2014-04-15 Richard Biener <rguenther@suse.de>
[official-gcc.git] / gcc / calls.c
blobf0c92ddc0182d3216a90e0dc06e9b3a40bca9a45
1 /* Convert function calls to rtl insns, for GNU C compiler.
2 Copyright (C) 1989-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "rtl.h"
25 #include "tree.h"
26 #include "stor-layout.h"
27 #include "varasm.h"
28 #include "stringpool.h"
29 #include "attribs.h"
30 #include "basic-block.h"
31 #include "tree-ssa-alias.h"
32 #include "internal-fn.h"
33 #include "gimple-expr.h"
34 #include "is-a.h"
35 #include "gimple.h"
36 #include "flags.h"
37 #include "expr.h"
38 #include "optabs.h"
39 #include "libfuncs.h"
40 #include "function.h"
41 #include "regs.h"
42 #include "diagnostic-core.h"
43 #include "output.h"
44 #include "tm_p.h"
45 #include "timevar.h"
46 #include "sbitmap.h"
47 #include "langhooks.h"
48 #include "target.h"
49 #include "cgraph.h"
50 #include "except.h"
51 #include "dbgcnt.h"
53 /* Like PREFERRED_STACK_BOUNDARY but in units of bytes, not bits. */
54 #define STACK_BYTES (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT)
56 /* Data structure and subroutines used within expand_call. */
58 struct arg_data
60 /* Tree node for this argument. */
61 tree tree_value;
62 /* Mode for value; TYPE_MODE unless promoted. */
63 enum machine_mode mode;
64 /* Current RTL value for argument, or 0 if it isn't precomputed. */
65 rtx value;
66 /* Initially-compute RTL value for argument; only for const functions. */
67 rtx initial_value;
68 /* Register to pass this argument in, 0 if passed on stack, or an
69 PARALLEL if the arg is to be copied into multiple non-contiguous
70 registers. */
71 rtx reg;
72 /* Register to pass this argument in when generating tail call sequence.
73 This is not the same register as for normal calls on machines with
74 register windows. */
75 rtx tail_call_reg;
76 /* If REG is a PARALLEL, this is a copy of VALUE pulled into the correct
77 form for emit_group_move. */
78 rtx parallel_value;
79 /* If REG was promoted from the actual mode of the argument expression,
80 indicates whether the promotion is sign- or zero-extended. */
81 int unsignedp;
82 /* Number of bytes to put in registers. 0 means put the whole arg
83 in registers. Also 0 if not passed in registers. */
84 int partial;
85 /* Nonzero if argument must be passed on stack.
86 Note that some arguments may be passed on the stack
87 even though pass_on_stack is zero, just because FUNCTION_ARG says so.
88 pass_on_stack identifies arguments that *cannot* go in registers. */
89 int pass_on_stack;
90 /* Some fields packaged up for locate_and_pad_parm. */
91 struct locate_and_pad_arg_data locate;
92 /* Location on the stack at which parameter should be stored. The store
93 has already been done if STACK == VALUE. */
94 rtx stack;
95 /* Location on the stack of the start of this argument slot. This can
96 differ from STACK if this arg pads downward. This location is known
97 to be aligned to TARGET_FUNCTION_ARG_BOUNDARY. */
98 rtx stack_slot;
99 /* Place that this stack area has been saved, if needed. */
100 rtx save_area;
101 /* If an argument's alignment does not permit direct copying into registers,
102 copy in smaller-sized pieces into pseudos. These are stored in a
103 block pointed to by this field. The next field says how many
104 word-sized pseudos we made. */
105 rtx *aligned_regs;
106 int n_aligned_regs;
109 /* A vector of one char per byte of stack space. A byte if nonzero if
110 the corresponding stack location has been used.
111 This vector is used to prevent a function call within an argument from
112 clobbering any stack already set up. */
113 static char *stack_usage_map;
115 /* Size of STACK_USAGE_MAP. */
116 static int highest_outgoing_arg_in_use;
118 /* A bitmap of virtual-incoming stack space. Bit is set if the corresponding
119 stack location's tail call argument has been already stored into the stack.
120 This bitmap is used to prevent sibling call optimization if function tries
121 to use parent's incoming argument slots when they have been already
122 overwritten with tail call arguments. */
123 static sbitmap stored_args_map;
125 /* stack_arg_under_construction is nonzero when an argument may be
126 initialized with a constructor call (including a C function that
127 returns a BLKmode struct) and expand_call must take special action
128 to make sure the object being constructed does not overlap the
129 argument list for the constructor call. */
130 static int stack_arg_under_construction;
132 static void emit_call_1 (rtx, tree, tree, tree, HOST_WIDE_INT, HOST_WIDE_INT,
133 HOST_WIDE_INT, rtx, rtx, int, rtx, int,
134 cumulative_args_t);
135 static void precompute_register_parameters (int, struct arg_data *, int *);
136 static int store_one_arg (struct arg_data *, rtx, int, int, int);
137 static void store_unaligned_arguments_into_pseudos (struct arg_data *, int);
138 static int finalize_must_preallocate (int, int, struct arg_data *,
139 struct args_size *);
140 static void precompute_arguments (int, struct arg_data *);
141 static int compute_argument_block_size (int, struct args_size *, tree, tree, int);
142 static void initialize_argument_information (int, struct arg_data *,
143 struct args_size *, int,
144 tree, tree,
145 tree, tree, cumulative_args_t, int,
146 rtx *, int *, int *, int *,
147 bool *, bool);
148 static void compute_argument_addresses (struct arg_data *, rtx, int);
149 static rtx rtx_for_function_call (tree, tree);
150 static void load_register_parameters (struct arg_data *, int, rtx *, int,
151 int, int *);
152 static rtx emit_library_call_value_1 (int, rtx, rtx, enum libcall_type,
153 enum machine_mode, int, va_list);
154 static int special_function_p (const_tree, int);
155 static int check_sibcall_argument_overlap_1 (rtx);
156 static int check_sibcall_argument_overlap (rtx, struct arg_data *, int);
158 static int combine_pending_stack_adjustment_and_call (int, struct args_size *,
159 unsigned int);
160 static tree split_complex_types (tree);
162 #ifdef REG_PARM_STACK_SPACE
163 static rtx save_fixed_argument_area (int, rtx, int *, int *);
164 static void restore_fixed_argument_area (rtx, rtx, int, int);
165 #endif
167 /* Force FUNEXP into a form suitable for the address of a CALL,
168 and return that as an rtx. Also load the static chain register
169 if FNDECL is a nested function.
171 CALL_FUSAGE points to a variable holding the prospective
172 CALL_INSN_FUNCTION_USAGE information. */
175 prepare_call_address (tree fndecl, rtx funexp, rtx static_chain_value,
176 rtx *call_fusage, int reg_parm_seen, int sibcallp)
178 /* Make a valid memory address and copy constants through pseudo-regs,
179 but not for a constant address if -fno-function-cse. */
180 if (GET_CODE (funexp) != SYMBOL_REF)
181 /* If we are using registers for parameters, force the
182 function address into a register now. */
183 funexp = ((reg_parm_seen
184 && targetm.small_register_classes_for_mode_p (FUNCTION_MODE))
185 ? force_not_mem (memory_address (FUNCTION_MODE, funexp))
186 : memory_address (FUNCTION_MODE, funexp));
187 else if (! sibcallp)
189 #ifndef NO_FUNCTION_CSE
190 if (optimize && ! flag_no_function_cse)
191 funexp = force_reg (Pmode, funexp);
192 #endif
195 if (static_chain_value != 0)
197 rtx chain;
199 gcc_assert (fndecl);
200 chain = targetm.calls.static_chain (fndecl, false);
201 static_chain_value = convert_memory_address (Pmode, static_chain_value);
203 emit_move_insn (chain, static_chain_value);
204 if (REG_P (chain))
205 use_reg (call_fusage, chain);
208 return funexp;
211 /* Generate instructions to call function FUNEXP,
212 and optionally pop the results.
213 The CALL_INSN is the first insn generated.
215 FNDECL is the declaration node of the function. This is given to the
216 hook TARGET_RETURN_POPS_ARGS to determine whether this function pops
217 its own args.
219 FUNTYPE is the data type of the function. This is given to the hook
220 TARGET_RETURN_POPS_ARGS to determine whether this function pops its
221 own args. We used to allow an identifier for library functions, but
222 that doesn't work when the return type is an aggregate type and the
223 calling convention says that the pointer to this aggregate is to be
224 popped by the callee.
226 STACK_SIZE is the number of bytes of arguments on the stack,
227 ROUNDED_STACK_SIZE is that number rounded up to
228 PREFERRED_STACK_BOUNDARY; zero if the size is variable. This is
229 both to put into the call insn and to generate explicit popping
230 code if necessary.
232 STRUCT_VALUE_SIZE is the number of bytes wanted in a structure value.
233 It is zero if this call doesn't want a structure value.
235 NEXT_ARG_REG is the rtx that results from executing
236 targetm.calls.function_arg (&args_so_far, VOIDmode, void_type_node, true)
237 just after all the args have had their registers assigned.
238 This could be whatever you like, but normally it is the first
239 arg-register beyond those used for args in this call,
240 or 0 if all the arg-registers are used in this call.
241 It is passed on to `gen_call' so you can put this info in the call insn.
243 VALREG is a hard register in which a value is returned,
244 or 0 if the call does not return a value.
246 OLD_INHIBIT_DEFER_POP is the value that `inhibit_defer_pop' had before
247 the args to this call were processed.
248 We restore `inhibit_defer_pop' to that value.
250 CALL_FUSAGE is either empty or an EXPR_LIST of USE expressions that
251 denote registers used by the called function. */
253 static void
254 emit_call_1 (rtx funexp, tree fntree ATTRIBUTE_UNUSED, tree fndecl ATTRIBUTE_UNUSED,
255 tree funtype ATTRIBUTE_UNUSED,
256 HOST_WIDE_INT stack_size ATTRIBUTE_UNUSED,
257 HOST_WIDE_INT rounded_stack_size,
258 HOST_WIDE_INT struct_value_size ATTRIBUTE_UNUSED,
259 rtx next_arg_reg ATTRIBUTE_UNUSED, rtx valreg,
260 int old_inhibit_defer_pop, rtx call_fusage, int ecf_flags,
261 cumulative_args_t args_so_far ATTRIBUTE_UNUSED)
263 rtx rounded_stack_size_rtx = GEN_INT (rounded_stack_size);
264 rtx call_insn, call, funmem;
265 int already_popped = 0;
266 HOST_WIDE_INT n_popped
267 = targetm.calls.return_pops_args (fndecl, funtype, stack_size);
269 #ifdef CALL_POPS_ARGS
270 n_popped += CALL_POPS_ARGS (*get_cumulative_args (args_so_far));
271 #endif
273 /* Ensure address is valid. SYMBOL_REF is already valid, so no need,
274 and we don't want to load it into a register as an optimization,
275 because prepare_call_address already did it if it should be done. */
276 if (GET_CODE (funexp) != SYMBOL_REF)
277 funexp = memory_address (FUNCTION_MODE, funexp);
279 funmem = gen_rtx_MEM (FUNCTION_MODE, funexp);
280 if (fndecl && TREE_CODE (fndecl) == FUNCTION_DECL)
282 tree t = fndecl;
284 /* Although a built-in FUNCTION_DECL and its non-__builtin
285 counterpart compare equal and get a shared mem_attrs, they
286 produce different dump output in compare-debug compilations,
287 if an entry gets garbage collected in one compilation, then
288 adds a different (but equivalent) entry, while the other
289 doesn't run the garbage collector at the same spot and then
290 shares the mem_attr with the equivalent entry. */
291 if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL)
293 tree t2 = builtin_decl_explicit (DECL_FUNCTION_CODE (t));
294 if (t2)
295 t = t2;
298 set_mem_expr (funmem, t);
300 else if (fntree)
301 set_mem_expr (funmem, build_simple_mem_ref (CALL_EXPR_FN (fntree)));
303 #if defined (HAVE_sibcall_pop) && defined (HAVE_sibcall_value_pop)
304 if ((ecf_flags & ECF_SIBCALL)
305 && HAVE_sibcall_pop && HAVE_sibcall_value_pop
306 && (n_popped > 0 || stack_size == 0))
308 rtx n_pop = GEN_INT (n_popped);
309 rtx pat;
311 /* If this subroutine pops its own args, record that in the call insn
312 if possible, for the sake of frame pointer elimination. */
314 if (valreg)
315 pat = GEN_SIBCALL_VALUE_POP (valreg, funmem, rounded_stack_size_rtx,
316 next_arg_reg, n_pop);
317 else
318 pat = GEN_SIBCALL_POP (funmem, rounded_stack_size_rtx, next_arg_reg,
319 n_pop);
321 emit_call_insn (pat);
322 already_popped = 1;
324 else
325 #endif
327 #if defined (HAVE_call_pop) && defined (HAVE_call_value_pop)
328 /* If the target has "call" or "call_value" insns, then prefer them
329 if no arguments are actually popped. If the target does not have
330 "call" or "call_value" insns, then we must use the popping versions
331 even if the call has no arguments to pop. */
332 #if defined (HAVE_call) && defined (HAVE_call_value)
333 if (HAVE_call && HAVE_call_value && HAVE_call_pop && HAVE_call_value_pop
334 && n_popped > 0)
335 #else
336 if (HAVE_call_pop && HAVE_call_value_pop)
337 #endif
339 rtx n_pop = GEN_INT (n_popped);
340 rtx pat;
342 /* If this subroutine pops its own args, record that in the call insn
343 if possible, for the sake of frame pointer elimination. */
345 if (valreg)
346 pat = GEN_CALL_VALUE_POP (valreg, funmem, rounded_stack_size_rtx,
347 next_arg_reg, n_pop);
348 else
349 pat = GEN_CALL_POP (funmem, rounded_stack_size_rtx, next_arg_reg,
350 n_pop);
352 emit_call_insn (pat);
353 already_popped = 1;
355 else
356 #endif
358 #if defined (HAVE_sibcall) && defined (HAVE_sibcall_value)
359 if ((ecf_flags & ECF_SIBCALL)
360 && HAVE_sibcall && HAVE_sibcall_value)
362 if (valreg)
363 emit_call_insn (GEN_SIBCALL_VALUE (valreg, funmem,
364 rounded_stack_size_rtx,
365 next_arg_reg, NULL_RTX));
366 else
367 emit_call_insn (GEN_SIBCALL (funmem, rounded_stack_size_rtx,
368 next_arg_reg,
369 GEN_INT (struct_value_size)));
371 else
372 #endif
374 #if defined (HAVE_call) && defined (HAVE_call_value)
375 if (HAVE_call && HAVE_call_value)
377 if (valreg)
378 emit_call_insn (GEN_CALL_VALUE (valreg, funmem, rounded_stack_size_rtx,
379 next_arg_reg, NULL_RTX));
380 else
381 emit_call_insn (GEN_CALL (funmem, rounded_stack_size_rtx, next_arg_reg,
382 GEN_INT (struct_value_size)));
384 else
385 #endif
386 gcc_unreachable ();
388 /* Find the call we just emitted. */
389 call_insn = last_call_insn ();
391 /* Some target create a fresh MEM instead of reusing the one provided
392 above. Set its MEM_EXPR. */
393 call = get_call_rtx_from (call_insn);
394 if (call
395 && MEM_EXPR (XEXP (call, 0)) == NULL_TREE
396 && MEM_EXPR (funmem) != NULL_TREE)
397 set_mem_expr (XEXP (call, 0), MEM_EXPR (funmem));
399 /* Put the register usage information there. */
400 add_function_usage_to (call_insn, call_fusage);
402 /* If this is a const call, then set the insn's unchanging bit. */
403 if (ecf_flags & ECF_CONST)
404 RTL_CONST_CALL_P (call_insn) = 1;
406 /* If this is a pure call, then set the insn's unchanging bit. */
407 if (ecf_flags & ECF_PURE)
408 RTL_PURE_CALL_P (call_insn) = 1;
410 /* If this is a const call, then set the insn's unchanging bit. */
411 if (ecf_flags & ECF_LOOPING_CONST_OR_PURE)
412 RTL_LOOPING_CONST_OR_PURE_CALL_P (call_insn) = 1;
414 /* Create a nothrow REG_EH_REGION note, if needed. */
415 make_reg_eh_region_note (call_insn, ecf_flags, 0);
417 if (ecf_flags & ECF_NORETURN)
418 add_reg_note (call_insn, REG_NORETURN, const0_rtx);
420 if (ecf_flags & ECF_RETURNS_TWICE)
422 add_reg_note (call_insn, REG_SETJMP, const0_rtx);
423 cfun->calls_setjmp = 1;
426 SIBLING_CALL_P (call_insn) = ((ecf_flags & ECF_SIBCALL) != 0);
428 /* Restore this now, so that we do defer pops for this call's args
429 if the context of the call as a whole permits. */
430 inhibit_defer_pop = old_inhibit_defer_pop;
432 if (n_popped > 0)
434 if (!already_popped)
435 CALL_INSN_FUNCTION_USAGE (call_insn)
436 = gen_rtx_EXPR_LIST (VOIDmode,
437 gen_rtx_CLOBBER (VOIDmode, stack_pointer_rtx),
438 CALL_INSN_FUNCTION_USAGE (call_insn));
439 rounded_stack_size -= n_popped;
440 rounded_stack_size_rtx = GEN_INT (rounded_stack_size);
441 stack_pointer_delta -= n_popped;
443 add_reg_note (call_insn, REG_ARGS_SIZE, GEN_INT (stack_pointer_delta));
445 /* If popup is needed, stack realign must use DRAP */
446 if (SUPPORTS_STACK_ALIGNMENT)
447 crtl->need_drap = true;
449 /* For noreturn calls when not accumulating outgoing args force
450 REG_ARGS_SIZE note to prevent crossjumping of calls with different
451 args sizes. */
452 else if (!ACCUMULATE_OUTGOING_ARGS && (ecf_flags & ECF_NORETURN) != 0)
453 add_reg_note (call_insn, REG_ARGS_SIZE, GEN_INT (stack_pointer_delta));
455 if (!ACCUMULATE_OUTGOING_ARGS)
457 /* If returning from the subroutine does not automatically pop the args,
458 we need an instruction to pop them sooner or later.
459 Perhaps do it now; perhaps just record how much space to pop later.
461 If returning from the subroutine does pop the args, indicate that the
462 stack pointer will be changed. */
464 if (rounded_stack_size != 0)
466 if (ecf_flags & ECF_NORETURN)
467 /* Just pretend we did the pop. */
468 stack_pointer_delta -= rounded_stack_size;
469 else if (flag_defer_pop && inhibit_defer_pop == 0
470 && ! (ecf_flags & (ECF_CONST | ECF_PURE)))
471 pending_stack_adjust += rounded_stack_size;
472 else
473 adjust_stack (rounded_stack_size_rtx);
476 /* When we accumulate outgoing args, we must avoid any stack manipulations.
477 Restore the stack pointer to its original value now. Usually
478 ACCUMULATE_OUTGOING_ARGS targets don't get here, but there are exceptions.
479 On i386 ACCUMULATE_OUTGOING_ARGS can be enabled on demand, and
480 popping variants of functions exist as well.
482 ??? We may optimize similar to defer_pop above, but it is
483 probably not worthwhile.
485 ??? It will be worthwhile to enable combine_stack_adjustments even for
486 such machines. */
487 else if (n_popped)
488 anti_adjust_stack (GEN_INT (n_popped));
491 /* Determine if the function identified by NAME and FNDECL is one with
492 special properties we wish to know about.
494 For example, if the function might return more than one time (setjmp), then
495 set RETURNS_TWICE to a nonzero value.
497 Similarly set NORETURN if the function is in the longjmp family.
499 Set MAY_BE_ALLOCA for any memory allocation function that might allocate
500 space from the stack such as alloca. */
502 static int
503 special_function_p (const_tree fndecl, int flags)
505 if (fndecl && DECL_NAME (fndecl)
506 && IDENTIFIER_LENGTH (DECL_NAME (fndecl)) <= 17
507 /* Exclude functions not at the file scope, or not `extern',
508 since they are not the magic functions we would otherwise
509 think they are.
510 FIXME: this should be handled with attributes, not with this
511 hacky imitation of DECL_ASSEMBLER_NAME. It's (also) wrong
512 because you can declare fork() inside a function if you
513 wish. */
514 && (DECL_CONTEXT (fndecl) == NULL_TREE
515 || TREE_CODE (DECL_CONTEXT (fndecl)) == TRANSLATION_UNIT_DECL)
516 && TREE_PUBLIC (fndecl))
518 const char *name = IDENTIFIER_POINTER (DECL_NAME (fndecl));
519 const char *tname = name;
521 /* We assume that alloca will always be called by name. It
522 makes no sense to pass it as a pointer-to-function to
523 anything that does not understand its behavior. */
524 if (((IDENTIFIER_LENGTH (DECL_NAME (fndecl)) == 6
525 && name[0] == 'a'
526 && ! strcmp (name, "alloca"))
527 || (IDENTIFIER_LENGTH (DECL_NAME (fndecl)) == 16
528 && name[0] == '_'
529 && ! strcmp (name, "__builtin_alloca"))))
530 flags |= ECF_MAY_BE_ALLOCA;
532 /* Disregard prefix _, __, __x or __builtin_. */
533 if (name[0] == '_')
535 if (name[1] == '_'
536 && name[2] == 'b'
537 && !strncmp (name + 3, "uiltin_", 7))
538 tname += 10;
539 else if (name[1] == '_' && name[2] == 'x')
540 tname += 3;
541 else if (name[1] == '_')
542 tname += 2;
543 else
544 tname += 1;
547 if (tname[0] == 's')
549 if ((tname[1] == 'e'
550 && (! strcmp (tname, "setjmp")
551 || ! strcmp (tname, "setjmp_syscall")))
552 || (tname[1] == 'i'
553 && ! strcmp (tname, "sigsetjmp"))
554 || (tname[1] == 'a'
555 && ! strcmp (tname, "savectx")))
556 flags |= ECF_RETURNS_TWICE | ECF_LEAF;
558 if (tname[1] == 'i'
559 && ! strcmp (tname, "siglongjmp"))
560 flags |= ECF_NORETURN;
562 else if ((tname[0] == 'q' && tname[1] == 's'
563 && ! strcmp (tname, "qsetjmp"))
564 || (tname[0] == 'v' && tname[1] == 'f'
565 && ! strcmp (tname, "vfork"))
566 || (tname[0] == 'g' && tname[1] == 'e'
567 && !strcmp (tname, "getcontext")))
568 flags |= ECF_RETURNS_TWICE | ECF_LEAF;
570 else if (tname[0] == 'l' && tname[1] == 'o'
571 && ! strcmp (tname, "longjmp"))
572 flags |= ECF_NORETURN;
575 return flags;
578 /* Similar to special_function_p; return a set of ERF_ flags for the
579 function FNDECL. */
580 static int
581 decl_return_flags (tree fndecl)
583 tree attr;
584 tree type = TREE_TYPE (fndecl);
585 if (!type)
586 return 0;
588 attr = lookup_attribute ("fn spec", TYPE_ATTRIBUTES (type));
589 if (!attr)
590 return 0;
592 attr = TREE_VALUE (TREE_VALUE (attr));
593 if (!attr || TREE_STRING_LENGTH (attr) < 1)
594 return 0;
596 switch (TREE_STRING_POINTER (attr)[0])
598 case '1':
599 case '2':
600 case '3':
601 case '4':
602 return ERF_RETURNS_ARG | (TREE_STRING_POINTER (attr)[0] - '1');
604 case 'm':
605 return ERF_NOALIAS;
607 case '.':
608 default:
609 return 0;
613 /* Return nonzero when FNDECL represents a call to setjmp. */
616 setjmp_call_p (const_tree fndecl)
618 if (DECL_IS_RETURNS_TWICE (fndecl))
619 return ECF_RETURNS_TWICE;
620 return special_function_p (fndecl, 0) & ECF_RETURNS_TWICE;
624 /* Return true if STMT is an alloca call. */
626 bool
627 gimple_alloca_call_p (const_gimple stmt)
629 tree fndecl;
631 if (!is_gimple_call (stmt))
632 return false;
634 fndecl = gimple_call_fndecl (stmt);
635 if (fndecl && (special_function_p (fndecl, 0) & ECF_MAY_BE_ALLOCA))
636 return true;
638 return false;
641 /* Return true when exp contains alloca call. */
643 bool
644 alloca_call_p (const_tree exp)
646 tree fndecl;
647 if (TREE_CODE (exp) == CALL_EXPR
648 && (fndecl = get_callee_fndecl (exp))
649 && (special_function_p (fndecl, 0) & ECF_MAY_BE_ALLOCA))
650 return true;
651 return false;
654 /* Return TRUE if FNDECL is either a TM builtin or a TM cloned
655 function. Return FALSE otherwise. */
657 static bool
658 is_tm_builtin (const_tree fndecl)
660 if (fndecl == NULL)
661 return false;
663 if (decl_is_tm_clone (fndecl))
664 return true;
666 if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
668 switch (DECL_FUNCTION_CODE (fndecl))
670 case BUILT_IN_TM_COMMIT:
671 case BUILT_IN_TM_COMMIT_EH:
672 case BUILT_IN_TM_ABORT:
673 case BUILT_IN_TM_IRREVOCABLE:
674 case BUILT_IN_TM_GETTMCLONE_IRR:
675 case BUILT_IN_TM_MEMCPY:
676 case BUILT_IN_TM_MEMMOVE:
677 case BUILT_IN_TM_MEMSET:
678 CASE_BUILT_IN_TM_STORE (1):
679 CASE_BUILT_IN_TM_STORE (2):
680 CASE_BUILT_IN_TM_STORE (4):
681 CASE_BUILT_IN_TM_STORE (8):
682 CASE_BUILT_IN_TM_STORE (FLOAT):
683 CASE_BUILT_IN_TM_STORE (DOUBLE):
684 CASE_BUILT_IN_TM_STORE (LDOUBLE):
685 CASE_BUILT_IN_TM_STORE (M64):
686 CASE_BUILT_IN_TM_STORE (M128):
687 CASE_BUILT_IN_TM_STORE (M256):
688 CASE_BUILT_IN_TM_LOAD (1):
689 CASE_BUILT_IN_TM_LOAD (2):
690 CASE_BUILT_IN_TM_LOAD (4):
691 CASE_BUILT_IN_TM_LOAD (8):
692 CASE_BUILT_IN_TM_LOAD (FLOAT):
693 CASE_BUILT_IN_TM_LOAD (DOUBLE):
694 CASE_BUILT_IN_TM_LOAD (LDOUBLE):
695 CASE_BUILT_IN_TM_LOAD (M64):
696 CASE_BUILT_IN_TM_LOAD (M128):
697 CASE_BUILT_IN_TM_LOAD (M256):
698 case BUILT_IN_TM_LOG:
699 case BUILT_IN_TM_LOG_1:
700 case BUILT_IN_TM_LOG_2:
701 case BUILT_IN_TM_LOG_4:
702 case BUILT_IN_TM_LOG_8:
703 case BUILT_IN_TM_LOG_FLOAT:
704 case BUILT_IN_TM_LOG_DOUBLE:
705 case BUILT_IN_TM_LOG_LDOUBLE:
706 case BUILT_IN_TM_LOG_M64:
707 case BUILT_IN_TM_LOG_M128:
708 case BUILT_IN_TM_LOG_M256:
709 return true;
710 default:
711 break;
714 return false;
717 /* Detect flags (function attributes) from the function decl or type node. */
720 flags_from_decl_or_type (const_tree exp)
722 int flags = 0;
724 if (DECL_P (exp))
726 /* The function exp may have the `malloc' attribute. */
727 if (DECL_IS_MALLOC (exp))
728 flags |= ECF_MALLOC;
730 /* The function exp may have the `returns_twice' attribute. */
731 if (DECL_IS_RETURNS_TWICE (exp))
732 flags |= ECF_RETURNS_TWICE;
734 /* Process the pure and const attributes. */
735 if (TREE_READONLY (exp))
736 flags |= ECF_CONST;
737 if (DECL_PURE_P (exp))
738 flags |= ECF_PURE;
739 if (DECL_LOOPING_CONST_OR_PURE_P (exp))
740 flags |= ECF_LOOPING_CONST_OR_PURE;
742 if (DECL_IS_NOVOPS (exp))
743 flags |= ECF_NOVOPS;
744 if (lookup_attribute ("leaf", DECL_ATTRIBUTES (exp)))
745 flags |= ECF_LEAF;
747 if (TREE_NOTHROW (exp))
748 flags |= ECF_NOTHROW;
750 if (flag_tm)
752 if (is_tm_builtin (exp))
753 flags |= ECF_TM_BUILTIN;
754 else if ((flags & (ECF_CONST|ECF_NOVOPS)) != 0
755 || lookup_attribute ("transaction_pure",
756 TYPE_ATTRIBUTES (TREE_TYPE (exp))))
757 flags |= ECF_TM_PURE;
760 flags = special_function_p (exp, flags);
762 else if (TYPE_P (exp))
764 if (TYPE_READONLY (exp))
765 flags |= ECF_CONST;
767 if (flag_tm
768 && ((flags & ECF_CONST) != 0
769 || lookup_attribute ("transaction_pure", TYPE_ATTRIBUTES (exp))))
770 flags |= ECF_TM_PURE;
772 else
773 gcc_unreachable ();
775 if (TREE_THIS_VOLATILE (exp))
777 flags |= ECF_NORETURN;
778 if (flags & (ECF_CONST|ECF_PURE))
779 flags |= ECF_LOOPING_CONST_OR_PURE;
782 return flags;
785 /* Detect flags from a CALL_EXPR. */
788 call_expr_flags (const_tree t)
790 int flags;
791 tree decl = get_callee_fndecl (t);
793 if (decl)
794 flags = flags_from_decl_or_type (decl);
795 else
797 t = TREE_TYPE (CALL_EXPR_FN (t));
798 if (t && TREE_CODE (t) == POINTER_TYPE)
799 flags = flags_from_decl_or_type (TREE_TYPE (t));
800 else
801 flags = 0;
804 return flags;
807 /* Precompute all register parameters as described by ARGS, storing values
808 into fields within the ARGS array.
810 NUM_ACTUALS indicates the total number elements in the ARGS array.
812 Set REG_PARM_SEEN if we encounter a register parameter. */
814 static void
815 precompute_register_parameters (int num_actuals, struct arg_data *args,
816 int *reg_parm_seen)
818 int i;
820 *reg_parm_seen = 0;
822 for (i = 0; i < num_actuals; i++)
823 if (args[i].reg != 0 && ! args[i].pass_on_stack)
825 *reg_parm_seen = 1;
827 if (args[i].value == 0)
829 push_temp_slots ();
830 args[i].value = expand_normal (args[i].tree_value);
831 preserve_temp_slots (args[i].value);
832 pop_temp_slots ();
835 /* If we are to promote the function arg to a wider mode,
836 do it now. */
838 if (args[i].mode != TYPE_MODE (TREE_TYPE (args[i].tree_value)))
839 args[i].value
840 = convert_modes (args[i].mode,
841 TYPE_MODE (TREE_TYPE (args[i].tree_value)),
842 args[i].value, args[i].unsignedp);
844 /* If the value is a non-legitimate constant, force it into a
845 pseudo now. TLS symbols sometimes need a call to resolve. */
846 if (CONSTANT_P (args[i].value)
847 && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
848 args[i].value = force_reg (args[i].mode, args[i].value);
850 /* If we're going to have to load the value by parts, pull the
851 parts into pseudos. The part extraction process can involve
852 non-trivial computation. */
853 if (GET_CODE (args[i].reg) == PARALLEL)
855 tree type = TREE_TYPE (args[i].tree_value);
856 args[i].parallel_value
857 = emit_group_load_into_temps (args[i].reg, args[i].value,
858 type, int_size_in_bytes (type));
861 /* If the value is expensive, and we are inside an appropriately
862 short loop, put the value into a pseudo and then put the pseudo
863 into the hard reg.
865 For small register classes, also do this if this call uses
866 register parameters. This is to avoid reload conflicts while
867 loading the parameters registers. */
869 else if ((! (REG_P (args[i].value)
870 || (GET_CODE (args[i].value) == SUBREG
871 && REG_P (SUBREG_REG (args[i].value)))))
872 && args[i].mode != BLKmode
873 && set_src_cost (args[i].value, optimize_insn_for_speed_p ())
874 > COSTS_N_INSNS (1)
875 && ((*reg_parm_seen
876 && targetm.small_register_classes_for_mode_p (args[i].mode))
877 || optimize))
878 args[i].value = copy_to_mode_reg (args[i].mode, args[i].value);
882 #ifdef REG_PARM_STACK_SPACE
884 /* The argument list is the property of the called routine and it
885 may clobber it. If the fixed area has been used for previous
886 parameters, we must save and restore it. */
888 static rtx
889 save_fixed_argument_area (int reg_parm_stack_space, rtx argblock, int *low_to_save, int *high_to_save)
891 int low;
892 int high;
894 /* Compute the boundary of the area that needs to be saved, if any. */
895 high = reg_parm_stack_space;
896 #ifdef ARGS_GROW_DOWNWARD
897 high += 1;
898 #endif
899 if (high > highest_outgoing_arg_in_use)
900 high = highest_outgoing_arg_in_use;
902 for (low = 0; low < high; low++)
903 if (stack_usage_map[low] != 0)
905 int num_to_save;
906 enum machine_mode save_mode;
907 int delta;
908 rtx addr;
909 rtx stack_area;
910 rtx save_area;
912 while (stack_usage_map[--high] == 0)
915 *low_to_save = low;
916 *high_to_save = high;
918 num_to_save = high - low + 1;
919 save_mode = mode_for_size (num_to_save * BITS_PER_UNIT, MODE_INT, 1);
921 /* If we don't have the required alignment, must do this
922 in BLKmode. */
923 if ((low & (MIN (GET_MODE_SIZE (save_mode),
924 BIGGEST_ALIGNMENT / UNITS_PER_WORD) - 1)))
925 save_mode = BLKmode;
927 #ifdef ARGS_GROW_DOWNWARD
928 delta = -high;
929 #else
930 delta = low;
931 #endif
932 addr = plus_constant (Pmode, argblock, delta);
933 stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr));
935 set_mem_align (stack_area, PARM_BOUNDARY);
936 if (save_mode == BLKmode)
938 save_area = assign_stack_temp (BLKmode, num_to_save);
939 emit_block_move (validize_mem (save_area), stack_area,
940 GEN_INT (num_to_save), BLOCK_OP_CALL_PARM);
942 else
944 save_area = gen_reg_rtx (save_mode);
945 emit_move_insn (save_area, stack_area);
948 return save_area;
951 return NULL_RTX;
954 static void
955 restore_fixed_argument_area (rtx save_area, rtx argblock, int high_to_save, int low_to_save)
957 enum machine_mode save_mode = GET_MODE (save_area);
958 int delta;
959 rtx addr, stack_area;
961 #ifdef ARGS_GROW_DOWNWARD
962 delta = -high_to_save;
963 #else
964 delta = low_to_save;
965 #endif
966 addr = plus_constant (Pmode, argblock, delta);
967 stack_area = gen_rtx_MEM (save_mode, memory_address (save_mode, addr));
968 set_mem_align (stack_area, PARM_BOUNDARY);
970 if (save_mode != BLKmode)
971 emit_move_insn (stack_area, save_area);
972 else
973 emit_block_move (stack_area, validize_mem (save_area),
974 GEN_INT (high_to_save - low_to_save + 1),
975 BLOCK_OP_CALL_PARM);
977 #endif /* REG_PARM_STACK_SPACE */
979 /* If any elements in ARGS refer to parameters that are to be passed in
980 registers, but not in memory, and whose alignment does not permit a
981 direct copy into registers. Copy the values into a group of pseudos
982 which we will later copy into the appropriate hard registers.
984 Pseudos for each unaligned argument will be stored into the array
985 args[argnum].aligned_regs. The caller is responsible for deallocating
986 the aligned_regs array if it is nonzero. */
988 static void
989 store_unaligned_arguments_into_pseudos (struct arg_data *args, int num_actuals)
991 int i, j;
993 for (i = 0; i < num_actuals; i++)
994 if (args[i].reg != 0 && ! args[i].pass_on_stack
995 && GET_CODE (args[i].reg) != PARALLEL
996 && args[i].mode == BLKmode
997 && MEM_P (args[i].value)
998 && (MEM_ALIGN (args[i].value)
999 < (unsigned int) MIN (BIGGEST_ALIGNMENT, BITS_PER_WORD)))
1001 int bytes = int_size_in_bytes (TREE_TYPE (args[i].tree_value));
1002 int endian_correction = 0;
1004 if (args[i].partial)
1006 gcc_assert (args[i].partial % UNITS_PER_WORD == 0);
1007 args[i].n_aligned_regs = args[i].partial / UNITS_PER_WORD;
1009 else
1011 args[i].n_aligned_regs
1012 = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
1015 args[i].aligned_regs = XNEWVEC (rtx, args[i].n_aligned_regs);
1017 /* Structures smaller than a word are normally aligned to the
1018 least significant byte. On a BYTES_BIG_ENDIAN machine,
1019 this means we must skip the empty high order bytes when
1020 calculating the bit offset. */
1021 if (bytes < UNITS_PER_WORD
1022 #ifdef BLOCK_REG_PADDING
1023 && (BLOCK_REG_PADDING (args[i].mode,
1024 TREE_TYPE (args[i].tree_value), 1)
1025 == downward)
1026 #else
1027 && BYTES_BIG_ENDIAN
1028 #endif
1030 endian_correction = BITS_PER_WORD - bytes * BITS_PER_UNIT;
1032 for (j = 0; j < args[i].n_aligned_regs; j++)
1034 rtx reg = gen_reg_rtx (word_mode);
1035 rtx word = operand_subword_force (args[i].value, j, BLKmode);
1036 int bitsize = MIN (bytes * BITS_PER_UNIT, BITS_PER_WORD);
1038 args[i].aligned_regs[j] = reg;
1039 word = extract_bit_field (word, bitsize, 0, 1, NULL_RTX,
1040 word_mode, word_mode);
1042 /* There is no need to restrict this code to loading items
1043 in TYPE_ALIGN sized hunks. The bitfield instructions can
1044 load up entire word sized registers efficiently.
1046 ??? This may not be needed anymore.
1047 We use to emit a clobber here but that doesn't let later
1048 passes optimize the instructions we emit. By storing 0 into
1049 the register later passes know the first AND to zero out the
1050 bitfield being set in the register is unnecessary. The store
1051 of 0 will be deleted as will at least the first AND. */
1053 emit_move_insn (reg, const0_rtx);
1055 bytes -= bitsize / BITS_PER_UNIT;
1056 store_bit_field (reg, bitsize, endian_correction, 0, 0,
1057 word_mode, word);
1062 /* Fill in ARGS_SIZE and ARGS array based on the parameters found in
1063 CALL_EXPR EXP.
1065 NUM_ACTUALS is the total number of parameters.
1067 N_NAMED_ARGS is the total number of named arguments.
1069 STRUCT_VALUE_ADDR_VALUE is the implicit argument for a struct return
1070 value, or null.
1072 FNDECL is the tree code for the target of this call (if known)
1074 ARGS_SO_FAR holds state needed by the target to know where to place
1075 the next argument.
1077 REG_PARM_STACK_SPACE is the number of bytes of stack space reserved
1078 for arguments which are passed in registers.
1080 OLD_STACK_LEVEL is a pointer to an rtx which olds the old stack level
1081 and may be modified by this routine.
1083 OLD_PENDING_ADJ, MUST_PREALLOCATE and FLAGS are pointers to integer
1084 flags which may may be modified by this routine.
1086 MAY_TAILCALL is cleared if we encounter an invisible pass-by-reference
1087 that requires allocation of stack space.
1089 CALL_FROM_THUNK_P is true if this call is the jump from a thunk to
1090 the thunked-to function. */
1092 static void
1093 initialize_argument_information (int num_actuals ATTRIBUTE_UNUSED,
1094 struct arg_data *args,
1095 struct args_size *args_size,
1096 int n_named_args ATTRIBUTE_UNUSED,
1097 tree exp, tree struct_value_addr_value,
1098 tree fndecl, tree fntype,
1099 cumulative_args_t args_so_far,
1100 int reg_parm_stack_space,
1101 rtx *old_stack_level, int *old_pending_adj,
1102 int *must_preallocate, int *ecf_flags,
1103 bool *may_tailcall, bool call_from_thunk_p)
1105 CUMULATIVE_ARGS *args_so_far_pnt = get_cumulative_args (args_so_far);
1106 location_t loc = EXPR_LOCATION (exp);
1107 /* 1 if scanning parms front to back, -1 if scanning back to front. */
1108 int inc;
1110 /* Count arg position in order args appear. */
1111 int argpos;
1113 int i;
1115 args_size->constant = 0;
1116 args_size->var = 0;
1118 /* In this loop, we consider args in the order they are written.
1119 We fill up ARGS from the front or from the back if necessary
1120 so that in any case the first arg to be pushed ends up at the front. */
1122 if (PUSH_ARGS_REVERSED)
1124 i = num_actuals - 1, inc = -1;
1125 /* In this case, must reverse order of args
1126 so that we compute and push the last arg first. */
1128 else
1130 i = 0, inc = 1;
1133 /* First fill in the actual arguments in the ARGS array, splitting
1134 complex arguments if necessary. */
1136 int j = i;
1137 call_expr_arg_iterator iter;
1138 tree arg;
1140 if (struct_value_addr_value)
1142 args[j].tree_value = struct_value_addr_value;
1143 j += inc;
1145 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
1147 tree argtype = TREE_TYPE (arg);
1148 if (targetm.calls.split_complex_arg
1149 && argtype
1150 && TREE_CODE (argtype) == COMPLEX_TYPE
1151 && targetm.calls.split_complex_arg (argtype))
1153 tree subtype = TREE_TYPE (argtype);
1154 args[j].tree_value = build1 (REALPART_EXPR, subtype, arg);
1155 j += inc;
1156 args[j].tree_value = build1 (IMAGPART_EXPR, subtype, arg);
1158 else
1159 args[j].tree_value = arg;
1160 j += inc;
1164 /* I counts args in order (to be) pushed; ARGPOS counts in order written. */
1165 for (argpos = 0; argpos < num_actuals; i += inc, argpos++)
1167 tree type = TREE_TYPE (args[i].tree_value);
1168 int unsignedp;
1169 enum machine_mode mode;
1171 /* Replace erroneous argument with constant zero. */
1172 if (type == error_mark_node || !COMPLETE_TYPE_P (type))
1173 args[i].tree_value = integer_zero_node, type = integer_type_node;
1175 /* If TYPE is a transparent union or record, pass things the way
1176 we would pass the first field of the union or record. We have
1177 already verified that the modes are the same. */
1178 if ((TREE_CODE (type) == UNION_TYPE || TREE_CODE (type) == RECORD_TYPE)
1179 && TYPE_TRANSPARENT_AGGR (type))
1180 type = TREE_TYPE (first_field (type));
1182 /* Decide where to pass this arg.
1184 args[i].reg is nonzero if all or part is passed in registers.
1186 args[i].partial is nonzero if part but not all is passed in registers,
1187 and the exact value says how many bytes are passed in registers.
1189 args[i].pass_on_stack is nonzero if the argument must at least be
1190 computed on the stack. It may then be loaded back into registers
1191 if args[i].reg is nonzero.
1193 These decisions are driven by the FUNCTION_... macros and must agree
1194 with those made by function.c. */
1196 /* See if this argument should be passed by invisible reference. */
1197 if (pass_by_reference (args_so_far_pnt, TYPE_MODE (type),
1198 type, argpos < n_named_args))
1200 bool callee_copies;
1201 tree base = NULL_TREE;
1203 callee_copies
1204 = reference_callee_copied (args_so_far_pnt, TYPE_MODE (type),
1205 type, argpos < n_named_args);
1207 /* If we're compiling a thunk, pass through invisible references
1208 instead of making a copy. */
1209 if (call_from_thunk_p
1210 || (callee_copies
1211 && !TREE_ADDRESSABLE (type)
1212 && (base = get_base_address (args[i].tree_value))
1213 && TREE_CODE (base) != SSA_NAME
1214 && (!DECL_P (base) || MEM_P (DECL_RTL (base)))))
1216 mark_addressable (args[i].tree_value);
1218 /* We can't use sibcalls if a callee-copied argument is
1219 stored in the current function's frame. */
1220 if (!call_from_thunk_p && DECL_P (base) && !TREE_STATIC (base))
1221 *may_tailcall = false;
1223 args[i].tree_value = build_fold_addr_expr_loc (loc,
1224 args[i].tree_value);
1225 type = TREE_TYPE (args[i].tree_value);
1227 if (*ecf_flags & ECF_CONST)
1228 *ecf_flags &= ~(ECF_CONST | ECF_LOOPING_CONST_OR_PURE);
1230 else
1232 /* We make a copy of the object and pass the address to the
1233 function being called. */
1234 rtx copy;
1236 if (!COMPLETE_TYPE_P (type)
1237 || TREE_CODE (TYPE_SIZE_UNIT (type)) != INTEGER_CST
1238 || (flag_stack_check == GENERIC_STACK_CHECK
1239 && compare_tree_int (TYPE_SIZE_UNIT (type),
1240 STACK_CHECK_MAX_VAR_SIZE) > 0))
1242 /* This is a variable-sized object. Make space on the stack
1243 for it. */
1244 rtx size_rtx = expr_size (args[i].tree_value);
1246 if (*old_stack_level == 0)
1248 emit_stack_save (SAVE_BLOCK, old_stack_level);
1249 *old_pending_adj = pending_stack_adjust;
1250 pending_stack_adjust = 0;
1253 /* We can pass TRUE as the 4th argument because we just
1254 saved the stack pointer and will restore it right after
1255 the call. */
1256 copy = allocate_dynamic_stack_space (size_rtx,
1257 TYPE_ALIGN (type),
1258 TYPE_ALIGN (type),
1259 true);
1260 copy = gen_rtx_MEM (BLKmode, copy);
1261 set_mem_attributes (copy, type, 1);
1263 else
1264 copy = assign_temp (type, 1, 0);
1266 store_expr (args[i].tree_value, copy, 0, false);
1268 /* Just change the const function to pure and then let
1269 the next test clear the pure based on
1270 callee_copies. */
1271 if (*ecf_flags & ECF_CONST)
1273 *ecf_flags &= ~ECF_CONST;
1274 *ecf_flags |= ECF_PURE;
1277 if (!callee_copies && *ecf_flags & ECF_PURE)
1278 *ecf_flags &= ~(ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
1280 args[i].tree_value
1281 = build_fold_addr_expr_loc (loc, make_tree (type, copy));
1282 type = TREE_TYPE (args[i].tree_value);
1283 *may_tailcall = false;
1287 unsignedp = TYPE_UNSIGNED (type);
1288 mode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
1289 fndecl ? TREE_TYPE (fndecl) : fntype, 0);
1291 args[i].unsignedp = unsignedp;
1292 args[i].mode = mode;
1294 args[i].reg = targetm.calls.function_arg (args_so_far, mode, type,
1295 argpos < n_named_args);
1297 /* If this is a sibling call and the machine has register windows, the
1298 register window has to be unwinded before calling the routine, so
1299 arguments have to go into the incoming registers. */
1300 if (targetm.calls.function_incoming_arg != targetm.calls.function_arg)
1301 args[i].tail_call_reg
1302 = targetm.calls.function_incoming_arg (args_so_far, mode, type,
1303 argpos < n_named_args);
1304 else
1305 args[i].tail_call_reg = args[i].reg;
1307 if (args[i].reg)
1308 args[i].partial
1309 = targetm.calls.arg_partial_bytes (args_so_far, mode, type,
1310 argpos < n_named_args);
1312 args[i].pass_on_stack = targetm.calls.must_pass_in_stack (mode, type);
1314 /* If FUNCTION_ARG returned a (parallel [(expr_list (nil) ...) ...]),
1315 it means that we are to pass this arg in the register(s) designated
1316 by the PARALLEL, but also to pass it in the stack. */
1317 if (args[i].reg && GET_CODE (args[i].reg) == PARALLEL
1318 && XEXP (XVECEXP (args[i].reg, 0, 0), 0) == 0)
1319 args[i].pass_on_stack = 1;
1321 /* If this is an addressable type, we must preallocate the stack
1322 since we must evaluate the object into its final location.
1324 If this is to be passed in both registers and the stack, it is simpler
1325 to preallocate. */
1326 if (TREE_ADDRESSABLE (type)
1327 || (args[i].pass_on_stack && args[i].reg != 0))
1328 *must_preallocate = 1;
1330 /* Compute the stack-size of this argument. */
1331 if (args[i].reg == 0 || args[i].partial != 0
1332 || reg_parm_stack_space > 0
1333 || args[i].pass_on_stack)
1334 locate_and_pad_parm (mode, type,
1335 #ifdef STACK_PARMS_IN_REG_PARM_AREA
1337 #else
1338 args[i].reg != 0,
1339 #endif
1340 reg_parm_stack_space,
1341 args[i].pass_on_stack ? 0 : args[i].partial,
1342 fndecl, args_size, &args[i].locate);
1343 #ifdef BLOCK_REG_PADDING
1344 else
1345 /* The argument is passed entirely in registers. See at which
1346 end it should be padded. */
1347 args[i].locate.where_pad =
1348 BLOCK_REG_PADDING (mode, type,
1349 int_size_in_bytes (type) <= UNITS_PER_WORD);
1350 #endif
1352 /* Update ARGS_SIZE, the total stack space for args so far. */
1354 args_size->constant += args[i].locate.size.constant;
1355 if (args[i].locate.size.var)
1356 ADD_PARM_SIZE (*args_size, args[i].locate.size.var);
1358 /* Increment ARGS_SO_FAR, which has info about which arg-registers
1359 have been used, etc. */
1361 targetm.calls.function_arg_advance (args_so_far, TYPE_MODE (type),
1362 type, argpos < n_named_args);
1366 /* Update ARGS_SIZE to contain the total size for the argument block.
1367 Return the original constant component of the argument block's size.
1369 REG_PARM_STACK_SPACE holds the number of bytes of stack space reserved
1370 for arguments passed in registers. */
1372 static int
1373 compute_argument_block_size (int reg_parm_stack_space,
1374 struct args_size *args_size,
1375 tree fndecl ATTRIBUTE_UNUSED,
1376 tree fntype ATTRIBUTE_UNUSED,
1377 int preferred_stack_boundary ATTRIBUTE_UNUSED)
1379 int unadjusted_args_size = args_size->constant;
1381 /* For accumulate outgoing args mode we don't need to align, since the frame
1382 will be already aligned. Align to STACK_BOUNDARY in order to prevent
1383 backends from generating misaligned frame sizes. */
1384 if (ACCUMULATE_OUTGOING_ARGS && preferred_stack_boundary > STACK_BOUNDARY)
1385 preferred_stack_boundary = STACK_BOUNDARY;
1387 /* Compute the actual size of the argument block required. The variable
1388 and constant sizes must be combined, the size may have to be rounded,
1389 and there may be a minimum required size. */
1391 if (args_size->var)
1393 args_size->var = ARGS_SIZE_TREE (*args_size);
1394 args_size->constant = 0;
1396 preferred_stack_boundary /= BITS_PER_UNIT;
1397 if (preferred_stack_boundary > 1)
1399 /* We don't handle this case yet. To handle it correctly we have
1400 to add the delta, round and subtract the delta.
1401 Currently no machine description requires this support. */
1402 gcc_assert (!(stack_pointer_delta & (preferred_stack_boundary - 1)));
1403 args_size->var = round_up (args_size->var, preferred_stack_boundary);
1406 if (reg_parm_stack_space > 0)
1408 args_size->var
1409 = size_binop (MAX_EXPR, args_size->var,
1410 ssize_int (reg_parm_stack_space));
1412 /* The area corresponding to register parameters is not to count in
1413 the size of the block we need. So make the adjustment. */
1414 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
1415 args_size->var
1416 = size_binop (MINUS_EXPR, args_size->var,
1417 ssize_int (reg_parm_stack_space));
1420 else
1422 preferred_stack_boundary /= BITS_PER_UNIT;
1423 if (preferred_stack_boundary < 1)
1424 preferred_stack_boundary = 1;
1425 args_size->constant = (((args_size->constant
1426 + stack_pointer_delta
1427 + preferred_stack_boundary - 1)
1428 / preferred_stack_boundary
1429 * preferred_stack_boundary)
1430 - stack_pointer_delta);
1432 args_size->constant = MAX (args_size->constant,
1433 reg_parm_stack_space);
1435 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
1436 args_size->constant -= reg_parm_stack_space;
1438 return unadjusted_args_size;
1441 /* Precompute parameters as needed for a function call.
1443 FLAGS is mask of ECF_* constants.
1445 NUM_ACTUALS is the number of arguments.
1447 ARGS is an array containing information for each argument; this
1448 routine fills in the INITIAL_VALUE and VALUE fields for each
1449 precomputed argument. */
1451 static void
1452 precompute_arguments (int num_actuals, struct arg_data *args)
1454 int i;
1456 /* If this is a libcall, then precompute all arguments so that we do not
1457 get extraneous instructions emitted as part of the libcall sequence. */
1459 /* If we preallocated the stack space, and some arguments must be passed
1460 on the stack, then we must precompute any parameter which contains a
1461 function call which will store arguments on the stack.
1462 Otherwise, evaluating the parameter may clobber previous parameters
1463 which have already been stored into the stack. (we have code to avoid
1464 such case by saving the outgoing stack arguments, but it results in
1465 worse code) */
1466 if (!ACCUMULATE_OUTGOING_ARGS)
1467 return;
1469 for (i = 0; i < num_actuals; i++)
1471 tree type;
1472 enum machine_mode mode;
1474 if (TREE_CODE (args[i].tree_value) != CALL_EXPR)
1475 continue;
1477 /* If this is an addressable type, we cannot pre-evaluate it. */
1478 type = TREE_TYPE (args[i].tree_value);
1479 gcc_assert (!TREE_ADDRESSABLE (type));
1481 args[i].initial_value = args[i].value
1482 = expand_normal (args[i].tree_value);
1484 mode = TYPE_MODE (type);
1485 if (mode != args[i].mode)
1487 int unsignedp = args[i].unsignedp;
1488 args[i].value
1489 = convert_modes (args[i].mode, mode,
1490 args[i].value, args[i].unsignedp);
1492 /* CSE will replace this only if it contains args[i].value
1493 pseudo, so convert it down to the declared mode using
1494 a SUBREG. */
1495 if (REG_P (args[i].value)
1496 && GET_MODE_CLASS (args[i].mode) == MODE_INT
1497 && promote_mode (type, mode, &unsignedp) != args[i].mode)
1499 args[i].initial_value
1500 = gen_lowpart_SUBREG (mode, args[i].value);
1501 SUBREG_PROMOTED_VAR_P (args[i].initial_value) = 1;
1502 SUBREG_PROMOTED_UNSIGNED_SET (args[i].initial_value,
1503 args[i].unsignedp);
1509 /* Given the current state of MUST_PREALLOCATE and information about
1510 arguments to a function call in NUM_ACTUALS, ARGS and ARGS_SIZE,
1511 compute and return the final value for MUST_PREALLOCATE. */
1513 static int
1514 finalize_must_preallocate (int must_preallocate, int num_actuals,
1515 struct arg_data *args, struct args_size *args_size)
1517 /* See if we have or want to preallocate stack space.
1519 If we would have to push a partially-in-regs parm
1520 before other stack parms, preallocate stack space instead.
1522 If the size of some parm is not a multiple of the required stack
1523 alignment, we must preallocate.
1525 If the total size of arguments that would otherwise create a copy in
1526 a temporary (such as a CALL) is more than half the total argument list
1527 size, preallocation is faster.
1529 Another reason to preallocate is if we have a machine (like the m88k)
1530 where stack alignment is required to be maintained between every
1531 pair of insns, not just when the call is made. However, we assume here
1532 that such machines either do not have push insns (and hence preallocation
1533 would occur anyway) or the problem is taken care of with
1534 PUSH_ROUNDING. */
1536 if (! must_preallocate)
1538 int partial_seen = 0;
1539 int copy_to_evaluate_size = 0;
1540 int i;
1542 for (i = 0; i < num_actuals && ! must_preallocate; i++)
1544 if (args[i].partial > 0 && ! args[i].pass_on_stack)
1545 partial_seen = 1;
1546 else if (partial_seen && args[i].reg == 0)
1547 must_preallocate = 1;
1549 if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode
1550 && (TREE_CODE (args[i].tree_value) == CALL_EXPR
1551 || TREE_CODE (args[i].tree_value) == TARGET_EXPR
1552 || TREE_CODE (args[i].tree_value) == COND_EXPR
1553 || TREE_ADDRESSABLE (TREE_TYPE (args[i].tree_value))))
1554 copy_to_evaluate_size
1555 += int_size_in_bytes (TREE_TYPE (args[i].tree_value));
1558 if (copy_to_evaluate_size * 2 >= args_size->constant
1559 && args_size->constant > 0)
1560 must_preallocate = 1;
1562 return must_preallocate;
1565 /* If we preallocated stack space, compute the address of each argument
1566 and store it into the ARGS array.
1568 We need not ensure it is a valid memory address here; it will be
1569 validized when it is used.
1571 ARGBLOCK is an rtx for the address of the outgoing arguments. */
1573 static void
1574 compute_argument_addresses (struct arg_data *args, rtx argblock, int num_actuals)
1576 if (argblock)
1578 rtx arg_reg = argblock;
1579 int i, arg_offset = 0;
1581 if (GET_CODE (argblock) == PLUS)
1582 arg_reg = XEXP (argblock, 0), arg_offset = INTVAL (XEXP (argblock, 1));
1584 for (i = 0; i < num_actuals; i++)
1586 rtx offset = ARGS_SIZE_RTX (args[i].locate.offset);
1587 rtx slot_offset = ARGS_SIZE_RTX (args[i].locate.slot_offset);
1588 rtx addr;
1589 unsigned int align, boundary;
1590 unsigned int units_on_stack = 0;
1591 enum machine_mode partial_mode = VOIDmode;
1593 /* Skip this parm if it will not be passed on the stack. */
1594 if (! args[i].pass_on_stack
1595 && args[i].reg != 0
1596 && args[i].partial == 0)
1597 continue;
1599 if (CONST_INT_P (offset))
1600 addr = plus_constant (Pmode, arg_reg, INTVAL (offset));
1601 else
1602 addr = gen_rtx_PLUS (Pmode, arg_reg, offset);
1604 addr = plus_constant (Pmode, addr, arg_offset);
1606 if (args[i].partial != 0)
1608 /* Only part of the parameter is being passed on the stack.
1609 Generate a simple memory reference of the correct size. */
1610 units_on_stack = args[i].locate.size.constant;
1611 partial_mode = mode_for_size (units_on_stack * BITS_PER_UNIT,
1612 MODE_INT, 1);
1613 args[i].stack = gen_rtx_MEM (partial_mode, addr);
1614 set_mem_size (args[i].stack, units_on_stack);
1616 else
1618 args[i].stack = gen_rtx_MEM (args[i].mode, addr);
1619 set_mem_attributes (args[i].stack,
1620 TREE_TYPE (args[i].tree_value), 1);
1622 align = BITS_PER_UNIT;
1623 boundary = args[i].locate.boundary;
1624 if (args[i].locate.where_pad != downward)
1625 align = boundary;
1626 else if (CONST_INT_P (offset))
1628 align = INTVAL (offset) * BITS_PER_UNIT | boundary;
1629 align = align & -align;
1631 set_mem_align (args[i].stack, align);
1633 if (CONST_INT_P (slot_offset))
1634 addr = plus_constant (Pmode, arg_reg, INTVAL (slot_offset));
1635 else
1636 addr = gen_rtx_PLUS (Pmode, arg_reg, slot_offset);
1638 addr = plus_constant (Pmode, addr, arg_offset);
1640 if (args[i].partial != 0)
1642 /* Only part of the parameter is being passed on the stack.
1643 Generate a simple memory reference of the correct size.
1645 args[i].stack_slot = gen_rtx_MEM (partial_mode, addr);
1646 set_mem_size (args[i].stack_slot, units_on_stack);
1648 else
1650 args[i].stack_slot = gen_rtx_MEM (args[i].mode, addr);
1651 set_mem_attributes (args[i].stack_slot,
1652 TREE_TYPE (args[i].tree_value), 1);
1654 set_mem_align (args[i].stack_slot, args[i].locate.boundary);
1656 /* Function incoming arguments may overlap with sibling call
1657 outgoing arguments and we cannot allow reordering of reads
1658 from function arguments with stores to outgoing arguments
1659 of sibling calls. */
1660 set_mem_alias_set (args[i].stack, 0);
1661 set_mem_alias_set (args[i].stack_slot, 0);
1666 /* Given a FNDECL and EXP, return an rtx suitable for use as a target address
1667 in a call instruction.
1669 FNDECL is the tree node for the target function. For an indirect call
1670 FNDECL will be NULL_TREE.
1672 ADDR is the operand 0 of CALL_EXPR for this call. */
1674 static rtx
1675 rtx_for_function_call (tree fndecl, tree addr)
1677 rtx funexp;
1679 /* Get the function to call, in the form of RTL. */
1680 if (fndecl)
1682 if (!TREE_USED (fndecl) && fndecl != current_function_decl)
1683 TREE_USED (fndecl) = 1;
1685 /* Get a SYMBOL_REF rtx for the function address. */
1686 funexp = XEXP (DECL_RTL (fndecl), 0);
1688 else
1689 /* Generate an rtx (probably a pseudo-register) for the address. */
1691 push_temp_slots ();
1692 funexp = expand_normal (addr);
1693 pop_temp_slots (); /* FUNEXP can't be BLKmode. */
1695 return funexp;
1698 /* Internal state for internal_arg_pointer_based_exp and its helpers. */
1699 static struct
1701 /* Last insn that has been scanned by internal_arg_pointer_based_exp_scan,
1702 or NULL_RTX if none has been scanned yet. */
1703 rtx scan_start;
1704 /* Vector indexed by REGNO - FIRST_PSEUDO_REGISTER, recording if a pseudo is
1705 based on crtl->args.internal_arg_pointer. The element is NULL_RTX if the
1706 pseudo isn't based on it, a CONST_INT offset if the pseudo is based on it
1707 with fixed offset, or PC if this is with variable or unknown offset. */
1708 vec<rtx> cache;
1709 } internal_arg_pointer_exp_state;
1711 static rtx internal_arg_pointer_based_exp (rtx, bool);
1713 /* Helper function for internal_arg_pointer_based_exp. Scan insns in
1714 the tail call sequence, starting with first insn that hasn't been
1715 scanned yet, and note for each pseudo on the LHS whether it is based
1716 on crtl->args.internal_arg_pointer or not, and what offset from that
1717 that pointer it has. */
1719 static void
1720 internal_arg_pointer_based_exp_scan (void)
1722 rtx insn, scan_start = internal_arg_pointer_exp_state.scan_start;
1724 if (scan_start == NULL_RTX)
1725 insn = get_insns ();
1726 else
1727 insn = NEXT_INSN (scan_start);
1729 while (insn)
1731 rtx set = single_set (insn);
1732 if (set && REG_P (SET_DEST (set)) && !HARD_REGISTER_P (SET_DEST (set)))
1734 rtx val = NULL_RTX;
1735 unsigned int idx = REGNO (SET_DEST (set)) - FIRST_PSEUDO_REGISTER;
1736 /* Punt on pseudos set multiple times. */
1737 if (idx < internal_arg_pointer_exp_state.cache.length ()
1738 && (internal_arg_pointer_exp_state.cache[idx]
1739 != NULL_RTX))
1740 val = pc_rtx;
1741 else
1742 val = internal_arg_pointer_based_exp (SET_SRC (set), false);
1743 if (val != NULL_RTX)
1745 if (idx >= internal_arg_pointer_exp_state.cache.length ())
1746 internal_arg_pointer_exp_state.cache
1747 .safe_grow_cleared (idx + 1);
1748 internal_arg_pointer_exp_state.cache[idx] = val;
1751 if (NEXT_INSN (insn) == NULL_RTX)
1752 scan_start = insn;
1753 insn = NEXT_INSN (insn);
1756 internal_arg_pointer_exp_state.scan_start = scan_start;
1759 /* Helper function for internal_arg_pointer_based_exp, called through
1760 for_each_rtx. Return 1 if *LOC is a register based on
1761 crtl->args.internal_arg_pointer. Return -1 if *LOC is not based on it
1762 and the subexpressions need not be examined. Otherwise return 0. */
1764 static int
1765 internal_arg_pointer_based_exp_1 (rtx *loc, void *data ATTRIBUTE_UNUSED)
1767 if (REG_P (*loc) && internal_arg_pointer_based_exp (*loc, false) != NULL_RTX)
1768 return 1;
1769 if (MEM_P (*loc))
1770 return -1;
1771 return 0;
1774 /* Compute whether RTL is based on crtl->args.internal_arg_pointer. Return
1775 NULL_RTX if RTL isn't based on it, a CONST_INT offset if RTL is based on
1776 it with fixed offset, or PC if this is with variable or unknown offset.
1777 TOPLEVEL is true if the function is invoked at the topmost level. */
1779 static rtx
1780 internal_arg_pointer_based_exp (rtx rtl, bool toplevel)
1782 if (CONSTANT_P (rtl))
1783 return NULL_RTX;
1785 if (rtl == crtl->args.internal_arg_pointer)
1786 return const0_rtx;
1788 if (REG_P (rtl) && HARD_REGISTER_P (rtl))
1789 return NULL_RTX;
1791 if (GET_CODE (rtl) == PLUS && CONST_INT_P (XEXP (rtl, 1)))
1793 rtx val = internal_arg_pointer_based_exp (XEXP (rtl, 0), toplevel);
1794 if (val == NULL_RTX || val == pc_rtx)
1795 return val;
1796 return plus_constant (Pmode, val, INTVAL (XEXP (rtl, 1)));
1799 /* When called at the topmost level, scan pseudo assignments in between the
1800 last scanned instruction in the tail call sequence and the latest insn
1801 in that sequence. */
1802 if (toplevel)
1803 internal_arg_pointer_based_exp_scan ();
1805 if (REG_P (rtl))
1807 unsigned int idx = REGNO (rtl) - FIRST_PSEUDO_REGISTER;
1808 if (idx < internal_arg_pointer_exp_state.cache.length ())
1809 return internal_arg_pointer_exp_state.cache[idx];
1811 return NULL_RTX;
1814 if (for_each_rtx (&rtl, internal_arg_pointer_based_exp_1, NULL))
1815 return pc_rtx;
1817 return NULL_RTX;
1820 /* Return true if and only if SIZE storage units (usually bytes)
1821 starting from address ADDR overlap with already clobbered argument
1822 area. This function is used to determine if we should give up a
1823 sibcall. */
1825 static bool
1826 mem_overlaps_already_clobbered_arg_p (rtx addr, unsigned HOST_WIDE_INT size)
1828 HOST_WIDE_INT i;
1829 rtx val;
1831 if (bitmap_empty_p (stored_args_map))
1832 return false;
1833 val = internal_arg_pointer_based_exp (addr, true);
1834 if (val == NULL_RTX)
1835 return false;
1836 else if (val == pc_rtx)
1837 return true;
1838 else
1839 i = INTVAL (val);
1840 #ifdef STACK_GROWS_DOWNWARD
1841 i -= crtl->args.pretend_args_size;
1842 #else
1843 i += crtl->args.pretend_args_size;
1844 #endif
1846 #ifdef ARGS_GROW_DOWNWARD
1847 i = -i - size;
1848 #endif
1849 if (size > 0)
1851 unsigned HOST_WIDE_INT k;
1853 for (k = 0; k < size; k++)
1854 if (i + k < SBITMAP_SIZE (stored_args_map)
1855 && bitmap_bit_p (stored_args_map, i + k))
1856 return true;
1859 return false;
1862 /* Do the register loads required for any wholly-register parms or any
1863 parms which are passed both on the stack and in a register. Their
1864 expressions were already evaluated.
1866 Mark all register-parms as living through the call, putting these USE
1867 insns in the CALL_INSN_FUNCTION_USAGE field.
1869 When IS_SIBCALL, perform the check_sibcall_argument_overlap
1870 checking, setting *SIBCALL_FAILURE if appropriate. */
1872 static void
1873 load_register_parameters (struct arg_data *args, int num_actuals,
1874 rtx *call_fusage, int flags, int is_sibcall,
1875 int *sibcall_failure)
1877 int i, j;
1879 for (i = 0; i < num_actuals; i++)
1881 rtx reg = ((flags & ECF_SIBCALL)
1882 ? args[i].tail_call_reg : args[i].reg);
1883 if (reg)
1885 int partial = args[i].partial;
1886 int nregs;
1887 int size = 0;
1888 rtx before_arg = get_last_insn ();
1889 /* Set non-negative if we must move a word at a time, even if
1890 just one word (e.g, partial == 4 && mode == DFmode). Set
1891 to -1 if we just use a normal move insn. This value can be
1892 zero if the argument is a zero size structure. */
1893 nregs = -1;
1894 if (GET_CODE (reg) == PARALLEL)
1896 else if (partial)
1898 gcc_assert (partial % UNITS_PER_WORD == 0);
1899 nregs = partial / UNITS_PER_WORD;
1901 else if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode)
1903 size = int_size_in_bytes (TREE_TYPE (args[i].tree_value));
1904 nregs = (size + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
1906 else
1907 size = GET_MODE_SIZE (args[i].mode);
1909 /* Handle calls that pass values in multiple non-contiguous
1910 locations. The Irix 6 ABI has examples of this. */
1912 if (GET_CODE (reg) == PARALLEL)
1913 emit_group_move (reg, args[i].parallel_value);
1915 /* If simple case, just do move. If normal partial, store_one_arg
1916 has already loaded the register for us. In all other cases,
1917 load the register(s) from memory. */
1919 else if (nregs == -1)
1921 emit_move_insn (reg, args[i].value);
1922 #ifdef BLOCK_REG_PADDING
1923 /* Handle case where we have a value that needs shifting
1924 up to the msb. eg. a QImode value and we're padding
1925 upward on a BYTES_BIG_ENDIAN machine. */
1926 if (size < UNITS_PER_WORD
1927 && (args[i].locate.where_pad
1928 == (BYTES_BIG_ENDIAN ? upward : downward)))
1930 rtx x;
1931 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
1933 /* Assigning REG here rather than a temp makes CALL_FUSAGE
1934 report the whole reg as used. Strictly speaking, the
1935 call only uses SIZE bytes at the msb end, but it doesn't
1936 seem worth generating rtl to say that. */
1937 reg = gen_rtx_REG (word_mode, REGNO (reg));
1938 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
1939 if (x != reg)
1940 emit_move_insn (reg, x);
1942 #endif
1945 /* If we have pre-computed the values to put in the registers in
1946 the case of non-aligned structures, copy them in now. */
1948 else if (args[i].n_aligned_regs != 0)
1949 for (j = 0; j < args[i].n_aligned_regs; j++)
1950 emit_move_insn (gen_rtx_REG (word_mode, REGNO (reg) + j),
1951 args[i].aligned_regs[j]);
1953 else if (partial == 0 || args[i].pass_on_stack)
1955 rtx mem = validize_mem (args[i].value);
1957 /* Check for overlap with already clobbered argument area,
1958 providing that this has non-zero size. */
1959 if (is_sibcall
1960 && (size == 0
1961 || mem_overlaps_already_clobbered_arg_p
1962 (XEXP (args[i].value, 0), size)))
1963 *sibcall_failure = 1;
1965 /* Handle a BLKmode that needs shifting. */
1966 if (nregs == 1 && size < UNITS_PER_WORD
1967 #ifdef BLOCK_REG_PADDING
1968 && args[i].locate.where_pad == downward
1969 #else
1970 && BYTES_BIG_ENDIAN
1971 #endif
1974 rtx tem = operand_subword_force (mem, 0, args[i].mode);
1975 rtx ri = gen_rtx_REG (word_mode, REGNO (reg));
1976 rtx x = gen_reg_rtx (word_mode);
1977 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
1978 enum tree_code dir = BYTES_BIG_ENDIAN ? RSHIFT_EXPR
1979 : LSHIFT_EXPR;
1981 emit_move_insn (x, tem);
1982 x = expand_shift (dir, word_mode, x, shift, ri, 1);
1983 if (x != ri)
1984 emit_move_insn (ri, x);
1986 else
1987 move_block_to_reg (REGNO (reg), mem, nregs, args[i].mode);
1990 /* When a parameter is a block, and perhaps in other cases, it is
1991 possible that it did a load from an argument slot that was
1992 already clobbered. */
1993 if (is_sibcall
1994 && check_sibcall_argument_overlap (before_arg, &args[i], 0))
1995 *sibcall_failure = 1;
1997 /* Handle calls that pass values in multiple non-contiguous
1998 locations. The Irix 6 ABI has examples of this. */
1999 if (GET_CODE (reg) == PARALLEL)
2000 use_group_regs (call_fusage, reg);
2001 else if (nregs == -1)
2002 use_reg_mode (call_fusage, reg,
2003 TYPE_MODE (TREE_TYPE (args[i].tree_value)));
2004 else if (nregs > 0)
2005 use_regs (call_fusage, REGNO (reg), nregs);
2010 /* We need to pop PENDING_STACK_ADJUST bytes. But, if the arguments
2011 wouldn't fill up an even multiple of PREFERRED_UNIT_STACK_BOUNDARY
2012 bytes, then we would need to push some additional bytes to pad the
2013 arguments. So, we compute an adjust to the stack pointer for an
2014 amount that will leave the stack under-aligned by UNADJUSTED_ARGS_SIZE
2015 bytes. Then, when the arguments are pushed the stack will be perfectly
2016 aligned. ARGS_SIZE->CONSTANT is set to the number of bytes that should
2017 be popped after the call. Returns the adjustment. */
2019 static int
2020 combine_pending_stack_adjustment_and_call (int unadjusted_args_size,
2021 struct args_size *args_size,
2022 unsigned int preferred_unit_stack_boundary)
2024 /* The number of bytes to pop so that the stack will be
2025 under-aligned by UNADJUSTED_ARGS_SIZE bytes. */
2026 HOST_WIDE_INT adjustment;
2027 /* The alignment of the stack after the arguments are pushed, if we
2028 just pushed the arguments without adjust the stack here. */
2029 unsigned HOST_WIDE_INT unadjusted_alignment;
2031 unadjusted_alignment
2032 = ((stack_pointer_delta + unadjusted_args_size)
2033 % preferred_unit_stack_boundary);
2035 /* We want to get rid of as many of the PENDING_STACK_ADJUST bytes
2036 as possible -- leaving just enough left to cancel out the
2037 UNADJUSTED_ALIGNMENT. In other words, we want to ensure that the
2038 PENDING_STACK_ADJUST is non-negative, and congruent to
2039 -UNADJUSTED_ALIGNMENT modulo the PREFERRED_UNIT_STACK_BOUNDARY. */
2041 /* Begin by trying to pop all the bytes. */
2042 unadjusted_alignment
2043 = (unadjusted_alignment
2044 - (pending_stack_adjust % preferred_unit_stack_boundary));
2045 adjustment = pending_stack_adjust;
2046 /* Push enough additional bytes that the stack will be aligned
2047 after the arguments are pushed. */
2048 if (preferred_unit_stack_boundary > 1)
2050 if (unadjusted_alignment > 0)
2051 adjustment -= preferred_unit_stack_boundary - unadjusted_alignment;
2052 else
2053 adjustment += unadjusted_alignment;
2056 /* Now, sets ARGS_SIZE->CONSTANT so that we pop the right number of
2057 bytes after the call. The right number is the entire
2058 PENDING_STACK_ADJUST less our ADJUSTMENT plus the amount required
2059 by the arguments in the first place. */
2060 args_size->constant
2061 = pending_stack_adjust - adjustment + unadjusted_args_size;
2063 return adjustment;
2066 /* Scan X expression if it does not dereference any argument slots
2067 we already clobbered by tail call arguments (as noted in stored_args_map
2068 bitmap).
2069 Return nonzero if X expression dereferences such argument slots,
2070 zero otherwise. */
2072 static int
2073 check_sibcall_argument_overlap_1 (rtx x)
2075 RTX_CODE code;
2076 int i, j;
2077 const char *fmt;
2079 if (x == NULL_RTX)
2080 return 0;
2082 code = GET_CODE (x);
2084 /* We need not check the operands of the CALL expression itself. */
2085 if (code == CALL)
2086 return 0;
2088 if (code == MEM)
2089 return mem_overlaps_already_clobbered_arg_p (XEXP (x, 0),
2090 GET_MODE_SIZE (GET_MODE (x)));
2092 /* Scan all subexpressions. */
2093 fmt = GET_RTX_FORMAT (code);
2094 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
2096 if (*fmt == 'e')
2098 if (check_sibcall_argument_overlap_1 (XEXP (x, i)))
2099 return 1;
2101 else if (*fmt == 'E')
2103 for (j = 0; j < XVECLEN (x, i); j++)
2104 if (check_sibcall_argument_overlap_1 (XVECEXP (x, i, j)))
2105 return 1;
2108 return 0;
2111 /* Scan sequence after INSN if it does not dereference any argument slots
2112 we already clobbered by tail call arguments (as noted in stored_args_map
2113 bitmap). If MARK_STORED_ARGS_MAP, add stack slots for ARG to
2114 stored_args_map bitmap afterwards (when ARG is a register MARK_STORED_ARGS_MAP
2115 should be 0). Return nonzero if sequence after INSN dereferences such argument
2116 slots, zero otherwise. */
2118 static int
2119 check_sibcall_argument_overlap (rtx insn, struct arg_data *arg, int mark_stored_args_map)
2121 int low, high;
2123 if (insn == NULL_RTX)
2124 insn = get_insns ();
2125 else
2126 insn = NEXT_INSN (insn);
2128 for (; insn; insn = NEXT_INSN (insn))
2129 if (INSN_P (insn)
2130 && check_sibcall_argument_overlap_1 (PATTERN (insn)))
2131 break;
2133 if (mark_stored_args_map)
2135 #ifdef ARGS_GROW_DOWNWARD
2136 low = -arg->locate.slot_offset.constant - arg->locate.size.constant;
2137 #else
2138 low = arg->locate.slot_offset.constant;
2139 #endif
2141 for (high = low + arg->locate.size.constant; low < high; low++)
2142 bitmap_set_bit (stored_args_map, low);
2144 return insn != NULL_RTX;
2147 /* Given that a function returns a value of mode MODE at the most
2148 significant end of hard register VALUE, shift VALUE left or right
2149 as specified by LEFT_P. Return true if some action was needed. */
2151 bool
2152 shift_return_value (enum machine_mode mode, bool left_p, rtx value)
2154 HOST_WIDE_INT shift;
2156 gcc_assert (REG_P (value) && HARD_REGISTER_P (value));
2157 shift = GET_MODE_BITSIZE (GET_MODE (value)) - GET_MODE_BITSIZE (mode);
2158 if (shift == 0)
2159 return false;
2161 /* Use ashr rather than lshr for right shifts. This is for the benefit
2162 of the MIPS port, which requires SImode values to be sign-extended
2163 when stored in 64-bit registers. */
2164 if (!force_expand_binop (GET_MODE (value), left_p ? ashl_optab : ashr_optab,
2165 value, GEN_INT (shift), value, 1, OPTAB_WIDEN))
2166 gcc_unreachable ();
2167 return true;
2170 /* If X is a likely-spilled register value, copy it to a pseudo
2171 register and return that register. Return X otherwise. */
2173 static rtx
2174 avoid_likely_spilled_reg (rtx x)
2176 rtx new_rtx;
2178 if (REG_P (x)
2179 && HARD_REGISTER_P (x)
2180 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (x))))
2182 /* Make sure that we generate a REG rather than a CONCAT.
2183 Moves into CONCATs can need nontrivial instructions,
2184 and the whole point of this function is to avoid
2185 using the hard register directly in such a situation. */
2186 generating_concat_p = 0;
2187 new_rtx = gen_reg_rtx (GET_MODE (x));
2188 generating_concat_p = 1;
2189 emit_move_insn (new_rtx, x);
2190 return new_rtx;
2192 return x;
2195 /* Generate all the code for a CALL_EXPR exp
2196 and return an rtx for its value.
2197 Store the value in TARGET (specified as an rtx) if convenient.
2198 If the value is stored in TARGET then TARGET is returned.
2199 If IGNORE is nonzero, then we ignore the value of the function call. */
2202 expand_call (tree exp, rtx target, int ignore)
2204 /* Nonzero if we are currently expanding a call. */
2205 static int currently_expanding_call = 0;
2207 /* RTX for the function to be called. */
2208 rtx funexp;
2209 /* Sequence of insns to perform a normal "call". */
2210 rtx normal_call_insns = NULL_RTX;
2211 /* Sequence of insns to perform a tail "call". */
2212 rtx tail_call_insns = NULL_RTX;
2213 /* Data type of the function. */
2214 tree funtype;
2215 tree type_arg_types;
2216 tree rettype;
2217 /* Declaration of the function being called,
2218 or 0 if the function is computed (not known by name). */
2219 tree fndecl = 0;
2220 /* The type of the function being called. */
2221 tree fntype;
2222 bool try_tail_call = CALL_EXPR_TAILCALL (exp);
2223 int pass;
2225 /* Register in which non-BLKmode value will be returned,
2226 or 0 if no value or if value is BLKmode. */
2227 rtx valreg;
2228 /* Address where we should return a BLKmode value;
2229 0 if value not BLKmode. */
2230 rtx structure_value_addr = 0;
2231 /* Nonzero if that address is being passed by treating it as
2232 an extra, implicit first parameter. Otherwise,
2233 it is passed by being copied directly into struct_value_rtx. */
2234 int structure_value_addr_parm = 0;
2235 /* Holds the value of implicit argument for the struct value. */
2236 tree structure_value_addr_value = NULL_TREE;
2237 /* Size of aggregate value wanted, or zero if none wanted
2238 or if we are using the non-reentrant PCC calling convention
2239 or expecting the value in registers. */
2240 HOST_WIDE_INT struct_value_size = 0;
2241 /* Nonzero if called function returns an aggregate in memory PCC style,
2242 by returning the address of where to find it. */
2243 int pcc_struct_value = 0;
2244 rtx struct_value = 0;
2246 /* Number of actual parameters in this call, including struct value addr. */
2247 int num_actuals;
2248 /* Number of named args. Args after this are anonymous ones
2249 and they must all go on the stack. */
2250 int n_named_args;
2251 /* Number of complex actual arguments that need to be split. */
2252 int num_complex_actuals = 0;
2254 /* Vector of information about each argument.
2255 Arguments are numbered in the order they will be pushed,
2256 not the order they are written. */
2257 struct arg_data *args;
2259 /* Total size in bytes of all the stack-parms scanned so far. */
2260 struct args_size args_size;
2261 struct args_size adjusted_args_size;
2262 /* Size of arguments before any adjustments (such as rounding). */
2263 int unadjusted_args_size;
2264 /* Data on reg parms scanned so far. */
2265 CUMULATIVE_ARGS args_so_far_v;
2266 cumulative_args_t args_so_far;
2267 /* Nonzero if a reg parm has been scanned. */
2268 int reg_parm_seen;
2269 /* Nonzero if this is an indirect function call. */
2271 /* Nonzero if we must avoid push-insns in the args for this call.
2272 If stack space is allocated for register parameters, but not by the
2273 caller, then it is preallocated in the fixed part of the stack frame.
2274 So the entire argument block must then be preallocated (i.e., we
2275 ignore PUSH_ROUNDING in that case). */
2277 int must_preallocate = !PUSH_ARGS;
2279 /* Size of the stack reserved for parameter registers. */
2280 int reg_parm_stack_space = 0;
2282 /* Address of space preallocated for stack parms
2283 (on machines that lack push insns), or 0 if space not preallocated. */
2284 rtx argblock = 0;
2286 /* Mask of ECF_ and ERF_ flags. */
2287 int flags = 0;
2288 int return_flags = 0;
2289 #ifdef REG_PARM_STACK_SPACE
2290 /* Define the boundary of the register parm stack space that needs to be
2291 saved, if any. */
2292 int low_to_save, high_to_save;
2293 rtx save_area = 0; /* Place that it is saved */
2294 #endif
2296 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
2297 char *initial_stack_usage_map = stack_usage_map;
2298 char *stack_usage_map_buf = NULL;
2300 int old_stack_allocated;
2302 /* State variables to track stack modifications. */
2303 rtx old_stack_level = 0;
2304 int old_stack_arg_under_construction = 0;
2305 int old_pending_adj = 0;
2306 int old_inhibit_defer_pop = inhibit_defer_pop;
2308 /* Some stack pointer alterations we make are performed via
2309 allocate_dynamic_stack_space. This modifies the stack_pointer_delta,
2310 which we then also need to save/restore along the way. */
2311 int old_stack_pointer_delta = 0;
2313 rtx call_fusage;
2314 tree addr = CALL_EXPR_FN (exp);
2315 int i;
2316 /* The alignment of the stack, in bits. */
2317 unsigned HOST_WIDE_INT preferred_stack_boundary;
2318 /* The alignment of the stack, in bytes. */
2319 unsigned HOST_WIDE_INT preferred_unit_stack_boundary;
2320 /* The static chain value to use for this call. */
2321 rtx static_chain_value;
2322 /* See if this is "nothrow" function call. */
2323 if (TREE_NOTHROW (exp))
2324 flags |= ECF_NOTHROW;
2326 /* See if we can find a DECL-node for the actual function, and get the
2327 function attributes (flags) from the function decl or type node. */
2328 fndecl = get_callee_fndecl (exp);
2329 if (fndecl)
2331 fntype = TREE_TYPE (fndecl);
2332 flags |= flags_from_decl_or_type (fndecl);
2333 return_flags |= decl_return_flags (fndecl);
2335 else
2337 fntype = TREE_TYPE (TREE_TYPE (addr));
2338 flags |= flags_from_decl_or_type (fntype);
2340 rettype = TREE_TYPE (exp);
2342 struct_value = targetm.calls.struct_value_rtx (fntype, 0);
2344 /* Warn if this value is an aggregate type,
2345 regardless of which calling convention we are using for it. */
2346 if (AGGREGATE_TYPE_P (rettype))
2347 warning (OPT_Waggregate_return, "function call has aggregate value");
2349 /* If the result of a non looping pure or const function call is
2350 ignored (or void), and none of its arguments are volatile, we can
2351 avoid expanding the call and just evaluate the arguments for
2352 side-effects. */
2353 if ((flags & (ECF_CONST | ECF_PURE))
2354 && (!(flags & ECF_LOOPING_CONST_OR_PURE))
2355 && (ignore || target == const0_rtx
2356 || TYPE_MODE (rettype) == VOIDmode))
2358 bool volatilep = false;
2359 tree arg;
2360 call_expr_arg_iterator iter;
2362 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
2363 if (TREE_THIS_VOLATILE (arg))
2365 volatilep = true;
2366 break;
2369 if (! volatilep)
2371 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
2372 expand_expr (arg, const0_rtx, VOIDmode, EXPAND_NORMAL);
2373 return const0_rtx;
2377 #ifdef REG_PARM_STACK_SPACE
2378 reg_parm_stack_space = REG_PARM_STACK_SPACE (!fndecl ? fntype : fndecl);
2379 #endif
2381 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
2382 && reg_parm_stack_space > 0 && PUSH_ARGS)
2383 must_preallocate = 1;
2385 /* Set up a place to return a structure. */
2387 /* Cater to broken compilers. */
2388 if (aggregate_value_p (exp, fntype))
2390 /* This call returns a big structure. */
2391 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
2393 #ifdef PCC_STATIC_STRUCT_RETURN
2395 pcc_struct_value = 1;
2397 #else /* not PCC_STATIC_STRUCT_RETURN */
2399 struct_value_size = int_size_in_bytes (rettype);
2401 if (target && MEM_P (target) && CALL_EXPR_RETURN_SLOT_OPT (exp))
2402 structure_value_addr = XEXP (target, 0);
2403 else
2405 /* For variable-sized objects, we must be called with a target
2406 specified. If we were to allocate space on the stack here,
2407 we would have no way of knowing when to free it. */
2408 rtx d = assign_temp (rettype, 1, 1);
2409 structure_value_addr = XEXP (d, 0);
2410 target = 0;
2413 #endif /* not PCC_STATIC_STRUCT_RETURN */
2416 /* Figure out the amount to which the stack should be aligned. */
2417 preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
2418 if (fndecl)
2420 struct cgraph_rtl_info *i = cgraph_rtl_info (fndecl);
2421 /* Without automatic stack alignment, we can't increase preferred
2422 stack boundary. With automatic stack alignment, it is
2423 unnecessary since unless we can guarantee that all callers will
2424 align the outgoing stack properly, callee has to align its
2425 stack anyway. */
2426 if (i
2427 && i->preferred_incoming_stack_boundary
2428 && i->preferred_incoming_stack_boundary < preferred_stack_boundary)
2429 preferred_stack_boundary = i->preferred_incoming_stack_boundary;
2432 /* Operand 0 is a pointer-to-function; get the type of the function. */
2433 funtype = TREE_TYPE (addr);
2434 gcc_assert (POINTER_TYPE_P (funtype));
2435 funtype = TREE_TYPE (funtype);
2437 /* Count whether there are actual complex arguments that need to be split
2438 into their real and imaginary parts. Munge the type_arg_types
2439 appropriately here as well. */
2440 if (targetm.calls.split_complex_arg)
2442 call_expr_arg_iterator iter;
2443 tree arg;
2444 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
2446 tree type = TREE_TYPE (arg);
2447 if (type && TREE_CODE (type) == COMPLEX_TYPE
2448 && targetm.calls.split_complex_arg (type))
2449 num_complex_actuals++;
2451 type_arg_types = split_complex_types (TYPE_ARG_TYPES (funtype));
2453 else
2454 type_arg_types = TYPE_ARG_TYPES (funtype);
2456 if (flags & ECF_MAY_BE_ALLOCA)
2457 cfun->calls_alloca = 1;
2459 /* If struct_value_rtx is 0, it means pass the address
2460 as if it were an extra parameter. Put the argument expression
2461 in structure_value_addr_value. */
2462 if (structure_value_addr && struct_value == 0)
2464 /* If structure_value_addr is a REG other than
2465 virtual_outgoing_args_rtx, we can use always use it. If it
2466 is not a REG, we must always copy it into a register.
2467 If it is virtual_outgoing_args_rtx, we must copy it to another
2468 register in some cases. */
2469 rtx temp = (!REG_P (structure_value_addr)
2470 || (ACCUMULATE_OUTGOING_ARGS
2471 && stack_arg_under_construction
2472 && structure_value_addr == virtual_outgoing_args_rtx)
2473 ? copy_addr_to_reg (convert_memory_address
2474 (Pmode, structure_value_addr))
2475 : structure_value_addr);
2477 structure_value_addr_value =
2478 make_tree (build_pointer_type (TREE_TYPE (funtype)), temp);
2479 structure_value_addr_parm = 1;
2482 /* Count the arguments and set NUM_ACTUALS. */
2483 num_actuals =
2484 call_expr_nargs (exp) + num_complex_actuals + structure_value_addr_parm;
2486 /* Compute number of named args.
2487 First, do a raw count of the args for INIT_CUMULATIVE_ARGS. */
2489 if (type_arg_types != 0)
2490 n_named_args
2491 = (list_length (type_arg_types)
2492 /* Count the struct value address, if it is passed as a parm. */
2493 + structure_value_addr_parm);
2494 else
2495 /* If we know nothing, treat all args as named. */
2496 n_named_args = num_actuals;
2498 /* Start updating where the next arg would go.
2500 On some machines (such as the PA) indirect calls have a different
2501 calling convention than normal calls. The fourth argument in
2502 INIT_CUMULATIVE_ARGS tells the backend if this is an indirect call
2503 or not. */
2504 INIT_CUMULATIVE_ARGS (args_so_far_v, funtype, NULL_RTX, fndecl, n_named_args);
2505 args_so_far = pack_cumulative_args (&args_so_far_v);
2507 /* Now possibly adjust the number of named args.
2508 Normally, don't include the last named arg if anonymous args follow.
2509 We do include the last named arg if
2510 targetm.calls.strict_argument_naming() returns nonzero.
2511 (If no anonymous args follow, the result of list_length is actually
2512 one too large. This is harmless.)
2514 If targetm.calls.pretend_outgoing_varargs_named() returns
2515 nonzero, and targetm.calls.strict_argument_naming() returns zero,
2516 this machine will be able to place unnamed args that were passed
2517 in registers into the stack. So treat all args as named. This
2518 allows the insns emitting for a specific argument list to be
2519 independent of the function declaration.
2521 If targetm.calls.pretend_outgoing_varargs_named() returns zero,
2522 we do not have any reliable way to pass unnamed args in
2523 registers, so we must force them into memory. */
2525 if (type_arg_types != 0
2526 && targetm.calls.strict_argument_naming (args_so_far))
2528 else if (type_arg_types != 0
2529 && ! targetm.calls.pretend_outgoing_varargs_named (args_so_far))
2530 /* Don't include the last named arg. */
2531 --n_named_args;
2532 else
2533 /* Treat all args as named. */
2534 n_named_args = num_actuals;
2536 /* Make a vector to hold all the information about each arg. */
2537 args = XALLOCAVEC (struct arg_data, num_actuals);
2538 memset (args, 0, num_actuals * sizeof (struct arg_data));
2540 /* Build up entries in the ARGS array, compute the size of the
2541 arguments into ARGS_SIZE, etc. */
2542 initialize_argument_information (num_actuals, args, &args_size,
2543 n_named_args, exp,
2544 structure_value_addr_value, fndecl, fntype,
2545 args_so_far, reg_parm_stack_space,
2546 &old_stack_level, &old_pending_adj,
2547 &must_preallocate, &flags,
2548 &try_tail_call, CALL_FROM_THUNK_P (exp));
2550 if (args_size.var)
2551 must_preallocate = 1;
2553 /* Now make final decision about preallocating stack space. */
2554 must_preallocate = finalize_must_preallocate (must_preallocate,
2555 num_actuals, args,
2556 &args_size);
2558 /* If the structure value address will reference the stack pointer, we
2559 must stabilize it. We don't need to do this if we know that we are
2560 not going to adjust the stack pointer in processing this call. */
2562 if (structure_value_addr
2563 && (reg_mentioned_p (virtual_stack_dynamic_rtx, structure_value_addr)
2564 || reg_mentioned_p (virtual_outgoing_args_rtx,
2565 structure_value_addr))
2566 && (args_size.var
2567 || (!ACCUMULATE_OUTGOING_ARGS && args_size.constant)))
2568 structure_value_addr = copy_to_reg (structure_value_addr);
2570 /* Tail calls can make things harder to debug, and we've traditionally
2571 pushed these optimizations into -O2. Don't try if we're already
2572 expanding a call, as that means we're an argument. Don't try if
2573 there's cleanups, as we know there's code to follow the call. */
2575 if (currently_expanding_call++ != 0
2576 || !flag_optimize_sibling_calls
2577 || args_size.var
2578 || dbg_cnt (tail_call) == false)
2579 try_tail_call = 0;
2581 /* Rest of purposes for tail call optimizations to fail. */
2582 if (
2583 #ifdef HAVE_sibcall_epilogue
2584 !HAVE_sibcall_epilogue
2585 #else
2587 #endif
2588 || !try_tail_call
2589 /* Doing sibling call optimization needs some work, since
2590 structure_value_addr can be allocated on the stack.
2591 It does not seem worth the effort since few optimizable
2592 sibling calls will return a structure. */
2593 || structure_value_addr != NULL_RTX
2594 #ifdef REG_PARM_STACK_SPACE
2595 /* If outgoing reg parm stack space changes, we can not do sibcall. */
2596 || (OUTGOING_REG_PARM_STACK_SPACE (funtype)
2597 != OUTGOING_REG_PARM_STACK_SPACE (TREE_TYPE (current_function_decl)))
2598 || (reg_parm_stack_space != REG_PARM_STACK_SPACE (current_function_decl))
2599 #endif
2600 /* Check whether the target is able to optimize the call
2601 into a sibcall. */
2602 || !targetm.function_ok_for_sibcall (fndecl, exp)
2603 /* Functions that do not return exactly once may not be sibcall
2604 optimized. */
2605 || (flags & (ECF_RETURNS_TWICE | ECF_NORETURN))
2606 || TYPE_VOLATILE (TREE_TYPE (TREE_TYPE (addr)))
2607 /* If the called function is nested in the current one, it might access
2608 some of the caller's arguments, but could clobber them beforehand if
2609 the argument areas are shared. */
2610 || (fndecl && decl_function_context (fndecl) == current_function_decl)
2611 /* If this function requires more stack slots than the current
2612 function, we cannot change it into a sibling call.
2613 crtl->args.pretend_args_size is not part of the
2614 stack allocated by our caller. */
2615 || args_size.constant > (crtl->args.size
2616 - crtl->args.pretend_args_size)
2617 /* If the callee pops its own arguments, then it must pop exactly
2618 the same number of arguments as the current function. */
2619 || (targetm.calls.return_pops_args (fndecl, funtype, args_size.constant)
2620 != targetm.calls.return_pops_args (current_function_decl,
2621 TREE_TYPE (current_function_decl),
2622 crtl->args.size))
2623 || !lang_hooks.decls.ok_for_sibcall (fndecl))
2624 try_tail_call = 0;
2626 /* Check if caller and callee disagree in promotion of function
2627 return value. */
2628 if (try_tail_call)
2630 enum machine_mode caller_mode, caller_promoted_mode;
2631 enum machine_mode callee_mode, callee_promoted_mode;
2632 int caller_unsignedp, callee_unsignedp;
2633 tree caller_res = DECL_RESULT (current_function_decl);
2635 caller_unsignedp = TYPE_UNSIGNED (TREE_TYPE (caller_res));
2636 caller_mode = DECL_MODE (caller_res);
2637 callee_unsignedp = TYPE_UNSIGNED (TREE_TYPE (funtype));
2638 callee_mode = TYPE_MODE (TREE_TYPE (funtype));
2639 caller_promoted_mode
2640 = promote_function_mode (TREE_TYPE (caller_res), caller_mode,
2641 &caller_unsignedp,
2642 TREE_TYPE (current_function_decl), 1);
2643 callee_promoted_mode
2644 = promote_function_mode (TREE_TYPE (funtype), callee_mode,
2645 &callee_unsignedp,
2646 funtype, 1);
2647 if (caller_mode != VOIDmode
2648 && (caller_promoted_mode != callee_promoted_mode
2649 || ((caller_mode != caller_promoted_mode
2650 || callee_mode != callee_promoted_mode)
2651 && (caller_unsignedp != callee_unsignedp
2652 || GET_MODE_BITSIZE (caller_mode)
2653 < GET_MODE_BITSIZE (callee_mode)))))
2654 try_tail_call = 0;
2657 /* Ensure current function's preferred stack boundary is at least
2658 what we need. Stack alignment may also increase preferred stack
2659 boundary. */
2660 if (crtl->preferred_stack_boundary < preferred_stack_boundary)
2661 crtl->preferred_stack_boundary = preferred_stack_boundary;
2662 else
2663 preferred_stack_boundary = crtl->preferred_stack_boundary;
2665 preferred_unit_stack_boundary = preferred_stack_boundary / BITS_PER_UNIT;
2667 /* We want to make two insn chains; one for a sibling call, the other
2668 for a normal call. We will select one of the two chains after
2669 initial RTL generation is complete. */
2670 for (pass = try_tail_call ? 0 : 1; pass < 2; pass++)
2672 int sibcall_failure = 0;
2673 /* We want to emit any pending stack adjustments before the tail
2674 recursion "call". That way we know any adjustment after the tail
2675 recursion call can be ignored if we indeed use the tail
2676 call expansion. */
2677 saved_pending_stack_adjust save;
2678 rtx insns;
2679 rtx before_call, next_arg_reg, after_args;
2681 if (pass == 0)
2683 /* State variables we need to save and restore between
2684 iterations. */
2685 save_pending_stack_adjust (&save);
2687 if (pass)
2688 flags &= ~ECF_SIBCALL;
2689 else
2690 flags |= ECF_SIBCALL;
2692 /* Other state variables that we must reinitialize each time
2693 through the loop (that are not initialized by the loop itself). */
2694 argblock = 0;
2695 call_fusage = 0;
2697 /* Start a new sequence for the normal call case.
2699 From this point on, if the sibling call fails, we want to set
2700 sibcall_failure instead of continuing the loop. */
2701 start_sequence ();
2703 /* Don't let pending stack adjusts add up to too much.
2704 Also, do all pending adjustments now if there is any chance
2705 this might be a call to alloca or if we are expanding a sibling
2706 call sequence.
2707 Also do the adjustments before a throwing call, otherwise
2708 exception handling can fail; PR 19225. */
2709 if (pending_stack_adjust >= 32
2710 || (pending_stack_adjust > 0
2711 && (flags & ECF_MAY_BE_ALLOCA))
2712 || (pending_stack_adjust > 0
2713 && flag_exceptions && !(flags & ECF_NOTHROW))
2714 || pass == 0)
2715 do_pending_stack_adjust ();
2717 /* Precompute any arguments as needed. */
2718 if (pass)
2719 precompute_arguments (num_actuals, args);
2721 /* Now we are about to start emitting insns that can be deleted
2722 if a libcall is deleted. */
2723 if (pass && (flags & ECF_MALLOC))
2724 start_sequence ();
2726 if (pass == 0 && crtl->stack_protect_guard)
2727 stack_protect_epilogue ();
2729 adjusted_args_size = args_size;
2730 /* Compute the actual size of the argument block required. The variable
2731 and constant sizes must be combined, the size may have to be rounded,
2732 and there may be a minimum required size. When generating a sibcall
2733 pattern, do not round up, since we'll be re-using whatever space our
2734 caller provided. */
2735 unadjusted_args_size
2736 = compute_argument_block_size (reg_parm_stack_space,
2737 &adjusted_args_size,
2738 fndecl, fntype,
2739 (pass == 0 ? 0
2740 : preferred_stack_boundary));
2742 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
2744 /* The argument block when performing a sibling call is the
2745 incoming argument block. */
2746 if (pass == 0)
2748 argblock = crtl->args.internal_arg_pointer;
2749 argblock
2750 #ifdef STACK_GROWS_DOWNWARD
2751 = plus_constant (Pmode, argblock, crtl->args.pretend_args_size);
2752 #else
2753 = plus_constant (Pmode, argblock, -crtl->args.pretend_args_size);
2754 #endif
2755 stored_args_map = sbitmap_alloc (args_size.constant);
2756 bitmap_clear (stored_args_map);
2759 /* If we have no actual push instructions, or shouldn't use them,
2760 make space for all args right now. */
2761 else if (adjusted_args_size.var != 0)
2763 if (old_stack_level == 0)
2765 emit_stack_save (SAVE_BLOCK, &old_stack_level);
2766 old_stack_pointer_delta = stack_pointer_delta;
2767 old_pending_adj = pending_stack_adjust;
2768 pending_stack_adjust = 0;
2769 /* stack_arg_under_construction says whether a stack arg is
2770 being constructed at the old stack level. Pushing the stack
2771 gets a clean outgoing argument block. */
2772 old_stack_arg_under_construction = stack_arg_under_construction;
2773 stack_arg_under_construction = 0;
2775 argblock = push_block (ARGS_SIZE_RTX (adjusted_args_size), 0, 0);
2776 if (flag_stack_usage_info)
2777 current_function_has_unbounded_dynamic_stack_size = 1;
2779 else
2781 /* Note that we must go through the motions of allocating an argument
2782 block even if the size is zero because we may be storing args
2783 in the area reserved for register arguments, which may be part of
2784 the stack frame. */
2786 int needed = adjusted_args_size.constant;
2788 /* Store the maximum argument space used. It will be pushed by
2789 the prologue (if ACCUMULATE_OUTGOING_ARGS, or stack overflow
2790 checking). */
2792 if (needed > crtl->outgoing_args_size)
2793 crtl->outgoing_args_size = needed;
2795 if (must_preallocate)
2797 if (ACCUMULATE_OUTGOING_ARGS)
2799 /* Since the stack pointer will never be pushed, it is
2800 possible for the evaluation of a parm to clobber
2801 something we have already written to the stack.
2802 Since most function calls on RISC machines do not use
2803 the stack, this is uncommon, but must work correctly.
2805 Therefore, we save any area of the stack that was already
2806 written and that we are using. Here we set up to do this
2807 by making a new stack usage map from the old one. The
2808 actual save will be done by store_one_arg.
2810 Another approach might be to try to reorder the argument
2811 evaluations to avoid this conflicting stack usage. */
2813 /* Since we will be writing into the entire argument area,
2814 the map must be allocated for its entire size, not just
2815 the part that is the responsibility of the caller. */
2816 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
2817 needed += reg_parm_stack_space;
2819 #ifdef ARGS_GROW_DOWNWARD
2820 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
2821 needed + 1);
2822 #else
2823 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
2824 needed);
2825 #endif
2826 free (stack_usage_map_buf);
2827 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
2828 stack_usage_map = stack_usage_map_buf;
2830 if (initial_highest_arg_in_use)
2831 memcpy (stack_usage_map, initial_stack_usage_map,
2832 initial_highest_arg_in_use);
2834 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
2835 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
2836 (highest_outgoing_arg_in_use
2837 - initial_highest_arg_in_use));
2838 needed = 0;
2840 /* The address of the outgoing argument list must not be
2841 copied to a register here, because argblock would be left
2842 pointing to the wrong place after the call to
2843 allocate_dynamic_stack_space below. */
2845 argblock = virtual_outgoing_args_rtx;
2847 else
2849 if (inhibit_defer_pop == 0)
2851 /* Try to reuse some or all of the pending_stack_adjust
2852 to get this space. */
2853 needed
2854 = (combine_pending_stack_adjustment_and_call
2855 (unadjusted_args_size,
2856 &adjusted_args_size,
2857 preferred_unit_stack_boundary));
2859 /* combine_pending_stack_adjustment_and_call computes
2860 an adjustment before the arguments are allocated.
2861 Account for them and see whether or not the stack
2862 needs to go up or down. */
2863 needed = unadjusted_args_size - needed;
2865 if (needed < 0)
2867 /* We're releasing stack space. */
2868 /* ??? We can avoid any adjustment at all if we're
2869 already aligned. FIXME. */
2870 pending_stack_adjust = -needed;
2871 do_pending_stack_adjust ();
2872 needed = 0;
2874 else
2875 /* We need to allocate space. We'll do that in
2876 push_block below. */
2877 pending_stack_adjust = 0;
2880 /* Special case this because overhead of `push_block' in
2881 this case is non-trivial. */
2882 if (needed == 0)
2883 argblock = virtual_outgoing_args_rtx;
2884 else
2886 argblock = push_block (GEN_INT (needed), 0, 0);
2887 #ifdef ARGS_GROW_DOWNWARD
2888 argblock = plus_constant (Pmode, argblock, needed);
2889 #endif
2892 /* We only really need to call `copy_to_reg' in the case
2893 where push insns are going to be used to pass ARGBLOCK
2894 to a function call in ARGS. In that case, the stack
2895 pointer changes value from the allocation point to the
2896 call point, and hence the value of
2897 VIRTUAL_OUTGOING_ARGS_RTX changes as well. But might
2898 as well always do it. */
2899 argblock = copy_to_reg (argblock);
2904 if (ACCUMULATE_OUTGOING_ARGS)
2906 /* The save/restore code in store_one_arg handles all
2907 cases except one: a constructor call (including a C
2908 function returning a BLKmode struct) to initialize
2909 an argument. */
2910 if (stack_arg_under_construction)
2912 rtx push_size
2913 = GEN_INT (adjusted_args_size.constant
2914 + (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype
2915 : TREE_TYPE (fndecl))) ? 0
2916 : reg_parm_stack_space));
2917 if (old_stack_level == 0)
2919 emit_stack_save (SAVE_BLOCK, &old_stack_level);
2920 old_stack_pointer_delta = stack_pointer_delta;
2921 old_pending_adj = pending_stack_adjust;
2922 pending_stack_adjust = 0;
2923 /* stack_arg_under_construction says whether a stack
2924 arg is being constructed at the old stack level.
2925 Pushing the stack gets a clean outgoing argument
2926 block. */
2927 old_stack_arg_under_construction
2928 = stack_arg_under_construction;
2929 stack_arg_under_construction = 0;
2930 /* Make a new map for the new argument list. */
2931 free (stack_usage_map_buf);
2932 stack_usage_map_buf = XCNEWVEC (char, highest_outgoing_arg_in_use);
2933 stack_usage_map = stack_usage_map_buf;
2934 highest_outgoing_arg_in_use = 0;
2936 /* We can pass TRUE as the 4th argument because we just
2937 saved the stack pointer and will restore it right after
2938 the call. */
2939 allocate_dynamic_stack_space (push_size, 0,
2940 BIGGEST_ALIGNMENT, true);
2943 /* If argument evaluation might modify the stack pointer,
2944 copy the address of the argument list to a register. */
2945 for (i = 0; i < num_actuals; i++)
2946 if (args[i].pass_on_stack)
2948 argblock = copy_addr_to_reg (argblock);
2949 break;
2953 compute_argument_addresses (args, argblock, num_actuals);
2955 /* If we push args individually in reverse order, perform stack alignment
2956 before the first push (the last arg). */
2957 if (PUSH_ARGS_REVERSED && argblock == 0
2958 && adjusted_args_size.constant > reg_parm_stack_space
2959 && adjusted_args_size.constant != unadjusted_args_size)
2961 /* When the stack adjustment is pending, we get better code
2962 by combining the adjustments. */
2963 if (pending_stack_adjust
2964 && ! inhibit_defer_pop)
2966 pending_stack_adjust
2967 = (combine_pending_stack_adjustment_and_call
2968 (unadjusted_args_size,
2969 &adjusted_args_size,
2970 preferred_unit_stack_boundary));
2971 do_pending_stack_adjust ();
2973 else if (argblock == 0)
2974 anti_adjust_stack (GEN_INT (adjusted_args_size.constant
2975 - unadjusted_args_size));
2977 /* Now that the stack is properly aligned, pops can't safely
2978 be deferred during the evaluation of the arguments. */
2979 NO_DEFER_POP;
2981 /* Record the maximum pushed stack space size. We need to delay
2982 doing it this far to take into account the optimization done
2983 by combine_pending_stack_adjustment_and_call. */
2984 if (flag_stack_usage_info
2985 && !ACCUMULATE_OUTGOING_ARGS
2986 && pass
2987 && adjusted_args_size.var == 0)
2989 int pushed = adjusted_args_size.constant + pending_stack_adjust;
2990 if (pushed > current_function_pushed_stack_size)
2991 current_function_pushed_stack_size = pushed;
2994 funexp = rtx_for_function_call (fndecl, addr);
2996 /* Figure out the register where the value, if any, will come back. */
2997 valreg = 0;
2998 if (TYPE_MODE (rettype) != VOIDmode
2999 && ! structure_value_addr)
3001 if (pcc_struct_value)
3002 valreg = hard_function_value (build_pointer_type (rettype),
3003 fndecl, NULL, (pass == 0));
3004 else
3005 valreg = hard_function_value (rettype, fndecl, fntype,
3006 (pass == 0));
3008 /* If VALREG is a PARALLEL whose first member has a zero
3009 offset, use that. This is for targets such as m68k that
3010 return the same value in multiple places. */
3011 if (GET_CODE (valreg) == PARALLEL)
3013 rtx elem = XVECEXP (valreg, 0, 0);
3014 rtx where = XEXP (elem, 0);
3015 rtx offset = XEXP (elem, 1);
3016 if (offset == const0_rtx
3017 && GET_MODE (where) == GET_MODE (valreg))
3018 valreg = where;
3022 /* Precompute all register parameters. It isn't safe to compute anything
3023 once we have started filling any specific hard regs. */
3024 precompute_register_parameters (num_actuals, args, &reg_parm_seen);
3026 if (CALL_EXPR_STATIC_CHAIN (exp))
3027 static_chain_value = expand_normal (CALL_EXPR_STATIC_CHAIN (exp));
3028 else
3029 static_chain_value = 0;
3031 #ifdef REG_PARM_STACK_SPACE
3032 /* Save the fixed argument area if it's part of the caller's frame and
3033 is clobbered by argument setup for this call. */
3034 if (ACCUMULATE_OUTGOING_ARGS && pass)
3035 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
3036 &low_to_save, &high_to_save);
3037 #endif
3039 /* Now store (and compute if necessary) all non-register parms.
3040 These come before register parms, since they can require block-moves,
3041 which could clobber the registers used for register parms.
3042 Parms which have partial registers are not stored here,
3043 but we do preallocate space here if they want that. */
3045 for (i = 0; i < num_actuals; i++)
3047 if (args[i].reg == 0 || args[i].pass_on_stack)
3049 rtx before_arg = get_last_insn ();
3051 /* We don't allow passing huge (> 2^30 B) arguments
3052 by value. It would cause an overflow later on. */
3053 if (adjusted_args_size.constant
3054 >= (1 << (HOST_BITS_PER_INT - 2)))
3056 sorry ("passing too large argument on stack");
3057 continue;
3060 if (store_one_arg (&args[i], argblock, flags,
3061 adjusted_args_size.var != 0,
3062 reg_parm_stack_space)
3063 || (pass == 0
3064 && check_sibcall_argument_overlap (before_arg,
3065 &args[i], 1)))
3066 sibcall_failure = 1;
3069 if (args[i].stack)
3070 call_fusage
3071 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[i].tree_value)),
3072 gen_rtx_USE (VOIDmode, args[i].stack),
3073 call_fusage);
3076 /* If we have a parm that is passed in registers but not in memory
3077 and whose alignment does not permit a direct copy into registers,
3078 make a group of pseudos that correspond to each register that we
3079 will later fill. */
3080 if (STRICT_ALIGNMENT)
3081 store_unaligned_arguments_into_pseudos (args, num_actuals);
3083 /* Now store any partially-in-registers parm.
3084 This is the last place a block-move can happen. */
3085 if (reg_parm_seen)
3086 for (i = 0; i < num_actuals; i++)
3087 if (args[i].partial != 0 && ! args[i].pass_on_stack)
3089 rtx before_arg = get_last_insn ();
3091 if (store_one_arg (&args[i], argblock, flags,
3092 adjusted_args_size.var != 0,
3093 reg_parm_stack_space)
3094 || (pass == 0
3095 && check_sibcall_argument_overlap (before_arg,
3096 &args[i], 1)))
3097 sibcall_failure = 1;
3100 /* If we pushed args in forward order, perform stack alignment
3101 after pushing the last arg. */
3102 if (!PUSH_ARGS_REVERSED && argblock == 0)
3103 anti_adjust_stack (GEN_INT (adjusted_args_size.constant
3104 - unadjusted_args_size));
3106 /* If register arguments require space on the stack and stack space
3107 was not preallocated, allocate stack space here for arguments
3108 passed in registers. */
3109 if (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
3110 && !ACCUMULATE_OUTGOING_ARGS
3111 && must_preallocate == 0 && reg_parm_stack_space > 0)
3112 anti_adjust_stack (GEN_INT (reg_parm_stack_space));
3114 /* Pass the function the address in which to return a
3115 structure value. */
3116 if (pass != 0 && structure_value_addr && ! structure_value_addr_parm)
3118 structure_value_addr
3119 = convert_memory_address (Pmode, structure_value_addr);
3120 emit_move_insn (struct_value,
3121 force_reg (Pmode,
3122 force_operand (structure_value_addr,
3123 NULL_RTX)));
3125 if (REG_P (struct_value))
3126 use_reg (&call_fusage, struct_value);
3129 after_args = get_last_insn ();
3130 funexp = prepare_call_address (fndecl, funexp, static_chain_value,
3131 &call_fusage, reg_parm_seen, pass == 0);
3133 load_register_parameters (args, num_actuals, &call_fusage, flags,
3134 pass == 0, &sibcall_failure);
3136 /* Save a pointer to the last insn before the call, so that we can
3137 later safely search backwards to find the CALL_INSN. */
3138 before_call = get_last_insn ();
3140 /* Set up next argument register. For sibling calls on machines
3141 with register windows this should be the incoming register. */
3142 if (pass == 0)
3143 next_arg_reg = targetm.calls.function_incoming_arg (args_so_far,
3144 VOIDmode,
3145 void_type_node,
3146 true);
3147 else
3148 next_arg_reg = targetm.calls.function_arg (args_so_far,
3149 VOIDmode, void_type_node,
3150 true);
3152 if (pass == 1 && (return_flags & ERF_RETURNS_ARG))
3154 int arg_nr = return_flags & ERF_RETURN_ARG_MASK;
3155 if (PUSH_ARGS_REVERSED)
3156 arg_nr = num_actuals - arg_nr - 1;
3157 if (arg_nr >= 0
3158 && arg_nr < num_actuals
3159 && args[arg_nr].reg
3160 && valreg
3161 && REG_P (valreg)
3162 && GET_MODE (args[arg_nr].reg) == GET_MODE (valreg))
3163 call_fusage
3164 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[arg_nr].tree_value)),
3165 gen_rtx_SET (VOIDmode, valreg, args[arg_nr].reg),
3166 call_fusage);
3168 /* All arguments and registers used for the call must be set up by
3169 now! */
3171 /* Stack must be properly aligned now. */
3172 gcc_assert (!pass
3173 || !(stack_pointer_delta % preferred_unit_stack_boundary));
3175 /* Generate the actual call instruction. */
3176 emit_call_1 (funexp, exp, fndecl, funtype, unadjusted_args_size,
3177 adjusted_args_size.constant, struct_value_size,
3178 next_arg_reg, valreg, old_inhibit_defer_pop, call_fusage,
3179 flags, args_so_far);
3181 /* If the call setup or the call itself overlaps with anything
3182 of the argument setup we probably clobbered our call address.
3183 In that case we can't do sibcalls. */
3184 if (pass == 0
3185 && check_sibcall_argument_overlap (after_args, 0, 0))
3186 sibcall_failure = 1;
3188 /* If a non-BLKmode value is returned at the most significant end
3189 of a register, shift the register right by the appropriate amount
3190 and update VALREG accordingly. BLKmode values are handled by the
3191 group load/store machinery below. */
3192 if (!structure_value_addr
3193 && !pcc_struct_value
3194 && TYPE_MODE (rettype) != VOIDmode
3195 && TYPE_MODE (rettype) != BLKmode
3196 && REG_P (valreg)
3197 && targetm.calls.return_in_msb (rettype))
3199 if (shift_return_value (TYPE_MODE (rettype), false, valreg))
3200 sibcall_failure = 1;
3201 valreg = gen_rtx_REG (TYPE_MODE (rettype), REGNO (valreg));
3204 if (pass && (flags & ECF_MALLOC))
3206 rtx temp = gen_reg_rtx (GET_MODE (valreg));
3207 rtx last, insns;
3209 /* The return value from a malloc-like function is a pointer. */
3210 if (TREE_CODE (rettype) == POINTER_TYPE)
3211 mark_reg_pointer (temp, MALLOC_ABI_ALIGNMENT);
3213 emit_move_insn (temp, valreg);
3215 /* The return value from a malloc-like function can not alias
3216 anything else. */
3217 last = get_last_insn ();
3218 add_reg_note (last, REG_NOALIAS, temp);
3220 /* Write out the sequence. */
3221 insns = get_insns ();
3222 end_sequence ();
3223 emit_insn (insns);
3224 valreg = temp;
3227 /* For calls to `setjmp', etc., inform
3228 function.c:setjmp_warnings that it should complain if
3229 nonvolatile values are live. For functions that cannot
3230 return, inform flow that control does not fall through. */
3232 if ((flags & ECF_NORETURN) || pass == 0)
3234 /* The barrier must be emitted
3235 immediately after the CALL_INSN. Some ports emit more
3236 than just a CALL_INSN above, so we must search for it here. */
3238 rtx last = get_last_insn ();
3239 while (!CALL_P (last))
3241 last = PREV_INSN (last);
3242 /* There was no CALL_INSN? */
3243 gcc_assert (last != before_call);
3246 emit_barrier_after (last);
3248 /* Stack adjustments after a noreturn call are dead code.
3249 However when NO_DEFER_POP is in effect, we must preserve
3250 stack_pointer_delta. */
3251 if (inhibit_defer_pop == 0)
3253 stack_pointer_delta = old_stack_allocated;
3254 pending_stack_adjust = 0;
3258 /* If value type not void, return an rtx for the value. */
3260 if (TYPE_MODE (rettype) == VOIDmode
3261 || ignore)
3262 target = const0_rtx;
3263 else if (structure_value_addr)
3265 if (target == 0 || !MEM_P (target))
3267 target
3268 = gen_rtx_MEM (TYPE_MODE (rettype),
3269 memory_address (TYPE_MODE (rettype),
3270 structure_value_addr));
3271 set_mem_attributes (target, rettype, 1);
3274 else if (pcc_struct_value)
3276 /* This is the special C++ case where we need to
3277 know what the true target was. We take care to
3278 never use this value more than once in one expression. */
3279 target = gen_rtx_MEM (TYPE_MODE (rettype),
3280 copy_to_reg (valreg));
3281 set_mem_attributes (target, rettype, 1);
3283 /* Handle calls that return values in multiple non-contiguous locations.
3284 The Irix 6 ABI has examples of this. */
3285 else if (GET_CODE (valreg) == PARALLEL)
3287 if (target == 0)
3288 target = emit_group_move_into_temps (valreg);
3289 else if (rtx_equal_p (target, valreg))
3291 else if (GET_CODE (target) == PARALLEL)
3292 /* Handle the result of a emit_group_move_into_temps
3293 call in the previous pass. */
3294 emit_group_move (target, valreg);
3295 else
3296 emit_group_store (target, valreg, rettype,
3297 int_size_in_bytes (rettype));
3299 else if (target
3300 && GET_MODE (target) == TYPE_MODE (rettype)
3301 && GET_MODE (target) == GET_MODE (valreg))
3303 bool may_overlap = false;
3305 /* We have to copy a return value in a CLASS_LIKELY_SPILLED hard
3306 reg to a plain register. */
3307 if (!REG_P (target) || HARD_REGISTER_P (target))
3308 valreg = avoid_likely_spilled_reg (valreg);
3310 /* If TARGET is a MEM in the argument area, and we have
3311 saved part of the argument area, then we can't store
3312 directly into TARGET as it may get overwritten when we
3313 restore the argument save area below. Don't work too
3314 hard though and simply force TARGET to a register if it
3315 is a MEM; the optimizer is quite likely to sort it out. */
3316 if (ACCUMULATE_OUTGOING_ARGS && pass && MEM_P (target))
3317 for (i = 0; i < num_actuals; i++)
3318 if (args[i].save_area)
3320 may_overlap = true;
3321 break;
3324 if (may_overlap)
3325 target = copy_to_reg (valreg);
3326 else
3328 /* TARGET and VALREG cannot be equal at this point
3329 because the latter would not have
3330 REG_FUNCTION_VALUE_P true, while the former would if
3331 it were referring to the same register.
3333 If they refer to the same register, this move will be
3334 a no-op, except when function inlining is being
3335 done. */
3336 emit_move_insn (target, valreg);
3338 /* If we are setting a MEM, this code must be executed.
3339 Since it is emitted after the call insn, sibcall
3340 optimization cannot be performed in that case. */
3341 if (MEM_P (target))
3342 sibcall_failure = 1;
3345 else
3346 target = copy_to_reg (avoid_likely_spilled_reg (valreg));
3348 /* If we promoted this return value, make the proper SUBREG.
3349 TARGET might be const0_rtx here, so be careful. */
3350 if (REG_P (target)
3351 && TYPE_MODE (rettype) != BLKmode
3352 && GET_MODE (target) != TYPE_MODE (rettype))
3354 tree type = rettype;
3355 int unsignedp = TYPE_UNSIGNED (type);
3356 int offset = 0;
3357 enum machine_mode pmode;
3359 /* Ensure we promote as expected, and get the new unsignedness. */
3360 pmode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
3361 funtype, 1);
3362 gcc_assert (GET_MODE (target) == pmode);
3364 if ((WORDS_BIG_ENDIAN || BYTES_BIG_ENDIAN)
3365 && (GET_MODE_SIZE (GET_MODE (target))
3366 > GET_MODE_SIZE (TYPE_MODE (type))))
3368 offset = GET_MODE_SIZE (GET_MODE (target))
3369 - GET_MODE_SIZE (TYPE_MODE (type));
3370 if (! BYTES_BIG_ENDIAN)
3371 offset = (offset / UNITS_PER_WORD) * UNITS_PER_WORD;
3372 else if (! WORDS_BIG_ENDIAN)
3373 offset %= UNITS_PER_WORD;
3376 target = gen_rtx_SUBREG (TYPE_MODE (type), target, offset);
3377 SUBREG_PROMOTED_VAR_P (target) = 1;
3378 SUBREG_PROMOTED_UNSIGNED_SET (target, unsignedp);
3381 /* If size of args is variable or this was a constructor call for a stack
3382 argument, restore saved stack-pointer value. */
3384 if (old_stack_level)
3386 rtx prev = get_last_insn ();
3388 emit_stack_restore (SAVE_BLOCK, old_stack_level);
3389 stack_pointer_delta = old_stack_pointer_delta;
3391 fixup_args_size_notes (prev, get_last_insn (), stack_pointer_delta);
3393 pending_stack_adjust = old_pending_adj;
3394 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
3395 stack_arg_under_construction = old_stack_arg_under_construction;
3396 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
3397 stack_usage_map = initial_stack_usage_map;
3398 sibcall_failure = 1;
3400 else if (ACCUMULATE_OUTGOING_ARGS && pass)
3402 #ifdef REG_PARM_STACK_SPACE
3403 if (save_area)
3404 restore_fixed_argument_area (save_area, argblock,
3405 high_to_save, low_to_save);
3406 #endif
3408 /* If we saved any argument areas, restore them. */
3409 for (i = 0; i < num_actuals; i++)
3410 if (args[i].save_area)
3412 enum machine_mode save_mode = GET_MODE (args[i].save_area);
3413 rtx stack_area
3414 = gen_rtx_MEM (save_mode,
3415 memory_address (save_mode,
3416 XEXP (args[i].stack_slot, 0)));
3418 if (save_mode != BLKmode)
3419 emit_move_insn (stack_area, args[i].save_area);
3420 else
3421 emit_block_move (stack_area, args[i].save_area,
3422 GEN_INT (args[i].locate.size.constant),
3423 BLOCK_OP_CALL_PARM);
3426 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
3427 stack_usage_map = initial_stack_usage_map;
3430 /* If this was alloca, record the new stack level for nonlocal gotos.
3431 Check for the handler slots since we might not have a save area
3432 for non-local gotos. */
3434 if ((flags & ECF_MAY_BE_ALLOCA) && cfun->nonlocal_goto_save_area != 0)
3435 update_nonlocal_goto_save_area ();
3437 /* Free up storage we no longer need. */
3438 for (i = 0; i < num_actuals; ++i)
3439 free (args[i].aligned_regs);
3441 insns = get_insns ();
3442 end_sequence ();
3444 if (pass == 0)
3446 tail_call_insns = insns;
3448 /* Restore the pending stack adjustment now that we have
3449 finished generating the sibling call sequence. */
3451 restore_pending_stack_adjust (&save);
3453 /* Prepare arg structure for next iteration. */
3454 for (i = 0; i < num_actuals; i++)
3456 args[i].value = 0;
3457 args[i].aligned_regs = 0;
3458 args[i].stack = 0;
3461 sbitmap_free (stored_args_map);
3462 internal_arg_pointer_exp_state.scan_start = NULL_RTX;
3463 internal_arg_pointer_exp_state.cache.release ();
3465 else
3467 normal_call_insns = insns;
3469 /* Verify that we've deallocated all the stack we used. */
3470 gcc_assert ((flags & ECF_NORETURN)
3471 || (old_stack_allocated
3472 == stack_pointer_delta - pending_stack_adjust));
3475 /* If something prevents making this a sibling call,
3476 zero out the sequence. */
3477 if (sibcall_failure)
3478 tail_call_insns = NULL_RTX;
3479 else
3480 break;
3483 /* If tail call production succeeded, we need to remove REG_EQUIV notes on
3484 arguments too, as argument area is now clobbered by the call. */
3485 if (tail_call_insns)
3487 emit_insn (tail_call_insns);
3488 crtl->tail_call_emit = true;
3490 else
3491 emit_insn (normal_call_insns);
3493 currently_expanding_call--;
3495 free (stack_usage_map_buf);
3497 return target;
3500 /* A sibling call sequence invalidates any REG_EQUIV notes made for
3501 this function's incoming arguments.
3503 At the start of RTL generation we know the only REG_EQUIV notes
3504 in the rtl chain are those for incoming arguments, so we can look
3505 for REG_EQUIV notes between the start of the function and the
3506 NOTE_INSN_FUNCTION_BEG.
3508 This is (slight) overkill. We could keep track of the highest
3509 argument we clobber and be more selective in removing notes, but it
3510 does not seem to be worth the effort. */
3512 void
3513 fixup_tail_calls (void)
3515 rtx insn;
3517 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
3519 rtx note;
3521 /* There are never REG_EQUIV notes for the incoming arguments
3522 after the NOTE_INSN_FUNCTION_BEG note, so stop if we see it. */
3523 if (NOTE_P (insn)
3524 && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG)
3525 break;
3527 note = find_reg_note (insn, REG_EQUIV, 0);
3528 if (note)
3529 remove_note (insn, note);
3530 note = find_reg_note (insn, REG_EQUIV, 0);
3531 gcc_assert (!note);
3535 /* Traverse a list of TYPES and expand all complex types into their
3536 components. */
3537 static tree
3538 split_complex_types (tree types)
3540 tree p;
3542 /* Before allocating memory, check for the common case of no complex. */
3543 for (p = types; p; p = TREE_CHAIN (p))
3545 tree type = TREE_VALUE (p);
3546 if (TREE_CODE (type) == COMPLEX_TYPE
3547 && targetm.calls.split_complex_arg (type))
3548 goto found;
3550 return types;
3552 found:
3553 types = copy_list (types);
3555 for (p = types; p; p = TREE_CHAIN (p))
3557 tree complex_type = TREE_VALUE (p);
3559 if (TREE_CODE (complex_type) == COMPLEX_TYPE
3560 && targetm.calls.split_complex_arg (complex_type))
3562 tree next, imag;
3564 /* Rewrite complex type with component type. */
3565 TREE_VALUE (p) = TREE_TYPE (complex_type);
3566 next = TREE_CHAIN (p);
3568 /* Add another component type for the imaginary part. */
3569 imag = build_tree_list (NULL_TREE, TREE_VALUE (p));
3570 TREE_CHAIN (p) = imag;
3571 TREE_CHAIN (imag) = next;
3573 /* Skip the newly created node. */
3574 p = TREE_CHAIN (p);
3578 return types;
3581 /* Output a library call to function FUN (a SYMBOL_REF rtx).
3582 The RETVAL parameter specifies whether return value needs to be saved, other
3583 parameters are documented in the emit_library_call function below. */
3585 static rtx
3586 emit_library_call_value_1 (int retval, rtx orgfun, rtx value,
3587 enum libcall_type fn_type,
3588 enum machine_mode outmode, int nargs, va_list p)
3590 /* Total size in bytes of all the stack-parms scanned so far. */
3591 struct args_size args_size;
3592 /* Size of arguments before any adjustments (such as rounding). */
3593 struct args_size original_args_size;
3594 int argnum;
3595 rtx fun;
3596 /* Todo, choose the correct decl type of orgfun. Sadly this information
3597 isn't present here, so we default to native calling abi here. */
3598 tree fndecl ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
3599 tree fntype ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
3600 int inc;
3601 int count;
3602 rtx argblock = 0;
3603 CUMULATIVE_ARGS args_so_far_v;
3604 cumulative_args_t args_so_far;
3605 struct arg
3607 rtx value;
3608 enum machine_mode mode;
3609 rtx reg;
3610 int partial;
3611 struct locate_and_pad_arg_data locate;
3612 rtx save_area;
3614 struct arg *argvec;
3615 int old_inhibit_defer_pop = inhibit_defer_pop;
3616 rtx call_fusage = 0;
3617 rtx mem_value = 0;
3618 rtx valreg;
3619 int pcc_struct_value = 0;
3620 int struct_value_size = 0;
3621 int flags;
3622 int reg_parm_stack_space = 0;
3623 int needed;
3624 rtx before_call;
3625 tree tfom; /* type_for_mode (outmode, 0) */
3627 #ifdef REG_PARM_STACK_SPACE
3628 /* Define the boundary of the register parm stack space that needs to be
3629 save, if any. */
3630 int low_to_save = 0, high_to_save = 0;
3631 rtx save_area = 0; /* Place that it is saved. */
3632 #endif
3634 /* Size of the stack reserved for parameter registers. */
3635 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
3636 char *initial_stack_usage_map = stack_usage_map;
3637 char *stack_usage_map_buf = NULL;
3639 rtx struct_value = targetm.calls.struct_value_rtx (0, 0);
3641 #ifdef REG_PARM_STACK_SPACE
3642 reg_parm_stack_space = REG_PARM_STACK_SPACE ((tree) 0);
3643 #endif
3645 /* By default, library functions can not throw. */
3646 flags = ECF_NOTHROW;
3648 switch (fn_type)
3650 case LCT_NORMAL:
3651 break;
3652 case LCT_CONST:
3653 flags |= ECF_CONST;
3654 break;
3655 case LCT_PURE:
3656 flags |= ECF_PURE;
3657 break;
3658 case LCT_NORETURN:
3659 flags |= ECF_NORETURN;
3660 break;
3661 case LCT_THROW:
3662 flags = ECF_NORETURN;
3663 break;
3664 case LCT_RETURNS_TWICE:
3665 flags = ECF_RETURNS_TWICE;
3666 break;
3668 fun = orgfun;
3670 /* Ensure current function's preferred stack boundary is at least
3671 what we need. */
3672 if (crtl->preferred_stack_boundary < PREFERRED_STACK_BOUNDARY)
3673 crtl->preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
3675 /* If this kind of value comes back in memory,
3676 decide where in memory it should come back. */
3677 if (outmode != VOIDmode)
3679 tfom = lang_hooks.types.type_for_mode (outmode, 0);
3680 if (aggregate_value_p (tfom, 0))
3682 #ifdef PCC_STATIC_STRUCT_RETURN
3683 rtx pointer_reg
3684 = hard_function_value (build_pointer_type (tfom), 0, 0, 0);
3685 mem_value = gen_rtx_MEM (outmode, pointer_reg);
3686 pcc_struct_value = 1;
3687 if (value == 0)
3688 value = gen_reg_rtx (outmode);
3689 #else /* not PCC_STATIC_STRUCT_RETURN */
3690 struct_value_size = GET_MODE_SIZE (outmode);
3691 if (value != 0 && MEM_P (value))
3692 mem_value = value;
3693 else
3694 mem_value = assign_temp (tfom, 1, 1);
3695 #endif
3696 /* This call returns a big structure. */
3697 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
3700 else
3701 tfom = void_type_node;
3703 /* ??? Unfinished: must pass the memory address as an argument. */
3705 /* Copy all the libcall-arguments out of the varargs data
3706 and into a vector ARGVEC.
3708 Compute how to pass each argument. We only support a very small subset
3709 of the full argument passing conventions to limit complexity here since
3710 library functions shouldn't have many args. */
3712 argvec = XALLOCAVEC (struct arg, nargs + 1);
3713 memset (argvec, 0, (nargs + 1) * sizeof (struct arg));
3715 #ifdef INIT_CUMULATIVE_LIBCALL_ARGS
3716 INIT_CUMULATIVE_LIBCALL_ARGS (args_so_far_v, outmode, fun);
3717 #else
3718 INIT_CUMULATIVE_ARGS (args_so_far_v, NULL_TREE, fun, 0, nargs);
3719 #endif
3720 args_so_far = pack_cumulative_args (&args_so_far_v);
3722 args_size.constant = 0;
3723 args_size.var = 0;
3725 count = 0;
3727 push_temp_slots ();
3729 /* If there's a structure value address to be passed,
3730 either pass it in the special place, or pass it as an extra argument. */
3731 if (mem_value && struct_value == 0 && ! pcc_struct_value)
3733 rtx addr = XEXP (mem_value, 0);
3735 nargs++;
3737 /* Make sure it is a reasonable operand for a move or push insn. */
3738 if (!REG_P (addr) && !MEM_P (addr)
3739 && !(CONSTANT_P (addr)
3740 && targetm.legitimate_constant_p (Pmode, addr)))
3741 addr = force_operand (addr, NULL_RTX);
3743 argvec[count].value = addr;
3744 argvec[count].mode = Pmode;
3745 argvec[count].partial = 0;
3747 argvec[count].reg = targetm.calls.function_arg (args_so_far,
3748 Pmode, NULL_TREE, true);
3749 gcc_assert (targetm.calls.arg_partial_bytes (args_so_far, Pmode,
3750 NULL_TREE, 1) == 0);
3752 locate_and_pad_parm (Pmode, NULL_TREE,
3753 #ifdef STACK_PARMS_IN_REG_PARM_AREA
3755 #else
3756 argvec[count].reg != 0,
3757 #endif
3758 reg_parm_stack_space, 0,
3759 NULL_TREE, &args_size, &argvec[count].locate);
3761 if (argvec[count].reg == 0 || argvec[count].partial != 0
3762 || reg_parm_stack_space > 0)
3763 args_size.constant += argvec[count].locate.size.constant;
3765 targetm.calls.function_arg_advance (args_so_far, Pmode, (tree) 0, true);
3767 count++;
3770 for (; count < nargs; count++)
3772 rtx val = va_arg (p, rtx);
3773 enum machine_mode mode = (enum machine_mode) va_arg (p, int);
3774 int unsigned_p = 0;
3776 /* We cannot convert the arg value to the mode the library wants here;
3777 must do it earlier where we know the signedness of the arg. */
3778 gcc_assert (mode != BLKmode
3779 && (GET_MODE (val) == mode || GET_MODE (val) == VOIDmode));
3781 /* Make sure it is a reasonable operand for a move or push insn. */
3782 if (!REG_P (val) && !MEM_P (val)
3783 && !(CONSTANT_P (val) && targetm.legitimate_constant_p (mode, val)))
3784 val = force_operand (val, NULL_RTX);
3786 if (pass_by_reference (&args_so_far_v, mode, NULL_TREE, 1))
3788 rtx slot;
3789 int must_copy
3790 = !reference_callee_copied (&args_so_far_v, mode, NULL_TREE, 1);
3792 /* If this was a CONST function, it is now PURE since it now
3793 reads memory. */
3794 if (flags & ECF_CONST)
3796 flags &= ~ECF_CONST;
3797 flags |= ECF_PURE;
3800 if (MEM_P (val) && !must_copy)
3802 tree val_expr = MEM_EXPR (val);
3803 if (val_expr)
3804 mark_addressable (val_expr);
3805 slot = val;
3807 else
3809 slot = assign_temp (lang_hooks.types.type_for_mode (mode, 0),
3810 1, 1);
3811 emit_move_insn (slot, val);
3814 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
3815 gen_rtx_USE (VOIDmode, slot),
3816 call_fusage);
3817 if (must_copy)
3818 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
3819 gen_rtx_CLOBBER (VOIDmode,
3820 slot),
3821 call_fusage);
3823 mode = Pmode;
3824 val = force_operand (XEXP (slot, 0), NULL_RTX);
3827 mode = promote_function_mode (NULL_TREE, mode, &unsigned_p, NULL_TREE, 0);
3828 argvec[count].mode = mode;
3829 argvec[count].value = convert_modes (mode, GET_MODE (val), val, unsigned_p);
3830 argvec[count].reg = targetm.calls.function_arg (args_so_far, mode,
3831 NULL_TREE, true);
3833 argvec[count].partial
3834 = targetm.calls.arg_partial_bytes (args_so_far, mode, NULL_TREE, 1);
3836 if (argvec[count].reg == 0
3837 || argvec[count].partial != 0
3838 || reg_parm_stack_space > 0)
3840 locate_and_pad_parm (mode, NULL_TREE,
3841 #ifdef STACK_PARMS_IN_REG_PARM_AREA
3843 #else
3844 argvec[count].reg != 0,
3845 #endif
3846 reg_parm_stack_space, argvec[count].partial,
3847 NULL_TREE, &args_size, &argvec[count].locate);
3848 args_size.constant += argvec[count].locate.size.constant;
3849 gcc_assert (!argvec[count].locate.size.var);
3851 #ifdef BLOCK_REG_PADDING
3852 else
3853 /* The argument is passed entirely in registers. See at which
3854 end it should be padded. */
3855 argvec[count].locate.where_pad =
3856 BLOCK_REG_PADDING (mode, NULL_TREE,
3857 GET_MODE_SIZE (mode) <= UNITS_PER_WORD);
3858 #endif
3860 targetm.calls.function_arg_advance (args_so_far, mode, (tree) 0, true);
3863 /* If this machine requires an external definition for library
3864 functions, write one out. */
3865 assemble_external_libcall (fun);
3867 original_args_size = args_size;
3868 args_size.constant = (((args_size.constant
3869 + stack_pointer_delta
3870 + STACK_BYTES - 1)
3871 / STACK_BYTES
3872 * STACK_BYTES)
3873 - stack_pointer_delta);
3875 args_size.constant = MAX (args_size.constant,
3876 reg_parm_stack_space);
3878 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
3879 args_size.constant -= reg_parm_stack_space;
3881 if (args_size.constant > crtl->outgoing_args_size)
3882 crtl->outgoing_args_size = args_size.constant;
3884 if (flag_stack_usage_info && !ACCUMULATE_OUTGOING_ARGS)
3886 int pushed = args_size.constant + pending_stack_adjust;
3887 if (pushed > current_function_pushed_stack_size)
3888 current_function_pushed_stack_size = pushed;
3891 if (ACCUMULATE_OUTGOING_ARGS)
3893 /* Since the stack pointer will never be pushed, it is possible for
3894 the evaluation of a parm to clobber something we have already
3895 written to the stack. Since most function calls on RISC machines
3896 do not use the stack, this is uncommon, but must work correctly.
3898 Therefore, we save any area of the stack that was already written
3899 and that we are using. Here we set up to do this by making a new
3900 stack usage map from the old one.
3902 Another approach might be to try to reorder the argument
3903 evaluations to avoid this conflicting stack usage. */
3905 needed = args_size.constant;
3907 /* Since we will be writing into the entire argument area, the
3908 map must be allocated for its entire size, not just the part that
3909 is the responsibility of the caller. */
3910 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
3911 needed += reg_parm_stack_space;
3913 #ifdef ARGS_GROW_DOWNWARD
3914 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
3915 needed + 1);
3916 #else
3917 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
3918 needed);
3919 #endif
3920 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
3921 stack_usage_map = stack_usage_map_buf;
3923 if (initial_highest_arg_in_use)
3924 memcpy (stack_usage_map, initial_stack_usage_map,
3925 initial_highest_arg_in_use);
3927 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
3928 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
3929 highest_outgoing_arg_in_use - initial_highest_arg_in_use);
3930 needed = 0;
3932 /* We must be careful to use virtual regs before they're instantiated,
3933 and real regs afterwards. Loop optimization, for example, can create
3934 new libcalls after we've instantiated the virtual regs, and if we
3935 use virtuals anyway, they won't match the rtl patterns. */
3937 if (virtuals_instantiated)
3938 argblock = plus_constant (Pmode, stack_pointer_rtx,
3939 STACK_POINTER_OFFSET);
3940 else
3941 argblock = virtual_outgoing_args_rtx;
3943 else
3945 if (!PUSH_ARGS)
3946 argblock = push_block (GEN_INT (args_size.constant), 0, 0);
3949 /* If we push args individually in reverse order, perform stack alignment
3950 before the first push (the last arg). */
3951 if (argblock == 0 && PUSH_ARGS_REVERSED)
3952 anti_adjust_stack (GEN_INT (args_size.constant
3953 - original_args_size.constant));
3955 if (PUSH_ARGS_REVERSED)
3957 inc = -1;
3958 argnum = nargs - 1;
3960 else
3962 inc = 1;
3963 argnum = 0;
3966 #ifdef REG_PARM_STACK_SPACE
3967 if (ACCUMULATE_OUTGOING_ARGS)
3969 /* The argument list is the property of the called routine and it
3970 may clobber it. If the fixed area has been used for previous
3971 parameters, we must save and restore it. */
3972 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
3973 &low_to_save, &high_to_save);
3975 #endif
3977 /* Push the args that need to be pushed. */
3979 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
3980 are to be pushed. */
3981 for (count = 0; count < nargs; count++, argnum += inc)
3983 enum machine_mode mode = argvec[argnum].mode;
3984 rtx val = argvec[argnum].value;
3985 rtx reg = argvec[argnum].reg;
3986 int partial = argvec[argnum].partial;
3987 unsigned int parm_align = argvec[argnum].locate.boundary;
3988 int lower_bound = 0, upper_bound = 0, i;
3990 if (! (reg != 0 && partial == 0))
3992 rtx use;
3994 if (ACCUMULATE_OUTGOING_ARGS)
3996 /* If this is being stored into a pre-allocated, fixed-size,
3997 stack area, save any previous data at that location. */
3999 #ifdef ARGS_GROW_DOWNWARD
4000 /* stack_slot is negative, but we want to index stack_usage_map
4001 with positive values. */
4002 upper_bound = -argvec[argnum].locate.slot_offset.constant + 1;
4003 lower_bound = upper_bound - argvec[argnum].locate.size.constant;
4004 #else
4005 lower_bound = argvec[argnum].locate.slot_offset.constant;
4006 upper_bound = lower_bound + argvec[argnum].locate.size.constant;
4007 #endif
4009 i = lower_bound;
4010 /* Don't worry about things in the fixed argument area;
4011 it has already been saved. */
4012 if (i < reg_parm_stack_space)
4013 i = reg_parm_stack_space;
4014 while (i < upper_bound && stack_usage_map[i] == 0)
4015 i++;
4017 if (i < upper_bound)
4019 /* We need to make a save area. */
4020 unsigned int size
4021 = argvec[argnum].locate.size.constant * BITS_PER_UNIT;
4022 enum machine_mode save_mode
4023 = mode_for_size (size, MODE_INT, 1);
4024 rtx adr
4025 = plus_constant (Pmode, argblock,
4026 argvec[argnum].locate.offset.constant);
4027 rtx stack_area
4028 = gen_rtx_MEM (save_mode, memory_address (save_mode, adr));
4030 if (save_mode == BLKmode)
4032 argvec[argnum].save_area
4033 = assign_stack_temp (BLKmode,
4034 argvec[argnum].locate.size.constant
4037 emit_block_move (validize_mem (argvec[argnum].save_area),
4038 stack_area,
4039 GEN_INT (argvec[argnum].locate.size.constant),
4040 BLOCK_OP_CALL_PARM);
4042 else
4044 argvec[argnum].save_area = gen_reg_rtx (save_mode);
4046 emit_move_insn (argvec[argnum].save_area, stack_area);
4051 emit_push_insn (val, mode, NULL_TREE, NULL_RTX, parm_align,
4052 partial, reg, 0, argblock,
4053 GEN_INT (argvec[argnum].locate.offset.constant),
4054 reg_parm_stack_space,
4055 ARGS_SIZE_RTX (argvec[argnum].locate.alignment_pad));
4057 /* Now mark the segment we just used. */
4058 if (ACCUMULATE_OUTGOING_ARGS)
4059 for (i = lower_bound; i < upper_bound; i++)
4060 stack_usage_map[i] = 1;
4062 NO_DEFER_POP;
4064 /* Indicate argument access so that alias.c knows that these
4065 values are live. */
4066 if (argblock)
4067 use = plus_constant (Pmode, argblock,
4068 argvec[argnum].locate.offset.constant);
4069 else
4070 /* When arguments are pushed, trying to tell alias.c where
4071 exactly this argument is won't work, because the
4072 auto-increment causes confusion. So we merely indicate
4073 that we access something with a known mode somewhere on
4074 the stack. */
4075 use = gen_rtx_PLUS (Pmode, virtual_outgoing_args_rtx,
4076 gen_rtx_SCRATCH (Pmode));
4077 use = gen_rtx_MEM (argvec[argnum].mode, use);
4078 use = gen_rtx_USE (VOIDmode, use);
4079 call_fusage = gen_rtx_EXPR_LIST (VOIDmode, use, call_fusage);
4083 /* If we pushed args in forward order, perform stack alignment
4084 after pushing the last arg. */
4085 if (argblock == 0 && !PUSH_ARGS_REVERSED)
4086 anti_adjust_stack (GEN_INT (args_size.constant
4087 - original_args_size.constant));
4089 if (PUSH_ARGS_REVERSED)
4090 argnum = nargs - 1;
4091 else
4092 argnum = 0;
4094 fun = prepare_call_address (NULL, fun, NULL, &call_fusage, 0, 0);
4096 /* Now load any reg parms into their regs. */
4098 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
4099 are to be pushed. */
4100 for (count = 0; count < nargs; count++, argnum += inc)
4102 enum machine_mode mode = argvec[argnum].mode;
4103 rtx val = argvec[argnum].value;
4104 rtx reg = argvec[argnum].reg;
4105 int partial = argvec[argnum].partial;
4106 #ifdef BLOCK_REG_PADDING
4107 int size = 0;
4108 #endif
4110 /* Handle calls that pass values in multiple non-contiguous
4111 locations. The PA64 has examples of this for library calls. */
4112 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4113 emit_group_load (reg, val, NULL_TREE, GET_MODE_SIZE (mode));
4114 else if (reg != 0 && partial == 0)
4116 emit_move_insn (reg, val);
4117 #ifdef BLOCK_REG_PADDING
4118 size = GET_MODE_SIZE (argvec[argnum].mode);
4120 /* Copied from load_register_parameters. */
4122 /* Handle case where we have a value that needs shifting
4123 up to the msb. eg. a QImode value and we're padding
4124 upward on a BYTES_BIG_ENDIAN machine. */
4125 if (size < UNITS_PER_WORD
4126 && (argvec[argnum].locate.where_pad
4127 == (BYTES_BIG_ENDIAN ? upward : downward)))
4129 rtx x;
4130 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
4132 /* Assigning REG here rather than a temp makes CALL_FUSAGE
4133 report the whole reg as used. Strictly speaking, the
4134 call only uses SIZE bytes at the msb end, but it doesn't
4135 seem worth generating rtl to say that. */
4136 reg = gen_rtx_REG (word_mode, REGNO (reg));
4137 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
4138 if (x != reg)
4139 emit_move_insn (reg, x);
4141 #endif
4144 NO_DEFER_POP;
4147 /* Any regs containing parms remain in use through the call. */
4148 for (count = 0; count < nargs; count++)
4150 rtx reg = argvec[count].reg;
4151 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4152 use_group_regs (&call_fusage, reg);
4153 else if (reg != 0)
4155 int partial = argvec[count].partial;
4156 if (partial)
4158 int nregs;
4159 gcc_assert (partial % UNITS_PER_WORD == 0);
4160 nregs = partial / UNITS_PER_WORD;
4161 use_regs (&call_fusage, REGNO (reg), nregs);
4163 else
4164 use_reg (&call_fusage, reg);
4168 /* Pass the function the address in which to return a structure value. */
4169 if (mem_value != 0 && struct_value != 0 && ! pcc_struct_value)
4171 emit_move_insn (struct_value,
4172 force_reg (Pmode,
4173 force_operand (XEXP (mem_value, 0),
4174 NULL_RTX)));
4175 if (REG_P (struct_value))
4176 use_reg (&call_fusage, struct_value);
4179 /* Don't allow popping to be deferred, since then
4180 cse'ing of library calls could delete a call and leave the pop. */
4181 NO_DEFER_POP;
4182 valreg = (mem_value == 0 && outmode != VOIDmode
4183 ? hard_libcall_value (outmode, orgfun) : NULL_RTX);
4185 /* Stack must be properly aligned now. */
4186 gcc_assert (!(stack_pointer_delta
4187 & (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1)));
4189 before_call = get_last_insn ();
4191 /* We pass the old value of inhibit_defer_pop + 1 to emit_call_1, which
4192 will set inhibit_defer_pop to that value. */
4193 /* The return type is needed to decide how many bytes the function pops.
4194 Signedness plays no role in that, so for simplicity, we pretend it's
4195 always signed. We also assume that the list of arguments passed has
4196 no impact, so we pretend it is unknown. */
4198 emit_call_1 (fun, NULL,
4199 get_identifier (XSTR (orgfun, 0)),
4200 build_function_type (tfom, NULL_TREE),
4201 original_args_size.constant, args_size.constant,
4202 struct_value_size,
4203 targetm.calls.function_arg (args_so_far,
4204 VOIDmode, void_type_node, true),
4205 valreg,
4206 old_inhibit_defer_pop + 1, call_fusage, flags, args_so_far);
4208 /* Right-shift returned value if necessary. */
4209 if (!pcc_struct_value
4210 && TYPE_MODE (tfom) != BLKmode
4211 && targetm.calls.return_in_msb (tfom))
4213 shift_return_value (TYPE_MODE (tfom), false, valreg);
4214 valreg = gen_rtx_REG (TYPE_MODE (tfom), REGNO (valreg));
4217 /* For calls to `setjmp', etc., inform function.c:setjmp_warnings
4218 that it should complain if nonvolatile values are live. For
4219 functions that cannot return, inform flow that control does not
4220 fall through. */
4221 if (flags & ECF_NORETURN)
4223 /* The barrier note must be emitted
4224 immediately after the CALL_INSN. Some ports emit more than
4225 just a CALL_INSN above, so we must search for it here. */
4226 rtx last = get_last_insn ();
4227 while (!CALL_P (last))
4229 last = PREV_INSN (last);
4230 /* There was no CALL_INSN? */
4231 gcc_assert (last != before_call);
4234 emit_barrier_after (last);
4237 /* Consider that "regular" libcalls, i.e. all of them except for LCT_THROW
4238 and LCT_RETURNS_TWICE, cannot perform non-local gotos. */
4239 if (flags & ECF_NOTHROW)
4241 rtx last = get_last_insn ();
4242 while (!CALL_P (last))
4244 last = PREV_INSN (last);
4245 /* There was no CALL_INSN? */
4246 gcc_assert (last != before_call);
4249 make_reg_eh_region_note_nothrow_nononlocal (last);
4252 /* Now restore inhibit_defer_pop to its actual original value. */
4253 OK_DEFER_POP;
4255 pop_temp_slots ();
4257 /* Copy the value to the right place. */
4258 if (outmode != VOIDmode && retval)
4260 if (mem_value)
4262 if (value == 0)
4263 value = mem_value;
4264 if (value != mem_value)
4265 emit_move_insn (value, mem_value);
4267 else if (GET_CODE (valreg) == PARALLEL)
4269 if (value == 0)
4270 value = gen_reg_rtx (outmode);
4271 emit_group_store (value, valreg, NULL_TREE, GET_MODE_SIZE (outmode));
4273 else
4275 /* Convert to the proper mode if a promotion has been active. */
4276 if (GET_MODE (valreg) != outmode)
4278 int unsignedp = TYPE_UNSIGNED (tfom);
4280 gcc_assert (promote_function_mode (tfom, outmode, &unsignedp,
4281 fndecl ? TREE_TYPE (fndecl) : fntype, 1)
4282 == GET_MODE (valreg));
4283 valreg = convert_modes (outmode, GET_MODE (valreg), valreg, 0);
4286 if (value != 0)
4287 emit_move_insn (value, valreg);
4288 else
4289 value = valreg;
4293 if (ACCUMULATE_OUTGOING_ARGS)
4295 #ifdef REG_PARM_STACK_SPACE
4296 if (save_area)
4297 restore_fixed_argument_area (save_area, argblock,
4298 high_to_save, low_to_save);
4299 #endif
4301 /* If we saved any argument areas, restore them. */
4302 for (count = 0; count < nargs; count++)
4303 if (argvec[count].save_area)
4305 enum machine_mode save_mode = GET_MODE (argvec[count].save_area);
4306 rtx adr = plus_constant (Pmode, argblock,
4307 argvec[count].locate.offset.constant);
4308 rtx stack_area = gen_rtx_MEM (save_mode,
4309 memory_address (save_mode, adr));
4311 if (save_mode == BLKmode)
4312 emit_block_move (stack_area,
4313 validize_mem (argvec[count].save_area),
4314 GEN_INT (argvec[count].locate.size.constant),
4315 BLOCK_OP_CALL_PARM);
4316 else
4317 emit_move_insn (stack_area, argvec[count].save_area);
4320 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
4321 stack_usage_map = initial_stack_usage_map;
4324 free (stack_usage_map_buf);
4326 return value;
4330 /* Output a library call to function FUN (a SYMBOL_REF rtx)
4331 (emitting the queue unless NO_QUEUE is nonzero),
4332 for a value of mode OUTMODE,
4333 with NARGS different arguments, passed as alternating rtx values
4334 and machine_modes to convert them to.
4336 FN_TYPE should be LCT_NORMAL for `normal' calls, LCT_CONST for
4337 `const' calls, LCT_PURE for `pure' calls, or other LCT_ value for
4338 other types of library calls. */
4340 void
4341 emit_library_call (rtx orgfun, enum libcall_type fn_type,
4342 enum machine_mode outmode, int nargs, ...)
4344 va_list p;
4346 va_start (p, nargs);
4347 emit_library_call_value_1 (0, orgfun, NULL_RTX, fn_type, outmode, nargs, p);
4348 va_end (p);
4351 /* Like emit_library_call except that an extra argument, VALUE,
4352 comes second and says where to store the result.
4353 (If VALUE is zero, this function chooses a convenient way
4354 to return the value.
4356 This function returns an rtx for where the value is to be found.
4357 If VALUE is nonzero, VALUE is returned. */
4360 emit_library_call_value (rtx orgfun, rtx value,
4361 enum libcall_type fn_type,
4362 enum machine_mode outmode, int nargs, ...)
4364 rtx result;
4365 va_list p;
4367 va_start (p, nargs);
4368 result = emit_library_call_value_1 (1, orgfun, value, fn_type, outmode,
4369 nargs, p);
4370 va_end (p);
4372 return result;
4375 /* Store a single argument for a function call
4376 into the register or memory area where it must be passed.
4377 *ARG describes the argument value and where to pass it.
4379 ARGBLOCK is the address of the stack-block for all the arguments,
4380 or 0 on a machine where arguments are pushed individually.
4382 MAY_BE_ALLOCA nonzero says this could be a call to `alloca'
4383 so must be careful about how the stack is used.
4385 VARIABLE_SIZE nonzero says that this was a variable-sized outgoing
4386 argument stack. This is used if ACCUMULATE_OUTGOING_ARGS to indicate
4387 that we need not worry about saving and restoring the stack.
4389 FNDECL is the declaration of the function we are calling.
4391 Return nonzero if this arg should cause sibcall failure,
4392 zero otherwise. */
4394 static int
4395 store_one_arg (struct arg_data *arg, rtx argblock, int flags,
4396 int variable_size ATTRIBUTE_UNUSED, int reg_parm_stack_space)
4398 tree pval = arg->tree_value;
4399 rtx reg = 0;
4400 int partial = 0;
4401 int used = 0;
4402 int i, lower_bound = 0, upper_bound = 0;
4403 int sibcall_failure = 0;
4405 if (TREE_CODE (pval) == ERROR_MARK)
4406 return 1;
4408 /* Push a new temporary level for any temporaries we make for
4409 this argument. */
4410 push_temp_slots ();
4412 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL))
4414 /* If this is being stored into a pre-allocated, fixed-size, stack area,
4415 save any previous data at that location. */
4416 if (argblock && ! variable_size && arg->stack)
4418 #ifdef ARGS_GROW_DOWNWARD
4419 /* stack_slot is negative, but we want to index stack_usage_map
4420 with positive values. */
4421 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
4422 upper_bound = -INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1)) + 1;
4423 else
4424 upper_bound = 0;
4426 lower_bound = upper_bound - arg->locate.size.constant;
4427 #else
4428 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
4429 lower_bound = INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1));
4430 else
4431 lower_bound = 0;
4433 upper_bound = lower_bound + arg->locate.size.constant;
4434 #endif
4436 i = lower_bound;
4437 /* Don't worry about things in the fixed argument area;
4438 it has already been saved. */
4439 if (i < reg_parm_stack_space)
4440 i = reg_parm_stack_space;
4441 while (i < upper_bound && stack_usage_map[i] == 0)
4442 i++;
4444 if (i < upper_bound)
4446 /* We need to make a save area. */
4447 unsigned int size = arg->locate.size.constant * BITS_PER_UNIT;
4448 enum machine_mode save_mode = mode_for_size (size, MODE_INT, 1);
4449 rtx adr = memory_address (save_mode, XEXP (arg->stack_slot, 0));
4450 rtx stack_area = gen_rtx_MEM (save_mode, adr);
4452 if (save_mode == BLKmode)
4454 arg->save_area
4455 = assign_temp (TREE_TYPE (arg->tree_value), 1, 1);
4456 preserve_temp_slots (arg->save_area);
4457 emit_block_move (validize_mem (arg->save_area), stack_area,
4458 GEN_INT (arg->locate.size.constant),
4459 BLOCK_OP_CALL_PARM);
4461 else
4463 arg->save_area = gen_reg_rtx (save_mode);
4464 emit_move_insn (arg->save_area, stack_area);
4470 /* If this isn't going to be placed on both the stack and in registers,
4471 set up the register and number of words. */
4472 if (! arg->pass_on_stack)
4474 if (flags & ECF_SIBCALL)
4475 reg = arg->tail_call_reg;
4476 else
4477 reg = arg->reg;
4478 partial = arg->partial;
4481 /* Being passed entirely in a register. We shouldn't be called in
4482 this case. */
4483 gcc_assert (reg == 0 || partial != 0);
4485 /* If this arg needs special alignment, don't load the registers
4486 here. */
4487 if (arg->n_aligned_regs != 0)
4488 reg = 0;
4490 /* If this is being passed partially in a register, we can't evaluate
4491 it directly into its stack slot. Otherwise, we can. */
4492 if (arg->value == 0)
4494 /* stack_arg_under_construction is nonzero if a function argument is
4495 being evaluated directly into the outgoing argument list and
4496 expand_call must take special action to preserve the argument list
4497 if it is called recursively.
4499 For scalar function arguments stack_usage_map is sufficient to
4500 determine which stack slots must be saved and restored. Scalar
4501 arguments in general have pass_on_stack == 0.
4503 If this argument is initialized by a function which takes the
4504 address of the argument (a C++ constructor or a C function
4505 returning a BLKmode structure), then stack_usage_map is
4506 insufficient and expand_call must push the stack around the
4507 function call. Such arguments have pass_on_stack == 1.
4509 Note that it is always safe to set stack_arg_under_construction,
4510 but this generates suboptimal code if set when not needed. */
4512 if (arg->pass_on_stack)
4513 stack_arg_under_construction++;
4515 arg->value = expand_expr (pval,
4516 (partial
4517 || TYPE_MODE (TREE_TYPE (pval)) != arg->mode)
4518 ? NULL_RTX : arg->stack,
4519 VOIDmode, EXPAND_STACK_PARM);
4521 /* If we are promoting object (or for any other reason) the mode
4522 doesn't agree, convert the mode. */
4524 if (arg->mode != TYPE_MODE (TREE_TYPE (pval)))
4525 arg->value = convert_modes (arg->mode, TYPE_MODE (TREE_TYPE (pval)),
4526 arg->value, arg->unsignedp);
4528 if (arg->pass_on_stack)
4529 stack_arg_under_construction--;
4532 /* Check for overlap with already clobbered argument area. */
4533 if ((flags & ECF_SIBCALL)
4534 && MEM_P (arg->value)
4535 && mem_overlaps_already_clobbered_arg_p (XEXP (arg->value, 0),
4536 arg->locate.size.constant))
4537 sibcall_failure = 1;
4539 /* Don't allow anything left on stack from computation
4540 of argument to alloca. */
4541 if (flags & ECF_MAY_BE_ALLOCA)
4542 do_pending_stack_adjust ();
4544 if (arg->value == arg->stack)
4545 /* If the value is already in the stack slot, we are done. */
4547 else if (arg->mode != BLKmode)
4549 int size;
4550 unsigned int parm_align;
4552 /* Argument is a scalar, not entirely passed in registers.
4553 (If part is passed in registers, arg->partial says how much
4554 and emit_push_insn will take care of putting it there.)
4556 Push it, and if its size is less than the
4557 amount of space allocated to it,
4558 also bump stack pointer by the additional space.
4559 Note that in C the default argument promotions
4560 will prevent such mismatches. */
4562 size = GET_MODE_SIZE (arg->mode);
4563 /* Compute how much space the push instruction will push.
4564 On many machines, pushing a byte will advance the stack
4565 pointer by a halfword. */
4566 #ifdef PUSH_ROUNDING
4567 size = PUSH_ROUNDING (size);
4568 #endif
4569 used = size;
4571 /* Compute how much space the argument should get:
4572 round up to a multiple of the alignment for arguments. */
4573 if (none != FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)))
4574 used = (((size + PARM_BOUNDARY / BITS_PER_UNIT - 1)
4575 / (PARM_BOUNDARY / BITS_PER_UNIT))
4576 * (PARM_BOUNDARY / BITS_PER_UNIT));
4578 /* Compute the alignment of the pushed argument. */
4579 parm_align = arg->locate.boundary;
4580 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward)
4582 int pad = used - size;
4583 if (pad)
4585 unsigned int pad_align = (pad & -pad) * BITS_PER_UNIT;
4586 parm_align = MIN (parm_align, pad_align);
4590 /* This isn't already where we want it on the stack, so put it there.
4591 This can either be done with push or copy insns. */
4592 emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), NULL_RTX,
4593 parm_align, partial, reg, used - size, argblock,
4594 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
4595 ARGS_SIZE_RTX (arg->locate.alignment_pad));
4597 /* Unless this is a partially-in-register argument, the argument is now
4598 in the stack. */
4599 if (partial == 0)
4600 arg->value = arg->stack;
4602 else
4604 /* BLKmode, at least partly to be pushed. */
4606 unsigned int parm_align;
4607 int excess;
4608 rtx size_rtx;
4610 /* Pushing a nonscalar.
4611 If part is passed in registers, PARTIAL says how much
4612 and emit_push_insn will take care of putting it there. */
4614 /* Round its size up to a multiple
4615 of the allocation unit for arguments. */
4617 if (arg->locate.size.var != 0)
4619 excess = 0;
4620 size_rtx = ARGS_SIZE_RTX (arg->locate.size);
4622 else
4624 /* PUSH_ROUNDING has no effect on us, because emit_push_insn
4625 for BLKmode is careful to avoid it. */
4626 excess = (arg->locate.size.constant
4627 - int_size_in_bytes (TREE_TYPE (pval))
4628 + partial);
4629 size_rtx = expand_expr (size_in_bytes (TREE_TYPE (pval)),
4630 NULL_RTX, TYPE_MODE (sizetype),
4631 EXPAND_NORMAL);
4634 parm_align = arg->locate.boundary;
4636 /* When an argument is padded down, the block is aligned to
4637 PARM_BOUNDARY, but the actual argument isn't. */
4638 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward)
4640 if (arg->locate.size.var)
4641 parm_align = BITS_PER_UNIT;
4642 else if (excess)
4644 unsigned int excess_align = (excess & -excess) * BITS_PER_UNIT;
4645 parm_align = MIN (parm_align, excess_align);
4649 if ((flags & ECF_SIBCALL) && MEM_P (arg->value))
4651 /* emit_push_insn might not work properly if arg->value and
4652 argblock + arg->locate.offset areas overlap. */
4653 rtx x = arg->value;
4654 int i = 0;
4656 if (XEXP (x, 0) == crtl->args.internal_arg_pointer
4657 || (GET_CODE (XEXP (x, 0)) == PLUS
4658 && XEXP (XEXP (x, 0), 0) ==
4659 crtl->args.internal_arg_pointer
4660 && CONST_INT_P (XEXP (XEXP (x, 0), 1))))
4662 if (XEXP (x, 0) != crtl->args.internal_arg_pointer)
4663 i = INTVAL (XEXP (XEXP (x, 0), 1));
4665 /* expand_call should ensure this. */
4666 gcc_assert (!arg->locate.offset.var
4667 && arg->locate.size.var == 0
4668 && CONST_INT_P (size_rtx));
4670 if (arg->locate.offset.constant > i)
4672 if (arg->locate.offset.constant < i + INTVAL (size_rtx))
4673 sibcall_failure = 1;
4675 else if (arg->locate.offset.constant < i)
4677 /* Use arg->locate.size.constant instead of size_rtx
4678 because we only care about the part of the argument
4679 on the stack. */
4680 if (i < (arg->locate.offset.constant
4681 + arg->locate.size.constant))
4682 sibcall_failure = 1;
4684 else
4686 /* Even though they appear to be at the same location,
4687 if part of the outgoing argument is in registers,
4688 they aren't really at the same location. Check for
4689 this by making sure that the incoming size is the
4690 same as the outgoing size. */
4691 if (arg->locate.size.constant != INTVAL (size_rtx))
4692 sibcall_failure = 1;
4697 emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), size_rtx,
4698 parm_align, partial, reg, excess, argblock,
4699 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
4700 ARGS_SIZE_RTX (arg->locate.alignment_pad));
4702 /* Unless this is a partially-in-register argument, the argument is now
4703 in the stack.
4705 ??? Unlike the case above, in which we want the actual
4706 address of the data, so that we can load it directly into a
4707 register, here we want the address of the stack slot, so that
4708 it's properly aligned for word-by-word copying or something
4709 like that. It's not clear that this is always correct. */
4710 if (partial == 0)
4711 arg->value = arg->stack_slot;
4714 if (arg->reg && GET_CODE (arg->reg) == PARALLEL)
4716 tree type = TREE_TYPE (arg->tree_value);
4717 arg->parallel_value
4718 = emit_group_load_into_temps (arg->reg, arg->value, type,
4719 int_size_in_bytes (type));
4722 /* Mark all slots this store used. */
4723 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL)
4724 && argblock && ! variable_size && arg->stack)
4725 for (i = lower_bound; i < upper_bound; i++)
4726 stack_usage_map[i] = 1;
4728 /* Once we have pushed something, pops can't safely
4729 be deferred during the rest of the arguments. */
4730 NO_DEFER_POP;
4732 /* Free any temporary slots made in processing this argument. */
4733 pop_temp_slots ();
4735 return sibcall_failure;
4738 /* Nonzero if we do not know how to pass TYPE solely in registers. */
4740 bool
4741 must_pass_in_stack_var_size (enum machine_mode mode ATTRIBUTE_UNUSED,
4742 const_tree type)
4744 if (!type)
4745 return false;
4747 /* If the type has variable size... */
4748 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
4749 return true;
4751 /* If the type is marked as addressable (it is required
4752 to be constructed into the stack)... */
4753 if (TREE_ADDRESSABLE (type))
4754 return true;
4756 return false;
4759 /* Another version of the TARGET_MUST_PASS_IN_STACK hook. This one
4760 takes trailing padding of a structure into account. */
4761 /* ??? Should be able to merge these two by examining BLOCK_REG_PADDING. */
4763 bool
4764 must_pass_in_stack_var_size_or_pad (enum machine_mode mode, const_tree type)
4766 if (!type)
4767 return false;
4769 /* If the type has variable size... */
4770 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
4771 return true;
4773 /* If the type is marked as addressable (it is required
4774 to be constructed into the stack)... */
4775 if (TREE_ADDRESSABLE (type))
4776 return true;
4778 /* If the padding and mode of the type is such that a copy into
4779 a register would put it into the wrong part of the register. */
4780 if (mode == BLKmode
4781 && int_size_in_bytes (type) % (PARM_BOUNDARY / BITS_PER_UNIT)
4782 && (FUNCTION_ARG_PADDING (mode, type)
4783 == (BYTES_BIG_ENDIAN ? upward : downward)))
4784 return true;
4786 return false;