PR target/47230
[official-gcc.git] / gcc / calls.c
blobe8456528a59e1c6065bc69a2db4d1e37f132bc6f
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);
1108 /* Count arg position in order args appear. */
1109 int argpos;
1111 int i;
1113 args_size->constant = 0;
1114 args_size->var = 0;
1116 /* In this loop, we consider args in the order they are written.
1117 We fill up ARGS from the back. */
1119 i = num_actuals - 1;
1121 int j = i;
1122 call_expr_arg_iterator iter;
1123 tree arg;
1125 if (struct_value_addr_value)
1127 args[j].tree_value = struct_value_addr_value;
1128 j--;
1130 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
1132 tree argtype = TREE_TYPE (arg);
1133 if (targetm.calls.split_complex_arg
1134 && argtype
1135 && TREE_CODE (argtype) == COMPLEX_TYPE
1136 && targetm.calls.split_complex_arg (argtype))
1138 tree subtype = TREE_TYPE (argtype);
1139 args[j].tree_value = build1 (REALPART_EXPR, subtype, arg);
1140 j--;
1141 args[j].tree_value = build1 (IMAGPART_EXPR, subtype, arg);
1143 else
1144 args[j].tree_value = arg;
1145 j--;
1149 /* I counts args in order (to be) pushed; ARGPOS counts in order written. */
1150 for (argpos = 0; argpos < num_actuals; i--, argpos++)
1152 tree type = TREE_TYPE (args[i].tree_value);
1153 int unsignedp;
1154 enum machine_mode mode;
1156 /* Replace erroneous argument with constant zero. */
1157 if (type == error_mark_node || !COMPLETE_TYPE_P (type))
1158 args[i].tree_value = integer_zero_node, type = integer_type_node;
1160 /* If TYPE is a transparent union or record, pass things the way
1161 we would pass the first field of the union or record. We have
1162 already verified that the modes are the same. */
1163 if ((TREE_CODE (type) == UNION_TYPE || TREE_CODE (type) == RECORD_TYPE)
1164 && TYPE_TRANSPARENT_AGGR (type))
1165 type = TREE_TYPE (first_field (type));
1167 /* Decide where to pass this arg.
1169 args[i].reg is nonzero if all or part is passed in registers.
1171 args[i].partial is nonzero if part but not all is passed in registers,
1172 and the exact value says how many bytes are passed in registers.
1174 args[i].pass_on_stack is nonzero if the argument must at least be
1175 computed on the stack. It may then be loaded back into registers
1176 if args[i].reg is nonzero.
1178 These decisions are driven by the FUNCTION_... macros and must agree
1179 with those made by function.c. */
1181 /* See if this argument should be passed by invisible reference. */
1182 if (pass_by_reference (args_so_far_pnt, TYPE_MODE (type),
1183 type, argpos < n_named_args))
1185 bool callee_copies;
1186 tree base = NULL_TREE;
1188 callee_copies
1189 = reference_callee_copied (args_so_far_pnt, TYPE_MODE (type),
1190 type, argpos < n_named_args);
1192 /* If we're compiling a thunk, pass through invisible references
1193 instead of making a copy. */
1194 if (call_from_thunk_p
1195 || (callee_copies
1196 && !TREE_ADDRESSABLE (type)
1197 && (base = get_base_address (args[i].tree_value))
1198 && TREE_CODE (base) != SSA_NAME
1199 && (!DECL_P (base) || MEM_P (DECL_RTL (base)))))
1201 mark_addressable (args[i].tree_value);
1203 /* We can't use sibcalls if a callee-copied argument is
1204 stored in the current function's frame. */
1205 if (!call_from_thunk_p && DECL_P (base) && !TREE_STATIC (base))
1206 *may_tailcall = false;
1208 args[i].tree_value = build_fold_addr_expr_loc (loc,
1209 args[i].tree_value);
1210 type = TREE_TYPE (args[i].tree_value);
1212 if (*ecf_flags & ECF_CONST)
1213 *ecf_flags &= ~(ECF_CONST | ECF_LOOPING_CONST_OR_PURE);
1215 else
1217 /* We make a copy of the object and pass the address to the
1218 function being called. */
1219 rtx copy;
1221 if (!COMPLETE_TYPE_P (type)
1222 || TREE_CODE (TYPE_SIZE_UNIT (type)) != INTEGER_CST
1223 || (flag_stack_check == GENERIC_STACK_CHECK
1224 && compare_tree_int (TYPE_SIZE_UNIT (type),
1225 STACK_CHECK_MAX_VAR_SIZE) > 0))
1227 /* This is a variable-sized object. Make space on the stack
1228 for it. */
1229 rtx size_rtx = expr_size (args[i].tree_value);
1231 if (*old_stack_level == 0)
1233 emit_stack_save (SAVE_BLOCK, old_stack_level);
1234 *old_pending_adj = pending_stack_adjust;
1235 pending_stack_adjust = 0;
1238 /* We can pass TRUE as the 4th argument because we just
1239 saved the stack pointer and will restore it right after
1240 the call. */
1241 copy = allocate_dynamic_stack_space (size_rtx,
1242 TYPE_ALIGN (type),
1243 TYPE_ALIGN (type),
1244 true);
1245 copy = gen_rtx_MEM (BLKmode, copy);
1246 set_mem_attributes (copy, type, 1);
1248 else
1249 copy = assign_temp (type, 1, 0);
1251 store_expr (args[i].tree_value, copy, 0, false);
1253 /* Just change the const function to pure and then let
1254 the next test clear the pure based on
1255 callee_copies. */
1256 if (*ecf_flags & ECF_CONST)
1258 *ecf_flags &= ~ECF_CONST;
1259 *ecf_flags |= ECF_PURE;
1262 if (!callee_copies && *ecf_flags & ECF_PURE)
1263 *ecf_flags &= ~(ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
1265 args[i].tree_value
1266 = build_fold_addr_expr_loc (loc, make_tree (type, copy));
1267 type = TREE_TYPE (args[i].tree_value);
1268 *may_tailcall = false;
1272 unsignedp = TYPE_UNSIGNED (type);
1273 mode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
1274 fndecl ? TREE_TYPE (fndecl) : fntype, 0);
1276 args[i].unsignedp = unsignedp;
1277 args[i].mode = mode;
1279 args[i].reg = targetm.calls.function_arg (args_so_far, mode, type,
1280 argpos < n_named_args);
1282 /* If this is a sibling call and the machine has register windows, the
1283 register window has to be unwinded before calling the routine, so
1284 arguments have to go into the incoming registers. */
1285 if (targetm.calls.function_incoming_arg != targetm.calls.function_arg)
1286 args[i].tail_call_reg
1287 = targetm.calls.function_incoming_arg (args_so_far, mode, type,
1288 argpos < n_named_args);
1289 else
1290 args[i].tail_call_reg = args[i].reg;
1292 if (args[i].reg)
1293 args[i].partial
1294 = targetm.calls.arg_partial_bytes (args_so_far, mode, type,
1295 argpos < n_named_args);
1297 args[i].pass_on_stack = targetm.calls.must_pass_in_stack (mode, type);
1299 /* If FUNCTION_ARG returned a (parallel [(expr_list (nil) ...) ...]),
1300 it means that we are to pass this arg in the register(s) designated
1301 by the PARALLEL, but also to pass it in the stack. */
1302 if (args[i].reg && GET_CODE (args[i].reg) == PARALLEL
1303 && XEXP (XVECEXP (args[i].reg, 0, 0), 0) == 0)
1304 args[i].pass_on_stack = 1;
1306 /* If this is an addressable type, we must preallocate the stack
1307 since we must evaluate the object into its final location.
1309 If this is to be passed in both registers and the stack, it is simpler
1310 to preallocate. */
1311 if (TREE_ADDRESSABLE (type)
1312 || (args[i].pass_on_stack && args[i].reg != 0))
1313 *must_preallocate = 1;
1315 /* Compute the stack-size of this argument. */
1316 if (args[i].reg == 0 || args[i].partial != 0
1317 || reg_parm_stack_space > 0
1318 || args[i].pass_on_stack)
1319 locate_and_pad_parm (mode, type,
1320 #ifdef STACK_PARMS_IN_REG_PARM_AREA
1322 #else
1323 args[i].reg != 0,
1324 #endif
1325 reg_parm_stack_space,
1326 args[i].pass_on_stack ? 0 : args[i].partial,
1327 fndecl, args_size, &args[i].locate);
1328 #ifdef BLOCK_REG_PADDING
1329 else
1330 /* The argument is passed entirely in registers. See at which
1331 end it should be padded. */
1332 args[i].locate.where_pad =
1333 BLOCK_REG_PADDING (mode, type,
1334 int_size_in_bytes (type) <= UNITS_PER_WORD);
1335 #endif
1337 /* Update ARGS_SIZE, the total stack space for args so far. */
1339 args_size->constant += args[i].locate.size.constant;
1340 if (args[i].locate.size.var)
1341 ADD_PARM_SIZE (*args_size, args[i].locate.size.var);
1343 /* Increment ARGS_SO_FAR, which has info about which arg-registers
1344 have been used, etc. */
1346 targetm.calls.function_arg_advance (args_so_far, TYPE_MODE (type),
1347 type, argpos < n_named_args);
1351 /* Update ARGS_SIZE to contain the total size for the argument block.
1352 Return the original constant component of the argument block's size.
1354 REG_PARM_STACK_SPACE holds the number of bytes of stack space reserved
1355 for arguments passed in registers. */
1357 static int
1358 compute_argument_block_size (int reg_parm_stack_space,
1359 struct args_size *args_size,
1360 tree fndecl ATTRIBUTE_UNUSED,
1361 tree fntype ATTRIBUTE_UNUSED,
1362 int preferred_stack_boundary ATTRIBUTE_UNUSED)
1364 int unadjusted_args_size = args_size->constant;
1366 /* For accumulate outgoing args mode we don't need to align, since the frame
1367 will be already aligned. Align to STACK_BOUNDARY in order to prevent
1368 backends from generating misaligned frame sizes. */
1369 if (ACCUMULATE_OUTGOING_ARGS && preferred_stack_boundary > STACK_BOUNDARY)
1370 preferred_stack_boundary = STACK_BOUNDARY;
1372 /* Compute the actual size of the argument block required. The variable
1373 and constant sizes must be combined, the size may have to be rounded,
1374 and there may be a minimum required size. */
1376 if (args_size->var)
1378 args_size->var = ARGS_SIZE_TREE (*args_size);
1379 args_size->constant = 0;
1381 preferred_stack_boundary /= BITS_PER_UNIT;
1382 if (preferred_stack_boundary > 1)
1384 /* We don't handle this case yet. To handle it correctly we have
1385 to add the delta, round and subtract the delta.
1386 Currently no machine description requires this support. */
1387 gcc_assert (!(stack_pointer_delta & (preferred_stack_boundary - 1)));
1388 args_size->var = round_up (args_size->var, preferred_stack_boundary);
1391 if (reg_parm_stack_space > 0)
1393 args_size->var
1394 = size_binop (MAX_EXPR, args_size->var,
1395 ssize_int (reg_parm_stack_space));
1397 /* The area corresponding to register parameters is not to count in
1398 the size of the block we need. So make the adjustment. */
1399 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
1400 args_size->var
1401 = size_binop (MINUS_EXPR, args_size->var,
1402 ssize_int (reg_parm_stack_space));
1405 else
1407 preferred_stack_boundary /= BITS_PER_UNIT;
1408 if (preferred_stack_boundary < 1)
1409 preferred_stack_boundary = 1;
1410 args_size->constant = (((args_size->constant
1411 + stack_pointer_delta
1412 + preferred_stack_boundary - 1)
1413 / preferred_stack_boundary
1414 * preferred_stack_boundary)
1415 - stack_pointer_delta);
1417 args_size->constant = MAX (args_size->constant,
1418 reg_parm_stack_space);
1420 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
1421 args_size->constant -= reg_parm_stack_space;
1423 return unadjusted_args_size;
1426 /* Precompute parameters as needed for a function call.
1428 FLAGS is mask of ECF_* constants.
1430 NUM_ACTUALS is the number of arguments.
1432 ARGS is an array containing information for each argument; this
1433 routine fills in the INITIAL_VALUE and VALUE fields for each
1434 precomputed argument. */
1436 static void
1437 precompute_arguments (int num_actuals, struct arg_data *args)
1439 int i;
1441 /* If this is a libcall, then precompute all arguments so that we do not
1442 get extraneous instructions emitted as part of the libcall sequence. */
1444 /* If we preallocated the stack space, and some arguments must be passed
1445 on the stack, then we must precompute any parameter which contains a
1446 function call which will store arguments on the stack.
1447 Otherwise, evaluating the parameter may clobber previous parameters
1448 which have already been stored into the stack. (we have code to avoid
1449 such case by saving the outgoing stack arguments, but it results in
1450 worse code) */
1451 if (!ACCUMULATE_OUTGOING_ARGS)
1452 return;
1454 for (i = 0; i < num_actuals; i++)
1456 tree type;
1457 enum machine_mode mode;
1459 if (TREE_CODE (args[i].tree_value) != CALL_EXPR)
1460 continue;
1462 /* If this is an addressable type, we cannot pre-evaluate it. */
1463 type = TREE_TYPE (args[i].tree_value);
1464 gcc_assert (!TREE_ADDRESSABLE (type));
1466 args[i].initial_value = args[i].value
1467 = expand_normal (args[i].tree_value);
1469 mode = TYPE_MODE (type);
1470 if (mode != args[i].mode)
1472 int unsignedp = args[i].unsignedp;
1473 args[i].value
1474 = convert_modes (args[i].mode, mode,
1475 args[i].value, args[i].unsignedp);
1477 /* CSE will replace this only if it contains args[i].value
1478 pseudo, so convert it down to the declared mode using
1479 a SUBREG. */
1480 if (REG_P (args[i].value)
1481 && GET_MODE_CLASS (args[i].mode) == MODE_INT
1482 && promote_mode (type, mode, &unsignedp) != args[i].mode)
1484 args[i].initial_value
1485 = gen_lowpart_SUBREG (mode, args[i].value);
1486 SUBREG_PROMOTED_VAR_P (args[i].initial_value) = 1;
1487 SUBREG_PROMOTED_UNSIGNED_SET (args[i].initial_value,
1488 args[i].unsignedp);
1494 /* Given the current state of MUST_PREALLOCATE and information about
1495 arguments to a function call in NUM_ACTUALS, ARGS and ARGS_SIZE,
1496 compute and return the final value for MUST_PREALLOCATE. */
1498 static int
1499 finalize_must_preallocate (int must_preallocate, int num_actuals,
1500 struct arg_data *args, struct args_size *args_size)
1502 /* See if we have or want to preallocate stack space.
1504 If we would have to push a partially-in-regs parm
1505 before other stack parms, preallocate stack space instead.
1507 If the size of some parm is not a multiple of the required stack
1508 alignment, we must preallocate.
1510 If the total size of arguments that would otherwise create a copy in
1511 a temporary (such as a CALL) is more than half the total argument list
1512 size, preallocation is faster.
1514 Another reason to preallocate is if we have a machine (like the m88k)
1515 where stack alignment is required to be maintained between every
1516 pair of insns, not just when the call is made. However, we assume here
1517 that such machines either do not have push insns (and hence preallocation
1518 would occur anyway) or the problem is taken care of with
1519 PUSH_ROUNDING. */
1521 if (! must_preallocate)
1523 int partial_seen = 0;
1524 int copy_to_evaluate_size = 0;
1525 int i;
1527 for (i = 0; i < num_actuals && ! must_preallocate; i++)
1529 if (args[i].partial > 0 && ! args[i].pass_on_stack)
1530 partial_seen = 1;
1531 else if (partial_seen && args[i].reg == 0)
1532 must_preallocate = 1;
1534 if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode
1535 && (TREE_CODE (args[i].tree_value) == CALL_EXPR
1536 || TREE_CODE (args[i].tree_value) == TARGET_EXPR
1537 || TREE_CODE (args[i].tree_value) == COND_EXPR
1538 || TREE_ADDRESSABLE (TREE_TYPE (args[i].tree_value))))
1539 copy_to_evaluate_size
1540 += int_size_in_bytes (TREE_TYPE (args[i].tree_value));
1543 if (copy_to_evaluate_size * 2 >= args_size->constant
1544 && args_size->constant > 0)
1545 must_preallocate = 1;
1547 return must_preallocate;
1550 /* If we preallocated stack space, compute the address of each argument
1551 and store it into the ARGS array.
1553 We need not ensure it is a valid memory address here; it will be
1554 validized when it is used.
1556 ARGBLOCK is an rtx for the address of the outgoing arguments. */
1558 static void
1559 compute_argument_addresses (struct arg_data *args, rtx argblock, int num_actuals)
1561 if (argblock)
1563 rtx arg_reg = argblock;
1564 int i, arg_offset = 0;
1566 if (GET_CODE (argblock) == PLUS)
1567 arg_reg = XEXP (argblock, 0), arg_offset = INTVAL (XEXP (argblock, 1));
1569 for (i = 0; i < num_actuals; i++)
1571 rtx offset = ARGS_SIZE_RTX (args[i].locate.offset);
1572 rtx slot_offset = ARGS_SIZE_RTX (args[i].locate.slot_offset);
1573 rtx addr;
1574 unsigned int align, boundary;
1575 unsigned int units_on_stack = 0;
1576 enum machine_mode partial_mode = VOIDmode;
1578 /* Skip this parm if it will not be passed on the stack. */
1579 if (! args[i].pass_on_stack
1580 && args[i].reg != 0
1581 && args[i].partial == 0)
1582 continue;
1584 if (CONST_INT_P (offset))
1585 addr = plus_constant (Pmode, arg_reg, INTVAL (offset));
1586 else
1587 addr = gen_rtx_PLUS (Pmode, arg_reg, offset);
1589 addr = plus_constant (Pmode, addr, arg_offset);
1591 if (args[i].partial != 0)
1593 /* Only part of the parameter is being passed on the stack.
1594 Generate a simple memory reference of the correct size. */
1595 units_on_stack = args[i].locate.size.constant;
1596 partial_mode = mode_for_size (units_on_stack * BITS_PER_UNIT,
1597 MODE_INT, 1);
1598 args[i].stack = gen_rtx_MEM (partial_mode, addr);
1599 set_mem_size (args[i].stack, units_on_stack);
1601 else
1603 args[i].stack = gen_rtx_MEM (args[i].mode, addr);
1604 set_mem_attributes (args[i].stack,
1605 TREE_TYPE (args[i].tree_value), 1);
1607 align = BITS_PER_UNIT;
1608 boundary = args[i].locate.boundary;
1609 if (args[i].locate.where_pad != downward)
1610 align = boundary;
1611 else if (CONST_INT_P (offset))
1613 align = INTVAL (offset) * BITS_PER_UNIT | boundary;
1614 align = align & -align;
1616 set_mem_align (args[i].stack, align);
1618 if (CONST_INT_P (slot_offset))
1619 addr = plus_constant (Pmode, arg_reg, INTVAL (slot_offset));
1620 else
1621 addr = gen_rtx_PLUS (Pmode, arg_reg, slot_offset);
1623 addr = plus_constant (Pmode, addr, arg_offset);
1625 if (args[i].partial != 0)
1627 /* Only part of the parameter is being passed on the stack.
1628 Generate a simple memory reference of the correct size.
1630 args[i].stack_slot = gen_rtx_MEM (partial_mode, addr);
1631 set_mem_size (args[i].stack_slot, units_on_stack);
1633 else
1635 args[i].stack_slot = gen_rtx_MEM (args[i].mode, addr);
1636 set_mem_attributes (args[i].stack_slot,
1637 TREE_TYPE (args[i].tree_value), 1);
1639 set_mem_align (args[i].stack_slot, args[i].locate.boundary);
1641 /* Function incoming arguments may overlap with sibling call
1642 outgoing arguments and we cannot allow reordering of reads
1643 from function arguments with stores to outgoing arguments
1644 of sibling calls. */
1645 set_mem_alias_set (args[i].stack, 0);
1646 set_mem_alias_set (args[i].stack_slot, 0);
1651 /* Given a FNDECL and EXP, return an rtx suitable for use as a target address
1652 in a call instruction.
1654 FNDECL is the tree node for the target function. For an indirect call
1655 FNDECL will be NULL_TREE.
1657 ADDR is the operand 0 of CALL_EXPR for this call. */
1659 static rtx
1660 rtx_for_function_call (tree fndecl, tree addr)
1662 rtx funexp;
1664 /* Get the function to call, in the form of RTL. */
1665 if (fndecl)
1667 if (!TREE_USED (fndecl) && fndecl != current_function_decl)
1668 TREE_USED (fndecl) = 1;
1670 /* Get a SYMBOL_REF rtx for the function address. */
1671 funexp = XEXP (DECL_RTL (fndecl), 0);
1673 else
1674 /* Generate an rtx (probably a pseudo-register) for the address. */
1676 push_temp_slots ();
1677 funexp = expand_normal (addr);
1678 pop_temp_slots (); /* FUNEXP can't be BLKmode. */
1680 return funexp;
1683 /* Internal state for internal_arg_pointer_based_exp and its helpers. */
1684 static struct
1686 /* Last insn that has been scanned by internal_arg_pointer_based_exp_scan,
1687 or NULL_RTX if none has been scanned yet. */
1688 rtx scan_start;
1689 /* Vector indexed by REGNO - FIRST_PSEUDO_REGISTER, recording if a pseudo is
1690 based on crtl->args.internal_arg_pointer. The element is NULL_RTX if the
1691 pseudo isn't based on it, a CONST_INT offset if the pseudo is based on it
1692 with fixed offset, or PC if this is with variable or unknown offset. */
1693 vec<rtx> cache;
1694 } internal_arg_pointer_exp_state;
1696 static rtx internal_arg_pointer_based_exp (rtx, bool);
1698 /* Helper function for internal_arg_pointer_based_exp. Scan insns in
1699 the tail call sequence, starting with first insn that hasn't been
1700 scanned yet, and note for each pseudo on the LHS whether it is based
1701 on crtl->args.internal_arg_pointer or not, and what offset from that
1702 that pointer it has. */
1704 static void
1705 internal_arg_pointer_based_exp_scan (void)
1707 rtx insn, scan_start = internal_arg_pointer_exp_state.scan_start;
1709 if (scan_start == NULL_RTX)
1710 insn = get_insns ();
1711 else
1712 insn = NEXT_INSN (scan_start);
1714 while (insn)
1716 rtx set = single_set (insn);
1717 if (set && REG_P (SET_DEST (set)) && !HARD_REGISTER_P (SET_DEST (set)))
1719 rtx val = NULL_RTX;
1720 unsigned int idx = REGNO (SET_DEST (set)) - FIRST_PSEUDO_REGISTER;
1721 /* Punt on pseudos set multiple times. */
1722 if (idx < internal_arg_pointer_exp_state.cache.length ()
1723 && (internal_arg_pointer_exp_state.cache[idx]
1724 != NULL_RTX))
1725 val = pc_rtx;
1726 else
1727 val = internal_arg_pointer_based_exp (SET_SRC (set), false);
1728 if (val != NULL_RTX)
1730 if (idx >= internal_arg_pointer_exp_state.cache.length ())
1731 internal_arg_pointer_exp_state.cache
1732 .safe_grow_cleared (idx + 1);
1733 internal_arg_pointer_exp_state.cache[idx] = val;
1736 if (NEXT_INSN (insn) == NULL_RTX)
1737 scan_start = insn;
1738 insn = NEXT_INSN (insn);
1741 internal_arg_pointer_exp_state.scan_start = scan_start;
1744 /* Helper function for internal_arg_pointer_based_exp, called through
1745 for_each_rtx. Return 1 if *LOC is a register based on
1746 crtl->args.internal_arg_pointer. Return -1 if *LOC is not based on it
1747 and the subexpressions need not be examined. Otherwise return 0. */
1749 static int
1750 internal_arg_pointer_based_exp_1 (rtx *loc, void *data ATTRIBUTE_UNUSED)
1752 if (REG_P (*loc) && internal_arg_pointer_based_exp (*loc, false) != NULL_RTX)
1753 return 1;
1754 if (MEM_P (*loc))
1755 return -1;
1756 return 0;
1759 /* Compute whether RTL is based on crtl->args.internal_arg_pointer. Return
1760 NULL_RTX if RTL isn't based on it, a CONST_INT offset if RTL is based on
1761 it with fixed offset, or PC if this is with variable or unknown offset.
1762 TOPLEVEL is true if the function is invoked at the topmost level. */
1764 static rtx
1765 internal_arg_pointer_based_exp (rtx rtl, bool toplevel)
1767 if (CONSTANT_P (rtl))
1768 return NULL_RTX;
1770 if (rtl == crtl->args.internal_arg_pointer)
1771 return const0_rtx;
1773 if (REG_P (rtl) && HARD_REGISTER_P (rtl))
1774 return NULL_RTX;
1776 if (GET_CODE (rtl) == PLUS && CONST_INT_P (XEXP (rtl, 1)))
1778 rtx val = internal_arg_pointer_based_exp (XEXP (rtl, 0), toplevel);
1779 if (val == NULL_RTX || val == pc_rtx)
1780 return val;
1781 return plus_constant (Pmode, val, INTVAL (XEXP (rtl, 1)));
1784 /* When called at the topmost level, scan pseudo assignments in between the
1785 last scanned instruction in the tail call sequence and the latest insn
1786 in that sequence. */
1787 if (toplevel)
1788 internal_arg_pointer_based_exp_scan ();
1790 if (REG_P (rtl))
1792 unsigned int idx = REGNO (rtl) - FIRST_PSEUDO_REGISTER;
1793 if (idx < internal_arg_pointer_exp_state.cache.length ())
1794 return internal_arg_pointer_exp_state.cache[idx];
1796 return NULL_RTX;
1799 if (for_each_rtx (&rtl, internal_arg_pointer_based_exp_1, NULL))
1800 return pc_rtx;
1802 return NULL_RTX;
1805 /* Return true if and only if SIZE storage units (usually bytes)
1806 starting from address ADDR overlap with already clobbered argument
1807 area. This function is used to determine if we should give up a
1808 sibcall. */
1810 static bool
1811 mem_overlaps_already_clobbered_arg_p (rtx addr, unsigned HOST_WIDE_INT size)
1813 HOST_WIDE_INT i;
1814 rtx val;
1816 if (bitmap_empty_p (stored_args_map))
1817 return false;
1818 val = internal_arg_pointer_based_exp (addr, true);
1819 if (val == NULL_RTX)
1820 return false;
1821 else if (val == pc_rtx)
1822 return true;
1823 else
1824 i = INTVAL (val);
1825 #ifdef STACK_GROWS_DOWNWARD
1826 i -= crtl->args.pretend_args_size;
1827 #else
1828 i += crtl->args.pretend_args_size;
1829 #endif
1831 #ifdef ARGS_GROW_DOWNWARD
1832 i = -i - size;
1833 #endif
1834 if (size > 0)
1836 unsigned HOST_WIDE_INT k;
1838 for (k = 0; k < size; k++)
1839 if (i + k < SBITMAP_SIZE (stored_args_map)
1840 && bitmap_bit_p (stored_args_map, i + k))
1841 return true;
1844 return false;
1847 /* Do the register loads required for any wholly-register parms or any
1848 parms which are passed both on the stack and in a register. Their
1849 expressions were already evaluated.
1851 Mark all register-parms as living through the call, putting these USE
1852 insns in the CALL_INSN_FUNCTION_USAGE field.
1854 When IS_SIBCALL, perform the check_sibcall_argument_overlap
1855 checking, setting *SIBCALL_FAILURE if appropriate. */
1857 static void
1858 load_register_parameters (struct arg_data *args, int num_actuals,
1859 rtx *call_fusage, int flags, int is_sibcall,
1860 int *sibcall_failure)
1862 int i, j;
1864 for (i = 0; i < num_actuals; i++)
1866 rtx reg = ((flags & ECF_SIBCALL)
1867 ? args[i].tail_call_reg : args[i].reg);
1868 if (reg)
1870 int partial = args[i].partial;
1871 int nregs;
1872 int size = 0;
1873 rtx before_arg = get_last_insn ();
1874 /* Set non-negative if we must move a word at a time, even if
1875 just one word (e.g, partial == 4 && mode == DFmode). Set
1876 to -1 if we just use a normal move insn. This value can be
1877 zero if the argument is a zero size structure. */
1878 nregs = -1;
1879 if (GET_CODE (reg) == PARALLEL)
1881 else if (partial)
1883 gcc_assert (partial % UNITS_PER_WORD == 0);
1884 nregs = partial / UNITS_PER_WORD;
1886 else if (TYPE_MODE (TREE_TYPE (args[i].tree_value)) == BLKmode)
1888 size = int_size_in_bytes (TREE_TYPE (args[i].tree_value));
1889 nregs = (size + (UNITS_PER_WORD - 1)) / UNITS_PER_WORD;
1891 else
1892 size = GET_MODE_SIZE (args[i].mode);
1894 /* Handle calls that pass values in multiple non-contiguous
1895 locations. The Irix 6 ABI has examples of this. */
1897 if (GET_CODE (reg) == PARALLEL)
1898 emit_group_move (reg, args[i].parallel_value);
1900 /* If simple case, just do move. If normal partial, store_one_arg
1901 has already loaded the register for us. In all other cases,
1902 load the register(s) from memory. */
1904 else if (nregs == -1)
1906 emit_move_insn (reg, args[i].value);
1907 #ifdef BLOCK_REG_PADDING
1908 /* Handle case where we have a value that needs shifting
1909 up to the msb. eg. a QImode value and we're padding
1910 upward on a BYTES_BIG_ENDIAN machine. */
1911 if (size < UNITS_PER_WORD
1912 && (args[i].locate.where_pad
1913 == (BYTES_BIG_ENDIAN ? upward : downward)))
1915 rtx x;
1916 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
1918 /* Assigning REG here rather than a temp makes CALL_FUSAGE
1919 report the whole reg as used. Strictly speaking, the
1920 call only uses SIZE bytes at the msb end, but it doesn't
1921 seem worth generating rtl to say that. */
1922 reg = gen_rtx_REG (word_mode, REGNO (reg));
1923 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
1924 if (x != reg)
1925 emit_move_insn (reg, x);
1927 #endif
1930 /* If we have pre-computed the values to put in the registers in
1931 the case of non-aligned structures, copy them in now. */
1933 else if (args[i].n_aligned_regs != 0)
1934 for (j = 0; j < args[i].n_aligned_regs; j++)
1935 emit_move_insn (gen_rtx_REG (word_mode, REGNO (reg) + j),
1936 args[i].aligned_regs[j]);
1938 else if (partial == 0 || args[i].pass_on_stack)
1940 rtx mem = validize_mem (copy_rtx (args[i].value));
1942 /* Check for overlap with already clobbered argument area,
1943 providing that this has non-zero size. */
1944 if (is_sibcall
1945 && (size == 0
1946 || mem_overlaps_already_clobbered_arg_p
1947 (XEXP (args[i].value, 0), size)))
1948 *sibcall_failure = 1;
1950 /* Handle a BLKmode that needs shifting. */
1951 if (nregs == 1 && size < UNITS_PER_WORD
1952 #ifdef BLOCK_REG_PADDING
1953 && args[i].locate.where_pad == downward
1954 #else
1955 && BYTES_BIG_ENDIAN
1956 #endif
1959 rtx tem = operand_subword_force (mem, 0, args[i].mode);
1960 rtx ri = gen_rtx_REG (word_mode, REGNO (reg));
1961 rtx x = gen_reg_rtx (word_mode);
1962 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
1963 enum tree_code dir = BYTES_BIG_ENDIAN ? RSHIFT_EXPR
1964 : LSHIFT_EXPR;
1966 emit_move_insn (x, tem);
1967 x = expand_shift (dir, word_mode, x, shift, ri, 1);
1968 if (x != ri)
1969 emit_move_insn (ri, x);
1971 else
1972 move_block_to_reg (REGNO (reg), mem, nregs, args[i].mode);
1975 /* When a parameter is a block, and perhaps in other cases, it is
1976 possible that it did a load from an argument slot that was
1977 already clobbered. */
1978 if (is_sibcall
1979 && check_sibcall_argument_overlap (before_arg, &args[i], 0))
1980 *sibcall_failure = 1;
1982 /* Handle calls that pass values in multiple non-contiguous
1983 locations. The Irix 6 ABI has examples of this. */
1984 if (GET_CODE (reg) == PARALLEL)
1985 use_group_regs (call_fusage, reg);
1986 else if (nregs == -1)
1987 use_reg_mode (call_fusage, reg,
1988 TYPE_MODE (TREE_TYPE (args[i].tree_value)));
1989 else if (nregs > 0)
1990 use_regs (call_fusage, REGNO (reg), nregs);
1995 /* We need to pop PENDING_STACK_ADJUST bytes. But, if the arguments
1996 wouldn't fill up an even multiple of PREFERRED_UNIT_STACK_BOUNDARY
1997 bytes, then we would need to push some additional bytes to pad the
1998 arguments. So, we compute an adjust to the stack pointer for an
1999 amount that will leave the stack under-aligned by UNADJUSTED_ARGS_SIZE
2000 bytes. Then, when the arguments are pushed the stack will be perfectly
2001 aligned. ARGS_SIZE->CONSTANT is set to the number of bytes that should
2002 be popped after the call. Returns the adjustment. */
2004 static int
2005 combine_pending_stack_adjustment_and_call (int unadjusted_args_size,
2006 struct args_size *args_size,
2007 unsigned int preferred_unit_stack_boundary)
2009 /* The number of bytes to pop so that the stack will be
2010 under-aligned by UNADJUSTED_ARGS_SIZE bytes. */
2011 HOST_WIDE_INT adjustment;
2012 /* The alignment of the stack after the arguments are pushed, if we
2013 just pushed the arguments without adjust the stack here. */
2014 unsigned HOST_WIDE_INT unadjusted_alignment;
2016 unadjusted_alignment
2017 = ((stack_pointer_delta + unadjusted_args_size)
2018 % preferred_unit_stack_boundary);
2020 /* We want to get rid of as many of the PENDING_STACK_ADJUST bytes
2021 as possible -- leaving just enough left to cancel out the
2022 UNADJUSTED_ALIGNMENT. In other words, we want to ensure that the
2023 PENDING_STACK_ADJUST is non-negative, and congruent to
2024 -UNADJUSTED_ALIGNMENT modulo the PREFERRED_UNIT_STACK_BOUNDARY. */
2026 /* Begin by trying to pop all the bytes. */
2027 unadjusted_alignment
2028 = (unadjusted_alignment
2029 - (pending_stack_adjust % preferred_unit_stack_boundary));
2030 adjustment = pending_stack_adjust;
2031 /* Push enough additional bytes that the stack will be aligned
2032 after the arguments are pushed. */
2033 if (preferred_unit_stack_boundary > 1)
2035 if (unadjusted_alignment > 0)
2036 adjustment -= preferred_unit_stack_boundary - unadjusted_alignment;
2037 else
2038 adjustment += unadjusted_alignment;
2041 /* Now, sets ARGS_SIZE->CONSTANT so that we pop the right number of
2042 bytes after the call. The right number is the entire
2043 PENDING_STACK_ADJUST less our ADJUSTMENT plus the amount required
2044 by the arguments in the first place. */
2045 args_size->constant
2046 = pending_stack_adjust - adjustment + unadjusted_args_size;
2048 return adjustment;
2051 /* Scan X expression if it does not dereference any argument slots
2052 we already clobbered by tail call arguments (as noted in stored_args_map
2053 bitmap).
2054 Return nonzero if X expression dereferences such argument slots,
2055 zero otherwise. */
2057 static int
2058 check_sibcall_argument_overlap_1 (rtx x)
2060 RTX_CODE code;
2061 int i, j;
2062 const char *fmt;
2064 if (x == NULL_RTX)
2065 return 0;
2067 code = GET_CODE (x);
2069 /* We need not check the operands of the CALL expression itself. */
2070 if (code == CALL)
2071 return 0;
2073 if (code == MEM)
2074 return mem_overlaps_already_clobbered_arg_p (XEXP (x, 0),
2075 GET_MODE_SIZE (GET_MODE (x)));
2077 /* Scan all subexpressions. */
2078 fmt = GET_RTX_FORMAT (code);
2079 for (i = 0; i < GET_RTX_LENGTH (code); i++, fmt++)
2081 if (*fmt == 'e')
2083 if (check_sibcall_argument_overlap_1 (XEXP (x, i)))
2084 return 1;
2086 else if (*fmt == 'E')
2088 for (j = 0; j < XVECLEN (x, i); j++)
2089 if (check_sibcall_argument_overlap_1 (XVECEXP (x, i, j)))
2090 return 1;
2093 return 0;
2096 /* Scan sequence after INSN if it does not dereference any argument slots
2097 we already clobbered by tail call arguments (as noted in stored_args_map
2098 bitmap). If MARK_STORED_ARGS_MAP, add stack slots for ARG to
2099 stored_args_map bitmap afterwards (when ARG is a register MARK_STORED_ARGS_MAP
2100 should be 0). Return nonzero if sequence after INSN dereferences such argument
2101 slots, zero otherwise. */
2103 static int
2104 check_sibcall_argument_overlap (rtx insn, struct arg_data *arg, int mark_stored_args_map)
2106 int low, high;
2108 if (insn == NULL_RTX)
2109 insn = get_insns ();
2110 else
2111 insn = NEXT_INSN (insn);
2113 for (; insn; insn = NEXT_INSN (insn))
2114 if (INSN_P (insn)
2115 && check_sibcall_argument_overlap_1 (PATTERN (insn)))
2116 break;
2118 if (mark_stored_args_map)
2120 #ifdef ARGS_GROW_DOWNWARD
2121 low = -arg->locate.slot_offset.constant - arg->locate.size.constant;
2122 #else
2123 low = arg->locate.slot_offset.constant;
2124 #endif
2126 for (high = low + arg->locate.size.constant; low < high; low++)
2127 bitmap_set_bit (stored_args_map, low);
2129 return insn != NULL_RTX;
2132 /* Given that a function returns a value of mode MODE at the most
2133 significant end of hard register VALUE, shift VALUE left or right
2134 as specified by LEFT_P. Return true if some action was needed. */
2136 bool
2137 shift_return_value (enum machine_mode mode, bool left_p, rtx value)
2139 HOST_WIDE_INT shift;
2141 gcc_assert (REG_P (value) && HARD_REGISTER_P (value));
2142 shift = GET_MODE_BITSIZE (GET_MODE (value)) - GET_MODE_BITSIZE (mode);
2143 if (shift == 0)
2144 return false;
2146 /* Use ashr rather than lshr for right shifts. This is for the benefit
2147 of the MIPS port, which requires SImode values to be sign-extended
2148 when stored in 64-bit registers. */
2149 if (!force_expand_binop (GET_MODE (value), left_p ? ashl_optab : ashr_optab,
2150 value, GEN_INT (shift), value, 1, OPTAB_WIDEN))
2151 gcc_unreachable ();
2152 return true;
2155 /* If X is a likely-spilled register value, copy it to a pseudo
2156 register and return that register. Return X otherwise. */
2158 static rtx
2159 avoid_likely_spilled_reg (rtx x)
2161 rtx new_rtx;
2163 if (REG_P (x)
2164 && HARD_REGISTER_P (x)
2165 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (x))))
2167 /* Make sure that we generate a REG rather than a CONCAT.
2168 Moves into CONCATs can need nontrivial instructions,
2169 and the whole point of this function is to avoid
2170 using the hard register directly in such a situation. */
2171 generating_concat_p = 0;
2172 new_rtx = gen_reg_rtx (GET_MODE (x));
2173 generating_concat_p = 1;
2174 emit_move_insn (new_rtx, x);
2175 return new_rtx;
2177 return x;
2180 /* Generate all the code for a CALL_EXPR exp
2181 and return an rtx for its value.
2182 Store the value in TARGET (specified as an rtx) if convenient.
2183 If the value is stored in TARGET then TARGET is returned.
2184 If IGNORE is nonzero, then we ignore the value of the function call. */
2187 expand_call (tree exp, rtx target, int ignore)
2189 /* Nonzero if we are currently expanding a call. */
2190 static int currently_expanding_call = 0;
2192 /* RTX for the function to be called. */
2193 rtx funexp;
2194 /* Sequence of insns to perform a normal "call". */
2195 rtx normal_call_insns = NULL_RTX;
2196 /* Sequence of insns to perform a tail "call". */
2197 rtx tail_call_insns = NULL_RTX;
2198 /* Data type of the function. */
2199 tree funtype;
2200 tree type_arg_types;
2201 tree rettype;
2202 /* Declaration of the function being called,
2203 or 0 if the function is computed (not known by name). */
2204 tree fndecl = 0;
2205 /* The type of the function being called. */
2206 tree fntype;
2207 bool try_tail_call = CALL_EXPR_TAILCALL (exp);
2208 int pass;
2210 /* Register in which non-BLKmode value will be returned,
2211 or 0 if no value or if value is BLKmode. */
2212 rtx valreg;
2213 /* Address where we should return a BLKmode value;
2214 0 if value not BLKmode. */
2215 rtx structure_value_addr = 0;
2216 /* Nonzero if that address is being passed by treating it as
2217 an extra, implicit first parameter. Otherwise,
2218 it is passed by being copied directly into struct_value_rtx. */
2219 int structure_value_addr_parm = 0;
2220 /* Holds the value of implicit argument for the struct value. */
2221 tree structure_value_addr_value = NULL_TREE;
2222 /* Size of aggregate value wanted, or zero if none wanted
2223 or if we are using the non-reentrant PCC calling convention
2224 or expecting the value in registers. */
2225 HOST_WIDE_INT struct_value_size = 0;
2226 /* Nonzero if called function returns an aggregate in memory PCC style,
2227 by returning the address of where to find it. */
2228 int pcc_struct_value = 0;
2229 rtx struct_value = 0;
2231 /* Number of actual parameters in this call, including struct value addr. */
2232 int num_actuals;
2233 /* Number of named args. Args after this are anonymous ones
2234 and they must all go on the stack. */
2235 int n_named_args;
2236 /* Number of complex actual arguments that need to be split. */
2237 int num_complex_actuals = 0;
2239 /* Vector of information about each argument.
2240 Arguments are numbered in the order they will be pushed,
2241 not the order they are written. */
2242 struct arg_data *args;
2244 /* Total size in bytes of all the stack-parms scanned so far. */
2245 struct args_size args_size;
2246 struct args_size adjusted_args_size;
2247 /* Size of arguments before any adjustments (such as rounding). */
2248 int unadjusted_args_size;
2249 /* Data on reg parms scanned so far. */
2250 CUMULATIVE_ARGS args_so_far_v;
2251 cumulative_args_t args_so_far;
2252 /* Nonzero if a reg parm has been scanned. */
2253 int reg_parm_seen;
2254 /* Nonzero if this is an indirect function call. */
2256 /* Nonzero if we must avoid push-insns in the args for this call.
2257 If stack space is allocated for register parameters, but not by the
2258 caller, then it is preallocated in the fixed part of the stack frame.
2259 So the entire argument block must then be preallocated (i.e., we
2260 ignore PUSH_ROUNDING in that case). */
2262 int must_preallocate = !PUSH_ARGS;
2264 /* Size of the stack reserved for parameter registers. */
2265 int reg_parm_stack_space = 0;
2267 /* Address of space preallocated for stack parms
2268 (on machines that lack push insns), or 0 if space not preallocated. */
2269 rtx argblock = 0;
2271 /* Mask of ECF_ and ERF_ flags. */
2272 int flags = 0;
2273 int return_flags = 0;
2274 #ifdef REG_PARM_STACK_SPACE
2275 /* Define the boundary of the register parm stack space that needs to be
2276 saved, if any. */
2277 int low_to_save, high_to_save;
2278 rtx save_area = 0; /* Place that it is saved */
2279 #endif
2281 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
2282 char *initial_stack_usage_map = stack_usage_map;
2283 char *stack_usage_map_buf = NULL;
2285 int old_stack_allocated;
2287 /* State variables to track stack modifications. */
2288 rtx old_stack_level = 0;
2289 int old_stack_arg_under_construction = 0;
2290 int old_pending_adj = 0;
2291 int old_inhibit_defer_pop = inhibit_defer_pop;
2293 /* Some stack pointer alterations we make are performed via
2294 allocate_dynamic_stack_space. This modifies the stack_pointer_delta,
2295 which we then also need to save/restore along the way. */
2296 int old_stack_pointer_delta = 0;
2298 rtx call_fusage;
2299 tree addr = CALL_EXPR_FN (exp);
2300 int i;
2301 /* The alignment of the stack, in bits. */
2302 unsigned HOST_WIDE_INT preferred_stack_boundary;
2303 /* The alignment of the stack, in bytes. */
2304 unsigned HOST_WIDE_INT preferred_unit_stack_boundary;
2305 /* The static chain value to use for this call. */
2306 rtx static_chain_value;
2307 /* See if this is "nothrow" function call. */
2308 if (TREE_NOTHROW (exp))
2309 flags |= ECF_NOTHROW;
2311 /* See if we can find a DECL-node for the actual function, and get the
2312 function attributes (flags) from the function decl or type node. */
2313 fndecl = get_callee_fndecl (exp);
2314 if (fndecl)
2316 fntype = TREE_TYPE (fndecl);
2317 flags |= flags_from_decl_or_type (fndecl);
2318 return_flags |= decl_return_flags (fndecl);
2320 else
2322 fntype = TREE_TYPE (TREE_TYPE (addr));
2323 flags |= flags_from_decl_or_type (fntype);
2325 rettype = TREE_TYPE (exp);
2327 struct_value = targetm.calls.struct_value_rtx (fntype, 0);
2329 /* Warn if this value is an aggregate type,
2330 regardless of which calling convention we are using for it. */
2331 if (AGGREGATE_TYPE_P (rettype))
2332 warning (OPT_Waggregate_return, "function call has aggregate value");
2334 /* If the result of a non looping pure or const function call is
2335 ignored (or void), and none of its arguments are volatile, we can
2336 avoid expanding the call and just evaluate the arguments for
2337 side-effects. */
2338 if ((flags & (ECF_CONST | ECF_PURE))
2339 && (!(flags & ECF_LOOPING_CONST_OR_PURE))
2340 && (ignore || target == const0_rtx
2341 || TYPE_MODE (rettype) == VOIDmode))
2343 bool volatilep = false;
2344 tree arg;
2345 call_expr_arg_iterator iter;
2347 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
2348 if (TREE_THIS_VOLATILE (arg))
2350 volatilep = true;
2351 break;
2354 if (! volatilep)
2356 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
2357 expand_expr (arg, const0_rtx, VOIDmode, EXPAND_NORMAL);
2358 return const0_rtx;
2362 #ifdef REG_PARM_STACK_SPACE
2363 reg_parm_stack_space = REG_PARM_STACK_SPACE (!fndecl ? fntype : fndecl);
2364 #endif
2366 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
2367 && reg_parm_stack_space > 0 && PUSH_ARGS)
2368 must_preallocate = 1;
2370 /* Set up a place to return a structure. */
2372 /* Cater to broken compilers. */
2373 if (aggregate_value_p (exp, fntype))
2375 /* This call returns a big structure. */
2376 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
2378 #ifdef PCC_STATIC_STRUCT_RETURN
2380 pcc_struct_value = 1;
2382 #else /* not PCC_STATIC_STRUCT_RETURN */
2384 struct_value_size = int_size_in_bytes (rettype);
2386 if (target && MEM_P (target) && CALL_EXPR_RETURN_SLOT_OPT (exp))
2387 structure_value_addr = XEXP (target, 0);
2388 else
2390 /* For variable-sized objects, we must be called with a target
2391 specified. If we were to allocate space on the stack here,
2392 we would have no way of knowing when to free it. */
2393 rtx d = assign_temp (rettype, 1, 1);
2394 structure_value_addr = XEXP (d, 0);
2395 target = 0;
2398 #endif /* not PCC_STATIC_STRUCT_RETURN */
2401 /* Figure out the amount to which the stack should be aligned. */
2402 preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
2403 if (fndecl)
2405 struct cgraph_rtl_info *i = cgraph_rtl_info (fndecl);
2406 /* Without automatic stack alignment, we can't increase preferred
2407 stack boundary. With automatic stack alignment, it is
2408 unnecessary since unless we can guarantee that all callers will
2409 align the outgoing stack properly, callee has to align its
2410 stack anyway. */
2411 if (i
2412 && i->preferred_incoming_stack_boundary
2413 && i->preferred_incoming_stack_boundary < preferred_stack_boundary)
2414 preferred_stack_boundary = i->preferred_incoming_stack_boundary;
2417 /* Operand 0 is a pointer-to-function; get the type of the function. */
2418 funtype = TREE_TYPE (addr);
2419 gcc_assert (POINTER_TYPE_P (funtype));
2420 funtype = TREE_TYPE (funtype);
2422 /* Count whether there are actual complex arguments that need to be split
2423 into their real and imaginary parts. Munge the type_arg_types
2424 appropriately here as well. */
2425 if (targetm.calls.split_complex_arg)
2427 call_expr_arg_iterator iter;
2428 tree arg;
2429 FOR_EACH_CALL_EXPR_ARG (arg, iter, exp)
2431 tree type = TREE_TYPE (arg);
2432 if (type && TREE_CODE (type) == COMPLEX_TYPE
2433 && targetm.calls.split_complex_arg (type))
2434 num_complex_actuals++;
2436 type_arg_types = split_complex_types (TYPE_ARG_TYPES (funtype));
2438 else
2439 type_arg_types = TYPE_ARG_TYPES (funtype);
2441 if (flags & ECF_MAY_BE_ALLOCA)
2442 cfun->calls_alloca = 1;
2444 /* If struct_value_rtx is 0, it means pass the address
2445 as if it were an extra parameter. Put the argument expression
2446 in structure_value_addr_value. */
2447 if (structure_value_addr && struct_value == 0)
2449 /* If structure_value_addr is a REG other than
2450 virtual_outgoing_args_rtx, we can use always use it. If it
2451 is not a REG, we must always copy it into a register.
2452 If it is virtual_outgoing_args_rtx, we must copy it to another
2453 register in some cases. */
2454 rtx temp = (!REG_P (structure_value_addr)
2455 || (ACCUMULATE_OUTGOING_ARGS
2456 && stack_arg_under_construction
2457 && structure_value_addr == virtual_outgoing_args_rtx)
2458 ? copy_addr_to_reg (convert_memory_address
2459 (Pmode, structure_value_addr))
2460 : structure_value_addr);
2462 structure_value_addr_value =
2463 make_tree (build_pointer_type (TREE_TYPE (funtype)), temp);
2464 structure_value_addr_parm = 1;
2467 /* Count the arguments and set NUM_ACTUALS. */
2468 num_actuals =
2469 call_expr_nargs (exp) + num_complex_actuals + structure_value_addr_parm;
2471 /* Compute number of named args.
2472 First, do a raw count of the args for INIT_CUMULATIVE_ARGS. */
2474 if (type_arg_types != 0)
2475 n_named_args
2476 = (list_length (type_arg_types)
2477 /* Count the struct value address, if it is passed as a parm. */
2478 + structure_value_addr_parm);
2479 else
2480 /* If we know nothing, treat all args as named. */
2481 n_named_args = num_actuals;
2483 /* Start updating where the next arg would go.
2485 On some machines (such as the PA) indirect calls have a different
2486 calling convention than normal calls. The fourth argument in
2487 INIT_CUMULATIVE_ARGS tells the backend if this is an indirect call
2488 or not. */
2489 INIT_CUMULATIVE_ARGS (args_so_far_v, funtype, NULL_RTX, fndecl, n_named_args);
2490 args_so_far = pack_cumulative_args (&args_so_far_v);
2492 /* Now possibly adjust the number of named args.
2493 Normally, don't include the last named arg if anonymous args follow.
2494 We do include the last named arg if
2495 targetm.calls.strict_argument_naming() returns nonzero.
2496 (If no anonymous args follow, the result of list_length is actually
2497 one too large. This is harmless.)
2499 If targetm.calls.pretend_outgoing_varargs_named() returns
2500 nonzero, and targetm.calls.strict_argument_naming() returns zero,
2501 this machine will be able to place unnamed args that were passed
2502 in registers into the stack. So treat all args as named. This
2503 allows the insns emitting for a specific argument list to be
2504 independent of the function declaration.
2506 If targetm.calls.pretend_outgoing_varargs_named() returns zero,
2507 we do not have any reliable way to pass unnamed args in
2508 registers, so we must force them into memory. */
2510 if (type_arg_types != 0
2511 && targetm.calls.strict_argument_naming (args_so_far))
2513 else if (type_arg_types != 0
2514 && ! targetm.calls.pretend_outgoing_varargs_named (args_so_far))
2515 /* Don't include the last named arg. */
2516 --n_named_args;
2517 else
2518 /* Treat all args as named. */
2519 n_named_args = num_actuals;
2521 /* Make a vector to hold all the information about each arg. */
2522 args = XALLOCAVEC (struct arg_data, num_actuals);
2523 memset (args, 0, num_actuals * sizeof (struct arg_data));
2525 /* Build up entries in the ARGS array, compute the size of the
2526 arguments into ARGS_SIZE, etc. */
2527 initialize_argument_information (num_actuals, args, &args_size,
2528 n_named_args, exp,
2529 structure_value_addr_value, fndecl, fntype,
2530 args_so_far, reg_parm_stack_space,
2531 &old_stack_level, &old_pending_adj,
2532 &must_preallocate, &flags,
2533 &try_tail_call, CALL_FROM_THUNK_P (exp));
2535 if (args_size.var)
2536 must_preallocate = 1;
2538 /* Now make final decision about preallocating stack space. */
2539 must_preallocate = finalize_must_preallocate (must_preallocate,
2540 num_actuals, args,
2541 &args_size);
2543 /* If the structure value address will reference the stack pointer, we
2544 must stabilize it. We don't need to do this if we know that we are
2545 not going to adjust the stack pointer in processing this call. */
2547 if (structure_value_addr
2548 && (reg_mentioned_p (virtual_stack_dynamic_rtx, structure_value_addr)
2549 || reg_mentioned_p (virtual_outgoing_args_rtx,
2550 structure_value_addr))
2551 && (args_size.var
2552 || (!ACCUMULATE_OUTGOING_ARGS && args_size.constant)))
2553 structure_value_addr = copy_to_reg (structure_value_addr);
2555 /* Tail calls can make things harder to debug, and we've traditionally
2556 pushed these optimizations into -O2. Don't try if we're already
2557 expanding a call, as that means we're an argument. Don't try if
2558 there's cleanups, as we know there's code to follow the call. */
2560 if (currently_expanding_call++ != 0
2561 || !flag_optimize_sibling_calls
2562 || args_size.var
2563 || dbg_cnt (tail_call) == false)
2564 try_tail_call = 0;
2566 /* Rest of purposes for tail call optimizations to fail. */
2567 if (
2568 #ifdef HAVE_sibcall_epilogue
2569 !HAVE_sibcall_epilogue
2570 #else
2572 #endif
2573 || !try_tail_call
2574 /* Doing sibling call optimization needs some work, since
2575 structure_value_addr can be allocated on the stack.
2576 It does not seem worth the effort since few optimizable
2577 sibling calls will return a structure. */
2578 || structure_value_addr != NULL_RTX
2579 #ifdef REG_PARM_STACK_SPACE
2580 /* If outgoing reg parm stack space changes, we can not do sibcall. */
2581 || (OUTGOING_REG_PARM_STACK_SPACE (funtype)
2582 != OUTGOING_REG_PARM_STACK_SPACE (TREE_TYPE (current_function_decl)))
2583 || (reg_parm_stack_space != REG_PARM_STACK_SPACE (current_function_decl))
2584 #endif
2585 /* Check whether the target is able to optimize the call
2586 into a sibcall. */
2587 || !targetm.function_ok_for_sibcall (fndecl, exp)
2588 /* Functions that do not return exactly once may not be sibcall
2589 optimized. */
2590 || (flags & (ECF_RETURNS_TWICE | ECF_NORETURN))
2591 || TYPE_VOLATILE (TREE_TYPE (TREE_TYPE (addr)))
2592 /* If the called function is nested in the current one, it might access
2593 some of the caller's arguments, but could clobber them beforehand if
2594 the argument areas are shared. */
2595 || (fndecl && decl_function_context (fndecl) == current_function_decl)
2596 /* If this function requires more stack slots than the current
2597 function, we cannot change it into a sibling call.
2598 crtl->args.pretend_args_size is not part of the
2599 stack allocated by our caller. */
2600 || args_size.constant > (crtl->args.size
2601 - crtl->args.pretend_args_size)
2602 /* If the callee pops its own arguments, then it must pop exactly
2603 the same number of arguments as the current function. */
2604 || (targetm.calls.return_pops_args (fndecl, funtype, args_size.constant)
2605 != targetm.calls.return_pops_args (current_function_decl,
2606 TREE_TYPE (current_function_decl),
2607 crtl->args.size))
2608 || !lang_hooks.decls.ok_for_sibcall (fndecl))
2609 try_tail_call = 0;
2611 /* Check if caller and callee disagree in promotion of function
2612 return value. */
2613 if (try_tail_call)
2615 enum machine_mode caller_mode, caller_promoted_mode;
2616 enum machine_mode callee_mode, callee_promoted_mode;
2617 int caller_unsignedp, callee_unsignedp;
2618 tree caller_res = DECL_RESULT (current_function_decl);
2620 caller_unsignedp = TYPE_UNSIGNED (TREE_TYPE (caller_res));
2621 caller_mode = DECL_MODE (caller_res);
2622 callee_unsignedp = TYPE_UNSIGNED (TREE_TYPE (funtype));
2623 callee_mode = TYPE_MODE (TREE_TYPE (funtype));
2624 caller_promoted_mode
2625 = promote_function_mode (TREE_TYPE (caller_res), caller_mode,
2626 &caller_unsignedp,
2627 TREE_TYPE (current_function_decl), 1);
2628 callee_promoted_mode
2629 = promote_function_mode (TREE_TYPE (funtype), callee_mode,
2630 &callee_unsignedp,
2631 funtype, 1);
2632 if (caller_mode != VOIDmode
2633 && (caller_promoted_mode != callee_promoted_mode
2634 || ((caller_mode != caller_promoted_mode
2635 || callee_mode != callee_promoted_mode)
2636 && (caller_unsignedp != callee_unsignedp
2637 || GET_MODE_BITSIZE (caller_mode)
2638 < GET_MODE_BITSIZE (callee_mode)))))
2639 try_tail_call = 0;
2642 /* Ensure current function's preferred stack boundary is at least
2643 what we need. Stack alignment may also increase preferred stack
2644 boundary. */
2645 if (crtl->preferred_stack_boundary < preferred_stack_boundary)
2646 crtl->preferred_stack_boundary = preferred_stack_boundary;
2647 else
2648 preferred_stack_boundary = crtl->preferred_stack_boundary;
2650 preferred_unit_stack_boundary = preferred_stack_boundary / BITS_PER_UNIT;
2652 /* We want to make two insn chains; one for a sibling call, the other
2653 for a normal call. We will select one of the two chains after
2654 initial RTL generation is complete. */
2655 for (pass = try_tail_call ? 0 : 1; pass < 2; pass++)
2657 int sibcall_failure = 0;
2658 /* We want to emit any pending stack adjustments before the tail
2659 recursion "call". That way we know any adjustment after the tail
2660 recursion call can be ignored if we indeed use the tail
2661 call expansion. */
2662 saved_pending_stack_adjust save;
2663 rtx insns;
2664 rtx before_call, next_arg_reg, after_args;
2666 if (pass == 0)
2668 /* State variables we need to save and restore between
2669 iterations. */
2670 save_pending_stack_adjust (&save);
2672 if (pass)
2673 flags &= ~ECF_SIBCALL;
2674 else
2675 flags |= ECF_SIBCALL;
2677 /* Other state variables that we must reinitialize each time
2678 through the loop (that are not initialized by the loop itself). */
2679 argblock = 0;
2680 call_fusage = 0;
2682 /* Start a new sequence for the normal call case.
2684 From this point on, if the sibling call fails, we want to set
2685 sibcall_failure instead of continuing the loop. */
2686 start_sequence ();
2688 /* Don't let pending stack adjusts add up to too much.
2689 Also, do all pending adjustments now if there is any chance
2690 this might be a call to alloca or if we are expanding a sibling
2691 call sequence.
2692 Also do the adjustments before a throwing call, otherwise
2693 exception handling can fail; PR 19225. */
2694 if (pending_stack_adjust >= 32
2695 || (pending_stack_adjust > 0
2696 && (flags & ECF_MAY_BE_ALLOCA))
2697 || (pending_stack_adjust > 0
2698 && flag_exceptions && !(flags & ECF_NOTHROW))
2699 || pass == 0)
2700 do_pending_stack_adjust ();
2702 /* Precompute any arguments as needed. */
2703 if (pass)
2704 precompute_arguments (num_actuals, args);
2706 /* Now we are about to start emitting insns that can be deleted
2707 if a libcall is deleted. */
2708 if (pass && (flags & ECF_MALLOC))
2709 start_sequence ();
2711 if (pass == 0 && crtl->stack_protect_guard)
2712 stack_protect_epilogue ();
2714 adjusted_args_size = args_size;
2715 /* Compute the actual size of the argument block required. The variable
2716 and constant sizes must be combined, the size may have to be rounded,
2717 and there may be a minimum required size. When generating a sibcall
2718 pattern, do not round up, since we'll be re-using whatever space our
2719 caller provided. */
2720 unadjusted_args_size
2721 = compute_argument_block_size (reg_parm_stack_space,
2722 &adjusted_args_size,
2723 fndecl, fntype,
2724 (pass == 0 ? 0
2725 : preferred_stack_boundary));
2727 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
2729 /* The argument block when performing a sibling call is the
2730 incoming argument block. */
2731 if (pass == 0)
2733 argblock = crtl->args.internal_arg_pointer;
2734 argblock
2735 #ifdef STACK_GROWS_DOWNWARD
2736 = plus_constant (Pmode, argblock, crtl->args.pretend_args_size);
2737 #else
2738 = plus_constant (Pmode, argblock, -crtl->args.pretend_args_size);
2739 #endif
2740 stored_args_map = sbitmap_alloc (args_size.constant);
2741 bitmap_clear (stored_args_map);
2744 /* If we have no actual push instructions, or shouldn't use them,
2745 make space for all args right now. */
2746 else if (adjusted_args_size.var != 0)
2748 if (old_stack_level == 0)
2750 emit_stack_save (SAVE_BLOCK, &old_stack_level);
2751 old_stack_pointer_delta = stack_pointer_delta;
2752 old_pending_adj = pending_stack_adjust;
2753 pending_stack_adjust = 0;
2754 /* stack_arg_under_construction says whether a stack arg is
2755 being constructed at the old stack level. Pushing the stack
2756 gets a clean outgoing argument block. */
2757 old_stack_arg_under_construction = stack_arg_under_construction;
2758 stack_arg_under_construction = 0;
2760 argblock = push_block (ARGS_SIZE_RTX (adjusted_args_size), 0, 0);
2761 if (flag_stack_usage_info)
2762 current_function_has_unbounded_dynamic_stack_size = 1;
2764 else
2766 /* Note that we must go through the motions of allocating an argument
2767 block even if the size is zero because we may be storing args
2768 in the area reserved for register arguments, which may be part of
2769 the stack frame. */
2771 int needed = adjusted_args_size.constant;
2773 /* Store the maximum argument space used. It will be pushed by
2774 the prologue (if ACCUMULATE_OUTGOING_ARGS, or stack overflow
2775 checking). */
2777 if (needed > crtl->outgoing_args_size)
2778 crtl->outgoing_args_size = needed;
2780 if (must_preallocate)
2782 if (ACCUMULATE_OUTGOING_ARGS)
2784 /* Since the stack pointer will never be pushed, it is
2785 possible for the evaluation of a parm to clobber
2786 something we have already written to the stack.
2787 Since most function calls on RISC machines do not use
2788 the stack, this is uncommon, but must work correctly.
2790 Therefore, we save any area of the stack that was already
2791 written and that we are using. Here we set up to do this
2792 by making a new stack usage map from the old one. The
2793 actual save will be done by store_one_arg.
2795 Another approach might be to try to reorder the argument
2796 evaluations to avoid this conflicting stack usage. */
2798 /* Since we will be writing into the entire argument area,
2799 the map must be allocated for its entire size, not just
2800 the part that is the responsibility of the caller. */
2801 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
2802 needed += reg_parm_stack_space;
2804 #ifdef ARGS_GROW_DOWNWARD
2805 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
2806 needed + 1);
2807 #else
2808 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
2809 needed);
2810 #endif
2811 free (stack_usage_map_buf);
2812 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
2813 stack_usage_map = stack_usage_map_buf;
2815 if (initial_highest_arg_in_use)
2816 memcpy (stack_usage_map, initial_stack_usage_map,
2817 initial_highest_arg_in_use);
2819 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
2820 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
2821 (highest_outgoing_arg_in_use
2822 - initial_highest_arg_in_use));
2823 needed = 0;
2825 /* The address of the outgoing argument list must not be
2826 copied to a register here, because argblock would be left
2827 pointing to the wrong place after the call to
2828 allocate_dynamic_stack_space below. */
2830 argblock = virtual_outgoing_args_rtx;
2832 else
2834 if (inhibit_defer_pop == 0)
2836 /* Try to reuse some or all of the pending_stack_adjust
2837 to get this space. */
2838 needed
2839 = (combine_pending_stack_adjustment_and_call
2840 (unadjusted_args_size,
2841 &adjusted_args_size,
2842 preferred_unit_stack_boundary));
2844 /* combine_pending_stack_adjustment_and_call computes
2845 an adjustment before the arguments are allocated.
2846 Account for them and see whether or not the stack
2847 needs to go up or down. */
2848 needed = unadjusted_args_size - needed;
2850 if (needed < 0)
2852 /* We're releasing stack space. */
2853 /* ??? We can avoid any adjustment at all if we're
2854 already aligned. FIXME. */
2855 pending_stack_adjust = -needed;
2856 do_pending_stack_adjust ();
2857 needed = 0;
2859 else
2860 /* We need to allocate space. We'll do that in
2861 push_block below. */
2862 pending_stack_adjust = 0;
2865 /* Special case this because overhead of `push_block' in
2866 this case is non-trivial. */
2867 if (needed == 0)
2868 argblock = virtual_outgoing_args_rtx;
2869 else
2871 argblock = push_block (GEN_INT (needed), 0, 0);
2872 #ifdef ARGS_GROW_DOWNWARD
2873 argblock = plus_constant (Pmode, argblock, needed);
2874 #endif
2877 /* We only really need to call `copy_to_reg' in the case
2878 where push insns are going to be used to pass ARGBLOCK
2879 to a function call in ARGS. In that case, the stack
2880 pointer changes value from the allocation point to the
2881 call point, and hence the value of
2882 VIRTUAL_OUTGOING_ARGS_RTX changes as well. But might
2883 as well always do it. */
2884 argblock = copy_to_reg (argblock);
2889 if (ACCUMULATE_OUTGOING_ARGS)
2891 /* The save/restore code in store_one_arg handles all
2892 cases except one: a constructor call (including a C
2893 function returning a BLKmode struct) to initialize
2894 an argument. */
2895 if (stack_arg_under_construction)
2897 rtx push_size
2898 = GEN_INT (adjusted_args_size.constant
2899 + (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype
2900 : TREE_TYPE (fndecl))) ? 0
2901 : reg_parm_stack_space));
2902 if (old_stack_level == 0)
2904 emit_stack_save (SAVE_BLOCK, &old_stack_level);
2905 old_stack_pointer_delta = stack_pointer_delta;
2906 old_pending_adj = pending_stack_adjust;
2907 pending_stack_adjust = 0;
2908 /* stack_arg_under_construction says whether a stack
2909 arg is being constructed at the old stack level.
2910 Pushing the stack gets a clean outgoing argument
2911 block. */
2912 old_stack_arg_under_construction
2913 = stack_arg_under_construction;
2914 stack_arg_under_construction = 0;
2915 /* Make a new map for the new argument list. */
2916 free (stack_usage_map_buf);
2917 stack_usage_map_buf = XCNEWVEC (char, highest_outgoing_arg_in_use);
2918 stack_usage_map = stack_usage_map_buf;
2919 highest_outgoing_arg_in_use = 0;
2921 /* We can pass TRUE as the 4th argument because we just
2922 saved the stack pointer and will restore it right after
2923 the call. */
2924 allocate_dynamic_stack_space (push_size, 0,
2925 BIGGEST_ALIGNMENT, true);
2928 /* If argument evaluation might modify the stack pointer,
2929 copy the address of the argument list to a register. */
2930 for (i = 0; i < num_actuals; i++)
2931 if (args[i].pass_on_stack)
2933 argblock = copy_addr_to_reg (argblock);
2934 break;
2938 compute_argument_addresses (args, argblock, num_actuals);
2940 /* Perform stack alignment before the first push (the last arg). */
2941 if (argblock == 0
2942 && adjusted_args_size.constant > reg_parm_stack_space
2943 && adjusted_args_size.constant != unadjusted_args_size)
2945 /* When the stack adjustment is pending, we get better code
2946 by combining the adjustments. */
2947 if (pending_stack_adjust
2948 && ! inhibit_defer_pop)
2950 pending_stack_adjust
2951 = (combine_pending_stack_adjustment_and_call
2952 (unadjusted_args_size,
2953 &adjusted_args_size,
2954 preferred_unit_stack_boundary));
2955 do_pending_stack_adjust ();
2957 else if (argblock == 0)
2958 anti_adjust_stack (GEN_INT (adjusted_args_size.constant
2959 - unadjusted_args_size));
2961 /* Now that the stack is properly aligned, pops can't safely
2962 be deferred during the evaluation of the arguments. */
2963 NO_DEFER_POP;
2965 /* Record the maximum pushed stack space size. We need to delay
2966 doing it this far to take into account the optimization done
2967 by combine_pending_stack_adjustment_and_call. */
2968 if (flag_stack_usage_info
2969 && !ACCUMULATE_OUTGOING_ARGS
2970 && pass
2971 && adjusted_args_size.var == 0)
2973 int pushed = adjusted_args_size.constant + pending_stack_adjust;
2974 if (pushed > current_function_pushed_stack_size)
2975 current_function_pushed_stack_size = pushed;
2978 funexp = rtx_for_function_call (fndecl, addr);
2980 /* Figure out the register where the value, if any, will come back. */
2981 valreg = 0;
2982 if (TYPE_MODE (rettype) != VOIDmode
2983 && ! structure_value_addr)
2985 if (pcc_struct_value)
2986 valreg = hard_function_value (build_pointer_type (rettype),
2987 fndecl, NULL, (pass == 0));
2988 else
2989 valreg = hard_function_value (rettype, fndecl, fntype,
2990 (pass == 0));
2992 /* If VALREG is a PARALLEL whose first member has a zero
2993 offset, use that. This is for targets such as m68k that
2994 return the same value in multiple places. */
2995 if (GET_CODE (valreg) == PARALLEL)
2997 rtx elem = XVECEXP (valreg, 0, 0);
2998 rtx where = XEXP (elem, 0);
2999 rtx offset = XEXP (elem, 1);
3000 if (offset == const0_rtx
3001 && GET_MODE (where) == GET_MODE (valreg))
3002 valreg = where;
3006 /* Precompute all register parameters. It isn't safe to compute anything
3007 once we have started filling any specific hard regs. */
3008 precompute_register_parameters (num_actuals, args, &reg_parm_seen);
3010 if (CALL_EXPR_STATIC_CHAIN (exp))
3011 static_chain_value = expand_normal (CALL_EXPR_STATIC_CHAIN (exp));
3012 else
3013 static_chain_value = 0;
3015 #ifdef REG_PARM_STACK_SPACE
3016 /* Save the fixed argument area if it's part of the caller's frame and
3017 is clobbered by argument setup for this call. */
3018 if (ACCUMULATE_OUTGOING_ARGS && pass)
3019 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
3020 &low_to_save, &high_to_save);
3021 #endif
3023 /* Now store (and compute if necessary) all non-register parms.
3024 These come before register parms, since they can require block-moves,
3025 which could clobber the registers used for register parms.
3026 Parms which have partial registers are not stored here,
3027 but we do preallocate space here if they want that. */
3029 for (i = 0; i < num_actuals; i++)
3031 if (args[i].reg == 0 || args[i].pass_on_stack)
3033 rtx before_arg = get_last_insn ();
3035 /* We don't allow passing huge (> 2^30 B) arguments
3036 by value. It would cause an overflow later on. */
3037 if (adjusted_args_size.constant
3038 >= (1 << (HOST_BITS_PER_INT - 2)))
3040 sorry ("passing too large argument on stack");
3041 continue;
3044 if (store_one_arg (&args[i], argblock, flags,
3045 adjusted_args_size.var != 0,
3046 reg_parm_stack_space)
3047 || (pass == 0
3048 && check_sibcall_argument_overlap (before_arg,
3049 &args[i], 1)))
3050 sibcall_failure = 1;
3053 if (args[i].stack)
3054 call_fusage
3055 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[i].tree_value)),
3056 gen_rtx_USE (VOIDmode, args[i].stack),
3057 call_fusage);
3060 /* If we have a parm that is passed in registers but not in memory
3061 and whose alignment does not permit a direct copy into registers,
3062 make a group of pseudos that correspond to each register that we
3063 will later fill. */
3064 if (STRICT_ALIGNMENT)
3065 store_unaligned_arguments_into_pseudos (args, num_actuals);
3067 /* Now store any partially-in-registers parm.
3068 This is the last place a block-move can happen. */
3069 if (reg_parm_seen)
3070 for (i = 0; i < num_actuals; i++)
3071 if (args[i].partial != 0 && ! args[i].pass_on_stack)
3073 rtx before_arg = get_last_insn ();
3075 if (store_one_arg (&args[i], argblock, flags,
3076 adjusted_args_size.var != 0,
3077 reg_parm_stack_space)
3078 || (pass == 0
3079 && check_sibcall_argument_overlap (before_arg,
3080 &args[i], 1)))
3081 sibcall_failure = 1;
3084 /* If register arguments require space on the stack and stack space
3085 was not preallocated, allocate stack space here for arguments
3086 passed in registers. */
3087 if (OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl)))
3088 && !ACCUMULATE_OUTGOING_ARGS
3089 && must_preallocate == 0 && reg_parm_stack_space > 0)
3090 anti_adjust_stack (GEN_INT (reg_parm_stack_space));
3092 /* Pass the function the address in which to return a
3093 structure value. */
3094 if (pass != 0 && structure_value_addr && ! structure_value_addr_parm)
3096 structure_value_addr
3097 = convert_memory_address (Pmode, structure_value_addr);
3098 emit_move_insn (struct_value,
3099 force_reg (Pmode,
3100 force_operand (structure_value_addr,
3101 NULL_RTX)));
3103 if (REG_P (struct_value))
3104 use_reg (&call_fusage, struct_value);
3107 after_args = get_last_insn ();
3108 funexp = prepare_call_address (fndecl, funexp, static_chain_value,
3109 &call_fusage, reg_parm_seen, pass == 0);
3111 load_register_parameters (args, num_actuals, &call_fusage, flags,
3112 pass == 0, &sibcall_failure);
3114 /* Save a pointer to the last insn before the call, so that we can
3115 later safely search backwards to find the CALL_INSN. */
3116 before_call = get_last_insn ();
3118 /* Set up next argument register. For sibling calls on machines
3119 with register windows this should be the incoming register. */
3120 if (pass == 0)
3121 next_arg_reg = targetm.calls.function_incoming_arg (args_so_far,
3122 VOIDmode,
3123 void_type_node,
3124 true);
3125 else
3126 next_arg_reg = targetm.calls.function_arg (args_so_far,
3127 VOIDmode, void_type_node,
3128 true);
3130 if (pass == 1 && (return_flags & ERF_RETURNS_ARG))
3132 int arg_nr = return_flags & ERF_RETURN_ARG_MASK;
3133 arg_nr = num_actuals - arg_nr - 1;
3134 if (arg_nr >= 0
3135 && arg_nr < num_actuals
3136 && args[arg_nr].reg
3137 && valreg
3138 && REG_P (valreg)
3139 && GET_MODE (args[arg_nr].reg) == GET_MODE (valreg))
3140 call_fusage
3141 = gen_rtx_EXPR_LIST (TYPE_MODE (TREE_TYPE (args[arg_nr].tree_value)),
3142 gen_rtx_SET (VOIDmode, valreg, args[arg_nr].reg),
3143 call_fusage);
3145 /* All arguments and registers used for the call must be set up by
3146 now! */
3148 /* Stack must be properly aligned now. */
3149 gcc_assert (!pass
3150 || !(stack_pointer_delta % preferred_unit_stack_boundary));
3152 /* Generate the actual call instruction. */
3153 emit_call_1 (funexp, exp, fndecl, funtype, unadjusted_args_size,
3154 adjusted_args_size.constant, struct_value_size,
3155 next_arg_reg, valreg, old_inhibit_defer_pop, call_fusage,
3156 flags, args_so_far);
3158 if (flag_use_caller_save)
3160 rtx last, datum = NULL_RTX;
3161 if (fndecl != NULL_TREE)
3163 datum = XEXP (DECL_RTL (fndecl), 0);
3164 gcc_assert (datum != NULL_RTX
3165 && GET_CODE (datum) == SYMBOL_REF);
3167 last = last_call_insn ();
3168 add_reg_note (last, REG_CALL_DECL, datum);
3171 /* If the call setup or the call itself overlaps with anything
3172 of the argument setup we probably clobbered our call address.
3173 In that case we can't do sibcalls. */
3174 if (pass == 0
3175 && check_sibcall_argument_overlap (after_args, 0, 0))
3176 sibcall_failure = 1;
3178 /* If a non-BLKmode value is returned at the most significant end
3179 of a register, shift the register right by the appropriate amount
3180 and update VALREG accordingly. BLKmode values are handled by the
3181 group load/store machinery below. */
3182 if (!structure_value_addr
3183 && !pcc_struct_value
3184 && TYPE_MODE (rettype) != VOIDmode
3185 && TYPE_MODE (rettype) != BLKmode
3186 && REG_P (valreg)
3187 && targetm.calls.return_in_msb (rettype))
3189 if (shift_return_value (TYPE_MODE (rettype), false, valreg))
3190 sibcall_failure = 1;
3191 valreg = gen_rtx_REG (TYPE_MODE (rettype), REGNO (valreg));
3194 if (pass && (flags & ECF_MALLOC))
3196 rtx temp = gen_reg_rtx (GET_MODE (valreg));
3197 rtx last, insns;
3199 /* The return value from a malloc-like function is a pointer. */
3200 if (TREE_CODE (rettype) == POINTER_TYPE)
3201 mark_reg_pointer (temp, MALLOC_ABI_ALIGNMENT);
3203 emit_move_insn (temp, valreg);
3205 /* The return value from a malloc-like function can not alias
3206 anything else. */
3207 last = get_last_insn ();
3208 add_reg_note (last, REG_NOALIAS, temp);
3210 /* Write out the sequence. */
3211 insns = get_insns ();
3212 end_sequence ();
3213 emit_insn (insns);
3214 valreg = temp;
3217 /* For calls to `setjmp', etc., inform
3218 function.c:setjmp_warnings that it should complain if
3219 nonvolatile values are live. For functions that cannot
3220 return, inform flow that control does not fall through. */
3222 if ((flags & ECF_NORETURN) || pass == 0)
3224 /* The barrier must be emitted
3225 immediately after the CALL_INSN. Some ports emit more
3226 than just a CALL_INSN above, so we must search for it here. */
3228 rtx last = get_last_insn ();
3229 while (!CALL_P (last))
3231 last = PREV_INSN (last);
3232 /* There was no CALL_INSN? */
3233 gcc_assert (last != before_call);
3236 emit_barrier_after (last);
3238 /* Stack adjustments after a noreturn call are dead code.
3239 However when NO_DEFER_POP is in effect, we must preserve
3240 stack_pointer_delta. */
3241 if (inhibit_defer_pop == 0)
3243 stack_pointer_delta = old_stack_allocated;
3244 pending_stack_adjust = 0;
3248 /* If value type not void, return an rtx for the value. */
3250 if (TYPE_MODE (rettype) == VOIDmode
3251 || ignore)
3252 target = const0_rtx;
3253 else if (structure_value_addr)
3255 if (target == 0 || !MEM_P (target))
3257 target
3258 = gen_rtx_MEM (TYPE_MODE (rettype),
3259 memory_address (TYPE_MODE (rettype),
3260 structure_value_addr));
3261 set_mem_attributes (target, rettype, 1);
3264 else if (pcc_struct_value)
3266 /* This is the special C++ case where we need to
3267 know what the true target was. We take care to
3268 never use this value more than once in one expression. */
3269 target = gen_rtx_MEM (TYPE_MODE (rettype),
3270 copy_to_reg (valreg));
3271 set_mem_attributes (target, rettype, 1);
3273 /* Handle calls that return values in multiple non-contiguous locations.
3274 The Irix 6 ABI has examples of this. */
3275 else if (GET_CODE (valreg) == PARALLEL)
3277 if (target == 0)
3278 target = emit_group_move_into_temps (valreg);
3279 else if (rtx_equal_p (target, valreg))
3281 else if (GET_CODE (target) == PARALLEL)
3282 /* Handle the result of a emit_group_move_into_temps
3283 call in the previous pass. */
3284 emit_group_move (target, valreg);
3285 else
3286 emit_group_store (target, valreg, rettype,
3287 int_size_in_bytes (rettype));
3289 else if (target
3290 && GET_MODE (target) == TYPE_MODE (rettype)
3291 && GET_MODE (target) == GET_MODE (valreg))
3293 bool may_overlap = false;
3295 /* We have to copy a return value in a CLASS_LIKELY_SPILLED hard
3296 reg to a plain register. */
3297 if (!REG_P (target) || HARD_REGISTER_P (target))
3298 valreg = avoid_likely_spilled_reg (valreg);
3300 /* If TARGET is a MEM in the argument area, and we have
3301 saved part of the argument area, then we can't store
3302 directly into TARGET as it may get overwritten when we
3303 restore the argument save area below. Don't work too
3304 hard though and simply force TARGET to a register if it
3305 is a MEM; the optimizer is quite likely to sort it out. */
3306 if (ACCUMULATE_OUTGOING_ARGS && pass && MEM_P (target))
3307 for (i = 0; i < num_actuals; i++)
3308 if (args[i].save_area)
3310 may_overlap = true;
3311 break;
3314 if (may_overlap)
3315 target = copy_to_reg (valreg);
3316 else
3318 /* TARGET and VALREG cannot be equal at this point
3319 because the latter would not have
3320 REG_FUNCTION_VALUE_P true, while the former would if
3321 it were referring to the same register.
3323 If they refer to the same register, this move will be
3324 a no-op, except when function inlining is being
3325 done. */
3326 emit_move_insn (target, valreg);
3328 /* If we are setting a MEM, this code must be executed.
3329 Since it is emitted after the call insn, sibcall
3330 optimization cannot be performed in that case. */
3331 if (MEM_P (target))
3332 sibcall_failure = 1;
3335 else
3336 target = copy_to_reg (avoid_likely_spilled_reg (valreg));
3338 /* If we promoted this return value, make the proper SUBREG.
3339 TARGET might be const0_rtx here, so be careful. */
3340 if (REG_P (target)
3341 && TYPE_MODE (rettype) != BLKmode
3342 && GET_MODE (target) != TYPE_MODE (rettype))
3344 tree type = rettype;
3345 int unsignedp = TYPE_UNSIGNED (type);
3346 int offset = 0;
3347 enum machine_mode pmode;
3349 /* Ensure we promote as expected, and get the new unsignedness. */
3350 pmode = promote_function_mode (type, TYPE_MODE (type), &unsignedp,
3351 funtype, 1);
3352 gcc_assert (GET_MODE (target) == pmode);
3354 if ((WORDS_BIG_ENDIAN || BYTES_BIG_ENDIAN)
3355 && (GET_MODE_SIZE (GET_MODE (target))
3356 > GET_MODE_SIZE (TYPE_MODE (type))))
3358 offset = GET_MODE_SIZE (GET_MODE (target))
3359 - GET_MODE_SIZE (TYPE_MODE (type));
3360 if (! BYTES_BIG_ENDIAN)
3361 offset = (offset / UNITS_PER_WORD) * UNITS_PER_WORD;
3362 else if (! WORDS_BIG_ENDIAN)
3363 offset %= UNITS_PER_WORD;
3366 target = gen_rtx_SUBREG (TYPE_MODE (type), target, offset);
3367 SUBREG_PROMOTED_VAR_P (target) = 1;
3368 SUBREG_PROMOTED_UNSIGNED_SET (target, unsignedp);
3371 /* If size of args is variable or this was a constructor call for a stack
3372 argument, restore saved stack-pointer value. */
3374 if (old_stack_level)
3376 rtx prev = get_last_insn ();
3378 emit_stack_restore (SAVE_BLOCK, old_stack_level);
3379 stack_pointer_delta = old_stack_pointer_delta;
3381 fixup_args_size_notes (prev, get_last_insn (), stack_pointer_delta);
3383 pending_stack_adjust = old_pending_adj;
3384 old_stack_allocated = stack_pointer_delta - pending_stack_adjust;
3385 stack_arg_under_construction = old_stack_arg_under_construction;
3386 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
3387 stack_usage_map = initial_stack_usage_map;
3388 sibcall_failure = 1;
3390 else if (ACCUMULATE_OUTGOING_ARGS && pass)
3392 #ifdef REG_PARM_STACK_SPACE
3393 if (save_area)
3394 restore_fixed_argument_area (save_area, argblock,
3395 high_to_save, low_to_save);
3396 #endif
3398 /* If we saved any argument areas, restore them. */
3399 for (i = 0; i < num_actuals; i++)
3400 if (args[i].save_area)
3402 enum machine_mode save_mode = GET_MODE (args[i].save_area);
3403 rtx stack_area
3404 = gen_rtx_MEM (save_mode,
3405 memory_address (save_mode,
3406 XEXP (args[i].stack_slot, 0)));
3408 if (save_mode != BLKmode)
3409 emit_move_insn (stack_area, args[i].save_area);
3410 else
3411 emit_block_move (stack_area, args[i].save_area,
3412 GEN_INT (args[i].locate.size.constant),
3413 BLOCK_OP_CALL_PARM);
3416 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
3417 stack_usage_map = initial_stack_usage_map;
3420 /* If this was alloca, record the new stack level for nonlocal gotos.
3421 Check for the handler slots since we might not have a save area
3422 for non-local gotos. */
3424 if ((flags & ECF_MAY_BE_ALLOCA) && cfun->nonlocal_goto_save_area != 0)
3425 update_nonlocal_goto_save_area ();
3427 /* Free up storage we no longer need. */
3428 for (i = 0; i < num_actuals; ++i)
3429 free (args[i].aligned_regs);
3431 insns = get_insns ();
3432 end_sequence ();
3434 if (pass == 0)
3436 tail_call_insns = insns;
3438 /* Restore the pending stack adjustment now that we have
3439 finished generating the sibling call sequence. */
3441 restore_pending_stack_adjust (&save);
3443 /* Prepare arg structure for next iteration. */
3444 for (i = 0; i < num_actuals; i++)
3446 args[i].value = 0;
3447 args[i].aligned_regs = 0;
3448 args[i].stack = 0;
3451 sbitmap_free (stored_args_map);
3452 internal_arg_pointer_exp_state.scan_start = NULL_RTX;
3453 internal_arg_pointer_exp_state.cache.release ();
3455 else
3457 normal_call_insns = insns;
3459 /* Verify that we've deallocated all the stack we used. */
3460 gcc_assert ((flags & ECF_NORETURN)
3461 || (old_stack_allocated
3462 == stack_pointer_delta - pending_stack_adjust));
3465 /* If something prevents making this a sibling call,
3466 zero out the sequence. */
3467 if (sibcall_failure)
3468 tail_call_insns = NULL_RTX;
3469 else
3470 break;
3473 /* If tail call production succeeded, we need to remove REG_EQUIV notes on
3474 arguments too, as argument area is now clobbered by the call. */
3475 if (tail_call_insns)
3477 emit_insn (tail_call_insns);
3478 crtl->tail_call_emit = true;
3480 else
3481 emit_insn (normal_call_insns);
3483 currently_expanding_call--;
3485 free (stack_usage_map_buf);
3487 return target;
3490 /* A sibling call sequence invalidates any REG_EQUIV notes made for
3491 this function's incoming arguments.
3493 At the start of RTL generation we know the only REG_EQUIV notes
3494 in the rtl chain are those for incoming arguments, so we can look
3495 for REG_EQUIV notes between the start of the function and the
3496 NOTE_INSN_FUNCTION_BEG.
3498 This is (slight) overkill. We could keep track of the highest
3499 argument we clobber and be more selective in removing notes, but it
3500 does not seem to be worth the effort. */
3502 void
3503 fixup_tail_calls (void)
3505 rtx insn;
3507 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
3509 rtx note;
3511 /* There are never REG_EQUIV notes for the incoming arguments
3512 after the NOTE_INSN_FUNCTION_BEG note, so stop if we see it. */
3513 if (NOTE_P (insn)
3514 && NOTE_KIND (insn) == NOTE_INSN_FUNCTION_BEG)
3515 break;
3517 note = find_reg_note (insn, REG_EQUIV, 0);
3518 if (note)
3519 remove_note (insn, note);
3520 note = find_reg_note (insn, REG_EQUIV, 0);
3521 gcc_assert (!note);
3525 /* Traverse a list of TYPES and expand all complex types into their
3526 components. */
3527 static tree
3528 split_complex_types (tree types)
3530 tree p;
3532 /* Before allocating memory, check for the common case of no complex. */
3533 for (p = types; p; p = TREE_CHAIN (p))
3535 tree type = TREE_VALUE (p);
3536 if (TREE_CODE (type) == COMPLEX_TYPE
3537 && targetm.calls.split_complex_arg (type))
3538 goto found;
3540 return types;
3542 found:
3543 types = copy_list (types);
3545 for (p = types; p; p = TREE_CHAIN (p))
3547 tree complex_type = TREE_VALUE (p);
3549 if (TREE_CODE (complex_type) == COMPLEX_TYPE
3550 && targetm.calls.split_complex_arg (complex_type))
3552 tree next, imag;
3554 /* Rewrite complex type with component type. */
3555 TREE_VALUE (p) = TREE_TYPE (complex_type);
3556 next = TREE_CHAIN (p);
3558 /* Add another component type for the imaginary part. */
3559 imag = build_tree_list (NULL_TREE, TREE_VALUE (p));
3560 TREE_CHAIN (p) = imag;
3561 TREE_CHAIN (imag) = next;
3563 /* Skip the newly created node. */
3564 p = TREE_CHAIN (p);
3568 return types;
3571 /* Output a library call to function FUN (a SYMBOL_REF rtx).
3572 The RETVAL parameter specifies whether return value needs to be saved, other
3573 parameters are documented in the emit_library_call function below. */
3575 static rtx
3576 emit_library_call_value_1 (int retval, rtx orgfun, rtx value,
3577 enum libcall_type fn_type,
3578 enum machine_mode outmode, int nargs, va_list p)
3580 /* Total size in bytes of all the stack-parms scanned so far. */
3581 struct args_size args_size;
3582 /* Size of arguments before any adjustments (such as rounding). */
3583 struct args_size original_args_size;
3584 int argnum;
3585 rtx fun;
3586 /* Todo, choose the correct decl type of orgfun. Sadly this information
3587 isn't present here, so we default to native calling abi here. */
3588 tree fndecl ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
3589 tree fntype ATTRIBUTE_UNUSED = NULL_TREE; /* library calls default to host calling abi ? */
3590 int count;
3591 rtx argblock = 0;
3592 CUMULATIVE_ARGS args_so_far_v;
3593 cumulative_args_t args_so_far;
3594 struct arg
3596 rtx value;
3597 enum machine_mode mode;
3598 rtx reg;
3599 int partial;
3600 struct locate_and_pad_arg_data locate;
3601 rtx save_area;
3603 struct arg *argvec;
3604 int old_inhibit_defer_pop = inhibit_defer_pop;
3605 rtx call_fusage = 0;
3606 rtx mem_value = 0;
3607 rtx valreg;
3608 int pcc_struct_value = 0;
3609 int struct_value_size = 0;
3610 int flags;
3611 int reg_parm_stack_space = 0;
3612 int needed;
3613 rtx before_call;
3614 tree tfom; /* type_for_mode (outmode, 0) */
3616 #ifdef REG_PARM_STACK_SPACE
3617 /* Define the boundary of the register parm stack space that needs to be
3618 save, if any. */
3619 int low_to_save = 0, high_to_save = 0;
3620 rtx save_area = 0; /* Place that it is saved. */
3621 #endif
3623 /* Size of the stack reserved for parameter registers. */
3624 int initial_highest_arg_in_use = highest_outgoing_arg_in_use;
3625 char *initial_stack_usage_map = stack_usage_map;
3626 char *stack_usage_map_buf = NULL;
3628 rtx struct_value = targetm.calls.struct_value_rtx (0, 0);
3630 #ifdef REG_PARM_STACK_SPACE
3631 reg_parm_stack_space = REG_PARM_STACK_SPACE ((tree) 0);
3632 #endif
3634 /* By default, library functions can not throw. */
3635 flags = ECF_NOTHROW;
3637 switch (fn_type)
3639 case LCT_NORMAL:
3640 break;
3641 case LCT_CONST:
3642 flags |= ECF_CONST;
3643 break;
3644 case LCT_PURE:
3645 flags |= ECF_PURE;
3646 break;
3647 case LCT_NORETURN:
3648 flags |= ECF_NORETURN;
3649 break;
3650 case LCT_THROW:
3651 flags = ECF_NORETURN;
3652 break;
3653 case LCT_RETURNS_TWICE:
3654 flags = ECF_RETURNS_TWICE;
3655 break;
3657 fun = orgfun;
3659 /* Ensure current function's preferred stack boundary is at least
3660 what we need. */
3661 if (crtl->preferred_stack_boundary < PREFERRED_STACK_BOUNDARY)
3662 crtl->preferred_stack_boundary = PREFERRED_STACK_BOUNDARY;
3664 /* If this kind of value comes back in memory,
3665 decide where in memory it should come back. */
3666 if (outmode != VOIDmode)
3668 tfom = lang_hooks.types.type_for_mode (outmode, 0);
3669 if (aggregate_value_p (tfom, 0))
3671 #ifdef PCC_STATIC_STRUCT_RETURN
3672 rtx pointer_reg
3673 = hard_function_value (build_pointer_type (tfom), 0, 0, 0);
3674 mem_value = gen_rtx_MEM (outmode, pointer_reg);
3675 pcc_struct_value = 1;
3676 if (value == 0)
3677 value = gen_reg_rtx (outmode);
3678 #else /* not PCC_STATIC_STRUCT_RETURN */
3679 struct_value_size = GET_MODE_SIZE (outmode);
3680 if (value != 0 && MEM_P (value))
3681 mem_value = value;
3682 else
3683 mem_value = assign_temp (tfom, 1, 1);
3684 #endif
3685 /* This call returns a big structure. */
3686 flags &= ~(ECF_CONST | ECF_PURE | ECF_LOOPING_CONST_OR_PURE);
3689 else
3690 tfom = void_type_node;
3692 /* ??? Unfinished: must pass the memory address as an argument. */
3694 /* Copy all the libcall-arguments out of the varargs data
3695 and into a vector ARGVEC.
3697 Compute how to pass each argument. We only support a very small subset
3698 of the full argument passing conventions to limit complexity here since
3699 library functions shouldn't have many args. */
3701 argvec = XALLOCAVEC (struct arg, nargs + 1);
3702 memset (argvec, 0, (nargs + 1) * sizeof (struct arg));
3704 #ifdef INIT_CUMULATIVE_LIBCALL_ARGS
3705 INIT_CUMULATIVE_LIBCALL_ARGS (args_so_far_v, outmode, fun);
3706 #else
3707 INIT_CUMULATIVE_ARGS (args_so_far_v, NULL_TREE, fun, 0, nargs);
3708 #endif
3709 args_so_far = pack_cumulative_args (&args_so_far_v);
3711 args_size.constant = 0;
3712 args_size.var = 0;
3714 count = 0;
3716 push_temp_slots ();
3718 /* If there's a structure value address to be passed,
3719 either pass it in the special place, or pass it as an extra argument. */
3720 if (mem_value && struct_value == 0 && ! pcc_struct_value)
3722 rtx addr = XEXP (mem_value, 0);
3724 nargs++;
3726 /* Make sure it is a reasonable operand for a move or push insn. */
3727 if (!REG_P (addr) && !MEM_P (addr)
3728 && !(CONSTANT_P (addr)
3729 && targetm.legitimate_constant_p (Pmode, addr)))
3730 addr = force_operand (addr, NULL_RTX);
3732 argvec[count].value = addr;
3733 argvec[count].mode = Pmode;
3734 argvec[count].partial = 0;
3736 argvec[count].reg = targetm.calls.function_arg (args_so_far,
3737 Pmode, NULL_TREE, true);
3738 gcc_assert (targetm.calls.arg_partial_bytes (args_so_far, Pmode,
3739 NULL_TREE, 1) == 0);
3741 locate_and_pad_parm (Pmode, NULL_TREE,
3742 #ifdef STACK_PARMS_IN_REG_PARM_AREA
3744 #else
3745 argvec[count].reg != 0,
3746 #endif
3747 reg_parm_stack_space, 0,
3748 NULL_TREE, &args_size, &argvec[count].locate);
3750 if (argvec[count].reg == 0 || argvec[count].partial != 0
3751 || reg_parm_stack_space > 0)
3752 args_size.constant += argvec[count].locate.size.constant;
3754 targetm.calls.function_arg_advance (args_so_far, Pmode, (tree) 0, true);
3756 count++;
3759 for (; count < nargs; count++)
3761 rtx val = va_arg (p, rtx);
3762 enum machine_mode mode = (enum machine_mode) va_arg (p, int);
3763 int unsigned_p = 0;
3765 /* We cannot convert the arg value to the mode the library wants here;
3766 must do it earlier where we know the signedness of the arg. */
3767 gcc_assert (mode != BLKmode
3768 && (GET_MODE (val) == mode || GET_MODE (val) == VOIDmode));
3770 /* Make sure it is a reasonable operand for a move or push insn. */
3771 if (!REG_P (val) && !MEM_P (val)
3772 && !(CONSTANT_P (val) && targetm.legitimate_constant_p (mode, val)))
3773 val = force_operand (val, NULL_RTX);
3775 if (pass_by_reference (&args_so_far_v, mode, NULL_TREE, 1))
3777 rtx slot;
3778 int must_copy
3779 = !reference_callee_copied (&args_so_far_v, mode, NULL_TREE, 1);
3781 /* If this was a CONST function, it is now PURE since it now
3782 reads memory. */
3783 if (flags & ECF_CONST)
3785 flags &= ~ECF_CONST;
3786 flags |= ECF_PURE;
3789 if (MEM_P (val) && !must_copy)
3791 tree val_expr = MEM_EXPR (val);
3792 if (val_expr)
3793 mark_addressable (val_expr);
3794 slot = val;
3796 else
3798 slot = assign_temp (lang_hooks.types.type_for_mode (mode, 0),
3799 1, 1);
3800 emit_move_insn (slot, val);
3803 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
3804 gen_rtx_USE (VOIDmode, slot),
3805 call_fusage);
3806 if (must_copy)
3807 call_fusage = gen_rtx_EXPR_LIST (VOIDmode,
3808 gen_rtx_CLOBBER (VOIDmode,
3809 slot),
3810 call_fusage);
3812 mode = Pmode;
3813 val = force_operand (XEXP (slot, 0), NULL_RTX);
3816 mode = promote_function_mode (NULL_TREE, mode, &unsigned_p, NULL_TREE, 0);
3817 argvec[count].mode = mode;
3818 argvec[count].value = convert_modes (mode, GET_MODE (val), val, unsigned_p);
3819 argvec[count].reg = targetm.calls.function_arg (args_so_far, mode,
3820 NULL_TREE, true);
3822 argvec[count].partial
3823 = targetm.calls.arg_partial_bytes (args_so_far, mode, NULL_TREE, 1);
3825 if (argvec[count].reg == 0
3826 || argvec[count].partial != 0
3827 || reg_parm_stack_space > 0)
3829 locate_and_pad_parm (mode, NULL_TREE,
3830 #ifdef STACK_PARMS_IN_REG_PARM_AREA
3832 #else
3833 argvec[count].reg != 0,
3834 #endif
3835 reg_parm_stack_space, argvec[count].partial,
3836 NULL_TREE, &args_size, &argvec[count].locate);
3837 args_size.constant += argvec[count].locate.size.constant;
3838 gcc_assert (!argvec[count].locate.size.var);
3840 #ifdef BLOCK_REG_PADDING
3841 else
3842 /* The argument is passed entirely in registers. See at which
3843 end it should be padded. */
3844 argvec[count].locate.where_pad =
3845 BLOCK_REG_PADDING (mode, NULL_TREE,
3846 GET_MODE_SIZE (mode) <= UNITS_PER_WORD);
3847 #endif
3849 targetm.calls.function_arg_advance (args_so_far, mode, (tree) 0, true);
3852 /* If this machine requires an external definition for library
3853 functions, write one out. */
3854 assemble_external_libcall (fun);
3856 original_args_size = args_size;
3857 args_size.constant = (((args_size.constant
3858 + stack_pointer_delta
3859 + STACK_BYTES - 1)
3860 / STACK_BYTES
3861 * STACK_BYTES)
3862 - stack_pointer_delta);
3864 args_size.constant = MAX (args_size.constant,
3865 reg_parm_stack_space);
3867 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
3868 args_size.constant -= reg_parm_stack_space;
3870 if (args_size.constant > crtl->outgoing_args_size)
3871 crtl->outgoing_args_size = args_size.constant;
3873 if (flag_stack_usage_info && !ACCUMULATE_OUTGOING_ARGS)
3875 int pushed = args_size.constant + pending_stack_adjust;
3876 if (pushed > current_function_pushed_stack_size)
3877 current_function_pushed_stack_size = pushed;
3880 if (ACCUMULATE_OUTGOING_ARGS)
3882 /* Since the stack pointer will never be pushed, it is possible for
3883 the evaluation of a parm to clobber something we have already
3884 written to the stack. Since most function calls on RISC machines
3885 do not use the stack, this is uncommon, but must work correctly.
3887 Therefore, we save any area of the stack that was already written
3888 and that we are using. Here we set up to do this by making a new
3889 stack usage map from the old one.
3891 Another approach might be to try to reorder the argument
3892 evaluations to avoid this conflicting stack usage. */
3894 needed = args_size.constant;
3896 /* Since we will be writing into the entire argument area, the
3897 map must be allocated for its entire size, not just the part that
3898 is the responsibility of the caller. */
3899 if (! OUTGOING_REG_PARM_STACK_SPACE ((!fndecl ? fntype : TREE_TYPE (fndecl))))
3900 needed += reg_parm_stack_space;
3902 #ifdef ARGS_GROW_DOWNWARD
3903 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
3904 needed + 1);
3905 #else
3906 highest_outgoing_arg_in_use = MAX (initial_highest_arg_in_use,
3907 needed);
3908 #endif
3909 stack_usage_map_buf = XNEWVEC (char, highest_outgoing_arg_in_use);
3910 stack_usage_map = stack_usage_map_buf;
3912 if (initial_highest_arg_in_use)
3913 memcpy (stack_usage_map, initial_stack_usage_map,
3914 initial_highest_arg_in_use);
3916 if (initial_highest_arg_in_use != highest_outgoing_arg_in_use)
3917 memset (&stack_usage_map[initial_highest_arg_in_use], 0,
3918 highest_outgoing_arg_in_use - initial_highest_arg_in_use);
3919 needed = 0;
3921 /* We must be careful to use virtual regs before they're instantiated,
3922 and real regs afterwards. Loop optimization, for example, can create
3923 new libcalls after we've instantiated the virtual regs, and if we
3924 use virtuals anyway, they won't match the rtl patterns. */
3926 if (virtuals_instantiated)
3927 argblock = plus_constant (Pmode, stack_pointer_rtx,
3928 STACK_POINTER_OFFSET);
3929 else
3930 argblock = virtual_outgoing_args_rtx;
3932 else
3934 if (!PUSH_ARGS)
3935 argblock = push_block (GEN_INT (args_size.constant), 0, 0);
3938 /* We push args individually in reverse order, perform stack alignment
3939 before the first push (the last arg). */
3940 if (argblock == 0)
3941 anti_adjust_stack (GEN_INT (args_size.constant
3942 - original_args_size.constant));
3944 argnum = nargs - 1;
3946 #ifdef REG_PARM_STACK_SPACE
3947 if (ACCUMULATE_OUTGOING_ARGS)
3949 /* The argument list is the property of the called routine and it
3950 may clobber it. If the fixed area has been used for previous
3951 parameters, we must save and restore it. */
3952 save_area = save_fixed_argument_area (reg_parm_stack_space, argblock,
3953 &low_to_save, &high_to_save);
3955 #endif
3957 /* Push the args that need to be pushed. */
3959 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
3960 are to be pushed. */
3961 for (count = 0; count < nargs; count++, argnum--)
3963 enum machine_mode mode = argvec[argnum].mode;
3964 rtx val = argvec[argnum].value;
3965 rtx reg = argvec[argnum].reg;
3966 int partial = argvec[argnum].partial;
3967 unsigned int parm_align = argvec[argnum].locate.boundary;
3968 int lower_bound = 0, upper_bound = 0, i;
3970 if (! (reg != 0 && partial == 0))
3972 rtx use;
3974 if (ACCUMULATE_OUTGOING_ARGS)
3976 /* If this is being stored into a pre-allocated, fixed-size,
3977 stack area, save any previous data at that location. */
3979 #ifdef ARGS_GROW_DOWNWARD
3980 /* stack_slot is negative, but we want to index stack_usage_map
3981 with positive values. */
3982 upper_bound = -argvec[argnum].locate.slot_offset.constant + 1;
3983 lower_bound = upper_bound - argvec[argnum].locate.size.constant;
3984 #else
3985 lower_bound = argvec[argnum].locate.slot_offset.constant;
3986 upper_bound = lower_bound + argvec[argnum].locate.size.constant;
3987 #endif
3989 i = lower_bound;
3990 /* Don't worry about things in the fixed argument area;
3991 it has already been saved. */
3992 if (i < reg_parm_stack_space)
3993 i = reg_parm_stack_space;
3994 while (i < upper_bound && stack_usage_map[i] == 0)
3995 i++;
3997 if (i < upper_bound)
3999 /* We need to make a save area. */
4000 unsigned int size
4001 = argvec[argnum].locate.size.constant * BITS_PER_UNIT;
4002 enum machine_mode save_mode
4003 = mode_for_size (size, MODE_INT, 1);
4004 rtx adr
4005 = plus_constant (Pmode, argblock,
4006 argvec[argnum].locate.offset.constant);
4007 rtx stack_area
4008 = gen_rtx_MEM (save_mode, memory_address (save_mode, adr));
4010 if (save_mode == BLKmode)
4012 argvec[argnum].save_area
4013 = assign_stack_temp (BLKmode,
4014 argvec[argnum].locate.size.constant
4017 emit_block_move (validize_mem
4018 (copy_rtx (argvec[argnum].save_area)),
4019 stack_area,
4020 GEN_INT (argvec[argnum].locate.size.constant),
4021 BLOCK_OP_CALL_PARM);
4023 else
4025 argvec[argnum].save_area = gen_reg_rtx (save_mode);
4027 emit_move_insn (argvec[argnum].save_area, stack_area);
4032 emit_push_insn (val, mode, NULL_TREE, NULL_RTX, parm_align,
4033 partial, reg, 0, argblock,
4034 GEN_INT (argvec[argnum].locate.offset.constant),
4035 reg_parm_stack_space,
4036 ARGS_SIZE_RTX (argvec[argnum].locate.alignment_pad));
4038 /* Now mark the segment we just used. */
4039 if (ACCUMULATE_OUTGOING_ARGS)
4040 for (i = lower_bound; i < upper_bound; i++)
4041 stack_usage_map[i] = 1;
4043 NO_DEFER_POP;
4045 /* Indicate argument access so that alias.c knows that these
4046 values are live. */
4047 if (argblock)
4048 use = plus_constant (Pmode, argblock,
4049 argvec[argnum].locate.offset.constant);
4050 else
4051 /* When arguments are pushed, trying to tell alias.c where
4052 exactly this argument is won't work, because the
4053 auto-increment causes confusion. So we merely indicate
4054 that we access something with a known mode somewhere on
4055 the stack. */
4056 use = gen_rtx_PLUS (Pmode, virtual_outgoing_args_rtx,
4057 gen_rtx_SCRATCH (Pmode));
4058 use = gen_rtx_MEM (argvec[argnum].mode, use);
4059 use = gen_rtx_USE (VOIDmode, use);
4060 call_fusage = gen_rtx_EXPR_LIST (VOIDmode, use, call_fusage);
4064 argnum = nargs - 1;
4066 fun = prepare_call_address (NULL, fun, NULL, &call_fusage, 0, 0);
4068 /* Now load any reg parms into their regs. */
4070 /* ARGNUM indexes the ARGVEC array in the order in which the arguments
4071 are to be pushed. */
4072 for (count = 0; count < nargs; count++, argnum--)
4074 enum machine_mode mode = argvec[argnum].mode;
4075 rtx val = argvec[argnum].value;
4076 rtx reg = argvec[argnum].reg;
4077 int partial = argvec[argnum].partial;
4078 #ifdef BLOCK_REG_PADDING
4079 int size = 0;
4080 #endif
4082 /* Handle calls that pass values in multiple non-contiguous
4083 locations. The PA64 has examples of this for library calls. */
4084 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4085 emit_group_load (reg, val, NULL_TREE, GET_MODE_SIZE (mode));
4086 else if (reg != 0 && partial == 0)
4088 emit_move_insn (reg, val);
4089 #ifdef BLOCK_REG_PADDING
4090 size = GET_MODE_SIZE (argvec[argnum].mode);
4092 /* Copied from load_register_parameters. */
4094 /* Handle case where we have a value that needs shifting
4095 up to the msb. eg. a QImode value and we're padding
4096 upward on a BYTES_BIG_ENDIAN machine. */
4097 if (size < UNITS_PER_WORD
4098 && (argvec[argnum].locate.where_pad
4099 == (BYTES_BIG_ENDIAN ? upward : downward)))
4101 rtx x;
4102 int shift = (UNITS_PER_WORD - size) * BITS_PER_UNIT;
4104 /* Assigning REG here rather than a temp makes CALL_FUSAGE
4105 report the whole reg as used. Strictly speaking, the
4106 call only uses SIZE bytes at the msb end, but it doesn't
4107 seem worth generating rtl to say that. */
4108 reg = gen_rtx_REG (word_mode, REGNO (reg));
4109 x = expand_shift (LSHIFT_EXPR, word_mode, reg, shift, reg, 1);
4110 if (x != reg)
4111 emit_move_insn (reg, x);
4113 #endif
4116 NO_DEFER_POP;
4119 /* Any regs containing parms remain in use through the call. */
4120 for (count = 0; count < nargs; count++)
4122 rtx reg = argvec[count].reg;
4123 if (reg != 0 && GET_CODE (reg) == PARALLEL)
4124 use_group_regs (&call_fusage, reg);
4125 else if (reg != 0)
4127 int partial = argvec[count].partial;
4128 if (partial)
4130 int nregs;
4131 gcc_assert (partial % UNITS_PER_WORD == 0);
4132 nregs = partial / UNITS_PER_WORD;
4133 use_regs (&call_fusage, REGNO (reg), nregs);
4135 else
4136 use_reg (&call_fusage, reg);
4140 /* Pass the function the address in which to return a structure value. */
4141 if (mem_value != 0 && struct_value != 0 && ! pcc_struct_value)
4143 emit_move_insn (struct_value,
4144 force_reg (Pmode,
4145 force_operand (XEXP (mem_value, 0),
4146 NULL_RTX)));
4147 if (REG_P (struct_value))
4148 use_reg (&call_fusage, struct_value);
4151 /* Don't allow popping to be deferred, since then
4152 cse'ing of library calls could delete a call and leave the pop. */
4153 NO_DEFER_POP;
4154 valreg = (mem_value == 0 && outmode != VOIDmode
4155 ? hard_libcall_value (outmode, orgfun) : NULL_RTX);
4157 /* Stack must be properly aligned now. */
4158 gcc_assert (!(stack_pointer_delta
4159 & (PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT - 1)));
4161 before_call = get_last_insn ();
4163 /* We pass the old value of inhibit_defer_pop + 1 to emit_call_1, which
4164 will set inhibit_defer_pop to that value. */
4165 /* The return type is needed to decide how many bytes the function pops.
4166 Signedness plays no role in that, so for simplicity, we pretend it's
4167 always signed. We also assume that the list of arguments passed has
4168 no impact, so we pretend it is unknown. */
4170 emit_call_1 (fun, NULL,
4171 get_identifier (XSTR (orgfun, 0)),
4172 build_function_type (tfom, NULL_TREE),
4173 original_args_size.constant, args_size.constant,
4174 struct_value_size,
4175 targetm.calls.function_arg (args_so_far,
4176 VOIDmode, void_type_node, true),
4177 valreg,
4178 old_inhibit_defer_pop + 1, call_fusage, flags, args_so_far);
4180 if (flag_use_caller_save)
4182 rtx last, datum = orgfun;
4183 gcc_assert (GET_CODE (datum) == SYMBOL_REF);
4184 last = last_call_insn ();
4185 add_reg_note (last, REG_CALL_DECL, datum);
4188 /* Right-shift returned value if necessary. */
4189 if (!pcc_struct_value
4190 && TYPE_MODE (tfom) != BLKmode
4191 && targetm.calls.return_in_msb (tfom))
4193 shift_return_value (TYPE_MODE (tfom), false, valreg);
4194 valreg = gen_rtx_REG (TYPE_MODE (tfom), REGNO (valreg));
4197 /* For calls to `setjmp', etc., inform function.c:setjmp_warnings
4198 that it should complain if nonvolatile values are live. For
4199 functions that cannot return, inform flow that control does not
4200 fall through. */
4201 if (flags & ECF_NORETURN)
4203 /* The barrier note must be emitted
4204 immediately after the CALL_INSN. Some ports emit more than
4205 just a CALL_INSN above, so we must search for it here. */
4206 rtx last = get_last_insn ();
4207 while (!CALL_P (last))
4209 last = PREV_INSN (last);
4210 /* There was no CALL_INSN? */
4211 gcc_assert (last != before_call);
4214 emit_barrier_after (last);
4217 /* Consider that "regular" libcalls, i.e. all of them except for LCT_THROW
4218 and LCT_RETURNS_TWICE, cannot perform non-local gotos. */
4219 if (flags & ECF_NOTHROW)
4221 rtx last = get_last_insn ();
4222 while (!CALL_P (last))
4224 last = PREV_INSN (last);
4225 /* There was no CALL_INSN? */
4226 gcc_assert (last != before_call);
4229 make_reg_eh_region_note_nothrow_nononlocal (last);
4232 /* Now restore inhibit_defer_pop to its actual original value. */
4233 OK_DEFER_POP;
4235 pop_temp_slots ();
4237 /* Copy the value to the right place. */
4238 if (outmode != VOIDmode && retval)
4240 if (mem_value)
4242 if (value == 0)
4243 value = mem_value;
4244 if (value != mem_value)
4245 emit_move_insn (value, mem_value);
4247 else if (GET_CODE (valreg) == PARALLEL)
4249 if (value == 0)
4250 value = gen_reg_rtx (outmode);
4251 emit_group_store (value, valreg, NULL_TREE, GET_MODE_SIZE (outmode));
4253 else
4255 /* Convert to the proper mode if a promotion has been active. */
4256 if (GET_MODE (valreg) != outmode)
4258 int unsignedp = TYPE_UNSIGNED (tfom);
4260 gcc_assert (promote_function_mode (tfom, outmode, &unsignedp,
4261 fndecl ? TREE_TYPE (fndecl) : fntype, 1)
4262 == GET_MODE (valreg));
4263 valreg = convert_modes (outmode, GET_MODE (valreg), valreg, 0);
4266 if (value != 0)
4267 emit_move_insn (value, valreg);
4268 else
4269 value = valreg;
4273 if (ACCUMULATE_OUTGOING_ARGS)
4275 #ifdef REG_PARM_STACK_SPACE
4276 if (save_area)
4277 restore_fixed_argument_area (save_area, argblock,
4278 high_to_save, low_to_save);
4279 #endif
4281 /* If we saved any argument areas, restore them. */
4282 for (count = 0; count < nargs; count++)
4283 if (argvec[count].save_area)
4285 enum machine_mode save_mode = GET_MODE (argvec[count].save_area);
4286 rtx adr = plus_constant (Pmode, argblock,
4287 argvec[count].locate.offset.constant);
4288 rtx stack_area = gen_rtx_MEM (save_mode,
4289 memory_address (save_mode, adr));
4291 if (save_mode == BLKmode)
4292 emit_block_move (stack_area,
4293 validize_mem
4294 (copy_rtx (argvec[count].save_area)),
4295 GEN_INT (argvec[count].locate.size.constant),
4296 BLOCK_OP_CALL_PARM);
4297 else
4298 emit_move_insn (stack_area, argvec[count].save_area);
4301 highest_outgoing_arg_in_use = initial_highest_arg_in_use;
4302 stack_usage_map = initial_stack_usage_map;
4305 free (stack_usage_map_buf);
4307 return value;
4311 /* Output a library call to function FUN (a SYMBOL_REF rtx)
4312 (emitting the queue unless NO_QUEUE is nonzero),
4313 for a value of mode OUTMODE,
4314 with NARGS different arguments, passed as alternating rtx values
4315 and machine_modes to convert them to.
4317 FN_TYPE should be LCT_NORMAL for `normal' calls, LCT_CONST for
4318 `const' calls, LCT_PURE for `pure' calls, or other LCT_ value for
4319 other types of library calls. */
4321 void
4322 emit_library_call (rtx orgfun, enum libcall_type fn_type,
4323 enum machine_mode outmode, int nargs, ...)
4325 va_list p;
4327 va_start (p, nargs);
4328 emit_library_call_value_1 (0, orgfun, NULL_RTX, fn_type, outmode, nargs, p);
4329 va_end (p);
4332 /* Like emit_library_call except that an extra argument, VALUE,
4333 comes second and says where to store the result.
4334 (If VALUE is zero, this function chooses a convenient way
4335 to return the value.
4337 This function returns an rtx for where the value is to be found.
4338 If VALUE is nonzero, VALUE is returned. */
4341 emit_library_call_value (rtx orgfun, rtx value,
4342 enum libcall_type fn_type,
4343 enum machine_mode outmode, int nargs, ...)
4345 rtx result;
4346 va_list p;
4348 va_start (p, nargs);
4349 result = emit_library_call_value_1 (1, orgfun, value, fn_type, outmode,
4350 nargs, p);
4351 va_end (p);
4353 return result;
4356 /* Store a single argument for a function call
4357 into the register or memory area where it must be passed.
4358 *ARG describes the argument value and where to pass it.
4360 ARGBLOCK is the address of the stack-block for all the arguments,
4361 or 0 on a machine where arguments are pushed individually.
4363 MAY_BE_ALLOCA nonzero says this could be a call to `alloca'
4364 so must be careful about how the stack is used.
4366 VARIABLE_SIZE nonzero says that this was a variable-sized outgoing
4367 argument stack. This is used if ACCUMULATE_OUTGOING_ARGS to indicate
4368 that we need not worry about saving and restoring the stack.
4370 FNDECL is the declaration of the function we are calling.
4372 Return nonzero if this arg should cause sibcall failure,
4373 zero otherwise. */
4375 static int
4376 store_one_arg (struct arg_data *arg, rtx argblock, int flags,
4377 int variable_size ATTRIBUTE_UNUSED, int reg_parm_stack_space)
4379 tree pval = arg->tree_value;
4380 rtx reg = 0;
4381 int partial = 0;
4382 int used = 0;
4383 int i, lower_bound = 0, upper_bound = 0;
4384 int sibcall_failure = 0;
4386 if (TREE_CODE (pval) == ERROR_MARK)
4387 return 1;
4389 /* Push a new temporary level for any temporaries we make for
4390 this argument. */
4391 push_temp_slots ();
4393 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL))
4395 /* If this is being stored into a pre-allocated, fixed-size, stack area,
4396 save any previous data at that location. */
4397 if (argblock && ! variable_size && arg->stack)
4399 #ifdef ARGS_GROW_DOWNWARD
4400 /* stack_slot is negative, but we want to index stack_usage_map
4401 with positive values. */
4402 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
4403 upper_bound = -INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1)) + 1;
4404 else
4405 upper_bound = 0;
4407 lower_bound = upper_bound - arg->locate.size.constant;
4408 #else
4409 if (GET_CODE (XEXP (arg->stack_slot, 0)) == PLUS)
4410 lower_bound = INTVAL (XEXP (XEXP (arg->stack_slot, 0), 1));
4411 else
4412 lower_bound = 0;
4414 upper_bound = lower_bound + arg->locate.size.constant;
4415 #endif
4417 i = lower_bound;
4418 /* Don't worry about things in the fixed argument area;
4419 it has already been saved. */
4420 if (i < reg_parm_stack_space)
4421 i = reg_parm_stack_space;
4422 while (i < upper_bound && stack_usage_map[i] == 0)
4423 i++;
4425 if (i < upper_bound)
4427 /* We need to make a save area. */
4428 unsigned int size = arg->locate.size.constant * BITS_PER_UNIT;
4429 enum machine_mode save_mode = mode_for_size (size, MODE_INT, 1);
4430 rtx adr = memory_address (save_mode, XEXP (arg->stack_slot, 0));
4431 rtx stack_area = gen_rtx_MEM (save_mode, adr);
4433 if (save_mode == BLKmode)
4435 arg->save_area
4436 = assign_temp (TREE_TYPE (arg->tree_value), 1, 1);
4437 preserve_temp_slots (arg->save_area);
4438 emit_block_move (validize_mem (copy_rtx (arg->save_area)),
4439 stack_area,
4440 GEN_INT (arg->locate.size.constant),
4441 BLOCK_OP_CALL_PARM);
4443 else
4445 arg->save_area = gen_reg_rtx (save_mode);
4446 emit_move_insn (arg->save_area, stack_area);
4452 /* If this isn't going to be placed on both the stack and in registers,
4453 set up the register and number of words. */
4454 if (! arg->pass_on_stack)
4456 if (flags & ECF_SIBCALL)
4457 reg = arg->tail_call_reg;
4458 else
4459 reg = arg->reg;
4460 partial = arg->partial;
4463 /* Being passed entirely in a register. We shouldn't be called in
4464 this case. */
4465 gcc_assert (reg == 0 || partial != 0);
4467 /* If this arg needs special alignment, don't load the registers
4468 here. */
4469 if (arg->n_aligned_regs != 0)
4470 reg = 0;
4472 /* If this is being passed partially in a register, we can't evaluate
4473 it directly into its stack slot. Otherwise, we can. */
4474 if (arg->value == 0)
4476 /* stack_arg_under_construction is nonzero if a function argument is
4477 being evaluated directly into the outgoing argument list and
4478 expand_call must take special action to preserve the argument list
4479 if it is called recursively.
4481 For scalar function arguments stack_usage_map is sufficient to
4482 determine which stack slots must be saved and restored. Scalar
4483 arguments in general have pass_on_stack == 0.
4485 If this argument is initialized by a function which takes the
4486 address of the argument (a C++ constructor or a C function
4487 returning a BLKmode structure), then stack_usage_map is
4488 insufficient and expand_call must push the stack around the
4489 function call. Such arguments have pass_on_stack == 1.
4491 Note that it is always safe to set stack_arg_under_construction,
4492 but this generates suboptimal code if set when not needed. */
4494 if (arg->pass_on_stack)
4495 stack_arg_under_construction++;
4497 arg->value = expand_expr (pval,
4498 (partial
4499 || TYPE_MODE (TREE_TYPE (pval)) != arg->mode)
4500 ? NULL_RTX : arg->stack,
4501 VOIDmode, EXPAND_STACK_PARM);
4503 /* If we are promoting object (or for any other reason) the mode
4504 doesn't agree, convert the mode. */
4506 if (arg->mode != TYPE_MODE (TREE_TYPE (pval)))
4507 arg->value = convert_modes (arg->mode, TYPE_MODE (TREE_TYPE (pval)),
4508 arg->value, arg->unsignedp);
4510 if (arg->pass_on_stack)
4511 stack_arg_under_construction--;
4514 /* Check for overlap with already clobbered argument area. */
4515 if ((flags & ECF_SIBCALL)
4516 && MEM_P (arg->value)
4517 && mem_overlaps_already_clobbered_arg_p (XEXP (arg->value, 0),
4518 arg->locate.size.constant))
4519 sibcall_failure = 1;
4521 /* Don't allow anything left on stack from computation
4522 of argument to alloca. */
4523 if (flags & ECF_MAY_BE_ALLOCA)
4524 do_pending_stack_adjust ();
4526 if (arg->value == arg->stack)
4527 /* If the value is already in the stack slot, we are done. */
4529 else if (arg->mode != BLKmode)
4531 int size;
4532 unsigned int parm_align;
4534 /* Argument is a scalar, not entirely passed in registers.
4535 (If part is passed in registers, arg->partial says how much
4536 and emit_push_insn will take care of putting it there.)
4538 Push it, and if its size is less than the
4539 amount of space allocated to it,
4540 also bump stack pointer by the additional space.
4541 Note that in C the default argument promotions
4542 will prevent such mismatches. */
4544 size = GET_MODE_SIZE (arg->mode);
4545 /* Compute how much space the push instruction will push.
4546 On many machines, pushing a byte will advance the stack
4547 pointer by a halfword. */
4548 #ifdef PUSH_ROUNDING
4549 size = PUSH_ROUNDING (size);
4550 #endif
4551 used = size;
4553 /* Compute how much space the argument should get:
4554 round up to a multiple of the alignment for arguments. */
4555 if (none != FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)))
4556 used = (((size + PARM_BOUNDARY / BITS_PER_UNIT - 1)
4557 / (PARM_BOUNDARY / BITS_PER_UNIT))
4558 * (PARM_BOUNDARY / BITS_PER_UNIT));
4560 /* Compute the alignment of the pushed argument. */
4561 parm_align = arg->locate.boundary;
4562 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward)
4564 int pad = used - size;
4565 if (pad)
4567 unsigned int pad_align = (pad & -pad) * BITS_PER_UNIT;
4568 parm_align = MIN (parm_align, pad_align);
4572 /* This isn't already where we want it on the stack, so put it there.
4573 This can either be done with push or copy insns. */
4574 emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), NULL_RTX,
4575 parm_align, partial, reg, used - size, argblock,
4576 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
4577 ARGS_SIZE_RTX (arg->locate.alignment_pad));
4579 /* Unless this is a partially-in-register argument, the argument is now
4580 in the stack. */
4581 if (partial == 0)
4582 arg->value = arg->stack;
4584 else
4586 /* BLKmode, at least partly to be pushed. */
4588 unsigned int parm_align;
4589 int excess;
4590 rtx size_rtx;
4592 /* Pushing a nonscalar.
4593 If part is passed in registers, PARTIAL says how much
4594 and emit_push_insn will take care of putting it there. */
4596 /* Round its size up to a multiple
4597 of the allocation unit for arguments. */
4599 if (arg->locate.size.var != 0)
4601 excess = 0;
4602 size_rtx = ARGS_SIZE_RTX (arg->locate.size);
4604 else
4606 /* PUSH_ROUNDING has no effect on us, because emit_push_insn
4607 for BLKmode is careful to avoid it. */
4608 excess = (arg->locate.size.constant
4609 - int_size_in_bytes (TREE_TYPE (pval))
4610 + partial);
4611 size_rtx = expand_expr (size_in_bytes (TREE_TYPE (pval)),
4612 NULL_RTX, TYPE_MODE (sizetype),
4613 EXPAND_NORMAL);
4616 parm_align = arg->locate.boundary;
4618 /* When an argument is padded down, the block is aligned to
4619 PARM_BOUNDARY, but the actual argument isn't. */
4620 if (FUNCTION_ARG_PADDING (arg->mode, TREE_TYPE (pval)) == downward)
4622 if (arg->locate.size.var)
4623 parm_align = BITS_PER_UNIT;
4624 else if (excess)
4626 unsigned int excess_align = (excess & -excess) * BITS_PER_UNIT;
4627 parm_align = MIN (parm_align, excess_align);
4631 if ((flags & ECF_SIBCALL) && MEM_P (arg->value))
4633 /* emit_push_insn might not work properly if arg->value and
4634 argblock + arg->locate.offset areas overlap. */
4635 rtx x = arg->value;
4636 int i = 0;
4638 if (XEXP (x, 0) == crtl->args.internal_arg_pointer
4639 || (GET_CODE (XEXP (x, 0)) == PLUS
4640 && XEXP (XEXP (x, 0), 0) ==
4641 crtl->args.internal_arg_pointer
4642 && CONST_INT_P (XEXP (XEXP (x, 0), 1))))
4644 if (XEXP (x, 0) != crtl->args.internal_arg_pointer)
4645 i = INTVAL (XEXP (XEXP (x, 0), 1));
4647 /* expand_call should ensure this. */
4648 gcc_assert (!arg->locate.offset.var
4649 && arg->locate.size.var == 0
4650 && CONST_INT_P (size_rtx));
4652 if (arg->locate.offset.constant > i)
4654 if (arg->locate.offset.constant < i + INTVAL (size_rtx))
4655 sibcall_failure = 1;
4657 else if (arg->locate.offset.constant < i)
4659 /* Use arg->locate.size.constant instead of size_rtx
4660 because we only care about the part of the argument
4661 on the stack. */
4662 if (i < (arg->locate.offset.constant
4663 + arg->locate.size.constant))
4664 sibcall_failure = 1;
4666 else
4668 /* Even though they appear to be at the same location,
4669 if part of the outgoing argument is in registers,
4670 they aren't really at the same location. Check for
4671 this by making sure that the incoming size is the
4672 same as the outgoing size. */
4673 if (arg->locate.size.constant != INTVAL (size_rtx))
4674 sibcall_failure = 1;
4679 emit_push_insn (arg->value, arg->mode, TREE_TYPE (pval), size_rtx,
4680 parm_align, partial, reg, excess, argblock,
4681 ARGS_SIZE_RTX (arg->locate.offset), reg_parm_stack_space,
4682 ARGS_SIZE_RTX (arg->locate.alignment_pad));
4684 /* Unless this is a partially-in-register argument, the argument is now
4685 in the stack.
4687 ??? Unlike the case above, in which we want the actual
4688 address of the data, so that we can load it directly into a
4689 register, here we want the address of the stack slot, so that
4690 it's properly aligned for word-by-word copying or something
4691 like that. It's not clear that this is always correct. */
4692 if (partial == 0)
4693 arg->value = arg->stack_slot;
4696 if (arg->reg && GET_CODE (arg->reg) == PARALLEL)
4698 tree type = TREE_TYPE (arg->tree_value);
4699 arg->parallel_value
4700 = emit_group_load_into_temps (arg->reg, arg->value, type,
4701 int_size_in_bytes (type));
4704 /* Mark all slots this store used. */
4705 if (ACCUMULATE_OUTGOING_ARGS && !(flags & ECF_SIBCALL)
4706 && argblock && ! variable_size && arg->stack)
4707 for (i = lower_bound; i < upper_bound; i++)
4708 stack_usage_map[i] = 1;
4710 /* Once we have pushed something, pops can't safely
4711 be deferred during the rest of the arguments. */
4712 NO_DEFER_POP;
4714 /* Free any temporary slots made in processing this argument. */
4715 pop_temp_slots ();
4717 return sibcall_failure;
4720 /* Nonzero if we do not know how to pass TYPE solely in registers. */
4722 bool
4723 must_pass_in_stack_var_size (enum machine_mode mode ATTRIBUTE_UNUSED,
4724 const_tree type)
4726 if (!type)
4727 return false;
4729 /* If the type has variable size... */
4730 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
4731 return true;
4733 /* If the type is marked as addressable (it is required
4734 to be constructed into the stack)... */
4735 if (TREE_ADDRESSABLE (type))
4736 return true;
4738 return false;
4741 /* Another version of the TARGET_MUST_PASS_IN_STACK hook. This one
4742 takes trailing padding of a structure into account. */
4743 /* ??? Should be able to merge these two by examining BLOCK_REG_PADDING. */
4745 bool
4746 must_pass_in_stack_var_size_or_pad (enum machine_mode mode, const_tree type)
4748 if (!type)
4749 return false;
4751 /* If the type has variable size... */
4752 if (TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
4753 return true;
4755 /* If the type is marked as addressable (it is required
4756 to be constructed into the stack)... */
4757 if (TREE_ADDRESSABLE (type))
4758 return true;
4760 /* If the padding and mode of the type is such that a copy into
4761 a register would put it into the wrong part of the register. */
4762 if (mode == BLKmode
4763 && int_size_in_bytes (type) % (PARM_BOUNDARY / BITS_PER_UNIT)
4764 && (FUNCTION_ARG_PADDING (mode, type)
4765 == (BYTES_BIG_ENDIAN ? upward : downward)))
4766 return true;
4768 return false;