adjust (); 2689 2690 /* Precompute any arguments as needed. */ 2691 if (pass) 2692 precompute_arguments (num_actuals, args); 2693 2694 /* Now we are about to start emitting insns that can be deleted 2695 if a libcall is deleted. */ 2696 if (pass && (flags & ECF_MALLOC)) 2697 start_sequence (); 2698 2699 if (pass == 0 && crtl->stack_protect_guard) 2700 stack_protect_epilogue (); 2701 2702 adjusted_args_size = args_size; 2703 /* Compute the actual size of the argument block required. The variable 2704 and constant sizes must be combined, the size may have to be rounded, 2705 and there may be a minimum required size. When generating a sibcall 2706 pattern, do not round up, since we'll be re-using whatever space our 2707 caller provided. */ 2708 unadjusted_args_size 2709 = compute_argument_block_size (reg_parm_stack_space, 2710 &adjusted_args_size, 2711 fndecl, fntype, 2712 (pass == 0 ? 0 2713 : preferred_stack_boundary)); 2714 2715 old_stack_allocated = stack_pointer_delta - pending_stack_adjust; 2716 2717 /* The argument block when performing a sibling call is the 2718 incoming argument block. */ 2719 if (pass == 0) 2720 { 2721 argblock = crtl->args.internal_arg_pointer; 2722 argblock 2723 #ifdef STACK_GROWS_DOWNWARD 2724 = plus_constant (argblock, crtl->args.pretend_args_size); 2725 #else 2726 = plus_constant (argblock, -crtl->args.pretend_args_size); 2727 #endif 2728 stored_args_map = sbitmap_alloc (args_size.constant); 2729 sbitmap_zero (stored_args_map); 2730 } 2731 2732 /* If we have no actual push instructions, or shouldn't use them, 2733 make space for all args right now. */ 2734 else if (adjusted_args_size.var != 0) 2735 { 2736 if (old_stack_level == 0) 2737 { 2738 emit_stack_save (SAVE_BLOCK, &old_stack_level); 2739 old_stack_pointer_delta = stack_pointer_delta; 2740 old_pending_adj = pending_stack_adjust; 2741 pending_stack_adjust = 0; 2742 /* stack_arg_under_construction says whether a stack arg is 2743 being constructed at the old stack level. Pushing the stack 2744 gets a clean outgoing argument block. */ 2745 old_stack_arg_under_construction = stack_arg_under_construction; 2746 stack_arg_under_construction = 0; 2747 } 2748 argblock = push_block (ARGS_SIZE_RTX (adjusted_args_size), 0, 0); 2749 if (flag_stack_usage_info) 2750 current_function_has_unbounded_dynamic_stack_size = 1; 2751 } 2752 else 2753 { 2754 /* Note that we must go through the motions of allocating an argument 2755 block even if the size is zero because we may be storing args 2756 in the area reserved for register arguments, which may be part of 2757 the stack frame. */ 2758 2759 int needed = adjusted_args_size.constant; 2760 2761 /* Store the maximum argument space used. It will be pushed by 2762 the prologue (if ACCUMULATE_OUTGOING_ARGS, or stack overflow 2763 checking). */ 2764 2765 if (needed > crtl->outgoing_args_size) 2766 crtl->outgoing_args_size = needed; 2767 2768 if (must_preallocate) 2769 { 2770 if (ACCUMULATE_OUTGOING_ARGS) 2771 { 2772 /* Since the stack pointer will never be pushed, it is 2773 possible for the evaluation of a parm to clobber 2774 something we have already written to the stack. 2775 Since most function calls on RISC machines do not use 2776 the stack, this is uncommon, but must work correctly. 2777 2778 Therefore, we save any area of the stack that was already 2779 written and that we are using. Here we set up to do this 2780 by making a new stack usage map from the old one. The 2781 actual save will be done by store_one_arg. 2782 2783 Another approach might be to try to reorder the argument 2784 evaluations to avoid this conflicting stack usage. */ 2785 2786 /* Since we will be writing into the entire argument area, 2787 the map must be allocated for its entire size, not just 2788 the part that is the responsibility of the caller. */ 2789 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) 2790 needed += reg_parm_stack_space; 2791 2792 #ifdef ARGS_GROW_DOWNWARD 2793 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use, 2794 needed + 1); 2795 #else 2796 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use, 2797 needed); 2798 #endif 2799 free (stack_usage_map_buf); 2800 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use); 2801 stack_usage_map = stack_usage_map_buf; 2802 2803 if (initial_highest_arg_in_use) 2804 memcpy (stack_usage_map, initial_stack_usage_map, 2805 initial_highest_arg_in_use); 2806 2807 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use) 2808 memset (&stack_usage_map[initial_highest_arg_in_use], 0, 2809 (highest_outgoing_arg_in_use 2810 - initial_highest_arg_in_use)); 2811 needed = 0; 2812 2813 /* The address of the outgoing argument list must not be 2814 copied to a register here, because argblock would be left 2815 pointing to the wrong place after the call to 2816 allocate_dynamic_stack_space below. */ 2817 2818 argblock = virtual_outgoing_args_rtx; 2819 } 2820 else 2821 { 2822 if (inhibit_defer_pop == 0) 2823 { 2824 /* Try to reuse some or all of the pending_stack_adjust 2825 to get this space. */ 2826 needed 2827 = (combine_pending_stack_adjustment_and_call 2828 (unadjusted_args_size, 2829 &adjusted_args_size, 2830 preferred_unit_stack_boundary)); 2831 2832 /* combine_pending_stack_adjustment_and_call computes 2833 an adjustment before the arguments are allocated. 2834 Account for them and see whether or not the stack 2835 needs to go up or down. */ 2836 needed = unadjusted_args_size - needed; 2837 2838 if (needed < 0) 2839 { 2840 /* We're releasing stack space. */ 2841 /* ??? We can avoid any adjustment at all if we're 2842 already aligned. FIXME. */ 2843 pending_stack_adjust = -needed; 2844 do_pending_stack_adjust (); 2845 needed = 0; 2846 } 2847 else 2848 /* We need to allocate space. We'll do that in 2849 push_block below. */ 2850 pending_stack_adjust = 0; 2851 } 2852 2853 /* Special case this because overhead of `push_block' in 2854 this case is non-trivial. */ 2855 if (needed == 0) 2856 argblock = virtual_outgoing_args_rtx; 2857 else 2858 { 2859 argblock = push_block (GEN_INT (needed), 0, 0); 2860 #ifdef ARGS_GROW_DOWNWARD 2861 argblock = plus_constant (argblock, needed); 2862 #endif 2863 } 2864 2865 /* We only really need to call `copy_to_reg' in the case 2866 where push insns are going to be used to pass ARGBLOCK 2867 to a function call in ARGS. In that case, the stack 2868 pointer changes value from the allocation point to the 2869 call point, and hence the value of 2870 VIRTUAL_OUTGOING_ARGS_RTX changes as well. But might 2871 as well always do it. */ 2872 argblock = copy_to_reg (argblock); 2873 } 2874 } 2875 } 2876 2877 if (ACCUMULATE_OUTGOING_ARGS) 2878 { 2879 /* The save/restore code in store_one_arg handles all 2880 cases except one: a constructor call (including a C 2881 function returning a BLKmode struct) to initialize 2882 an argument. */ 2883 if (stack_arg_under_construction) 2884 { 2885 rtx push_size 2886 = GEN_INT (adjusted_args_size.constant 2887 + (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype 2888 : TREE_TYPE (fndecl))) ? 0 2889 : reg_parm_stack_space)); 2890 if (old_stack_level == 0) 2891 { 2892 emit_stack_save (SAVE_BLOCK, &old_stack_level); 2893 old_stack_pointer_delta = stack_pointer_delta; 2894 old_pending_adj = pending_stack_adjust; 2895 pending_stack_adjust = 0; 2896 /* stack_arg_under_construction says whether a stack 2897 arg is being constructed at the old stack level. 2898 Pushing the stack gets a clean outgoing argument 2899 block. */ 2900 old_stack_arg_under_construction 2901 = stack_arg_under_construction; 2902 stack_arg_under_construction = 0; 2903 /* Make a new map for the new argument list. */ 2904 free (stack_usage_map_buf); 2905 stack_usage_map_buf = XCNEWVEC (char, highest_outgoing_arg_in_use); 2906 stack_usage_map = stack_usage_map_buf; 2907 highest_outgoing_arg_in_use = 0; 2908 } 2909 /* We can pass TRUE as the 4th argument because we just 2910 saved the stack pointer and will restore it right after 2911 the call. */ 2912 allocate_dynamic_stack_space (push_size, 0, 2913 BIGGEST_ALIGNMENT, true); 2914 } 2915 2916 /* If argument evaluation might modify the stack pointer, 2917 copy the address of the argument list to a register. */ 2918 for (i = 0; i < num_actuals; i++) 2919 if (args[i].pass_on_stack) 2920 { 2921 argblock = copy_addr_to_reg (argblock); 2922 break; 2923 } 2924 } 2925 2926 compute_argument_addresses (args, argblock, num_actuals); 2927 2928 /* If we push args individually in reverse order, perform stack alignment 2929 before the first push (the last arg). */ 2930 if (PUSH_ARGS_REVERSED && argblock == 0 2931 && adjusted_args_size.constant != unadjusted_args_size) 2932 { 2933 /* When the stack adjustment is pending, we get better code 2934 by combining the adjustments. */ 2935 if (pending_stack_adjust 2936 && ! inhibit_defer_pop) 2937 { 2938 pending_stack_adjust 2939 = (combine_pending_stack_adjustment_and_call 2940 (unadjusted_args_size, 2941 &adjusted_args_size, 2942 preferred_unit_stack_boundary)); 2943 do_pending_stack_adjust (); 2944 } 2945 else if (argblock == 0) 2946 anti_adjust_stack (GEN_INT (adjusted_args_size.constant 2947 - unadjusted_args_size)); 2948 } 2949 /* Now that the stack is properly aligned, pops can't safely 2950 be deferred during the evaluation of the arguments. */ 2951 NO_DEFER_POP; 2952 2953 /* Record the maximum pushed stack space size. We need to delay 2954 doing it this far to take into account the optimization done 2955 by combine_pending_stack_adjustment_and_call. */ 2956 if (flag_stack_usage_info 2957 && !ACCUMULATE_OUTGOING_ARGS 2958 && pass 2959 && adjusted_args_size.var == 0) 2960 { 2961 int pushed = adjusted_args_size.constant + pending_stack_adjust; 2962 if (pushed > current_function_pushed_stack_size) 2963 current_function_pushed_stack_size = pushed; 2964 } 2965 2966 funexp = rtx_for_function_call (fndecl, addr); 2967 2968 /* Figure out the register where the value, if any, will come back. */ 2969 valreg = 0; 2970 if (TYPE_MODE (rettype) != VOIDmode 2971 && ! structure_value_addr) 2972 { 2973 if (pcc_struct_value) 2974 valreg = hard_function_value (build_pointer_type (rettype), 2975 fndecl, NULL, (pass == 0)); 2976 else 2977 valreg = hard_function_value (rettype, fndecl, fntype, 2978 (pass == 0)); 2979 2980 /* If VALREG is a PARALLEL whose first member has a zero 2981 offset, use that. This is for targets such as m68k that 2982 return the same value in multiple places. */ 2983 if (GET_CODE (valreg) == PARALLEL) 2984 { 2985 rtx elem = XVECEXP (valreg, 0, 0); 2986 rtx where = XEXP (elem, 0); 2987 rtx offset = XEXP (elem, 1); 2988 if (offset == const0_rtx 2989 && GET_MODE (where) == GET_MODE (valreg)) 2990 valreg = where; 2991 } 2992 } 2993 2994 /* Precompute all register parameters. It isn't safe to compute anything 2995 once we have started filling any specific hard regs. */ 2996 precompute_register_parameters (num_actuals, args, ®_parm_seen); 2997 2998 if (CALL_EXPR_STATIC_CHAIN (exp)) 2999 static_chain_value = expand_normal (CALL_EXPR_STATIC_CHAIN (exp)); 3000 else 3001 static_chain_value = 0; 3002 3003 #ifdef REG_PARM_STACK_SPACE 3004 /* Save the fixed argument area if it's part of the caller's frame and 3005 is clobbered by argument setup for this call. */ 3006 if (ACCUMULATE_OUTGOING_ARGS && pass) 3007 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock, 3008 &low_to_save, &high_to_save); 3009 #endif 3010 3011 /* Now store (and compute if necessary) all non-register parms. 3012 These come before register parms, since they can require block-moves, 3013 which could clobber the registers used for register parms. 3014 Parms which have partial registers are not stored here, 3015 but we do preallocate space here if they want that. */ 3016 3017 for (i = 0; i < num_actuals; i++) 3018 { 3019 if (args[i].reg == 0 || args[i].pass_on_stack) 3020 { 3021 rtx before_arg = get_last_insn (); 3022 3023 if (store_one_arg (&args[i], argblock, flags, 3024 adjusted_args_size.var != 0, 3025 reg_parm_stack_space) 3026 || (pass == 0 3027 && check_sibcall_argument_overlap (before_arg, 3028 &args[i], 1))) 3029 sibcall_failure = 1; 3030 } 3031 3032 if (args[i].stack) 3033 call_fusage 3034 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[i].tree_value)), 3035 gen_rtx_USE (VOIDmode, args[i].stack), 3036 call_fusage); 3037 } 3038 3039 /* If we have a parm that is passed in registers but not in memory 3040 and whose alignment does not permit a direct copy into registers, 3041 make a group of pseudos that correspond to each register that we 3042 will later fill. */ 3043 if (STRICT_ALIGNMENT) 3044 store_unaligned_arguments_into_pseudos (args, num_actuals); 3045 3046 /* Now store any partially-in-registers parm. 3047 This is the last place a block-move can happen. */ 3048 if (reg_parm_seen) 3049 for (i = 0; i < num_actuals; i++) 3050 if (args[i].partial != 0 && ! args[i].pass_on_stack) 3051 { 3052 rtx before_arg = get_last_insn (); 3053 3054 if (store_one_arg (&args[i], argblock, flags, 3055 adjusted_args_size.var != 0, 3056 reg_parm_stack_space) 3057 || (pass == 0 3058 && check_sibcall_argument_overlap (before_arg, 3059 &args[i], 1))) 3060 sibcall_failure = 1; 3061 } 3062 3063 /* If we pushed args in forward order, perform stack alignment 3064 after pushing the last arg. */ 3065 if (!PUSH_ARGS_REVERSED && argblock == 0) 3066 anti_adjust_stack (GEN_INT (adjusted_args_size.constant 3067 - unadjusted_args_size)); 3068 3069 /* If register arguments require space on the stack and stack space 3070 was not preallocated, allocate stack space here for arguments 3071 passed in registers. */ 3072 if (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))) 3073 && !ACCUMULATE_OUTGOING_ARGS 3074 && must_preallocate == 0 && reg_parm_stack_space > 0) 3075 anti_adjust_stack (GEN_INT (reg_parm_stack_space)); 3076 3077 /* Pass the function the address in which to return a 3078 structure value. */ 3079 if (pass != 0 && structure_value_addr && ! structure_value_addr_parm) 3080 { 3081 structure_value_addr 3082 = convert_memory_address (Pmode, structure_value_addr); 3083 emit_move_insn (struct_value, 3084 force_reg (Pmode, 3085 force_operand (structure_value_addr, 3086 NULL_RTX))); 3087 3088 if (REG_P (struct_value)) 3089 use_reg (&call_fusage, struct_value); 3090 } 3091 3092 after_args = get_last_insn (); 3093 funexp = prepare_call_address (fndecl, funexp, static_chain_value, 3094 &call_fusage, reg_parm_seen, pass == 0); 3095 3096 load_register_parameters (args, num_actuals, &call_fusage, flags, 3097 pass == 0, &sibcall_failure); 3098 3099 /* Save a pointer to the last insn before the call, so that we can 3100 later safely search backwards to find the CALL_INSN. */ 3101 before_call = get_last_insn (); 3102 3103 /* Set up next argument register. For sibling calls on machines 3104 with register windows this should be the incoming register. */ 3105 if (pass == 0) 3106 next_arg_reg = targetm.calls.function_incoming_arg (args_so_far, 3107 VOIDmode, 3108 void_type_node, 3109 true); 3110 else 3111 next_arg_reg = targetm.calls.function_arg (args_so_far, 3112 VOIDmode, void_type_node, 3113 true); 3114 3115 /* All arguments and registers used for the call must be set up by 3116 now! */ 3117 3118 /* Stack must be properly aligned now. */ 3119 gcc_assert (!pass 3120 || !(stack_pointer_delta % preferred_unit_stack_boundary)); 3121 3122 /* Generate the actual call instruction. */ 3123 emit_call_1 (funexp, exp, fndecl, funtype, unadjusted_args_size, 3124 adjusted_args_size.constant, struct_value_size, 3125 next_arg_reg, valreg, old_inhibit_defer_pop, call_fusage, 3126 flags, args_so_far); 3127 3128 /* If the call setup or the call itself overlaps with anything 3129 of the argument setup we probably clobbered our call address. 3130 In that case we can't do sibcalls. */ 3131 if (pass == 0 3132 && check_sibcall_argument_overlap (after_args, 0, 0)) 3133 sibcall_failure = 1; 3134 3135 /* If a non-BLKmode value is returned at the most significant end 3136 of a register, shift the register right by the appropriate amount 3137 and update VALREG accordingly. BLKmode values are handled by the 3138 group load/store machinery below. */ 3139 if (!structure_value_addr 3140 && !pcc_struct_value 3141 && TYPE_MODE (rettype) != BLKmode 3142 && targetm.calls.return_in_msb (rettype)) 3143 { 3144 if (shift_return_value (TYPE_MODE (rettype), false, valreg)) 3145 sibcall_failure = 1; 3146 valreg = gen_rtx_REG (TYPE_MODE (rettype), REGNO (valreg)); 3147 } 3148 3149 if (pass && (flags & ECF_MALLOC)) 3150 { 3151 rtx temp = gen_reg_rtx (GET_MODE (valreg)); 3152 rtx last, insns; 3153 3154 /* The return value from a malloc-like function is a pointer. */ 3155 if (TREE_CODE (rettype) == POINTER_TYPE) 3156 mark_reg_pointer (temp, BIGGEST_ALIGNMENT); 3157 3158 emit_move_insn (temp, valreg); 3159 3160 /* The return value from a malloc-like function can not alias 3161 anything else. */ 3162 last = get_last_insn (); 3163 add_reg_note (last, REG_NOALIAS, temp); 3164 3165 /* Write out the sequence. */ 3166 insns = get_insns (); 3167 end_sequence (); 3168 emit_insn (insns); 3169 valreg = temp; 3170 } 3171 3172 /* For calls to `setjmp', etc., inform 3173 function.c:setjmp_warnings that it should complain if 3174 nonvolatile values are live. For functions that cannot 3175 return, inform flow that control does not fall through. */ 3176 3177 if ((flags & ECF_NORETURN) || pass == 0) 3178 { 3179 /* The barrier must be emitted 3180 immediately after the CALL_INSN. Some ports emit more 3181 than just a CALL_INSN above, so we must search for it here. */ 3182 3183 rtx last = get_last_insn (); 3184 while (!CALL_P (last)) 3185 { 3186 last = PREV_INSN (last); 3187 /* There was no CALL_INSN? */ 3188 gcc_assert (last != before_call); 3189 } 3190 3191 emit_barrier_after (last); 3192 3193 /* Stack adjustments after a noreturn call are dead code. 3194 However when NO_DEFER_POP is in effect, we must preserve 3195 stack_pointer_delta. */ 3196 if (inhibit_defer_pop == 0) 3197 { 3198 stack_pointer_delta = old_stack_allocated; 3199 pending_stack_adjust = 0; 3200 } 3201 } 3202 3203 /* If value type not void, return an rtx for the value. */ 3204 3205 if (TYPE_MODE (rettype) == VOIDmode 3206 || ignore) 3207 target = const0_rtx; 3208 else if (structure_value_addr) 3209 { 3210 if (target == 0 || !MEM_P (target)) 3211 { 3212 target 3213 = gen_rtx_MEM (TYPE_MODE (rettype), 3214 memory_address (TYPE_MODE (rettype), 3215 structure_value_addr)); 3216 set_mem_attributes (target, rettype, 1); 3217 } 3218 } 3219 else if (pcc_struct_value) 3220 { 3221 /* This is the special C++ case where we need to 3222 know what the true target was. We take care to 3223 never use this value more than once in one expression. */ 3224 target = gen_rtx_MEM (TYPE_MODE (rettype), 3225 copy_to_reg (valreg)); 3226 set_mem_attributes (target, rettype, 1); 3227 } 3228 /* Handle calls that return values in multiple non-contiguous locations. 3229 The Irix 6 ABI has examples of this. */ 3230 else if (GET_CODE (valreg) == PARALLEL) 3231 { 3232 if (target == 0) 3233 { 3234 /* This will only be assigned once, so it can be readonly. */ 3235 tree nt = build_qualified_type (rettype, 3236 (TYPE_QUALS (rettype) 3237 | TYPE_QUAL_CONST)); 3238 3239 target = assign_temp (nt, 0, 1, 1); 3240 } 3241 3242 if (! rtx_equal_p (target, valreg)) 3243 emit_group_store (target, valreg, rettype, 3244 int_size_in_bytes (rettype)); 3245 3246 /* We can not support sibling calls for this case. */ 3247 sibcall_failure = 1; 3248 } 3249 else if (target 3250 && GET_MODE (target) == TYPE_MODE (rettype) 3251 && GET_MODE (target) == GET_MODE (valreg)) 3252 { 3253 bool may_overlap = false; 3254 3255 /* We have to copy a return value in a CLASS_LIKELY_SPILLED hard 3256 reg to a plain register. */ 3257 if (!REG_P (target) || HARD_REGISTER_P (target)) 3258 valreg = avoid_likely_spilled_reg (valreg); 3259 3260 /* If TARGET is a MEM in the argument area, and we have 3261 saved part of the argument area, then we can't store 3262 directly into TARGET as it may get overwritten when we 3263 restore the argument save area below. Don't work too 3264 hard though and simply force TARGET to a register if it 3265 is a MEM; the optimizer is quite likely to sort it out. */ 3266 if (ACCUMULATE_OUTGOING_ARGS && pass && MEM_P (target)) 3267 for (i = 0; i < num_actuals; i++) 3268 if (args[i].save_area) 3269 { 3270 may_overlap = true; 3271 break; 3272 } 3273 3274 if (may_overlap) 3275 target = copy_to_reg (valreg); 3276 else 3277 { 3278 /* TARGET and VALREG cannot be equal at this point 3279 because the latter would not have 3280 REG_FUNCTION_VALUE_P true, while the former would if 3281 it were referring to the same register. 3282 3283 If they refer to the same register, this move will be 3284 a no-op, except when function inlining is being 3285 done. */ 3286 emit_move_insn (target, valreg); 3287 3288 /* If we are setting a MEM, this code must be executed. 3289 Since it is emitted after the call insn, sibcall 3290 optimization cannot be performed in that case. */ 3291 if (MEM_P (target)) 3292 sibcall_failure = 1; 3293 } 3294 } 3295 else if (TYPE_MODE (rettype) == BLKmode) 3296 { 3297 rtx val = valreg; 3298 if (GET_MODE (val) != BLKmode) 3299 val = avoid_likely_spilled_reg (val); 3300 target = copy_blkmode_from_reg (target, val, rettype); 3301 3302 /* We can not support sibling calls for this case. */ 3303 sibcall_failure = 1; 3304 } 3305 else 3306 target = copy_to_reg (avoid_likely_spilled_reg (valreg)); 3307 3308 /* If we promoted this return value, make the proper SUBREG. 3309 TARGET might be const0_rtx here, so be careful. */ 3310 if (REG_P (target) 3311 && TYPE_MODE (rettype) != BLKmode 3312 && GET_MODE (target) != TYPE_MODE (rettype)) 3313 { 3314 tree type = rettype; 3315 int unsignedp = TYPE_UNSIGNED (type); 3316 int offset = 0; 3317 enum machine_mode pmode; 3318 3319 /* Ensure we promote as expected, and get the new unsignedness. */ 3320 pmode = promote_function_mode (type, TYPE_MODE (type), &unsignedp, 3321 funtype, 1); 3322 gcc_assert (GET_MODE (target) == pmode); 3323 3324 if ((WORDS_BIG_ENDIAN || BYTES_BIG_ENDIAN) 3325 && (GET_MODE_SIZE (GET_MODE (target)) 3326 > GET_MODE_SIZE (TYPE_MODE (type)))) 3327 { 3328 offset = GET_MODE_SIZE (GET_MODE (target)) 3329 - GET_MODE_SIZE (TYPE_MODE (type)); 3330 if (! BYTES_BIG_ENDIAN) 3331 offset = (offset / UNITS_PER_WORD) * UNITS_PER_WORD; 3332 else if (! WORDS_BIG_ENDIAN) 3333 offset %= UNITS_PER_WORD; 3334 } 3335 3336 target = gen_rtx_SUBREG (TYPE_MODE (type), target, offset); 3337 SUBREG_PROMOTED_VAR_P (target) = 1; 3338 SUBREG_PROMOTED_UNSIGNED_SET (target, unsignedp); 3339 } 3340 3341 /* If size of args is variable or this was a constructor call for a stack 3342 argument, restore saved stack-pointer value. */ 3343 3344 if (old_stack_level) 3345 { 3346 rtx prev = get_last_insn (); 3347 3348 emit_stack_restore (SAVE_BLOCK, old_stack_level); 3349 stack_pointer_delta = old_stack_pointer_delta; 3350 3351 fixup_args_size_notes (prev, get_last_insn (), stack_pointer_delta); 3352 3353 pending_stack_adjust = old_pending_adj; 3354 old_stack_allocated = stack_pointer_delta - pending_stack_adjust; 3355 stack_arg_under_construction = old_stack_arg_under_construction; 3356 highest_outgoing_arg_in_use = initial_highest_arg_in_use; 3357 stack_usage_map = initial_stack_usage_map; 3358 sibcall_failure = 1; 3359 } 3360 else if (ACCUMULATE_OUTGOING_ARGS && pass) 3361 { 3362 #ifdef REG_PARM_STACK_SPACE 3363 if (save_area) 3364 restore_fixed_argument_area (save_area, argblock, 3365 high_to_save, low_to_save); 3366 #endif 3367 3368 /* If we saved any argument areas, restore them. */ 3369 for (i = 0; i < num_actuals; i++) 3370 if (args[i].save_area) 3371 { 3372 enum machine_mode save_mode = GET_MODE (args[i].save_area); 3373 rtx stack_area 3374 = gen_rtx_MEM (save_mode, 3375 memory_address (save_mode, 3376 XEXP (args[i].stack_slot, 0))); 3377 3378 if (save_mode != BLKmode) 3379 emit_move_insn (stack_area, args[i].save_area); 3380 else 3381 emit_block_move (stack_area, args[i].save_area, 3382 GEN_INT (args[i].locate.size.constant), 3383 BLOCK_OP_CALL_PARM); 3384 } 3385 3386 highest_outgoing_arg_in_use = initial_highest_arg_in_use; 3387 stack_usage_map = initial_stack_usage_map; 3388 } 3389 3390 /* If this was alloca, record the new stack level for nonlocal gotos. 3391 Check for the handler slots since we might not have a save area 3392 for non-local gotos. */ 3393 3394 if ((flags & ECF_MAY_BE_ALLOCA) && cfun->nonlocal_goto_save_area != 0) 3395 update_nonlocal_goto_save_area (); 3396 3397 /* Free up storage we no longer need. */ 3398 for (i = 0; i < num_actuals; ++i) 3399 free (args[i].aligned_regs); 3400 3401 insns = get_insns (); 3402 end_sequence (); 3403 3404 if (pass == 0) 3405 { 3406 tail_call_insns = insns; 3407 3408 /* Restore the pending stack adjustment now that we have 3409 finished generating the sibling call sequence. */ 3410 3411 pending_stack_adjust = save_pending_stack_adjust; 3412 stack_pointer_delta = save_stack_pointer_delta; 3413 3414 /* Prepare arg structure for next iteration. */ 3415 for (i = 0; i < num_actuals; i++) 3416 { 3417 args[i].value = 0; 3418 args[i].aligned_regs = 0; 3419 args[i].stack = 0; 3420 } 3421 3422 sbitmap_free (stored_args_map); 3423 internal_arg_pointer_exp_state.scan_start = NULL_RTX; 3424 VEC_free (rtx, heap, internal_arg_pointer_exp_state.cache); 3425 } 3426 else 3427 { 3428 normal_call_insns = insns; 3429 3430 /* Verify that we've deallocated all the stack we used. */ 3431 gcc_assert ((flags & ECF_NORETURN) 3432 || (old_stack_allocated 3433 == stack_pointer_delta - pending_stack_adjust)); 3434 } 3435 3436 /* If something prevents making this a sibling call, 3437 zero out the sequence. */ 3438 if (sibcall_failure) 3439 tail_call_insns = NULL_RTX; 3440 else 3441 break; 3442 } 3443 3444 /* If tail call production succeeded, we need to remove REG_EQUIV notes on 3445 arguments too, as argument area is now clobbered by the call. */ 3446 if (tail_call_insns) 3447 { 3448 emit_insn (tail_call_insns); 3449 crtl->tail_call_emit = true; 3450 } 3451 else 3452 emit_insn (normal_call_insns); 3453 3454 currently_expanding_call--; 3455 3456 free (stack_usage_map_buf); 3457 3458 return target; 3459 } 3460 3461 /* A sibling call sequence invalidates any REG_EQUIV notes made for 3462 this function's incoming arguments. 3463 3464 At the start of RTL generation we know the only REG_EQUIV notes 3465 in the rtl chain are those for incoming arguments, so we can look 3466 for REG_EQUIV notes between the start of the function and the 3467 NOTE_INSN_FUNCTION_BEG. 3468 3469 This is (slight) overkill. We could keep track of the highest 3470 argument we clobber and be more selective in removing notes, but it 3471 does not seem to be worth the effort. */ 3472 3473 void 3474 fixup_tail_calls (void) 3475 { 3476 rtx insn; 3477 3478 for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) 3479 { 3480 rtx note; 3481 3482 /* There are never REG_EQUIV notes for the incoming arguments 3483 after the NOTE_INSN_FUNCTION_BEG note, so stop if we see it. */ 3484 if (NOTE_P (insn) 3485 && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG) 3486 break; 3487 3488 note = find_reg_note (insn, REG_EQUIV, 0); 3489 if (note) 3490 remove_note (insn, note); 3491 note = find_reg_note (insn, REG_EQUIV, 0); 3492 gcc_assert (!note); 3493 } 3494 } 3495 3496 /* Traverse a list of TYPES and expand all complex types into their 3497 components. */ 3498 static tree 3499 split_complex_types (tree types) 3500 { 3501 tree p; 3502 3503 /* Before allocating memory, check for the common case of no complex. */ 3504 for (p = types; p; p = TREE_CHAIN (p)) 3505 { 3506 tree type = TREE_VALUE (p); 3507 if (TREE_CODE (type) == COMPLEX_TYPE 3508 && targetm.calls.split_complex_arg (type)) 3509 goto found; 3510 } 3511 return types; 3512 3513 found: 3514 types = copy_list (types); 3515 3516 for (p = types; p; p = TREE_CHAIN (p)) 3517 { 3518 tree complex_type = TREE_VALUE (p); 3519 3520 if (TREE_CODE (complex_type) == COMPLEX_TYPE 3521 && targetm.calls.split_complex_arg (complex_type)) 3522 { 3523 tree next, imag; 3524 3525 /* Rewrite complex type with component type. */ 3526 TREE_VALUE (p) = TREE_TYPE (complex_type); 3527 next = TREE_CHAIN (p); 3528 3529 /* Add another component type for the imaginary part. */ 3530 imag = build_tree_list (NULL_TREE, TREE_VALUE (p)); 3531 TREE_CHAIN (p) = imag; 3532 TREE_CHAIN (imag) = next; 3533 3534 /* Skip the newly created node. */ 3535 p = TREE_CHAIN (p); 3536 } 3537 } 3538 3539 return types; 3540 } 3541 3542 /* Output a library call to function FUN (a SYMBOL_REF rtx). 3543 The RETVAL parameter specifies whether return value needs to be saved, other 3544 parameters are documented in the emit_library_call function below. */ 3545 3546 static rtx 3547 emit_library_call_value_1 (int retval, rtx orgfun, rtx value, 3548 enum libcall_type fn_type, 3549 enum machine_mode outmode, int nargs, va_list p) 3550 { 3551 /* Total size in bytes of all the stack-parms scanned so far. */ 3552 struct args_size args_size; 3553 /* Size of arguments before any adjustments (such as rounding). */ 3554 struct args_size original_args_size; 3555 int argnum; 3556 rtx fun; 3557 /* Todo, choose the correct decl type of orgfun. Sadly this information 3558 isn't present here, so we default to native calling abi here. */ 3559 tree fndecl ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */ 3560 tree fntype ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */ 3561 int inc; 3562 int count; 3563 rtx argblock = 0; 3564 CUMULATIVE_ARGS args_so_far_v; 3565 cumulative_args_t args_so_far; 3566 struct arg 3567 { 3568 rtx value; 3569 enum machine_mode mode; 3570 rtx reg; 3571 int partial; 3572 struct locate_and_pad_arg_data locate; 3573 rtx save_area; 3574 }; 3575 struct arg *argvec; 3576 int old_inhibit_defer_pop = inhibit_defer_pop; 3577 rtx call_fusage = 0; 3578 rtx mem_value = 0; 3579 rtx valreg; 3580 int pcc_struct_value = 0; 3581 int struct_value_size = 0; 3582 int flags; 3583 int reg_parm_stack_space = 0; 3584 int needed; 3585 rtx before_call; 3586 tree tfom; /* type_for_mode (outmode, 0) */ 3587 3588 #ifdef REG_PARM_STACK_SPACE 3589 /* Define the boundary of the register parm stack space that needs to be 3590 save, if any. */ 3591 int low_to_save = 0, high_to_save = 0; 3592 rtx save_area = 0; /* Place that it is saved. */ 3593 #endif 3594 3595 /* Size of the stack reserved for parameter registers. */ 3596 int initial_highest_arg_in_use = highest_outgoing_arg_in_use; 3597 char *initial_stack_usage_map = stack_usage_map; 3598 char *stack_usage_map_buf = NULL; 3599 3600 rtx struct_value = targetm.calls.struct_value_rtx (0, 0); 3601 3602 #ifdef REG_PARM_STACK_SPACE 3603 reg_parm_stack_space = REG_PARM_STACK_SPACE ((tree) 0); 3604 #endif 3605 3606 /* By default, library functions can not throw. */ 3607 flags = ECF_NOTHROW; 3608 3609 switch (fn_type) 3610 { 3611 case LCT_NORMAL: 3612 break; 3613 case LCT_CONST: 3614 flags |= ECF_CONST; 3615 break; 3616 case LCT_PURE: 3617 flags |= ECF_PURE; 3618 break; 3619 case LCT_NORETURN: 3620 flags |= ECF_NORETURN; 3621 break; 3622 case LCT_THROW: 3623 flags = ECF_NORETURN; 3624 break; 3625 case LCT_RETURNS_TWICE: 3626 flags = ECF_RETURNS_TWICE; 3627 break; 3628 } 3629 fun = orgfun; 3630 3631 /* Ensure current function's preferred stack boundary is at least 3632 what we need. */ 3633 if (crtl->preferred_stack_boundary < PREFERRED_STACK_BOUNDARY) 3634 crtl->preferred_stack_boundary = PREFERRED_STACK_BOUNDARY; 3635 3636 /* If this kind of value comes back in memory, 3637 decide where in memory it should come back. */ 3638 if (outmode != VOIDmode) 3639 { 3640 tfom = lang_hooks.types.type_for_mode (outmode, 0); 3641 if (aggregate_value_p (tfom, 0)) 3642 { 3643 #ifdef PCC_STATIC_STRUCT_RETURN 3644 rtx pointer_reg 3645 = hard_function_value (build_pointer_type (tfom), 0, 0, 0); 3646 mem_value = gen_rtx_MEM (outmode, pointer_reg); 3647 pcc_struct_value = 1; 3648 if (value == 0) 3649 value = gen_reg_rtx (outmode); 3650 #else /* not PCC_STATIC_STRUCT_RETURN */ 3651 struct_value_size = GET_MODE_SIZE (outmode); 3652 if (value != 0 && MEM_P (value)) 3653 mem_value = value; 3654 else 3655 mem_value = assign_temp (tfom, 0, 1, 1); 3656 #endif 3657 /* This call returns a big structure. */ 3658 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE); 3659 } 3660 } 3661 else 3662 tfom = void_type_node; 3663 3664 /* ??? Unfinished: must pass the memory address as an argument. */ 3665 3666 /* Copy all the libcall-arguments out of the varargs data 3667 and into a vector ARGVEC. 3668 3669 Compute how to pass each argument. We only support a very small subset 3670 of the full argument passing conventions to limit complexity here since 3671 library functions shouldn't have many args. */ 3672 3673 argvec = XALLOCAVEC (struct arg, nargs + 1); 3674 memset (argvec, 0, (nargs + 1) * sizeof (struct arg)); 3675 3676 #ifdef INIT_CUMULATIVE_LIBCALL_ARGS 3677 INIT_CUMULATIVE_LIBCALL_ARGS (args_so_far_v, outmode, fun); 3678 #else 3679 INIT_CUMULATIVE_ARGS (args_so_far_v, NULL_TREE, fun, 0, nargs); 3680 #endif 3681 args_so_far = pack_cumulative_args (&args_so_far_v); 3682 3683 args_size.constant = 0; 3684 args_size.var = 0; 3685 3686 count = 0; 3687 3688 push_temp_slots (); 3689 3690 /* If there's a structure value address to be passed, 3691 either pass it in the special place, or pass it as an extra argument. */ 3692 if (mem_value && struct_value == 0 && ! pcc_struct_value) 3693 { 3694 rtx addr = XEXP (mem_value, 0); 3695 3696 nargs++; 3697 3698 /* Make sure it is a reasonable operand for a move or push insn. */ 3699 if (!REG_P (addr) && !MEM_P (addr) 3700 && !(CONSTANT_P (addr) 3701 && targetm.legitimate_constant_p (Pmode, addr))) 3702 addr = force_operand (addr, NULL_RTX); 3703 3704 argvec[count].value = addr; 3705 argvec[count].mode = Pmode; 3706 argvec[count].partial = 0; 3707 3708 argvec[count].reg = targetm.calls.function_arg (args_so_far, 3709 Pmode, NULL_TREE, true); 3710 gcc_assert (targetm.calls.arg_partial_bytes (args_so_far, Pmode, 3711 NULL_TREE, 1) == 0); 3712 3713 locate_and_pad_parm (Pmode, NULL_TREE, 3714 #ifdef STACK_PARMS_IN_REG_PARM_AREA 3715 1, 3716 #else 3717 argvec[count].reg != 0, 3718 #endif 3719 0, NULL_TREE, &args_size, &argvec[count].locate); 3720 3721 if (argvec[count].reg == 0 || argvec[count].partial != 0 3722 || reg_parm_stack_space > 0) 3723 args_size.constant += argvec[count].locate.size.constant; 3724 3725 targetm.calls.function_arg_advance (args_so_far, Pmode, (tree) 0, true); 3726 3727 count++; 3728 } 3729 3730 for (; count < nargs; count++) 3731 { 3732 rtx val = va_arg (p, rtx); 3733 enum machine_mode mode = (enum machine_mode) va_arg (p, int); 3734 int unsigned_p = 0; 3735 3736 /* We cannot convert the arg value to the mode the library wants here; 3737 must do it earlier where we know the signedness of the arg. */ 3738 gcc_assert (mode != BLKmode 3739 && (GET_MODE (val) == mode || GET_MODE (val) == VOIDmode)); 3740 3741 /* Make sure it is a reasonable operand for a move or push insn. */ 3742 if (!REG_P (val) && !MEM_P (val) 3743 && !(CONSTANT_P (val) && targetm.legitimate_constant_p (mode, val))) 3744 val = force_operand (val, NULL_RTX); 3745 3746 if (pass_by_reference (&args_so_far_v, mode, NULL_TREE, 1)) 3747 { 3748 rtx slot; 3749 int must_copy 3750 = !reference_callee_copied (&args_so_far_v, mode, NULL_TREE, 1); 3751 3752 /* If this was a CONST function, it is now PURE since it now 3753 reads memory. */ 3754 if (flags & ECF_CONST) 3755 { 3756 flags &= ~ECF_CONST; 3757 flags |= ECF_PURE; 3758 } 3759 3760 if (MEM_P (val) && !must_copy) 3761 { 3762 tree val_expr = MEM_EXPR (val); 3763 if (val_expr) 3764 mark_addressable (val_expr); 3765 slot = val; 3766 } 3767 else 3768 { 3769 slot = assign_temp (lang_hooks.types.type_for_mode (mode, 0), 3770 0, 1, 1); 3771 emit_move_insn (slot, val); 3772 } 3773 3774 call_fusage = gen_rtx_EXPR_LIST (VOIDmode, 3775 gen_rtx_USE (VOIDmode, slot), 3776 call_fusage); 3777 if (must_copy) 3778 call_fusage = gen_rtx_EXPR_LIST (VOIDmode, 3779 gen_rtx_CLOBBER (VOIDmode, 3780 slot), 3781 call_fusage); 3782 3783 mode = Pmode; 3784 val = force_operand (XEXP (slot, 0), NULL_RTX); 3785 } 3786 3787 mode = promote_function_mode (NULL_TREE, mode, &unsigned_p, NULL_TREE, 0); 3788 argvec[count].mode = mode; 3789 argvec[count].value = convert_modes (mode, GET_MODE (val), val, unsigned_p); 3790 argvec[count].reg = targetm.calls.function_arg (args_so_far, mode, 3791 NULL_TREE, true); 3792 3793 argvec[count].partial 3794 = targetm.calls.arg_partial_bytes (args_so_far, mode, NULL_TREE, 1); 3795 3796 if (argvec[count].reg == 0 3797 || argvec[count].partial != 0 3798 || reg_parm_stack_space > 0) 3799 { 3800 locate_and_pad_parm (mode, NULL_TREE, 3801 #ifdef STACK_PARMS_IN_REG_PARM_AREA 3802 1, 3803 #else 3804 argvec[count].reg != 0, 3805 #endif 3806 argvec[count].partial, 3807 NULL_TREE, &args_size, &argvec[count].locate); 3808 args_size.constant += argvec[count].locate.size.constant; 3809 gcc_assert (!argvec[count].locate.size.var); 3810 } 3811 #ifdef BLOCK_REG_PADDING 3812 else 3813 /* The argument is passed entirely in registers. See at which 3814 end it should be padded. */ 3815 argvec[count].locate.where_pad = 3816 BLOCK_REG_PADDING (mode, NULL_TREE, 3817 GET_MODE_SIZE (mode) <= UNITS_PER_WORD); 3818 #endif 3819 3820 targetm.calls.function_arg_advance (args_so_far, mode, (tree) 0, true); 3821 } 3822 3823 /* If this machine requires an external definition for library 3824 functions, write one out. */ 3825 assemble_external_libcall (fun); 3826 3827 original_args_size = args_size; 3828 args_size.constant = (((args_size.constant 3829 + stack_pointer_delta 3830 + STACK_BYTES - 1) 3831 / STACK_BYTES 3832 * STACK_BYTES) 3833 - stack_pointer_delta); 3834 3835 args_size.constant = MAX (args_size.constant, 3836 reg_parm_stack_space); 3837 3838 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) 3839 args_size.constant -= reg_parm_stack_space; 3840 3841 if (args_size.constant > crtl->outgoing_args_size) 3842 crtl->outgoing_args_size = args_size.constant; 3843 3844 if (flag_stack_usage_info && !ACCUMULATE_OUTGOING_ARGS) 3845 { 3846 int pushed = args_size.constant + pending_stack_adjust; 3847 if (pushed > current_function_pushed_stack_size) 3848 current_function_pushed_stack_size = pushed; 3849 } 3850 3851 if (ACCUMULATE_OUTGOING_ARGS) 3852 { 3853 /* Since the stack pointer will never be pushed, it is possible for 3854 the evaluation of a parm to clobber something we have already 3855 written to the stack. Since most function calls on RISC machines 3856 do not use the stack, this is uncommon, but must work correctly. 3857 3858 Therefore, we save any area of the stack that was already written 3859 and that we are using. Here we set up to do this by making a new 3860 stack usage map from the old one. 3861 3862 Another approach might be to try to reorder the argument 3863 evaluations to avoid this conflicting stack usage. */ 3864 3865 needed = args_size.constant; 3866 3867 /* Since we will be writing into the entire argument area, the 3868 map must be allocated for its entire size, not just the part that 3869 is the responsibility of the caller. */ 3870 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))) 3871 needed += reg_parm_stack_space; 3872 3873 #ifdef ARGS_GROW_DOWNWARD 3874 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use, 3875 needed + 1); 3876 #else 3877 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use, 3878 needed); 3879 #endif 3880 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use); 3881 stack_usage_map = stack_usage_map_buf; 3882 3883 if (initial_highest_arg_in_use) 3884 memcpy (stack_usage_map, initial_stack_usage_map, 3885 initial_highest_arg_in_use); 3886 3887 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use) 3888 memset (&stack_usage_map[initial_highest_arg_in_use], 0, 3889 highest_outgoing_arg_in_use - initial_highest_arg_in_use); 3890 needed = 0; 3891 3892 /* We must be careful to use virtual regs before they're instantiated, 3893 and real regs afterwards. Loop optimization, for example, can create 3894 new libcalls after we've instantiated the virtual regs, and if we 3895 use virtuals anyway, they won't match the rtl patterns. */ 3896 3897 if (virtuals_instantiated) 3898 argblock = plus_constant (stack_pointer_rtx, STACK_POINTER_OFFSET); 3899 else 3900 argblock = virtual_outgoing_args_rtx; 3901 } 3902 else 3903 { 3904 if (!PUSH_ARGS) 3905 argblock = push_block (GEN_INT (args_size.constant), 0, 0); 3906 } 3907 3908 /* If we push args individually in reverse order, perform stack alignment 3909 before the first push (the last arg). */ 3910 if (argblock == 0 && PUSH_ARGS_REVERSED) 3911 anti_adjust_stack (GEN_INT (args_size.constant 3912 - original_args_size.constant)); 3913 3914 if (PUSH_ARGS_REVERSED) 3915 { 3916 inc = -1; 3917 argnum = nargs - 1; 3918 } 3919 else 3920 { 3921 inc = 1; 3922 argnum = 0; 3923 } 3924 3925 #ifdef REG_PARM_STACK_SPACE 3926 if (ACCUMULATE_OUTGOING_ARGS) 3927 { 3928 /* The argument list is the property of the called routine and it 3929 may clobber it. If the fixed area has been used for previous 3930 parameters, we must save and restore it. */ 3931 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock, 3932 &low_to_save, &high_to_save); 3933 } 3934 #endif 3935 3936 /* Push the args that need to be pushed. */ 3937 3938 /* ARGNUM indexes the ARGVEC array in the order in which the arguments 3939 are to be pushed. */ 3940 for (count = 0; count < nargs; count++, argnum += inc) 3941 { 3942 enum machine_mode mode = argvec[argnum].mode; 3943 rtx val = argvec[argnum].value; 3944 rtx reg = argvec[argnum].reg; 3945 int partial = argvec[argnum].partial; 3946 unsigned int parm_align = argvec[argnum].locate.boundary; 3947 int lower_bound = 0, upper_bound = 0, i; 3948 3949 if (! (reg != 0 && partial == 0)) 3950 { 3951 rtx use; 3952 3953 if (ACCUMULATE_OUTGOING_ARGS) 3954 { 3955 /* If this is being stored into a pre-allocated, fixed-size, 3956 stack area, save any previous data at that location. */ 3957 3958 #ifdef ARGS_GROW_DOWNWARD 3959 /* stack_slot is negative, but we want to index stack_usage_map 3960 with positive values. */ 3961 upper_bound = -argvec[argnum].locate.slot_offset.constant + 1; 3962 lower_bound = upper_bound - argvec[argnum].locate.size.constant; 3963 #else 3964 lower_bound = argvec[argnum].locate.slot_offset.constant; 3965 upper_bound = lower_bound + argvec[argnum].locate.size.constant; 3966 #endif 3967 3968 i = lower_bound; 3969 /* Don't worry about things in the fixed argument area; 3970 it has already been saved. */ 3971 if (i < reg_parm_stack_space) 3972 i = reg_parm_stack_space; 3973 while (i < upper_bound && stack_usage_map[i] == 0) 3974 i++; 3975 3976 if (i < upper_bound) 3977 { 3978 /* We need to make a save area. */ 3979 unsigned int size 3980 = argvec[argnum].locate.size.constant * BITS_PER_UNIT; 3981 enum machine_mode save_mode 3982 = mode_for_size (size, MODE_INT, 1); 3983 rtx adr 3984 = plus_constant (argblock, 3985 argvec[argnum].locate.offset.constant); 3986 rtx stack_area 3987 = gen_rtx_MEM (save_mode, memory_address (save_mode, adr)); 3988 3989 if (save_mode == BLKmode) 3990 { 3991 argvec[argnum].save_area 3992 = assign_stack_temp (BLKmode, 3993 argvec[argnum].locate.size.constant, 3994 0); 3995 3996 emit_block_move (validize_mem (argvec[argnum].save_area), 3997 stack_area, 3998 GEN_INT (argvec[argnum].locate.size.constant), 3999 BLOCK_OP_CALL_PARM); 4000 } 4001 else 4002 { 4003 argvec[argnum].save_area = gen_reg_rtx (save_mode); 4004 4005 emit_move_insn (argvec[argnum].save_area, stack_area); 4006 } 4007 } 4008 } 4009 4010 emit_push_insn (val, mode, NULL_TREE, NULL_RTX, parm_align, 4011 partial, reg, 0, argblock, 4012 GEN_INT (argvec[argnum].locate.offset.constant), 4013 reg_parm_stack_space, 4014 ARGS_SIZE_RTX (argvec[argnum].locate.alignment_pad)); 4015 4016 /* Now mark the segment we just used. */ 4017 if (ACCUMULATE_OUTGOING_ARGS) 4018 for (i = lower_bound; i < upper_bound; i++) 4019 stack_usage_map[i] = 1; 4020 4021 NO_DEFER_POP; 4022 4023 /* Indicate argument access so that alias.c knows that these 4024 values are live. */ 4025 if (argblock) 4026 use = plus_constant (argblock, 4027 argvec[argnum].locate.offset.constant); 4028 else 4029 /* When arguments are pushed, trying to tell alias.c where 4030 exactly this argument is won't work, because the 4031 auto-increment causes confusion. So we merely indicate 4032 that we access something with a known mode somewhere on 4033 the stack. */ 4034 use = gen_rtx_PLUS (Pmode, virtual_outgoing_args_rtx, 4035 gen_rtx_SCRATCH (Pmode)); 4036 use = gen_rtx_MEM (argvec[argnum].mode, use); 4037 use = gen_rtx_USE (VOIDmode, use); 4038 call_fusage = gen_rtx_EXPR_LIST (VOIDmode, use, call_fusage); 4039 } 4040 } 4041 4042 /* If we pushed args in forward order, perform stack alignment 4043 after pushing the last arg. */ 4044 if (argblock == 0 && !PUSH_ARGS_REVERSED) 4045 anti_adjust_stack (GEN_INT (args_size.constant 4046 - original_args_size.constant)); 4047 4048 if (PUSH_ARGS_REVERSED) 4049 argnum = nargs - 1; 4050 else 4051 argnum = 0; 4052 4053 fun = prepare_call_address (NULL, fun, NULL, &call_fusage, 0, 0); 4054 4055 /* Now load any reg parms into their regs. */ 4056 4057 /* ARGNUM indexes the ARGVEC array in the order in which the arguments 4058 are to be pushed. */ 4059 for (count = 0; count < nargs; count++, argnum += inc) 4060 { 4061 enum machine_mode mode = argvec[argnum].mode; 4062 rtx val = argvec[argnum].value; 4063 rtx reg = argvec[argnum].reg; 4064 int partial = argvec[argnum].partial; 4065 #ifdef BLOCK_REG_PADDING 4066 int size = 0; 4067 #endif 4068 4069 /* Handle calls that pass values in multiple non-contiguous 4070 locations. The PA64 has examples of this for library calls. */ 4071 if (reg != 0 && GET_CODE (reg) == PARALLEL) 4072 emit_group_load (reg, val, NULL_TREE, GET_MODE_SIZE (mode)); 4073 else if (reg != 0 && partial == 0) 4074 { 4075 emit_move_insn (reg, val); 4076 #ifdef BLOCK_REG_PADDING 4077 size = GET_MODE_SIZE (argvec[argnum].mode); 4078 4079 /* Copied from load_register_parameters. */ 4080 4081 /* Handle case where we have a value that needs shifting 4082 up to the msb. eg. a QImode value and we're padding 4083 upward on a BYTES_BIG_ENDIAN machine. */ 4084 if (size < UNITS_PER_WORD 4085 && (argvec[argnum].locate.where_pad 4086 == (BYTES_BIG_ENDIAN ? upward : downward))) 4087 { 4088 rtx x; 4089 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT; 4090 4091 /* Assigning REG here rather than a temp makes CALL_FUSAGE 4092 report the whole reg as used. Strictly speaking, the 4093 call only uses SIZE bytes at the msb end, but it doesn't 4094 seem worth generating rtl to say that. */ 4095 reg = gen_rtx_REG (word_mode, REGNO (reg)); 4096 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1); 4097 if (x != reg) 4098 emit_move_insn (reg, x); 4099 } 4100 #endif 4101 } 4102 4103 NO_DEFER_POP; 4104 } 4105 4106 /* Any regs containing parms remain in use through the call. */ 4107 for (count = 0; count < nargs; count++) 4108 { 4109 rtx reg = argvec[count].reg; 4110 if (reg != 0 && GET_CODE (reg) == PARALLEL) 4111 use_group_regs (&call_fusage, reg); 4112 else if (reg != 0) 4113 { 4114 int partial = argvec[count].partial; 4115 if (partial) 4116 { 4117 int nregs; 4118 gcc_assert (partial % UNITS_PER_WORD == 0); 4119 nregs = partial / UNITS_PER_WORD; 4120 use_regs (&call_fusage, REGNO (reg), nregs); 4121 } 4122 else 4123 use_reg (&call_fusage, reg); 4124 } 4125 } 4126 4127 /* Pass the function the address in which to return a structure value. */ 4128 if (mem_value != 0 && struct_value != 0 && ! pcc_struct_value) 4129 { 4130 emit_move_insn (struct_value, 4131 force_reg (Pmode, 4132 force_operand (XEXP (mem_value, 0), 4133 NULL_RTX))); 4134 if (REG_P (struct_value)) 4135 use_reg (&call_fusage, struct_value); 4136 } 4137 4138 /* Don't allow popping to be deferred, since then 4139 cse'ing of library calls could delete a call and leave the pop. */ 4140 NO_DEFER_POP; 4141 valreg = (mem_value == 0 && outmode != VOIDmode 4142 ? hard_libcall_value (outmode, orgfun) : NULL_RTX); 4143 4144 /* Stack must be properly aligned now. */ 4145 gcc_assert (!(stack_pointer_delta 4146 & (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1))); 4147 4148 before_call = get_last_insn (); 4149 4150 /* We pass the old value of inhibit_defer_pop + 1 to emit_call_1, which 4151 will set inhibit_defer_pop to that value. */ 4152 /* The return type is needed to decide how many bytes the function pops. 4153 Signedness plays no role in that, so for simplicity, we pretend it's 4154 always signed. We also assume that the list of arguments passed has 4155 no impact, so we pretend it is unknown. */ 4156 4157 emit_call_1 (fun, NULL, 4158 get_identifier (XSTR (orgfun, 0)), 4159 build_function_type (tfom, NULL_TREE), 4160 original_args_size.constant, args_size.constant, 4161 struct_value_size, 4162 targetm.calls.function_arg (args_so_far, 4163 VOIDmode, void_type_node, true), 4164 valreg, 4165 old_inhibit_defer_pop + 1, call_fusage, flags, args_so_far); 4166 4167 /* Right-shift returned value if necessary. */ 4168 if (!pcc_struct_value 4169 && TYPE_MODE (tfom) != BLKmode 4170 && targetm.calls.return_in_msb (tfom)) 4171 { 4172 shift_return_value (TYPE_MODE (tfom), false, valreg); 4173 valreg = gen_rtx_REG (TYPE_MODE (tfom), REGNO (valreg)); 4174 } 4175 4176 /* For calls to `setjmp', etc., inform function.c:setjmp_warnings 4177 that it should complain if nonvolatile values are live. For 4178 functions that cannot return, inform flow that control does not 4179 fall through. */ 4180 4181 if (flags & ECF_NORETURN) 4182 { 4183 /* The barrier note must be emitted 4184 immediately after the CALL_INSN. Some ports emit more than 4185 just a CALL_INSN above, so we must search for it here. */ 4186 4187 rtx last = get_last_insn (); 4188 while (!CALL_P (last)) 4189 { 4190 last = PREV_INSN (last); 4191 /* There was no CALL_INSN? */ 4192 gcc_assert (last != before_call); 4193 } 4194 4195 emit_barrier_after (last); 4196 } 4197 4198 /* Now restore inhibit_defer_pop to its actual original value. */ 4199 OK_DEFER_POP; 4200 4201 pop_temp_slots (); 4202 4203 /* Copy the value to the right place. */ 4204 if (outmode != VOIDmode && retval) 4205 { 4206 if (mem_value) 4207 { 4208 if (value == 0) 4209 value = mem_value; 4210 if (value != mem_value) 4211 emit_move_insn (value, mem_value); 4212 } 4213 else if (GET_CODE (valreg) == PARALLEL) 4214 { 4215 if (value == 0) 4216 value = gen_reg_rtx (outmode); 4217 emit_group_store (value, valreg, NULL_TREE, GET_MODE_SIZE (outmode)); 4218 } 4219 else 4220 { 4221 /* Convert to the proper mode if a promotion has been active. */ 4222 if (GET_MODE (valreg) != outmode) 4223 { 4224 int unsignedp = TYPE_UNSIGNED (tfom); 4225 4226 gcc_assert (promote_function_mode (tfom, outmode, &unsignedp, 4227 fndecl ? TREE_TYPE (fndecl) : fntype, 1) 4228 == GET_MODE (valreg)); 4229 valreg = convert_modes (outmode, GET_MODE (valreg), valreg, 0); 4230 } 4231 4232 if (value != 0) 4233 emit_move_insn (value, valreg); 4234 else 4235 value = valreg; 4236 } 4237 } 4238 4239 if (ACCUMULATE_OUTGOING_ARGS) 4240 { 4241 #ifdef REG_PARM_STACK_SPACE 4242 if (save_area) 4243 restore_fixed_argument_area (save_area, argblock, 4244 high_to_save, low_to_save); 4245 #endif 4246 4247 /* If we saved any argument areas, restore them. */ 4248 for (count = 0; count < nargs; count++) 4249 if (argvec[count].save_area) 4250 { 4251 enum machine_mode save_mode = GET_MODE (argvec[count].save_area); 4252 rtx adr = plus_constant (argblock, 4253 argvec[count].locate.offset.constant); 4254 rtx stack_area = gen_rtx_MEM (save_mode, 4255 memory_address (save_mode, adr)); 4256 4257 if (save_mode == BLKmode) 4258 emit_block_move (stack_area, 4259 validize_mem (argvec[count].save_area), 4260 GEN_INT (argvec[count].locate.size.constant), 4261 BLOCK_OP_CALL_PARM); 4262 else 4263 emit_move_insn (stack_area, argvec[count].save_area); 4264 } 4265 4266 highest_outgoing_arg_in_use = initial_highest_arg_in_use; 4267 stack_usage_map = initial_stack_usage_map; 4268 } 4269 4270 free (stack_usage_map_buf); 4271 4272 return value; 4273 4274 } 4275 4276 /* Output a library call to function FUN (a SYMBOL_REF rtx) 4277 (emitting the queue unless NO_QUEUE is nonzero), 4278 for a value of mode OUTMODE, 4279 with NARGS different arguments, passed as alternating rtx values 4280 and machine_modes to convert them to. 4281 4282 FN_TYPE should be LCT_NORMAL for `normal' calls, LCT_CONST for 4283 `const' calls, LCT_PURE for `pure' calls, or other LCT_ value for 4284 other types of library calls. */ 4285 4286 void 4287 emit_library_call (rtx orgfun, enum libcall_type fn_type, 4288 enum machine_mode outmode, int nargs, ...) 4289 { 4290 va_list p; 4291 4292 va_start (p, nargs); 4293 emit_library_call_value_1 (0, orgfun, NULL_RTX, fn_type, outmode, nargs, p); 4294 va_end (p); 4295 } 4296 4297 /* Like emit_library_call except that an extra argument, VALUE, 4298 comes second and says where to store the result. 4299 (If VALUE is zero, this function chooses a convenient way 4300 to return the value. 4301 4302 This function returns an rtx for where the value is to be found. 4303 If VALUE is nonzero, VALUE is returned. */ 4304 4305 rtx 4306 emit_library_call_value (rtx orgfun, rtx value, 4307 enum libcall_type fn_type, 4308 enum machine_mode outmode, int nargs, ...) 4309 { 4310 rtx result; 4311 va_list p; 4312 4313 va_start (p, nargs); 4314 result = emit_library_call_value_1 (1, orgfun, value, fn_type, outmode, 4315 nargs, p); 4316 va_end (p); 4317 4318 return result; 4319 } 4320 4321 /* Store a single argument for a function call 4322 into the register or memory area where it must be passed. 4323 *ARG describes the argument value and where to pass it. 4324 4325 ARGBLOCK is the address of the stack-block for all the arguments, 4326 or 0 on a machine where arguments are pushed individually. 4327 4328 MAY_BE_ALLOCA nonzero says this could be a call to `alloca' 4329 so must be careful about how the stack is used. 4330 4331 VARIABLE_SIZE nonzero says that this was a variable-sized outgoing 4332 argument stack. This is used if ACCUMULATE_OUTGOING_ARGS to indicate 4333 that we need not worry about saving and restoring the stack. 4334 4335 FNDECL is the declaration of the function we are calling. 4336 4337 Return nonzero if this arg should cause sibcall failure, 4338 zero otherwise. */ 4339 4340 static int 4341 store_one_arg (struct arg_data *arg, rtx argblock, int flags, 4342 int variable_size ATTRIBUTE_UNUSED, int reg_parm_stack_space) 4343 { 4344 tree pval = arg->tree_value; 4345 rtx reg = 0; 4346 int partial = 0; 4347 int used = 0; 4348 int i, lower_bound = 0, upper_bound = 0; 4349 int sibcall_failure = 0; 4350 4351 if (TREE_CODE (pval) == ERROR_MARK) 4352 return 1; 4353 4354 /* Push a new temporary level for any temporaries we make for 4355 this argument. */ 4356 push_temp_slots (); 4357 4358 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL)) 4359 { 4360 /* If this is being stored into a pre-allocated, fixed-size, stack area, 4361 save any previous data at that location. */ 4362 if (argblock && ! variable_size && arg->stack) 4363 { 4364 #ifdef ARGS_GROW_DOWNWARD 4365 /* stack_slot is negative, but we want to index stack_usage_map 4366 with positive values. */ 4367 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS) 4368 upper_bound = -INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1)) + 1; 4369 else 4370 upper_bound = 0; 4371 4372 lower_bound = upper_bound - arg->locate.size.constant; 4373 #else 4374 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS) 4375 lower_bound = INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1)); 4376 else 4377 lower_bound = 0; 4378 4379 upper_bound = lower_bound + arg->locate.size.constant; 4380 #endif 4381 4382 i = lower_bound; 4383 /* Don't worry about things in the fixed argument area; 4384 it has already been saved. */ 4385 if (i < reg_parm_stack_space) 4386 i = reg_parm_stack_space; 4387 while (i < upper_bound && stack_usage_map[i] == 0) 4388 i++; 4389 4390 if (i < upper_bound) 4391 { 4392 /* We need to make a save area. */ 4393 unsigned int size = arg->locate.size.constant * BITS_PER_UNIT; 4394 enum machine_mode save_mode = mode_for_size (size, MODE_INT, 1); 4395 rtx adr = memory_address (save_mode, XEXP (arg->stack_slot, 0)); 4396 rtx stack_area = gen_rtx_MEM (save_mode, adr); 4397 4398 if (save_mode == BLKmode) 4399 { 4400 tree ot = TREE_TYPE (arg->tree_value); 4401 tree nt = build_qualified_type (ot, (TYPE_QUALS (ot) 4402 | TYPE_QUAL_CONST)); 4403 4404 arg->save_area = assign_temp (nt, 0, 1, 1); 4405 preserve_temp_slots (arg->save_area); 4406 emit_block_move (validize_mem (arg->save_area), stack_area, 4407 GEN_INT (arg->locate.size.constant), 4408 BLOCK_OP_CALL_PARM); 4409 } 4410 else 4411 { 4412 arg->save_area = gen_reg_rtx (save_mode); 4413 emit_move_insn (arg->save_area, stack_area); 4414 } 4415 } 4416 } 4417 } 4418 4419 /* If this isn't going to be placed on both the stack and in registers, 4420 set up the register and number of words. */ 4421 if (! arg->pass_on_stack) 4422 { 4423 if (flags & ECF_SIBCALL) 4424 reg = arg->tail_call_reg; 4425 else 4426 reg = arg->reg; 4427 partial = arg->partial; 4428 } 4429 4430 /* Being passed entirely in a register. We shouldn't be called in 4431 this case. */ 4432 gcc_assert (reg == 0 || partial != 0); 4433 4434 /* If this arg needs special alignment, don't load the registers 4435 here. */ 4436 if (arg->n_aligned_regs != 0) 4437 reg = 0; 4438 4439 /* If this is being passed partially in a register, we can't evaluate 4440 it directly into its stack slot. Otherwise, we can. */ 4441 if (arg->value == 0) 4442 { 4443 /* stack_arg_under_construction is nonzero if a function argument is 4444 being evaluated directly into the outgoing argument list and 4445 expand_call must take special action to preserve the argument list 4446 if it is called recursively. 4447 4448 For scalar function arguments stack_usage_map is sufficient to 4449 determine which stack slots must be saved and restored. Scalar 4450 arguments in general have pass_on_stack == 0. 4451 4452 If this argument is initialized by a function which takes the 4453 address of the argument (a C++ constructor or a C function 4454 returning a BLKmode structure), then stack_usage_map is 4455 insufficient and expand_call must push the stack around the 4456 function call. Such arguments have pass_on_stack == 1. 4457 4458 Note that it is always safe to set stack_arg_under_construction, 4459 but this generates suboptimal code if set when not needed. */ 4460 4461 if (arg->pass_on_stack) 4462 stack_arg_under_construction++; 4463 4464 arg->value = expand_expr (pval, 4465 (partial 4466 || TYPE_MODE (TREE_TYPE (pval)) != arg->mode) 4467 ? NULL_RTX : arg->stack, 4468 VOIDmode, EXPAND_STACK_PARM); 4469 4470 /* If we are promoting object (or for any other reason) the mode 4471 doesn't agree, convert the mode. */ 4472 4473 if (arg->mode != TYPE_MODE (TREE_TYPE (pval))) 4474 arg->value = convert_modes (arg->mode, TYPE_MODE (TREE_TYPE (pval)), 4475 arg->value, arg->unsignedp); 4476 4477 if (arg->pass_on_stack) 4478 stack_arg_under_construction--; 4479 } 4480 4481 /* Check for overlap with already clobbered argument area. */ 4482 if ((flags & ECF_SIBCALL) 4483 && MEM_P (arg->value) 4484 && mem_overlaps_already_clobbered_arg_p (XEXP (arg->value, 0), 4485 arg->locate.size.constant)) 4486 sibcall_failure = 1; 4487 4488 /* Don't allow anything left on stack from computation 4489 of argument to alloca. */ 4490 if (flags & ECF_MAY_BE_ALLOCA) 4491 do_pending_stack_adjust (); 4492 4493 if (arg->value == arg->stack) 4494 /* If the value is already in the stack slot, we are done. */ 4495 ; 4496 else if (arg->mode != BLKmode) 4497 { 4498 int size; 4499 unsigned int parm_align; 4500 4501 /* Argument is a scalar, not entirely passed in registers. 4502 (If part is passed in registers, arg->partial says how much 4503 and emit_push_insn will take care of putting it there.) 4504 4505 Push it, and if its size is less than the 4506 amount of space allocated to it, 4507 also bump stack pointer by the additional space. 4508 Note that in C the default argument promotions 4509 will prevent such mismatches. */ 4510 4511 size = GET_MODE_SIZE (arg->mode); 4512 /* Compute how much space the push instruction will push. 4513 On many machines, pushing a byte will advance the stack 4514 pointer by a halfword. */ 4515 #ifdef PUSH_ROUNDING 4516 size = PUSH_ROUNDING (size); 4517 #endif 4518 used = size; 4519 4520 /* Compute how much space the argument should get: 4521 round up to a multiple of the alignment for arguments. */ 4522 if (none != FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval))) 4523 used = (((size + PARM_BOUNDARY / BITS_PER_UNIT - 1) 4524 / (PARM_BOUNDARY / BITS_PER_UNIT)) 4525 * (PARM_BOUNDARY / BITS_PER_UNIT)); 4526 4527 /* Compute the alignment of the pushed argument. */ 4528 parm_align = arg->locate.boundary; 4529 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward) 4530 { 4531 int pad = used - size; 4532 if (pad) 4533 { 4534 unsigned int pad_align = (pad & -pad) * BITS_PER_UNIT; 4535 parm_align = MIN (parm_align, pad_align); 4536 } 4537 } 4538 4539 /* This isn't already where we want it on the stack, so put it there. 4540 This can either be done with push or copy insns. */ 4541 emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), NULL_RTX, 4542 parm_align, partial, reg, used - size, argblock, 4543 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space, 4544 ARGS_SIZE_RTX (arg->locate.alignment_pad)); 4545 4546 /* Unless this is a partially-in-register argument, the argument is now 4547 in the stack. */ 4548 if (partial == 0) 4549 arg->value = arg->stack; 4550 } 4551 else 4552 { 4553 /* BLKmode, at least partly to be pushed. */ 4554 4555 unsigned int parm_align; 4556 int excess; 4557 rtx size_rtx; 4558 4559 /* Pushing a nonscalar. 4560 If part is passed in registers, PARTIAL says how much 4561 and emit_push_insn will take care of putting it there. */ 4562 4563 /* Round its size up to a multiple 4564 of the allocation unit for arguments. */ 4565 4566 if (arg->locate.size.var != 0) 4567 { 4568 excess = 0; 4569 size_rtx = ARGS_SIZE_RTX (arg->locate.size); 4570 } 4571 else 4572 { 4573 /* PUSH_ROUNDING has no effect on us, because emit_push_insn 4574 for BLKmode is careful to avoid it. */ 4575 excess = (arg->locate.size.constant 4576 - int_size_in_bytes (TREE_TYPE (pval)) 4577 + partial); 4578 size_rtx = expand_expr (size_in_bytes (TREE_TYPE (pval)), 4579 NULL_RTX, TYPE_MODE (sizetype), 4580 EXPAND_NORMAL); 4581 } 4582 4583 parm_align = arg->locate.boundary; 4584 4585 /* When an argument is padded down, the block is aligned to 4586 PARM_BOUNDARY, but the actual argument isn't. */ 4587 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward) 4588 { 4589 if (arg->locate.size.var) 4590 parm_align = BITS_PER_UNIT; 4591 else if (excess) 4592 { 4593 unsigned int excess_align = (excess & -excess) * BITS_PER_UNIT; 4594 parm_align = MIN (parm_align, excess_align); 4595 } 4596 } 4597 4598 if ((flags & ECF_SIBCALL) && MEM_P (arg->value)) 4599 { 4600 /* emit_push_insn might not work properly if arg->value and 4601 argblock + arg->locate.offset areas overlap. */ 4602 rtx x = arg->value; 4603 int i = 0; 4604 4605 if (XEXP (x, 0) == crtl->args.internal_arg_pointer 4606 || (GET_CODE (XEXP (x, 0)) == PLUS 4607 && XEXP (XEXP (x, 0), 0) == 4608 crtl->args.internal_arg_pointer 4609 && CONST_INT_P (XEXP (XEXP (x, 0), 1)))) 4610 { 4611 if (XEXP (x, 0) != crtl->args.internal_arg_pointer) 4612 i = INTVAL (XEXP (XEXP (x, 0), 1)); 4613 4614 /* expand_call should ensure this. */ 4615 gcc_assert (!arg->locate.offset.var 4616 && arg->locate.size.var == 0 4617 && CONST_INT_P (size_rtx)); 4618 4619 if (arg->locate.offset.constant > i) 4620 { 4621 if (arg->locate.offset.constant < i + INTVAL (size_rtx)) 4622 sibcall_failure = 1; 4623 } 4624 else if (arg->locate.offset.constant < i) 4625 { 4626 /* Use arg->locate.size.constant instead of size_rtx 4627 because we only care about the part of the argument 4628 on the stack. */ 4629 if (i < (arg->locate.offset.constant 4630 + arg->locate.size.constant)) 4631 sibcall_failure = 1; 4632 } 4633 else 4634 { 4635 /* Even though they appear to be at the same location, 4636 if part of the outgoing argument is in registers, 4637 they aren't really at the same location. Check for 4638 this by making sure that the incoming size is the 4639 same as the outgoing size. */ 4640 if (arg->locate.size.constant != INTVAL (size_rtx)) 4641 sibcall_failure = 1; 4642 } 4643 } 4644 } 4645 4646 emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), size_rtx, 4647 parm_align, partial, reg, excess, argblock, 4648 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space, 4649 ARGS_SIZE_RTX (arg->locate.alignment_pad)); 4650 4651 /* Unless this is a partially-in-register argument, the argument is now 4652 in the stack. 4653 4654 ??? Unlike the case above, in which we want the actual 4655 address of the data, so that we can load it directly into a 4656 register, here we want the address of the stack slot, so that 4657 it's properly aligned for word-by-word copying or something 4658 like that. It's not clear that this is always correct. */ 4659 if (partial == 0) 4660 arg->value = arg->stack_slot; 4661 } 4662 4663 if (arg->reg && GET_CODE (arg->reg) == PARALLEL) 4664 { 4665 tree type = TREE_TYPE (arg->tree_value); 4666 arg->parallel_value 4667 = emit_group_load_into_temps (arg->reg, arg->value, type, 4668 int_size_in_bytes (type)); 4669 } 4670 4671 /* Mark all slots this store used. */ 4672 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL) 4673 && argblock && ! variable_size && arg->stack) 4674 for (i = lower_bound; i < upper_bound; i++) 4675 stack_usage_map[i] = 1; 4676 4677 /* Once we have pushed something, pops can't safely 4678 be deferred during the rest of the arguments. */ 4679 NO_DEFER_POP; 4680 4681 /* Free any temporary slots made in processing this argument. Show 4682 that we might have taken the address of something and pushed that 4683 as an operand. */ 4684 preserve_temp_slots (NULL_RTX); 4685 free_temp_slots (); 4686 pop_temp_slots (); 4687 4688 return sibcall_failure; 4689 } 4690 4691 /* Nonzero if we do not know how to pass TYPE solely in registers. */ 4692 4693 bool 4694 must_pass_in_stack_var_size (enum machine_mode mode ATTRIBUTE_UNUSED, 4695 const_tree type) 4696 { 4697 if (!type) 4698 return false; 4699 4700 /* If the type has variable size... */ 4701 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) 4702 return true; 4703 4704 /* If the type is marked as addressable (it is required 4705 to be constructed into the stack)... */ 4706 if (TREE_ADDRESSABLE (type)) 4707 return true; 4708 4709 return false; 4710 } 4711 4712 /* Another version of the TARGET_MUST_PASS_IN_STACK hook. This one 4713 takes trailing padding of a structure into account. */ 4714 /* ??? Should be able to merge these two by examining BLOCK_REG_PADDING. */ 4715 4716 bool 4717 must_pass_in_stack_var_size_or_pad (enum machine_mode mode, const_tree type) 4718 { 4719 if (!type) 4720 return false; 4721 4722 /* If the type has variable size... */ 4723 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) 4724 return true; 4725 4726 /* If the type is marked as addressable (it is required 4727 to be constructed into the stack)... */ 4728 if (TREE_ADDRESSABLE (type)) 4729 return true; 4730 4731 /* If the padding and mode of the type is such that a copy into 4732 a register would put it into the wrong part of the register. */ 4733 if (mode == BLKmode 4734 && int_size_in_bytes (type) % (PARM_BOUNDARY / BITS_PER_UNIT) 4735 && (FUNCTION_ARG_PADDING (mode, type) 4736 == (BYTES_BIG_ENDIAN ? upward : downward))) 4737 return true; 4738 4739 return false; 4740 } 4741